Вы находитесь на странице: 1из 328

Graduate Texts in Mathematics 120

Editorial Board
J.H. Ewing F.W. Gehring P.R. Halmos
Graduate Texts in Mathematics

I TAKEUWZARING. Introduction to Axiomatic Set Theory. 2nd ed


2 OXTOBY Measure and Category. 2nd cd
3 SCHIAEFFER. Topological Vector Spaces.
4 HILTON/STAMMBACH A Course in Homological Algebra
5 MAC LANE Categories for the Working Mathematician.
6 HUGHES/PIPER. Projective Planes.
7 StsRRF.. A Course in Arithmetic.
8 TAKEUTUZARING Axiomatic Set Theory.
9 HUMPHREYS Introduction to Lie Algebras and Representation Theory.
10 COHEN A Course in Simple Homotopy Theory.
I l CONWAY. Functions of One Complex Variable 2nd ed
12 BE.Ais. Advanced Mathematical Analysis
13 ANDERSON /FULLER Rings and Categories of Modules.
14 GoLUBITSKY/GUII_LEMIN. Stable Mappings and Their Singularities.
15 BERBERIAN Lectures in Functional Analysis and Operator Theory.
16 WINTER. The Structure of Fields.
17 ROSENBLATT_ Random Processes 2nd ed.
18 HALMOS. Measure Theory.
19 HALMOS. A Hilbert Space Problem Book 2nd ed., revised.
20 HUSEMrx_LEk. Fibre Bundles. 2nd ed
21 HUMPHREYS Linear Algebraic Groups.
22 BARNES/MACK. An Algebraic Introduction to Mathematical Logic
23 GREus. Linear Algebra. 4th ed
24 HOLMEs. Geometric Functional Analysis and its Applications.
25 HEwmYSTROMaERu. Real and Abstract Analysis.
26 MANES, Algebraic Theories
27 KELLEY General Topology.
28 ZARtsKIVSAMUtt.. Commutative Algebra. Vol. I.
29 ZARtSKIISAMUEL. Commutative Algebra. Vol. II.
30 JACOBSON. Lectures in Abstract Algebra I: Basic Concepts.
31 lacoBSON Lectures in Abstract Algebra 11: Linear Algebra.
32 lacoisoN. Lectures in Abstract Algebra III: Theory of Fields and Galois Theory.
33 Hu scH Differential Topology.
34 SPIT/Ht. Principles of Random Walk 2nd ed.
35 WERMER. Banach Algebras and Several Complex Va riables. 2nd cd.
36 KELLEY/NAMioxA et al. Linear Topological Spaces
37 MONK. Mathematical Logic.
38 GRAUERT/FRJTLSCHE. Several Complex Variables.
39 AavESON. An Invitation to C'-Algebras.
40 KEMENY/SNELL/KNAPP. Denumerable Markov Chains. 2nd cd
41 ArusroL.. Modular Functions and Dirichlet Series in Number Theory.
42 SERRE. Linear Representations of Finite Groups.
43 GELLMAtt/lERisoN. Rings of Continuous Functions
44 K i.ir e. Elementary Algebraic Geometry.
45 Loevr Probability Theory I. 4th cd.
46 LoEvE Probability Theory Il. 4th ed.
47 MOtsE. Geometric Topology in Dimensions 2 and 3.
fO4fYY44 :11101f7
William P. Ziemer

Weakly Differentiable
Functions
Sobolev Spaces and Functions of
Bounded Variation

Springer-Verlag
New York Berlin Heidelberg
London Pa ri s Tokyo Hong Kong
William P. Ziemer
Department of Mathematics
Indiana University
Bloomington, IN 47405
USA

Editorial Board
J. H. Ewing F. W. Gehring P. R. Halmos
Department of Department of Department of
Mathematics Mathematics Mathematics
Indiana University University of Michigan Santa Clar a University
Bloomington, IN 47405 Ann Arbor, MI 48109 Santa Clara, CA 95053
USA USA USA

With l illustration

Mathematics Subject Classifications (1980)• 46-E35, 26-830, 31-B15

Library of Congress Cataloging-in-Publication Data


Ziemer, William P.
Weakly differentiable functions: Sobolev spaces and functions of
bounded variation William P. Ziemer.
p. cm.—(Graduate texts in mathematics; 120)
Bibliography: p.
Includes index
ISBN 0-387.97017-7
1 Sobolev spaces. 2 Functions of bounded variation 1. Title
II. Series.
QA323 Z53 1989
515' .73—dc20 89-10072

Printed on acid-free paper.

C) 1989 by Springcr-Vcrlag New York Inc


All rights reserved This work may not be translated or copied in whole or in part without the
written permission of the publisher (Springer-Verlag, 175 Fifth Avenue. New York, NY 10010,
USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connec-
tion with any form of information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use of general descriptive names, trade names, trademarks, etc. to this publication, even if
the former are not especially identified, is not to be taken as a sign that such names, as under-
stood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely by
anyone

Camera-ready copy prepared using LaTEX.


Printed and bound by R R Donnelley & Sons, Harrisonburg, Virginia
Printed in the United States of America.

987654321

ISBN 0-3R7-97017-7 Springer-Verlag New York Berlin Heidelberg


ISBN 3-540-97017-7 Springer-Verlag Berlin Heidelberg New York
To Su z anne
Preface

The term "weakly differentiable functions" in the title refers to those inte-
grable functions defined on an open subset of R" whose partial derivatives
in the sense of distributions are either L" functions or (signed) measures
with finite total variation. The former class of functions comprises what
is now known as Sobolev spaces, though its origin, traceable to the early
1900s, predates the contributions by Sabolev. Both classes of functions,
Sobolev spaces and the space of functions of bounded variation (BV func-
tions), have undergone considerable development during the past 20 years.
From this development a rather complete theory has emerged and thus has
provided the main impetus for the writing of this book. Since these classes
of functions play a significant role in many fields, such as approximation
theory, calculus of variations, partial differential equations, and non-linear
potential theory, it is hoped that this monograph will be of assistance to a
wide range of graduate students and researchers in these and perhaps other
related areas. Some of the material in Chapters 1 - 4 has been presented in
a graduate course at Indiana University during the 1987-88 academic year,
and I am indebted to the students and colleagues in attendance for their
helpful comments and suggestions.
The major thrust of this book is the analysis of pointwise behavior of
Sobolev and BV functions. I have not attempted to develop Sobolev spaces
of fractional order which can be described in terms of Bessel potentials,
since this would require an effort beyond the scope of this book. Instead,
I concentrate on the analysis of spaces of integer order which is largely
accessible through real variable techniques, but does not totally exclude
the use of Bessel potentials. Indeed, the investigation of pointwise behavior
requires an analysis of certain exceptional sets and they can be conveniently
described in terms of elementary aspects of Bessel capacity.
The only prerequisite for the present volume is a standard graduate
course in real analysis, drawing especially from Lebesgue point theory and
measure theory. The material is organized in the following manner. Chap-
ter 1 is devoted to a review of those topics in real analysis that are needed
in the sequel. Included here is a brief overview of Lebesgue measure, L"
spaces, Hausdorff measure, and Schwartz distributions. Also included are
sections on covering theorems and Lorentz spaces--the latter being neces-
sary for a treatment of Sobolev inequalities in the case of critical indices.
Chapter 2 develops the basic properties of Sobolev spaces such as equiva-
lent formulations of Sobolev functions and their behavior under the opera-

viii Preface

Lions of truncation, composition, and change of variables. Also included is a


proof of the Sobolev inequality in its simplest form and the related Rellich-
Kondrachov Compactness Theorem. Alternate proofs of the Sobolev in-
equality are given, including the one which relates it to the isoperimetric
inequality and provides the best constant. Limiting cases of the Sobolev
inequality are discussed in the context of Lorentz spaces.
The remaining chapters are central to the book. Chapter 3 develops the
analysis of pointwise behavior of Sobolev functions. This includes a dis-
cussion of the continuity properties of functions with first derivatives in
D' in terms of Lehesgue points, approximate continuity, and fine conti-
nuity, as well as an analysis of differentiability properties of higher order
Sobolev functions by means of L"-derivatives. Here lies the foundation for
more delicate results, such as the comparison of I/-derivatives and dis-
tributional derivatives, and a result which provides an approximation for
Sobolev functions by smooth functions (in norm) that agree with the given
function everywhere except on sets whose complements have small capacity.
Chapter 4 develops an idea due to Norman Meyers. He observed that
the usual indirect proof of the Poincaré inequality could be used to es-
tablish a Poincaré-type inequality in an abstract setting. By appropriately
interpreting this inequality in various contexts, it yields virtually all known
inequalities of this genre. This general inequality contains a term which in-
volves an element of the dual of a Sobolev space. For many applications,
this term is taken as a measure; it therefore is of interest to know precisely
the class of measures contained in the dual of a given Sobolev space. For-
tunately, the Hedberg-Wolff theorem provides a characterization of such
measures.
The last chapter provides an analysis of the pointwise behavior of BV
functions in a manner that runs parallel to the development of Lebesgue
point theory for Sobolev functions in Chapter 3. While the Lebesgue point
theory for Sobolev functions is relatively easy to penetrate, the corre-
sponding development for BV functions is much more demanding. The
intricate nature of BV functions requires a more involved exposition than
does Sobolev functions, but at the same time reveals a rich and beautiful
structure which has its foundations in geometric measure theory. After the
structure of 13V functions has been developed, Chapter 5 returns to the
analysis of Poincaré inequalities for BV functions in the spirit developed
for Sobolev functions, which includes a characterization of measures that
belong to the dual of $V.
In order to place the text in better perspective, each chapter is con-
cluded with a section on historical notes which includes references to all
important and relatively new results. In addition to cited works, the Bib-
liography contains many other references related to the material in the
text. Bibliographical references are abbreviated in square brackets, such as
[DLL. Equation numbers appear in parentheses; theorems, lemmas, corollar-
ies,and remarks are numbered as a.b.c where b refers to section b in chapter
Preface ix

a, and section a.b refers to section b in chapter a.


I wish to thank David Adams, Robert Glassey, Tero Kilpelâinen,
Christoph Neugehauer, Edward Stredulinsky, Tevan Trent, and William
K. Ziemer for having critically read parts of the manuscript and supplied
many helpful suggestions and corrections.

WILLIAM P. ZIEMER
Contents

Preface vii

1 Preliminaries 1

1.1 Notation 1
Inner product of vectors
Support of a function
Boundary of a set
Distance from a point to a set
Characteristic function of a set
Multi-indices
Partial derivative operators
Function spaces—continuous, Holder continuous,
Holder continuous derivatives
1.2 Measures on R" 3
Lebesgue measurable sets
Lebesgue measurability of Borel sets
Sustin sets
1.3 Covering Theorems 7
Hausdorff maximal principle
General covering theorem
Vitali covering theorem
Covering lemma, with n-balls whose radii vary in
Lipschitzian way
Besicovitch covering lemma
Besicovitch differentiation theorem
1.4 Hausdorff Measure 15
Equivalence of Hausdorff and Lebesgue measures
Hausdorff dimension
1.5 LP-Spaces 18
Integration of a function via its distribution
function
Young's inequality
Holder's and Jensen's inequality
1.6 Regularization 21
LP-spaces and regularization
xii Contents

1.7 Distributions 23
Functions and measures, as distributions
Positive distributions
Distributions determined by their local behavior
Convolution of distributions
Differentiation of distributions
1.8 Lorentz Spaces 26
Non-increasing rearrangement of a function
Elementary properties of rearranged functions
Lorentz spaces
O'Neil's inequality, for rearranged functions
Equivalence of LP-norm and (p, p)-norm
Hardy's inequality
Inclusion relations of Lorentz spaces
Exercises 37
Historical Notes 39

2 Sobolev Spaces and Their Basic Properties 42

2.1 Weak Derivatives 42


Sobolev spaces
Absolute continuity on lines
L'-norin of difference quotients
Truncation of Sobolev functions
Composition of Sobolev functions
2.2 Change of Variables for Sobolev Functions 49
Rademacher's theorem
Bi-Lipschitzian change of variables
2.3 Approximation of Sobolev Functions by Smooth
Functions 53
Partition of unity
Smooth functions are dense in Wk P,

2.4 Sobolev Inequalities 55


Soholev's inequality
2.5 The Rellich-Kondrachov Compactness Theorem 61
Extension domains
2.6 Bessel Potentials and Capacity 64
Riesz and Bessel kernels
Bessel potentials
Bessel capacity
Basic properties of Bessel capacity
Capacitability of Suslin sets
Minimax theorem and alternate formulation of
Bessel capacity
Contents xiii

Metric properties of Bessel capacity


2.7 The Best Constant in the Sobolev Inequality 76
Co-area formula
Sobolev's inequality and isoperimetric inequality
2.8 Alternate Proofs of the Fundamental Inequalities 83
Hardy- Littlewood-Wiener maximal theorem
Sabolev's inequality for Riesz potentials
2.9 Limiting Cases of the Sobolev Inequality 88
The case kp = n by infinite series
The best constant in the case kp = n
An L°°-bound in the limiting case
2.10 Lorentz Spaces, A Slight Improvement 96
Young's inequality in the context of Lorentz spaces
Sobolev's inequality in Lorentz spaces
The limiting case
Exercises 103
Historical Notes 108

3 Pointwise Behavior of Sobolev Functions 112

3.1 Limits of Integral Averages of Sobolev Functions 1I2


Limiting values of integral averages except for
capacity null set
3.2 Densities of Measures 116
3.3 Lebesgue Points for Sobolev Functions 118
Existence of Lebesgue points except for capacity
null set
Approximate continuity
Fine continuity everywhere except for capacity null set
3.4 LP-Derivatives for Sobolev Functions 126
Existence of Taylor expansions D'
3.5 Properties of LP-Derivatives 130
The spaces T k , tk, Tk,p , t k, P
The implication of a function being in T k '' at all
points of a closed set
3.6 An LP-Version of the Whitney Extension Theorem 136
Existence of a CO° function comparable to the
distance function to a closed set
The Whitney extension theorem for functions in
Tk.P and tk.P
3.7 An Observation on Differentiation 142
3.8 Rademacher's Theorem in the LP-Context 145
A function in Tk.P everywhere implies it is in
t k 'p almost everywhere
xiv Contents


3.9 The Implications of Pointwise Differentiability 146
Comparison of L"-derivatives and distributional
derivatives
If u E tkPF(x) for every z, and if the
V'-derivatives are in LP, then u E Wk,P

3.10 A Lusin-Type Approximation for Sobolev Functions 153
Integral averages of Sobolev functions are uniformly
close to their limits on the complement of sets
of small capacity
Existence of smooth functions that agree with Sobolev
functions on the complement of sets of
small capacity

3.11 The Main Approximation 159
Existence of smooth functions that agree with
Sobolev functions on the complement of sets of
small capacity and are close in norm

Exercises 168

Historical Notes 175


4 Poincaré Inequalities—A Unified Approach 177

4.1 Inequalities in a General Setting 178
An abstract version of the Poincaré inequality

4.2 Applications to Sobolev Spaces 182
An interpolation inequality

4.3 The Dual of Wm'P(S2) 185
The representation of (W' (1))'

4.4 Some Measures in (Wâ '"(11))' 188
Poincaré inequalities derived from the abstract
version by identifying Lebesgue and Hausdorff
measure with elements in (Wm'P(SZ))"
The trace of Sobolev functions on the boundary of
Lipschitz domains
Poincaré inequalities involving the trace of
a Sobolev function

4.5 Poincaré Inequalities 193
Inequalities involving the capacity of the set on
which a function vanishes

4.6 Another Version of Poincaré's Inequality 196
An inequality involving dependence on the set on
which the function vanishes, not merely on its
capacity
4.7 More Measures in (Wm P(S2))e, 198
Soholev's inequality for Riesz potentials involving
Contents xv

measures other than Lebesgue measure


Characterization of measures in (Wm.P(R")).
4.8 Other Inequalities Involving Measures in (WkiP)` 207
Inequalities involving the restriction of Hausdorff
measure to lower dimensional manifolds
4.9 The Case p = 1 209
Inequalities involving the L'-norm of the gradient
Exercises 214
Historical Notes 217

5 Functions of Bounded Variation 220

5.1 Definitions 220


Definition of BV functions
The total variation measure IIDulI
5.2 Elementary Properties of 13V Functions 222
Lower seinicontinuity of the total variation measure
A condition ensuring continuity of the total -
variation measure
5.3 Regularization of BV Functions 224
Regularization does not increase the 13V norm
Approximation of BV functions by smooth functions
Compactness in L 1 of the unit ball in BV
5.4 Sets of Finite Perimeter 228
Definition of sets of finite perimeter
The perimeter of domains with smooth boundaries
Isoperimetric and relative isoperimetric inequality for
sets of finite perimeter
5.5 The Generalized Exterior Normal 233
A preliminary version of the Gauss-Green theorem
Density results at points of the reduced boundary
5.6 Tangential Properties of the Reduced Boundary and the
Measure-Theoretic Normal 237
Blow-up at a point of the reduced boundary
The measure-theoretic normal
The reduced boundary is contained in the
measure-theoretic boundary
A lower bound for the density of IIDXEii
Hausdorff measure restricted to the reduced boundary
is bounded above by IIDXEDD
5.7 Rectifiability of the Reduced Boundary 243
Countably (n - 1)-rectifiable sets
Countable (n - 1)-rectifiability of the
measure-theoretic boundary
xvi Contents


5.8 The Gauss-Green Theorem 246
The equivalence of the restriction of Hausdorff
measure to the measure-theoretic boundary
and if DXEII
The Gauss-Green theorem for sets of finite perimeter

5.9 Pointwise Behavior of BV Functions 249
Upper and lower approximate limits
The Boxing inequality
The set of approximate jump discontinuities

5.10 The Trace of a BV Function 255
The bounded extension of BV functions
Trace of a BV function defined in terms of the
upper and lower approximate limits of the
extended function
The integrability of the trace over the
measure-theoretic boundary

5.11 Sobolev-Type Inequalities for BV Functions 260
Inequalities involving elements in (BV(SZ))•

5.12 Inequalities Involving Capacity 262
Characterization of measure in (BV (SZ) )'
Poincaré inequality for BV functions

5.13 Generalizations to the Case p > 1 270

5.14 Trace Defined in Terms of Integral Averages 272

Exercises 277

Historical Notes 280

Bibliography 283

List of Symbols 297

Index 303
1

Preliminaries
Beyond the topics usually found in basic real analysis, virtually all of the
material found in this work is self-contained. In particular, most of the in-
formation contained in this chapter will be well-known by the reader and
therefore no attempt has been made to make a complete and thorough pre-
sentation. Rather, we merely introduce notation and develop a few concepts
that will be needed in the sequel.

1.1 Notation
Throughout, the symbol Si will generally denote an open set in Euclidean
space R" and 0 will designate the empty set. Points in - Rn are denoted by
z = (z,, ... , z n ), where z 1 E R 1 , 1 < i < n_ If x, y E R", the inner product
of r and y is

x • y= Exiy,
i=1
and the norm of z is
1x1 = (z • z)'/ 2 .
If u: SZ — R 1 is a function defined on SZ, the support of u is defined by

spt u = 1 n {z: u(x) 0},


where the closure of a set S C R" is denoted by S. If S C 11, S compact
and also 3 c 0, we shall write S cc SZ. The boundary of a set S is defined
by
ôS=S n (R n — S).
For E C R" and z E R", the distance from z to E is

d(z, E) = inf{Iz – yI . y E E }.
It is a simple exercise (see Exercise 1.I) to show that

Id(r, E) – d(y, E)I < Ix – yi

whenever z, y E Rn. The diameter of a set E C R" is defined by

diam(E) = sup{Ir – yI x, y E E),


1. Preliminaries

and the characteristic function E is denoted by xE. The symbol

B(x, r) _ {y : Ix — yi < r}
denotes the open ball with center x, radius r and

a(x, r) _ {p : tx -. yi < r}
will stand for the closed ball. We will use a(n) to denote the volume of the
ball of radius 1 in R. If a = (a l , ... ,a n ) is an n-tuple of non-negative
integers, a is called a multi-index and the length of a is
n

iai = E1=l ai.


If x = (x l ... , x n ) E Rte, we will let
,

^a = xa1 l • x2
2 •' •
^an
n

and a! = ar!a2! • • • a, a !. The partzal derivative operators are denoted by


D, = a fax, for 1 < i < n, and the higher order derivatives by

D° =D,'.. D = a x Q'...a x
1 nQR •
The gradient of a real-valued function u is denoted by

Du(x) = (D i u(x), ..., D n u(x)).


If k is a non-negative integer, we will sometimes use Dku to denote the
vector D k u = {Dau}^ a ^ = k.
We denote by C ° (1) the space of continuous functions on O. More gen-
erally, if k is a non-negative integer, possibly oo, let

C k (SZ) = {u : °
RL, Dau E C (St), p< lai < k },

Co (SZ) = Ck(Sa) fl {u : spt u compact, spt u C iZ),


and

Ck`(SZ) = ck(fZ) n {u : Da u has a continuous extension to SZ o < lcxi < k}•


,

Since Ii is open, a function u E C k (SZ) need not be bounded on U. However,


if u is bounded and uniformly continuous on fZ, then u can be uniquely
extended to a continuous function on i We will use C k (1; Rm) to denote
the class of functions u: f/ Rm defined on SZ whose coordinate functions
belong to C k (SZ). Similar notation is used for other function spaces whose
elements are vector-valued.
1.2. Measures on R71 3

if 0 < a < 1, we say that u is Hdlder continuous on SI with exponent a


if there is a constant C such that

Iu(x) — u(y)I < CIx — y1', x,y E S1.


We designate by C°'° (Sl) the space of all functions u satisfying this condi-
tion on a In case a = 1, the functions are called Lipschitz and the constant
C is denoted by Lip(u). For functions that possess some differentiability,
we let

ck k(Sl) = C°*°(S2) n { u : D/3 u E C"(S?), 0 < 101 < k).


Note that Ck•°(S2) is a Banach space when provided with the norm

sup sup IDA u(x) D^ü(g)i + max sup 1,603 u(x)I.


19I=k s,yEn Ix vI °<v31<k xEf


r#y

1.2 Measures on Rn
For the definition of Lebesgue outer measure, we consider closed n-dimen-
sional intervals
I = {x .a;< x < , i--1,...,n}
;

and their volumes


v(I) = ll (b,_ ai).
The Lebesgue outer measure of an arbitrary set E C R" is defined by
00
v( 4)
AEI = inf E
k=1
: E C J Ik , Ik an interval
k =1
(1.2.1)

A set E is said to be Lebesgue measurable if

BAI = IA n E1+IA n (R"—E)I (1.2.2)

whenever A C R".
The reader may consult a standard text on measure theory to find that
the Lebesgue measurable sets form a o- algebra, which we denote by A; that
is
(i) 0,R" E A.

(ii) If E 1 , Ea , . . . E A, then
00
UE 1 € A. (1.2.3)
;=i
1. Preliminaries

(iii) If E E .A, then R "- E E ,4.


Observe that these conditions also imply that A is also closed under count-
able intersections. It follows immediately from (1.2.2) that sets of measure
zero are measurable. Also recall that if E l , E2, ... are pairwise disjoint
measurable sets, then
00
E (1.2.4)
^Ei' .

i=1
Moreover, if E1 C E2 C ... are measurable, then
00
U;_i
Ei = ili rn'Ei I ^0o
(1.2.5)

and if E l D E2 D ..., then

fl Ed — lim 'Ed^
00

(1.2.6)
i^00
^^1
provided that lEkI < oo for some k.
Up to this point, we find that Lebesgue measure possesses many of the
continuity properties that are essential for fruitful applications in analysis.
However, at this stage we do not yet know whether the o-algebra, A, con-
tains a sufficiently rich supply of sets to be useful. This possible objection
is met by the following result.

1,2.1. Theorem. Each closed set C C R" is Lebesgue measurable.


In view of the fact that the Borel subsets of R" form the smallest a-
algebra that contains the closed sets, we have

1.2.2. Corollary. The Borel sets of R" are Lebesgue measurable.

Proof of Theorem 1.2.1. Because of the subadditivity of Lebesgue mea-


sure, it suffices to show that for a closed set C C R",
A[ > IA nC'+ 'A n (R" —C)) (1.2.7)
whenever A C W'. This will follow from the following property of Lebesgue
outer measure, which follows easily from (1.2.1):
'A U BI ='AÎ + IBI (1.2.8)
whenver A, B E R" with d(A,B) = inf {'x — y' : x E A, g E B} > 0. Indeed,
it is sufficient to establish that 'AU B1 > IA' + I B'. For this purpose, choose
e > O and let
00

A U B C U Ik where
k=I
1.2. Measures on R" 5
00

E v(4) < IAUBI + e. (1.2.9)

Because d(A, B) > 0, there exists disjoint open sets U and V such that

ACU, BCV. (1.2.10)

Clearly, the covering of AU B by {Ik} can be modified so that, for each k,

IkCUUV (1.2.11)

and that (1.2.9) still remains valid. However, (1.2.10) and (1.2.11) imply
00
> v(Ik) > IAI + 1BI.
^ -1

In order to prove (1.2.7), consider A C Rn with IAI < co and let Ct =


{x : d(x,C) < 1/i}. Note that
d(A —C=,A n C) > 0
and therefore, from (1.2.8),

lAI ? I(A — C;) u (A n C)I > IA — C;I + IA n CI. (1.2.12)

The proof of (1.2.7) will be concluded if we can show that

Urn IA—C,1—IA—CI
i-.00
Note that we cannot invoke (1.2.5) because it is not known that A — C; is
measurable since A is an arbitrary set, perhaps non-measurable. Let
1
T =An x. 1 <d(x,C)< (1.2.13)
+1
;

and note that since C is closed,

A— Ci(A—C)U (ÛT; )
(1.2.14)
^=1

which in turn, implies


00

I A - C I < I A- C,I (1.2.15)


i =i

Hence, the desired conclusion will follow if it can be shown that


00

>IT1I< oo. (1.2.16)


==1
6

1. Preliminaries

> 2. Thus, we

p
To establish this, first observe that d(11, Ti ) > 0 if Ii —
obtain from (1.2.8) that for each positive integer m,
m
Tz _ ^ iTzi l <— lAl < oo,
^

E T2i— U T21-1 <lAl<oo.


é =1 i-1
This establishes (1.2.16) and thus concludes the proof. ❑

1.2.3. Remark. Lebesgue measure and Hausdorff measure (which will be


introduced in Section 1.4) will meet most of the applications that occur
in this book, although in Chapter 5, it will be necessary to consider more
general measures. We say that is a measure on R" if assigns a non-
negative (possibly infinite) number to each subset of R" and p(0) = 0. It
is also accepted terminology to call such a set function an outer measure.
Following (1.2.2), a set E is called p-measurable if
µ(A) = p(A n E)+ p(An (R n — E))
whenever A C R" . A measure p on R" is called a Bores measure if every
Borel set is µ-measurable. A Borel measure with the properties that each
subset of R" is contained within a Borel set of equal p measure and that
p(K) < oo for each compact set K C R" is called a Radon measure.
Many outer measures defined on R n have the property that the Borel sets
are measurable. However, it is sometimes necessary to consider a larger a-
algebra of sets, namely, the Suslin sets, (often referred to as analytic sets).
They have the property of remaining invariant under continuous mappings
on R", a property not enjoyed by the Borel sets. The Suslin sets of R" can
be defined in the following manner. Let N denote the space of all infinite
sequences of positive integers topologized by the metric

j_
00
[! 2 -ilai b:l —

1+1a,— b,)

where (ai l and {b t ) are elements of


product topology. If
N. Let R" x N be endowed with the

p:Rn xN—.R"

is the projection defined by p(x,a) = x, then a Suslin set of 11" can be


defined as the image under p of some closed subset of R" x N.
The main reason for providing the preceding review of Lebesgue measure
is to compare its development with that of Hausdorff measure, which is
not as well known as Lebesgue measure but yet is extremely important in
geometric analysis and will play a significant role in the development of
this monograph.
1.3. Covering Theorems 7

1.3 Covering Theorems


Before discussing Hausdorff measure, it will be necessary to introduce sev-
eral important and useful covering theorems, the first of which is based on
the following implication of the Axiom of Choice.

Hausdorff Maximal Principle. If E is a family of sets (or a collection


of families of sets) and if {UF : F E .9 E E for any subfamily .F of E
with the property that

FCG or GCF whenever F,G E.F,

then there exists E C E which is maximal In the sense that it is not a subset
of any other member of E.

The following notation will be used. If B is a closed ball of radius r, let


B denote the closed ball concentric with B with radius 5r.

1.3.1. Theorem. Let ç be a family of closed balls with

R - sup{diamB : B E Q} < oo.

Then there is a subfamily .F C ç of pairwise disjoint elements such that

{UB: B E Ç}C{UB: B E .F} .

In fact, for each B E ç there exists B 1 E F such that B n B 1 o Q and


B C Bl .

Proof. We determine .F as follows. For j = 1, 2, ... let

ÿj = ç n B : B < diam B<---)


-1 ,
41 21

and observe that ç = U7 1 ç, . Now proceed to define Yi C çj inductively


as follows.
Let .F1 C ÇÇ1 be an arbitrary maximal subcollection of pairwise disjoint
elements. Such a collection exits by the Hausdorff maximal principle. As-
suming that .F 1 , F2 ,... , îi-1 have been chosen, let Yi be a maximal pair-
wise disjoint subcollection of
j-1
çj n B : B n B' =0 whenever B' E .F: . (1.3.1)

Thus, for each B E ç3, j ? 1, there exists B 1 E U; =1 .7; such that Bn B 1 o


0. For if not, the family F7 consisting of B along with all elements of .Fj
S 1. Preliminaries

would be a pairwise disjoint subcollection of (1.3.1), thus contradicting the


maximality of Moreover,

diam B< 1 =2R < 2 diem B 1


2 2
which implies that B C È. Thus,

{UB: BE ç;} C UB:B E U .^^ ,


i=i

and the conclusion holds by taking

u.F';.
00

.F= CI
i -1

1.3.2. Definition. A collection ç of closed balls is said to cover a set


E C R" finely if for each x E E and each e > 0, there exists B(x,r) E ç
and r < C.

1.3.3. Corollary. Let E C Rn be a set that is covered finely by g, where


Ç and ,F are as in Theorem 1.3.1. Then,

E -{UB:BE.r}c{UB:BEF-.F`}

for each finite collection .F' C

Proof. Since Rn -- {UB : B Er} is open, for each z E E - {UB : B E .F' }


there exists B E ç such that z E B and B n[{UB : B E .F'} = O. From
Theorem 1.3.1, there is B 1 E .F such that B fl B 1 0 0 and B 1 J B. Now
B 1 it F' since B n B 1 ,- 0 and therefore

xE{UB.BE .F^-^`}.

The next result addresses the question of determining an estimate for


the amount of overlap in a given family of closed balls. This will also be
considered in Theorem 1.3.5, but in the following we consider closed balls
whose radii vary in a Lipschitzian manner. The notation Lip(h) denotes
the Lipschitz constant of the mapping h.

1.3.4. Theorem. Let SCITC R" and suppose h : U -, (0, oo) is Lipschitz
with Lip(h) < A. Let a > 0, fi > 0 with as < 1 and a/3 < 1. Suppose the
collection of closed n-balls {B(a, h(s)) : s E S} is disjointed. Let

= S n Is : B(x, ah(z)) n B(s, /3h(sp 0 0).


1.3. Covering Theorems 9

Then

(1 - A0)/( 1 + Aa) < h(x)/h(s) < (1 + A03)/(1 - Act) (1.3.2)

whenever s E 52 and

card(S= ) < ja + 03+1)(1


1)(1 + Aa)(1 - )Q) -1 r n [(1 + A0)/(1 - »J))"

where card(S.) denotes the number of elements in Sz .

Proof. If s E Ss , then clearly ix - sf < ah(x) + aka) and therefore


!h(x) - h(s)I < -- si < Aah(x) + A/3h(s),

(1 - Afi)h(s) < (1 + Aa)h(x),


(1 - Arx)h(x) < (1 + A13)h(a). (1.3.3)
Now,

^x - + h(s) < ah(x) + (A + 1 )h(s)


< ah(x) + ( 0 + 1)[(1 + Aa) /(1 - A,Q)]h(x)
= ryh(x)

where ry = ar + (0+ 1)(1 + Aor)/(1 - Afl). Hence

B(s, h(s)) C B(x, ryh(x)) whenever s E Sx .

Since {B(s, h(s))) is a disjoint family,

L I$(8, h(s))I < IB(a, 7h(x))j


.ES:

or from (1.3.3)

card(SZ)a(n)[(l+Aa)(1-A0)-1h(x))" < E a(n)h(s)" < a(n)[7h(x))". Q


DES =

We now consider an arbitrary collection of closed balls and find a sub-


cover which is perhaps not disjoint, but whose elements have overlap which
is controlled.

1.3.5. Theorem. There is a positive number N > 1 depending only on n


so that any family B of dosed balls in R" whose cardinality is no less than
N and R = sup{r : B(a, r) E B) < oo contains disjointed subfamilies B i ,
82, ... , BN such that if A is the set of centers of balls in B, then

Ac U{uB:B E B{}.
i-1
1. Preliminaries

10

Proof.
Step I. Assume A is hounded.
Choose B 1 = B(a l , r l ) with r 1 > 1R. Assuming we have chosen B 1 , ...,
B,_, in 8 where j > 2 choose B, inductively as follows. 1f A3 = A N
U_ i Bs = 0, then the process stops and we set J = j. If A; 0, continue
by choosing Eli = B(a1, r,) E B so that a, E Ai and

r 2 > 4u s p{r : B(a,r) E 8 , a E A i l. (1.3.4)

If Ai 0 for all j, then we set J = +oo. In this case lirr, ? moo, r^ = 0


because A is bounded and the inequalities
r; 2 ri ri
Iai — ail>ri= — 3 + 2, for i<j,
3
imply that
{B(a1,ri/3) : 1 < j < J} is disjointed.

(1.3.5)
In case J < oo, we clearly have the inclusion
AC{UB,:1<j<J}.

(1.3.6)

This is also true in case J = +oo, for otherwise there would exist B(a, r) E
B with a E nl°_ 1 A l and an integer j with r, < 3r/4, contradicting the
choice of B1 .
Step II. We now prove there exists an integer M (depending only on n)
such that for each k with 1 < k < J, M exceeds the number of balls B i
with 1<i< k and B1 n Bk0 0.
First note that if r ; < 10rk, then
B(a„ ri /3) C B(ak, 15rk)
because if z E r; /3),
Iz—akI <Iz ail+Ias akI
— —

< lOrk /3 + r; + r k
< 43rk/3 < 15rk.
Hence, there are at most (60)" balls B; with
1<i< k, B, n Bk 0, and r i <IOrk
because, for each such i,
B(as , ri/3) C B(a k , 15rk),
and by (1.3.4) and (1.3.5)

IB(as, rs / 3 )I = iBi I ^ ( )r > IB1I ^ ()"


!

r 60n I B(a k , 15 rk)I-


1.3. Covering Theorems 11

To complete Step II, it remains to estimate the number of points in the set

I ={i . 1 <i< k,Bt f1 Bk 0,rt > IOrk}.

For this we first find an absolute lower bound on the angle between the two
vectors
a, — ak and a s — ak
corresponding to i,j E I with i < j. Assuming that this angle a < 7112,
2,
consider the triangle

and assume for notational convenience that rk = 1, d = lai — aki. Then

10< r i < lai — akI <r ; +1 and lai — ajI > ri

because j E I, a k Bs , BJ fl Bk 0, and as V B. Also

10 <rt < d< r t +1 < 3 rt +1

because j E I, ak it Bi , BJ fl Bk 0, and (1.3.4) applies to ri.


The law of cosines yields

ia i —akI 2 +d2 iai -- a;l 2 (r,+1) 2 +d 2 —r;


cos a = <

21ai — akld 2rtd


_ 2r,+1+d2 _ 1 1 d 1 1 4ri 1
gr i d d + — + 2rid 2ri < d + gr i d + 6r; + 2rt
1 1 4 1
<.822;
< 10+200++ 20
hence lcv > arccos .822 > O. Consequently, the rays determined by a s — ak
and a i — ak intersect the boundary of B(ak,1) at points that are separated
by a distance of at least /2(1 — cow). Since the boundary of B(ak, 1)
12 1. Preliminaries

has finite H"-l measure, the number of points in I is no more than some
constant depending only on n.
Step Ill. Choice of BI! ... , BM in case A is bounded.
With each positive integer j f we define an integer Aj such that Aj = j
whenever 1 <j ~ M and for j > M we define A;+l inductively as follows.
From Step II there is an integer A,;+I e {I, 2, ... , M} such that

Bj + 1 n {UBi : 1 < i < i, Ai = ,\ j + l} = 0.


Now deduce from (1.3.6) that the unions of the disjointed families

covers A.
Step IV. The case A is unbounded.
For each positive integer I, apply Step III with A replaced by Et =
A n {x : 3(t - l)R < Ixl < 31R} and 8 replaced by the subfamily Ct of 8
of balls with centers in Et. We obtain disjointed subfamilies 8f, ... ,B'M of
Ct such that
14
Et c U{UB : BE Bf}·
.=1
Since P n Q =0 whenever P E 8 t , Q E B'" and m > I + 2, the theorem
follows with
U8
00 00

8 1= U8
t~l
2t-l
1 , ... , 8 M=
t=1
2i
14 -
1

00 00

81.1+1 = U 8~t, . .. ,8214 = U 8~L


t=1 t~l

and N = 2M. o
We use this result to establish the following covering theorem which
contains the classical result of Vitali involving Lebesgue measure. An in-
teresting and novel aspect of the theorem is that the set A is not assumed
to be I-£-measurable. The thrust of the proof is that the previous theorem
allows 118 to obtain a disjoint 8ubfamily that provides a fixed percentage of
the 1-£ measure of the original set.

1.3.6. Theorem. Let ~be a Radon measure on R n and Juppose :F is a


family of closed ball3 that covers a set A C IF finely, where ~(A) < 00.
Then there exists a countable disjoint Bubfamily 9 of F such that

Jl(A - tuB : BEg}) = o.


1.3. Covering Theorems 13

Proof. Choose e > 0 so that e < 1/N, where N is the constant that appears
in the previous theorem. Then .F has disjointed subfamilies B,, ... , Bnr such
that

A C U{UB: BE B;}
i-,
and therefore

A(A) < p({u(A n B) : B E Bi)).

Thus, there exists 1 < k < N such that

p({U(A n B): B E > 1/N

which implies

p(A - {UB : B E Bk)) < (1— 1/N)p(A).

Hence, there is a finite subfamily Bk, of Bk such that

p(A - {UB : B E Bk 1 }) < (1 — 1/N + -€)14(A).

Now repeat this argument by replacing A with AL = 1 - {UB : B E Bk 1 }


and .F with .F E _ .F n {B : B n {UB : B E Bk,) = 0) to obtain a finite
disjointed subfamily Bk 2 of .F1 such that

p(A - {UB : B E BO) < (1 — 1/N + C);4111).


)

Thus,
p(A-{UB: B EBk, UBk 2 })<(1-1/N+e) 2 i(A).
Continue this process to obtain the conclusion of the theorem with
00
ç =U Bk, o
,-t

1.3.7. Lemma. Let p. and v be Radon measures on R". For each positive
number a let
Ea = x: sup ^ [$ T ))
(x ' >
r >0 l (x, )1
a .

Then, p(E a.) > av(Ea ).

Proof. By restricting our attention to bounded subsets of Ea , we may


assume that p(Ea ), v(Ea ) < oo. Let U D Ea be an open set. For e > 0
and for each z: E Ea , there exists a sequence of closed balls B(s, r 1 ) C U
withr i -'Osuchthat

µ[B(x, ri)] > (a + €)v[B(x , ri)].


14 1. Preliminaries

This produces a family .7" of closed balls that covers E„, finely. Hence, by
Theorem 1.3.6, there exists a disjoint subfamily Ç that covers v almost all
of E.. Consequently

(a + e)v(Ea ) < (a + e) E v(B) < E p(B) < p(U).


BEç BEST

Since E and U are arbitrary, the conclusion follows. D

If f is a continuous function, then the integral average of f over a ball of


small radius is nearly the same as the value of f at the center of the ball.
A remarkable result of real analysis states that this is true at (Lebesgue)
almost all points whenever f is integrable. The following result provides a
proof relative to any Radon measure. The notation

f(y) CLAW
B (s,r)

denotes
,IB(x, r)]-1 f (y) dµ01)•
B(z,r)

1.3.8. Theorem. Let p be a Radon measure on R" and f a locally Inte-


grable function on R" with respect to p. Then

]i m
r^d i
f (y) dp(y) = f (x)
B(= r)

for p almost all x E R" .

Proof. Note that

I (s,r) f (y)dp(v) — f (x) B( 2,r)


If (y) — g(y)Idgy)

+ I9(v) — f(x)Idµ(y)
B(z,r)
and if g is continuous, the last term converges to Ig(x) — f(x)I as r —, 0 .

Letting L(x) denote the upper limit of the term on the left, we obtain

L(x) < supI f(v) — 9(01001) +19(x) — f(x)I .


r >D B (s,r)

Hence,

{x : L(x) > a} C x : sup If(y) — g(y)Idp(y) > a / 2


r >ü B (s,r)
1.4. Hausdorff Measure 15

Li {x : Ig(x) - f (x)I > a/2},


and therefore, by the previous lemma,

p({x : L(x) > rx}) < 2/a1 f - gldN + 2/a I f - gidu•


R
Rn R
Since dR" If — gldp can be made arbitrarily small with appropriate choice
of g, cf. Section 1.6, it follows that µ({x : L(x) > a}) - 0 for each a > O.

1.3.9. Remark. If p and y are Radon measures with p absolutely con-


tinuous with respect to y, then the Radon-Nikodym theorem provides
f E L 1 (R", v) such that

p(E) = f (x) dv(x).


E

The results above show that the Radon-Nikodym derivative f can be taken
as the derivative of p with respect to v; that is,

h m 1` (B (x r)1 f(x)
r-0 v(B(x, r)J
,
[ (

for v almost all x E Rn.

1.4 Hausdorff Measure


The purpose here is to define a measure on R" that will assign a reason-
able notion of "length," "area" etc. to sets of appropriate dimension. For
example, if we would like to define the notion of length for an arbitrary set
E C R" , we might follow (1.2.1) and let

A(E) = inf E diam.4 : E C U A , . ; ;

; -i i =i
However, if we take n = 2 and E = {(t,sin(1/t)) : 0 < t < I}, it is easily
seen that A(E) < co whereas we should have À(E) = oo. The difficulty with
this definition is that the approximating sets A ; are not forced to follow
the geometry of the curve. This i8 changed in the following definition.

1.4.1. Definition. For each -y > 0, E > 0, and E C R", let


00 00

H£ (E) = infE a(7)2r"diarn(A;r : E C UA;, diam A< < c .


._1 ■• x
16 1. Preliminaries

Because HE (E) is non-decreasing in r, we may define the 7 dimensional


Hausdorif measure of E as

H7(E) = l âHgE)• (1.4.1)

In case 7 is a positive integer, or(1) denotes the volume of the unit ball
in R 7 . Otherwise, a(7) can be taken as an arbitrary positive constant.
The reason for requiring 41(1) to equal the volume of the unit ball in /V
when Y is a positive integer is to ensure that 1-1 7 (E) agrees with intuitive
notions of "-y-dimensional area" when E is a well-behaved set. For example,
it can be shown that H" agrees with the usual definition of n-dimensional
area on an n-dimensional C i submanifold of Rn+ k , k > O. More generally,
if f: Rn —. R n+k is a univalent, Lipschitz map and E C R n a Lebesgue
measurable set, then
= H n [ f (E)]
fE J
where J f is the square root of the sum of the squares of the n x n deter-
minants of the Jacobian matrix. The reader may consult [F4, Section 3.2]
for a thorough treatment of this subject. Here, we will merely show that
H defined on Rn is equal to Lebesgue measure.
"

1.4.2. Theorem. If E C Rn , then fi n (E) = 1E1.

Proof. First we show that

HE (E) < lEI for every e > O.


Consider the case where IEI = 0 and E is bounded. For each ii > 0, let
U E be an open set with WI < r). Since U is open, U can be written as
the union of closed balls, each of which has diameter less than e. Theorem
1.3.1 states that there is a subfamily .2 of pairwise disjoint elements such
that
U C {UB : $ E .^ }.

Therefore,

Ç HE (U) <
1/:(E)
1
E H£ (B.) < E 2-na(n)(diamBi)n
BEEF B; EN

E
B,E7
2 - "a(n)5 n (cliamB;)"

=5" E1B11
B,EF
< 5"IUI < 5 " 7/)
which proves that H"(E) = 0 since e and ri are arbitrary. The case when E
is unbounded is easily disposed of by considering E ( B(0,1), i = 1, 2, ....
1.4. Hausdorff Measure 17

Each of these sets has zero n-dimensional Hausdorff measure, and thus so
does E.
Now suppose E is an arbitrary set with IEI < oo. Let U J E be an open
set such that
lUl < IEI + >7. (1.4.2)
Appealing to Theorem 1.3.6, it is possible to find a family .7" of disjoint
closed balls B1, B2 i ..., such that U°_ 1 B ; C U, diatn Bi < e, i = 1, 2, ...,
and
00

E— B, —0. (1.4.3)
i=1
Let E' V U;_ 1 (EnBi ) and observe that E = (E— E')UE' with IE — E`l =
O. Now apply (1.4.1) and (1.4.2) to conclude that
00

H£ (E') < E 2a(n)(diam Bi n

i=1
00

DBil
.=1

UB
00

1=1
=lUl<1El + 17 .

Because e and n are arbitrary, it follows that H"(E') < IEj. However,
II" (E) < HIE —E')+H" (Ea ) with H" (E—E') = 0 because I E—E'I = O.
Therefore, H"(E) < 1E1.
In order to establish the opposite inequality, we will employ the isodi-
arnetric inequality which states that among all sets E C R" with a given
diameter, d, the ball with diameter d has the largest Lebesgue measure;
that is,
IEI < 2 — "a(n)(diarnE)'s (1.4.4)
whenever E C R. For a proof of this fact, see [F4, p. 197]. From this the
desired inequality follows immediately, for suppose
00

E 21 (n)(diam Ej) " <117(E) + rt


i=1

where E C U°2 1 Ei . Applying (1.4.3) to each E i yields

IEI < EIEl < E 2 -"a(ra)(diam Eir < IÎNE) +n,


1=1
18 1. Preliminaries

which implies, 1E1 < H"(E) silice e and i are arbitrary.

1.4.3. Remark. The reader can easily verify that the outer measure, H'',
has many properties in common with Lebesgue outer measure. For example,
(1.2.4), (1.2.5), and (1.2.6) are also valid for H'' as well as the analog
of Corollary 1.2.2. However, a striking difference between the two is that
'E) < oo whenever E is bounded whereas this may be false for H''(E). One
important ramification of this fact is the following. A Lebesgue measurable
set, E, can be characterized by the fact that for every e > 0, there exists
an open set U D E such that

lU — EE < e. (1.4.5)

This regularity property cannot hold in general for /P.


The fact that IP (E) may be possibly infinite for bounded sets E can be
put into better perspective by the following fact that the reader can easily
verify. For every set E, there is a non-negative number, d = d(E), such
that
H''(E) = 0 if 7 > d
H^ (E)= oo if 7< d.
The number d(E) is called the Hausdorff dimension of E.

Finally, we make note of the following elementary but useful fact. Sup-
pose f; R k —► R k +n is a Lipschitz map with Lip(f) = M. Then for any set
ECR k
Hk1 f (E)] < MH k (E). (1.4.6)
In particular, sets of zero k-dimensional Hausdorff measure remain invari-
ant under Lipschitz maps.

1.5 LP Spaces
For 1 < p < oo, E.,(11) will denote the space consisting of all measurable
functions on SI that are p"-power integrable on each compact subset of SZ.
LP(11) is the subspace of functions that are p th -power integrable on Si In
case the underlying measure is i rather than Lebesgue measure, we will
employ the notation L o,([; p.) and U(11, ,L) respectively. The norm on
LP(S1) is given by
1/p
lIt llp;c^ = (LI lulpdx (1.5.1)

and in case p — oc, it is defined as

iu ll^.n = esse sup lul (1.5.2)


1.5. L" Spaces 19

Analogous definitions are used in the case of LP(f2; p) and then the norm
is denoted by
IItIIp,N.S-
The notation f u(x) dx or sometimes simply f u dx will denote integration
f
with respect to Lebesgue measure and u dp the integral with respect to
the measure p. Strictly speaking, the elements of L"(11) are not functions
but rather equivalence classes of functions, where two functions are said
to be equivalent if they agree everywhere on SI except possibly for a set of
measure zero. The choice of a particular representative will be of special
importance later in Chapters 3 and 5 when the pointwise behavior of func-
tions in the spaces W k .P(SZ) and BV (0) is discussed. Recall from Theorem
1.3.8 that if u E L 1 (R" ), then for almost every zo E R", there is a number
z such that
u(y)dy --+ z as r 0+ ,
e(zo,r)
where f- denotes the integral average. We define u(xo) = z, and in this
way a canonical representative of u is determined. In those situations where
no confusion can occur, the elements of LP(n) will be regarded merely as
functions defined on O.
The following lemma is very useful and will be used frequently through-
out.

1.5.1. Lemma. If u > 0 is measurable, p > 0, and Et = {x : u(x) > t},


then 00 00
p (1.5.3)
u(x)pdx —IEtldtP ` tp-1lEtldt.
0
More generally, if µ is a measure defined on some a-algebra of Rn, u > 0
is a p-measurable function, and t is the countable union of sets of finite
p measure, then

j
00 00
updµ = N(Et)dtP = p (1.5.4)
n o

The proof of this can be obtained in at least two ways. One method is to
employ Fubini's Theorem on the product space SZ x [0, oo). Another is to
observe that (1.5.3) is immediate when u is a simple function. The general
case then follows by approximating u from below by simple functions.
The following algebraic and functional inequalities will be frequently used
throughout the course of this book.
Cauchy's inequality: if e > 0, a, b E R 1 , then

(1.5.5)
lab! < la^^ +
2E Ib;^
1. Preliminaries

20

and more generally, Young's inequality:

IeaI + [b/ e }p ,
IabI < (1.5.6)
P P'
where p > 1 and lip + 1 /p' = 1.
From Young's inequality follows Holder's inequality

I uv dx < ;fi, p > 1, (1.5.7)

which holds for functions u E L'(1), t1 E LP (ft). In case p = 1, we '

take p' = co and IIvIIp ;n = esse sup M. Holder's inequality can be ex-
,

tended to the case of k functions, uh ... uk lying respectively in spaces


,

L1" (0), ... ,17k (SZ) where


^` 1
(1.5.8)

By an induction argument and (1.5.7) it follows that

I u l . . . ukdx 5_1101 ;0 . • • Iludlpkst- (1.5.9)

One important application of (1.5.7) is Minkowski's inequality, which states


that (1.5.3) yields a norm on 11(11). That is,

lIu + vIIp;fl <— Iiullp;n + IIt1Ilp;1 (1.5.10)

for p > 1. Employing the notation

fo u dx = ISZI -f u dx,
n
another consequence of Holder's inequality is
1/p 1/q
u p dx < u'dx (1.5.11)
n n
whenever 1 < p < q and SZ C R" a measurable set with ISZI < oo.
We also recall Jensen's inequality whose statement involves the notion
of a convex function. A function A: R" —. R I is said to be convex if

A[(1 — t)xt + tx2J < (1 — t)A(xl) + tA(x2),


whenever x , z2 E R" and 0 < t < 1. Jensen's inequality states that if A is
a convex function on R" and E C R" a bounded measurable set, then

n f(x)dx < A(f (x)1^ (1.5.12)


E E
1.6. Regularization 21

whenever f E L' (E).


A further consequence of Holder's inequality is

I11411pIlullr's u E Lr(o), (1.5.13)

where p < q < r, and 1/q _ A/p+(1—A)/r. In order to see this, let a = aq,
= (1 — A)q and apply Holder's inequality to obtain

lulQdx = lu l° lu rdx < ula rdZ (10 lulf^V dZ 1/3r


1
T n I

where z = p/Aq and y = r/(1 A)q.


When endowed with the norm defined in (1.5.1), L 3'(11), 1 < p < oo, _
is a Banach space; that is, a complete, linear apace. If 1 < p < oo, it is
also separable. The normed dual of V'(1) consists of all bounded linear
functionals on LP(0) and is isometric to L" (f1) provided p < oo. Hence,
L" (0) is reflexive for 1 < p < oo. We recall the following fundamental result
concerning reflexive Banach spaces, which is of considerable importance in
the case of L"(1).

1.5.2. Theorem. A Banach space is reflexive if and only if its closed urrit
ball is weakly sequentially compact.

1.6 Regularization
Let cp be a non-negative, real-valued function in Co (R") with the property
that
cp(x)dt = 1, spt cp C B(0, 1), (1.6.1)

An example of such a function is given by

^(^) _ Cexp[-1/(1 — 1x1 2 )1 if 1x1 < 1


(1.6.2)
0 ifixl>x

where C is chosen so that fR „ cp = 1. For e > 0, the function (Mx)


e ncp(x/e) belongs to Co (R") and spt C B(0, c). (p c is called a regular-

izer (or mollifier) and the convolution

ue (x) = SPE * u(x) = cpc(x y)u(y)dy (1.6.3)



Rn

defined for functions u for which the right side of (1.6.3) liar meaning,
is called the regularization (mollification) of u. Regularization has several
important and useful properties that are summarized in the following the-
orem.
22 1. Preliminaries

1.8.1. Theorem.
(i) If u E Ltoc (R"), then for every e > 0, u E E CO°(R") and D°'(cp E *u)
(Dav e ) * u for each multi-index a.

(ii) n c (x) u(x) whenever r as a Lebesgue point for u. In case u is


continuous then u c converges uniformly to u on compact subsets of
R'
(iii) If u E !J(R"), 1 < p < oo, then o f E L"(Rn), Iluelip < Ilulip, and
lime —.0 — ullp = O.

Proof. For the proof of (i), it suffices to consider Iarl 1, since the case of
general a can be treated by induction. Let el, ... , e" be the standard basis
of R" and observe that
h
Il e (X + hei) — ti c (x) = ; çp F (x — z + te ; )u(z)dtdz
R"0 D
h
D,^pE(x — z + te,)u(z)dzdt.
= 4 Rn
As a function of t, the inner integral on the right is continuous, and thus
(1) follows.
In case (ii) observe that

1 u E (x) — u(x)l <f wr (x — y)Iu(y) — u(x)Idy

< sup cPE—" f Iu(x) — u(y)Idy —' 0


13(x ,t)
as E —, 0 whenever x is a Lebesgue point for u. Clearly the convergence
is locally uniform if u is continuous because u is uniformly continuous on
compact sets.
For the proof of (iii), Holder's inequality yields

Iuc(x)I = Sp c (x — y)u(y)dy

lip' 1/p
< co (x — y)dy SPF(x — y)lu(y)I p dy

The first factor on the right is equal to 1 and hence, by Fubini's theorem,

lue rdx < ^PF(x — y)Iu(y)lpdydx


R^ R^ R"
(Mx — y)lu(y)I p dxdy
f..L.
Iu(y)I p dy.
f"
1.7. Distributions 23

Consequently,
IIuE Ilp llullp• (1.6.4)
To complete the proof, for each rî > O let v E C0(R n ) be such that

Ilu vllP<
-
^-
(1.6.5)

Because u has compact support, it follows from (ii) that Ilv -- v e li p < ri for
e sufficiently small. Now apply (1.6.4) and (1.6.5) to the difference y -- u
and obtain

Ilu -ue llp<-Il u-v llp+llU- vE lip +IlvE tic lip <_ 317.

Hence u, u in LP(Rn) as e O. ❑

1.8.2. Remark. If u E L'(11), then u E (x) = (p c * u(x) is defined provided


• E S2 and e < dist(x, OD). It is a simple matter to verify that Theorem
1.6.1 remains valid in this case with obvious modification. For example, if
u E C(R) and S2' CC S2, then u E converges uniformly to u on S2' as -' O.
Also note that (iii) of Theorem 1.6.1. implies that mollification does
not increase the norm. This is intuitively clear since the norm must take
into account the extremities of the function and mollification, which is an
averaging operation, does not increase the extremities.

1.7 Distributions
In this section we present a very brief review of some of the elementary
concepts and techniques of the Schwartz theory of distributions [SCH} that
will be needed in subsequent chapters. The notion of weak or distributional
derivative will be of special importance.

1.7.1. Definition. Let SZ C Rn be an open set. The space (S2) is the


set of all (iv in Co (1l) endowed with a topology so that a sequence {}
converges to an element cp in g(1) if and only if
(i) there exists a compact set K C St such that spt Sp; C K for every i,
and

(ii) lirn, . D°(#0= = Dapp uniformly on K for each multi-index cr.

The definition above does not attempt to actually define the topology
on gl(11) but merely states a consequence of the rigorous definition which
requires the concept of "generalized sequences" or "nets," a topic that
we do not wish to pursue in this brief treatment. For our purposes, it
will suffice to consider only ordinary sequences. It turns out that £ (S2)
is a topological vector space with a locally convex topology but is not
24 1. Preliminaries

a normable space. The dual space, .Y'(fl), of 2(S/) is called the space of
(Schwartz) distributions and is given the weak'-topology. Thus, T; E 0(0)
converges to T if and only if T,(cp) -' T(cp) for every cp E . '(fi).
We consider some important examples of distributions. Let 1.1 be a Radon
measure on S/ and define the corresponding distribution by

T (cP) = f o(x)dµ

for all E 2(0). Clearly T is a linear functional on .Y(fl) and 1T(rp) <
Iµ4(spt (P)Il from which it is easily seen that T is continuous, and thus a
distribution. In this way we will make an identification of Radon measures
and the associated distributions.
Similarly, let f E Li ,(fl), p > 1, and consider the corresponding signed
measure defined for all Borel sets E C R" by

µ(E) = f (r)dx
E
and pass to the associated distribution

f (cP) = ço(x)f (x)dx.


R*
In the sequel we shall often identify locally integrable functions with their
corresponding distributions without explicitly indicating the identification.

1.7.2. Remark. We recall two facts about distributions that will be of


importance later. A distribution T on an open set SZ is said to be positive if
TOp) > 0 whenever cp > 0, cp E . (1l). A fundamental result in distribution
theory states that a positive distribution is a measure. Of course, not all
distributions are measures. For example, the distribution defined on Re by

T (cp) = f (p'(z)dx

is not a measure since it is not continuous on 0(0) when endowed with


the topology of uniform convergence on compact sets.
Another important fact is that distributions are determined by their local
behavior. By this we mean that if two distributions T and S on CI have
the property that for every x E fZ there is a neighborhood U such that
T ((p) = S((p) for all ç E . (1l) supported by U, then T = S. For example,
,

this implies that if {SZ a } is a family of open sets such that Ufl a = Si and
T is a distribution on fl such that T is a measure on each fL, then T is a
measure on O. This also implies that if a distribution T vanishes on each
open set of some family it then vanishes on the union of all elements of
.F. The support of a distribution T is thus defined as the complement of
the largest open set on which T vanishes.
1.7. Distributions 25

We now proceed to define the convolution of a distribution with a test


function Sp E O(SZ). For this purpose, we introduce the notation ( 03(x) =
Sp(—x) and r2 ço(y) = Sp(y — r). The convolution of a distribution T defined
on R" with yo E (SZ) is a function of class C°° given by

T * v(x) = T(rzy3 ). (1.7.1)

An obvious but important observation is

T * sp(0) = T(roç) = T(0).


If the distribution T is given by a locally integrable function f then we
have
(T * (P)(z) = f f (x - y)cP(y)db
which is the usual definition for the convolution of two functions. It is easy
to verify that
T*( *V)= (T * (p) * b
whenever cp, t/1 E^.

Let T be a distribution on an open set 0. The partial derivative of T is


defined as
D=T(v) = T (D.ço)

for cP E OW). Since Dop E 0(n) it is clear that D,T is again a distribu-
tion. Since the test functions W are smooth, the mixed partial derivatives
are independent of the order of differentiation:

D;Dj cP = Dl Dop

and therefore the same equation holds for distributions:

D; Di T = D • D 1 T.

Consequently, for any multi-index or the corresponding derivative of T is


given by the equation

DcT(cp) = ( -1 )' a1 T (D a (p).

Finally, we note that a distribution on SI can be multiplied by smooth


functions. Thus, if T E 93"(f1) and f E co 0(a), then the product fT is a
distribution defined by

(fT)(ço) = T(fçp), cP E r(SZ)-


The Leibniz formula is easily seen to hold in this context (see Exercise 1.5).
The reader is referred to [SCHJ for a complete treatment of this topic.
26 1. Preliminaries

1.8 Lorentz Spaces


We have seen in Lemma 1.5.1 that if f E L' (le), f > 0, then its integral
is completely determined by the measure of the sets {x : f (x) > t}, t E W.
The non-increasing rearrangement of f , (defined below) can be identified
f
with a radial function having the property that for all t E R V , {x : f(x) >
t} is a ball centered at the origin with the same measure as {x : f (x) > t).
f f
Consequently, f and have the same integral. Because can be thought
of as a function of one variable, it is often easier to employ than f . We
introduce a class of spaces called Lorentz spaces which are more general
but closely related to L" spaces. Their definition is based on the concept
of non-increasing rearrangement. Later in Chapter 2, we will extend basic
Sobolev inequalities in an L" setting to that of Lorentz spaces.

1.8.1. Definition. If f is a measurable function defined on R", let


E; ={x:If(x)1>s}, (1.8.1)

and let the distribution function of f be denoted by


(1.8.2)
af(s)=IE f I-
Note that the distribution function of f is non-negative, non-increasing,
and continuous from the right. With the distribution function we associate
the non-increasing rearrangement of f on (0, oo) defined by

f'(t) = inf{s > 0 : af(9) < t}. (1.8.3)


Clearly f' is non-negative and non-increasing on (0, oo). Further, if off is
continuous and strictly decreasing, then f' is the inverse of af, that is,
f' = cr f 1 . It follows immediately from the definition of f'(t) that

f . (af(s)) < s (1.8.4)

and because ci f is continuous from the right, that

af(f + (t)) t. (1.8.5)

These two facts lead immediately to the following propositions.

1.8.2. Proposition. f' is continuous from the right.


Proof. Clearly, f' (t) > f• (t + h) for all h > O. if f' were not continuous
at t, there would exist y such that f'(t) > y > f'(t+h) for all h > O. But
then, (1.8.5) would imply that a f (y) < a f (f' (t + h)) < t+ h for all h > O.
1.8. Lorentz Spaces

27

Thus, a f (y) < t and therefore, f'(t) < y, a contradiction.

1.8.3. Proposition. af . (s) a1(s) for all s > O.

Proof. Because f' is non-increasing, it follows from the definition of a f. (t)


that
a f•(8) = sup{t > 0 . f* (t) > s}. (1.8.6)
Hence, f'(a f (8)) < .q implies a f (s) > a1• (s). For the opposite inequal-
ity, note from (1.8.6) that if t > a f. (s), then f' (t) < s and consequently,
a f(s) < af(f +(t)) < t, by (1.8.5). Thus, af(s) < a f. (s) and the proposi-
tion is established. O

1.8.4. Proposition. Let { ff ) be a sequence of measurable functions on Rn


such that {I f,I} is a non-decreasing sequence. If If (x)I = lirn i . œ I f=(x)( for
each r E En, then a f, and fi increase to a f and f ' respectively.

Proof. Clearly
EP C Ef and = Ef
for each s and therefore a f,(8) --> ar f(s) as s — oo. It follows from definition
of non-increasing rearrangement, that f, (t) < f ' 1 (t) < f'(t) for each t
and i = 1,2, .... Let g(t) = f, (t). Since f; (t) < g(t) it follows from
(1.8.5) that a11 [g(t)] < af [f: (01 < t. Therefore,

o f [g(t)1 = Urn a1► [g(t)] < t

which implies that f'(t) < g(t). But g(t) < fi(t) and therefore the proof
is complete. ❑

1.8.5. Theorem. If f E L", 1 < p < oo, then

I/p 1/p
^fI p = [f œ [f'(t)1 dt (1.8.7)

Proof. This follows immediately from Lemma 1.5.1 and the fact that f
and f' have the same distribution function (Proposition 1.8.3). ❑

We now introduce Lorentz spaces and in order to motivate the following


definition, we write (1.8.7) in a more suggestive form as

I^f Îi p = ( f 0o
r t1/Pf(t)lPdt/1
)

)
1/13
28 L Preliminaries

It is sometimes more convenient to work with the average of f' than


with f' itself. Thus, we define

f"(t) = t 0
f " (r)dr.

1.8.6. Definition. For 1 < p < oo and 1 < q < oo, the Lorentz space
L(p, q) is defined as

L(p, q) = {f : f measurable on Rr, lI f II(p,4) < co} (1.8.8)

where II f It(p,4) is defined by

71
[tl/vf"(t)Iq,44 1/4
1 <p<oo,1 < g< ao
II/II(p,a) = j
.°00
sup t'/p f.. (t), 1 <p< oo, q=oo.
c>o

It will be shown in Lemma 1.8.10 that

L(p p) = L.
,
(1.8.9)

The norm above could be defined with f" replaced by f' in case p > 1
and 1 < q < oo. This alternate definition remains equivalent to the original
one in view of Hardy's inequality (Lemma 1.8.11) and the fact that f" > f'
(since f' is non-increasing). For p > 1, the space L(p, oo) is known as the
Marcinkiewicz space and also as Weak LP. In case p = 1, we clearly have
L(1, oo) = V. With the help of Lemma 1.5.1, observe that

1
t o0

f'(r)dr = tf'(t) + cxf(s)ds


f• (I)
and therefore
f" (t) = f' (t) + a f(s)ds. (1.8.10)
t f' (t)
For our applications it will be necessary to know how the non-increasing
rearrangement behaves relative to the operation of convolution. The next
two lemmas address this question. Because g" is non-increasing, note that
in the following lemma, the first and second conclusions are most interesting
when t < r and t > r, respectively.

1.8.7. Lemma. Let f and g be measurable functions on Rn where sup{ f (x):


r E R n } < a and f vanishes outside of a measurable set E with 1Ei r.
Let h = f * g. Then, for € > 0,
h"(t) < ac rg"(r)
1.8. Lorentz Spaces 29

and
h+'(t) < a rg'* (t).

Proof. For a > 0, define

if Ig(x)I < a
ga(x) g(x)
asgng(x) if fg(x)I > a

and let
g a (x) = g(x) _ gs(x)•
Then, define functions h i and h2 by

h=f *g —f *ga+f *g a
—h l +h 2 .

From elementary estimates involving the convolution and Lemma 1.5.1,


we obtain
0o
sup {h z (x) : x E R") < sup{ f (x) : x E E }Il gl l l < a a9(s)ds (1.8.11)
a

because ga(x) = 0 whenever Ig(x)I < a. Also

sup{h l (x) : x E R ) <


_ "
Ilf ll1 sup{ ga (x) : x E E} < ara, (1.8.12)

and
02111 <_ IIfllillgalli < a
a
a9(s)ds. (1.8.13)

Now set a = g`(r) in (1.8.11) and (1.8.12) and obtain


e
h"(t) = t

h+ (y)dy < 016

<_ Ilhlll^ +Ilhzl6


00

arg'(r) + a rx g (.․ )ds


9'(r)
coo
< a r?(r) +a g (s)d8
f •(r)
[

= a rg" (r) .

The last equality follows from (1.8.10) and thus, the first inequality of the
lemma is established.
1. Preliminaries

30

To prove the second inequality, set a = g' (r) and use (1.8.12) and (1.8.13)
to obtain
e
th "(t) = h'(yt)dyr hi(y!)dy + h2(y)dy
0 0
f
0
vo
t11hl lly° +h;(i) dy t ll h jll 0 + 11h2111
00
< ta rg'(t) + ac r as(s)ds
9' (T)
ao
< a r t9*(t) +
[ a s (a)ds
9 '(r)
< artg" (t)
by (1.8.10). 0

1.8.8. Lemma. If h, f , and g a re measurable functions such that h = f *g,


then for any t > 0
00
h "(t) S tf"(t)g+w(t) +f'(u)g*(r^)dr^.
f

Proof. Fix t > 0.


Select a doubly infinite sequence {ys } whose indices ranges from -oo to
+oo such that
Yo = f* (t)
— y:+1
yi <
Vim y, - oc
1 -1 00

Let

f(z) = fi(z)
i=t ‘00
-

where
0 if If (z)1 < yi-t
f1(z) = f(z) - y=-l sgn f(z) if ys_1 < If(z)I S yc
Ys — Ye —i sgn f (z) if y. < f (z)i•
Clearly, the series converges absolutely and therefore,
00

h=f •g-

o
( icxD
E

' )
*g

= E fi *g+ *g
,--on 6_i

-h l + h 2
1.8. Lorentz Spaces 31

with
h'*(t) < hi'(t) + h2 (t).
-

To evaluate hnt) we use the second inequality of Lemma 1.8.7 with


E, _ {z: If(z)I> y i _ i ) = E and a=y,-y i _ i to obtain
00

h's*(t) < E(yt - 3h-1)af(y,-1)g"(t)


s-1
00
l
— 9-'(t) Eckf(yti-1!(yi — yi-1)•
i= 1

The series on the right is an infinite Riemann sum for the integral

a f (y)dy,

and provides an arbitrarily close approximation with an appropriate choice


of the sequence { y, } . Therefore,
00
g
h2'(t) < *`(t) a f(y)dy. (1.8.14)
f •(i)
By the first inequality of Lemma 1.8.7,
00

hi *(t) 5 E(yi - y;-1)af (yi-1)9' + (af(ys-1))


1= 1

The sum on the right is an infinite Riemann sum tending (with proper
choice of y ; ) to the integral,

J.
f' (e)
a f (y)9 * * (af (y))d11.

We shall evaluate the integral by making the substitution y = f' (u) and
then integrating by parts. In order to justify the change of variable in the
integral, consider a Riemann sum
00

Eaf(Yi-09.6(af(Y8-1))(Yi — y+-1)
=t

that provides a close approximation to

1
f' (t)
af(y)9 '- (CI f (y))dy•

By adding more points to the Riemann sum if necessary, we may assume


that the left-hand end point of each interval on which cr f is constant is
32 1. Preliminaries

included among the yi. Then, the Riemann sum is not changed if each yi
that is contained in the interior of an interval on which a f is constant, is
deleted. It is now an easy matter to verify that for each of the remain-
ing y i there is precisely one element, ui, such that y i = f'(ui) and that
a f(f'(u i )) = u i . Thus, we have
ao
Eafoh-og"(af(y._,))(vs-v,_,)
i =1
00
`
Eui-00'( u i-1)(f'(us) - 1 y(uti -1))

which, by adding more points if necessary, provides a close approximation


to 00
— ug" (u)df• (u).
t

Therefore, we have

f' (f)
(t) < (If (y)g .' (a f (g))dEl
o
00
= _ u9..(,a)df.(.n)
t
00
= _ug .. (u)
f • (u)ir + f f'(u)g'(u)du
00
< tg * ' (Or (t) + f f+(u)g'(u)^u. (1.8.15)

To justify the integration by parts, let A be an arbitrarily large number


and choose u i such that t = u i < u 2 < ... < A. Observe that

tg .• (t) f■(t) _
^

Ag ..

(A)f'(A) E
1=1
^

+ E f'(u)[g(u,+1)u1+j g " (u,)uc]


5 -1

^
1
= r
ui +1g " (ui +1) {f (?ti+ I ) — f (%)1
'
.

1=1
+ „,+ .
Ef (ua) [

u,
g' (Odd
1-
1
i
Ç Eth+o-scui+otf.(u,i+i) - f'(ui)1
i =1

1.8. Lorentz Spaces 33

+ E f'(ui)g'(ui)[ui+1 ^ ui]•
i=1

This shows that


a a
A9 '(A)f*(A) — tg" (Of' (t) c ug " (u)d f* (u) + f *(u)g' (u)du.
t
To establish the opposite inequality, write
7

ag' * Nf s (A) — tg" (Or (t) = E uir(ui)[f'(u; +1) - f 'lui)i

^
J
+ r,,
f'(ut+1)lJ*'(u;♦1)u. +1 - g' *(u,)ui]
i= 1
J
= Eu, g * ( u i) [ f * ( u i+j) - f'(ui)1
;=1

+ Ef'+') f g
^

'(r)dr
^ [

> E uig "(u+)[f'(ui+,) — f`(ui)1


i-r
^

+ E f (u i+ i )g s (Ili+ )[u i+i — ui^•


i=1

Now let A 0o to obtain the desired equality. Thus, from (1.8.15), (1.8.14),

and (1.8.10),

0o ea
(t) (t) < g "(t) { tf*() +^ (y)dy
^ + f (u)g' (u)du
f• (e) r

< t f"(t)g "(t) + f °13 f'(u)g*(u)du. 0

1.8.9. Lemma. Under the hypotheses of Lemma 1.8.8,


co
h .. (t) <
f f *. (u)g" (u)du.

Proof. We may as well assume the integral on the right is finite and then
conclude
lirn uf'*(u)g "(u) = 0. (1.8.16)
u^vo
34 1. Preliminaries

By Lemma 1.8.8 and the fact that f` < f", we have


00
h "(t) < t f «.(t) g ..(t) +f'(u)g'(u)du
f

< t f .. (t)g .. (t) + f DID f.. (u)g'(u)du. (1.8.17)

Note that since f' and g' are non-increasing,

a.(,u) ' ^ (u) — (u)]


â4L r
fa*

and

du ug" (u) = g' (u)


for almost all (in fact, all but countably many) u. Since f" and g" are
absolutely continuous, we may perform integration by parts and employ
(1.8.16) and (1.8.17) to obtain

h"(t) < t f..(t)g "(t) + uf'«(u)g..(u)Ir

00
+ (f(u) — f'(u))g"(u)du
t
^

= (f" (u) — f'(u))g" (u)du


t
00
f** •• D
it

We conclude this section by proving some lemmas that provide a com-


parison between various Lorentz spaces. We begin with the following that
compares 17 and L(p,p).

1.8.10. Lemma. If 1 < p < oc and lip + 1/p' = 1, the n

Ilfllp < IIfI102,0 <_ p'Ilfllp•

Proof. Since f' < f",

IIfIIp = (f*(^)1pdt =
°° (tlip f'(t))p dt < f [t'/Pf»(t)1!
0 o /j
= (I1fII(p,p))p.
1.8. Lorentz Spaces 35

The second inequality follows immediately from the definition of f''(t) and
the inequality

°° xl/pfox n di] 1/P too dx 1/p


f(t)dt ç p' o rx 1/
[J0 x x af(4P x j

which is a consequence of the following lemma with r = p - 1. ❑

The next result is a classical estimate, known as Hardy's inequality, which


gives information related to Jensen's inequality (1.5.12). If f is a non-
negative measurable function defined on the positive real numbers, let
s
F(x) =-- (t)dt, x > 0.
-f f
Jensen's inequality gives an estimate of the pth power of F; Hardy's in-
equality gives an estimate of the weighted integral of the pth power of F.

1.8.11. Lemma (Hardy). If 1 < p < oo, r > 0 and f is a non-negative


measurable function on (0, oo), then with F defined as above,

i
co cc
F^x)Pxp -r -ldx < (= f (t)ptF-r-ldt
l r r)
a

Proof. By an application of Jensen's inequality (1.5.12) with the measure


t ( r/p )-1 dt, we obtain
s P f? ( P
f(t)d = f(t)tl- (r /P)tfT/p)1dt
o
s
<\ xr(1 -1/p) ( Ef(t)Y)tP_T_1+Tdt.
o

Then by Fubirri's theorem,

f
0o Z

f (t)dt s -r -ldx
0

f
(

< (! P_' ) 00 s 1 - (r/p) If (t)]Ptp-"* - 1 +(r/P) dt dx


P 1 oo
^, -
(12) fo t ptp- r -1 +(r /p)Ut
0011 (}1 x-1-tr/P)dx dt

I 1 F °O ( / jt r -I
� ^f(t)t^p - dt. D
- tpl
r o
36 1. Preliminaries

The following two lemmas provide some comparison between the spaces
L(p,q) and L(p, r).

1.8.12. Lemma.

**( <(1) 1/9


f/
lIfll(P,Q) ^ e 1 /e ^If I;)
` p xl/P — x /p .

Proof.

f..(t)^Qdt
(II /II(p,Q)) 4 = m [eh'
0
2:[ f+*(t)14t(4/P)-1dt
z
0
z
^ [ f ** ( x )]9t (9/p)' l dt
0
P/4
[f ** (^)^ 4 x
_^
The first inequality follows by solving for f'*(x) and the second by ob-
< ql/ 4 < e lks
serving that ( 2 ) 1/9 ❑
P

1.8.13. Lemma (Calderón). If 1 < p < oc and 1 < q < r < co, then

II /II(p.r) Ç ( i ) I II II(P,4) Ç el fe ll f 11(F.4)•


(1/q)—(1/r)

Proof.
00
(11116.0Y — [f**(y)lrz ( r/P )-l dz
0
00
= [t** (x)1 Q [f ** (x)jr -4 x (
r/P )-1 d.Z
0
1/4 I r -4
^ iIf II (p, 9) x (r/P)- l dx
< ^[f ** (X)? 1 /p
o [ 't vil
i

= (II/II(p.q))r-4(li/IIcP,9))q,

(q)("q)-

and the first inequality follows by taking the r th root of both sides. The
second follows by the same reasoning as in the previous lemma. ❑
Exercises 37

Exercises
1.1. Prove that if E C R" is an arbitrary set, then the distance function
to E is Lipschitz with constant 1. That is, if d(x) = d(x, E), then
id(x) — d(y)I < ix — yI for all x, y E R".

1.2. (a) Prove that if E is a set with HQ(E) < oo, then HA(E) = 0 for
every f3 > a.
(b) Prove that any set E C R" has a unique Hausdorff dimension.
See Remark 1.4.3.

1.3. Give a proof of Lemma 1.5.1. More generally, prove the following:
Let So: [0, oo] —' [0, oo] be a monotonic function which is absolutely
continuous on every closed interval of finite length. Then, under the
conditions of Lemma 1.5.1, prove that
^
Sp o 11,0=f(E)co'(t)dt.

1.4. Prove that C"(11) is a Banach space with the norm defined in Sec-
-

tion 1.1.

1.5. Let f E Cô (Rn) and T a distribution. Verify the Leibniz formula

a^
D'( fT)_^ Dg f D^ AT —

A<a Q^(a Q)^

where we say ,3 < a provided f3 < at for 1 < i < n.

1.6. Prove that if T is a distribution and cp E Cfl (R"), then T * (P E


Cô (Rn) and D(T * Sp) = (DT) * ço where D denotes any partial
derivative of the first order. This may be accomplished by analyzing
difference quotients and using the fact that rh(DT) = D(rhT).

1.7. Lemma 1.8.13 shows that if 1 < p < oo and 1 < q < r < oo, then

L(p q) C L(p,q) C L(p, r) C L(p, oo).


,

Give examples that show the above inclusions are strict.

1.8. As we have noted in Remark 1.4.3, the measure FP does not satisfy
the regularity property analogous to (1.4.5). However, it does have
other approximation properties. Prove that if A C R" is an arbitrary
set, there exists a G6-set G D A such that

HI (A) = H " ( G ).
38 1. Preliminaries

It can also be shown (although the proof is not easy) that if A is a


Suslin set, then

H 7 (A) = sup{H'(K) : K C A, K compact, H'r(K) < co).

See [F4, 2.10.48].

1.9. Prove the statement that leads to (1.8.10), namely, if f E L'(R Th ),


then

1 f '(r)dr = t f'(t) + a f(s)d9.


f ' (t)

Hint: Consider the graph of f and employ Lemma 1.5.1.

1.10. Another Hausdorff-type measure often used in the literature is Haus-


dor f spherical measure, HS. It is defined in the same manner as Ill
(see Definition 1.4.1) except that the sets A; are taken as n-balls.
Clearly, 11 1'(E) < HS(E) for any set E. Prove that HS(E) = 0
whenever H 7 (E) = 0.

1.11. Suppose u is a function defined on an open set SZ C R" with the


property that it is continuous almost everywhere. Prove that u is
measurable.

1.12. Using only basic information, prove that the class of simple functions
is dense in the Lorentz space L(p, q).

1.13. Let p be a Radon measure on R". As an application of Theorem 1.3.6


prove that any open set U C R" is essentially (with respect to p) the
disjoint union of n-balls. That is, prove that there is a sequence of
disjoint n-balls B ; C U such that

[u _ û niJ =0.
t-1

1.14. Let p be a Radon measure on R". Let I be an arbitrary index set


and suppose for each a E I, that Ea is an p-measurable set with the
property that
g[Ea n B(x, r)1
lim —1
r-'0 tt[B(z r)],

for every x E Ea . Prove that UaEIEQ is p-measurable.

1.15. From Exercise 1.1 we know that the distance function, d, to an arbi-
trary set E is Lipschitz with constant 1. Looking ahead to Theorem
2.2.1, we then can conclude that d is differentiable alma't everywhere.
Historical Notes 39

Prove that if E is a closed set and d is differentiable at a point x « E,


then there exists a unique point «(x) E E nearest x. Also prove that

Dd(x) = x (x)
x d
()

1.16. (a) If f is a continuous function defined on R", prove that its non-
increasing rearrangement f' is also continuous. Thus, continu-
ous functions remain invariant under the operation of rearrange-
ment.
(b) Now prove that Lipschitz functions also remain invariant under
rearrangement. For this it will be necessary to use the Brunn-
Minkowski inequality. It states that if E and F are nonernpty
subsets of R" , then

IE + F ill" > 1Eil/n + PFI 1/"

whereE+F= {x+y:XEE,yE F}.


(c) Looking ahead to Chapter 2, prove that if f E W 1 •P(R"), then
f' E W 149 (R"). Use part (b) and Theorem 2.5.1.
(d) Show by an example that C 1 (R") does not remain invariant
under the operation of rearrangement.

1.17. Let u E C ° (R 1 ). For each h 0, let u h be the function defined by

u(x + h) — u(x)
uh(x) = h

Prove that uh --e u' in the sense of distributions.

1.18. Let fu,) be a sequence in fP (R") that converges weakly to u in


LP(R"), p > 1. That is,

lim u, v dz --►uv dx
i —oo R^ R "

for every v E IP' (R"). Prove that D'u, --r D'u in the sense of
distributions for each multi-index a.

Historical Notes
1.2. The notion of measures leas two fundamental applications: one can be
used for estimating the size of sets while the other can be used to define
integrals. In his 1894 thesis, E. Borel (cf. [BO)) essentially introduced what
is now known as Lebesgue outer measure to estimate the size of sets to assist
his investigation of certain pathological functions. Lebesgue [LEI' used
40 1. Preliminaries

measures as a device to construct his integral. Later, when more general


measures were studied, Radon (1913) for example, emphasized measure
as a countably additive set function defined on a o-ring of sets whereas
Carathéodory (1914) pursued the notion of outer measures defined on all
sets.
1.3. The material in this section represents only a very small portion of the
literature devoted to differentiation theory and the related subject of cover-
ing theorems. Central to this theory is the celebrated theorem of Lebesgue
[LE2J which states that a locally integrable function can be represented
by the limit of its integral averages over concentric balls whose radii tend
to zero. Theorem 1.3.8 generalizes this result to the situation in which
Lebesgue measure is replaced by a Radon measure. This result and the
covering theorems (Theorems 1.3.5 and 1.3.6) which lead to it are due to
Besicovitch, [BE1), [BE2]. The proof of Theorem 1.3.5 was communicated
to the author by Robert Hardt. The original version of Theorem 1.3.6 is due
to Vitali [VI) who employed closed cubes and Lebesgue measure. Lebesgue
(LE2) observed that the result is still valid if cubes are replaced by gen-
eral sets that are "regular" when compared to cubes. A sequence of sets
{Ek } is called regular at a point z o if z o E flk 1Ek, diam(Ek) —+ 0 and
lim infk_. Q p(Ek) > 0 where p(Ek ) is defined as the infimum of the numbers
ICI/IEkI with G ranging over all cubes containing Ek. In particular, one is
allowed to consider coverings by nested cubes or balls that are not neces-
sarily concentric. However, in the case when Lebesgue measure is replaced
by a Radon measure, Theorem 1.3.6 no longer remains valid if the balls in
the covering are allowed to become too non-concentric. At about the time
that Besicovitch made his contributions, A.P. Morse developed a theory
which allowed coverings by a general class of sets rather than by concentric
closed balls. The following typifies the results obtained by Morse [MSE2):
Let A C R" be a bounded set. Suppose for each z E A there is a set H(x)
satisfying the following two properties: (1) there exist M > 0 independent
of r and r(x) > 0 such that

B(x, r(x)) C H(x) C B(x, Mr(x));

(ii) H(x) contains the convex hull of the set {y} U B(x, r(z)) whenever
y E H(x). Then a conclusion similar to that in Theorem 1.3.5 holds.
Another useful covering theorem due to Whitney [WHI states than an
open set in R" can be covered by non-overlapping cubes that become
smaller as they approach the boundary. Theorem 1.3.5 is a similar result
where balls are used instead of cubes and where the requirement of disjoint-
ness is replaced by an estimate of the amount of overlap. This treatment
is found in [F4, Section 3.1).
Among the many results concerning differentiation with respect to irreg-
ular families is the following interesting theorem proved in [JMZ]: Suppose

Historical Notes 41

u is a measurable function defined on R" such that

f 'ul(1 +log + Pul)n -l dx < oo.

Then, for almost every x E Rn,

lim III ` Iu(y) - u(x)Idy = 0


1

where the limit is taken over all bounded open intervals I containing the
point x. This result is false if u is assumed only to be integrable. Such ir-
regular intervals are useful in applications concerning parabolic differential
equations, where it is natural to consider intervals of the form C x 10, r 2 J,
where C is an (n - 1)-cube of side-length r.
For further information pertaining to differentiation and coverings, the
reader may consult [DG], [F4, Section 2.8].
1.4. Carathéodory [CAY] was the first to introduce "Hausdorff" measure in
his work on the general theory of outer measure. He only developed linear
measure in Rn although he indicated how k-dimensional measure could be
defined for integer values of k. k-dimensional measure for general positive
values of k was introduced by Hausdorff [HAU] who illustrated the use
of these measures by showing that the Cantor ternary set has fractional
dimension log 2/ log 3.
1.7. There are various ways of presenting the theory of distributions, but
the method employed in this section is the one that reflects the original
theory of Schwartz [SCI-1] which is based on the duality of topological vector
spaces. The reader may wish to consult the monumental work of Gelfand
and his collaborators which contains a wealth of material on "generalized
functions" [GE1], [GE2], [GE3], [GE4], [GEM.
1.8. Fundamental to the notion of Lorentz spaces is the classical concept
of the non-increasing rearrangement of a function which, in turn, is based
upon a notion of symmetrization which transforms a given solid in R 3
into a ball with the same volume. There are a variety of symmetrization
procedures including the one introduced by J. Steiner [ST] in 1836 which
changes a solid into one with the same volume and at least one plane of
symmetry. The reader may consult the works by Pôlya and Szegô [PS] or
Burago and Zalgaller [BUZ] for excellent accounts of isoperimetric inequal-
ities and their connection with symmetrization techniques. In 1950 G.G.
Lorentz [LO1], 1LO21, first discussed the spaces that are now denoted by
L(p, 1) and L(p, oo). Papers by Hunt [HU] and O'Neil [0] present interest-
ing developments of Lorentz spaces. Much of this section is based on the
work of O'Neil and the main results of this section, Lemmas 1.8.7-1.8.9,
were first proved in [O]. The reader may consult LCA21, [CA31, 1LP], [PE]
for further developments in this area.
2

Sobolev Spaces and Their


Basic Properties
This chapter is concerned with the fundamental properties of Sobolev
spaces including the Sobolev inequality and its associated imbedding the-
orems. The basic Sobolev inequality is proved in two ways, one of which
employs the co-area formula (Section 2.7) to obtain the best constant in the
inequality. This method relates the Sobolev inequality to the isoperinietric
inequality.
The point-wise behavior of Sobolev functions will be discussed in Chap-
ters 3 and 4 and this will entail a method of defining Sobolev functions
on large sets, sets larger than the complement of sets of Lebesgue measure
zero. It turns out that the appropriate null sets for this purpose are de-
scribed in terms of sets of Bessel capacity zero. This capacity is introduced
and developed in Section 2.6 but only to the extent needed for the analysis
in Chapters 3 and 4. The theory of capacity is extensive and there is a vast
literature that relates Bessel capacity to non-linear potential theory. It is
beyond the scope of this book to give a thorough treatment of this topic.
One of the interesting aspects of Sobolev theory is the behavior of the
Sobolev inequality in the case of critical indices. In order to gain a better
appreciation of this phenomena, we will include a treatment in the context
of Lorentz spaces.

2.1 Weak Derivatives


Let u E LL(n). For a given multi-index a, a function y E LL(11) is called
the a th weak derivative of u if

Spvdx = ( -1 )i' 1 f uD ° Spdx (2.1.1)

for all W E (SZ). y is also referred to as the generalized derivative of u


and we write v = D°u. Clearly, D Q u is uniquely determined up to sets
of Lebesgue measure zero. We say that the ath weak derivative of u is a
measure if there exists a regular Borel (signed) measure g on SZ such that

i ^pdµ = (-1)i' 1
n
uD°`Spdx (2.1.2)
2.1. Weak Derivatives 43

for all ' E Co (1/). In most applications, lal = 1 arid then we speak of u
whose partial derivatives are measures.

2.1.1. Definition. For p > 1 and k a non-negative integer, we define the


Sobolev space

Wk'p(1) = L"(1 Z) n {u : Dau E Lp(1), lal < k].


- (2.1.3)

The space Wk•p(f) is equipped with a norm

Ilullk.p;sl = f E IDa^;lp^
1 j <k
a
(2.1.4)

which is clearly equivalent to

E IID°ullp,n• (2.1.5)
IQ!<k

It is an easy matter to verify that W k ip(1) is a Banach space. The space


Wô'p(fl) is defined as the closure of Co (11) relative to the norm (2.1.4).
We also introduce the space BV(0) of integrable functions whose partial
derivatives are (signed measures) with finite variation; thus,

BV(S1) = L l (1Z) n f u : D'u is a measure, jD°ul(11) < oo, lal = 11.


(2.1.6)
A norm on BV(11) is defined by

Ilullev(n) = Ilulll;n + E IDauI(n). (2.1.7)


lai.1

2.1.2. Remark. Observe that if u E Wk'p(11) t.) BV (SI), then u is deter-


mined only up to a set of Lebesgue measure zero. We agree to call these
functions u continuous, bounded, etc. if there is a function û such that
TA = u a.e. and t has these properties.
-

We will show that elements in Wk.P(SZ) have representatives that permit


us to regard them as generalizations of absolutely continuous functions
on R'. First, we prove an important result concerning the convergence of
regularizers of Sobolev functions.

2.1.3. Lemma. Suppose u E W k,P (SZ ), p > 1. Then the regularizers of u


(see Section 1.6), u e , have the property that

li ô NE - u I'Ik.p ;n , = 0
44 2. Sobolev Spaces and Their Basic Properties

whenever f' CC 11. In case S2 = Rn, then lirn._.o Igu e - uP[k, F = O.

Proof. Since 1' is a bounded domain, there exists ea > 0 such that c o <
dist(ir, 8S1). For e < c o , differentiate under the integral sign and refer to
(2.1.1) to obtain for x E re and [a[ < k,

Dac e (x) = e-nf Dx ço u(y)dy

= (-1) 1'a-n
1 Dÿ (P (L=1) u(y)dy
n a
= e-n
D°`u(y)dy
n
_ (D a u)E (x)
for each x E f2'. The result now follows from Theorem 1.6.1(iii). C^

Since the definition of a Sobolev function requires that its distributional


derivatives belong to L", it is natural to inquire whether the function pos-
sesses any classical differentiability properties. To this end, we begin by
showing that its partial derivatives exist almost everywhere. That is, in
keeping with Remark 2.1.2, we will show that there is a function u such
that Ti = u a.e. and that the partial derivatives of Ft exist almost every-
where. However, the result does not give any information concerning the
most useful concept of total differential, the linear approximation of the
difference quotient. This topic will be pursued in Chapter 1

2.1.4. Theorem. Suppose u E LP(S2). Then u E W 1 "( )) , p > 1, if and


only if u has a representative i that is absolutely continuous on almost
all line segments in I2 parallel to the coordinate axes and whose (classical)
partial derivatives belong to LP(S2).

Proof. First, suppose u E W 1, P(12). Consider a rectangular cell in 1


R = [a1, b1] x ... x ia n , b^]
all of whose side lengths are rational. We know from Lemma 2.1.3 that the
regularizers of u converge to u in the W 1 () norm. Thus, writing x E R
as x = (z, x i ) where x E R" -1 and ri E [ai , bi], 1 < i < n, it follows from
Fubini's Theorem that there is a sequence {Ell -' 0 such that
bt
Tim Iuk(x,xi) — utx xi)I ' + IDuk(x,zi) — Du(x zi)I edx1 , =ü
k--•oo a
,

for almost all x. Here, we denote li ck = uk. Since uk is smooth, for each
such x and for every q > 0, there is M > 0 such that for b E [a i , b i ],

f
bt
u k (x b) y uk ( x, a +)1 <_ ID t1k(2, x i )Id^ i
t
2.1. Weak Derivatives 45

b,
< I Du (x, x; ) I d^ +1/
;

ad

for k > M. If {uk(i, a,)} converges as k oo, (which may be assumed


without loss of generality), this shows that the sequence {uk} is uniformly
bounded on [a ; , bd. Moreover, as a function of xi, the u k are absolutely
continuous, uniformly with respect to k, because the Li convergence of
Du k to Du implies that for each e > 0, there is a 45 > 0 such that
fE < e whenever 11 1 (E) < b for all positive integers k.
Thus, by the Arzelà-Ascoli theorem, {uk} converges uniformly on [a 1 , b11
to an absolutely continuous function that agrees almost everywhere with
u. This shows that u has the desired representative on R. The general case
follows from the familiar diagonalization process.
Now suppose that u has such a representative û. Then iacp also possesses
the absolute continuity properties of û whenever cp E Co (S2). Thus, for
,

1 < 1 < n, it follows that

f iD 1 codx = — f D,rtSp dz

on almost every line segment in fl whose end-points belong to R n -- spt (P


and is parallel to the i th coordinate axis. Fubini's Theorem thus implies
that the weak derivative Diu has Dui as a representative. D

2.1.5. Remark. Theorem 2.1.4 can be stated in the following way. If u E


LP(12), then u E W i 'P(fI) if and only if u has a representative i such that
E W 1 'P(A) for almost all hne segments A in S2 parallel to the coordinate
axes and /DUI E LP(S1). For an equivalent statement, an application of
Fubini's Theorem allows us to replace almost all line segments A by almost
all k-dimensional planes Ak in 1 that are parallel to the coordinate k-planes.
It is interesting to note that the proof of Theorem 2.1.4 reveals that
the regularizers of u converge everywhere on almost all lines parallel to the
coordinate axes. Hu were not an element of W 14 (t1), but merely an element
of L' (e) Fubini's theorem would imply that the convergence occurs only
,

H'-a.e. on almost all lines. Thus, the assumption u E W t.P(1l) implies


that the regularizers converge on a relatively large set of points. This is an
interesting facet of Sobolev functions that will be pursued later in Chapter
3.
Recall that if u E L 1 '(R"), then IIu(x + h) — u(x)I6 —► 0 as h O. A
similar result provides a very useful characterization of Wl.a(R").

2.1.6. Theorem. Let 1 < p < oo. Then u E W I 'P(R") if and only if
U E L1'(R") and

fj u(x+h)—u(x)
dx
I " e/P =
+ h) — u.(z)1In
2. Sobolev Spaces and Their Basic Properties

46

remains bounded for all h E R".

Proof. First assume u E Co ( R" ). Then

ux+h 1 Ihl
( ) —u (x ) _ 9-
11 Du ^+ t • ^ dt,
Ih4 o I I Ihl
so by Jensen's inequality (1.5.12),
P
u,(x + h) — u(x) ^ ' hl
1Du + t—h dt.
h - khl o II
Therefore,

1 Ihl h p
IIu x + h) — u(x)IT^ Ihlp Du x + t dxdi,
{

IhI o Rn Ih I
or
IIu(x + h) — u(x)IIp <— Ihi IIDuII p .
By Lemma 2.1.3, this holds whenever u E W 1, p(R").
Conversely, if e; is the ill unit basis vector, then the sequence
u(x + e, /k) — u(x)
1/k
is bounded in LP(R"). Hence, by Theorem 1.5.2, there exists a subsequence
(which will be denoted by the full sequence) and u, E L"(R") such that

u(x + e, / k) — u(x)
u;
1/k
weakly in L (R" ). Thus, for (,o E gv ,

uopdx= lirn [ u x+ ei/k)_u(x)


( ] ()d r
k-.00 f " /
1k
f ço(x — ei /k) — cP(x)
= Gm u(x) dx
k^ooR " j 1/k

-- — uD cp dx.
;

R "

This shows that


D i u =u;
in the sense of distributions. Hence, u E W 1 "p(R"). o

2.1.7. Definition. For a measurable function u: SZ - F R 1 , let

u + = max{u, 0}, u - = min{u, 0}.


2.1. Weak Derivatives 47

2.1.8. Corollary. Let u E W""p(S2), p > 1. Then u+, u - E W l 'p(S2) and

Du if u >0
Du+ =
0 if u<0

Du - _ f 4 if u
Du if u < 0.
Proof. Because u has a representative that has the absolute continuity
properties stated in Theorem 2.1.4, it follows immediately that v.+, u - E
W 1, p(11). The second part of the theorem is reduced to the observation
that if f is a function of one variable such that f' exists a.e., then (f + )' _
f' ' xuf>o)• D
2.1.9. Corollary. If SI is connected, u E W' (S2), p > 1, and Du = 0 a.e.
on Sl, then u is constant on S2.

Proof. Appealing to Theorem 2.1.4, we see that u has a representative


that assumes a constant value on almost all line segments in Sl parallel to
the coordinate axes. ❑

2.1.10. Remark. The corollary states that elements of W1 F(Sl) remain


,

invariant under the operation of truncation. One of the interesting aspects


of the theory is that this, in general, is no longer true for the space W k .P(0).
Motivated by the observation that u+ = H o u where H is defined by

o
H(t)= t^
Ot < 0 —

we consider the composition H o u where H is a smooth function. It was


shown in [MA21 and [MA3] that it is possible to smoothly truncate non-
negative functions in W 2 ' 1'. That is, if H E C°o(11. 1 ) and

sup lt 2-1 H ( ' ) (t)l < M < oo

for j = 1, 2, then there exists C = C(p, M) such that for any non-negative
v E Cp (R")
IlDaH(u)Ilp < ClID2vIl,
for 1 < p < n/2 and any multi-index a with lai = 2. Here D 2 v denotes the
vector whose components consist of all second derivatives of u. However,
it is surprising to find that this is not true for all spaces Wk , p. Indeed, it
was established in [DA1] that if 1 < p < n/k, 2 < k < n, or 1 < p < n/k,
k = 2, and H E C°°(R 1 ) with H (k) (t) > 1 for [ti < 1, then there exists a
function u E W k iP(R") n C°O(R?) such that H(u) « W k ip(R"). The most
general result available in the positive direction is stated in terms of Riesz
potentials, I * f (see Section 2.6), where f is a non-negative function in
Q

48 2. Sobolev Spaces and Their Basic Properties

L'. The following result is due to Dahlberg [DA2]. Let 0 < a < n and
1 < p < n/a. Let H E C°° (R 1 ) have the property that
sup It2-1H(i) (t)I < M < oc
t>o
for j = 0,1, ... , a*, where a' is the smallest integer > a. If f E LP(R")
and f > 0, then there exists g E IP(R") such that
H(Ic,*g)=Ia *g a.e.

and IIgIIp < CII f lip where C = C(a, p, n, M). The case of integral a was
treated in [AD4] and in this situation the result can be formulated as

I D"[H(Ia *f)1IIp CIIfIIp


for any multi-index 7 with I -II = k.
To continue our investigation of the calculus of Sobolev functions, we con-
sider the problem of composition of a suitable function with u E W l .p(SZ).
Before doing so, we remind the reader of the analogous problem in Real
Variable theory. In general, if f and g are both absolutely continuous func-
tions, then the composition, f og, need not be absolutely continuous. Recall
that a function, f, is absolutely continuous if and only if it is continuous, of
bounded variation, and has the property that If(E)1 = 0 whenever IEI = O.
Thus, the consideration that prevents f o g from being absolutely continu-
ous is that fog need not be of bounded variation. A result of Valle Poussin
[PO] states that f o g is absolutely continuous if and only if f' o g g' is
integrable. An analogous result is valid in the context of Sobolev theory, cf.
[MM1], IMM2), but we will consider only the case when the outer function
is Lipschitz.

2.1.11. Theorem. Let f : R 1 -, R` be a Lipschitz function and u E


W 1, P(0), p > 1. If f o u E IP(1 ), then f o u E Wl'p(1) and for almost all
E SZ,
D(f o u)(x) = flu(x)]. Du(r).

Proof. By Theorem 2,1,4, we may assume that u is absolutely continuous


on almost all line segments in ft. Select a coordinate direction, say the
1 th , an d consider the partial derivative operator, D i . On almost all line
segments, A, in ft parallel to the it" coordinate axis, fou is clearly absolutely
continuous because f is Lipschitz. Moreover,

Di( f o u)(x) = f'[u(x)) . Diu(x) (2.1.8)

holds at all x E A such that D,u(r) and flu(r)1 both exist. Note that if
Diu(x) = 0, then D i (f o u)(x) = 0 because
I f [u(x + heh) - f [u(x)]1 (r + h h) - u(x)1
<M1'
11 II
2.2. Change of Variables for Sobolev Functions 49

where M is the Lipschitz constant of f and e, is the i th coordinate vector.


Thus, letting N = A n {x : D i u(x) = 0), we have that (2.1.8) holds on N.
Now let

P= (A- N) n{x:au(x) exists and D ; te(x)# 0}

and note that P U N occupies H 1 -almost all of A. From classical consid-


erations, we have that if S C P and H 1 [u(S)1 = 0, then H 1 (S) = O. In
particular, if we let E = {y : f'(y) fails toexist}, then H 1 [u -1 (E)nPJ = 0.
Since (2.1.8) holds if x E A - u - ' (E) n P and D i u(x) exists, it follows
therefore that (2.1.8) holds at H 1 -almost all points of A. At all such x, we
may conclude that

_
AU a u)(x)I" < M P I D,u(x)IF. (2.1.9)

Once it is known that the set of x E 11 for which (2.1.8) holds is a mea-
surable set, we may apply Fubini's Theorem to conclude that f o u sat-
isfies the hypotheses of Theorem 2.1.4. This is a consequence of the fact
that the functions on both sides of (2.1.8) are measurable. In particular,
f' a u is measurable because f' agrees with one of its Borel measurable
Dini derivates almost everywhere. O

2.2 Change of Variables for Sobolev Functions


In addition to the basic facts considered in the previous section, it is also
useful to know what effect a change of variables has on a Sobolev function.
For this purpose, we consider a bi-Lipschitzian map

T: .

That is, for some constant M, we assume that both T and T` 1 satisfy,

IT(x) - T(y)I < Mix for all x, y E Si,

IT -1 (x') - T -1 (11)1 < Mix' - y'1 for all x i , y' E fr. (2.2.1)
In order to proceed, we will need an important result of Rademacher which
states that a Lipschitz map T: R" -s Rtm is differentiable at almost all
points in R". That is, there is a set E C R with IEi = 0 such that for
"

each x E Rn - E, there is a linear map dT(x): R Mil (the differential


"

of T at x) with the property that

17 (x + y) - T(x) - dT(x, y)I


1
= 0. {2.2.2}
o
lo
^ Iui
2. Sobolev Spaces and Their Basic Properties

50

In order to establish (2.2.2) it will he sufficient to prove the following


result.

2.2.1. Theorem. If f : R n — R 1 is Lipschitz, then for almost all x E R" ,


-Fy) — f(x)—Df(x)'31
lim f(x _ü.
y ^o lyf
Proof. For ty E R" with H = I, and x E W', let 7(t) = f (x + tv). Since f
is Lipschitz, 7 is differentiable for almost all t.
Let df (x, v) denote the directional derivative of f at x. Thus, df (x, y) =
7'(0) whenever 7'(0) exists. Let

N„ = R n {x : df (x, v) fails to exist}


"

Note that

+ tu) — f (x) > iv) (x)}


N„ = x : lim sup f (x t lirn inf f (x + t — f
t ^o ^
1-00

and is therefore a Borel measurable set. However, for each line A whose
direction is V, we have HI (N„ fl A) = 0, because f is Lipschitz on A.
Therefore, by Fubini's theorem, IN„' = O. Note that on each line A parallel
to VV

1. df (x, v)Sp(x)dx = — f f (x)dSp(x, v)dx

for cp E Co (R"). Because Lebesgue measure remains invariant under or-


thogonal transformations, it follows by Fubini's Theorem that

R df (x, v)Sp(x)dx = — f (x)dcp(x, v)dx


R "

= - f (x)Dcp(x) - v dx
R "

=— f (x)Dj çp(x) • v, dx
j=1 R"

-E Djf(x)cP(x)•Vjdx

f
R"
J=1
^o(x)D f (x) - v dx.
"

Because this is valid for all cp E Ca (Rn), we have that

df (x, v) = D f (x) • v a.e. x E R11 .


, (2.2.3)
2.2. Change of Variables for Sobolev Functions 51

Now let u1, u2,... be a countable dense subset of Sn -1 and observe that
there is a set E with 1E1 = 0 such that

df (x, vk) = D f (x) . vk (2.2.4)

for all x E R" — E, k --- 1, 2,....


We will now show that our result holds at all points of R" — E. For this
purpose, let x E R" — E, lui = 1, t > 0 and consider the difference quotient

Q(x, v, t) - 1(x +
hi) — f (x) D (x) • v.
t
For v, v' E Sn -1 and t > 0 note that

f (x + tv') + (v - v') • D f (x)I


IQ(z, v, t) — Q(x, v , t)I w If (x + tu) -
t
'

< Mlv — v'I + 1v — v'l . IDf(x)I Ç M(n + 1)1v — v i l (2.2.5)


where M is the Lipschitz constant of f. Since the sequence {v i } is dense in
Sn -1 , there exists an integer K such that

Iv — vkl < 2(n + 1)M for some k E {1, 2, .. ., K} (2.2.6)

whenever y E Si -1 . For x o E R" - E, we have from (2.2.4) the existence


of 6 > 0 such that
^
IQ(x o , vk, t)1 < 2 for O < t < 6, k E {1, 2, . .. , K}. (2.2.7)

Since
IQ(xo v, t)l _
,
< I Q(xo, vk, t)I + IQ(xo, v, t) -- Q(xo, vk, t)I
for k E {1, 2, . . . , K }, it follows from (2.2.7), (2.2.5), and (2.2.6) that

IQ(xo, v, t)l <2 + 2e = e

whenever Iv' = 1 and 0 < t < 6.

Recall that if L: R" , R" is a linear mapping and E C R" a measurable


set, then
IL(E)i = I det Li 1E1.
It is not difficult to extend this result to more general transformations.
Indeed, if T: R" -, R" is Lipschitz, we now know from Theorem 2.2.1
that T has a total differential almost everywhere. Moreover, if T is also
univalent, one can show that

11"[T(E)1 = L JT da for every measurable set E, (2.2.8)


52 2. Sobolev Spaces and Their Basic Properties

where JT is the Jacobian of T. From this follows the general transformation

I
formula
f o T J2` dx = f dx (2.2.9)
T(E)

whenever f is a measurable function. We refer the reader to 1F4; 3.2.31 for


a proof.
We are now in a position to discuss a bi-Lipschitzian change of coordi-
nates for Sobolev functions.

2.2.2. Theorem. L et T: R" --+ R" be a bi-Lipachatzian mapping as in


(2.2.1). if u E p> 1, then y = u o T E W l, p(V), V a' T -1 (f/),
and
Du[T(x)] • dT (x, = Dv (x) • (2.2.10)
for a.e x E SZ and for all E R".

Proof. Let u e be a sequence of regularizers for u, defined on SZ' CC SI, (see


Section 1.6). Then v e u E o T is Lipschitz on V' = T -1 (SZ') and because
v E is differentiable almost everywhere (Theorem 2.2.1), it follows that

D i v,(x) = E D,u£[T(x))D, (x)


j=1
(2.2.11)

for a.e. s E V'. Here we have used the notation T = (T 1 , T 2 , ... ,T") where
the T3 are the coordinate functions of T. They too are Lipschitz. (2.2.11)
holds at all points x at which the right side is meaningful, i.e., at all points
at which T is differentiable. If M denotes the Lipschitz constant of T, we
have from (2.2.11) that

CDv e (x)1 < n 2 MIDu c [T(x)11 for a.e. z E V'. (2.2.12)

In view of the fact that

M —n < JT(x) < M" for a.e. x E R n ,


(2.2.12) implies that there exists a constant C = C(n, M) such that

I DvL (x)IP < CI Du e [T(x)]1" . JT(x),

and therefore
jDvfrdx < C f
V' ^
from (2.2.9). A quick review of the above analysis shows that in fact, we
have
^Dv^ - Dv E ^ (Fdx < C ^Du^ - Du E ^ ^pds. (2.2.13)
v! n
2.3. Approximation of Sobolev Functions by Smooth Functions 53

Also,
Ive – ve 'Prix < lue – ue , IPdx. (2.2.14)
v' f2
From 2.2.11 we see that the regularizers te e converge to u in the norm of
W""F(fl') whenever 11' CC n. Thus, (2.2.13) and (2.2.14) imply that {y e }
is a Cauchy sequence in W L PP (st' ), and thus converges to some element
v E W 1 4P(V') with

IIv111.P;V <_ cllulli,p,n < ^Il^ali^;n. (2.2.15)

Since u E (x) — u(x) for a.e. x E SI, it is clear that y is defined on V with
v = u o T. Moreover, v E W `.P(V') whenever V' CC V and (2.2.15) shows
that, in fact, y E W" P(V). Finally, observe that (2.2.10) holds by letting
e— 0in(2.2.11).
, O

2.3 Approximation of Sobolev Functions by


Smooth Functions
From Theorem 1.6.1, we see that for each u E W k 'P(11), there is a sequence
of Cfl (S1) functions, fu e l, such that u r u in Wk4 4 (SZ 1 ) for SZ' CC OE The
purpose of the next important result is to show that a similar approximation
exists on all of SZ and not merely on compact subsets of SZ.
We first require a standard result which concerns the existence of a C°°
partition of unity subordinate to an open cover.

2.3.1. Lemma. Let E C Rn and let ÿ be a collection of open sets U such


that E C {UU : U E G). Then, there exists a family F of non-negative
functions f E Ca (Rn) such that 0 < f < 1 and
{i) for each f e F, there exists U e Ç such that spt f C U,

(ii) if K C E is compact, then spt f n K 0 for only finitely many


f E .^,

(iii)EfEF f (x) = 1 for each r E E.


Proof. Suppose first that E is compact, so that there exists a positive
integer N such that E C U; 1 U; U1 E Ç. Clearly, there exist compact sets
,

E C U, such that E C UN 1 E; . By regularizing XE, , the characteristic


;

function of Ei, there exists y, E Co (U,) such that g, > 0 on E1. Let g =
r+N
2..1=1 g, and note that g E C°° (R") and that g > 0 on some neighborhood of
E. Consequently, it is not difficult to construct a function h E C°°(R") such
that h > 0 everywhere and that h = g on E. Now let .7- = { f; : f = g;/h,
;

1 < _ < N) to obtain the desired result in case E is compact.


2. Sobolev Spaces and Their Basic Properties

54

If E is open, let

E, = E n B(0,:) n x : dist(x, 0E) > .

Thus, Ei is compact and E = U=_ 1 E, . Let ÿ i be the collection of all open


sets of the form
U n {int E i+1 - E,_2}
where U E Ç. (We take ED = E_, = 0). The elements of G, provide an
open cover for E, - int E,_ 1 and therefore possess a partition of unity
with finitely many elements. Let
00

8(x) = E s(x)
i =l gEf,
and observe that only finitely many positive terms are represented and that

_.
s(x) > 0 for x E E. A partition of unity for the open set E is obtained by
defining

f f (x)= 0..1 for some g E .F, if ,}


f(x) = 0 if x E.
If E C R n is arbitrary, then any partition of unity for the open set
{UU : U E G} provides one for E. ❑

Clearly, the set

S =Ck(Sa)n {u. ^lull^ ,p. n <oo}

is contained in WkiP(f2) and therefore, since Wk.P(12) is complete, S C


W''P(1l). The next result shows that S = Wk.F(û).

2.3.2. Theorem. The space

Ilü llk,p;^ < oc}


is dense in W k .P(SZ).

Proof. Let S2, be subdornains of Sl such that f2, CC f2,+1 and u;_°_ 1 SZ, = SZ.
Let F be a partition of unity of fl subordinate to the covering {52, +1 -SZ,- 1 },
i = 0,1, ..., where no and SZ- r are taken as the null set. Thus, if we let f,
denote the sum of the finitely many f E .7. with spt f C Sal+1 - f2,_ 1 , then
fi E Cp (fli+1 - ,li-1) and
00

E f 1 on a (2.3.1)
i^l
2.4. Sobolev Inequalities 55

Choose e > 0. For u E W k iP(S2), there exists e i > 0 such that

spt((Jiu)E,) C fi+1 - ^^-1 (2.3.2)

II(fiu }E, — , ^iullkt2 < e2".


With v ; _ (fiu) e , (2.3.2) implies that only a finite number of the v i can
fail to vanish on any given fi' CC 0, and therefore v - a°'1 vi is defined
and belongs to C°° (0). For z E Di, we have

u(x) = E fl (x)u(x),
; -1

Y(T) = E(fiu),, (x) by (2.3.2)


j =1

and consequently,
^

Ilu - vllk,P;n, < E II(f;u)f, - f,ullk,p;‘, < E .

i =1
The conclusion follows from the Monotone Convergence theorem. ❑

The approximating space C°'°(SI) n (u : ^lul^ kn < ool admits functions


that are not smooth across the boundary of SZ and therefore it is natural to
ask whether it is possible to approximate functions in Wk P(SZ) by a nicer
,

space, say
C°°(a) n {u : (IuIIk.P;st < oo}. (2.3.3)
In general, this is easily seen to be false by considering the domain n defined
as an n-ball with its equatorial (n-1)-plane deleted. The function u defined
by u - 1 on the top half-ball and u —I on the bottom half-ball is clearly
an element of WkiP(SZ) that cannot be closely approximated by an element
in (2.3.3). The difficulty here is that the domain lies on both sides of part
of its boundary. If the domain SZ possesses the segment property, it has
been shown in EAR2, Theorem 3.1$] that the space (2.3.3) is then dense in
W k, P(ft). A domain SZ has the segment property if for each x E OP, there is
an r > 0 and a vector v s E Rn such that if y E 5In$(x, r), then y+tv s E SZ
for all 0 < t < 1.

2.4 Sobolev Inequalities


One of the main objectives of this monograph is to investigate the many
inequalities that allow the L"-norm of a function to be estimated by the
norm of its partial derivatives. In this section the Sobolev inequality, which
56 2. Sobolev Spaces and Their Basic Properties

is of fundamental importance, will be established for functions in the space


W "(1). We will return to the topic of Sobolev-type inequalities in Chapter
4.

2.4.1. Theorem. Let SZ C Rn, n > 1, be an open domain. There is a


constant C = C(n,p) such that if n > p, p > 1, and u E WW'p(1/), then

IlUllnp/(n-p);fl C^IDuI^P,^,.

If p > n and SZ bounded, then u E C(SZ) and


sup Iul < 1/PIiDullp.n.

Proof. First assume that u E Cfl (SZ) and that p = 1. Clearly, for each i,
1 < i < n,
z,
lu(z)I lDiu(xi, . • - t, • ,x n )Idt ,

- 00
ith
where t occupies the component of the vector in the integrand. Therefore

n 1/n-1
o0
/n-1 < IDiuldzi . (2.4.1)
lu(x)ln
i =1 foe
If this inequality is integrated with respéct to the first variable, z 1 , and
then Holder's inequality is applied, we obtain
+ 00
/ n -1 dz1
lu(x )ln

0o i/(n-1)

< (f ID l u tt , z2 7 . . . , z„ ) Idt

00 n 1/(n-])
I Diu #^ i
dx1
- oo ^-2 (f:va

00 1/(ri-1)
< ID1i1(i,x2,...Zn)IdE
- oo
n o0 00 1/(n -1)
^ ID,uld^idzl . (2.4.2)
i=2 -oo - 00
Continuing this procedure and thus integrating (2.4.1) successively with
respect to each variable, we obtain
n t/(n -i)
n /( n )dx < I D, ujdx
l u( x )!
--,

R "
i=1 R"
2.4. Sobolev Inequalities 57

and therefore, using the fact that the geometric mean is dominated by the
arithmetic mean,

(rial ) l i n < —
n
En aJ, a? ?0,
J= 1
we have
)1/n
alulln/("-1) < i_1 ( L "
IDiu l dx E lD,uldx
R* i=i

5. ^IlDulll. (2.4.3)
n
This establishes the result in case p = 1. The result in full generality can
be obtained from (2.4.3) by replacing lui by powers of Ili!. Thus, if q > 1,

ii lu g l IIn/(n-1) <
n
L .
fID(lur)lidx

< q luI¢—'IDuldx
R^

ll l u r ^ 1 l6'11Dut!p,
q■/71'

by Holder's inequality. Now let q = (n - 1)p/(n - p) to obtain the desired


result for the case 1 < p < n and u E C4 (1l). Now assume u E WW'p(a)
and [et (u,} be a sequence of functions in C'°(1) converging to u strongly
in WW'' (iz). Then, with p' = np/(n - p), an application of the inequality
to ut - u3 yields
liui - uj lip. Cil !Ai - 741 111,p-
Thus, u, u in L" (fl) and the desired result follows. This completes the
proof in case 1 < p < n.
ln case p > n and SI bounded, let {u,} be a sequence such that u, E
Co (11) and u i — u W s "(11). The proof is thus reduced to the case when
u E C7(0). Now select x E R and because u has compact support, note
"

that
Iu(x)I < IDu(r)ldII 1 (r) (2A.4)
a=
where A x is any ray whose end-point is x. Let S" -1 (x) denote the (n - 1)-
sphere of radius 1 centered at r and denote by A = (B) the ray with end-point
x that passes through B, where B E S n—L (x). By integrating (2.4.4) over
S" - '(x) we obtain

f~ - (z)
lu(7)IdH"-1(e) <
5"-1(x)
IDu(^,)Id^l (r)dHn -s(9)
f = (e)
58 2. Sobolev Spaces and Their Basic Properties

= lDR (1)l r n-1 dH 1 ( r ) dH n-1( e )


fi -1 W a^(8) r
_ IDu(^)I1 dy (2.4.5)
Rn lx yl
where r = Ix -- y1. Thus, for any x E Rn,

1/p
w(n - 1 )1/445- IIDull') (f Ix - 1/1 (1-n)p' dy , (2.4.6)
to

where w(n - 1) = H'9Sn -1 ]. We estimate the potential on the right side


of (2.4.6) in the following way. Let B(x, R) be the ball such that 18(x, R)l =
Ispt ul. Observe that for each y E spt u - B(x, R) and z E B(x, R) - spt u,
we have
Ix
— yl(1-11)p' < Z — I Z
I(1-71)F'

and because Isptu - B(x, R)I = IB(x, R) - spt ul, it therefore follows that

I x - y l(1-n)p' dy <
Y1 (1-n)p 'dy.
apt u- E1(z,R) f (r,R)—spt u
I' -

Consequently,

Ix — yl t ( - n) p ' dy <• ix — yl(1—n)e'dy . (2.4.7)


apt ts f (z,R)

However,

1/p'
— 01—n)P1 d = ( -1 ( n )R7) 11p
Ix 1J 7 Cl
'
(2.4.8)
B(x,R)

where 'y = (1 - n)p' + n and cx(n) is the volume of the unit n-ball. But
a(n)Rn = lspt ul and therefore

(a(n)R7) 1/p' = a(n - ')Inl s ptul'/' -11 p (2.4.9)


.

The second inequality of the theorem follows from (2.4.9), (2.4.8), and
(2.4.6). To show that u E C(SZ) when p > n, let (u,} E CQ (Q) be a
sequence converging to u in W"(). Apply the second inequality of the
theorem to the difference u = - ui and obtain that lui} is fundamental in
the sup norm on.. ❑

The first part of Theorem 2.4.1 states that the LP . norm of u can be
bounded by IIuIl1,p , the Sobolev norm of u, where p• = np/(n - p). It is
2.4. Sobolev Inequalities 59

possible to bound a higher LP norm of u by utilizing higher order deriva-


tives of u as shown in the next theorem. Observe that the proof is slightly
different from that of Theorem 2.4.1 in case k = 1, p > n.

2.4.2. Theorem. Let S2 C R" be an open set. There is a constant C


C(n, k,p) such that if kp < n, p > 1, and u E Wô1p(1l), then

Hub' CIIuIIk,p;i , where p* = np/(n — kp). (2.4.10)

If kp > n, then u E CM) and

k— ^

s u p Iul < CIKI''^ p ` E (diam K)) 1a1


c^}
IlDaullp;K
1Ql=o
—1
+ (diam(K)) k k 1 )1
1 — ^ IlD k ullp,K (2.4.11)
( kp
where K = spt u and C = C(k,p,n).

Proof. When kp < n, the proof proceeds by induction on k. Observe that


Theorem 2.4.1 establishes the case k = 1.
Now assume for every y E Wô ' 3) (SZ) that

IIUII¢k-^ <- Cll^llk W t^p (2.4.12)

where
qk—Y = np/(n — kp + p)•
An application of (2.4.12) to y = D i u, 1 < j Ç n, yields

IID,u11Qt -^ <— CIID;ullk—i,p < ^Ilull^ , p - (2.4.13)

However, (2.4.12) holds with y replaced by u and this, combined with


(2.4.13), implies
1111 11 1,9 k _ 1 <- Cllu l lk,p. (2.4.14)
Since kp < n, we have qk_ i < n and therefore, Theorem 2.4.1 implies

Hu h <_ Cl l ulli.gk_, (2.4.15)

where q = ng k _ 1 /(n — qk- 1 ) = np/(n — kp). (2.4.14) and (2.4.15) give the
desired conclusion.
In order to treat the case kp > n, first assume u E C(fl) and for each
y E 3) use the Taylor expansion of u to obtain, with the notation of Section
1.1,
u(v) = (v) + Rz(u)

60 2. Sobo!ev Spaces and Their Basic Properties

where
k-1
P=(y) _E 1i D ° u(z)(y - x)a
ial = o Cr.
and
1
( y) _ kE
Rz ar
lai=k
^
[f0.-ok-iDo.(0 - t)x + ty)dt (y - x)a.

To estimate In(y)I, note that

IKI lu(y)I <_ [IP=(y)I + IR=(y)I) dx (2.4.16)

and employ Holder's inequality to obtain

k-1
%---k a
IP={v)Idx < Dan(x)(y - x) dx
x x lai=oa^

k - 1 1
< IKIIIp' E (diam K) 1°1 a.^ IlDaul^p.
;x (2.4.17)
10.1=o
=o
Similarly, to estimate the remainder term, we have

IRx(y)Idx <- (diam(K))kk E 1 (1 - t )k-1


fx lal=k
a. o K
• IDau((1 -- t)x + ty)Idxdt
1
1< (diam(K))kk
E ^ (1 - t) k-1 (l - t) -n
lark
a.
° fK`
- IDau(z)Idzdt,

where K, = T,(K) and T1 (x) = (1 - t)x +ty. Note that IK,I = (1 - t) n IKI.
Consequently, by Holder's inequality and kp > n, we obtain

I„ IRs(3+)Idx < IKI /p'(diam(K)) k laI-k


L E k
a.

1 (1 - t)k-1(1 - t)"IID ^llp;x(


^

a 1- t)n/P'dt

_
-L
n
< IK1 1 ^p (diam(K)) k k E (k IID°ullp;x,
lark 1
^. p

which, along with (2.4.16) and (2.4.17), establishes the desired inequality.
If u E tiVo'' )), let (u,) be a sequence of smooth functions converging to
2.5. The Rellich-Kondrachov Compactness Theorem 61

u in W "(f2). The application of (2.4.1) to each u i thus establishes the


inequality for u E Wô'p(13). To conclude that u E 0(f2), apply (2.4.11) to
the difference u i - ui and obtain that {u i } is fundamental in the sup norm
on S3. O

2.4.3. Remark. An important case to consider in the previous two the-


orems is fi = R". In this situation, Wk.P(R") = Wô''(R") (see Exercise
2.1) and therefore the results apply to Wk p(Rn). ,

Observe that for p > n, the proof of Theorem 2.4.1 as well as that of
Theorem 2.4.2 yields more than the fact that u is bounded. Indeed, u is
Holder continuous, which we state as a separate result.

2.4.4. Theorem. If u E Wo' p (1 ), p > n, then u E C ° ''(SZ), where ar =


1 - n/p .

Proof. Assume u E Cô (f1) and select x E 1. Let B = B(x, r) be an


arbitrary ball and choose z E B fl a Then,

lu(x) - tt(z)l < !Du(r)Ix$(r)dH'a` 1 (r)


A.(8)

where A 1 (0) is the ray whose end-point is x and passes through the point
B, 0 E S" -1 (x). Proceeding as in (2.4.5) and (2.4.6), we obtain

) 1 /p'
w(n - 1)1u(x) - u(z)I _< IIDulI p Ix -- Y1(1-n)P dy (2.4.18)
13

But,
1 /p'

U, yl(l-n)Fldy
r
= l7'la(n))1/P'rl-nip

where y and a(n) are as in (2.4.8). Since the smooth functions are dense
in Wô' P (13), we find that (2.4.18) holds for u E WW'p(II) and for almost all
x, z. ❑

An interesting aspect of the Sobolev inequality is the limiting case kp


n. This will be considered separately in Chapter 2, Section 2.4.

2.5 The Rellich—Kondrachov Compactness


Theorem
As a result of the inequalities proved in the previous section, it follows
that the Sobolev spaces Wfl'P(S3) are continuously imbedded in if (0)
where p' = np/(n - kp), if kp < n. In case kp > n, the imbedding is
62 2. Sobolev Spaces and Their Basic Properties

into the space C°(st), and if kp > n + mp, it can easily be shown that
the imbedding is into Cm (M. In this section it will be shown that the
imbedding possesses a compactness property if we allow a slightly larger
target space. Specifically, we will show that the injection map from WW'p(11)
into either L 9 (0), q < p' , or Cm(SZ) has the property that the closure of an
arbitrary closed set in Wt P (0) is compact in the range space. That is, the
image sets are precompact. We recall here that a set S in a metric space
is said to be totally bounded if for each e > 0, there are a finite number of
points in S such that the union of balls of radius e with centers at these
points contains S.

2.5.1. Theorem. Let SZ C Rn be a bounded domain. Then, if kp < n and


p > 1, WtP(11) is compactly imbedded in L 4 (0) where q < npf (n - kp). If
kp > n + mp, Wt P(11) is compactly imbedded in Cm (S2).

Proof. Consider the first part of the theorem and let B C Wt P (f) be
a bounded set. We will show that B is a compact set in L4 (11). Since
Cr (f2) is dense in WtP(fl), we may assume without loss of generality that
B C Co (11), For convenience, we will also assume that Ilullk,p;n < 1 for all
u E B.
For e > 0, let u e be the regularization of u. That is, ti e = u * Sp, where
cp, is the regularizer (see Sectipn 1.6). If u E B, then

lue(x) I <Iu(x - y)I^E(y)dy


B ( 0 ,£)
< c—n sup
cpllUlll
E - n sup {v(y) : y E R"),
and

_
IDue(x)I < Iu(x — y)I ID(Pe(y)Idy
B{ox)
€—n—i
sup{ID'(y)I : y E R" } IIuIIi
< e-n-1 sup{lDcp(y)I . y E Rn).
Therefore, if we let Bf = {u£ : u E B), it follows that BE is a bounded,
equicontinuous subset of C ° (S2). With the help of Arzela's theorem, it fol-
lows that BB is precompact in Li (SZ). Next, observe that

Iu(x) — 11 E(X)1 <


f (0,a)
Iu(x) — u(x — y)I^E(y)dY

_
1
< fmomo jDu o #y(t) •' '^ ' (t)IcpE(rJ)dtdy

< IDu(x - ty)I iyf4Pf (y)dtdy


f(0,e) 0 ]
2,5. The Rellich-Kondrachov Compactness Theorem 63

where 'y(t) = t(x -- y) + (1 - t)x = x - ty. Consequently, Fubini's theorem


leads to
p
f" 1 11 (X) - u^(x)Idx <
J3(0,0 ff" IDu(x - ty)I IyIcPE(y)dxdtdy

<^ IC l IDuIdr^e.
Thus, B is contained within an r-neighborhood of BE in L 1 (SZ). Since BE is
precompact in L 1 (fl) it is totally bounded. That is, for every r > 0, there
exist a finite number of balls in L 1 (11) of radius r whose union contains
B,. Hence, B is totally bounded and therefore precompact in L 1 (W. This
establishes the theorem in case q = 1.
If 1 < q < rap/(rn - kp), refer to (1.5.13) to obtain

iIAIIq <_ IIuII11IuII n; / n_kp)


(

where
A = 1 /g"(nrkp)/np.
1 - ( n - kp)/r^p
Then, by Theorem (2.4.2)

Huh < cII uMuIIl;PA


which implies that bounded sets in Wô''(f1) are totally bounded in Lq(SZ)
and therefore precompact.
The second part of the theorem follows immediately from Theorem 2.4.4
and Arzela's theorem in case k = 1. The general case follows from repeated
applications of this and Theorem 2.4.1. 0

2.5.2. Remark. The results of Sections 2.4 and 2.5 are stated in terms
of functions in Wa ' p (SZ). A natural and important question is to identify
those domains S2 for which the results are valid for functions in Wk.P(II).
One answer can be formulated in terms of those domains of having the
property that there exists a bounded linear operator

L : Wk,p(H) —P Wag .p(R') (2.5.1)

such that L(u)Ii= = u for all u E Wkie(f). We say that fZ is an (k,p)-


extension domain for Wk.P(i) if there exists an extension operator for
_
W k .P(12) with 1 < p < oo, k a aeon-negative integer. We will refer to
this definition extensively in Chapter 4, and if the context makes it clear
what indices k and p are under consideration, for brevity we will use the
term extension domain rather than (k, p)-extension domain. Clearly, the
results of the previous two sections are valid for u E Wk P(f) when S2
,

is a bounded extension domain. Indeed, by Lemma 2.3.1 there exists a


64 2. Sobolev Spaces and Their Basic Properties

function f E Co (R") such that f ' 1 on a Thus, if u E W 1' 4'01), then


f • L(u) E W "(SZ') where i2' is some bounded domain containing spt f. It
is now an easy matter to check that the results of the previous two sections
are valid for the space Wk1P(1) by employing Wâ111').
A fundamental result of Calderon-Stein states that every Lipschitz do-
main is an extension domain. An open set fi is a Lipschitz domain if its
boundary can be locally represented as the graph of a Lipschitz function de-
fined on some open ball of Rn'. This result was proved by Calderon [CA1]
when I < p < n and Stein [ST) extended Calderôn's result to p = 1, co.
Later, Jones (JO] introduced a class of domains that includes Lipschitz
domains, called (e, ô) domains, which he proved are extension domains
for Sobolev functions. A domain SZ is called an (e, ô) domain if whenever
x, y E R" and ix - yi < 6, there is a rectifiable arc 7 C St joining x to y
and satisfying
length 7 < F 1 Ix — v i
-

and
d(z, R" - SI) > fix zl ill
z1 for all z on 7.
ix - vi
Among the interesting results he obtained is the following: If 11 C R 2 is
finitely connected, then 0 is an extension domain if and only if it is an
(e, 6) domain for some values of e, b > 0.

2.6 Bessel Potentials and Capacity


In this section we introduce the notion of capacity which is critical in
describing the appropriate class of null sets for the treatment of pointwise
behavior of Sobolev functions which will be discussed in the following chap-
ter. We will not attempt a complete development of capacity and non-linear
potential theory which is closely related to the theory of Sobolev spaces,
for these topics deserve a treatment that lies beyond the scope of this expo-
sition. Instead, we will develop the basic properties of Bessel capacity and
refer the reader to other sources for further information, cf. [HM], [ME1],
[AD6).
The Riesz kernel, IQ , 0 < a < n, is defined by
llzla n
Ia (x) = 7 (a)
-

where
7r n / 2 2 a r(a/2)
7(a) =
F(n/2 - a/2) .
The Riesz potential of a function f is defined as the convolution

I. * f(x) 1 f (04
=-- 7(a) hin n
ix — Yin
2.6. Bessel Potentials and Capacity 65

The precise value of 7(ac) is not important for our purposes except for the
role it plays in the Riesz composition formula:

I * I0 IQ+s, a >0, Q>0, a + <n


Q

cf. [ST, p. 1181.


Observe that I. * f is lower semicontinuous whenever f > O. Indeed, if
s, z, then (x; - yia - " f (y) -, Ix -- yIQ - "f (y) for all y E R n , and lower
semicontinuity thus follows from Fatou's lemma.
The Riesz potential leads to many important applications, but for the
purpose of investigating Sobolev functions, the Bessel potential is more
suitable. For an analysis of the Bessel kernel, we refer the reader to 1ST,
Chapter 5} or [DO, Part III1 and quote here without proof the facts relevant
to our development.
The Bessel kernel, g Q , a > 0, is defined as that function whose Fourier
transform is
ga(x) = (27r) - "/ 2 (l + IxI2)-Q/2
where the Fourier transform is


A') = (21)-n/2 a-+z f(1/)dv. (2.6.1)

It is known that ga is a positive, integrable function which is analytic except


at r = O. Similar to the Riesz kernel, we have

90 * gee = 90+0, a, /3 > 0. (2.6.2)

There is an intimate connection between Bessel and Riesz potentials


which is exhibited by ga near the origin and infinity. Indeed, an analysis
shows that for some C > 0,
CIxl(I/2){a-n-1)e-Is) as 1x1 ~ oc.
9Q(x) N

Here, a(x) b(z) means that a(x)/b(x) is bounded above and below for
all large Izj. Moreover, it can be shown that
a
I as IaJ -+O
9a(x ) = xI (n
7(a)
IrIQ-n)
+o

if 0 < a < n. Thus, it follows for some constants C 1 and C2, that

g a (x) < C1 ae
IxIn -

for all x E R". Moreover, it also can be shown that

- ^ -c, l=I (2.6.4)


I D9Q (x)I e
66 2. Sobolev Spaces and Their Basic Properties

From our point of view, one of the most interesting facts concerning
Bessel potentials is that they can be employed to characterize the Sobolev
spaces W k 'p(R"). This is expressed in the following theorem where we
employ the notation

L ° 'p (R n ), a > 0, 1 < p < co

to denote all functions u such that

u= 9Q*f

for some f E Lp (R") .

2.8.1. Theorem. If k is a positive integer and l < p < oo, then


Lk,P(Rn) = Wk,p(Rtz).

Moreover, if u E L k 'p(R") with u = go * f, then

GI -1 11/11p S Ilullk,p < MG


where C = C(a, p, n).

Remark. The equivalence of the spaces Li" and Wk,p fails when p = 1 or
p=oo.
It is also interesting to observe the following dissimilarity between Bessel
and Riesz potentials. In view of the fact that IIgo III < C, Young's inequality
for convolutions implies

IlsQ * flip Cllfllp, 1 < P oo. (2.6.5)

On the other hand, we will see in Theorem 2.8.4 that the Riesz potential
satisfies
IIIQ *fllq SCIIfflp, p>1 (2.6.6)
where q np f (n — ap). However, an inequality of type (2.6.6) is possible
for only such q, cf. (Exercise 2.19), thus disallowing an inequality of type
(2.6.5) for la and for every f E L".
We now introduce the notion of capacity, which we develop in terms of
the Bessel and Riesz potentials.

2.6.2. Definition. For a > 0 and p > 1, the Bessel capacity is defined as

B p(E) = inf { flf IIp : g a * f ? 1 on E, f ? 0),


,„

whenever E C R. In case a = 0, we take Bo , p as Lebesgue measure. The


Riesz capacity, R C , p , is defined in a similar way, with g o replaced by IQ.
2.6. Bessel Potentials and Capacity 67

Since g a (x) < I0 (x), x E R", it follows immediately from definitions


that for 0 < cx < n, 1 < p < n, there exists a constant C = C(a, p, n) such
that
Ram(E) < CBQ , p (E), whenever E c R". (2.6.7)
Moreover, it can easily be shown that

110„ p (E) = 0 if and only if B 0 , p (E) = 0, (2.6.8)

(Exercise 2.5).
We now give some elementary properties of capacity.

2,6.3. Lemma. For 0 < a < n and 1 < p < oo, the following hold:

(i) 13„„,p(0) = 0,
(ii) If E l C E2, then 130 , p (E 1 ) < B a , p (E 2 ),

(iii) If E; C R' , j = 1, 2, ..., then


00

Ba,pû: 1 E. i=i

Proof. (i) and (ii) are trivial to verify. For the proof of (iii), we may assume
that E°_° 1 13 0 (E1 ) < oo. Since each term in the series is finite, for each
e > 0 there is a non-negative function fi E IP(Rn) such that

ÿa * fi > 1 on E1, Ilfillp < Ba,p(Ei) + 2 - te.


Let f(x) = sup{ f1 (a) : _ = 1, 2, ...}. Clearly, g e, * f > 1 on U;_ 1 E, and
f (z)" < fi (x)F. Therefore,
00 00 00

Bar , p Û Ei <_ llf I^ ^E E Ba,p(Ei) + F. ❑

i =t i-1 i -1

Another useful characterization of capacity is t:.e following:

B a , p (E) = inf{ inf gQ * f = {sup inf y a * f (x)} -1)


J xEE f xEE

where f E IP(R"), f > 0 and11 flip _< 1 (Exercise 2.4).


Although Lemma 2.6.3 states that Ba , p is an outer measure, it is fruitless
to attempt a development in the context of measure theory because it can
be shown that there is no adequate supply of measurable sets. Rather, we
will establish other properties that show that the appropriate context for
68 2. Sobolev Spaces and Their Basic Properties

B. is the theory of capacity, as developed by Brelot, Choquet, [BRIT


[CH).

2.8.4. Lemma. If IL} is a sequence in I7(R") such that — f II -' 0


as i -, oo, p > I, then there is a subsequence {f, } such that

9o* fi,(x) 9a*f(x)

for Ba , p -q.e. x E Rn.

(We employ the time-honored convention of stating that a condition


holds BQ , p -q.e., an abbreviation for B a , p -quasi everywhere, if it holds at all
points except possibly for a set of /3 0 , p -capacity zero.)

Proof. It follows easily from the definition of B a , p capacity that if f E


IF(R" ), then Igo * f(x)1 < oo for B Q , p -q. e. x E Rn . Thus, for E > 0,

Ba,p(ix : fga * fi(x) ` 9a * f (x)I e} = BQ,p({x : I9a * (fi -- f)(x)I E})


-' IIfi — fllp.
Consequently, there exists a subsequence {A}} and a sequence of sets E
such that

Igo * fi, (l) - 9a * f(Z)I 1, x E R" - E ^,

with
Ba,p(Ej) < e2 -3 .
Hence, ga * f; -► g a * f uniformly on R" - U? _ 1 E3 , where Ba,, p (U °_ 1 E) ) _<
?

e. Now a standard diagonaiization process yields the conclusion. D

2.8.5. Lemma. If { f } is a sequence in L" (R71 ), p > 1, such that fi -e f


;

weakly in L"(R" ), then

lim inf 9Q * fi (x) < g a * f (x) < lim sup 9 a * f i ( (2.6.10)

for 13„, p -q.e. x E R". If :n a ddition, it is assumed that each f > 0, then
9a * f (x) < lim inf g a * fi (x) for a E R" (2.6.1 1)
ti^oo

and
g o * f (x) = Lim inf 9 a * fi (x) (2.6. I2)
i.00
for B a , p -q.e. x E Rn.

Proof. Under the assumption that f — f weakly in V'(Rn), by the


;

Banach-Saks theorem there exists a subsequence of IL} (which will be


2.6. Bessel Potentials and Capacity 69

denoted by the full sequence) such that

gi - i-1 E f f
3= i

converges strongly in I'(R") to f . Lemma 2.6.4 thus yields a subsequence


of {g; } (denoted by the full sequence) such that

9Q * f(x) = f lirn
^oo
g o, * gi(x)

for Ba, p-q.e. z E Rn. However, for each z E R73 ,

lim inf g a * fc(x) < lim g o * Mx),


s-.00 - i -.00
which establishes the first inequality in (2.6.10). The second part of (2.6.10)
follows from the first by replacing f and f by --L and -f respectively.
;

In the complement of any ball, B, containing the origin, IIg 0 IPp. ; R'_B <
co, by (2.6.3). Thus, (2.6.11) follows from the weak convergence of f; to f.
(2.6.12) follows from (2.6.11) and (2.6.10). D

2.6.6. Lemma. For every set E C Rn


B 0 , p (E) = inf { B a , p (U) : U D E,U open).
.

Proof. Since g o is continuous away from the origin, the proof of the lower
semicontinuity of ga *f when f ? 0 is similar to that for the Riesz potential
given at the beginning of this section. The lemma follows immediately from
this observation.

The lemma states that B a , p is outer regular. To obtain inner regularity


on a large class of sets, we will require the following continuity properties
of B0 , p .

2.6.7. Theorem. If {E; } is a sequence of subsets of Rn, then


Bar ,p ^ lim inf E } < lim inf Ba p (E;). (2.6.13)
; ,

—• 00 —*CO

If E10 E2C ..., then

Ba,p u E; = lim B Q . p (E;). (2.6.14)


-^
i_-.1.
If IC I D K2 D ... are compact sets, then
(00
p K; = lirn Ba,p(K;). (2.6.15)
i__.00
c= ^
70 2. Sobolev Spaces and Their Basic Propel. tm..

Proof. For the proof of (2.6.14) assume that the limit is finite and let f,
be a non-negative function in L" (Rh) such that ga * fi > 1 on Ei with

Ilfillp < BQ.p(Et) + 1/i. (2.6.16)

Since Il fi if p is a bounded sequence of real numbers, Theorem 1.5.2 asserts


the existence of f E LP(Rn) and a subsequence of {AI} that converges
weakly to f. Hence, (2.6.12) implies that there exists a set B C E = U°__ L Ei
with B a , p (E — B) = 0 such that ga *f > 1 on B. Therefore,

Ba,p(E) = Ba,p(B) < Ilf llp


<limif llf1llp
< lim B a , p ( Ei ),
i—.00

from (2.6.16). If

un
^ o0

Ai= Ek,
J= i k= ^

then {A,} is an ascending sequence of sets whose union equals lirn inf E,.
Therefore, since Ai C Ei for i > 1, (2.6.14) implies (2.6.13) because
00

Ba,p inf E,} = Ba , p (U Ai)


i^oo
i =f
= Jim B a p (A,)
i^ao

< lim inf B a , p (E,).


i—oo

Finally, it {K;} is a descending sequence of compact sets, Lemma 2.6.6


provides an open set U D fls_ 1 K i such that

Ba,p(U) < Ba,p +f

for an arbitrarily chosen e > O. However, K, C U for all sufficiently large à


and consequently B a , p (K,) < B a , p (U). (2.6.15) is now immediate and the
proof is complete, y ❑

(2.6.14) states that Bail) is left-continuous on arbitrary sets whereas


(2.6.15) states that B a , p is right continuous on compact sets. The impor-
tance of these two facts is seen in a fundamental result of Choquet [CH,
Theorem 11 which we state without proof.

2.6.8. Theorem. Let C be a non-negative set function defined on the Bord


sets in R n with the following properties:
2.6. Bessel Potentials and Capacity 71

(i) C(0) = 0,
(ii) If B1 C B2 are Borel sets, then C(B 1 ) < C(B2 ),

(iii) If (13; } is a sequence of Borel sets, then C (U°° 1 BI ) < Er° C(B ),
y ;

(iv) C is left continuous on arbitrary sets and right continuous on compact


sets.
Then, for any Suslin set A C Rn ,
sup{C(K) : K C A, K compact} = inf{C(U) : U D A, U open}.

Any set A for which the conclusion of the theorem applies is called C-
capacitable. In view of Lemma 2.6.3 and Theorem 2.6.8, the following is
immediate.

2.6.9. Corollary. All Suslin sets are B a , p -capacitable.

The usefulness of Theorem 2.6.8 and its attending corollary is quite clear,
for it reduces many questions concerning capacity to the analysis of its
behavior on compact sets.
We now introduce what will eventually result in an equivalent formula-
tion of Bessel capacity.

2.6.10. Definition. For 1 < p < oo, and E C Rn a Suslin set, let M(E)
denote the class of Radon measures thon Rn such that p(R" — E) = O. We
define

bQ,p(E) = sup{p(Rn) } (2.6.17)


where the supremum is taken over all A E M(E) such that
Ilga * 1.4p' < 1. (16.18)

Clearly,
b,„, p (E) = (inf{Ilga * ^I^p'}) (2.6.19)
where the infimum is taken over all u E M(E) with v(R") = 1. We have
that
{L .

II9Q * r.IÎ p ' = sup 9a *T/• f dZ: f> o, 1IfI1p Ç 1

sup
R^
g^ *f du: f?o, VG ^
l
J1
and thus obtain

b a , F (E) -1 = ( inîsuJ 9a * f dy (2.6.20)


72 2. Sobolev Spaces and Their Basic Properties

where y E M(E), v(R") = 1, and f > 0 with IIf IIp < 1.


Recall from (2.6.9) that if E C R", then

13., p (E) = (sup inf 9. * f (x) } _p


f xEE

where f E LP(R"), f > 0 and I fli p < 1. By considering measures concen-


trated at points, this is easily seen to be

BQ , p (E) -l ip — sup inf g a * f dv (2.6.21)


f y

where f and v are the same as in (2.6.20).


We would like to conclude that there is equality between (2.6.20) and
(2.6.21). For this purpose, assume E C R" is a compact set and let

F( f, v) = f g * f dv
Q (2.6.22)

where f E L"(R"), f > 0, I f lIp < 1 and v E M(E), v(R") = 1. Clearly


F is linear in each variable and is lower semicontinuous in v relative to
weak convergence. Since the spaces in which f and v vary are compact we
may apply the following minimax theorem, which we state without proof,
to obtain our conclusion, EFA1.

Minimax Theorem. Let X be a compact Hausdorff space and Y an ar-


bitrary set. Let F be a real-valued function on X x Y such that, for every
y E Y, F(x, y) is lower semicontinuous on X. If F is convex on X and
concave on Y, then

inf sup F(x, y) = sup inf F(x, yy).


sEX yEY yEY zEX

We thus obtain the following result.

2.6.11. Lemma. If K C Rn is compact, then



[ba,p(K)? = Ba,p(K). (2.6.23)

Our next task is to extend (2.6.23) to a more general class of sets. For
this purpose, observe that if E C Rn is a Suslin set, then

b Q , p (E) = sup{b a , p (K) : K C E, K compact). (2.6.24)

To see this, for each Suslin set E, let p E M(E) with IIg. * µll p ' < 1. If
K C E is compact, then v — uI K has the property that v E .A4 (E) with
Ilga *vllp < 1. Since p is a regular measure, we have
,

p(E) = sup{p(K) K c E,K compact},


2.6. Bessel Potentials and Capacity 73

and therefore

bk, p (E) = sup{bk, p (K) : K C E, K compact}. (2.6.25)

From (2.6.25), (2.6.23), and Corollary 2.6.9 we conclude the following.

2.6.12. Theorem. If E C R" is a Suslin sct, then

[byi.p (E)jp = Bck,e(E)•

Thus far, we have developed the set-theoretic properties of Ba , p . We now


will investigate its metric properties.

2.8.13. Theorem. For p > 1, ap < n, there exists a constant C =


C(a, p, n) such that
C -l r n-"
P < BQ,P[B(x r)] < Cr"'"
,

whenever z E Rn and 0 < r < 1/2.

Proof. Without loss of generality, we will prove the theorem only for B(0, r)
and write B(r) = B(0, r). Let f E I1(R"), f > 0, have the property that

g * f > 1 on B(2).
Q (2.6.26)

By a change of variable, this implies

gQ r y f ^M)r - ndy^ 1 (2.6.27)


Rn

for x E B(2r). From (2.6.3) and (2.6.4), there exists C = C(a, p, n) such
that

Cl x - yl a-n e - 'Ix- E/! < ga (z - 3/ ) < Ctx -- yIQ - ne -12- v 1 ,

and therefore

-y)< GIx — yIQ-nrn-Qe-1t-U1r 1


90 (L---■-
r

< Clx - y la-n r e -vi (7- 1/2)


n-ck -21x
< C 2 rn — Qgc,(z (r < 1/2).
— y)
Consequently, from (2.6.27),

C2 g0 (x - y) f ( r7-1 ) r - ady > 1 for z E B(2r), (r < 1/2).


R
74 2. Sobolev Spaces and Their Basic Properties

However,

Hence,
J R. [C3r-cif ()}'dy = C
2P r n-°' p 11f11p.

Ba,P[B(2r)i < C2prn-ap ll f IIp, r < 1/2,


for every f E L P (R") satisfying (2.6.26). Thus,
2P r' -aP
$a,P[B( 2 r)] < C Ba,p[B( 2 )) , r < 1 / 2,

from which the conclusion follows.


For the proof of the first inequality of the theorem, let f E LP(R' ),
f > 0, be such that g.* f > 1 on B(r). Then

I B(r)I f s f dx < IB(r)I l/4 119a * f1Iq'


B (r)
where q = p` = np/(n —ap). It follows from (2.6.3) that g a < CIa . Because
there is no danger of a circular argument, we employ the Sobolev inequality
for Riesz potentials (Theorem 2.8.4) to obtain
r n-aP < CIA'
Taking the infimum over all such f establishes the desired inequality. ❑

The case ap > n requires special treatment.

2.6.14. Theorem. If p > 1, ap _ n and 0 < r < I, there exists C =


C(n, r) such that
C` 1 (logr -1 ) 1- P < B.,P[B(x , r)) < C(logr -11 )
L_P

whenever O<r <ir< I and x R' 1 .

Proof. As in the proof of the previous theorem, it suffices to consider only


the case z = O. Let be a Radon measure such that µ[R" — B(r)) = 0 and
119a * < 1, where we write B(r) = B(x,r). Because of the similarity
between the Riesz and Bessel kernels discussed at the beginning of this
section, there exists a constant C independent of r such that

(IQ * µ)Pax<C (9a*µ) 1 dx<C.


f (1) R"

If ly1 < r and 1x1 ? r, then Ix — y1 < (xi + 1Y1 < 1x1 + r < 21x1 and
therefore

(1,
P'
C> _ (Ia*µ)Pdx— Ix - yl a- "dtA(y) dx
r <^x^< 1 f^I1^1

> C1[p(Rn))p ,

f
1x1 —n dx
<,=,«
= C1[u(R°)lps [logr -i ^.
2.6. Bessel Potentials and Capacity 75

Thus, by Theorem 2.6.12, it follows that


) 1- p.
B°,p[B(r)] < C(logr -1
To establish the opposite inequality, let a r denote the restriction of
Lebesgue measure to B(r). Since g a < I° , we have

ga * Ar(x) < C f ix - yla`ndy. (2.6.28)


B(r)
If 1x1 < r/r , NI < r, then Ix - yi < cr where c(r) = 1 + 1/i. That is,
B(r) C B(x, cr). Therefore,
"dy
lx - 1,1`2- - vl °-n dy
f (r) B(s,cr)

< C(r)r°
which, by (2.6.28), implies

9a * A r (x) < C(f)r ° if lx' < r/F. (2.6.29)

If l y j < r and r/f < Izi < 1, then Ex - yi > IxI - lyi > lx^ - r > e(F)1x1,
where now c(f) = 1 - F. Hence,

9a * Ar (x) < C Î x - y iandy < C1 rn fxIa-n if r/t < Ix' < 1. (2.6.30)
B(r)
If lx1 > 1, then (2.6.3) yields

go * Ar(x) < Cr n e -Izt .


-

(2.6.31)

Thus, (2.6.29), (2.6.30), and (2.6.31) yield

1i * A. li p, < Crn (log r-2)1/p

Appealing again to Theorem 2.6.12, we establish the desired result. ❑


2.6.15. Remark. in case ap > n, it is not difficult to show that there is a
constant C = C(ar, p, n) such that

B a , p (E) > C

whenever E 0. See Exercise 2.6.


Because Ba , p [B(x, r)) .:r T r". " one would expect that Bessel capacity
and Hausdorff measure are related. This is indeed the case as seen by the
following theorem that we state without proof, [ME1), [HM). See Exercises
2.15 and 2.16.

2.6.16. Theorem. If p > 1 and ap < n, then B0 (H) = 0 if H" - °p(E) <
oo. Conversely, if 13., p (E) = 0, then H" -aP+E (E) = 0 for every e > O.
76 2. Sobolev Spaces and Their Basic Properties

2.7 The Best Constant in the Sobolev Inequality


There is a fundamental relationship between the classical isoperimetric in-
equality for subsets of Euclidean space and the Sobolev inequality in the
case p = 1. Indeed, it was shown in [FF] that the former implies the latter
and, as we shall see in Remark 2.7.5 below, the converse is easily seen to
hold.
We will give a method that gives the best constant in the Sobolev In-
equality (Theorem 2.4.1), by employing an argument that depends critically
on a suitable interpretation of the total variation for functions of several
variables. This is presented in Theorem 2.7.1 and equality (2.7.1) if referred
to as the co-area formula. Tbis is a very useful tool in analysis that has
seen many applications. We will give a proof for only smooth functions but
this will be sufficient for our purposes.

2.7.1. Theorem. Let u E Co (R"). Then

I„
+oo
[Duldx =Hn -1 [u -1 (t) n S^]dt. (2.7.1)
f 00

Before giving the proof of this theorem, let us first consider some of
its interpretations. In case n = 1, the integrand on the right-hand side
involves Hausdorff 0-dimensional measure, H ° . H ° (E) is merely the number
of points (including oo) in E and thus, the integrand on the right side of
(2.7.1) gives the number of points m the set u -1 (t) n O. This is equivalent
to the number of times the graph of u, when considered as a subset of
R 2 = {(x, y)}, intersects the line y = t. In this case (2.7.1) becomes

L u'1 f N(y)dy (2.7.2)

where N(y) denotes the number of points in u -1 (y) n Il. (2.7.2) is known
as the Banach Indicatrix formula, [SK, p. 284
The Morse-Sand Theorem [MSE1], [SA], states that a real-valued func-
tion u of class Cn defined on Rn has the property that H 1 [u(N)] = 0 where
N = {x : Du(x) = 0}. For example, if we consider a function u E Co (R2),
an application of the Implicit Function theorem implies that u -1 (t) n S2 is
a 1-dimensional class C 2 manifold for a.e. t. In this case, H 1 (u -1 (t) r1 S2]
is the length of the curve obtained by intersecting the graph of u in Ra by
f
the hyperplane z = t. Thus, the variation of u, n IDuodx, is obtained by
integrating the length of the curves, f ri u ' (t), with respect to t.
-

The co-area formula is known to be valid for Lipschitzian functions. (We


will see in Chapter 5, that another version is valid for BV functions.) The
proof in its complete generality requires a delicate argument from geometric
measure theory that will not be given here. The main obstacle in the proof
2.7. The Best Constant in the Sobolev Inequality 77

is to show that if u is Lipschitz, then

H"-1[u-1(^) n N]dt = 0
f'
where N = {x : D(x) = 0}. Once this has been established, the remainder
of the proof follows from standard arguments. Because our result assumes
that u E C", we avoid this difficulty by appealing to the Morse-Sard the-
orem referred to above. In preparation for the proof, we first require the
following lemma.

2.7.2. Lemma. If U C R" is a bounded, open set with C 2 boundary, then

jdiv
sup dx : cp E C (R";R"), supp] <_ 1 = H" 1 [8U].
i
{

Proof. By the Gauss-Green theorem,

1 divSp dxj,(x).
!c r^(x)dH"
U
I (x)

where v is the unit exterior normal. Hence

sup div (09 dx :çP E CQ (R"; R"), suP iWi < 1 < Hit-1 (8U).
u

To prove the opposite inequality, note that v is a C 1 vector field of unit


length defined on ôU and so may be extended to a C' vector field V
defined on R" such that IV (x)I < 1 for all x E R", cf. Theorem 3.6.2. If
tke Cô (R") and 10l <1, then with ço=tkV,we have

J div Sp dx =t/J(y)dH" - 1 (y)


fu
so that

sup div cp : cp E Co (R"; R"), sup 11,01 < 1


u

> supt, dH"- 1 : t € Co (R"), sup I I < 1 = Hn -1 (8U). O


ou

Proof of Theorem 2.7.1. We first consider linear maps L: R" - R 1 .


Then there exists an orthogonal transformation f : R" -+ R" and a non-
singular transformation g such that f (N 1 ) = R 1 , f (N) _ R"'', (N =
ker L) and
L —gopo f
78 2. Sobolev Spaces and Their Basic Properties

where p: R" -. R 1 is the projection. For each y E R 1 , p -1 (y) is a hy-


perplane that is a translate of the subspace p -1 (0). The inverse images
p -1 (y) decompose Rn into parallel (n - 1)-dimensional slices and an easy
application of Fubini's theorem yields

IEI = R Hn-1[E np 1 (y)]dy (2.7.3)


e
whenever E is a measurable subset of R". Therefore

H"-1[E n p
If(E)1 — IEI = R1
-1 (y)]dy

1[f
"- (E) n p - 1 (v)Ïdy
-fRI H
-1 (p
-H n-1 (E n f -1 (y))1dy•
fe
Now use the change of variables z = g(y) and observe that the last integral
above becomes

Hn -1 [E n f -1 (p - 1
Ig'IIEI - R^
(g -1 (z)))Jdz

-
f„, H "-1[En L-1(z)]dz. (2.7.4)

But WI = IDLI and thus (2.7.4) establishes Theorem 2.7.1 for linear maps.
We now proceed to prove the result for general u as stated in the theorem.
Let N = {x : Du(x) = 0} and for each t E R 1 , let

Et-R"n{x: u(x) >t}

and define a function fe : R" -+ R 1 by

ft = )(Et, if t >0
XR" -E, if t < 0.
-

Thus,
u(x) =ft (x)dt, x E R".
R1
Now consider a test function Sp E Cô (Rn N), such that sup IV, < 1.
-

Then, by Fubini's theorem,

t (x)Sp(x)dtdx
f" u(x)co(x)dx
R" R
=f
I

f
= fe(x)cp(x)dxdt.
, R"
(2.7.5)
2.7. The Best Constant in the Sobolev Inequality 79

Now (2.7.5) remains valid if cp is replaced by any one of its first partial
derivatives. Since Du # 0 in the open set R" - N, the Implicit Function
theorem implies that u -1 (t) n (R" -N) is an (n - 1)-manifold of class Cn.
In addition, since spt cp C R" - N, it follows from the Divergence theorem
that
div tip dx = yo(r) • v(x)dHn ' (x).
JôE e )fl(R" -N)
-

Ee

Therefore, if cp is now taken as cp E Cô (R" - N; R n ) with sup kpl < 1, we


have

- Du gp dx = u div cp dx = div tic) d


R" R " Et
R ' E
= cp(x) v(x)d1-1n -1 (x)dt
Rt fR" -N)nBEt
H " 'RR" — N) n u -1 (t)ldt
JR'
-

fin - 1 [u E (t)]dt. (2.7.6)


f
-

however, the sup of (2.7.6) over all such cp equals

Du^dx =
R" -N R"
In order to prove the opposite inequality, let Lk: R" -e R 1 be piecewise
linear maps such that

lire ILk - uldx = 0 (2.7.7)


k-1°°o R "

and
^l iyrn IDLk id x = ^ 1Duldx. (2.7.8)
R n R
Let
E^ — R" n {z : L k (x) > t ?,
Xé = XE •
From (2.7.7) it follows that there is a countable set S C R 1 such that

0 (2.7.9)
k'oo f a lX t — Xé ldx =
lim

whenever t ft S. By the Morse -Sard theorem and the Implicit Function


theorem, we have that u -1 (t) is a closed manifold of class C" for all t E
R 1 - T where 1-1 1 (T) = O. Redefine the set S to also include T. Thus, for
80 2. Sobolev Spaces and Their Basic Properties

t g S, and e > 0, refer to Lemma 2.7.2 to find (P E C7( R n ; Rn) such that
IV, < 1 and
lin-i[u-1(t)1 — divcpdx < (2.7.10)
l
E : 2.
Let M = fjr, jdiv coldx and choose ko such that for k > k o ,

iXt
2 — X:
M idx < .
R"
For k > k o ,

div cP dx — div çP dx < M iXt — Xi 1dz < 2 . (2.7.11)


f E t
t R"

Therefore, from (2.7.10) and (2.7.11)

H "-1[u-1(t)] <L
t d1 dx+

< div 5pdx+ e


E^k

— (P• v dHn -1 + e
8E=
< fIn-t[Lk-1 (t)1 + e.

Thus, for t ft S,

li n -1 [tr i (t)] < lip inf fi n -1 [L i

Fatou's lemma, (2.7.8), and (2.7.4) imply

lin -1 {u -1 (t}]dt < lirn inf Hn-'[Lk'(t)]dt


f ^ -
k' Ri

<lim inf IDLk^dx


.

- k ^ oo R"

_'Duldx. CI
R"

Theorem 2.7.1 is a special case of a more general version developed by


Federer [F1] which we state without proof.

2.7.3. Theorem. If X and Y are separable Riemannian manifolds of class


1 with
dim X = m > k = dim Y
2.7. The Best Constant in the Sobolev Inequality 81

and f : X --+ Y is a Lipschitzian map, then

I, ^

f(x)dHm(=) =
Y
n f -1 (31))dH k (g)

whenever A C X is an Hm-measurable set. Moreover, if g is an Hu'


integrable function on X, then

g(x).If
(x)dHm (x) =
J f -'(!I)
g(x)dHm-k(a)dHk (31).

Here, .1 f (x) denotes the square root of the sum of squares of the deter-
minant of the k x k minors of Jacobian matrix of f at z.
The proof of Theorem 2.7.1 above is patterned after the one by Fleming
and Rishel [FR] which establishes a similar result for BV functions. Their
result will be presented in Chapter 5.
We now give another proof of Theorem 2.4.1 that yields the best constant
in the case p 1.

2.7.4. Theorem. If u E Co (R"), then

IIUIIn / {n - 1) < n - 1 ^(n) - 1^


nIID^II.

Proof. For t > Q, let

A t = {x : Iu(x)I > = {x : Iu(x)I = t}

and let u t be the function obtained from u by truncation at heights t and


—t. If
1(0 = -1),
then clearly
I^t +^ I <_ Iutl + hXa.
f (t + h) < f(t) + htA1l(n-Min (2.7.12)
for h > Q. It follows from the Morse-Sard theorem that for a.e. t > 0, B t is
an (n — 1)-dimensional manifold of class oo and therefore, an application
of the classical isoperimetric inequality yields
-1 )/n < n -l a ( n )- 1 /nHn-1(B t ) . (2.7.13)
IAtI(n

It follows from (2.7.12) that f is an absolutely continuous function with


f'(t) < IAt+(n -1)/n
82 2. Sobolev Spaces and Their Basic Properties

for a.e. t. Therefore, with the aid of (2.7.13), it follows that


(n-1)/n
Iuln/(n-1)dx = f ( Do ) -- Ï (o)
00

= f`(t)dt
0
< n -1 a(n) -l in Nn-1[13t1dt.
f0 00

The co-area formula, Theorem 2.7.1, shows that the last integral equals

l DuldX,
R^
thus establishing the theorem.

From the inequality

IIuIIn/(n-1) < nla(n)-l/nlIDull (2.7.14)

one can deduce the inequality

IIuIÏP• < np(n - 1)/(n - p) II D u IIP (2.7.15)

by replacing u in (2.7.14) by uQ where q = p(n - 1)/(n -- p). Then


(n-1)/n
^nP/tn-P)d^ < n'a(n) -1 /nq lu1 4 - i lDulda

(P 1 )/ v
5_ n -l a(n) -E/n q unP/(n-P)dx
-

IIDuIIP

by Holder's inequality.
Of course, one cannot expect the constant in (2,7.15) to be optimal.
Indeed, Talenti [TA) has shown that the best constant C(n, p) is
1- ( 1 /p) r(1 + 1/n
C(n,p) =7r-1/2n-1/2
P - 1
n - P L (n/2))r(n)
r(n/pr(I^ n- (n/p))
where 1 < p < n.

2.7.5. Remark. The proof of Theorem 2.7.4 reveals that the classical
isoperimetric inequality implies the validity of the Sobolev inequality when
p = 1. It is not difficult to see that the converse is also true.
To that end let K C Rn be a compact set with smooth boundary. Let
dK(z) denote the distance from x to K,

dK(x) = inf{lx - y) . y E K }.
2.8. Alternate Proofs of the Fundamental Inequalities 83

It is well-known and easy to verify that dK(x) is a Lipschitz function with


Lipschitz constant 1. (See Exercise 1.1.) Moreover, Rademacher's theorem
(Theorem 2.2.1) implies that dK is totally differentiable at almost every
point x with IDdk (x) f = 1 for a.e. x E f?. For each h > 0, let

Fh(x) = 1 - min[dK(x), h} • h -1

and observe that Fh is a Lipschitz function such that


(1) Fh(x) =Iif x E K

(ii) Fh(x) 0 if dK(x) ? h

(iii) IDFh(z)i < h - for a.e. x E Rn.


'

By standard smoothing techniques, Theorem 2.7.4 is valid for Fh because


Fh is Lipschitz. Therefore

t{x :0 < dK(x) < NI


(11(1)(n-1)(n) < n-'a(n)-1fn
h
Since lDdK (x)I = 1 for a.e. x E R", the co-area formula for Lipschitz maps,
Theorem 2.7.3, implies that

I{x : 0 < dK(x) < h}I 1 IDdKIdz


h h o<a,<<hf {

I h
Hr [dxi (t))dt
0
= Hn - ' [di' (t h ))

where 0 < t f, < h. Because K is smoothly bounded, it follows that


Hn -1[dxl(th)) -0 H" -1 (8K) as h -' 0
and thus, the isoperimetric inequality is established.
Of course, by appealing to some of the more powerful methods in geo-
metric measure theory, the argument above could be employed to cover the
case where the compact set K is a Lipschitz domain. By appealing to the
properties of Minkowski content, cf. [F4, Section 3.2.391, it can be shown
that the above proof still remains valid.

2.8 Alternate Proofs of the Fundamental


Inequalities
In this section another proof of the Sobolev inequality (2.4.10) is given
which is based on the Hardy-Littlewood-Wiener maximal theorem. This
84 2. Sobolev Spaces and Their Basic Properties

approach will be used in Section 2.9, where the inequality will be treated
in the case of critical indices, kp = n.
We begin by proving the Hardy-Littlewood-Wiener maximal theorem.

2.8.1. Definition. Let f be a locally integrable function defined on R'a.


The maximal function of f, M(f), is defined by

M(f )(x) = supIf(y)Idy : r > 0 .


B (r,r)

2.8.2. Theorem. If f E L p (R n ), 1 < p < oo, then M(f) E I7(Rn) and


there exists a constant C = C(p, n) such that

IIM(f)llp < CIJfllp.

Proof. For each t ER', let A t ={x:M(f)(x)>t}. Prom Definition 2.8.1


it follows that for each x E A t , there exists a ball with center x E A t , such
that
Ifl dy > t. (2.8.1)
iBs=x
If we let F be the family of n-balls defined by F _ {El. : x E At}, then The-
orem 1.3.1 provides the existence of a disjoint subfamily {B1, B2, ... , Bk, .. .}
such that
00

EIBkI?5 - nlAtl
k =1
and therefore, from (2.8.1),
DO

IIfII I ? lfldy > t E l$kI ? t5IAtI ,


fu r . 1 Bk k=1
or n
lAt f ç 5 Ill
7 I11 whenever t E R I . (2.8.2)

We now assume that 1 < p < oo, for the conclusion of the theorem
obviously holds in case p = oo. For each t E R', define

ft(x) = f (x) if I f(x)I > t/2


0 if If(x)1 < t/2.
Then, for all x,
If(x)1 < I f:(x)I + t/2,
M(f)(x) < M(f t )(x) + t/2
and thus,
{x : M(f)(x) > t} C {M(ft )(x) > t/2 }.
2.8. Alternate Proofs of the Fundamental Inequalities 85

Applying (2.8.2) with f replaced by ft yields

2 • t5R 5^
'Ad <_ III (ft)(r) > t/2}I <
_ If:ldy = 2 If ldy.
R n t ffill>t/2)
(2.8.3)
Now, from Lemma 1.5.1, and (2.8.3),
00
L " M fY =Adt
o
00
_ p tp-lIAeIdt
0
00
< p2 • 5" 9 -2 I f(d^ dt
o flifi>02}

p2p • 5" tv-2 f{ifi>e} If ldx dt
o
00
= p 2P • 5" tP-2p({1f Î > t})dt
0

where p is a measure defined by p(E) = fE !Mr for every Borel set E.


Thus, appealing again to Lemma 1.5.1, we have
p2p 5n oo
(Mf)pdx = o f
i
u(iifl > t }) dtp
-

fRn p -1
p 2p.5" 1

p — 1 JR» If lp du
=
p 2p . 5"
__
p 1 IRR
Since p > 1. This establishes the theorem. ❑

For 0 < a < n, we recall from Section 2.6 the definition of the Riesz
potential of f of order a:

L* f(x) = ici (x)= 1 f (y)du .


^`(^) hin — yin —a
The following lemma is the final ingredient necessary to establish the
Sobolev inequality for Riesz potentials.
2.8.3. Lemma. If 0 < a < n, 0 > 0, and 6 > 0, then there is a constant
C = C(n) such that for each z E R",
If(y)Idu < C6°M(f)(z)
(i) yin-" —
B(:.6) Ix —
86 2. Sobolev Spaces and Their Basic Properties

(ii)
If(y)Idy < C 6-13
M(f)(x).
I
1" -B(z,6) x 11I

Proof. Only (i) will be proved since the proof of (ii) is similar. For r E Rn
and a > 0, let the annulus be denoted by

b ô ë b
..4 x, 2k ' 2k + i =
B x B x, 2k +1
2k -
,

and note that

I If(^)Id y
(z,6) Ix —
yln -a = ^ À(=
k-o
vo
,

6 a-n
^ ^)
;^ Ix
If(y)Id^
yIn a-

< 2k^.1 e(T.*)Iflix


k_p
ao ( i )a—n ^
=a(n)
i L(Z,f1dT
^
(

k0 1
^ C 6a M(f)(z)+
where a(n) denotes the volume of the unit n-ball. This proves (i). ❑

We now will see that the Sobolev inequality for Riesz potentials is an
easy consequence of the above results.

2.8.4. Theorem. Let a > 0, 1 < p < oo, and ap < n. Then, there is a
constant C = C(n,p) such that
np
Ifla(1 )IIps 5 P. = n ap s

whenever f E L(Rn ).

Proof. For 6 > 0, Holder's inequality implies that


1/p'
If(ll)I n-1-p'(n-Q) dé
R _$( x 5fix
,
I yin-a
dy < w n ( - 1 )IIf IIp (16 °13r
where r = Ix - gel. The integral on the right is dominated by ë" -(n fp ) since
ap < n, and therefore, by Lemma 2.8.3(i),

Ila(f)(x)I < C [ 6 M(f)(x) + t a-(Tt ) IIfIIp] • (2.8.4)

I f we choose
d=
II/11p ) ,

( M(f)(x)\-Din

2.8. Alternate Proofs of the Fundamental Inequalitics 87

then (2.8.4) becomes

Ifa(f)(x)I<_ C[M(f)(x)r (°P/n) Ilf llp /n


or,
— CIm(f)(x)]PIIIIIpaP / n ) P • .
Ifa(f)(x)IP. <
An application of Theorem 2.8.2 now yields the desired conclusion. Q

2.8.5. Remark. If we are willing to settle for a slightly weaker result in


Theorem 2.8.4, an easy proof is available that also provides an estimate of
the constant C that appears on the right-hand side of the inequality. Thus,
if I/ is a domain with finite measure, f E LP(Q), and p < q < p', we can
obtain a bound on IIII(f )IIq by a method that essentially depends only on
Holder's inequality.
For this purpose, let T = 1 — (p — q) and note that because q < p',
-n
I x - yI a € Lr(n) (2.8.5)
for each fixed x E R. As in the proof of (2.4.7), if I 13(x, R)I IS11, then
WR ( - 1)R(a-tt)r +n
A
I^ _ ^I(a-ri}rd^
fB (s,R)
k—y
,(a-n)r dy = l
(cY -n)r+ n

w(n -1)IS/I''' Oa, r, 3 0 - (2.8.6)


' Rot - n)r + nia(n)7 =-
where 7 = Rot — n)r)/n + 1. For each fixed x, observe that

YI (a-n) 1/4
Ix - If M I = ( Ix - 1/1 (a-n)r If (OP)
{a -n)r/PI 1 -
•( Ix - yI If (Y)Ipa (2.8.7)
where b = P i -q 1 . Because A + iq + b = 1, we may apply Holder's inequality
P
to the three factors on the right side of (2.8.7) to obtain
1/q
I fa(f)(x)I < (f I x - yt(°',n'rIf(v)iPdu
1/p` 6
bI(a-o)rdEl
• Ix - If(y)^pdv •
n n

Therefore, by Fubini's theorem and (2.8.6),

fa (raf )Qdx [f
n c
ix - ^I (a
t
-
n)r
If (y)IPdzdy

•C(a, 6,n)p, • IIIIIp 6


< c(ct, 6 , n) • li f
- IT - C(a, 6, n)^, • 11/11: 96 -

88 2. Sobolev Spaces and Their Basic Properties

Thus,

IIIQf1f4 < C(a, 15, H)('Iq)+(`Ip') I1flfp


< C(a, ë, 0 )
1/r
fIf lfp-

2.8.6. Remark. It is an easy matter to see that Theorem 2.8.4 provides


another proof of Theorem 2.4.2. Indeed, if u E Co (R" ), recall from (2.4.5)
that for every x E R",

fu(x)I 5 C(n)Il(IDuI)- (2.8.8)

In fact, if we employ the Riesz composition formula which states that

* Ip = fQ+A ,
a + Li < n,

an application of (2.8.8) to the derivatives of u gives the estimate

Iu(x)I < C(n k)Ik(ID k ui)•


,

From Theorem 2.8.4 we have

fllk(IDkuI)IIp. < CflDkulip


if kp < n. Thus,
Ilullp•<_ CIID k uIIp <_ Cllullk.p
which is the conclusion in (2.4.10) when SZ = R".
Of course, one could also employ Theorem 2.6.1 which states that each
u E W k IP(R") can be represented as u = gk * f for some f E U' (R" ), where
IIf IIp Ilullk,p;R° . Then, in view of the fact that gk < CIk, (2.4.10) follows
from Theorem 2.8.4.

2.9 Limiting Cases of the Sobolev Inequality


In previous sections all Sobolev-type inequalities were established under
the restriction kp n. We now treat the case kp = n in the context of
Riesz potentials and since the Riesz kernel I , is defined for all positive
numbers a, we will therefore replace the integer k by a.
When ap = n, one might hope that t o * f is bounded because p* --' o0
as ap -' n. However, while boundedness is trivially true when n = 1, it
is false when n > 1. As an example, consider u(x) = I log lxII' -2 / ( " - L ) ;
clearly u E W 1 ""(B(O,r)) for r < 1, but u L°'°(B(0,r)). Although an
L estimate cannot, in general, be obtained it is possible to obtain results
that provide a good substitute. Our first result below offers exponential
integrability as a substitute for boundedness. We begin with a simple and
2.9. Limiting Cases of the Sabolev Inequality 89

elegant proof of this fact which follows easily from the estimate discussed
in Remark 2.8.5.

2.9.1. Theorem. Let f E L"(ft), p > 1, and define

9=In/p* f.
Then there are constants C1 and C2 depending only on p and n such that
P
exp dx < Ca . (2.9.1)
st Cl Ilf IIP
Proof. Let p < q < oo and recall from Remark 2.8.5 the estimate
1l)"
Il'af Ilq < C (Q, o, Ilf llp ,
(2.9.2)
where T=1—( 1 1 '

cd(n — 1)Ift17
C(a ' S' (1) [(a — n)r + nja(n) 7 '

and
(a — n)r 1
7= +
n
In the present situation, cp = n, and therefore

(oc n)r+n— pq
TIP —

+(P— g)
Thus, we can write

C(a, ô, S-2) = K l l SZ l 7 pq +(p — q) < )

KIS2I7q
(
np
where K1 and K are constants that depend only on p and n. Thus, since
74/r = I, from (2.9.2) we have

IglQdx < C q/r Ili llp


i2
+ ( q/P' ) •
(qK)q/riot llflip = (qK) 1 IIcI Ilfllp.
Now replacement of q by p'q (which requires that q > p — 1) yields

IgIP i4 dx < (F'gK) 1 +gInl Ilfl1p'g.

In preparation for an expression involving an infinite series, substitute an


integer k, k > p — 1, for q to obtain
,k
1 ^ i9Î kk (K p'
Lk ^llfllP
A dx < p'K
(k 1).
' ISZI
CP
k
2. Sobolev Spaces and Their Basic Properties

90

for any constant C > O. Consequently,


0 0 p' Te oo k , k

f ^ 1 f191 {{ dz < p'KI^ ^ k Kp


k,ko
k^ Cllflkp I k (k - 1 +. C131 )
—ko
)

where k o = [p]. The series on the right converges if CP' > elf); and thus
the result follows from (1.5.12) when applied to the terms involving k < k e,
and the monotone convergence theorem.

By appealing to a different method, we will give another proof of expo-


nential integrability that gives a slightly stronger result than the one just
obtained.

2.9.2. Theorem. Let f E L"(R" ), spt f c B where B is a ball of radius R,


and let p = n jot > 1. Then, for any e > O, there zs a constant C = C(c, n, p)
such that
n Tn/p ( f)( X ) p dxC.
exp e (2.9.3)
Ilf llp
]

^ un-1

Proof. Clearly, we may assume that 11f Ilp = 1. Then,

ndy
Ia(f)(s) - f(v)I x — y1a ndy + f(y)lx — E/la
-

f (2,6)
-

B— B(2.6)

where x E B and 0 < 6 < R. By Lemma 2.8.3(i), the first integral on the
right is dominated by C6kM(f)(z). By Holder's inequality and the fact
that 11f IIp = 1, the second integral on the right can be estimated as follows;
if r = Ix - vi, then

R 1111;
f (âl)Iz — glla-"dy < w(n —1)f
fB—B(x,6) 6
= [w(n - 1) log(R/6)1 10 ' .

Thus
IIQ(f)(x)I C6°M( f )(z) + (w(n - 1) log(R/6)) 1 /n'.
If we choose
P = rnin(EC -1 [M( f )(r)] -1 , R ° ),

then we have
,
II„(f)(x)I < c +[w(n - 1) log + (^ -1/°`C` /° M(f)(x) 1 /°`)] 1/p a
or +D'
(Ia(f)(x) - e) w(n - 1 )n -1 log + (R"e - "C"M(f)(x) p )
2.9. Limiting Cases of the Sobolev Inequality 91

since ap = n. Because II f II p = 1, the conclusion now follows immediately


from Theorem 2.8.2. D

2.9.3. Remark. Inequality (2.9.3) clearly implies that if A < n/w(n - 1),
then there is a constant C = C(Q, n, p) such that

exp dx < C (2.9.4)


[°111fUlipXx)r
thus recovering inequality (2.9.1).
Although it is of independent mathematical interest to determine the
best possible constants in inequalities, in some applications the sharpness
of the constant can play a critical role.
The sharpness of the Sobolev imbedding theorem in the case of critical
indices has had many different approaches. For example, in [HMTI, it was
shown that the space W01 10) could not be embedded in the Orlicz space
L,p (1l) where çp(t) = exp(Itln/(n -1 ) - 1). On the other hand, with this
Sobolev space, it was shown by Moser IMOS1 that (2.9.4) remains valid for
/3 = n/w(n - 1); that is, a can be taken to be zero in (2.9.3). Recently,
Adams [AD8I has shown that (2.9.4) is valid for f3 = n/w(n - 1) with no
restriction on a.
Theorems 2.9.1 and 2.9.2 give one version of a substitute for boundedness
in the case ap = n. We now present a second version which was developed
by Brezis and Wainger [BWI.
For this, recall the definition of the Bessel kernel, g a , introduced in (2.6.1)
by means of its Fourier transform;

9 0 (x) = 2 7r -n/2 (1 + Ix 2)-a/2,

Also, recall that the space of Bessel potentials, L°•p(R"), is defined as all
functions u such that u = ga * f where f E L 13 (R"). The norm in this space
is defined as IIuIIa,p = II fli p . Also, referring to Theorem 2.6.1, we have in
the case a is a positive integer, that this norm is equivalent to the Sobolev
norm of u.
For the development of the next result, we will assume that the reader
is familiar with the fundamental properties of the Fourier transform.

2.9.4. Theorem. Let u E Lt.¢(R") with lq > n, 1 < q < co and let
op = n, 1 < p < 00. If Ilulin.p < 1, then

IIu II^ <_ C [ 1 + log l /p1 (1 + IIuIIt ^) •


^ ^
(2.9.5)

Proof. Because Co (R") is dense in 1,1 . 9 (R") relative to its norm and
also in the topology induced by uniform convergence on compact sets, it is
sufficient to establish (2.9.5) for u E Co (R").
92 2. Sobolev Spaces and Their Basic Properties

Let cp, n E C°°(R") be functions with spt tp compact, cp (27r)-n f2


on some neighborhood of the origin, and Sp + iJ = (27r)_n/ 2 on R". Since
u E Ca (R"), u may be written in terms of the inverse Fourier transform

1
a$
u(s) = e'z
" .11 û(092 (âi1 R)dy + 4u(^)rl(^/R)dl!
=

= ui(x) + u2(x), (2.9.6)


where R> 2 is a positive constant to be determined later.
The proof will be divided into two parts. In Part 1, the following inequal-
ity will be established,

Ilulll^ <_ C(log R)


^

`^ p

while in Part 2, it will be shown that

IIuaII^ < CR - 4 4 II14


for some 5 > O. The conclusion of the theorem will then follow by taking

R max( 2 ,IIUII46 ).

Proof of Part 1. We proceed to estimate u 1 as follows:

e i = .1 (27r)"/ 2 0. (p(y1R) d y
u1(x) _ (21r)--n J2 IfI 2 ) a,2 (Y1)
+ 11,I2)o/2
= f *xR(x)
where
f(y) = (27r)n/ 2 (1 + m 2 ) 0/2 u(y)
and
ç(y/R) (2.9.7)
KR (b) — ( 1 .

+ I1+I 2 ) 012
Note that u = g o * f (see Section 2.6) and therefore

Ilflip — Ilulla,p <_ 1 .

Consequently, in order to establish Part 1 it will be sufficient to show that

II^R4Ip< ç C (iog R) l /p 1 , R > 2.

We now define a function L such that L = cp. Note that L is a rapidly


decreasing function and thus, in particular, L E L 1 (R") fl C°°(R"). Let

LR(x) = R "L(Rx).
2.9. Limiting Cases of the Sobolev Inequality 93

From (2.9.7) we have that


( 2,r) -71/2
KR(y) = LR(y) • 9a(li)
and
(27)-n/2KR = LR * 9a
LetB(R) be the ball of radius R centered at the origin. Define two functions
Gâ and Gâ by
Gp(s) = 9 0 (x)XB(R- 1 ) (x)

Ga (x) = 9a(x) — G (z)' a


Then
( 27 )
-n/2 KR(s) = LR * Gâ(s) + LR * G^(x).
An application of Young's inequality yields

HLR * GâIID' <_ IILRIIpf • IIGQII1 (2.9.8)

and it is easily verified that

IILRIfn, ;
CrIp

while from (2.6.3) and ap = n, it follows that

IIGQ 111 <_ CR - n/P.


Similarly, from (2.6.3) we see that

9a(x) Ç C1lxI°` - ne -Czlsk

and therefore
)lip'

ÎrGkirp' < C ^,IxI^^-


1 /R<Isj<1
nlp' dx

1/P
+C e—C' !TIP' dx
11<lx1<co

< C (log R)'/p' + C


< C(log R)IIp'

since R > 2. Hence,


IILR * GkII p' < C(logR)1/721 (2.9.9)

because
IILRII'1= IILIIi = C < oo.
94 2. Sobolev Spaces and Their Basic Properties

Thus, from (2.9.8), and (2.9.9) we have

IIKRIIP' < C(log R) 1 /P', (2.9.10)

thereby establishing Part 1.

Proof of Part 2. We write u 2 as follows:

=( -n/2 e ^ I I)"2
u2(x) 27 ) iXil û(11)( 27r) n/2 ( 1 + Iy12)t/ chi
(1 + lyl)

g * KR(x) (2.9.11)
where
^?(y/ R) (2.9.12)
KR(y) _(1 + 11/I2) e/2
and
'4 (v ) = ü(v)(270 n / 2 ( 1 + Iu1 2 ) e / 2 .
By assumption, u E L t. g(R n ), and therefore it follows from definition that
g E LQ(R") with u = ye * g. In order to establish Part 2, it suffices to show
that
-6
IIKRI^Q, < CR for some ë > O.
First, consider the case q = 1. Since eq > n by assumption, we have
l > n. Now write

ei" g(y/Rd Irl(y/R)Idb


IKR(x)I =If <
(1 +10 2 ) 1 / 2 1 — IyI
Recall that t vanishes in some neighborhood of the origin, say for all y such
that 'y1 < ER. Thus, for all x,

I KR( x )l
I 77(yIR)1dy
R"—B(Q,eR) M
00
< C r"-^-edr
ER
< CR "

since I > n. Thus, Part 2 is established if q = 1.


Now consider q > 1, so that q' < oo and without loss of generality, let
l G n. Since cp + q = (27r) - "/ 2 on Rn, (2.9.12) can be written as

0(YiR ) I2
4
( 2 1r) ni2 KR(y) — 1
( 1 + ICI

(1+1y17)t 2 )

Thus, we have
KR(x) = ge(x) — ge * LR(x) (2.9.13)

2.9. Limiting Cases of the Sobolev Inequality 95

where
LR(x) = RT L(Rx), L(y) = 9 9 (y)•
We can rewrite (2.9.13) as

KR(x) = f [Mx) - 9i(x - y)1LR(y)dy (2.9.14)

because
(270 —n/2 =
cp(0) = (2 7 )—n/2 e io.yL( y )d

y•
Rn
To estimate f 19t (x -- y) - gt (z) ri dz we write
f n
i9t(x y) - 9t(x)^Q^dx =
— I9t(x - y) - 9c(x)r" dx
1ri521v1

+I9t(x- y) -gt(x)f9^dx
Irl>?I41
=--I1+I2.
-

Now

Îl < C I
,

)
frI < 2IyI
1

x1521v1

<C 19t(x-y)J (Ix +1gt(x)1Qdx


fix - til< 3 1Yr j=15 2 1y/
< CIyjit n)q'+n,
-

by (2.6.3). To estimate I 2 , note that gt is smooth away from the origin,


and therefore we may write Igt(x - y) — 9t(x)1 < ID gt (z)1 - iyl where z =
t(x - y) + (1 - t)x = x - ty for some t E [0, 11. Since, Izt ? 1/(2x1) when
1x1 > 2114 we have, with the help of (2.6.4),

P-clz11x1(t-n-" Iyllrdx
I9t(x - y) - 9t(x)IQ'dx < C

-c lzllxl (t- n -r) Q'dx


< IvV"e
R"
< CIy1 4 .
Consequently, combining the estimates for I I and I2 , we have

II9t(x - y) — 9t(x)I1q' < CIyIS (2.9.15)


where ë = [(t + 1 - n)q' + nJ/q' > O. Referring to (2.9.14), we estimate
[IKRIIq' with the aid of Minkowski's inequality and (2.9.15) as follows:
1 /q' [ /g'
IKR(x)IQ'dx <_ 1 I91(x —
y) — gt(x )1 a'dx ILR(y)Idy
2. Sobolev Spaces and Their Basic Properties

96

Cf ILR(9)t 19I 6 dy

= CR" 6 L(Ry)I 1R91 a d9



< CR -6 .

f
The integral IL(z)I Izl 6 dz is finite because L = cp and thus L is rapidly
decreasing. The proof of Part 2 is now complete and the combination of
Parts 1 and 2 completes the proof of the theorem. ❑

2.10 Lorentz Spaces, A Slight Improvement


In this section we turn to the subject of Lorentz spaces which was intro-
duced in Chapter 1, Section 8. We will show that the Sobolev inequality
for Riesz potentials (Theorem 2.8.4) as well as the development in Chapter
2, Section 9, can be improved by considering Lorentz spaces instead of LP
spaces.
We begin by proving a result that is similar to Young's inequality for
convolut ions.

2.10.1. Theorem. If h = f * g, where

f € L(P1, 4i), g E L(P2, 42), and 1 1


+ >1,
Pi P2
then h E L(r, a) where
1 1 1
—+--1= -
P1 P2 r
and s > 1 is any number such that

— 1 > 1,
1 + ____
41 42 s
M oreover,
II h II( r , d) < 3TII f IIc7)„g,)11 9 11(p2,q2 ).
Proof. Let us suppose that q 1 , q 2 , s are all different from oc. Then, by
Lemma 1.8.9,

F
00 OD
°° Z ^^r f
(1 I h ^I(r, ․ )) ° = (-
x
* :(t)g" (t)dt dr
i
o 0 s l

—o
_
^
f” 1 g " ^ (2.10.1)
o y
[-d--1-r- u u
u d^d 1 .
f1.
2.10. Lorentz Spaces, A Slight Improvement 97

The last equality is by the change of variables x = 1/y, t 1/u. Now use
Hardy's inequality (Lemma 1.8.11) to obtain,

°° f 1 y ( 1) .. 1 du dy '

0 Y t /r 0 g
f" (;) ‘7 y

[ I
oo
< r' dY.
2
o Y
y '_('/r) f(h/y)g(h/y)
y ' '

00
' r (xt+(1/r) {..(^1^..^x` ^sd2
o J 1 1 x'
by letting y = 1/x.
Since s/q l + 5/q2 > 1, we may find positive numbers m t , m 2 such that
1 s
11 and 1 < s 1
m1 m2 m1 qt m2 q2
Therefore q i Ç sm li q z G sm 2 . An application of Hdlder's inequality with
indices rn li na g , yields
< rsop [x!/Pi f•*(x)Ta jT1/pzg*s(x)]a
xl/m' J ll xi/m2
(IIhII(r,a))a dx
U
o o 1
dx / m i
< re /Pi
[x f■slx)Jsm^
0 J z
(x/pz 4 ..( x )^s.n 2 _
d.
[Zooi
II
= r" ( IIf Il(Pi ,am►))'(II9II(P,,.mz) )s.

Thus, by Lemma 1.8.13

IIhII(r,a) < rllf Ile pr ,amAgIIp,,amz)


< e 1 /e e l /e rllf II(p l , g , ) Ilsll(p,, 0 )
<_ 3 r. IIf 11(p, .g1)IIgIl(p2,42)•
Similar reasoning leads to the desired result in case one or more of (it,
42, 8 are oo. ❑

As an application of Theorem 2.10.1, consider the kernel 4,(x) = 1xia--n


which is a constant multiple of the Riesz kernel that was introduced in
Chapter 2, Section 6. For simplicity of notation in this discussion, we omit
the constant 7(a) -1 that appears in the definition of /„,(x). Observe that
the distribution function of l a is given by
a-n > t}I
ara (t) = I{x : Ixl
= (Ix : IxI < t1/(a-n)}1
a(n)tre /(Q-n)
=
2. Sobolev Spaces and Their Basic Properties

98

and because I is the inverse of the distribution function, we have 1V() =


( t ){
a—n)/n. It follows immediately from definition that
WC)
n)

I:* (t) = n t(a—n)/n


)

and therefore that I. E L(n/(n - at), oo). If we form the convolution la * f


where f E LP = L(p,p), then Theorem 2.10.1 states that

la * f E L(q,p)

where
1
-_- 1
+ n- a 1 oc
- 1 =---.
q P n p n
Moreover, it follows from Lemma 1.8.13 that L(q,p) C L(q, q) and thus we
have an improvement of Theorem 2.8.4 which allows us to conclude only
that fa * f E L 4 . As a consequence of Theorem 2.10.1, we have the following
result that is analogous to Theorem 2.8.4.

2.10.2. Theorem. If f E L(p,q) and 0 < a < n/p, then

Ia * f E L(r,q)

and

Ilia * f ll(r, g ) Illall(n/(n—a),00) IIf II(p,g)


=GII fII (p,g)
where
1 1 a
r - p n.

We now consider the limiting case of 1 /p i +1/p2 = 1 in Theorem 2.10.1.


In preparation for this, we first need the following lemma.

2.10.3. Lemma. Lei cp be a measurable function defined on (0, 1) such


that tcp(t) E Lp(0, 1; dt/t), p > 1. Then,
1
-1
II( 1 + I log ti) 40(8)616{0,1,d1/0<_ —1)—1 IIt9(t)IILp(o,1;dt/t)•
: p

Proof. By standard limit procedures, we may assume without loss of gen-


erality that w E L 1 (0,1) is non-negative and bounded. Let
p
1 ^ 1 (1 +I logtl)!p cp(s)ds
o (J 1 i
2.10. Lorentz Spaces, A Slight Improvement 99

Since

f = 1
i
(j ' co(s)ds ) d(1 -- log t)—p+'

integration by parts yields


1
=
P j t
cp(s)ds (1 — log t}

and Holder's inequality implies


1
I < ppl II( 1 +I logtl) -' t ^P( 8 )d81I L n ^o ,i;dt t)1I t ^(t)11^^(o,i;dt/t) ,

from which the conclusion follows. ❑

2.10.4. Remark. Before proving the next theorem, let us recall the fol-
lowing elementary proof concerning convolutions. If f,g E L 1 (R") we may
conclude that

I f (r — g)g(y)I dxd y = Ig(g)1 - If(x , y )ldidy


R" j" R" R"

=R" 1001 lIfllidy


= IIIlli llglli < oo•
Thus, the mapping y — f (x — y)g(y) E L 1 (R") for almost all x E R" and
f * g E L 1 (Rn).
In the event that one of the functions, say f, is assumed only to be an
element of L(p, q), p > 1, q > 1, while g E L 1 (R"), then the convolution
need not belong to L 1 (R"), but it will at least be defined. To see this, let

1 if f (z) > 1
f i (x) = If (x) if — 1 < f (x) < 1
—1 if f(x) < —1

and let 12 = f — f 1 . Then f 1 * g is defined because f is bounded. We '

will now show that 12 E L' (R") thus implying that 12 * g is defined and
therefore, similarly for f * g. In order to see that 12 = L' (R") let

0f( x) if I f (x)I > 1


f2(x) = i f j f (x) f < 1.
Clearly ah (a) = ar f (1) if 0 < s < 1 and ci f2 (s) = a f (8) if a > 1. Conse-
quently
f2 (af(1)) = inf{.g : a-2 (s) < a1(1)} _ O.
Thus, since A. is non-increasing, f2 (t) vanishes for all t > a f (1) and it
is easy to see that r (t) = JO) for all t < a f(1). We may assume

100 2. Sobolev Spaces and Their Basic Properties

that q > p for, if q < p, then Lemmas 1.8.13 and 1.8.10 imply that
f E L(p, q) C 14,p). Consequently, by Young's inequality for convolu-
tions, j * g is defined. In fact, f * g E L". With q > p and a = a f(1), we
have
°o f#.(t)q °° f.(t)g
00 >
l
fa f`(t)Q
fo -ti (g/p) (Ît
> 0tl- (Q /p)dt > t1-(g/p) t
a
> a(g(P)-1 f*(t}Qdt
0
a
= a (9 /p )-2 12 (t) Q dt
0
00
= a ( Q /p )-1 (t)gdt
0

where i?, E [0, a]. Since j2 (t) vanishes for t > a whereas f2 (t) > 1 for t < a,
it therefore follows that

f j2(t)dt < 0 /(t)dt

thus showing that /2 is integrable because /2 and f2 have the same distri-
bution function. Therefore 1 2 is integrable.

2.10.5. Lemma. Let 1 < p < oo, 1 < qi < oo, 1 < qz < oo be such
that 1/q i + 1/q2 < 1 and set 1/r= 1/q 1 + 1/q2. Assume j E L(p, q 1 ) and
g E (p',q2) fl L 1 (1V) and let u = f * g. Then
t/T
.. t) T dt
19111)
I_ i + I log II j — cIIfIlcp.gs) • (11911(pi.o)
(

t `

where C depends only on p, q 1 , and q2.

Proof. Note from the preceeding remark, that u is defined. For simplicity
we set II f II(p,q,) and III! = Ilgli(pi )+IIgll,. Also, for notational convenience
.

in this discussion, we will insert a factor of (q/p) 1 / 4 in the definition of the


III II (p,q) ; thus,
1/q o0
(t) 1 <p<oo,0<q <oo
P o

1 !q
--)
9 sup tl/p f #* (t), 1 <p<oo,q=oo.
P t >o
We distinguish two cases:
(i) r<co
2.10. Lorentz Spaces, A Slight Improvement 101

(ii) r = oo (i.e. q 1 = q2 = oo).

(i) The case r Coo. Recall from Lemma 1.8.8 that for every t > 0,
00
u` • (t) < t f"(t)g"(t) f
+f'(s)g'(s)ds. (2.10.2)

Clearly, the following inequalities hold for every a > 0'

r(s) < f"(s) <_


s
/P Cif II(p, g i)> (2.10.3)

g'(s) < s' "(s) < szl /p, Ifsl1(P',g2)? (2.10.4)

9• (8) s g •` ( s ) <_ 9119f1 1. (2.10.5)

For t < 1, it follows from (2.10.2), (2.10.3), and (2.10.4) that,


1
1 1
u'`(t) <t t fi(s)g•(s)ds
%p t1/p^ CCf1i( p,as )IIgI1( p)g2 ) + e
00
+f"(s)g`(s)ds.
f (2.10.6)

From (2.10.3) and (2.10.5) we have that

1
00

f `(s)g•(s)d8 < P11Ifil(pm) - II9II1)-


This in conjunction with (2.10.6) yields
1

u'•(t) < P1If 11 II9Il + f * (s)9 (s)ds.


M (2.10.7)
r
By Lemma 2.10.3, (2.10.7), and (2.10.4) we have

11( 1 + 1 log t1) - 1 u" • (t)IIL' (o,l;dt/t) <_ 01/11 11911 + cliff•(t)s'(t)IIL-(0.1;dt/t)
cliff! 11911 + CiIt 1 /P f+(t)t 1 /P'
9 6 ( t)IILr(0,r dt/t) ;

C11111 11911 + C111 II(P,q,)119II(p',92).


(ii) The case r o). By (2.10.7), (2.10.3), and (2.10.4) we have, for t < 1,

u" ( 1 ) <_ p11111 11911 + 1 log tl IIII1(P,00)Il9ll(p',00)


and therefore
11(1 + I log tI) ' u•' (t)IiL—(0,1)
-
<_ Cilf i1 il9il
2.10.6. Theorem. Let 1 < p < oo, 1 < Q 1 < oo, 1 < q 2 < oo be such
that 1/q1 + 1/q2 < 1 and set 1 f r 1/q1 + 1/g2. Assume f E L(p, qi) and
g E L(p', Q 1 ) fl L 1 and [et u f * g. Then
102 2. Sobolev Spaces and Their Basic Properties

(i) if r < oo, e a i"i E L L(R n ) for every A > 0.

(ii) if r < oo, there exists positive numbers C = C(p, q 1 , 42) and M =
M(I(2I) such that

J e c lui rg dx < M

for every f and g with Ilf II (p,gi) I and 11911(^ .q, ) + II9II1 <_ 1 .

Before proceeding with the proof, let us see how this result extends the
analogous one established in Theorem 2.9.1. To make the comparison, take
one of the functions in the above statement of the theorem, say f, as
the Riesz kernel, I n /p . As we have seen from the discussion preceeding
Theorem 2.10.2, f E L(p', oo). The other function g is assumed to be
an element of LP(11) where S 2 is a bounded set. Thus, by Lemma 1.8.10,
-

g E L(p, p) fl L 1 (Rn ). In this context, q 1 = oo and q3 = p > 1 thus proving


that this result extends Theorem 2.9.1.

Proof. Consider part (1) first. Because u' (t) is non-increasing we have that

Iu'(r)I'' t (1 - log s } r - d9 < u (s) T (1 — l og s) -r


da
0 s —
o s
for every t < 1. The first integral equals (1 - log t) - r }1 /r - 1 with r > 1
and Lemma 2.10.5 implies that 1(t) -, 0 as t -' 0 where 1(t) denotes the
second integral. Note that there exists a constant K = K(r) such that

/(r-1)
l u'(t) I*' < K(1 + I logtl)I(t)I

0 < t < 1. (2.10.8)

With Q C Rn any bounded measurable set, we have with the help of


Lemma 1.5.1,

i
l41 eXp(Clu to
exp(CIu(x)I)r ^ = , (t)I^T ' dt = exptClu'(t)I) r dt
f 0

IQI
+ (2.10.9)
J o
exp(CIu'(t)I)T dt /

where 0 < to < IQI. Because u* is non-increasing, it is only necessary to


show that the first integral is finite. For this purpose, choose t o < 1 so that
Cr' KI(t o ) 1 /(r -1 ) < 1. Then, from (2.10.8),

exp(Clu`(t)I) r < (e/t) Q


where a = C'' i KI{t o } 1 /(r -1 ) Thus, part (i) of the theorem is established.
Exercises 103

For the proof of (ii), Lemma 2.10.5 and the fact that u 0 (t) is non-
increasing allow us to conclude that for 0 < t <1,
i
le(t)I r j ( 1 _log ) Gill IV • 11911 *
-

where Il f II = Ill 11(p,Q,) and 11911 =1191642) + bill. Therefore,

Iu'Mir' < K(1 — log t)IIfPIr'11911r#-


Similar to part (i), the proof of (ii) now follows from (2.10.9) by choosing
KC<1. ❑
Exercises

2.1. Prove that W' (R") = W:'p(R").

2.2. If f and g are integrable functions defined on R" such that

fcPdx = f gcodx

for every function cp € Co (R"), prove that f = g almost everywhere


on R".

2.3. Prove the following extension of the Rellich-Kondrachov compactness


theorem. If St is a domain having the extension property, then
Wk+rn,p(1) Wk,q(0)

is a compact imbedding if mp < n, 1 <q<np/(n—tnp) and ma


non-negative integer.

2.4. Verify the following equivalent formulation of Bessel capacity:


} - P

Ba.p(E) _ inf 9a * f(z)


z EE
... p

{ suPf Q * f(2)
EE

where f E IP(R'a), f > 0, and IIf 11p ? 1


2.5. Prove that the Riesz and Bessels capacities have the same null sets;
that is, RQ , p (E) = 0 if and only if B Q , p (E) = 0 for every set E C R„_
104 2. Sobolev Spaces and Their Basic Properties

2.6. Show that there is a constant C = C(a, p, n) such that

Ba,p(E) > C

provided cep > n and E is non-empty.

2.7. As an extension of Corollary 2.1.9, prove that if 1 is connected and


u E Wk P(11) has the property that Dau = 0 almost everywhere on
,

n, for all lal = k, then u is a polynomial of degree at most k - 1,


2.8. Let 1 < p, kp < n. If K C R" is a compact set, let

yk, p (K) = ; u > 1 on a neighborhood of K,

u E Cfl (R"))
With the aid of Theorem 2.6.1, prove that there exists a constant
C = C(p, n) such that

C -' Bk, (K) < l'k,p( K ) < CBk,p(K) .

2.9. Show that for each compact set K C R",

inf IDurdx . u > 1 on K, u E Co (R") = 4.


Rn

2.10. Prove that there exists a sequence of piecewise linear maps

Lk: R" -. RI

satisfying (2.7.7) and (2.7.8). See the discussion in Exercise 5.2.

2.11. Suppose that u: R" -0 R' is Liptichitz with pull ov► < co. Define
a u : R -' R' by a u (t) = f {x : u(x) > t}1. Since a u is non-increasing,
'

it is differentiable almost everywhere.

(a) Prove that for almost all t,

-otu(t) dH "-' .
IDur l
>u (t)

(b) Prove that equality holds in (a) if

[ix . Du(x)=0}I=0.

2.12. Theorem 2.8.4 gives the potential theoretic version of Theorem 2.4.2,
but observe that the latter is true for p = 1 whereas the former is
false in this case. To see this, choose f ; > 0 with fir, f,dx - 1 and
Exercises 105

spt fi C B(0, 1/i), Prove that Ia * f; l


a uniformly on I?" — B(0, r)
for every r > 0. Thus conclude that

(lQ * fi)" /(n- ^ ) dx > (fa* fi )"/ ( n -Q)


fn f(0,r)

The right side tends to

( fa) n /(n -a) dx = C


t (O,r) B(o,r)
I:C} ndx.
-

But
I x rdx — oo as r 0.
B(0,r)

2.13. Show that. Theorem 2.8.4 is false when ap = n. For this consider
o( 1 +E)/n
^zI ^1og 51) !x^ < 1
f (x) =
o 'xi >

where e > O. Then f E L" since ap = n but / 0 * f (0) = oo whenever


a(1 + e)/n < 1.

2.14. Prove the following extension of Theorem 2.8.2. Suppose Ill log(2 +
if i) is integrable over the unit ball B. Then M f E L i (B). To prove
this, note that (with the notation of Theorem 2.8.2)

1
0o
M < 1B} +Mdx < ^B^ + fAtldt +
A ^ 1

Now use (2.8.3) and Exercise 1.3.

2.15. There is a variety of methods available to treat Theorem 2,6.16. Here


is one that shows that B 1 , p (K) = 0 if Nn p < oo, 1 < p < n.
-

STEP 1. Use Exercise 2.8 to replace B 1 , p (K) by 7 1 , p (K).


STEP 2. There exists C = C(n, k) such that for any open set U D K,
there exist an open set V D K and u E W "(R") such that

(i) u > 0
(ii) spt u C U
(iii) KcV C {x:u(x)=1)
(iv) fir, ÎDui dx < C.
106 2. Sobolev Spaces and Their Basic Properties

To prove Step 2, first observe that HS P(K) < oo (see Exercise -

1.10). Since K is compact, there exists a finite sequence of open halls


{B(r i )}, i = 1, 2, ... , m, such that
m rn

KC B(r i ) C B(2r,) C U
i = l i =1
and 00

L a(n - p)ri -p < HS P(K) + 1. -

i= i
Let V = Um 1 B(r i ) and define u; to be that piecewise linear function
such that ui=1o ❑ B(r,),u i =Oon R"- B(2r;). Let

u=max{u,.a=1,2,..., m}

to establish Step 2.
STEP 3. For each positive integer k, let

Uk— x: d(x,K)< 1k .

Employ Step 2 to find corresponding u k such that

IDukI dx < C

for k =1,2, ....


STEP 4. Use Theorem 2.5.1 to find a subsequence {uk} and u E
Wo'P(R") such that uk -' u weakly in Wo'p(Rn) and u k -i u strongly
in LP. Hence, conclude that u = 1 almost everywhere on K and that
u=OonR" -K.
STEP 5. Conclude from Theorem 2.1.4 that IKI = 0 and therefore
that u=0.
STEP 6. Use the Banach-Saks theorem to find a subsequence {uk)
such that
1
tlj = E
'

uk
k =1
converges strongly to u in WW' p (R"). Thus, IIDv;IIp - 0 as j --• oo.
But
71,P(K) Ç IDvJI p dZ
R "

for each j = 1, 2, ....


Exercises 107

2.16. In this problem we sketch a proof of the fact that y i , p (K) = 0 implies
11"-p+e(K) _ 0 whenever c > 0. The proof requires some elementary
results found in subsequent chapters.
STEP 1. For each positive integer i, there exists u, E Co (Rn) such
that u ; > t on a neighborhood of K and

ID dx < I .
R" 2'

Let v = Ei ° 1 u, and conclude that v E W L 7 (R" ). Also Mote that


K C interior {x : v(x) > k) whenever k > 1. Therefore, for r E K,
lim inf v(x, r) = o0
r-.0

where
v(x, r) = v(y)dy.
B(s,r)

STEP 2. For all x E K and e>0,

lim sup rp - T ! IDviPdy = oo.


r-.0 .l B(x,r)

If this were not true, there would exist k < co such that

r p -n -E IDv1Pdy < k

for all small r > 0. For all such r, it follows from a classical version
of the Poincaré inequality (Theorem 4.2.2) that

113(.,r) Iv(U) — v(t,r)Ipdy < Crp - n 1DvIpdy < Cr`.


B(s,r)

Thus conclude that

17(x, r/2) — v(x, r)j < CrEIP

for all small r > 0. Therefore, the sequence {1(z, 1/2 3 )} has a finite
limit, contradicting the conclusion of Step 1.
STEP 3. Use Lemma 3.2.1 to reach the desired conclusion.

2.17. At the end of Section 2.3 we refer to [AR2) for the result that C°° (SZ)
is dense in W k, P(SZ) provided 0 possesses the segment property. Prove
this result directly if the boundary of 0 can be locally represented as
the graph of a Lipschitz function.

2.18. Show that C°. 1 (Q) = W''°°(l) whenever 11 is a domain in R.


108 2. Sobolev Spaces and Their Basic Properties

2.19. This problem addresses the issue raised in (2.6.6). If f E LP (R") and
ap < n, then Theorem 2.8.4 states that

Ilia *file <_ Cilfllp


where q = p'. A simple homogeneity argument shows that in order
for this inequality to hold for all f E D' , it is necessary for q = p`.
For 6 > 0, let Tb f (x) = f (6x). Then

ll^a * (T6f)II4 ^^"^pllf^lP


and
* ( Tbf )— ra Td (ta* f )'

Hence,
-a - n /g
llia * (Tbf)IIg - a b }4ia * f Ilgi
thus requiring
1 1 a
Q P n

Historical Notes
2.1. It is customary to refer to the spaces of weakly differentiable func-
tions as Sobolev spaces, although various notions of weak differentiability
were used before Sobolev's work, [SO2]; see also [S01], [S03]. Beppo Levi
in 1906 and Tonelli [TO] both used the class of functions that are abso-
lutely continuous on almost all lines parallel to the coordinate axes, the
property that essentially characterizes Sobolev functions (Theorem 2.1.4).
Along with Sobolev, Catkin [CA] and Morrey [MOI] developed many of
the properties of Sobolev functions that are used today. Although many
authors contributed to the theory of Sobolev spaces, special note should be
made of the efforts of Aronszajn and Smith, [ARS1], [ARS2J, who made a
detailed study of the pointwise behavior of Sobolev functions through their
investigations of Bessel potentials.
2.2. Theorem 2.2.1 was originally proved by Rademacher [RA]. The proof
that is given is attributed to C.B. Morrey [MO1, Theorem 3.1.6]; the proof
we give appears in [SI. In our development, Rademacher's theorem was
used to show that Sobolev functions remain invariant under composition
with bi-Lipschitzian transformations. However, it is possible to obtain a
stronger result by using different techniques as shown in [Z3]. Suppose
T : R" -i R" is a bi-measurable homeomorphism with the property that
it and its inverse are in Wi o'' (R", R" ), p > n - 1. If u E W1 f (R") where
p` = p[p - (n - 1)] -1 , then u o T E W101: (R"). With this it is possible to
show that if u E Wi ô: (R") and T is a K-quasiconformal mapping, then
uoTE W1 (R").
Historical Notes 109

2.3. Theorem 2.3.2 is due to Meyers and Serrin, [MSE].


2.4. Theorem 2.4.1 is the classical Sobolev inequality [SO1], [SO2], which
was also developed by Gagliardo [GA1], Morrey IMO1], and Nirenberg
[NI2]. The proof of Theorem 2.4.1 for the case p < n is due to Nirenberg
[NI2).
2.5. Theorem 2.5.1 originated in a paper by Rellich [RE] in the case p =
2 and by Kondrachov [KN] in the general case. Generally, compactness
theorems are of importance in analysis, but this one is of fundamental
importance, especially in the calculus of variations and partial differential
equations. There are variations of the Rellich-Kondrachov result that yield
a slightly stronger conclusion. For example, we have the following result
due to Frehse [FRE]: Let 1 C R" be a bounded domain and suppose
u E W' P(c1), 1 < p < n, is a bounded sequence of functions with the
;

property that for each i = 1, 2, ...,

IDu i IF -2 Du; dx < be Wilco

for all cp E W 1 '( 1l) n L°°(1). Then there exist u E W "P(Il) and a subse-
,

quence such that u -, u strongly in W"'(0), whenever q < p.


;

2.6. Potential theory is an area of mathematics whose origins can be traced


to the 18th century when Lagrange in 1773 noted that gravitational forces
derive from a function. This function was labeled a potential function by
Green in 1828 and simply a potential by Gauss in 1840. In 1782 Laplace
showed that in a mass free region, this function satisfies what is now known
as Laplace's equation. The fundamental principles of this theory were de-
veloped during the 19th century through the efforts of Gauss, Dirichlet,
Riemann, Schwarz, Poincaré, Kellogg, and many others, and they consti-
tute today classical potential theory. Much of the theory is directed to the
understanding of boundary value problems for the Laplace operator and its
linear counterparts. With the work of H. Cartan [CAR1], [CAR2], in the
early 1940s, began an important new phase in the development of potential
theory with an approach based on a Hilbert-space structure of sets of mea-
sures of finite energy. Later, J. Deny [DE] enriched the theory further with
the concepts and techniques of distributions. At about the same time, po-
tential theory and a general theory of capacities were being developed from
the point of view of an abstract structure based on a set of fundamental
axioms. Among those who made many contributions in this direction were
Brelot [BRT], Choquet [CH], Deny, Hervé, Ninorniya, and Ohtsuka. The
abstract theory of capacities is compatible with the recent development of
capacities associated with non-linear potential theory which, among other
applications, is used to study questions related to non-linear partial differ-
ential equations. The first comprehensive treatment of non-linear potential
theory and its associated Bessel capacity was developed by Meyers [ME1],
110 2. Sobolev Spaces and Their Basic Properties

Havin and Maz'ya NM], and Resetnjak [RES). Most of the material in
Section 2.6 has been adopted from IME1).
2.7. The co-area formula as stated in Theorem 2.7.3 was proved by Fed-
erer in [F11. In the case m 2, k = 1, Kronrod [KR] used the right side of
(2.7.1) to define the variation of a function of two variables. Fleming and
Rishel [ FR] established a version of Theorem 2.7.1 for 13V functions. An-
other version resembling the statement in Theorem 2.7.3 for BV functions
appears in [F4, Section 4.5.9].
The proof of the best possible constant in the Sobolev inequality (The-
orem 2.7.4) is due to Fleming and Federer [FF]. Their result can be stated
as follows:
na(n) = sup flDull
Ilulln/(n-1)
where the supremum is taken over all u E Co (Rn). Talenti [TAI extended
this result to the case p > 1 by determining the constant C(n, p) defined
by
IlDullp
C(n,p) = sup
IIuIIp.
He showed that

1,( 1 + pr(n)
1/n
1—(1/P)
r -1/2 n -1/9 P - 1
C(n,p) =
n-p { r()r(1+n- p)

He also showed that if the supremum is taken over all functions which
decay rapidly at infinity, the function u that attains the supremum in the
definition of C(n, p) is of the form

u(x) = (a+blzyPf ( P - 1))1—nfP


where a and b are positive constants. This leads to the following obser-
vation: in view of the form of the extremal function, it follows that if S2
is a bounded domain and if u E W "(1) has compact support, then by
extending u to be zero outside of St, we have

IIuIIp• < C(n , P)IIDuIIP.


Brezis and Lieb [BL] provide a lower bound for the difference of the two
sides of this inequality for p = 2. They show that there is a constant C(S1, n)
such that
C(O, n)IIull,00) +1114 -< C(n,p)IIDullâ
where q = n/(n - 2) and IIuII(q,00) denotes the weak Lq-norm of u (see
Definition 1.8.6).
2.8. The maximal theorem 2.8.2 was initially proved by Hardy-Littlewood
[HLI for n = 1 and for arbitrary n by Wiener [WI]. The proofs of Theorem
2.8.4 and its preceding lemma are due to Hedberg [HE1].
Historical Notes 111

2.9. The proof of exponential integrability in Theorem 2.9.1 is taken from


IGTI while the improved version that appears in Theorem 2.9.2 was proved
by Hedberg [11E1]. The question concerning sharpness of this inequality has
an interesting history. Trudinger (TR1 proved (2.9.1) for Sobolev functions
in W k .P, nip = k, with the power p' replaced by n'. However, when nip > 1,
Strichartz [STR1 noted that Trudinger's result could be improved with the
appearance of the larger power p'. The reason why Trudinger's proof did
not obtain the optimal power is that the case of k > 1 was reduced to the
case of k = 1 by using the result that if u E W k " ) , k > 2, kp = n, then
u E W". However, in this reduction argument, some information is lost
because if u E W k •e then u is actually in a better space than W l in. In fact,
by appealing to Theorem 2.10.3, we find that the first derivatives are in the
Lorentz space L(n, p) C L". This motivated Brezis and Wainger to pursue
the matter further in [BM where Theorem 2.9.4 and other interesting
results are proved. The sharpness of the Sobolev imbedding theorem in the
case of critical indices was also considered in [HMT1, where it was shown
that the space W(l) could not be imbedded in the Orlicz space L,(1)
where çp(t) = exp(Iiln/(R -1 ) — 1).
The other question of sharpness of the inequality pertains to the constant
/3 that appears in (2.9.4). It was shown in [MOST that (2.9.4) remains valid
for fi = n/w(n — 1) in the case of Sobolev functions that vanish on the
boundary of a domain. The optimal result has recently been proved by
Adams [ADS] where (2.9.4) has been established for Q = n/w(n — 1) and
all a > O.
2.10. Most of the material in this section was developed by Brezis and
Wainger [BWI although Theorems 2.10.1 and 2.10.2 and due to O'Neil [O1.
3

Pointwise Behavior of
Sobolev Fonctions
In this chapter the pointwise behavior of Sobolev functions is investigated.
Since the definition of a function u E Wk.P(S2) requires that the kth-order
distributional derivatives of u belong to (0), it is therefore natural to
inquire whether the function u possesses some type of regularity (smooth-
ness) in the classical sense. The main purpose of this chapter is to show
that this question can be answered in the affirmative if interpreted ap-
propriately. Although it is evident that Sobolev functions do not possess
smoothness properties in the usual classical sense, it will be shown that if
u E W k 'P(Rn), then u has derivatives of order k when computed in the
metric induced by the LP-norm. That is, it will be shown for all points
x in the complement of some exceptional set, there is a polynomial Px of
degree k such that the L"-norm of the integral average of the remainder
lu -- Ps I over a ball B(x, r) is o(r k ). Of course, if u were of class C k , then
the LP-norm could be replaced by the sup norm.
We will also investigate to what extent the converse of this statement is
true. To this end, it will be shown that if u has derivatives of order k in the
17-sense at all points in an open set SI, and if the derivatives are in IP(Sl),
then u E Wk1P(0). This is analogous to the classical fact that if a function
u defined on a bounded interval is differentiable at each point and if u' is
integrable, then u is absolutely continuous. In order to further pursue the
question of regularity, it will be established that u can be approximated in a
strong sense by functions of class CI, I < k. The approximants will have the
property that they are close to u in the Sobolev norm and that they agree
pointwise with u on large sets. That is, the sets on which they do not agree
will have small capacity, thus establishing a Lusin-type approximation for
Sobolev functions.

3.1 Limits of Integral Averages of


Sobolev Functions
In this and the next two sections, it will be shown that a Sobolev function
u E W k iP(Sï) can be defined everywhere, except for a set of capacity zero,
in terms of its integral averages. This result is analogous to the one that
3.1. Limits of Integral Averages of Soholev Functions 113

holds for integrable functions, namely, if u E L 1 , then

(u(y) - u(x)(dy -, 0 as r -, 0
B(a,r)

for almost all x E Rn. Since our result deals with Sobolev functions, the
proof obviously will require knowledge of the behavior of the partial deriva-
tives of u. The development we present here is neither the most efficient nor
elegant. These qualities have been sacrificed in order give a presentation
that is essentially self-contained and clearly demonstrates the critical role
played by the gradient of u in order to establish the main result, Theorem
3.3.3. Later, in Section 3.10, we will return to the subject of Lebesgue points
and prove a result (Theorem 3.10.2) that extends Theorem 3.3.3. Its proof
will employ the representation of Sobolev functions as Bessel potentials
(Theorem 2.6.1) and the Hardy-Littlewood maximal theorem (Theorem
2.8.2).
In this first section, it will be shown that the limit of integral averages
of Sobolev functions exist at all points except possibly for a set of capacity
zero. We begin by proving a lemma that relates the integral average of u
over two concentric balls in terms of the integral of the gradient.

3.1.1. Lemma. Let


u E WLp[B(xo,r)1, p > 1, where xo E R" and r > 0.
Let 0 < 6 < r. Then
1 —n
r—"
B(so.r)
u(i1)dy—b—"
L06) U(11)6111= nr f
B (z o ,r)
[Du(11)' (11 - xo)1d11

16
- - -n
n f B( 0 ,6)
[Du() • (y - zo)1dy
'
1
n J (so,r) — B(so,a)
11/ — zo1"[Du(Y) • (y - xo)1d1l. (3.1.1)

Proof. Define p on R 1 by

µ(t) = t—n — p - n b < t< r


0 oTh_ptt > r.
{ 6
Define a vector field V by V(y) = µ(1y - xoI)(FI - z0). Since u is the
strong limit of smooth functions defined on B(zo, r) (Theorem 2.3.2), an
application of the Gauss-Green theorem implies

(3.1.2)
J (sp,r) u(v)div v (v)dv = -Du(g)
B (xD,r)
• (y - ro)p(Iy = xoI)dy.

An easy calculation of div V establishes equation (3.1.1). ❑


114 3. Pointwise Behavior of Sobolev Functions

3.1.2. Lemma. Let t be a positive real number such that Ip < n, p > 1,
and let u E W l 'p(Rn). Then

Iv -- IDu(y)
x1 Idy (3.1.3)
JR .n Iv — X1 1-7 1 11 (014 5
Rn
1-71+1

for all x E Rn.

Proof. (i) We suppose first of all that u vanishes outside a hounded set.
Let z E IV and for each positive integer j, define a C °° vector field Vl on
Rn by
1 (1/2)(t -n)
V, (y) = - + I Y _xI2 (y
[
x)• ]

Since Iul E W i rP(R°) (by Corollary 2.1.8), lul is therefore the strong limit
of smooth functions with compact support. Therefore, by the Gauss-Green
theorem,

j Rn
JR
div V;(y)lu(y)Idy = rV(y)' D(IuI)(3►)^dy.
n

Moreover, since ID(ful)I = IDul a.e.,

divV; (y)lu(y)Idy < V,(y)IDu(y)Idy. (3.1.4)


Rn Rn

By calculating the divergence on the left-hand side of (3.1.4) one obtains

L
1 (1/2)(t-n--2) n
ly — z1 2 fly — 1 1 2 + -jlu(y)Idy
[

n J+
(1/2)(t—n)
< 1 + ly — x 2 I 1y — lI IDu(y)Idy•
RR 3
The inequality (3.1.3) now follows, in this case, when j oc.
(ii) The general case. Let ri be a C°° function on R, such that 0 < n < 1,
rt(t) = 1 when t < 1 and 17(t) = 0 when t > 2. Define
-1
ui(Y) = u(y)i,(.l' IyI)
for y E R''. By applying (i) to u 3 and then letting j --+ oo, one can verify
(3.1.3) in the general case. D

3.1.3. Lemma. Let I be a positive real number and k a positive integer


such that (k + t —1)p < n. Then there exists a constant C = C(n, k, t) such
that
t- n ID a
I y ^ xl n I u(y)I dy <
_ G ^ f ^ lu xI +k U(y)Idy
Rn ^a^ =k
3.1. Limits of Integral Averages of Sobolev Functions 115

for all x C R' and all u E W k •p(R n ).

This follows from Lemma 3.1.2 by mathematical induction. ID

We are now in a position to prove the main theorem of this section


concerning the existence of integral averages of Sobolev functions.

3.1.4. Theorem. Let k be a positive integer such that kp < n, p > 1, let 0
be a non-empty open subset of Rn and let u E W 'p(SZ). Then there exists k
a subset E of SZ, such that

Bk ,P(E ) = 0
and
lira u(y)dy (3.1.5)
6-.0+ B(r,6)
exists for all x E ( - E.

Proof. (1) We suppose first of all that SI = R' . Define

g(y) _ E I rru(y)1

(3.1.6)
101=k
for y E R". Then g E LP(R R ). Let E be the set of all those points x of R n
for which
(Ik * g)(x) = co. (3.1.7)
Then, from the definition of Riesz capacity (Definition 2.6.2),

Rk.p(E) = 0,
and therefore from (2.6.7),

Bk,p(E) = O.

Consider x E R'ti — E. By (3.1.1),

u(y)dy = n
f ( s, i ) [Du(y) • (y — x)1dy
^- n
u(y)dy —
f (x,l} f (x,6)
-- 1 a-n
[Du(3I) •• (t! — x))dy
a(=,6)

— 1 ly — xr [Du(y) - (y — x)]dy•
nf <ly—sl<1
n (3.1.8)

When k = 1, it follows from (3.1.6) and (3.1.7) that

x1 1- "iDu(y)idy (3.1.9)
Iv — < 00 .
B (s,l)
116 3. Pointwise Behavior of Sobolev Functions

When k > 1, it follows from Lemma 3.1.3 with t = 1 and k — 1 substituted


for k, that

Iv — slt-'ID,u(y)Idy <— C ly — XI k_n [D9Diu(y)1Idy ,


ffin n
lal=k-1

which, by (3.1.6) and (3.1.7), is finite. Thus (3.1.9) still holds when k > 1.
By (3.1.9)

Urn ly — ri -n (Du(y) ' (y — (3.1.10)


6-.0+ 6<iy-zl<1

exists. It also follows from (3.1.9) that

Iy — xl l- "IDu(y)Idy
a l ô = 0,
e (:, 6 )
hence
b [Du(y) • (y — x)jdy --y 0 (3.1.11)
f6 (x,6)
as b -' 0+. It now follows from (3.1.8), (3.1.10), and (3.1.11) that the limit
in (3.1.5) exists.
(ii) The general case. Let II be an open set of R" . There exists an in-
creasing sequence {cp j ) of non-negative CO° functions on R", with compact
supports, and spt (p i C 1 for all j such that the interiors of the sets

{r : x E R" and Spy (z) = 1}

tend to S2 as j —► oo. Define

u(x)
u3 (x) = 0 (x) E
x41 ft.

By applying (i) to each of the functions u T , one can easily prove the theorem
in this case. ❑

3.2 Densities of Measures


Here some basic results concerning the densities of arbitrary measures are
established that will be used later in the development of Lebesgue points
for Sobolev functions.

3.2.1. Lemma. Let p > 0 be a Radon measure on R". Let 0 < A < co
and 0 < ci < n. Suppose far an arbitrary Bo re l set A C R" that
B z, rr))
lire sup (z >
r^0 T
3.2. Densities of Measures 117

for each x E A. Then there is a constant C = C(a, n) such that

p(A) > C\H' (A).

Proof. Assume p(A) < oo and choose e > O. Let U D A be an open set
with p(U) < oo. Let Ç be the family of all dosed balls B(x, r) C U such
that
01
x E A, 0 < r < e/2, p[B(â' > A.
r
Clearly, Ç covers A finely and thus, by Corollary 1.3.3, there is a disjoint
subfamily .F C Ç such that
AC[U{B:BE.F'}]U[U{B:BEf - P}]

(
whenever .F' is a finite subfamily of Y. Thus, by Definition 1.4.1,

H5a(A) < C c2y^ C5' E 6(B) 1

2
BEY' BEY-Y•
where 5(B) denotes the diameter of the ball B. Since .F C Ç and .F is
disjoint, we have

E 5 ( V <cA_1 E p(B)
BEY BEY
< Ca `p(U) < oo.

( o( B) 1
-

Since
cE 2
BEY-Y•
can be made arbitrarily small with an appropriate choice of .F', we conclude
HsF (A) < CA -3 p(U).
Since i is a Radon measure, we have that µ(A) = inf(p(U) : U D A, U
open). Thus, letting a --+ 0, we obtain the desired result. ❑

3.2.2. Lemma. Let > 0 be a Radon measure on R' 1 that is absolutely


continuous with respect to Lebesgue measure. Let

A = R" fl x : lim sup p[B(x, r)] > 0 .


r—o r
Then, Ha(A) = 0 whenever 0 < a < n.

Proof. The result is obvious for i = 0, so choose 0 < a < n. For each
positive integer i let
p[B(x'r)] > i-i
Ai = R" fl x : lx, < i, lirn sup
}

r—.0 r°
3. Pointwise Behavior of Sobolev Functions

118

and conclude from the preceding lemma that

µ(Ai) > Ci - 'H °'(A i ). (3.2.1)

Since A i is bounded, p(Ai) < co. Therefore H a (Ai) < oo from (3.2.1).
Since cY < n, H" (Ai) = 0 and therefore IAi(= 0 from Theorem 1.4.2. The
absolute continuity of µ implies µ(A i ) = 0 and consequently H°' (A ) = 0 ;

from (3.2.1). But A = U°_ I A i , and the result follows. O

3.2.3. Corollary. Suppose u E LP(R n ), 1 < p < oo, and let 0 < a < n. If
E is defined by

E = x: lirn sup r ' f >0 ,


B (x,r)
-

r— 0
,

then H°(E) = O.

Proof. This follows directly from Lemma 3.2.2 by defining a measure p as

n
p(A) --1u1 p dx.
A

3.3 Lebesgue Points for Sobolev F unctions


We will now prove the principal result of the first three sections (Theorem
3.3.3) which is concerned with the existence of Lebesgue points for Sobolev
functions. We will show that if u E WI P(R"), then

lim iu(y) - u(x)rdx = 0


r•-•0

for Bk, p -q.e. x E R". This is stronger than the conclusion reached in Theo-
rem 3.1.4, which only asserts the existence of the limit of integral averages.
However, in case u E W''P(Rn), the existence of the limit of integral av-
erages implies the one above concerning Lebesgue points. In this case, we
can use the fact that u - p4 E WIôp(R") for each real number p and then
apply Theorem 3.1.4 to conclude that

^^ô M y) — pidy
B (s,r)

exists for B 1 , p -q.e. x E R". Of course, the exceptional set here depends on p.
The object of Exercise 3.1 is to complete this argument. This approach fails
to work if u E W k iP(R n ) since it is not true in general that lui E Wk'P(Rn),
3.3. Lebesgue Points for Sobolev Functions 119

cf. Remark 2.1.10.

3.3.1. Lemma. Let k be a non-negative integer and A, p real numbers such


that p>1,kp< n, and k<11 <n/p. If

u E W k 'p(R n ), (3.3.1)

then
ba--k u(31)dy --► 0 (3.3.2)
B(x,a)
as b 0 +, for all x E Rn except for a set E with Ba, p (E) = O.

Theorem 3.1.4 states that the integral averages converge to a finite value
at all points in the complement of a B k , p -null set. This lemma offers a slight
variation in that the integral averages when multiplied by the factor 6A - k
converge to 0 on a larger set, the complement of a BA null set. At some
points of this larger set, the integral averages may converge to infinity, but
at a rate no faster than 5 k-A .

Proof of Lemma 3.3.1. (1) Suppose k = O. It follows from Corollary 3.2.3


that
bpa^^ Ita(y)Ipdy - 6 0 (3.3.3)
e(s,6)
as b 0+, for all x E Rn except for a set E with Hn -a p(E) = 0. From
the definition of Hausdorff measure, for e > 0 there is a countable number
of sets {E,} such that E C UEi and E(diam E i )n - aa < e. Each Ei is
contained in a ball Bi of radius r i where r ; = diarn Ei. Therefore, with the
aid of Theorem 2.6.13,

Ba,p(E) < E BÀ,p(E,) < CE r, -ap < Ce.


ti=1 i=1
Since a is arbitrary, we have that BA, p (E) = 0.
Now consider x E Rn - E. From Holder's inequality, there is a constant
C = C(n,p) such that
l fp

u(y)dy < C bap-n


15A —n
Iu(y)I°dy . (3.3.4)
f (s,6)

B(s,6)

(3.3.2) now follows from (3.3.3) and (3.3.4), and (i) is established.
(ii) Now suppose k > O. Let E be the set of all x for which

I3/ — MI A —n E Ib°u(1!)I dy = 00• (3.3.5)


Iai = k
120 3. Pointwisc Behavior of Sobolev F'nnctions

Then R,, p (E) = 0 and therefore Ba, p (E) = 0 by (2.6.8).


Consider z E R n — E. When k = 1, it follows from (3.3.5) that

f n(x,i)
ly
- xla - k+ l - niDu(y)Idy < oc. (3.3.6)

When k> 1, we replace t byÀ- k+ 1 and k by k-- 1 in Lemma 3.1.3 and


again derive (3.3.6) from (3.3.5). For z E fi n — E, we now show that

ba -k Iv _
xl l- nlDu(y)Idy -■ 0, (3.3.7)
f <ly-xj<1

as b j 0. Let r E (0, 1) be arbitrary. Clearly

45.x -k , xl l- nlDu(y)Idy --► 0


Iv (3.3.8)
f <jg-s(<1
as 10. When 0 < b< rwe have
aJl -k xl l- ^IDu(y)Idv Ç +I -n
^31 - xI^ -k lDu(y)Idy.
6<^y-x^<r 13(x ,r)
(3.3.9)
It follows from (3.3.6) that the right-hand side of (3.3.9) approaches zero
as r j 0. (33.7) now follows from (3.3.8) and (3.3.9).
Clearly,

fs(x-,1) Iu — x I IDu(y) Idv < oc. (3.3.10)

Since JI - k - n < 0, it follows that

6A-k-n k+ln
I^— xl I Du(p)idy <Iv — xl lDu(v)Idy ,

B(x,6) f ( x ,6)
so that by (3.3.6),

b a-k-n Iv - zI I Du(y)Idy 0 (3.3.11)


f(x.6)
as b j 0. By putting r = 1 in Lemma 3.1.1, one can obtain (3.3.2) from
Lemma 3.1.1, (3.3.10), (3.3.11), and (3.3.7). ❑

3.3.2. Theorem. Let 1, k be integers such that k > 1, 0 < k < l and
1p < n, p > 1. Let u E WkiP(Rn) and for each z E R n and r > 0 put

u x ,r u(y)dy.
fB(x,r)

Then
r(t-k)p u(j) -- u = . r l p dv --► 0
4 (m,r) (3.3.12)

3.3. Lebesgue Points for Sobolev Functions 121

as r 0, for all x E R" except for a set E with Bt, p (E) = O.


.

Proof. We proceed by induction on k. Suppose to begin with that k O.


It follows from Corollary 3.2.3 that


r tp—" lu(y)Ipdy -' 0 (3.3.13)
B(x,r)

for all x E R", except for a set E' with H" — P(E') = 0 and therefore '

B,, p (E') = 0. (3.3.14)

We now have, for x E R" — E',

1/p 1/p
r(1- (n ))
/P
^u(y) - uz , r l p d^f < r(t-("/p)) ^u(y)^pd^1
B (x,r ) L(X) [

1/p
^- r(t- (n /P)) Iux.r 1 dy . (3.3.15)
[
L(,)
But by Lemma 3.3.1,
rt u x , r - ► 0 (3.3.16)
as r . 0, for all z E R" except for a set E" with Bc p (E") = 0. (3.3.12) now
follows from (3.3.13), (3.3.15), and (3.3.16) in the case k = O.
Now suppose that k > 0 and that the theorem has been proved for all
functions of Wk -1 ip(R"). Let u E W k 1P(Rn). By the Poincar6 inequality,
which we shall prove in a more general setting later in Chapter 4 (for
example, see Theorem 4.4.2),

r(t-k)p-n
J(x,r)
ju(y) f ux,rIPdy < Cirif-(k-1AP_"
f (z.r)
IDu(y)Ipdv,
(3.3.17)
for all z E R", where C depends only on n. By the induction assumption,
there exists a set F', with
Bt,p ( F' ) = 0 (3.3.18)
and
r,i1-- (k -1)1v --n (3.3.19)
IDiu(y) - (D+u)z,rlPdy -. 0
B(aC,r)
asr10, forall xE R" -F'. But
1 /p 1/p

1D:u(N)^pdy < 1/31 04(y) - (Dtu)=,rl P dtl


[
L(X.r) [
fB,?)

122 3. Pointwise Behavior of Sobolev Functions

1/p

+ l(Diu)x,r l j [ dv } , (3.3.20)
6( x,r)

and by Lemma 3.3.1


re-(k--1)(Diu)s,r —• 0 (3.3.21)
as r J. 0, for all x E R", except for a set F" with

Bc p (F") = O.

(3.3.12) now follows from (3.3.17), (3.3.19), (3.3.20), and (3.3.21). This
completes the proof. ❑

3.3.3. Theorem. Let k be a positive integer such that kp < n, let 1-2 be an
open set of R" and let u E W k ip (S2) . Then


lu(y) — u(x)Ipdy --, o (3.3.22)
s(s,r)

as r j 0, for all z E S2, except for a set E with Bk, p (E) = 0.

Proof. (i) When SZ = R", (3.3.22) follows from Theorem 3.3.2 and Theorem
3.1.4.
(ii) When SZ is arbitrary, the theorem can be derived from (i) as in the
proof of Theorem 3.1.4. ❑

3.3.4. Corollary. Let k be a positive integer such that kp < n, let it be an


open set of R n and let u E W k.P(SZ). Then

lim lu(y)Ipdy exists and = lu(x)I' (3.3.23)


r-.0+ g(x,r)

for all x E Si, except for a set E with B k , p (E) = O.

3.3.5. Remark. Theorem 3.3.3 states that on the average, the oscillation
of u at x is approximately equal to u(x) at Bk , p -q.e. z E Si This can
also be stated in terms of the classical concept of approximate continuity,
which will be used extensively in Chapter 5. A function u is said to be
approximately continuous at x o if there exists a measurable set A such
that
t/3(xo, r) n AI
— 1
= (3.3.24)
1 ô B ( x o , r ) I
and u is continuous at zo relative to A. It is not difficult to show that if
u has a Lebesgue point at x o then u is approximately continuous at zo.
A proof of this is given in Remark 4.4.5. Thus, in particular, Theorem
3.3.3 implies that u E W 1, p(R") is approximately continuous at /31, p -q.e.
x R".
3.3. Lebesgue Points for Sobolev Functions 123

Approximate continuity is a concept from measure theory. A similar con-


cept taken from potential theory is fine continuity and is defined in terms
of thin sets. A set A C Rn is said to be thin at x 0 relative to the capacity
Bk, p if
Bic (A n B(xo, r)11 1/(P-13 dr
1
— <oo. (3.3.25)
Bk,pjB(xo, r)1 r
[

o
A function u is finely continuous at x 0 if there exists a set A that is thin
at x 0 and
lim u(x) = u(x o ).
Z +t o --

xvA
It follows from standard arguments in potential theory that A can be taken
as a measurable set. In the case of the capacity, B1,2, which is equivalent
to Newtonian capacity, these definitions are in agreement with those found
in classical potential theory. In view of the fact that
71/(n-kP)
IAI GIBk.P(A)1

for any set A C Rn, it follows that (3.3.25) implies

11m IB(xo, r) n (Rn - A)I ' 1,


r---'d IB(xd, r)I
and therefore fine continuity implies approximate continuity.
We now will show that the approximate continuity property of Sobolev
functions can be replaced by fine continuity. First, we need the following
lemma.

Lemma. If {A 1 } is a sequence of sets each of which is thin at x 0 , then


there exists a sequence of real numbers {r 1 } such that
00

U
i-1
Ai n B(x , r i )
Q

is thin at x 0 .

Proof. Because A ; is thin at x0, it follows that there exists a sequence


{r i } -, 0 such that

Bk, p[A11r fl B(xo, r0)I 0 as i oo.


Bk,PI B(x0, ri )]
We may assume the r, to have been chosen so that

T' Bk,PIAi n B(z o , r)) 1 /(P-') dr _


< 2(^+t ?,

d Bk,y[B (xo, r)^ r


124 3. Pointwise Behavior of Sobolev Functions

Then,

I ' [Bk. p [Ai n B(xo, r) n B(x o , 1 1/(P-1) dr


Bk, F [B(xo, r))
1/(p-1) dr
fo r, $k.p[A+ n B(xo, r) n B(xo, r^ )j
Bk,p[B(xo, r)1 r
-1)
i [Bk, p [A, fl B(x o , r) n B(x o , r ; )1 1/(p dr
+ r, Bk,p[B(xa, r)] r
1 1 1 '/(p-1) dr
< 2 -(i +l) + Bk .p[A, n B(x o , r;)['/(p -1)
r, CrTh_"P r [

1
]

< + B k ,p [A ; n B(x o , r,)1'/ ( p -1) G 1 1 1


)jl/(P- 1)
Bkm[B(Z0, ri

< 2' for r s sufficiently small.


Since capacity is countably subadditive, the result easily follows. ❑

For ease of exposition, we now restrict our attention to u E W 1 ' 13 (R").


Again, we see the important role played by the growth of the gradient in
order to obtain some regularity at a given point.

Theorem. Let xo E R", p > 1, and suppose u E W' '(R") has the property
that
"(p-1)
dr
J: '
r p—" IDuI
B(xo,r) r
dx < oc.

Suppose also that


[irn u(y)dy = u(xo ).
r—.o J(z o ,r)
Then u is finely continuous at x o .

Proof. For each e > 0, let

A(x o , e) = Rn n {x : I u(x) — 4l(x 0 )I > E}.

For r > 0, let


yr (x) = pr (x) [u(x) - FL(2r)]
where (so, is a smooth function such that ipr - 1 on B(x o , r), apt 4p r C
B(x o , 2r), < Cr -1 and where

T../(2r) = u dx.
B(x0,2r)
3.3. Lebesgue Points for Sobolev Functions 125

Because of the assumption û(2r) u(x0) as r -, 0, note that for all


sufficiently small r,

v r (x) > e/2 for x E A(x o , e) fl B(r o , r).

Therefore, by appealing to Exercise 2.8, which allows B1 1p to be expressed


in terms of a variational integral, there exists C = C(p, n) such that

-1 )p
B1.1,1A(xo e) B(xo, r)) < C(2e
, ^Dvr^pd^
B(so,2r)

< C(2e -1 )F ^Du^ p dz


B(so,2r)
r-1)p
+ (C2e- 1 lu - û(2r)1 pdx. (3.3.26)
B(so,2r)

An application of Poincaré's inequality (cf. Theorem 4.4.2) yields

lu - û(2r)Ipdx < Cr" IDujpdx


f(xo,2r) f(zo,2r)

and therefore (3.3.26) can be written as

B1.p(A(ro, E) n B(xo, r)1 < Ce-1r-n


rn-p 13(z0,2r) f Duf pdx,

which directly implies that A(xo, e) is thin at z 0 . Now let e3 be a sequence


tending to 0. By the preceding lemma, there is a decreasing sequence 5 -. 0
such that

U EA(xo ei) n B(xo> r )j


00

A= , 2

^
-1

is thin at x0 . Clearly
lim u(x) = u(X0)
s oso

sER" —A

and the theorem is established. ❑

It can be shown that


1 1/(p-1)
- n dr
rp IDuIpdx — < co (3.3.27)
0 B(sa,r) r

for B 1 , p -q.e. xp E R" cf. 1ME31. Therefore, with Theorem 3.3.3, we obtain
the following.

Corollary. If u E W l "p(R") then u finely continuous al all points except


for a set of B lip capacity zero.
126 3. Pointwise Behavior of Sobolev Functions

Observe that Corollary 3.2.3 implies

lim sup rp "I DuI pdx = 0


— (3.3.28)
r -•0 B (xo,r)

for H" — p-a.e. x0 E R. Although (3.3.27) implies (3.3.28) for each z0, the
exceptional set for the former is larger than that for the latter.

3.4 II-Derivatives for Sobolev Functions


In the previous three sections, the continuity properties of Sobolev func-
tions were explored through an investigation of Lebesgue points and fine
continuity. We now proceed to analyze their differentiability properties. We
begin by proving that Sobolev functions can be expanded in a finite Tay-
lor series such that for all points in the complement of an exceptional set,
the integral average of the remainder term tends to 0, (Theorem 3.4.2). In
keeping with the spirit of this subject, it will be seen that the exceptional
set has zero capacity. Obse rv e that Theorem 3.3.3 provides the first step
in this direction if we interpret the associated polynomial as one of degree
0 and the remainder at x as lu(y) — u(x)I.
_
When k, m are integers such that 0 < m < k, (k — rn)p < n and
u E Wl u 1 (Rn), it follows from Theorem 3.1.4 that there exists a subset E
of R" such that
B,c_,,,,p(E) = 0 (3.4.1)
and
11m D°^u(y)dy (3.4.2)
r^0+ B(z,r)

exists for all x E R" — E and for each multi-index ar with 0 < led < rn.
Thus, for all such x, we are able to define the Taylor polynomial em) in
the usual way:

Pâm)(!1) = L ! Dau(x)(y — x)a• (3.4.3)


° +Q) m

(Recall the notation introduced in Section 1.1.) Observe that when 7R is a


Cm function on RP, Taylor's theorem can be expressed in the form
a m--2
u(U) = Rz m-1) (10 + rn E ° (1 — t }

iaE =m a.
• D"u[(1 — t)x + ty]dti (y x)". — (3.4.4)

3.4.1. Theorem. Let 1 < i < k and suppose (k — rn)p < n. Let u E
3.4. L'-Derivatives for Sobolev Functions 127

W k"(R") and E be the set described to (3.4.1) and (3.4.2). Then


1/p 1
(1 _ t )m-1
L(z,r)
lu(v) — Pzm) (v)I Pdy < Tm E C! a
{
iai =m
1/p

I D ° u(y) - D aru(x) IPdy dt (3.4.5)


B(z,tr)
[ rn
and
1/p 1

[
L(z ) r) l
(
u(y)- p=m -t)(b)^pd^1
< Tm
Iai=m
a! o ^ t
(^ - m

1/p

• t'„ ID°u(y)Ipdv dt, (3.4.6)


B (z.tr)

for all x ER" except for a set E'iE with Bk -„a , p (E')= O.

Proof. (i) Suppose first of all that u is a Cm function on fr. Let r E R",
r > 0 and put B _ B(x, r). Let cp be a function of L" (B) with Ijcp^j< < 1
where p' is the conjugate of p. By (3.4.3) and (3.4.4),

i [ 1411) - R1m)(y)14:0)dy - E
la+=m
7:1 (1 -
7
t)m -1

{D°' u((1 - t)x + ty) - Dau(x)}(y - x)°cp(y)dy dt.


[113
Hence, by Holder's inequality,
^

IJ ru(y) — R1m (0 1 cP(04 < r


)
-

„` ^

E
iaj =m
a .
f (1 _ t )m -1

1/p
• IDau((1 - t)x + ty) - Dau(x)IPdy dt.
B
By making the substitution z = x + t(y - x) in the right-hand side and
then taking the supremum over all cp, one obtains (3.4.5).
The inequality (3.4.6) can be derived similarly.
(ii) Now let u be an arbitrary function of W k,P(R „ ). By Theorem 3.3.3,
there exists a set E' D E, with Bk_„,, p (E') = 0 such that

lim ID4u(y) - Dau(x)IPdy = 0 (3.4.7)


4-' 0+ B(z,6)
128 3. Pointwise Behavior of Sobolev Functions

when 0< Pal <m and x E R"—E'.


Consider r E R" — E'. There exists a constant M (depending on s), such
that
IDau(y)Ipdy <_ M (3.4.8)
B(z,6)
for all lai = m and all b > O. Let {y,,} be a sequence of regularizers as
discussed in Section 1.6. Thus, cp, E Cfl (Rn),

cPE(x)dx = 1, (3.4.9)
Rn

apt cPc C B(0, e) and


sup w,(x) < CC" (3.4.10)
'ER"
for all e (where C depends only on n), while

(Sp, * Dau)(x) — D°u(x) (3.4.11)

as i 0, for 0<'al <m and x E R"—E'. Put u, =cp,*u. Each u, E


C °° (Rn )nW k ip(R"). Let us denote by (3.4.5), and (3.4.6), the inequalities
(3.4.5) and (3.4.6) with u replaced by u,. Since u, is smooth we know
that (3.4.5), and (3.4.6), are valid. By (3.4.11) and Fatou's lemma, the
lower-limit as e 1. 0 of the left-hand side of (3.4.5), and (3.4.6), is greater
than or equal to the left-hand side of (3.4.5) and (3.4.6). The result of the
theorem will thus follow from Theorem 1.6.1(iî) and Lebesgue's Dominated
Convergence theorem when we show for each a with lai = m and r > 0
fixed, that the following function of t, 0 < t < 1, is bounded; that is,

t-"
fB(x,gr)
1D a u,(y)1Pdy < Mr" (3.4.12)

where M is independent of e.
We now proceed to establish (3.4.12). For any measurable subset E of
Rn, we have (when Ian = m)
p

E if
IDauf(y)ipdy = ço,(y — z)Dr4i(z)dz dy,
E
R"

hence by (3.4.10)
P
ID a uE (y)I p dy < C P C ( " ) I D °`u(z)Idz dy. (3.4.13)
1 E B(y,c)
Thus, when p > 1, we have by Holder's inequality

P 1
H 4
-

E B(U,^) IDau(Z)'pdz

E IDau^(y)Ipdy<CpE-(+^p) B (U.e)
3.4. LP-Derivatives for Sobolev Functions 129

so that

1 ID°u e (y)1Pdy < Ce'


!
IDQu(z)^Pdzdy,
E f (y,f)
(3.4.14)

where C depends only on n and p. When p = 1, (3.4.14) follows from


(3.4.13).
When tr < 3E, we let E be the ball with center x and radius tr. Since
B(y,e) C B(x 4E) when y E B(x tr), (3.4.14) implies that
, ,

-n ID°u(z)Ipdzdy.
ID ° u E (y)I P dy < CE
f (x,tr) B(x,tr)B(s,4e)

It now follows from (3.4.8) and (3.4.12) holds in the case where tr < 3c.
When tr > 3E, we have

I D ° uc(1!)Î P dy = f ID«üc(y)lpdy+ ID ° lie (Or dy


f (x,tr) B(z,3e) f f<1y-rl <tr
and a double application of (3.4.14) yields

f(x,tr)
iD a ue(y)I P dy < CE_ -n
ID"u(z)rdzdy
B(x,3e)f(x,4F)
+ CE -n ID°u(z)1pdzdy
fe<ly-2I<tr B(y,e)

and by (3.4.8)

< C't"rn + Ce -n 1D°`u(w + y)Ipdwdy


f e<' V -sl<tr f (O,E)

< C' tn rn + C^t• rn

f B(0 e) 2e
1
D°u(y)1Pdyy dw

so that
fD°ue(y)IPdy Ci"tnrn.
f(x,tr)
Thus (3.4.12) is established. Q

3.4.2. Theorem. Let 0 < m < k and suppose (k - rn)p < n. Let u E
W k •P(Rn). Then,
1/p

r -m
[fB( x . r )
l u( Y ) - PIm ( y) I p dy
) ^ 0

as r l 0, for all x E Rn, except for a set F with

Bk-m,p(F) = O.
130 3. Pointwise Behavior of Sobolev Functions

This is the main result of this section. In particular, it states that the
integral average over a ball of radius r of the remainder term involving the
formal Taylor polynomial of degree k tends to 0 as r —. 0 at a speed greater
than r ig at almost every point. If a Taylor polynomial of smaller degree is
considered, the integral average tends to 0 at perhaps a slower speed, but
on a larger set.

Proof of Theorem 3.4.2. When m = 0, the theorem reduces to Theorem


3.3.3. Suppose via > O. By Theorem 3.3.3,

ID Q u(y) — DQu(x)Iedy —' O (3.4.15)


B(s,r)

as r j 0, for all laf = m for all z E R", except for a set F with
Bk_mp(F) = O.

Consider r E R" — F and an a with fal ï m. Define

f
1/p
9(r) = r _n
[
ID°'te(y) — D °u(x)f pdy (3.4.16)
L()
for r > O. By (3.4.15),î(r) -- Oasr J.
, hence

I.
1
(1 — t)m -2 7)(tr)dt —■ 0 (3.4.17)

r j O. The required result now follows from (3.4.16), (3.4.17), and Theorem
3.4.1. ❑

3.5 Properties of 11-Derivatives


In this section we consider arbitrary functions that possess formal Taylor se-
ries expansions and investigate their relationship with those functions that
have Taylor series expansions in the metric of LP, such as those discussed
in the previous section.

3.5.1. Definition. Let E C Rn . A bounded function u defined on E


belongs to Tk(E), k > 0, if there is a positive number M and for each
z E E there is a polynomial P2 (•) of degree less than k of the form
ua(x) (y— x)°t, (u o (15.1)
Px(y) _ = u)
a.
In*I>o
whose coefficients u Q satisfy

luck(s)( < M for z E E, 0 < lai < k,


3.5. Properties of Dl-Derivatives 131

and
un(y) = D ° PS(y) + Ra(x, y)
whenever x, y E E and where Ra (x, y) < Cly — xl k— W, 0 < iai < k.
The class tk (E) is defined as all functions u on E such that for each
E E there is a polynomial P= (-) of degree less than or equal to k of the
form (3.5.1) such that for 0 < iaj < k,

Er PIA) = D °`Pt(y) + Ra(z, y),

whenever x, y E E with IR,(x, y)1 < Cly -- xi k-1 'I and

lim R°' (xky )^ i ` 0


-

y ly i
uniformly on E.
As a mnemonic, Tk(E) and t k (E) may be considered as classes of func-
tions that possess formal Taylor series expansions relative to E whose re-
mainder terms tend to 0 "big O" or "little O," respectively.

3.5.2. Remark. Clearly, if u E T k (E) then u o, is locally Lipschitz on E,


0 < lai < k. If E is an open set, note that the derivatives Dku exist on E,
0 <jal< k, and that

D a u(x) D'PP (x) - u a (x) for x E E.

Since IDaPP (x)I < M for x E E, it follows that u E Wiâ-' '(E) for every
p > 1. The space tk (E) may be considered as the class of functions on E
that admit formal Taylor series expansions of degree k. Of course, if E were
open and u E C k (E), then u would have an expansion as in Definition 3.5.1
with
Pr(y) = E 1 D°u(x)(y xr.
0<la 1<k

Moreover, if u E C k (R") and E C R", then the restriction of u to E,


ulE, belongs to tk(F) for each compact set F C E. One of the reasons for
identifying the class t k (E) is that it applies directly to the Whitney exten-
sion theorem [W11), which we state here without proof. We will provide a
different version in Section 3.6.

3.5.3. Whitney Extension Theorem. Let E C R" be compact. If u E


t k (E), k > 4 an integer, then there exists û E C k (R") such that for 0 <
Pi < k

D4317(x) = DsPP (x) for all r E E.


In view of this result, it follows that u E t c (E) if and only if u is the
restriction to E of a function of class Ck(Rn).
132 3. Pointwise Behavior of Sobolev Functions

We now introduce another class of functions similar to those introduced


in Definition 3.5.1 but different in the respect that the remainder term
is required to have suitable decay relative to the LP-norm instead of the
LQ°-norm. The motivation for this definition is provided by the results
established in Section 4 concerning Taylor expansions for Sobolev functions.

3.5.4. Definition. For 1 < p < oo, k a non-negative integer, and z E


R", T k 1P(x) will denote those functions u E LP for which there exists a
polynomial Pr (.) of degree less than k and a constant M = Mfr, u) such
that for 0<r<oo
1/p
Iu(y) — Px(y)Îpd}f < Mr k _ (3.5.2)
(
L(XF)

When p = oo, the left side of (3.5.2) is interpreted to mean esssup"B(=,r)


lu(y) — Pr (y)I. T' (x) is a Banach space if for each u E T k 'p(x) the norm
of u, IlulI T k .P(x), is defined as the sum of NG, the absolute value of the
coefficients of Px , and the smallest value of M in (3.5.2).

3.5.5. Definition. A function u E 74, P(x) belongs to t k IP(x) if there is a


polynomial of degree less than or equal to k such that
1/p

in(y) — P2(y)1PdY = o(rk) as r --• 0. (3-5.3)


B(z,r)

Note that if u E Tk'P(x) the polynomial Pr is uniquely determined. To


see this write
u(y) = Ps(y) + R=(y)
where
1/p

1Rz(0Ipdy < Mr k .

(IB(z,r)

If Ps were not uniquely determined, we would have u(y) = Q x (y) + R r (y),


where R 2 satisfies an integral inequality similar to that of R x .
Let SA (y) = P2 (y) —Q x (y). In order to show that S = E 0, first note that

B
1 1B
ISz(y)ldy <
(z,r) /p
(z,r) B(x,r)
FSz(y)Ipdy < Cr', 0 < r < oo.

Now let L s be the sum of terms of S x of lowest order and let Mx = SE L E . —

Thus, L s has the property that for each A E R 1 , L s (Ay +x) = A° L z (y + x ),


where a is an integer, 0 < a < k 1. Since Ms is a polynomial of degree

at most k 1, we have

1MZ(b)Idy <Cr k-1 , 0 < r <oo.


3.5. Properties of LP-Derivatives 133

It follows from the inequality jLgy), < 1S z (y)1 + IM=(Y)1 that

r° lLx(y)Idy =
8(x,1)

-I^x(y)Idy
B (a ,r)
< Cr k + Cr k-i , 0 < r < oo.

This is impossible for all small r > 0 if a < k - I and L = is non-zero. If


a = k - 1, then Ms - 0 and the term Crk -1 above can be replaced by 0.
A similar argument holds in case u E t k ie(x).
Obviously, t k (E) C t k IP(x) and Tk(E) C T k i'(x) whenever r E E and
p > 1. We now consider the question of the reverse inclusion. For this
purpose, we first need the following lemma.

3.5.6. Lemma. Lei k be a non-negative integer. Then there exists cp E


Co (R n ) with spt cP C { Ix' < 1) such that for every polynomial P on Rn
of degree < k and every e > 0,

ci)£ * P=P

where cp e (x) = e ' ço(x/e).

Proof. Let V = C'°(B) where B is the closed unit ball centered at the
origin and let W denote the vector space of all m-tuples (y} whose compo-
nents are indexed by multi-indices a = (al, ar ts ... , a„) with 0 < lai < k.
The number m is determined by k and n. Define a linear map T: V -e W
by
T() = cp(x)xadz ;
R"
thus,
y et =Sp(x)xadx
^ n

where 0 < 'al < k and x° = xi' x2 7 .. -xn^.


Note that vector space, range T, has the property that range T = W for
if not, there would exist a vector, a = {a} orthogonal to range T. That
is,
E a a y, = 0 whenever y = { y a } E range T.
This implies,

cp(x)E a.xadx = 0 whenever Sp E V.


134 3. Pointwise Behavior of Sobolev Functions

Select rl E V such that rl > 0 in (z :1x1 < 1). Now define t' by = E a c,xa11
and note that 1i E V. Therefore,

/R. (E a^x°) 2 rJ(x)dz = 0,

which implies E a Q x° = 0 whenever 1x1 < 1. But this implies that all m
numbers a s = 0, a contradiction. Thus, range T = W. In particular, this
implies there is Sp E V such that

çp(x)dx = 1, (p(x)zadx ï 0, 0 < lai < k.


fR R"
Since any polynomial Q of degree no greater than k is of the form

Q(z) = E bQz ar
o<iaI<k

it follows that
çp(z)Q(z)dz Q(0).

Given a polynomial P = P(x) as in the statement of the lemma, let z =


(x — y)/e and set Q(z) = P(x — €z) to obtain the desired result. ❑

The next theorem is the main result of this section. Roughly speaking,
it states that if a function possesses a finite Taylor expansion in the 1/-
sense at all points of a compact set E, then it has a Taylor expansion in
the classical sense on E. It is rather interesting that we are able to deduce
a Lw-conclusion from a V'-hypothesis. A critical role is played by the
existence of a smoothing kernel (P that leaves all polynomials of a given
degree invariant under the action of convolution.

3.5.7. Theorem. Let E C Rn be dosed and suppose u E T k 'P(x), 1 < p <


oo, k > 0, with IiurlTk.o(r) < M for all z E E. Then u E Tk (E). Also, if E
is compact and if u E t k .P(x) for all x E E with (3.5.3) holding uniformly
on E, then u E t k (E).

In view of Whitney's Extension theorem (Theorem 3.5.3), note that a


function satisfying the second part of the theorem is necessarily the restric-
tion of a function of class C k (R"). In the next section, we will investigate
Whitney's theorem in the context of LP.

Proof of Theorem 3.5.7. Let Sp E Co (R") be the function obtained in


Lemma 3.5.6 such that
SP, * P(x) = P(z) (3.5.4)
3.5. Properties of IF-Derivatives 135

whenever P is a polynomial of degree less than k, e > 0, and x € R". Note


that (3.5.4) implies

D a çpe * P(x) = cpe * 13' P(x) = D°`P(x). (3.5.5)

Since u E TSCiP(x o ) for all x o E E, we have for xo and x E E,

u(y) = P:o(y) + R(xo, y) (3.5.6)

and
u(y) = Px (y) + R(x,y) (3.5.7)
where
ilp
IR(x', y)Ipdy S Mr k , (3.5.8)
( fB(. r)

with x` either x0 or x. Now let e = lx — xol and for 0 < 1131 < k consider

I = DI3 cp e * u(x).

For each fixed z E R", define R as R s (x) = R(z, x) whenever x E R.


;

From (3.5.6) and (3.5.5) it follows that

I = D 0 cp e * Pzo (x) + Dpcp, * R zo (x)


= D I3 Pzo (x) + 13''(p * Rso (x)•

Similarly, using (3.5.7) and (3.5.5), we have

I = D'P= (x) + D 0 ç0 f * Rz (x)


= uo(x) + Ds(p e * Rz(x)•
Therefore,

D I3 Ps(x) = D A Pxo(x) + [D 13 (pe * (Rzo — Rz)1(x)


= D Q P=o (x)
+ f (n+IDDK3cP (2; y) [

IR(xo, y) — R(x, y)]dy•


F

Because cp =— 0 on 1x1 > 1, the last integral is taken over B(x, e). Since
B(x, e) C B(xo, 2e), the integral is dominated by

C 1R(xo, y)Idy + 1R(x, y)Idy e — I 01 (3.5.9)


IB(tp,2e) $ (x,e)

where C depends on an upper bound for jD 13 cpI. Jensen's inequality and


(3.5.8) implies that (3.5.9) is bounded by CMe k-1131 , thus proving u E
Tk(E).

136 3. Pointwise Behavior of Sobolev Functions

A similar proof establishes the second assertion of the theorem. Indeed,


as before we obtain

D a Pr(x) = D Q Pro (x) + (D R cPf * (Rs o - Rr)J(x)


= D 3 P=0 (x)

+ e-("+I,QI}DB(p (x F y} (R(xo, y) R(x, y)ldy -

< C IR(xo, y)Idy IR(x, y)jcly C —IAI.


B(x0,2e) B(x,c)

Since (3.5.3) is assumed to hold uniformly on E, for q > O arbitrary, the


last expression is dominated by i£k^IQ) xulk-I61 provided Ix - xoI
is sufficiently small. The compactness of E is used in this case to ensure
that IRQ(x, y)I < Cox - ya k- I 131 whenever x, y E E.
-

3.6 An LP-Version of the Whitney Extension


Theorem
We now return to the Whitney Extension Theorem (Theorem 3.5.3) that
was stated without proof in the previous section. It states that for a corn-
pact set E C R", a function u is an element of t k (E) if and only if it is the
restriction to E of a function of class C k (R"). The result we establish here,
which was first proved in [CZ], is slightly stronger in that the full strength
of the hypothesis u E tk(E) is not required. Instead, our hypothesis requires
that u E t k up(x) for all x E E with (3.5.3) holding uniformly on E.
We begin by proving a lemma that establishes the existence of a smooth
function which is comparable to the distance function to an arbitrary closed
set.

3.6.1. Lemma. Let A C R" be closed and for x E R" let d(x) = d(x, A)
denote the distance from x to A. Let U = {x : d(x) < 1}. Then there is a
function ê E C °° (U — A) and a positive number M = M(n) such that

M - 'd(x) < b(x) < Md(x), r: E U — A,

IDa6(x)I < C(a)d(x)'`iul, x E U - A, > O.

Proof. Let h(x) = Z0 d(x), x E U - A, and consider a cover of U - A by


closed balls {B(x, h(x))}, with center x and radins h(x), x EU_- A. From
Theorem 1.3.1 there is a countable set S C U - A such that {B($, h(s)) :
s E S} is disjointed and
R" — A D {UB(8,5h(s)) : s ES } ❑ U — A.
3.6. An LP-Version of the Whitney Extension Theorem 137

With a = 13 = 10 and A = 15 , we infer from Lemma 1.3.4 that

3 h(x)/h(s) < 3 for s E Sz . (3.6.1)

Let 0(x) = H°(S M ) < C(n) and let rj : RI -, 10,11 be of class C°° with

rgt) =l for t <I, n(t) = 0 for t > 2.


Now define lb E C°°(R") by i(x) =i(lx1) and y, EC°°(U) by

v,(x) = h(8)0 ( 5h
for s E S, x E U.
s8)
Note that apt v, C R(s,10h(s)), v = h(s) on B(s, 5h(s)) and from (3.6.1)
that
ilrvs(x)1 < h(s)N(a)[5h(s)j - M"M
< N(a)h(x)1 -l`if for E Ss ,
where N(a) is a bound for IDp i1, 1/31 < 101. Now define

b(x) = v,(x) = E v8(x) for x E U.


'ES 'ES:

Clearly,
d(x) ' h(x) < 9(x)d(x)
a(x) < 3e(x)h(x) = 2
60 0
and
iD°a(i)i < 5 - lQ 1 3 1 a 1-1 0(x)N(a)h(x) 1- faM, for x E U - A. ❑

The following is only a prelude to the D'-version of the Whitney exten-


sion theorem, although its proof supplies all of the necessary ingredients.
Its hypothesis only invokes information pertaining to the spaces Tk.P (x)
(bounded difference quotients) and not the spaces t k iP(x) (differentiabil-
ity). In particular, the theorem states that if u is Lipschitz on A (the case
when k = 1) then u can be extended to a Lipschitz function on an open
set containing A. This fact is also contained in the statement of Theorem
3.5.7.

3.6.2. Theorem. Let A c R° be closed and let U = {x : d(x, A) < 11.


If u E Lp(U), 1 < p < oo, and there is a positive constant M such that
< M for all x E A, where k is a non-negative integer, then there
exists û E C - ^ (U) such that DIû(x) = D PP (x) for x E A, 0 <101 < k.
'

Proof. Let b denote the function determined in Lemma 3.6.1. Define Ti = u


on A and for x E U- A let

u(x) = cooa(=) su(r) (3.6.2)


138 3. Pointwise Behavior of Sobolev Functions

where cp is the function determined by Lemma 3.5.6 and where

40 6(x)(Y) = b(xrnV' b(x) •

Thus, ü is defined at x as the convolution of SPa(=) and u evaluated at z.


Because both cP and ô are of class C°° it is easily verified that TIE C°O(U-
A). For x E U, let z' be a point in A such that 4x - x' 1 = d(z) = d(x, A).
Because u E T k 'p(x) we may write

u(x) = Ps . (x) + R. (x) (3.6.3)

where
1 /p

IR1. (z)W p dz < Mr k .


B(x',r)
By substituting this expression into (3.6.2), we obtain

D" (x) = DA [(P6(s) * P=' (x)] + D13 kP6(x) * R.. (z)]

= (DpcP5(2) * P1- (z) + f R 13 (x, y)R'. (y)dy

0
= cp6(x) * (D P2•)(z) + f Rp(x , y)R1 . (y)db (3.6.4)

where R13(x, y) D 13 {6(z) - nw[(x - y)5(x) -1 ]}. Applying Lemma 3.5.6 to


the first term on the right side of (3.6.4) we obtain

Dpri(x) = D 0 Pz . (z) + f Rs(x, y)R7• (y)dy. (3.6.5)

We wish to estimate the remainder term in (3.6.5) which requires an anal-


ysis of RS (z, y). It can be shown that

IR13(x, y)1 < C(0)d(x)-n-101

and consequently

1 f Rs(x, y)Rz• (y)dy < C(0)d(x)



-n-1 13 1
iRz- (y)Idy.
B( =,b)
(3.6.6)

Because 6(x) is comparable to d(x) (Lemma 3.6.1) and Ix - x' I = d(x), it


follows that i(x, ô(x)) C B(x', Kd(x)) for some K > O. Therefore from
(3.6.3) and Holder's inequality,

f (s•,Kd(x))
IRx• (y)Idy < M[Kd(x)} n+k (3.6.7)
3.6. An Imo-Version of the Whitney Extension Theorem 139

which along with (3.6.5) and (3.6.6) implies,

(3.6.8)

DAu(x) - D 09 P2 . (x) = Sp(x`, x)

where
< C(Q, k)MIx -
x+lk—IQI
1S0(x`,x)(
.

We emphasize here that for given x E U — A, (3.6.8) is valid only for x' E A
such that d(x) = ix - x* 1. We now proceed to establish the estimate for
arbitrary x' E A.
By assumption IiüfITh.,(x) < M for all x E A. Therefore, we may apply
Theorem 3.5.7 to conclude that u E T k (A). Thus, if xi E A,

Px-(x') u(e)

and

Da Pz . (ac') DaP=; (x') + x'), 0 < lal < k (3.6.9)

where
IR0(4, x')I C(a, k)M1x' - x l lk - lal.
By Taylor's theorem for polynomials, it follows that

k-1-1091 1
D^ Ps • (x) _ E — D°+aPs. ( x +)( x - x')a.
a!

Thus, by (3.6.9) and Taylor's theorem,

k -1 -1091
DRP=• ( x ) _ LrDfi+aP=` (x • ) + ^+S(zi, x+)1(x — x' a )

L--4
lal= 0
k- L -Ipi k- 1 - (101 +1091) 1
r 7' — z'1 7
E
)

u
1aj=o
a!
=a
+ Rail? (xT, x•)) (x - x')a. (3.6.10)

By Taylor's theorem, it follows that


DpPz (x) _ E t Do+QPx, (xi)(x - xi ) a .
lal>o
Therefore, since

^ x -x'j <ix -x11 and ix' - <- -x1+ix-x,l <21x- x,^,


140 3. Pointwise Behavior of Sobolev Functions

(3.6.10) becomes (after some algebraic simplification)


(k-))-I$I 1 .
DI3 P . ( x) - E c7 D0+a Pz'
(xi)(x — zir = O(lx — x11 k-Is1 )
.
Ial=a
It follows from (3.6.8), that
(k-1)--101
Dpû(x) -
E 1
l Dp+aPz;
(xi)(x — xi)° = O( I x — xE1kWIs1)
1a1= 0
or
./Y3 û(x) - D"PP ; (x) = 0( lx - x7lk - 1 131 ), (3.6.11)
Thus, (3.6.11) holds whenever act € A and x E U - A and Theorem 3.5.7
implies that it also holds with Dsû(x) replaced by us(x) whenever x E
A. This implies that DOT is a continuous extension of up and that this
extension has a Taylor series expansion about each point in A. Since û E
C°° (U — A) it now follows that ii E C k-1 (U).
In order to prove that ii E C k-1,1 (U) it suffices to show that DQû is
Lipschitz, ff1 = k - 1. We know from (3.6.11) that if a E A, and 101 = k -1

ID 8 û(x) - D'û(a)1 < C(k)Mlx - al (3.6.12)

for z E U. Therefore, it is necessary to consider only the case z, y E U - A.


First suppose lx - yl > id(y) and let a E A be such that d(y) = la - yl.
Then, la - y1 < 21x - yl and

Ix - al <Ix-v1+ly-al S 31x_yl.
Thus, utilizing (3.6.12),

IDsU(x) - D ° u(y)I < I D' (x) - D 13 ic(a)1 + ID a u(v) - Dtt(a)1


< 1D Q17 (x) - D' i(a)1 + ID' (v) - D /31 (a)1
< C(k)M[Ix - al + Iv - al)
< 5C(k)Mlx - vl.
Finally, suppose Ix - yl < Id(y) and d(y) = la - yl. Using (3.6.5) with
101= k - 1 and the Mean Value theorem, we have

ID a ri(x) - D' (y)I = f R(a, z) IRt(x, z) - R1(v, z)J dz

Ix — yl f IDxRs(xo, z}l 'Mat x)ldz (3.6.13)

where xs is a point on the line segment joining x and y. Now apt Rs(xo, y) C
B(x a , 6(x o )) and ô(x o ) < Cd(xo). Thus,

ID=Rt(xo, z)I < C(Q)d(xa)-n-k,


3.6. An LP-Version of the Whitney Extension Theorem 141

(Pi = k - 1). Therefore, (3.6.13) implies

!D1(x) - Dike(y)1 5 C(Q)Ix - yld(xo) -n-k IR(a, z)Idz.


B(xo,Cd( =o))
(3.6.14)
Since Lip(d) = 1, we have

2d(x o ) > d(x) > d(3I) - ix - Ili > d(11) > ix - 1/1.

If z E U, I z - xoI < Cd(xo), then

Iz - aI <lz — xol+Ixo — al
< Cd(x o ) +Ix o — al
< Cd(x o ) + Ixo - yf + d(y)
< Cd(xo) + Ix - 11I + d(1!)
< Cd(x 0 ) + d(x0) + 2 d(x0).
That is,
B(x01Cd(xo)) C B(a, (C + 3 )d(xo)).
Therefore, reference to (3.6.7) implies

IR(a, z)I dz < C(d(xo))"+k


B(xo,Cd(xo))
and this, along with (3.6.14) completes the proof.

This proof leads directly to the following which is the Whitney extension
theorem in the context of t k iP(x) spaces.

3.6.3. Theorem. Let A C R" be closed and let U = {x : d(x, A) < 11. If
u E 17(U), 1 < p < oo, and u E tk'F(x) for all x E A with (3.5.3) holding
uniformly on A, then there exists û E C k (U) such that DO (x) = DO 1 3,(x)
for z E A, O<WI<k.

Proof. The proof is essentially the same as the one above with only minor
changes necessary. For example, the polynomials in (3.6.8) and (3.6.9) are
now of degree k and the remainders can be estimated, respectively, by

ISO (x' , x )I < o(Ix - x•Ik-1Ø1)


and
IRa(xi, x i )I < o(Ix. - x l I k-Ial ),
thus allowing (3.6.11) to be replaced by
D(x) - DBP=i (x) = o(ix - xi i &-1141 ).
The remainder of the argument proceeds as before.
142 3. Pointwise Behavior of Sobolev Functions

3.7 An O bservation on Di fferentiation


We address the technicality of showing that IuIITk,v(z) is a measurable
function of x and then establish a result in differentiation theory that will
be needed later in the sequel.

3.7.1. Lemma. Let u E T k .p(x) for all r in a measurable set E. Then,


IIt IITc.o(z) is a measurable function of x.

Proof. Recall that the norm IIuIrTR.a(z) is the sum of the numbers llullp,
ID°PP (x)I, 0 <'^I < k - 1, and the eh root of

r -kp
sup lu(y) — Pz(y)I p dy.
r>Q B(x,r)

Also recall that DaP1 (x) = u Q (x). To show that DaP1 (x) is measurable
in z consider the function cp of Lemma 3.5.6 and define

uc(x) = (Pc * u(x).


If we write u(y) - Pz (y) - R z (y), then

D ° u,(x) = D ° (çof * Pz)(x) + D a (cPE * R2)(x)


= D a PP(x) + f e -( " +Iak) D a SP [J Rx(x - y)dy.

The above integral is dominated by

Ce- ("+Ian) y)Idy < CE-(n+1aj)ek+n


IR=(x -
f(x,c)

= Ce k- r a i -0 0 as e -• 0.

This shows that DaPz (x) is the limit of smooth functions Dau ( (x) for all
x E E, and is therefore measurable. The remainder of the proof is easy to
establish. ❑
3.7.2. Lemma. Let u E LP(R n ), 1 < p < co, be such that for some C,
a > 0 and all r > 0,
1/p
lu(y)Ipdv <- Cr a ,
B(z,r)

for all x in a measurable set E c R". Then, for almost all z E E,


1/p

lu(y)jpdy = o(r°) as r 1 O.
( (r.r)
3.7. An Observation on Differentiation 143

Proof. Without loss of generality we may assume that E is bounded and


that u has compact support. Given e > 0, let A C E be a closed set such
that I — AI < e. Let U be the open set defined by

U ={x:d(x,A)< 1}.

It will suffice to establish the conclusion for almost all z E A.


First, observe that the hypotheses imply that

lien Iu(y)Idy = 0
(x,r)
r~0 B

for z E A and therefore, u = 0 almost everywhere on A.


Let h(x) = ôd(x,A). Recall from Theorem 1.3.1 that there is a count-
able set S C U — A such that {B(s, h(s)) : s E S) is disjointed and

{UB(s, 5h(s)) : s € S} J U - A.

Therefore, since u = 0 almost everywhere on A,

lu(y)1 dy d
dx < ydx I^(y)I
+a
A U ^^ — yln A U —A Ix — yl^ +a
<L lu(y)Idy
E B(s,5h(s)) I X —
sES
yln +a dx

=E dx
my)! f (3.7.1)
f(s• 5h(s3) A Ix — iJl n+a d y'

Let x, E A be such that Is — x,I = d(s, A) = d(s). Hence, B(s,5h(s)) C


B(x,, I s -- x,I + 5h(s)) and Is— x, I = d(s) = 10h(s). By Jensen's inequality
and the hypothesis of the lemma

1 /p l/p

lul < lulp < C Iulp < Ch(s)a.


IB(c5h(a)) (
fB..5h(s)) B(x+,15h(s))
(3.7.2)
Now for z E A,y E B(s,5h(s)), we have

Ix — vI? Ix — sI —
Is — yI>d(s) - 5h(s)= 5 h(s).

Hence, for y E B(s, 5h(s)) we estimate by spherical coordinates with origin


at y,

1. <
dx
C
— yln +a
Ix
°°
i 5^(a)i
< C(a)h(s) — °.
—^—
r'dr
144 3. Pointwise Behavior of Sobolev Functions

This, along with (3.7.2) yields

I11`(v)
yin --tr... < C ()()
Q h ^n
A iBr o,,sh(,)) Ix —
Since {B(8, h(9)) : s E S} is disjointed, it follows from (3.7.1) that

lu(y)I+a dydx < C(a) E h(s)" < co


IA U I z ^11 sES
-

and therefore,
l u(y)I dv < CO
U Ix — yÎ n+a
for almost every x E A. Clearly,

lu(y )Idv
J " -U Ix ^ vl +e+a
< o0

for all x E A, and therefore

lu(y)Idy
Ix — É/1 " +a < o0
R"

for almost all x E A.


An analysis of the argument shows that this was established by using
only the fact that

B(x,r lul
)
< Cr a .

If we apply the above argument with u = IuI', our hypothesis becomes

Iv(y)Idy < Cr"

for all x E E and therefore

Iv(y)Idy
00 >
_ l u(y)I P dy
^ x''L+ap g^
xi n+ap
üq R" I R" I tl r

for almost all x E E. But, for all such x, and for c > 0,

Iu(y) Ipdv < F for all small r > O.


B(s,r) I v — xl"+pa
That is,
)1/p
l u(y) Ipdv < era
B(2,0
for all small r. O
3.8. Rademacher's Theorem in the LP-Context 145

3.8 Rademacher's Theorem in the 17-Context


Recall the fundamental result of Rademacher which states that a Lipschitz
function defined on Rn has a total differential at almost all points (Theorem
2.2.1). We rephrase this result in terms of the present setting by replacing
the hypothesis that u is Lipschitz by u E T k ' p (x) for all r in some set E.
If k = 1 and p = oo, this yields the usual Rademacher hypothesis. The
conclusion we will establish is that u E tkiP(x) for almost all z E E.

3.8.1. Theorem. Let u E T k 'p(r) for all r E E, where E C R" is measur-


able, k a non-negative integer and 1 < p < oc. Then u E tk ,p(x) for almost
ail x E E.

Proof. By Lemma 3.7.1 and Lusin's theorem we may assume that E is


compact and that IIuIjr.,P( s ) < M for all z E E. Since u E T k 'D(x) for
x E E, we may write u(y) = P1 (y) + R z (y) where Pz is a polynomial of
degree less than k and where
lip

IR x (y)1'dy < Mr k , r > 0. (3.8.1)


in(r,r)

From Theorem 3.6.2 it follows that there exists an open set U D E and
riE C k- '' 1 (U) such that

Di(z) = Df3 Pz (x), 0 < 1131 < k. (3.8.2)

Because Ft is of class Ck - 13 it follows from Theorem 2.1.4 that r E W of (Rn )


and therefore we may apply Theorem 3.4.2. Thus, for almost all z E Rn,
there is a polynomial Q S of degree at most k such that i(y) = Q x (y)+R x (y)
where
1/p
iRx (y)1pdy = o(r k ) as r 10. (3.8.3)
(.4.(x,r)
Because ü E C k-I (R n ), the argument following Definition 3.5.5 implies
that
D13 17(x) = Df3 Q 1 (x), O <— IAI < k. (3.8.4)
Therefore, in view of (3.8.1), (3.8.2), and (3.8.3)
I/p

(fB(x,r)
in — i,IP < Cr k

for almost all x E E. Appealing to Lemma 3.7.2 we have


1 /p
lu —ûI 13 = o(rk) as r1O
(fB(r,r)
146 3. Pointwise Behavior of Sobolev Functions

for almost all x E E. Consequently, for all such z,

1/p l/p
^u(y) — = (y)I p d Iu(Y) — U(y)I p du
B(x,r) B(x,r)
/P

+ I u(y) z(âl)I P dY
(41(x,r)

< o(r k ) as r —. 0,

thus establishing the result. ❑

3.9 The Implications of Pointwise Differentiability


We have seen in Section 4 of this chapter that Sobolev functions possess
LP-derivatives almost everywhere. This runs parallel to the classical result
that an absolutely continuous function f on the real line is differentiable
almost everywhere. Of course, the converse is false. However, if it is assumed
that f' exists everywhere and that 11 1 is integrable, them f is absolutely
,

continuous (Exercise 3.16). It is natural, therefore, to inquire whether this


result has a counterpart in the multivariate LP theory. It will be shown
that this question has an affirmative answer. Indeed, we will establish that
if a function has an LP derivative everywhere except for a small exceptional
set, and if the coefficients of the associated Taylor polynomial are in 17,
then the function is in a Sobolev space.
We begin the investigation by asking the following question. Suppose
u E L P (R") has LP-derivatives at x E R"; that is, suppose u E tk'P(x)
where k is a positive integer. Then, is it possible to relate the distributional
derivatives of u (which always exist) to the LP-derivatives of u? The first
step in this direction is given by the following lemma. First, recall that
u E tk , P(x) if there is a polynomial Pi of degree k such that

1/p

(JiB (r,r) lu(y) — Pz(v)I P dy = o(r k ) as r 0, (3.9.1)

and u E TkiP(x) if there is a polynomial P, of degree less than k and a


number M > 0 such that
/p

(1U(Y) — Px(y)j p dy < Mr k , 0 < r < oc.


f (r,r)

3.9.1. Lemma. Suppose u E L7 (R"), p > 1.


3.9. The Implications of Pointwise Differentiability 147

(i) If u E T k .P(x), then

lim i^nf Sp i * D°'u(y) > — oo,

with !x -- yi < t, and where Dau denotes the distributional derivative


of u,0<fal <k;

(ii) If u E t k 'P(x), then

lim sup (p i * D'u(y) = D a Ps (x),

with Îx — yi<t, 0 <lai < k.


The function (p i above is a mollifier as described in Section 1.6. Since
(Pt E Co (R"), its convolution with a distribution T is again a smooth
function. Moreover, for small t and lv —xi < t, the quantity (p i *T(y) gives
an approximate description of the behavior of T in a neighborhood of x.
Indeed, if T is a function, then

lim sup Sp i * T(y) = T(x)


t^o
1y—x1<t

whenever x is a Lebesgue point for T. This will be established in the proof of


Lemma 3.9.3. Very roughly then, the statement in (ii) of the above lemma
states that, on the average, the behavior of the distribution Dau near x
is reflected in the value of the coefficient, D°Px (x), of the Taylor series
expansion.
Let F(y, t) = cp t s u(y). F is thus a function defined on a subset of R"+' ,
namely R" x (0, oo) and is smooth in y. The lower and upper limits stated
in (i) and (ii) above can be interpreted as non-tangential approach in R"+ 1
of (y, t) to the point (s, 0) whch is located on the hyperplane t = 0.

Proof of Lemma 3.9.1. Proof of (ii). Let

u(y) = P7(y) + R (Y) and Ft (V) = F(y, t)•


Then
D°Ft(y) = b a Nvt * u}(y) = D °'cpt * u(y).
Therefore

DaFi (x + h) = f DacPt(x + h — y)u(y)dy

= f D a Spt (x + h — y)Ps (y)dy + f D a cpt (x + h — 3/)Rx (y)d1!

^
(wt * Ps)(x + h) + D°cp t (x + h — y)Rs (y)dy. (3.9.2)
148 3. Pointwise Behavior of Sobolev Functions

There is a constant C = C(l DSpl) such that

I D`k 't(x + h - y)I < Cry"

for IaI = k. Consequently, for h E B(0, t), it follows that

If D ° Wt(x + h - y)Rx(y)dy < Ct-n-k lRx(Y)Idy


f(x+h,t)
k
< Ct' IR1(y)Idy --. 0 as t j 0,
a (z,si)

by (3.9.1). Writing Pi in terms of its Taylor series, we have

Px(y) D^ Px(x)(y - x)a


1°1=a

and therefore D'P1 (y) = D' P1 (x) for all y E R if (al = k. Hence, "

cp e * Da 13,(x+ h) = D°P2 (x), and reference to (3.9.2) yields

lim sup çp t * Dau(x + h) = P1 (x), 0 < Ihl < t, (3.9.3)

thus establishing (ii) if Pal = k. However, if 0 < l < k, then u E t t .P(x) and
the associated polynomial is

D°Px (x)(h -- x)a


E
faI o
-
a!

Thus, applying (3.9.3) to this case leads to the proof of (ii).


The proof of (i) is similar and perhaps simpler. The only difference is
that because Pz is of degree at most k -1, we have

D a F: (x + h) = 0 + f D ° vt(h - EI)Rz(y)dy

if Ial = k. The integral is estimated as before and its absolute value is seen
to be bounded for all t > 0, thus establishing (i). ❑

The next two lemmas, along with the preceding one, will lead to the
main result, Theorem 3.9.4.

3.9.2. Lemma. Let T be a distribution and suppose for all x in an open


set S2 C R" that

lim inf cp e * T(y) > - oo, Ix -


yl <_t,
t -•0
tES
3.9. The Implications of Pointwise Differentiability 149

where Sc (0, oo) is a countable set having 0 as its only limit point. Let C
be a closed set such that C n St O. Then there exist N > 0 and an open
o
set I i C Si with C fl S21 0 such that SPt * T(1i) > —N > —oo whenever
Iv xl Çt,xECf1Stl,tE S.

Proof. Let
F.(x) = inf (SPt * T(y) :ix — < t, t E S}.
Then, F.(x) > —oo for x E Ii since 0 is the only limit point of S and it is
easy to verify that F. is upper semicontinuous. Thus, the sets

Cn ftn ix. F.(x)> —i ), i = 1,2, ...,

are dosed relative to C n ft and their union is C n Cl. Since C n fI is of the


second category in itself, the Baire Category theorem implies that one of
these sets has a non-empty interior relative to C n ft. ❑

One of the fundamental results in distribution theory is that a non-


negative distribution is a measure. The following lemma provides a gener-
alization of this fact.

3.9.3. Lemma. Let b > 0, N > 0, and suppose S is as in Lemma 3.9.2. If


T is a distribution in an open set fZ such that

(Pt *T(x)>—N>—oo for x ESl , t ESfl (OA


and
lim sup Sp i * T(x) > 0 for almost all xo Eft,
elo
^x—xo I<t

then T 18 a non-negative measure in ft.

Proof. Let E O(1l), tsb > 0, and recall from Section 1.7, that the convo-
lution Spt * T is a smooth function defined by

SP t * T x = T(T:43 t)
( )

where tPe(i1) = too t (—y) and TxSFe(y) = Ot(y — x). Then,


TOP *Ot) =T *(i!1 *Vt)( 0 )
_ (T *SPt)*'1;( 0 )
= f T * SPe (`g1)til(t!)dl!

= f T * tPt(11)0(âi)dg.
150 3. Pointwise Behavior of Sobolev Functions

Now V)* CPt —, t/) in g(S1) as t —> 0+. Moreover, since 7/) is non-negative
and coot * T(x) _
> —N for x E SZ and t E S n (0, b), it follows with the help
of Fatou's lemma, that

T(0) = elô T(0** (430


tES
> lim inf T * SPt (l0(11)dEl
to R
tES

> lim ^rtf T * ePt(Y).0(3I)dY


n
tES

— N(y)dy•
Thus, the distribution T + N has the property that
(T+N)(7i)> 0 for E 2(f1), ip > 0.

That is, T + N is a non-negative measure on IZ, call it A. Let u = v + or


where v is absolutely continuous with respect to Lebesgue measure and or
is singular. Clearly, SZ is the union of a countable number of sets of finite
v measure. Thus, by the Radon-Nikodym theorem, there exists f E L'(S2)
such that
v(E) = f (x)dx
E
for every measurable set E C O. Since T + N = p, it follows that

çot * T (x) + N = çot * (T + N) (x) = Soot * µ(x)


(3.9.4)
n ipt (x - Of (WY + n ço (x — 11)da(y),

for x E ft Because or is a singular measure, a result from classical differen-


tiation theory states that

Urn a[$(xo, t)] _


IB(xo,t)I °
for almost all x o E n, cf. [SA, Lemma 7.1]. Therefore, at all such r0 with
Ix — xol <t,

f vt(ac — y)da(y) Soe(x — 1/)da(y)


f(z,t)

< I ^Pt(x — g1)da(y)


?
a[(B(xo, 2 t)]
—< ChPIIoo IB(xo, t)I
—• 0 as t —, 0* with Ix — xoI < t.
3.9. The Implications of Pointwise Differentiability 151

To treat the other term in (3.9.4), recall that f has a Lebesgue point at
almost all x o € D. That is,

If (I!) — f(xo)1dy —0 as r -+ 0+
B(zo.r)

Therefore,

B(x.t) f (Owe (x — y)dil — f (xo) _ f if (y) — f (xo)1(Pt(x — y)dv


(x,t)

^ Cq P1100
( 11(v) — f (2o)1dy 0 as t -• 0, - ro i Ç t.
B(z o ,2t)
Consequently,
N < lim sup (p t * T(r) + N = f (xo)
tlo
Ix-xoI<t
for almost all r o E Sl. This implies that

w(E) > NIEI

for all measurable E C Sl. Since p(E) > v(E) it follows that the measure
p - N = T is non-negative. ❑

3.9.4. Theorem. Lei T be a distribution in an open set Sl C R" and let


f E LL (n). Assume

lim sup SPt * T(y) > f(x), - S t,


do

for almost all x E 1, and

lir lô f too t * T(y) > -oo, lx — y j < t ,

tES

for all x E OE Then T - f is a non-negative measure in Sl.

Proof. We first assume that f = 0. Lemma 3.9.2 implies that every open
subset of contains an open subset Sl' such that for some N > 0, cp t *T (x) >
-N for z€ T,tE S. Lemma3.9.3impliesthat Tis ameasurein11'.
Let (I I be the union of all open sets Si' c SZ such that T is a non-negative
measure on I". From Remark 1.7.2 we know that T is a measure in D I .
We wish to show that 121 = Si. Suppose not. Applying Lemma 3.9.2 with
C = R" — il l , there is an open set 12' C Sl such that w e * T(x) > -N
for y E C fl S2', Ix - vi < t, and t E S. Let Sl 2 = it i U Si' and note that
0 2 — f2 1 = C fl Sl'. Let
fl 3 = SZ x n {x : d(x, R" -- n2) > e}
152 3. Pointwise Behavior of Sobolev Functions

for some e > O. Take e sufficiently small so that 11 3 n (Rn — 121) A 0.


Consider Sp i * T(x) for x E S23 and t < e. Now T is a non-negative measure
in Of. Therefore, if d(x, Rn — no > t, SPt *2-(x) > O. On the other hand,
if d(x, Rn — 11 1 ) < t, there exists y E R" — [2 2 such that Ix — yl < t.
Since B(x, t) C SZ 2 , it follows that y E S22 — SZ1 = C n ST. Consequently,
cp t *T (x) > —N. Hence, Spt * T(x) is bounded below for x E Q3 r t E Sn(0, e),
and thus T is a measure in SZ 3 by Lemma 3.9.3. But S23 n (Rn — Sl i ) 0 0
thus contradicting the definition of O 1 .
For the case f 96 0, for each N > O define

N, f (x) ? N
fw(x)= f(x), — N<f(x)<N
—N, f (x) < —N
and let R be the distribution defined by R = T — fN. Clearly R satisfies
the same conditions as did T when f was assumed to be identically zero.
Therefore, R is a non-negative measure in S2. Thus, for E Cr 1l), lb >15,

R(0) = T (b) — f both dx > O.

Letting N --' oo, we have that

T(?i) — f f idx ^ O.

That is, T — f is a non-negative measure in 0. D

Now that Theorem 3.9.4 is established, we are in a position to consider


the implications of a function u with the property that u E Tk p(x) for
,

every x E 0, where S2 is an open subset of R". From Theorem 3.8.1 we


have that u E t k .P(x) for almost all x E S2. Moreover, in view of Lemma
3.9.1 (ii), it follows that whenever u E tk.P(x),

Urn sup * Dau(y) = D °'Px (x), Ix — yl < t,


tlo

for 0 < IQl < k. For convenience of notation, let u ,(z) = DaPP (x), and
assume u Q E (11). Then Theorem 3.9.4 implies that the distribution
Dan — u. a is a non-negative measure. Similar reasoning applied to the func-
tion —u implies that Da (—u) — (—u a ) is a non-negative measure or equiv-
alently, that Dau — u a is a non-positive measure. Thus, we conclude that
Dau = u a, almost everywhere in OE That is, the distributional derivatives
of u are functions in L"(SZ). In summary, we have the following result.

3.9.5. Theorem. Let 1 < p < oo and let k be a non-negative integer. If


u E T k .P(x) for every x E S2 and the LP-derivatives, u a , belong to L"(11),
3.10. A Lusin-Type Approximation for Sobolev Functions 153

0 < f cx} < k, then u E Wk p(SZ).


,

Clearly, the hypothesis that the LP-derivatives belong to LP(0) is neces-


sary. On the other hand, we will be able to strengthen the result slightly
by not requiring that u E T k IP(x) for all x E Q. The following allows an
exceptional set.

3.9.6. Corollary. Let K C R" be compact and let SZ = R" — K. Suppose


1-I n-1 [7r,(K)) = 0 where the 7r, : Rn , i = 1, 2, ... , n, a re n inde-
pendent orthogonal projections. Assume u E Tk P(x) for all s E fZ and that
u s E IP(ft), 0 < lal < k. Then u E W P(R").

Proof. Assume initially that the projections 7r ; are given by

7ri(x) = (x1, • • • , x2, ... , x n )

where (ac t , ... , x„ ... , x n ) denotes the (n— 1)-tuple with the z,-component
dcleted. Theorem 3.9.5 implies that u E Wk1e(11). In view of the assumption
on K, reference to Theorem 2.1.4 shows that u E W L .P(R ") since u has a
representative that is absolutely continuous on almost all lines parallel to
the coordinate axes. Now consider D'u, oaf = 1. Since Daft E Wk-1,p0 1) -

a similar argument shows that Dail E W"P(R") and therefore that u E


W 2 P(R"). Proceeding inductively, we have that u E W k ie(R1 ).
Recall from Theorem 2.2.2 that u E W k iP(R") remains in the space
Wk•P(R") when subjected to a linear, non-singular change of coordinates.
Thus, the initial restriction on the projections in is not necessary and the
proof is complete. D

In the special case of k = 1, it is possible to obtain a similar result


that does not require the exceptional set K to be compact. We state the
following [BAZ, Theorem 4.51, without proof.

3.9.7. Theorem. Let K C R" be a Borel set and suppose H" - 1 [7r; (K)] = 0
where the r, : R" Rn -1 i = 1, 2, . _ . , n, arc n independent orthogonal
,

projections. Let f2 = R" -- K and assume u E L o c (fl) has the property that
its partial derivatives exist at each point of iZ and that they arc in L ôc( 1 ).
Then u E Wloc (R") .

3.10 A Lusin-Type Approximation for Sobolev


Functions
Lusin's Theorem states that a measurable function on a compact inter-
val agrees with a continuous function except perhaps for a closed set of
arbitrarily small measure. By analogy, it seems plausible that a Sobolev
154 3. Pointwise Behavior of Sobolev Functions

function u E Wk,P(11) should agree with a function of class C k (1/) except


for a set of small measure. Moreover, if the requirement concerning the
degree of smoothness is lessened, perhaps it could be expected that there
is a larger set on which there is agreement. That is, one could hope that u
agrees with a function of class Cl (n), 0 < l < k, except for a set of small
Bk_r, p -capacity. Finally, because Sobolev functions can be approximated
in norm by functions of class C k (11), it is also plausible that the Lusin-type
approximant could be chosen arbitrarily close to u in norm. The purpose
of this and the next section is to show that all of this is possible.
In this section, we begin by showing that if u E W"(R"), then u agrees
with a function, y, of class C t on the complement of an open set of ar-
bitrarily small Bk_,, p -capacity. In the next section, it will also be shown
that Flu -- vH,, p can be made small. The outline of the proof of the exis-
tence of y is as follows. If u E W k .p(R n ) and 0 < t < k, then Theorem
3.4.2 implies that u E t e 'p(x) for all x except for a set of Bk_t, p -capacity O.
This means that the remainder terms tends to 0 (with appropriate speed)
at Bk_t, p -q.e. x E R. We have already established that if a function u
has an 11-derivative of order t at all points of a closed set A (that is, if
u E tt 'p(x) for each x E A) and if the remainder term tends to 0 in L" uni-
formly on A, then there exists a function v E C t (Rn) which agrees with u
on A (Theorem 3.6.3). Thus, to establish our result, we need to strengthen
Theorem 3.4.2 by showing that the remainder tends uniformly to 0 on the
complement of sets of arbitrarily small capacity. This will be accomplished
in Theorem 3.10.4 below.
In the following, we will adopt the notation

Mp,Ru(x) = sup iu(y)I dy


0<r<R D(z,r)

whenev e r u E I.p(Rn ), 1 < p < oo, and 0< R< co.

3.10.1. Theorem. If 1 < p < oo and k is a non-negative integer such that


kp < n, then there is a constant C = C(k, p, n) such that

Bk,p[ {x : Mp,Ru(x) > t }] <_ t Ilullk,p (3.10.1)

whenever u E Wkip(Rn) and R < 1.

Proof. We use Theorem 2.6.1 to represent u as u = gk * f where f E LP(Rn


and Iluiik, p N VG- Thus, it is sufficient to establish (3.10.1) with iluf lk,P
replaced by jl f ji p . Since lug < g k * If i, we may assume f > 0. Let

Et = f x : Mp ,Ru(x) > t}
and choose r E Et . For notational convenience, we will assume that x = 0
3.10. A Lusin-Type Approximation for Sobolev Functions 155

and denote B(0, r) B(r). Thus, there exists 0 < r < R < 1 such that

Iu(y)Ipdy > tF
B(r)
or
P
gk(y — w)f (w)dw dy > tP.
1,30.) R^
_
Utilizing the simple inequality (a + b)P < 2P -1 (aP + b") whenever a, b > 0,
it therefore follows that either
P
gk(y — w) f (w)dw dy > 2 1 `P9 (3.10.2)
B(r) (j wI< 2 r

or
F
gk(y — w) f (w)dw dy > (3.10.3)
B(r) (
iWI>2T

If y E B(r), then from Lemma 2.8.3(i) and the fact that gk < CIk, (2.6.3),
we obtain

f w^<2r
9k(y — w)f (w)dw < C —
f (w)
I
I y wI
-k dw

< Cr k Mf (y),
where C = C(k, n). T hu.g, in case (3.10.2) holds, we have

tP Ç Cr k P M f (y)Pdy (3.10.4)
B(r)

where C = C(k, p, n).


We will now establish the estimate

g k (y — w) f (w)dw < C inf gk(y — w) f (w)dw (3.10.5)


fwI>2r yEB(r) 1 wf>2r

for all y E B(r). Recall that r < 1. Now if y and w are such that I yI < r <
2r<IwI<2,we have

3 IwI ? Iw! + IyI ? Iw — yI ? IwI — IyI ? Iw! — IwI


2
Consequently, if y 1 and y2 are any two points of B(r), refer to (2.6.3) and
the inequality preceding it to conclude that for some constant C = C(k, n)
C C
9k(w — vi)
lw — vt ln^ ICI"-k
k

C <
Cg k (w y2)• (3.10.6)
Iw -` 1/21"
156 3. Pointwise Behavior of Sobolev Functions

If }w` > 2 and y E B(r), then IwS > 1w1 + 1 ? 1w — yi > Iwl EvI

Iwl — 1 > (w1/2. Therefore, in this case we also have

gk(w - yi) <_ Cgk(w — g2)• (3.10.7)

Our desired estimate (3.10.5) follows from (3.10.6) and (3.10.7). Thus, in
case (3.10.3) holds, there is a constant C _ C(k, p, n) such that
P
t p < C inf gk(w —y)f (w)dw
y€ 13 (r) I wl>2 r
<C inf (9k * f(v)) p .

To summarize the results of our efforts thus far, for each z E B = there
exists 0 < r < 1 such that either

tp < Cr kp1^f f (y)pdy (3.10.8)


B (r,r)
or
t < C inf gk * f (y). (3.10.9)
yE B (x,r)

Let G i be the family of all closed balls for which (3.10.8) holds. By Theorem
1.3.1, there exists a disjoint subfamily Y such that

Bk, p ({UB : B E g1}] < Bk, p !{UB : B E f]

Ç E Bk,p(B)
BEf
5. E (50 n-kp (by Theorem 2.6.13)
B(x,r)EF

<^ Mf (y) p dY
EB
BEf

< li f i^ p (by Theorem 2.8.2). (3.10.10)


PI
Let ÿ2 be the family of closed balls for which (3.10.9) holds, then the def-
inition of Bessel capacity implies that Bk, p [{ UB : B E Gill < (C/tp)II f lif.
Thus

Bk ,p lEej < Bk, p [{UB : B E gill + Bk, p [{UB : B E gall < -tF Ilf lip ,

which establishes our result. ❑

We now have the necessary information to prove that integral averages


of Sobolev functions can be made uniformly small on the complement of
3.10. A Lusin-Type Approximation for Sobolev Functions 157

sets of small capacity. This result provides an alternate proof of Theorem


3.3.3, as promised in the introduction to Section 1 of this chapter.

3.10.2. Theorem. Let 1 < p < oo and k be a non-negative integer such


that kp Ç n. If u E W k, P(R"), then for every E > 0 there exists an open
set U C R" with B k , p (U) < e such that

lu(v) -u(x)(Pdy -. 0
B(x,r)

uniformly on Rf2 - U as r J. 0.

Proof. With the result of Theorem 3.3.3 in mind, we define

Aru(x) = Iu(v) - u(x)IPdy


13(z ,r)

for x E R" and r > 0. Select i< such that 0 < i= < 1. Since u E W k.P (Rn ),
there exists g E Co (Rn) such that

IIt - gIIi,p < EP+ '/ 2 .


Set h = u = g. Then

Aru(x) 2p -1 [Arg(x) + Arh(x)}+

A r h(x) < 2P-1 Ih(g1)I p dg + ih(x)I p


B (z,r)

and therefore,

Aru(x) S C Arg(x) + Ih(y)Ipdy + Ih(x)I P


B(x,r)

where C = C(p). Consequently, for each x E R",

sup A r u(x) < C sup Arg(x) + Mp,RIhl(x) + Jh(x)I p •


0<r<R 0<r<R

Since g has compact support, it is uniformly continuous on R" and therefore


there exists 0 < R < 1 such that

sup CA r g(x) <T.


0<r<R

whenever x E R ". Therefore,

x : sup A,.u(x) > 3T C {x : CMp,RI hI (x) > T} U {x : CIh(x)IP > T}


0 <r <R
3. Pointwise Behavior of Sobolev Functions

158

C {x : CMp ,RI hI (x) > is } U {x : (C7 -1 ) 1 IFIh(x)I > 11.


Since h E W k .'(R"), by Theorem 2.6.1 we can write h = yk * f, where
liflip i1hllk.p. Now

{x : (Ce E)"Ih(x)I > 1) c {x :


- ( -1)' /p 9k *111(x) > 1)

+^ _
and therefore, by the preceding theorem and the definition of capacity, we
obtain a constant C = C(k, p, n) such that

Bk,p({x : sup A r u(x) > 3F- }] <


0<r<R
C[E 9hFIt,,
- ^Ihllk,p]
Fptl
< C2E_F
2
< Ce.

For each positive integer i and e as in the statement of the theorem, let
F,=C -l et - ' to obtain 0<R; <1 such that

Bk ,n [f. : sup A r u(x) > 3T<


0<r<R,
< e2 - `.

Let
00

U = ^ x : sup A r z1(2) > 3 €;


i=1 0<r<Rs

to establish the conclusion of the theorem. ❑

3.10.3. Remark. If we are willing to accept a slightly weaker conclusion in


Theorem 3.10.2, the proof becomes less complicated. That is, if we require
only that

B(x,r)

Iu(y) - U(X)14 0

uniformly on Rn - U as r j 0, rather than

lu(y) - u(x)Ipoli -'


0,
9(x,r)

then an inspection of the proof reveals that it is only necessary to show

Bk, F f{x ; Mu(x) > t}] < IIuIIk,p'


To prove this, let u = g k * f, where II flip ^' 11e4k, p and define

if 1271 < T
r r (s) =
an otherwise.
3.11. The Main Approximation 159

Then

lu(y)Idy = r r * IuI(x)

< r,* (gk * If 1)(x)


< gk * Ml f I (x),
which implies Mu < g k * MI f I. From the definition of capacity,

Bk.p[{x : Mu(x) ? t}] Bk.pRx : gk * Mill(x) ? t})


• t °IIM III IIp
-

• Ct p lI f g, by Theorem 2.8.2,
• Ct -p llulik,p•
As an immediate consequence of Theorem 3.10.2 and the proof of The-
orem 3.4.2, we obtain the following theorem which states that Sobolev
functions are uniformly differentiable on the complement of sets of small
capacity.

3.10.4. Theorem, Let & k be non-negative integers such that l < k and
(k - t)p < n. Let u E W k .n(Rn). Then, for each e > 0, there exists an open
set U with Bk_cp(U) < e such that
I/p
r -t [

lu(g) — Pr ll (b ) Indg -- 0

uniformlyon Rn - U as r10.

Finally, as a direct consequence of Theorems 3.10.4 and 3.6.3, we have


the following.

3.10.5. Theorem. Let t, k be Taon- negative integers such that l < k and
(k - e)p < n. Let u E W k Ip(R n ) and e > 0_ Then there exists an open set
U C R" and a C I function y on Rn, such that
Bk_ ,,
(U) < e
an d
D"v(x) = D "u(x)
for all x E Rn -U and 0< IcYI < B.

3.11 The Main Approximation


We conclude the approximation procedure by proving that the smooth
function y obtained in the previous theorem can be modified so as to be
close to u in norm.
160 3. Pointwise Behavior of Sobolev Functions

In addition to some preliminary lemmas, we will need the following ver-


sion of the Poincaré inequality which will be proved in Theorem 4.5.1.

3.11.1. Theorem. Let a E (0,1), t a positive integer, and 1 < p < oo.
Then there exists a constant C = C(o, 1, p, n) such that for every non-empty
bounded convex subset II of Rn with diameter p and every u E W 1, p(fÛ)
for which
^SZ fI {x : u(x) = 0) 1> Q^SZ^,
we have the inequality

I Iu(x)1pdx < Cpt E


I0E=t
n
^ku(zedx.
^

3.11.2. Lemma. Let t be a positive integer and let u be a function W t "p(R'n)


which vanishes outside a bounded open set U. Let 6,a E (0, 1) and let

inf IK(xlt) n (Rn — U) >


E = ôU f1 x : (3.11.1)
a<t<6 in

where K(x, t) denotes the closed cube with center x and side-length t. Let
in be a positive integer such that m < t and let e > O. Then there ezrsts a
function v E W m'p (Rn) and an open set V such that

( 1 ) Ilu — VJ^m,p < ei

(ii) E C V and v(x)= 0 whenxE VU (R"--U).

Proof. For A E (0,1), let Ka denote the set of all closed cubes of the form

((i1 — 1 )A, i1A] x ((i2 — 1)A, i2AJ x ... x Rin — 1)À, i n A]


where i1 i i 2 , ... , i n are arbitrary integers. Let A < 36 and let

K1, K2,..., Kr

be those cubes of 1Ca that intersect E. Let a i be the center of K; and let

pi K(ai, 4A).
Let ( be a C°'O function on Rn, such that 0 < < 1, ((z) = 0, when
x E K(0,1) and ((x) = t when z ft K(0,3/2). Define

VA(X) = n(x) ^ ^ (
z _ ai )

(3.11.2)
t =1
Za

for x E R n . Clearly va(x) = 0 when d(x,E) < IA, so that, for any A, we
can define y by y = va and find an open set V satisfying (ii).
3.11. The Main Approximation 161

We keep i fixed for the moment and estimate

flu - UaIlt,p;P, • (3.11.3)

We observe that there exists a constant r, depending only on n and such


that at most T of the cubes Pi intersect Pt (including Pi ). Denote these by

Pill P12 ,...,Pi.

where s < T . Then, for x E P,

vA(x) = u(x)w(x), (3.11.4)

where
a
W ( x) = j^ ^[( z - Qik)/24 (3.1 1.5)
k=1
Now, for r E P, and any multi-index a with 0 <10/1 < t, we have

I D' w(x)I < A1A -1°1 ,

where A l depends only on t and n. Hence for almost all x E P; and any
multi-index 7 with 0 < IryI < t, we have

IYI
ID 7 Va( 1 )1 A2 E L I D13u(x ,, ) (3.11.6)
r= o 1A1 =r
where A2 depends only on t and n.
Let y be a point where K1 intersects E. Clearly, there is a subcube Q; of
P; with center y and edge length 3A. By (3.11.1), u and hence its derivatives
are zero on a subset Z of Q. with

IZI > a(3A)"• , (3.11.7)

By applying the Poincaré inequality to the interior of the convex set P, we


obtain, when },Q} < t,

fpi llY3 u(x)rdx < A3ap(1-1131) E 1D ( u(x)Iedx (3.11.8)


P^
It1=1

where A3 = A3(t, Q, p, n). But, with a suitable constant A3, (3.11.8) will
still hold when 1/3 1 = 1. By (3.11.6) and (3.11.8) (since A < 1)

ID 7 v x (x)I"dx < A4 IDcu(x)IFdr (3.11.9)


Pf P i
ItI=1
162 3. Pointwise Behavior of Sobolev Functions

for 0 < < t, where A4 = A4(1, p, a n). Let ,

XA= U Pi'
s=1

Then r
ID'rva(x )Ipd^ < A4 E ^ ID^u(^)Ipd;^
X,,
ItI=t z =1 P ;

for 0 <1/I < t. But each point of X), belongs to at most r of the cubes P;,
hence

L. 1D 7 ua(4 p dz < rA4 E


k^k=^
xa
ID4u(x)Iedz, (3.11.10)

for 0 <171 < L. Now

tlu VAIIt,p 2" E I D7va(r)Ipdi + X ID 7 u(:^)Ied.r ,


0[1ryl<t X a a

so that by (3.11.10)

IIu — Va 114 < AS E ID u(x) Ipdz


. (3.1 ] .11)
< <1 LAU
0 171

where A5 - A5 (l, p, a, n). But

X),nUcUr1 {x: d(z,ÔU) < 2/a}.

Hence I(X , fl U)I —► 0 as A j. O. Therefore by (3.11.11)


)

IIu — vallt,p -, 0
asA--►0+.
The required function y is now obtained by putting y = Va, with suffi-
ciently small A. ❑

3.11.3. Lemma. Let 0 < A < n. Then there exists a constant C = C(A n) ,

such that
Ix — yl"dx < Cty — zl a ^n (3.11.12)
B(z,5)
for all y,z E R' and all 6> 0.

Proof. We first show that there exists a constant C, such that (3.11.12)
holds when y _ 0 and z is arbitrary.
3.11. The Main Approximation 163

When Izl > 36, we have

Izl <_ Ix! + Iz -xI <1x1+ Ix! + 314

so that Iz! < 2141, IxI A-n < Alzla^" and (3.11.12) holds.
,

When IzI < 36,

A-n
ë -n IxI dx 5_ &-n IxlA-ndx = Cb a-n
B(:,6) 13(0,46)

Hence, it is clear that (3.1112) holds with y = O.

Ô-n f
Since we have shown that

IxII-ndx < CIzIa -n


(s,6)

for all z E R n , the general result follows by a change of variables; that is,
replace z by z - y. ❑

Throughout the remainder of this section, it will be more convenient


to employ the Riesz capacity, Rk, p , rather than the Bessel capacity, Bk.p.
This will have no significant effect on the main result, Theorem 3.11.6. See
Remark 3.11.7.

3.11.4. Lemma. Let k be a non-negative real number such that kp < n.


Let U be a bounded non-empty open subset of Rn and F a subset of ôU
with the property that for each x E F, there is a t E (0,1) for which
IU nB(z,t)I >
(3.11.13)
I$(x,t)I -
where a E (0, 1). Then there exists a constant C C(n, p, k) such that

Rk, p (U U F) < Ccr - pRk, p (U). (3.11.14)

Proof. Let a, U, and F be as described above. The cases k = 0 and k > 0


are treated separately.
( i) We consider first the case where k > 0. Let be a non-negative
function iii LP(Rn) with the property that

Ix - yl k _n i(y)d1/ > 1 (3.11.15)


I(k) f R
for all x E U. Let C1 be the constant of Lemma 3.11.3. It can be assumed
that C 1 > 1. Consider a point b E F and let t be such that (3.11.13) holds
for z = b. By Lemma 3.11.3,
Ci ly - blk-n >,
Ix - yl k-n dx
B(b,i)

164 3. Pointwise Behavior of Sobolev Functions

so that

c:Iy - bI k-n ,(y)dy


L " L (b t) [ Le. Ix -- ylk-n^(y)dy dz

and by (3.11.15),
> -y(k)t —n IU fl B(b, t)I.
Hence by (3.11.13),

rCa
Iy - bl k-n ^(^)db ? a.
7(k)JRh1+

Put r? = C2(7 -1 0. Then


1
7(k) R" Iy — xIk-ni(y)dy ? 1

for all x E F, and therefore,


(C.2 ) I'
Rk,(F) < 11711; = iS IIp.

Thus
(C2)P

Rk,p(F) < Rk ,p (U)•
a
The required inequality now follows.
(ii) Now let k = 0, so that Rk, p becomes Lebesgue measure. Let 8 be
the collection of all closed balls B with center in F and radius between O
and 1 such that
lung > a. (3.11.16)
II
Hence, by Theorem 1.3.1, there exists sequence {Br }, B r E B, such that
B r fl B. = 4 when r s and
oc,
FC V Br.
r=t
Thus
ill<E or' = 5n EIBrI
r -1 r=1
and by (3.11.16)
00

< 5nQf1 E IU fl Br I < 5nalUI.


r=1
Since a < 1, the required inequality follows. 0
3.11. The Main Approximation 165

3.11.5. Lemrna. Let p > 1, k a non-negative real number such that kp < n
and I a positive integer. There exists a constant C = C(n, p, k, l) such that
for each bounded non-empty open subset U of Rn , each u E W 1 'p(R")
which vanishes outside U and every e > 0 there exists a C°° function v on
R n with the properties
(i) I}u — viit,p < c ,

( 1i ) R k,p( S p t y) < CRk,p(U) and

(iii) spty C V = R n n{x: d(x,U) <e}.

Proof. Let U, u, and e be as described above. Since U # 0, it follows that


Rk, p (U) > O. Let
IB(x , t) U}] > 1
=âU n x: inf (3.11.17)

E
o <t<1/2 in 2
-

Then E is closed. By Lemma 3.11.2 there exists a function vo E W t,p(Rn )


and an open set Vo such that

IIu - v ollt, p < 2^, (3.11.18)

ECVo and v o (x)=0 when x EVo U (R"—U). Set

F= aU - E.
Then, for each x E F there exists t E (0,1/2] such that

IU nB(x, t)I (3.11.19)


IB(x, t)I — cr
where a = 1 -- 1/(2a(n)). Let C, be the constant appearing in Lemma
3.11.4. Then
Rk,p(U U F) < 2CR
k p(U), (3.11.20)
,

where C = 2Cla p. Let


B = R " n {x:v o (x)# 0}.


Then B C U U F and hence Rk, p (B) < i CRk, p (U), so that there exists an
open set W with B C W and

Rk, p (W) < CRk, p (U).

By applying a suitable mollifier to yp we can obtain a C°° function v with


sptvCVnWand
1
Ilvo — vllt,p < 2e- (3.11.21)
166 3. Pointwise Behavior of Sobolev Functions

It follows from (3.11.18) and (3.11.21) that y has the required proper-
ties. D

We are now in a position to prove the main theorem.

3.11.6. Theorem. Let t, m be positive integers with m < t, (t - m)p < n


and let Si be a non-empty open subset of R". Then, for u E We (n) and
each e > 0, there exists a Cm function v on Si such that if

F = SZ n {x : u(x)o v(x)},

then
Ri-,,,p(F) < e and Ilu --- vll,n,p < e.

Proof. It can be assumed that the set A = SZn{x : u(x) # 0} is not empty.
Initially, it will be assumed that SZ = Rn and A bounded. We will show
that there exists a Cm function y on R" satisfying the conclusion of the
theorem and that spt v is contained in the set V = Rn n {x : d(x., A) < e}.
Let C be the constant of Lemma 3.11.5. Let u be defined by its values
at Lebesgue points everywhere on 0 except for a set E with Bt, p (E) =
R,, p (E) = O. By Theorem 3.10.5 there exists an open set U of Rn and a
Cm function h on R", such that U D E,
F
Rt-m,p(U) < 1 + C (3.11.22)

and
h(x) = u(x)
for all x E Rn — U. We may assume that spt h C V and U C V. By
substituting t - m for k and u - h for u in Lemma 3.11.5, we obtain a C°°
function cP on R" such that

Ilu- h -wllm,p< C, (3.11.23)

R t — m , p( s P t SP) _< CRt-m,p(U), (3.11.24)


and
spt cp C V. (3.11.25)
Put y = h +cp. Then the second part of the theorem follows from (3.11.23).
Clearly,

F c R" n j {x : h(x) 0 u(x) } u spt (pl c U u spt cp, (3.11.26)

so that by (3.11.24)

Rt-m,p(F) ç (1+ C)Rt-m,p(U).


3.11. The Main Approximation 167

Thus, the first part of the conclusion follows from (3.11.22). Since spt h and
spt cP are both contained in V, it follows that spt v c V.
We now consider the genera] case when SZ is an arbitrary open subset of
R'. Let (Cl r° o be an infinite sequence of non-empty compact sets, such
that
C; C Int C‘ .4. 1 (3.11.27)
for i a non-negative integer and

hm C, = SZ. (3.11.28)
1—.00
Put C_ 1 = 0. For each i > 0, let (p, be a C°° function on Rn such that
0 <(P.*< 1 ,
C 1 C int{x : cp,(x) = 11, (3.11.29)
and
sptcP, C Int C, +l• (3.11.30)
Put
^Go = (Po and fir, = (pi — gyp,_ 1 (3.11.31)
when i > 1. Then each t' is C°° on R" with compact support and

spt C (IntCs+l) — C1_ 1 . (3.11.32)

Hence, for each x E SZ, i(x) 0 for at most two values of i. Therefore
00

E 0, ( x ) = 1 (3.11.33)
s-o
for all xE U. For each i=0,1,2,... define

_f when x ESZ
u(x),M')
u=(x) (3.11.34)
{
when x SZ.

By the conclusion of our theorem proved under the assumption that SZ =


R", there exists for each i > 0 a Cm function v, on Rn with compact
support such that
II ZI, Il m. P < 21 +1 '
^

-
tl ,

and
R i - m,p( Fi) < 2:+1

where
F = R" fl {x : ui (x) # vi(x)I•
Moreover,
spt v, c (Int C; +l ) — Cc-1 • ( 3. 1 1 . 37)
168 3. Pointwise Behavior of Sobolev Functions

For each x E SZ, there are at most two values of i for which v i (x) 0 O.
Hence we can define 00
v(x) = E.,(x)

for x E SZ. It is easily seen that F C U=°_ 0 F;, hence

Rt_m,p(F) < e.
Also OC

E oui U < E.

a
i!o

3.11.7. Remark. We have seen from earlier work in Section 2.6, that
Rk, p < CBk, p and that Rk, p and Bk,p have the same null sets. However, it
also can be shown that Bk, p Ç C[Rk p + (Rk, p )n' (n -kp) t for kp < n, cf. [A5].
Therefore, the Riesz capacity in the previous theorem can he replaced by
Bessel capacity.

Exerc i ses
3.1. Prove that the statement

lim u(y)dy = u(x)


r^0
B(x,r)

for B 1 , p -q.e. x E R" and any u E W 1 "F(R") implies the apparently


stronger statement

lirn
r ^0 8(z,r)
1u(y) — u(x)Idy = o

for 13 14,-q.e. x E R". See the beginning of Section 3.3.

3.2. It was proved in Theorem 2.1.4 that a function u E W'•P(Rn) has


a representative that is absolutely continuous on almost all line seg-
ments parallel to the coordinate axes. If a restriction is placed on p,
more information can be obtained. For example, if it is assumed that
p < n — 1, then u is continuous on almost all hyperplanes parallel
to the coordinate planes. To prove this, refer to Theorem 3.10.2 to
conclude that there is a sequence of integral averages

Ak(x) = u(y)dy
B(s,)

which, for each r > 0, converges uniformly to u on the complement


of an open set U whose B 1 , p -capacity is less than e. Hence E
Exercises 169

f,>oUU is a set of /31, p -capacity O. It follows from Theorem 2.16.6


(or Exercise 2.16) that the projection of E onto a coordinate axis
has linear measure O. Note that u is continuous on 7r -1 (t), t ¢ ir(E)
where it denotes the projection. Corresponding results for p > n - k ,

k an integer, can be easily stated and proved.

3.3. At the beginning of Section 3.9, an example is given which shows that
u need not be hounded when u E 1,471 ."(B(0, r)}, r < 1. This example
can be easily modified to make the pathology even more striking. Let
u(x) = log log( 1 /IxI) for small ix' and otherwise defined so that u is
positive, smooth and has compact support. Now let

v z =_ E 2 -k u(2 - rk)
( )

k=1
where {r k } is dense in Rn. Then y E W 1, n(R") and is unbounded in
a neighborhood of each point.

3.4. Use (2.4.18) to show that if u E Wiop (R"), p > n, then u is classically
differentiable almost everywhere.

3.5. Verify that IIuIj Tk.p(z), which is discussed in Definition 3.5.4, is in fact
a norm.

3.6. If u E W 1 .n(R" ), the classical Lebesgue point theorem states that

li rn Iu(x) — u(xo)Idx = 0 (*)


r-0 B(xo,r)

for a.e. x o . Of course, u E L 1 (R") is sufficient to establish this result.


Since u E W 1 .p(R11 ), this result can be improved to the extent that
(*) holds for /3 1 , p -q.e. x o E R" (Theorem 3.3.3). Give an example
that shows this result is optimal. That is, show that in general it is
necessary to omit a B 1 , p -null set for the validity of (*).

3.7. Prove that (3.3.22) can be improved by replacing p by p' = np/(n -


kp).

3.8. A measurable function u is said to have a Lebesgue point at so if

le m
r D--
Iu(y) - u(x o )jdy = O.
B(zo,r)

A closely related concept is that of approximate continuity. A mea-


surable function u is said to be approximately continuous at s o if
there exists a measurable set E with density 1 at xo such that u is
continuous at xo relative to E. Show that if u has a Lebesgue point
170 3. Pointwise Behavior of Sobolev Functions

at x 0 , then u is approximately continuous at x 0 . See Remark 4.4.5.


Show that the converse is true if u is bounded and that it is false
without this assumption.
Another definition of approximate continuity is the following. u is
approximately continuous at xo if for every e > 0, the set

A _
, {x : tu(x) — u(x o )I >_ e}
has density 0 at xo. A is said to have density 0 at x o if
f

}im I
A E n B(xo , r)I = 0.
r ~ 0 I B(xo , r)I

Prove that the two definitions of approximate continuity are equiva-


lent.

3.9. The definition of an approximate total differential is analogous to


that of approximate continuity. If u is a real valued function defined
on a subset of R", we say that a linear function L : R" --■ RW is an
approximate differential of u at 10 if for every e > 0 the set

A, = { x: lu(x) — u(xo) — L(x — xo) > €


Ix — xol
has density 0 at xo. Prove the analog of Exercise 3.8; show that if u
is an element of t 1 "(xo), then u has an approximate total differential
at x o .

3.10. The definition of an approximate total differential given in Exercise


3.9 implies that the difference quotient

lu(x) — u(xo) — L(x — xo)


Ix — xol

approaches 0 as x --► x0 through a set E whose density at 1 0 is


1. In some applications, it is necessary to have more information
concerning the set E. For example, if u E W I. P(R"), p > n — 1, then
it can be shown that u has a regular approximate total differential at
almost all points x 0 . The definition of this is the same as that for an
approximate total differential, except that the set E is required to be
the union of boundaries of concentric cubes centered at x o . To prove
this, consider

u(so)
^o (t+ z) = u(x0 + tz)t — L(x),
and define
'Ys o (t) = sup{Iut(z)i : z E OC}
exercises 171

where C is a cube centered at xo. Since x o is fixed throughout the


argument, let u t (z) = u to (t, z).
STEP 1. For each xo and each cube C with xo as center, observe that
u t E W 1 "P(C) for all sufficiently small t > O. With

ckzo(t) = (lutI F + IHurI P )dx


C

prove that czo (t) —* 0 as t ---t 0 for almost all z o .


STEP 2. Show that u has a regular approximate differential at all z o
that satisfy the conclusion of Step 1 and for which Du(z o ) exists. For
this purpose, let L(z) = Du(x o ) • z. Since u t E W 1 'P(C), it follows
that ut E W 1 "P(K r ) for almost all r > 0 where Kr is the boundary of
a cube of side length 2r. Moreover, from Exercise 3.2, we know that
u t is continuous on all such K. Let

4/t(r) _ JKr U
utlF + ' Du t I)dH n1 .

Let Et = [1/2,11P (r : Sp t (r) < a zp (t) 1 / 2 } and conclude that


1/2 .
1([1/2,11— Et)I < azo(t)

STEP 3. Use the Sobolev inequality to prove that for z E Kr , r E E1 ,

and ,D = (n— 1)/p

1/p
iu t (z)I < Mr -11 Iut[ P dH "- 1
KT
1/p
-1
+ Mr' IDut1pdHn
Kr
/P - Mr1-13S001(r)1/P
< Mr-13Spt(r)1
< [M 213 + Mlazo (t) 1 /2P ,

where M = M(p, n).


STEP 4. Thus, for z E K r and r E Et ,

yzo (t • r) = r -1 sup {lu t (z)j : z E Kr }


< 2[M2 s + 11 • a xo (t) 1 / 2P.

STEP 5. For each positive integer i, let t i = 2 - ` and let Et , be the


associated set as in Step 2. Set A = u Et and note that 0 is a
point of right density for A (Step 1) and that 720 (i) --+ 0 as t ---+ 0,
t E A.
172 3. Pointwise Behavior of Sobolev Functions

3.11. Prove that Ti defined in (3.6.2) belongs to C'(U - A).

3.12. Give an example which shows that the uniformity condition in the
second part of the statement in Theorem 3.5.7 is necessary.

3.13. In this and the next exercise, it will be shown that a function with
minimal differentiability hypotheses agrees with a C 1 function on a
set of large measure, thus establishing an extension of Theorem 3.11.6.
For simplicity, we only consider functions of two variables and begin
by outlining a proof of the following classical fact: If u is a measur-
able function whose partial derivatives exist almost everywhere on a
measurable set E, then u has an approximate total differential almost
everywhere on E. See Exercise 3.9 for the definition of approximate
total differential.
STEP 1. By Lusin's theorem, we may assume that E is dosed and
that u is its partial derivatives are continuous on E.
STEP 2. For each (x, y) E E consider the differences

i(x, y; h, k) = lu(x + h, y + k) - ri(x, y) - hptu(x, y) - kDzu(x, y)l

a l (x, y; h) = lu(x + h, y) - u(x, y) - hD l u(x, y)l


4 2(x, y; k) = lu(x, y + k) - n(x, y) - kD2u(x, y)}
where D 1 = 8/ax and D2 = a fay. Choose positive numbers e, r.
Using the information in Step 1, prove that there exists a > 0 such
that the set A C E consisting of all points (x, y) with the property
that
l{x+ h : Ai(x,yy;h) < rh, (x+h,y) E E,
a< x< b, lb - al <o- , Ihl < lb -- al}l> (1 ` e)1b- al
satisfies I E - Al < e. Perhaps the following informal description of A
will be helpful. For fixed (x, y), let us agree to call a point (x + h, y)
"good" if i 1 (x, y; h) < rh and (x + h, y) E E. The set A consists
of those points (x, y) with the property that if Iz is any interval
parallel to the x-axis containing z whose length is less than o-, then
the relative measure of the set of good points in Ix is large.
STEP 3. Now repeat the analysis of Step 2 with E replaced with A to
obtain a positive number a l < o' and a closed set B C A, IA - BI < c
which consists of all points (x, y) with the property that

I {y + k : 02(x,y; k) <Tk, (x,y +k) E A,

a < y -< b, lb - al <17 1 , lkl < lb- al}' > (1 - E)lb - aI.
Exercises 173

STEP 4. Let o2 < aj be such that

IDIu(x +h2, y+k2) — Dlu(x+ h i , y+kt)1 < T


for any 2 points (x+h2, y+k2), (x+hi, y+k,) in E with 1h2 — hil <11 2,
Ike — k11 <al.
STEP 5. Choose (xo, yo) E B and let R = [a l , bul x [a 2 , b 2 J be any
rectangle containing (x o , y o ) whose diameter is less than a2 < a l < a.
Let

E2 = {(yo + k) : A2(xo, yo; k) < T1kI,

(xo, yo + k) E A, a 2 < y o + k < b211

and for each (y o + k)

E1(yo + k) _ {(xo + h) : O1(xo, yo + k; h) < 7Ih1,

(xo+h,yo+ k) E E}.

Now for any (h, k) such that Vo + k E E2 and x o + h E E1(yo + k),


we have (xo + h, yo + k) E E n R and therefore

A(xo, WI; h, k) <Di (xo, Yo; h)+ 42(xo, Yo; k)


+ Ihi ID 1144, yo + k) — D1u(xo, yo)I
< T(IhI+ Ik1).

From this conclude that

1B n R n {(xo + h, yo + k) : A (x0, yo; h, k) < 21. (Ihl + IkI)}i


> (1 — E) 2 01 — (2 1)(b2 — ai) = ( 1 — e) 2 IR1-
STEP 6. Take R to be a square with (xo,Vo) as center and appeal to
Exercise 3.8 to reach the desired conclusion.

3.14. We continue to outline the proof that a function whose partial deriva-
tives exist almost everywhere agrees with a C' function on a set of
large measure. Let u be a real valued function defined on a measurable
set E C R", and for each positive number M and x E E let

A(x, M) = En {y : Iu(y) — u(x)I < MIy — xI}.

If A(x, M) has density 1. at z, u is said to be of approximate linear


distortion at x. Our objective is to show that if u is of approximate
linear distortion at each point E, then there exists sets Ek such that
E = Uk 1 Ek and u is Lipechitzian on each of the sets Ek.
174 3. Pointwise Behavior of Soholev Functions

STEP 1. If z 1. and z 2 are any two points of R", then

_ I B(xl, ix2 — z, l) n B(x2, Ix9 - x11)1


a
— IB(x1, Ix2 — x11)1 + IB(x2,1X2 — Z11)1
is a positive number less than 1 which is independent of the choice of
z 1 and z2.
STEP 2. For each positive integer k, let Ek be the set of those points
x E E such that lu(x)l < k and that if r is any number such that
0 < r < 1/k, then

IA(x, k) n B(x, r )I >


1 — rx.
IB(x, r)I
Prove that E = Uk Ek.
STEP 3. In order to show that u is Lipschitz on Ek, choose any two
points z 1 , z2 E Ek. If (x2 — x 1 I > 1/k, then

Iu(x2) — u(xl)1 < 2k 2 Ix2 — x11.


Thus, assume that
0 <Iz 2 -z 1 I< k1 .
Let
A1 = A(xl, k) n T3(x1, Ix2 - x1 1),
A2 = A(x2, k) n B(zz, Ix2 xi l). -

Prove that IA1 n A l l > O. If ac' E A1 n A2 1 show that

Iu(z") — u(z;)l < klz' — x,l, i = 1,2


Ix" x. l< Ix2 — z l I, i= 1, 2.

Now conclude that

1u(x2) - u( x 1)1 < 2 k1z 2 - x 1 I.

STEP 4. 1f u has partial derivatives almost everywhere, appeal to the


previous exercise to conclude that u is of approximate linear distor-
tion at almost every point. Now refer to Theorem 3.11.6 to find a C'
function that agrees with u on a set of arbitrarily large measure.

3.15. Suppose u E 1471 .P (Rn). Prove that for $ 11p -q.e. x E R te , u is abso-
lutely continuous on almost every ray A s whose endpoint is z.

3.16. Let f be a measurable function defined on [0, II having the property


that f' exists everywhere on [0, 1] and that If'l is integrable. Prove
that f is an absolutely continuous function.
Historical Notes 175

Historical Notes
3.1. The idea that an integrable function has a representative that can be
expressed as the limit of integral averages originates with Lebesgue [LE21.
The set of points for which the limit of integral averages does not exist (tlie
exceptional set) is of measure zero. Several authors were aware that the
exceptional sets associated with Sobolev functions or Riesz potentials were
much smaller than sets of measure zero, cf. [DL], [ARS1j, [FU], [FL], [C11.
However, optimal results for the exceptional sets in terms of capacity were
obtained in [FZ], [BAZ], [ME2], [CFR1. The development in this section is
taken from IMIZ1.
3.2. The results in this section are merely a few of the many measure
theoretic density theorems of a general nature; see [F, Section 2.10.91 for
more.
3.3. Theorem 3.3.3 was first established in [FZ] for the case k = 1, and for
general k in [BAZ], [ME2], and [CFR]. The concepts of thinness and fine
continuity are found in classical potential theory although their develop-
ment in the context of nonlinear potential theory was advanced significantly
in [AM], [HE2], [HW], [ME31. The proof of the theorem in Remark 3.3.5
was communicated to the author by Norman Meyers.
3.4. Derivatives of a function at a point in the LI-sense were first studied
in depth by Calderon and Zygmund [CZ]. They also proved Theorem 3.4.2
where the exceptional set was obtained as a set of Lebesgue measure zero.
The proof of the theorem with the exceptional set expressed in terms of
capacity appears in [BAZ], [ME2], and [CFR].
3.5. The spaces Tk(E), t k (E), Tk.P(s), and tk•e(x) were first introduced
in [CZ] where also Theorem 3.5.7 was proved. These spaces introduce but
one of many methods of dealing with the notion of "approximate differen-
tiability." For other forms of approximate differentiability, see [F, Section
3.1.2], [RR].
3.6 - 3.8. The material in these sections is adopted from [CZ]. It should
be noted that Theorem 8 in [CZ] is slightly in error. The error occurs
in the following part of the statement of their theorem: "If in addition
f E tt(xa) for all x 0 E Q, then f E b u (Q)." The difficulty is that for
this conclusion to hold, it is necessary that condition (1.2) in [CZ] holds
uniformly. Indeed, the example in [WH] can be easily modified to show
that this uniformity condition is necessary. Theorem 3.6.3 gives the correct
version of their theorem. In order for this result to be applicable within the
framework of Sobolev spaces, it is necessary to show that Sobolev functions
are uniformly differentiable on the complement of sets of small capacity.
This is established in Theorem 3.10.4.
In comparing Whitney's Extension theorem (Theorem 3.5.3) with the
LP - version (Theorem 3.63), observe that the latter is more general in the
176 3. Pointwise Behavior of Sobolev Functions

sense that the remainder term of the function u in question is required to


approach zero only in LP and not in L°O. On the other hand, the function
is required to be defined only on the set E in Whitney's theorem, while the
condition u E t iP(x) in Theorem 3.5.3 implies that u E L1[B(x, r)] (for all
small r > 0) thus requiring u to be defined in a neighborhood of r.
3.9. Theorem 3.9.4 and the preceding lemmas are due to Calderon [CA4];
the remaining results are from [BAZ].
3.10. The main result of this section is Theorem 3.10.2 which easily implies
that Sobolev functions are uniformly differentiable in LP on the complement
of sets of small capacity. The proof is due to Lars Hedberg. Observe that
the proof of Theorem 3.10.2 becomes simpler if we are willing to accept a
subsequence {r , } such that
;

u(y) — u(x) jpdy —• 0


B(x,r, 1 )
uniformly on JP —U where Bk, p (U) < e. This can be proved by the methods
of Lemma 2.6.4. However, this result would not be strong enough to apply
Theorem 3.6.3, thus not making it possible to establish the approximation
result in Theorem 3.10.5.
3.11. These results appear in [MIZ]. The main theorem (Theorem 3.11.6)
is analogous to an interesting result provcd by J.H. Michael [MI] in the
setting of area theory. He proved that a measurable function f defined on
a closed cube Q C R" can be approximated by a Lipschitz function g such
t hat
t{x : f ( 2 ) g(x)} I < e
and IA( f , Q) — A(g, Q)I < e where A(f , Q) denotes the Lebesgue area of f
on Q. Theorem 3.11.6 was first proved by Liu [LI) in the case m = t.
4

Poincaré Inequalities
A Unified Approach
In Chapter 2, basic Sobolev inequalities were established for functions in
the space Wô 'p(St). We recall the following fundamental result which is a
particular case of Theorem 2.4.2.

4.1.1. Theorem. Let 0 Rn be an open set and 1< p<n. There is a


constant C = C(p, n) such that if u E W' F (SI), then

Ilullp•;n < CIIDuIII.p;n (4.1.1)

where p' =np/(n - p).

Clearly, inequality (4.1.1) is false in case u is the function that is identi-


cally equal to a non-zero constant, thereby ruling out the possibility that
it may hold for all u E W1'F(1). One of the main objectives of this chapter
is to determine the extent to which the hypothesis that u is "zero on the
boundary of Stn can be replaced by others. It is well known that there are
a variety of hypotheses that imply (4.1.1). For example, if we assume that
SZ is a bounded, connected, extension domain (see Remark 2.5.2) and that
u is zero on a set S with ISI = a > 0, then it can be shown that (4.1.1)
remains valid where the constant C now depends on a, n, and 1. This in-
equality and others similar to it, are known as Poincaré-type inequalities.
We will give a proof of this inequality which is based on an argument that
is fundamental to the development of this chapter. A general and abstract
version of this argument is given in Lemma 4.1.3.
There is no loss of generality in proving the inequality with p' replaced
by p. The proof proceeds as follows and is by contradiction. If (4.1.1) were
false for the class of Sobolev functions that vanish on a set whose measure
is greater than a, then for each integer i there is such a function u; with
the property that
IIuiIIp;n > iIIDu iIIP;^.
Clearly, we may assume that Hu; Ij 1, p p = 1. But then, there exist a sub-
sequence (denoted by the full sequence) and u E W l'p(Q) such that u ;
tends weakly to u in W 1 "p(11). By the Rellich-Kondrachov compactness
theorem (Theorem 2.5.1, see also Remark 2.5.2) u tends strongly to u in
;

L'(0). Since IIu ; fI1, p; n = Zit follows that IIDu,UI p; n --+ 0 and therefore that
II Duli p; n = O. Corollary 2.1.9 thus implies that u is constant on S2. This
178 4. Poincaré Inequalities—A Unified Approach

constant is not 0 since Ilull p; ^ = 1. Now each ui is 0 on a set S, whose mea-


sure is no less than a. The strong convergence of ui to u in LP(S2) implies
that (for a subsequence) u, —' u almost everywhere on S = rlj"_ U°°_ ; Si.
Since ISI > a > 0, this contradicts the conclusion that u is equal to a
non-zero constant on Il.
A close inspection of the proof reveals that the result also remains valid if
we assume fI u(x)ds = 0 rather than u 0 on a set of positive measure. In
this chapter we will show that these two inequalities and many other related
ones follow from a single, comprehensive inequality obtained in Theorem
4.2.1.

4.1 Inequalities in a General Setting


We now proceed to establish an abstract version of the argument given
above which will lead to the general form of the Poincaré inequality, The-
orem 4.2.1.

4.1.2. Definition. If X is a Banach space and Y C X a .subspace, then a


bounded linear map L : X --+ Y onto Y is called a projection if 1! o L = L.
Note that
(4.1.2)
for there exists x E X such that L(x) = y and y = L(x) = LIL(x)1 = L(y).

4.1.3. Lemma. Let X o be a normed linear space with norm It HD and let
X C X0 be a Banach space with norm II Il• Suppose It II = II lia + II Ili where
I^ Ill is a semi-norm and assume that bounded sets in X are precompact in
Xo . Let Y = X n Ix : IIxlII = 01. If L : X -; Y is a projection, there is a
constant C independent of L such that

Ilx — L(z)Ilo < CIILII (4.1.3)


for all x E X .

Proof. First, select a particular projection L' : X —, Y. We will prove that


there is a constant C' = C' (#I L' II) such that

Ilx L'(x)llo <_ ^'Ilxllr,



(4.1.4)

for all x E X. We emphasize that this part of the proof will produce a
constant that depends on L' .
If (4.1.4) were false there would exist x i E X such that

llx■ — L (^s)Ilo > =Ilxilti,


'
i = 1,2,... .
4.1. Inequalities in a General Setting 179

Replacing r i by xi/Uri — L i (xi)11 0 it follows that

Ilxi — L'(x.)II o 1 and Ilx^lli ^ 0.

Let z,-r,— L'(x t ). Then


Ilzilf 1= IIx — L'(:r )IIi + If L'(x^)II^
IIx=111
since Il L' (xi) II 1 = O. Hence, z, is a bounded sequence in X and therefore,
by assumption, there exist a subsequence (which we still denote by { z i } )
and z E X o such that II z i -- z Ila —00. Since I1 zi jf 1 —0. 0 it follows that z, is a
Cauchy sequence in X and therefore It; — zII —, 0. Note that IIzIIo = 1 and
Ilzll^ = O. Thus z O, z E Y, and L'(z)=z by(4.1.2). But L'(z ; )— , L'(z)
and L'(z,) = 0, a contradiction.
The next step is to prove (4.1.3) for any projection L where C does not
depend on L. Let L : X — Y be a projection and observe that

r— L(r) = x — L'(x) — L(x — L'(x)).


Hence, by (4.1.4),

llx — L(r)Ilo <_ 11x — L'(x)Ilo + IIL(x — L (z))Ilo '

< C llxlll +
' — L`(r))fl
C'Iirlll + IILII fl(r — L'(x))II
< C r llrlll + IILII (II(x — L'(z))Ilo + I141}
since 1IL'(x)Ii i = 0. Appealing again to (4.1.4) we obtain,

lix — L(x)110 < C ' llxlli + IILII [C'llxll l + Il x ll ll


(C' + (C' + 1 )IILII)ItxlIi.

Since L is a projection, 1141 > 1 and the result now follows. O

We now will apply this result in the context of Sobolev spaces. In par-
ticular it will be convenient to take X = WmiP(Tl).
For notational simplicity, in the following we will let the characteristic
function of S) be denoted by 1. That is, let xn = 1. Also, let Pk(R") denote
the set of all polynomials in R" of degree k.

4.1.4. Lemma. Let k and m be integers with 0 < k < m and p > 1.
Let S? C R " be an open set of finite Lebesgue measure and suppose T E
(W m-k m(SZ))' has the property that T(1) O. Then there is a projection
L : Wm'P(SZ) —. Pk(R") such that for cash u E W m 'D(SZ) and all < k,

T(Dr u ) = T(D°iP) (4.1.5)


180 4. Poincaré Inequalities—A Unified Approach

where P L(u). Moreover, L has the form

4%0= E T(Pa(Du))x«
«1Sk

where P« E Pk(R"), D u = (D i u, D 2 u, . . . ,13.u), and


k +1
(111)
II n I ? C•GIT

C = C(k,P, PI)•

Proof. If P E Pk(R') then P has the form P(x) = EÉ,71=0 a.: 7 and
therefore
D0P(0) = a!a a
for any multi-index a. Consequently, by Taylor's theorem for polynomials,

k-1«1 Da+o p(0 ) ^A


D a P(x) = E
A'
IQ1=o
or
k-1n1
(a + Q)!
D"P(x) = E
1o1 =o
610-1-(3 (3! rA

In particular,
D'P( x ) = aaa !

if IcI = k. Thus, in order to satisfy (4.1.5), the coefficients a s of the poly-


nomial must satisfy
T(Dau)
as = , (4.1.6)
a!T(1)
if IcI = k. Similarly, if lai - k -- 1 then

al + E as + p (a +
D°P(x) = a,
0) 1 x A .
IQ1=1
p!
Consequently by using (4.1.6), (4.1.5) will hold if

(a + f3 )! T (xs )
a s = T (D"u) - a s +p a!
a.T(1) /3! T(1)
101=1
where lal = k - 1. Proceeding recursively, for any lai < k we have
IQ1
T(D°U) (a + p)! T (x^)
-

a -k a«+A (4.1.7)
« a!T(1) a!/3! T(1)
1A1=1
4.1. Inequalities in a General Setting 181

It is easily verified that L is a projection since L(u) = P implies Da [L(u)] =


DaP for any multi-index a. But then,
T(DQu) = T(DQP) = T[Da(Lu)]
and reference to (4.1.7) yields the desired conclusion.
In order to estimate the norm of L, let u E Wm.p(1) with Ilulf m , p; Ç < 1.
Then IILII <_ IIL(u)Ilm,p;n = IIPIIm,p;fl where P(x) = > ki ,rl _ o a.y x 1'. Now

111316,pin < C(I1I) E la-rI.


I-v1= 0
To estimate the series, first consider IaQI, lai = k. Note that for ial = k
and any non-negative integer t,

1171 •T(1)r +1
C(l, p, iHI) IITI I (4.1.8)
a&T(1)1+t 1)
a!T(1) ( )

because T(1) In particular, this holds for = k. Hence from


(4.1.6) it follows that
k+1
< I ^I (4.1.9)
1 '2 ' 1 a&T
< C(k^P, ICI)

If lal _ k-1, k > 1, then from (4.1.7), (4.1.9) and the fact that IITII/T( 1 ) >
I0I -1 Ip
^ITII + C(k 1 1 1) E aQ+ Q ?ITII
IaQI < a.T(1) ,
,
I0I=1 (1)

< C(k, p, I0I) TII 2I


Tw (
)

^. k+ 1
< C'(k,p, I^I) II l#
T (Z)
In general, if lai = k — i , k > ti, we have

!co <C(k,p,InI) IITII =+1


T(1)

_
< C'(k,p, If2I)IITII
T (1) k+1 -
Proceeding in this way, we find that
k+1
fill
^

II ^II <_ C(k,p, I^I) T


)
182 4. Poincaré Inequalities—A Unified Approach

In the preceding analysis, if we knew that the distribution T was a non-


negative measure p, then we would be able to improve the result. Indeed,
suppose the measure satisfies the inequality

fn edp(x) < Mp(Q) (4.1.I0)

for every Ic < k. Of course, such an M exists if either SZ or spt a is hounded.


Then the estimate of IILII becomes sharper because, with T = p, the term
T(7 0 )/T(1) in (4.1.7) is bounded above by M, thus implying that

II
IILII < C(k,p,M)
()
Hence, we have the following corollary.

4.1.5. Corollary. Let k and m be integers with 0 < k < m and p > 1.
Let SZ C R" be an open set and suppose p E (W m-k .p(1l)) . is a non-
negative non-trivial measure satisfying (4.1.10). Then there is a projection
L : —> Pk(Rn) such that for each u E Wm.p(1) and all ice' < k,

p(D"u) = p(DaP) (4.1.11)

where P = L(u). Moreover, L has the form

L(u) = E p ( Pa( Du )) xa
iai<k

where P, E Pk(R11 ), Du = (D i u, Dzu, ... , D"u), and

IILII < C - (Ilpll


p(f)) '

C = C(k, p, M).

4.2 Applications to Sobolev Spaces


We now consider some of the consequences of the previous two results when
applied in the setting of Sobolev spaces. Thus, if 0 < k < m are integers,
p > 1 and SZ C R" is a bounded, connected, extension domain (see Remark
2.5.2), we employ Lemma 4.1.3 with X = Wm , p(11) and X 0 = W47k .P(0). It
follows from the Rellich-Konrachov imbedding theorem (see Exercise 2.3)
that bounded sets in Wm.P(S)) are precompact in W k, P(11). Set IIuIIo =
Il ull k JAI and IlullL = IlD k+i ullm-(k+1),p;f2 where Dk+lu is considered as the
vector {Dau) lai = k + 1. Clearly, (lull = Hullo + Hulk is au equivalent
norm on W m (0) . Moreover, it follows from Exercise 2.7 that llull1 = 0 if
4.2. Applications to Sobolev Spaces 183

and only if u E Pk(R n ). If T E (Wm - kip(SZ))' with T(I) 96 0, then Lemma


4.1.4 asserts that there is a projection L : Wm'P(SZ) Pk (Rn) such that
fli ) k+]
IILII C
T(1)
(

Therefore, Lemma 4.1.3 implies

IIu - L(u)Ilk,p;s1 CIILII IID k+ `ullm- (k +1),p;n


k+1
^C 11111

IIDk +lUl l m— (k +l),D;f1


T(1)

These observations are summarized in the following theorem

4.2.1. Theorem. Suppose 0 < k < m are integers and p > 1. Let SZ C Rn
be a bounded, connected extension domain. Let T E (Wm - k'P(S2)) 4 be such
that T(1) O. Then, if L : Wm'P(l) Pk (R1z) is the projection associated
with T,
k+ 1
11Thl
IIu - L(u)IIk,D;n <_ C (
(1) )
II Dk+1Ullm--(k+1),p;r1 (4.2.1)

where C = C(k, p, 12).

It will now be shown that the norm on the left side of (4.2.1) can be
replaced by the LP . -norm of u -- L(u), where p' = np/(n - mp). For this
we need the following lemma.

4.2.2. Lemma. Suppose m > 1 is an integer and p > 1. Let SZ C Rn be a


bounded extension domain. Then for each integer k, I < k < m - 1, and
e > 0 there is a constant C = C(n, m, p, k, r, 1) such that

llD k l4IP;n < CIIuII F ;c + ellDmullD;si, u E W m ' A ( 1 l) (4.2.2)

whenever u E W m .P(11).

Proof. We proceed by contradiction. If the result were not true, then for
each positive integer i there would exist u i E Wm'P(12) such that

IlD k w llp;^ > + (4.2.3)

replacing ui by ui/Iluillm, a we may assume that IIu^IIm p n = 1, i = , ;

1, 2, .... Hence, from Exercise 2.3 there is u E Wm.P(S2) and a subsequence


(which we assume without loss of generality is the full sequence) such that
u i — u strongly in Wm -1 "(f2). In particular ui — u in L"(11). Since

II!D^`u^llD;n _< Ilu: ll m 1,v;n _< IIu^II^,D;^,


-
(4.2.4)
184 4. Poincaré Inequalities—A Unified Approach

it follows from (9.2.3) that u, —, 0 in LP(1) and therefore u = O. But then


ui —, 0 in Wm-i p(S2) and consequently IID k n i lip;n —' 0 by (4.2.4). This
,

implies that IIDmu, ll p p —, 0 or u i —, 0 in Wm'p(SZ), a contradiction to the


fact that IIu,II m , p; n =1.

If v E Wm'p(RR) has compact support, then it follows from the funda-


mental Sabolev inequalities, namely Theorem 2.4.2, 2.9.1, and 2.4.4 that

I IU IIp• <_ GlI v l lm ,p


where p' is defined by
1 1 m
if mp < n,
P .
^ P n
1< p` <oo if rnp = ra,
and
p' = oo if mp > n.
Since Si C R" is an extension domain, u E Wm'D(1) has an extension
to v E W m .P(R' 1 ) with compact support such that IlVIIm,p < (; II^ IIm,p;i2•
Therefore,

Ilullp• ; CI l !) IIF-
.

< CIIVIIm,p
< Cllüllm,p;n
< C EII^IIn + IlDmu llp;nl (4.2.5)
by Lemma 4.2.2. Now apply this to (4.2.1) while observing that D'(L(u))
0, lot' = m, and obtain

lin — L(u){ 1p•;0 — L(n)Ilp;n + IlDmullp,nl


< C IITII k+i IID k + i ^Ilm- ( k+^y,P;s^'
T(1)
We have thus established the following result.

4.2.3. Corollary. With the hypotheses of Theorem 4.2.1,


k+1
(1177
Ilu — L(n)llp•:n <_ C T(1)1 IIDfc+'Ullm-(k+i)p;n.
/)

4.2.4. Remark. In many applications it is of interest to know when L(u) =


O. In this connection we remind the reader the coefficients of the polynomial
L(u) are given by (4.1.7) and will be zero if T(D°u) = 0 for 0 < Ial <
k. The question of determining conditions under which L(u) = 0 will be
pursued in Sections 4.4 and 4.5.
4.3. The Dual of Wm'P(fl) 185

4.3 The Dual of WmtP(1)


In order to obtain more information from inequality (4.2.1) it will be helpful
to have a representation of (Wm'P(1l))', the dual of Wm'P(11). This is easily
accomplished by regarding Wm.P(Ç) as a closed subspace of the cartesian
product of LJ'(SZ).
To this end let
N=N(n,rn)= E
0<1a1<m
be the number of multi-induces a with 0 < lal < m. Let
N
LN(S/) _ 11L(c).
LN(SZ) is endowed with the norm
m 1/p

^ Ilvallp n;
if 1 <p<oo
IIUIIp, N;^2 = 1 aE— o
m ax IIUII00;fi if p = oo
o <1ai<m

where y = {v a } E LN(SZ).

4.3.1. Theorem. Let SI C Rn be an open set. Then each linear functional


T E (Wm.a(SZ))', 1 < p < oo, can be represented as
m
T(u) _ E v a (x)D° u(2)dx for u
1 E W m 's'(SZ), (4.3.1)
1a1_0
,
where y = {v a } E LN(SZ).

Proof. Clearly, the right side of (4.3.1) defines an element T E (Wm.p(SZ))'


with
11Th
see (2.1,5). In order to express T(u) in the form of (4.3.1) first observe
that W in I P (n) can be identified as a subspace of LN((). The operator
D: W"(f2)— LN(1) defined by
D(u) = {D°u}, 0 < lai < m
has a closed range since Wm•P(fI) is complete. Define a linear functional
T' on the range of D by
T' f D(u)J = T(u), u E Wm.P(ft).
186 4. Poincaré Inequalities —A Unified Approach

By the Hahn-Banach theorem, there is a norm preserving extension T'


of T' to all of LN (I1). By the Riesz Representation theorem, there exists
i1 = { v a } E L%;(1) such that

m
T' (w) = E f v (x)w a a (r)dx
rai_o
whenever w = {wa } E LN(SZ). Thus, if u E W m ''(0), we may regard

Du = {D a u) E LN(1Z) and therefore

_
T(u) = T' ID(u)] = T'(Du)
m

n ya(x)Day.(x ) dZ. ❑

1 (2 1= 0a
In the event that 1Z C R" is a bounded extension doinaiii, the repre-
sentation of (Wm°P(SZ))' is slightly simpler, as described in the following
result.

4.3.2. Theorem. If SZ C R'' is a bounded extension domain and 1 < p <

E
oo, then each element T E (Wm•P(Ç1))' can be represented as

T(u) = f vu + (
v0D ° u)dx
(4.3.2)
2
l° 1-m

where y, v a E Lp`(S2), lcxl = m .

Proof. The proof is almost the same as in Theorem 4.3.1 except that now
Wm.P(S?) can be identified with a subspace of LN (ft) where N = k(m) + 1,
and k(m) = the number of multi-indices a such that lad = M. Thus u E
Wm.P(SZ) is identified with (u, {Dau}1a1=m). In view of Lemma 4.2.2 this
provides an isometric embedding of Wm'P(12) into LN(SZ). ❑

It is useful to regard the restriction of the linear functional T in Theorems


4.3.1 and 4.3.2 to the space .95(1) as a distribution. Indeed, if Sp E g(ft)
is a Schwartz test function (see Section 1.7), then from (4.3.1) we have
m
T(Sp) _ E j i, a D° Spdx (4.3.3)
Ial= 0
where v ° E Lp' (II). In the language of distributions, this states that T is a
distribution in Ç with

T = E -1 101I7'va ( ) (4.3.4)
1a1= 0
4.3. The Dual of Wm'p(SZ) 187

where v a E LP G (SI). Similarly, if T is the functional in Theorem 4.3.2, then

T = y + E (-1)I0ID°`ya (4.3.5)
Ian=m

where u, v a E L (il). However, not every distribution T of the form (4.3.4)


or (4.3.5) is necessarily in (Wm.p(S2))'. In case one deals with Wâ ' p (SZ)
instead of Wm'p(SZ), distributions of the form (4.3.4) or (4.3.5) completely
describe the dual space, for if T is a distribution as in (4.3.4), for example,
then it possesses a unique extension to Wp "(û). To see this, consider
u E Wô ' p (f2) and let {vi} be a sequence in (SZ) such that çp, —, u in
Wo p (SZ). Then

la1= f
IT(92i ) — 4.02)1 =-- E vaDacp i — ya Da vjt
101.4)
m

E 11Da(cpi — Spj)II r 1IV IIp',n


I a1.0
r I
Il^t — ^PjIlm^pIIvaIIp',o --t O

as i , 3 a oo. Thus, T (cp ) converges to a limit, denoted by t(u), which is


- ;

well-defined. f' is clearly linear and bounded, for if Bp i -. u in W 'P(SZ),


then

IT(u)I — +moo
llrn IT(Pi)1 lun
f^oo
IITII II^i1lm,p
= IITII II tIIm,p•
The norm IITII in this context is defined relative to the space Wô `"(SZ).
These remarks are formalized in the following theorem.

4.3.3. Theorem. Let 1 < p < oo. If SZ c R" is an open set, then the dual
spare (Wô '"(SZ))' consists of all distributions T of the form
m
T = E (-1) Ial D a v a
Ial=O

where v a E L" (fi). If SZ is a bounded extension domain, then (Wo 'p(0))'


consists of those T such that

T = v + E (-1)Ia!Dckva

v, v a E (S/) •
The dual space (Wô 'P(f2))' is denoted by W-m,p'(A).
188 4. Poincaré Inequalities—A Unified Approach

4.4 Some Measures in (W(cZ))*


We now exploit Theorem 4.2.1 and its Corollary 4.2.3 to derive some of the
most basic and often used Poincaré-type inequalities. These inequalities
are obtained below by considering Lebesgue measure and its variants as
elements of (Wm°p(0))'.
In order to demonstrate the method that employs the results of Section
4.2, we begin by reproving the inequality

_ CIID m uilp
Il D k ull p < (4.4.1)

for u E Co (R"), where 0 < k < m are integers and p > 1_ Suppose that the
support of u is contained in some ball: spt u C B(0, r). Let S2 = B(0, 2r).
With this choice of f2, we wish to apply Corollary 4.2.3 by selecting T so
that the associated projection L will have the property that L(u) = O.
Then by appealing to (4.2.2), we will have established (4.4.1). Define T E
(W m + km (f 1 )) • by
T(w) _vw dx
n
for w E Wm - kmo(0), where v ! XB(0,20- B(0,r)• Since spt u C B(0, r),

T(Dau) = 0 for 0 < lai < k

and therefore L(u) = 0 by Remark 4.2.4. Hence, (4.4.1) is established.


In case Il is a bounded open set and u E Co (f2), a similar result can
be established by defining u to be identically zero on the complement of 0
and by considering a bail B(0, r) that contains Il. Since Cr (0) is dense
in WtP(f2) the following is immediate. (Of course, this result also follows
from the inequalities established in Chapter 2.)

4.4.1. Theorem. Let SI C R" be a bounded set. Let 0 < k < m be integers
and p > 1. Then, there is a constant C = C(k, m, p, diam 1 2) such that
-

II D k ullp,o < CII D m ullp ;n.


for u E Wô 'P(S2).

A slight variation of the preceding argument leads to the following results.

4.4.2. Theorem. Suppose 0 < k < m are integers and p > 1. Let S2 be a
bounded extension domain. Suppose u E Wm.P(11) has the property that

1 Daudx_0 for 0<lai<k,

where E c Sl is a measurable set of positive Lebesgue measure. Then,


IIt II k,p;O < CIID k4- Mini-(k+1),pin
4.4. Some Measures in (Wô 'F(f l))' 189

where C = C(k, m, p, 0, IEI).

Proof. Define T E (Wm`k.p(1))' by

T(w) = f wdz, w E W m-k, P(0).


E
Then T(1) 0,
T(D°u)=0 for 05IcI< k,
and therefore by Remark 4.2.4 the associated functional L has the property
that L(u) = O. The result now follows from Theorem 4.2. 1. ❑

4.4.3. Corollary. If u E W m.P (fû) has the property that Dck u = 0 almost
everywhere on E for 0 5 loll < k, then

Ilullk.p;n < CIIDk+1Ullm-(k+14;f1.

Theorem 4.4.2 provides a Poincaré-type inequality provided the integral


averages of the derivatives of u over a set E of positive measure are zero. In
the next result, the integral average hypothesis is replaced by one involving
the generalized notion of median of a function. If the sets A and B below
are of equal measure, then we could think of 0 as being the median of u
over A U B.

4.4.4. Theorem. Let fi E Rn be a bounded extension domain and let


u E W 1 'e(0), p > 1. Suppose u > 0 on A and u < 0 on B, where A and B
a re measurable subsets of D of positive Lebesgue measure. Then

Ilu llp.n < C IIDüIIP;c^


where C = C(p,n, IAI,IBI)•

Proof. Let
=u ds and Q = u dx
A B
and define T E ( W i .p(Sl))' by

T(w) = f vwdx, w E W i ^p(f2)

where v = (1/a4A — (1/0)XB. Then T(n) = 0 and the result follows from
Theorem 4.2.1 and Remark 4.2.4. 0

4.4.5. Remark. In the remainder of this section, we will include a small


development of the notion of trace of a Sobolev function on the boundary
of a Lipschitz domain as well as some related Sobolev-type inequalities
190 4. Poincaré Inequalities—A Unified Approach

(Theorem 4.4.6 and its corollary). This material will be subsumed in the
development of BV functions in Chapter 5, but we include it here for the
benefit of the reader who does not wish to pursue the BV theory.
If Sl C R" is a bounded Lipschitz domain and u E W 1 "F(Sl), 1 < p < oo,
it is possible to give a pointwise definition of u on Of) in the following way.
Since fl is an extension domain, let ù denote an extension of u to all of
Jr where û E W''n(R"). Therefore, it has a Lebesgue point everywhere on
R" except possibly for a set of B1, p -capacity zero (Theorem 3.3.3). Since
p > I, we know from Theorem 2.6.16 that sets of B 1 , F -capacity zero are of
H" -1 -measure zero and therefore it is defined H" -1 -almost everywhere on
an. We define the trace of u on On by setting u = té on OSi.
We now show that this definition is independent of the extension ü. For
this purpose, we first show that at each Lebesgue point x o of û, there is a
measurable set A such that the Lebesgue density of A at xo is 1 and that
°
ii is continuous at x relative to A. Since

I^(x) - Ü(xo)Idx 0 as r -■ 0,

for each positive integer i, there is a number r i such that the set E i =
R" fl {x : Iu(x) - u(x o )1 > 1/4 has the property that

IB(x o , r) nE I •
< 2 -1 , for r < r i . (4.4.2)
IB(xo, r)I
We may assume that the sequence {ri} is strictly decreasing. Let
00
E = UB(xo,rj) - g(xo, ri_ 1 )] n Ei .
i-1

We now will show that the Lebesgue density of E at x o is zero, that is

lirn IB(xa, r) n EI = 0. (4.4.3)


r---.0 IB(x, r)I
Choose a small r > 0 and let k be that unique index such that rk+1 <
r < r k . For notational simplicity, let B(r) = B(x o , r). Then from (4.4.2) it
follows that

IB(r) fl El < U [(B(r) n Ei) n ($(ri) - B(ri+1))j


+ =k
ao
< 2 -k IB(r)I + E 2 'IB(ri)I-

i =k +1

< 2 kiB(r)I + 2 k1B(r)1


-
-
4.4. Some Measures in (Wo"(f ))' 191

which establishes (4.4.3). Clearly, if we set A = Rn - E we have that u is


continuous at xo relative to A and that the Lebesgue density of A at z o is
one.
Because 1? is a Lipschitz domain, the boundary of SZ is locally repre-
sentable as the graph of a Lipschitz function. Thus, the boundary can
be expressed locally as {(x, f (x)) : x E U), where U is an open ball in
Rn-1 and f is a Lipschitz function. Recall from Theorem 2.2.1 that a Lip-
schitz function is differentiable almost everywhere. Moreover, the function
7 : R"' -+ R" defined by

7(x) = (x, f (x))


is Lipschitz and carries sets of Lebesgue measure zero in Rn -1 into sets of
H" 1 -measure zero in R". Consequently, an possesses a tangent plane at
all H"" 1 -almost all points of 31l. From this it is not difficult to show that

I _1
inn B (xo, r) n DI 2 ,
01
for H" -1 -almost all x o E an. Since the Lebesgue density of A at xo is equal
to one, it follows that

lirn
fB(zo r) n ' sin A I - 2.
1
/
r^U I B(xo+ r)I
Also, because u is continuous at z o relative to A, it is clear that

litn u(x) = ü (x 0 ).
X ---DZa
rE fi n A

This shows that the value of ti(r o ) is determined by u in SZ, thus proving
that the trace of u on the boundary of SZ is independent of the extension
ü.
In the statement of the next theorem, we will let denote the restriction
of (n - 1)-dimensional Hausdorff measure to Oft That is #(A) = H" -1 (An
80) whenever A C R".

4.4.6. Theorem. Let SZ c R" be a bounded Lipschitz domain and suppose


u E W'P'(fï), 1 < p Cao. Let

c(u) =udHn -1 = udµ.


a sp a rt
Then µ E ( W I 'P(SZ))" and
1/P' f/n
lu - c(u)lo . dx < C(f IDa^lpdx
Jn ,
(in
([
192 4. Poincaré Inequalities—A Unified Approach

where p' = np/(n - p) and C = C(n,p, fZ).


Proof. Because fZ is a bounded Lipschitz domain, u has an extension ü
to all of R n such that IIÙII1 ,p < By multiplying i by a function
E Co (Rn) with too _ 1 on SI, we may assume that ti has compact support.
In order to show that E (W 1 .p(cl))*, we will first prove that

f vdp < CIIvII 1,p (4.4.4)

whenever v is a non-negative function in Co (Rn). from Lemma 1.5.1, we


have 00
v dµ = f(E µ i )dt (4.4.5)

where Et = {x: v(x) > t). By the Morse-Sard theorem, for almost all t,
Et is bounded by a smooth manifold. We now borrow an essentially self-
contained result of Chapter 5. That is, we employ Lemma 5.9.3 and Remark
5.4.2 to conclude that for all such t, E t can be covered by balls B(x 1 , r^ )
such that 00

E t„n -1 ^ ^:f^n-1^U-1(t)7 (4.4.6)


i=1
where C is a constant depending only on n. Because Of/ is locally the graph
of a Lipschitz function, it follows from (1.4.6) that there is a constant C
such that p(B(x, r)) < Crn' 1. Thus, from (4.4.6) it follows that
00

ti(Et) < EL(B(zj,ri))


1 =1
00
C G,C`` r n -1 < CH n-1 (v -! (t)).
L
L^ t
t= 1

Appealing to (4.4.5) and co-area formula (Theorem 2.7.1), we have


00
v dN = f µ(E: )dt
00
�C Hn - 1 (v -1 (t))dt
0
= CIIflvII1
• C(sZ)IIDvllp
< cllUlll,p,
thus establishing (4.4.4).
If y is now assumed to be a bounded, non-negative function in W 1, p(Rn),
we may apply (4.4.4) to the mollified function v E . From Theorem 1.6.I we
4.5. Poincaré Inequalities 193

have that llv€ — vil lip —, 0 and that v,(x) v(x) whenever x is a Lebesgue
point for v. From Theorem 3.3.3 we know that y has a Lebesgue point at all
x except possibly for a set of B l ,p capacity zero, therefore of H'-measure
0, and therefore of it-measure O. Consequently, by Lebesgue's dominated
convergence theorem,

f V c dL —+ fv dµ.

It now follows that (4.4.4) is established whenever y is a non-negative,


bounded function in W 1 'P(Rn).
If we drop the assumption that y is bounded, then we may apply (4.4.4)
to the functions
if v(x) > k
vk(x) — k
v(x) if v(x) < k.
It follows from Corollary 2.1.8 that vk E W 1,P (R") for k = 1, 2, .... Thus,
an application of the Monotone Convergence theorem yields (4.4.4) for non-
negative functions in W 1, P(R"), in particular, for ü+ and ü`. Hence (4.4.4)
is established for u.
From Remark 4.4.5 we have that u = u Hn — l-almost everywhere on 81?,
and therefore

J udi — udµ
Ç CiRul11 ,P
Gllulll,p;n.
Thus, we have shown that ti E (W 14)(1l))*, and reference to Corollary
4.2.3 completes the proof. ❑

The following is an immediate consequence of Theorem 4.4.6.

4.4.7. Corollary. If SZ is a bounded Lipschitz domain and u E W 1 .P(11),


p> 1,then

and
f u dHn -1 S C ^llüll p • in + 11Dullp,nl

Ilullp.;u <_ C [ itLiip c


; dHn-1
js1
As mentioned in the beginning of Remark 2.4.5, these inequalities will
be extended to the situation when u E BV, thus including the case p = 1.

4.5 Poincaré Inequalities


Here we further develop the results in Section 4.2 to obtain Poincaré-type
inequalities for which the term L(u) in inequality (4.2.1) is zero. We will
194 4. Poincaré Inequalities—A Unified Approach

show that this term vanishes provided the set {x : u(s) = 0} is sufficiently
large when measured by an appropriate capacity.
First, recall from Corollary 2.6.9 that i f' A C R" is a Suslin set, then

B,, p (A) = sup{B,, p (K) : K C A,K compact}.

Moreover, if' K C Rn is compact Theorem 2.6.12 implies that there is a


non-negative measure p such that spt P C K,

119c * PIIp' < 1,


and
P(R") = [Bcp(K)1 1I P
Now consider u E Wm -k m(f2) where fZ C Rn is a bounded extension
domain. Then u has an extension û defined on Rn such that IIÙII m _k p < ,

CIlullm_k,p;ç. Without loss of generality, we may assume that u has com-


pact support. From Theorem 2.6.1 it follows that ü has the representation

u =9m-k * f
where f E I1(R n ) and IJ f II
suppose that p is a non-negative measure with the properties that
Now suppose
spt p C 1 and
9m-k*P E LP' (R")
where k is an integer, 0 < k G i. Observe that p can be considered as an
element of (Wm - k(f2))' for if we define T : Wm k(SI) -' R' by
-

T(u) = f u dµ,

then,

f pdp = füdµ
f— 9m - k * f dFt

- 9m--k * p • f dx, by Fubini's theorem,


S N9m-k * Pulp' if lip, by Iidlder's inequality,
* ullp+Ilullm -k,p'
Ç CII9rn -k * lltilim-k,p;0- (4.5.1)

Thus, p E (Wm-kip(f2))`.
4.5. Poincaré Inequalities 195

This leads to another application of Theorem 4.2.1 which allows the


main constant in the inequality to be estimated by the capacity of the set
on which u vanishes.

4.5.1. Theorem. Let SI E R" be a bounded extension domain and let


A C R" be a Suslin set with Bm _k, p (A) >0 where0<k<m are integers
and p > 1. Then, there exists a projection L: Wm"p(s1) lk(R") such that

1Iu - L(u)Ilk,p,n Ç C (Bm-k,p(A)) -1/F IID 144- 'uIIm-(k+1),p ; n


where C = C(k, rn,p, 0).

Proof. From the above discussion, there exists a non-negative measure p


such that is supported in A,

119m-k * <1
and
p(fin) ? -
2
If we set T = p in Theorem 4.2.1, we have T(1) = p(R") > 0 and from
(4.5.1) that
11Th <_ CIIgm-k * pIIp' <_ C.
The result now follows from Theorem 4.2.1 and Corollary 4.1.5. ❑

4.5.2. Corollary. Let u E Wm'p(0) and let


N = f1 {x: D °iu(x) = 0 for all 0< Iocl <k} .

If B m _k, p (N) > 0 then

II^IIk,P;^ C(Bra_k,p(N)) -
LIP IID k+l ullm-(k+l),p;n
and
k+1
IIuIIp•,O < C (Bin-k,p(N)) Jl 1P D ullm-(k+1),pm.
Proof. The coefficients of the polynomial L(u) in Theorem 4.5.1 depend
upon
T(Dau) = D°`u dp

for 0 < lkl < k, and thus are all zero, (see Remark 4.2.4). The second
inequality follows from Corollary 4.2.3. ❑

Because of the importance of the case m = 1, k = 0, we state the Poincaré


inequality separately in this situation.

4.5.3. Corollary. If u E W l p(SZ), then


,

^I^1Ip•;^ ^ C (Bl,p(N)) - ' l° IIDu^I°,^ (4.5.2)


196 4, Poincaré Inequalities—A Unified Approach

where N — {x : u(x) = 0}.

4.6 Another Version of Poincaré's Inequality


We can improve the inequalities of Corollary 4.5.2 if we allow dependence
on the set N and not merely on its capacity. In particular, if j, k, an d m are
integers such that 0 < j < k < m, then the assumption B,n _k, p (N) > 0 will
be replaced by the weaker one, Bm _( k _ j ), p (N) > 0, provided an additional
condition is added which requires dependence on the set N in the resulting
inequality.
To make this precise let 0 be a bounded extension domain and let N C S2
be a Suslin set with the property that

Bin - (k -j),p(N - Z) > 0

for every set Z of the form

Z n {x: DaP(x)- 0, OO P E Pk( R")}.

These sets comprise a subclass of the class of algebraic varieties. Thus, for
any algebraic variety of the form (4.6.2), we require some subset of N of
positive capacity to lie in the complement of Z.
Let M(N) denote the set of all non-negative Radon measures com-
pactly supported in N such that

9m-(k-j) * E L P. (R " ).
Consider all functionals of the form

T(u) = fDau dµ,, u E M(N), (4.6.3)

where Ia l < k— j. We will verify that all such T are elements of (Wm P (0) r.
,

Let u E Wm.P(SZ). Since SZ is an extension domain, there is an extension û


of u to R" with IIÜIIm,p < from Theorem 2.6.1, we know that û
IIu
has the representation fa = g," * f, where f E Lp(R") and II fli p ^' IIm,p
Since is supported in N C 1, it follows that

Daudµ = f D°lidµ.

But D'û E Wm ial.P(R") and therefore


-

Da ii = 9m--,a1 * g

4.6. Another Version of Poincaré's Inequality 197

where g E L' (R n ). By Fubini's theorem we have

f Daü dp = f 9m-1a1 * 9dµ


f 9rn_lol *p.gdx.

It follows from (2.6.2) that

9m-101* th = ge * (gm-(k-1) * P)i

where l = (k - j) - 'al. Since gm _(k_ i ) * a E LP' (R") by assumption and


ge E L 1 (R'a ), it follows from Young's inequality that gT1-la1 * L E LI? (Rn).
Therefore an application of Holder's inequality yields

DaU dp = f 9m-IoI * tz • g dx

Ilgm-iai * ilIIp` IIgilP


fl9m - 1oI * PUP' IID"uIlm - Iol.p.
However,
iiDaullm-iar,p Ilullm.p - CIIUIIm,p,
thus proving that T E (W m 'P (OW .
Let V C (Wm•P(1))* be the space spanned by all such functionals T as
defined in (4.6.3). Let

Vo={TIPk: TEV},
so that Vo C N. Observe that
dim Vo = dimPk(R")

or
V0 = [Pk(R")] * ,
for if this were not true, there would exist OOPE Pk(R") such that
T(P) = 0 for every T E V. This would imply

D a Pdp=O ► Ic:tI <k - j

for all i E M(N). That is, from Theorem 2.6.12, this would imply

Er 13 (x) = 0,
for Bm _( k _ j )-q.e. z E N and loci < k - j, a contradiction to (4.6.1) and
(4.6.2). Therefore dim Vo = dimPk(R") or Vo = IPk(R")]*. This implies
the existence of Ta E V such that
Ta (xS) - 6a ,B. (4.6.4)
198 4. Poincaré Inequalities--A Unified Approach

Hence, if we define L ; Wm'p(0) --► Pk(R") by


rn

L(u) = E T (u)x°,
or (4.6.5)
1a1 =o
then (4.6.4) shows that L is a projection. An appeal to (4.1.3) results in
the following theorem.

4.6.1. Theorem. Let it C R" be a bounded extension domain. Suppose


N C St is a Suslin set such that (4.6.1) is satisfied. Then, with L given by
(4.6.5), there is a constant C = C(j, k, m, N,1) such that

IIu - L(u)Ilk,p;fi < CII Dk+ I tillnt-(k+1),p;n.


The special nature of the projection L is what makes this result inter-
esting. For example if we assume that Dau = 0 on N except possibly for
a Bm _(k_,), p -null set, then all T of the form (4.6.3) are zero and therefore
L(u) = O. The following is a consequence.

4.6.2. Corollary. Let n C R" be a bounded extension domain. Suppose


N cn is a Suslin set such that (4.6.1) is satisfied. If u E W m •p(SZ) is such
that
D"u(x) = 0 for B m _( k _ 3 ), p -q.e. x E N
and all O < la! < k - j, then

IIuIIk,p;n CllD
k+1 fllm- (k+1),p;t2

where C = C(j, k, m, N, Il).

4.7 More Measures in (


The general inequality (4.2.1) involves a projection operator L : Wm•p(SZ)
Pk(R") which is determined by an element T E (W m-k ip(SZ)) • It is there-
fore of importance to have an ample supply of elements in the dual of
Wm - k , P(0) that are useful in applications. In Section 4.4 we have already
seen that Lebesgue measure (more precisely, suitably normalized measures
which are absolutely continuous with respect to Lebesgue measure) and
normalized (n-1)-dimensional Hausdorff measure belong to (Wm - kiP(S1))*.
The fact that these measures are elements of (Wm -k •p(Çl))* allowed us to
deduce interesting Poincaré-type inequalities. In this section we will per-
form a finer analysis to establish that a large class of measures belong
to (Wm k'p(û))•, including those that are obtained as the restriction of
-

Hausdorff measure to sub-manifolds of appropriate dimension in R".


4.7. More Measures in (Wm'p(S2))' 199

We begin with a result that provides a generalization of the Sobolev


inequality for Riesz potentials and also gives us a method of exhibiting a
large class of measures that are elements of (Wm -k .D(S1))'. It will depend
on the Marcinkiewicz Interpolation Theorem which we state here without
proof.
Let (p o , q o ) and (p i , q i ) be pairs of numbers such that 1 < p; < q; < CO,
i = 0,1, po < pi, and qo 0 q t . Let p be a Radon measure defined on R"
and suppose T is an additive operator defined on Co (Rn) whose values are
µ-measurable functions. The operator T is said to be of weak-type (p i , q,)
if there is a constant C i such that for any f E Co (R'), and a > 0,

1.4({x :1(Tf)(x)I > a1) < (ct -i Ctflf IIp. ) 4 ' •

4.7.1. Theorem (Marcinkiewicz Interpolation Theorem). Suppose T is


simultaneously of weak-types (po , qo ) and (pi, qi). If 4 < 0 < 1, and

1 -B ^
1Ip_ +
Po Pi
1 - 1 B B
I4 + go Pi ^

then T is of strong type (p, q); that is,

Il^f ll q;,i <_ CCo -B Cl Ilflip, f E Co (R"),


where C = C(p„ q i , 9), i = 0, 1.

We are now in a position to prove the basic estimate of this section which
is expressed in ternis of the Riesz kernel, Ik, that was introduced in Section
2.6.

4.7.2. Theorem. Let µ be a Radon measure on RR such that for all x E R"
and 0 < r < oo, there is a constant M with the property that

01B(x, r)1 < Mr'

where a-q/p(n - kp),k> 0, 1 < p < q < oo, and kp < n. If f EI"(Rri),
then

(1
i/4
* fl°dp < CM'/gI1f11D
11k
where C = C(k, p, q, n).

This inequality is obviously an extension of the Sobolev inequality for


Riesz potentials that was established in Theorem 2.8.4. In that situation,
the measure is taken as Lebesgue measure. In Remark 4.7.3, we will
200 4. Poincaré Inequalities—A Unified Approach

discuss further what other measures play an important role in the inequality
of Theorem 4.7.2.

Proof o f' Theorem 4.7.2. For t > 0 let

At= {y: 'k * III(Y)>t}.


Our objective is to estimate p(A t ) in terms of III Let p t = µ I A t . Then

tp(At) Ik * If I di. = Ik * IÎI dfpe


fg

L . he kt(x)II (x)I dx (4.7.1)

where the last equality is a consequence of Fubini's theorem. Referring to


Lemma 1.5.1 it follows that

Ik * Pe (x) = it [B (r,rh/(k_')] dr
7( k)j °°
°°
= ( n — kf) t (B( x r)) r k_hI_ 1 d r .
7(k)
For R > 0 which will be specified later, (4.7.1) becomes
J R
dx dr
ti(A) I^ (x) ,r)]rk
I t[$(x
ar(k) Rn
_
a-
+ n k c) If(x)Ipt(B(x, r)]r k- ldx dr
rfL
= I 1 + 12 . (4.7.2)

Since µ(B(x, r)] < Mr° by hypothesis, the first integral, I t , is estimated
by observing that

pt[B(x,r)] A t(B(x, r)]1!p'(Mra)1/P

and then applying Holder's inequality to obtain

jl < (n — k) R l/p
I1111 M'1p
(fR.ILL[B(x,r)]dx) rk-n-1+(a1p)dr.
7(k) o
(4.7.3)
We now will evaluate
Ftt[B(x, r)]dx•

For this purpose, consider the diagonal

D = (Rn x Rn) n {(x,b): x= y}


4.7. More Measures in (Wm'P(Si))' 201

and define for r > 0 ,

D r = (R n x R n ) (1 {(x> y) Iz -
yI < r}.
Finally, let F = xDr.. Then, by Fubini's theorem,

P t f B(x, r) l dx = dJ^cty)dx
RR R " f (s,r)
= F(x, i!)di^e (b)dx
R^ R"

=
Lin L. F(x, P)dx dpi (^)

= I B(Y , r)I dµt(U)


R"
cr(n)r n µ(At).

Therefore (4.7.3) yields

p(n - k) n - a )p
^ < II /Ilpml/pa(n)^/p'µ(At}t/P' Rk - (

7(k)[kp (n^a)l
Similarly, by employing the elementary estimate

1jt(B(x, r)] < pt[B(x, r)]1/p'µ(At)1/P }

we have
_ ao 1/p'
I^ < (n k)k) II /IIpp(At)' / P ^l,t(B(x, r)]dx r^` -n -ldr
< r p(
^±'(
k- k) k
R R"

-n /P
f pl7{ ) IIfIIpti(At)ck(n)l/P'Rk

Hence
( ? 1P
Il + < P(n - k) ^(n) 11 P ' M 1 /PP(A)
t 1 /P'Rk_n_a
Ii.f II p
[


-
•y(k) kP - (n - a)

+ A(At)R k -n /P
n - kp
In order for this inequality to achieve its maximum effectiveness, we seek
that value of R for which the right-hand side attains a minimum. An ele-
mentary calculation shows that
1la
R =
N(At)
M ►
202 4. Poincaré Inequalities— A Unified Approach

and the value of the right-hand side for this value of R is

p(n )a ^(n) i/p` M 1Iv


a(n)11P1 F^(At) i-i l^ll f llp-
7(k)( n k
P)(P n+
ri )

Consequently, from (4.7.2)


t p (A t )i/q i
tA(At)p(At)119-1
< p(n - k)a a(n)1/p'Mi/v
^y(k)(n- kp)(kp-n+a)
((A e )'""(At) III ii )
a(n)IIp M il4
llt'lp. ( 4.7.4)
Ç 7(k)(n - kP)(q - P)

Expression (4.7.4) states that the Riesz potential operator Ik is of weak


type (p, q) whenever p and q are numbers such that

1<p< q<oo, kp<n. (4.7.5)

Hence, if (po, qo), (p, q) and (p i , qi) are pairs of numbers such that (po. qo),
(pi, qi) satisfy (4.7.5) and for 0 < 0 < 1,

1/p- 1- B +
0
Po Pi
1/ _ 1-B + B
q -
qo Pi
then the Marcinkiewicz Interpolation Theorem states that Ik is of type
(p, q), with
111k* .6l4:µ <_ GM 119 l1 f Ilµ,
thus establishing our result. D

4.7.3. Remark. The number a that appears in the statement of Theorem


4.7.2 is equal to n when q = np/(n - kp) = p'. In this case the conditions
of the theorem are satisfied by any measure p that is absolutely continuous
with respect to Lebesgue and that has bounded density. In particular, if
we take p as Lebesgue measure, we can recover Theorem 2.3.6, which is
Sobolev's inequality for Riesz potentials.
Theorem 4.7.2 also provides an inequality for Riesz potentials restricted
to a lower dimensional submanifold MA of Rn. For example, if Ma is a corn-
pact, smooth A-dimensional submanifold of Rn, then it is easy to verify that
A-dimensional Hausdorff measure restricted to M satisfies the condition of
Theorem 4.7.2. That is, if we define p by

p(E) = H(E n MA),


4.7. More Measures in (Wm.p(SZ))` 203

for every Bore] set E C Rn, then there is a constant M such that

p[B(x, r)1 < Mr A (4.7.6)


for every ball B(x, r) C R". Now let f E I"(R") and consider the potential
u = Ik * f.
By Theorem 4.7.2 we have
i/A•
u A. dp < Cil fli p
where a` = J1p/(n — kp), n — A < kp < n. In other words,
3/ A'

U .
Iu.(x) I A. dHa(x) <— Cilfllp (4.7.7)

where C = C(k, p, A, n, M). Note that the constant C depends on M which,


in turn, theory, it is sometimes possible to obtain an equality similar to
(4.7.7) where the constant is independent of the manifold.
Inequality (4.7.7) is valid for Riesz potentials u = Ik * f and thus does
not automatically include Sobolev functions. However, it is immediate that
Theorem 4.7.2 and (4.7.7) apply as well to Sobolev functions u E W k, p(Rn)
because Theorem 2.6.2 states that u can be represented as

u= 9k*f
where g k is the Besse] kernel, f E LP(R") and Ilf ljp , Ilullk,p. Moreover,
we know from (2.6.3) that there is a constant C such that gk(x) < CIk(x),
x 0.
To reassure ourselves that the integral on the left-side of (4.7.7) is mean-
ingful, recall from Theorem 3.3.3 that u is defined pointwise everywhere on
R" except possibly for a set A with Rk, p (A) = O. Therefore, by Theorem
2.6.16, 11 71—k P +e (A) = 0 for every e > 0. By assumption, A > n — kp and
consequently H A (A) = O. Thus, u is defined HA almost everywhere on MA
which is in accord with inequality (4.7.7).
Also, we observe that if p is a non-negative measure on R R with compact
support, and otherwise satisfies the conditions of Theorem 4.7.2, then p E
(Wk.p(1))` whenever II is a bounded extension domain. To see this, let
u E Wk , p(SZ) and let u be an extension of u to R" such that II/14,p < _
Cllullk,pio. Because u E W k, P(R" ), we have
û- 9k * f
where Ilf lip ilüiik,p• Hence, by Theorem 4.7.2 and the fact that spt p is
compact,

f 'O 5_ C
p lu lQdp
lIa
_<_ C IJf Ilp
204 4. Poincaré Inequalities—A Unified Approach

= CIIu IJk.p
< Cll^ llk,p;n.
This establishes the following result.

4.7.4. Theorem. Let 1 C Rn be a bounded extension domain and suppose


p is a compactly supported Radon measure on R" with the property that if
e > 0, there is a constant M such that
kp+e,
p(B(x, r)) < Mr"
-

for all z R" and all r > 0, where kp< n,p> 1. Then p E (Wk p(0))8 .
,

This result obviously is not sharp and thereby invites the question of
determining an optimal condition for p to be an element of (W k .p(Rn))'.
By using a different approach, it is possible to find a condition, related
to the one in the theorem above, that provides a characterization of those
Radon measures that are elements of (W k ip(R"))'_
For this purpose, we need a few preliminaries. Ifµ is a Radon measure,
we will use the fractional maximal operator

Mkp(x) = sup{r k "µ[B(x, r)] : r > 0}.


-

There is an obvious relationship between the Riesz potential of and the


fractional maximal operator: Mkp(x) < CIk * p(x) for every x E R'1 , where
C = C(k, n). The opposite inequality in integrated form is not so obvious
and is implied by a result due to [MW]. It states that for every 1 < p < oc
and 0 < k < n, there exists C = C(k, p, n) such that

IIIk *tiIIp <_ CII kp1I . (4.7.8)

The (k, p) energy of


- is defined as
,
ek.p(µ) _ (9k * p)p dz.
R*

Since the Bessel kernel is dominated by the Riesz kernel, we have

tk.p(p) < C f " * p) p dx


-

Rn
(fk * p) • ( Ik * 11)1/(p-l)dx
-- Ik * (Ik *
µ) 1 / ( p -
1} dµ, by Fubini's theorem. (4.7.9)
f"
The expression
Ik * (Ik * /(p -1)
4.7. More Measures in (W m 'P(f ))' 205

is called the non-linear potential of A.

4.7.5. Theorem. Let p > 1 and kp < n. If is a Radon measure, then


µ E (W k 'D(ft"))' if and only if
1/(p-1) dr
(
IB(vir)1
pT \ dµ(y) < oc. (4.7.10)
r"—
Rn U

Proof. In order to avoid technical details involving the behavior at infinity,


we will give a proof for measures with compact support.
If it is such a measure with 9k * µ E L", then by Fubini's theorem and
with u = gk * f we can write

f udµ= 9k*fdA

f gk *u f dx
II9k* i4
^^f IIP
< CIfgk * IIIp ' 1Iulik,p ,

which implies that /.4 E (W k 'P (R") )' . Conversely, if E (W k, p (R") )', then
the reflexivity of LP implies that gk * µ E LP ' . Therefore µ is an element
of (W k ' p (R n )) • if and only if 119k * µ4 p' < oo, i.e., if and only if the (k,p)-
energy of µ is finite.
We proceed to find a (sharp) condition on µ that will ensure the finiteness
of its (k,p)-energy. For each r > 0,

1 PEB(x'o)1i
• i/p'
2r p
dt
A[B(x, r)l ^C
rn — k f to k —
t
p' I /p'
(f:
µ[B(^,t)1 dt
—<C (, r tn —k t
Thus,
,
^ 1Ip
Jl^t kij(
^< u[B(z, t)^ p dt
)_ C t" - k t
o
Therefore,
,

tic .p(u) ^ ^ (Ik * µ)p dx


R"

5.0 (Mkµ)P dx, by (4.7.8),


^

R*
p ^ dt
< C r
p
t+[B(x, t)^
tn —k t
dx.

R"

206 4. Poincaré Inequalities—A Unified Approach

Now to evaluate the last term, we have

p[B(x, t)]prd z = A[B(x, t)] 11(p-1) u[B(x, t)jdx


fR' R"

< il[B(x, t))1/(p-1) dth(y) dx


R" B(z,t)

< L . fa(o) p[B(y,2t)l11(p-1)diL(y) dx

F(x ) y)F4B(11, 2 t)] 1/(n-1) dp(y) dx


R" R"

< 4B(y, 2t)] 1/(P-1) IB(x, t)Idp(y)


R"
where F(x, y) = XD r and D 1 = Rn x R" n ((x, y) : la - yI < t}. Therefore,
°O (tk ") t!(P-1)
£k,p(µ) < C p' " p[B(y, 2t)] 1B(x, t)Idj(y) dt
D R

[B(y, t)] 1/(F-1) dt dA(y)
<C to-kp t

R" 0 oo
Since has compact support and finite total mass, it is evident that the
expression on the right side of the above inequality is finite if and only if
( [B(Ylt)] ) l?(r -1) dt
1
t"-kp Td(y) < oc.
R" o
This establishes the sufficiency of condition (4.7.10).
For the proof of necessity, we employ the estimate

gk(x) ? Clxi k- "e -21x1 for x E R", z #a (4.7_11)


(see Section 2.6). As in (4.7.9), we have

t"k,p(N) = (9k * Fl) p dx k * (9k *


R JR 'l
To estimate the last integral, let f = (gk * p) 1 /( -1 ) and use Lemma 1.5.1
and (4.7.11) to obtain
00 dr
9k * .f (x)•? G f (y)dy rk-"e-9r
o B(z,r) r

Clearly, for r > 0,

f(y) ?9k(y - x)dµ(z)


B (y,r)
4.8. Other Inequalities Involving Measures in (W k IT 207

and therefore,

00
k-n -2r dr
9k * f (x) ? C f (y)dy r e
o B(s,r) r
a, 1/(P-1)
dr
^^ ( r n 9k(y - z)d►l(z) dy Tk-n e- 2r—
0 BfV.r1 T
(

r
)1/(p-1)
>
C ^, n rk- ne e - 2r dT
-ardF^(z) dy Tk -n
r
o B(y,r)

i/(P-I)
> C fo w F^rB(^J, T)1 e-2P
,r dr
- Tn - kp r

This implies

I 1 ^ ^ ^ I4B(Y f))) 1/(P_l) e_2Plr dlL(y)


^k(l^)
(

tic,
R" o Tn-kp r

>c l^(^1 i/(p-1) - 2p,
' dg y)
^ e r ^r
T^Bn-kpr)^
r

L
,^I l 4B(Yr)] 1 /(P -1) dr
2 (

? Ce-
P — du(ll)• D
T R" r

4.8 Other Inequalities Involving Measures in


(Wk,p)*

We now return to the inequality (4.2.1) for another application. It states


that
I on ii
Hu - L(u)Ilk,p;ç <_ C

where T E (Wm k, P( 11) )' and L : WmiP(SZ) -' Pk(Rn) is the associated
-

projection, L(u) has the form

T(Pa(D)u)Xa
L(u) =
1
Ek

where PQ is a polynomial of degree Ici whose argument is D = (D 1 , ... , D n ).


In Corollary 4.5.2 we found that L(u) 0 if B m _k, p (N) > 0 where

N=ft fl{x: D°u(x)= 0 for all 0< la' < k).

This was proved by the establishing the existence of a measure p > 0


supported in N with p E (Wm k.P(11))• By taking T = p it clearly follows
-
208 4. Poincaré Inequalities—A Unified Approach

that
L(u) = E f P° (D)udp = 0. (4.8.1)

Now, if p is taken as any non-negative measure in (Wm - k.P(11))* with the


property that
D°u dp= 0 for all 0<IaI< k,

then (4.8.1) holds. This observation along with Theorem 4.2.1 and Corol-
lary 4.2.3 yield the following result.

4.8.1. Theorem. Let p > 1 and suppose 0 < k < m are integers. Let
S1 C R" be a bounded extension domain. If u is a non-negative measure on
R n such that p E (Wm - k , P(fl))*, µ(Rn) # 0 and

f D°u dµ = 0 for ail 0 < k,

then
IlUIIk,p;i2 < CIIDk+1uIlm—(k+l),p;S2
and
CIID lc+l ullm_(k+1) PP
I

where C = (k, p, rrb, µ, Sl).


In particular, with k = 0 and m = 1, we have

II^IIp';n < ^IIDuIIP;n


if p E (Wh P(11))* and
f u d = o.
From the preceding section we have found that a non-negative measure
µ with compact support belongs to (Wm -k m(ft))* if, for some e > 0,

sup n [ B(xk )ja E


)
: R", r > 0 < oo. (4.8.2)
r a

Consequently, if a is an integer such that a > n - (m - k)p + c and M A is


a smooth compact manifold of dimension a, then H a I Ma is a measure in
(Wm - k , e(11))•. As an immediate consequence of Theorem 4.8.1, we have

4.8.2. Corollary. Let A be an integer such that A > n - (m - k)p + e


where p > 1 and e > O. Suppose MA is a smooth compact submanifold
4.9. The Case p - 1 209

of dimension JI of R", SZ C R" is an extension domain and suppose u E


Wmip(1) has the property that

JM
AnR D"u dH a = 0 for all 0 < < k

where H a (M P n SI) # O. Then

Ilulik.p;n < CIID k+l ullm—(k+1),p;ç

and
Hullap';ct < CIID k+l ulfm— (k +l),pm
where C = C(k, p, rn, M A , SZ).

4.9 The Case p = 1


The development thus far in this chapter has excluded the case p = 1,
a situation which almost always requires special treatment in LP-theory.
Our objective here is to extend Theorem 4.7.2 to include the case p = 1.
Since the analysis will depend upon estimates involving Wulf , it is not
'

surprising that the co-area formula (Theorem 2.7.1) will play a critical role.
We begin with the following lemma that se rves as a first approximation to
Theorem 4.7.2 in the case p 1. We will return to this later (in Chapter
5) for a complete development in the setting of BV functions.

4.9.1. Lemma. Let > 0 be a Radon measure on R" and q a number


such that 1 < q < n/n — 1. If

B ( x,r
Il rq( _l) ) 1!
sup : x E R", r> 0 < M

for some M > 0, then there exists C = C(q, n) such that

1/q
uQdµ CM1/61IIDuII 1 (4.9.1)
R"

whenever u E Co (R").

Proof. First consider q =1 and refer to Lemma 1.5.1 to conclude that

f r<dµf4Et)dt
_F ( 4 .9 . 2 )

whenever u E Co (R") is non-negative. Here Et = {x : u(x) > t}. Because


u is continuous, OEt C u -1 (t) for each t > 0; moreover the smoothness
210 4. Poincaré Inequalities -A Unified Approach

of u and the Morse-Sard theorem states that u. -1 (t) is a smooth (n - 1)-


manifold for almost all t > O. Consequently, Lemma 5.9.3 and Remark 5.4.2
imply that for all such t there exists a covering of E t by a sequence of balls
B(x i , r i ) such that
m
Er:1-1<cHn-ipEo <cHn-i(0,-1 ( 0})
i =1

where C = C(n). Hence,


00
1i(Et) < µ{$(Ti,ri)1
,-1

<rnE
i=1
< C H 1 [{u -1 (t)}n

where C = C(n). Referring to (4.9.2) and the co-area formula (Theorem


2.7.1) we have
co
f udjc = p(E t )dt
0
00
< CMHn -1 ({u - '(t)})dt
0
= CM I Dul dx .

R"

If u is not non-negative, write lui = u+ - u - , and apply the preceding


argument to u+ and u - to establish our result for q = 1.
Now consider q > 1 and let g E Lg (p), g > 0. Then, Holder's inequality
'

implies
1/q'

9 dµ < ggrdJ1 p[E(i, r)]114


B(x,r) (fB(x,r)
<_ ^1 / g I1911g' ;^Tn -^
Thus, gp is a Radon measure which satisfies the conditions of the lemma
for q = 1. Consequently, if u E Co (R") we have

J IttIgdµ < CM 1
14
Ilsllg';^•
for all g E Lq (u), g > O. However, by the Riesz Representation theorem,
^n
I^uld^
!

Ilullq,^. = sup R"


IuIg dp : II91I q' < 1, g> 0
4.9. The Case p = 1 211

and our result is established. a

4.9.2. Remark. The restriction q < n/n - 1 in the lemma is not essential.
If q > n/n - 1, the lemma would require a Radon measure ti to satisfy

4B(x, r)) < Mr'


for all x E R n , all r > 0, and some m > n. However, there is no non-trivial
Radon measure with this property. In order to see this, let U C R" be a
bounded open set. Choose e > 0 and for each x E_U, consider the family
Gm of closed balls B(x, r) such that 0 < r < e and B(x, r) C U. Defining

ÿ={B: B E ç= ,x E U}
we see that Q covers U and thus, by Theorem 1.3.1 there is a disjointed
subfamily C Ç such that

U CU{B :BE Ç}cU{B: B E.1"}.

Hence, denoting the radius of B, by r ; , we have


00

u(U) < E gift) Ç M E 5r,)m (

BE.* i=1
Er n.
00
< M5m C m -n
i=1
< M5 m c m-n lUl .

Since U is bounded and e is arbitrary, it follows that ,.(U) = O. D

Our next objective in this section is to extend inequality (4.9.1) by re -

placing on the right side by (IDtuIIi. For this purpose it will be


necessary to first establish the following lemma.

4.9.3. Lemma. Lei > 0 be a Radon measure on R", t < n, 1 < q <
(n-1 +1)(n - t) and r -1 = 1 - (q - 1)(n -1)/n. Then there ta a constant
C such that for all z € R" and r > 0,
rL—(n+l) II
II
l s *^ r;B{s,r) <(: SUp{r(i-n)Qµ[B(x, r)j : x E R", r > 0}.

Proof. It will suffice to prove the lemma for x = O. An application of


Minkowski's inequality for integrals yields
i/r
dp(y) r
dx
(
18(0.?) (
L(O.2?) Ix _ Yi " -1

212 4. Poincaré Inequalities—A Unified Approach

1/r

f B(0.2r) f
_ (n-1} 7
(O,r) 4x ^I
dx
du(F1)• (4.9.3)

Observe that

fB0 1 i x
dx
_ V(n- 1) r
J (0,r)r1B(y,r) Ix -
dx
YI N-1)r
dx
+ 1)r
B(O r)
, -
B(y r) Ix -
,
YIN
The first integral can be estimated by
dx < Crn-r(n-1)

I (n-1)r
fB(//,r) I x `- 11I
and the second integral is dominated by r n-r(n-I) Here we have used the
fact that (n - 1)r < n. Thus,

dx < Crn-i(n-1)
,

/Bo o-) Ix - VI N-11r


and therefore from (4.9.3)
r 1/r
r t-(n+1)
dp(3i)
d^
( far) 9(0,2r) ix - YI N-1)
< Cr(C-n)gµ[B(O, 2r)j.
(4.9.4)
If 1x1 < r and M > 2r, then IyI < 21v - xl. Consequently,
r 1/r
r l-(n+1) dµ(y) dx
L(D,r) Li^ 2 r Ix - yl (nT1)
( (

< Gr (")-(Q-1)(n-1) f dµ(11)


4-1
(4.9.5)
11ip.2 r I11I •
Appealing to Lemma 1.5.1, we have
o0
OM <
(n - 1) µ[B(0, t)^t ndt.
JY I>2r lyl ^-i
-

2r
Now define a measure v on B 1 by y= t (n-l) Q -n dt and write

1413(0, t))i -n dt = µ[B(0, t )] t (t-')gd v


fr jr
o0
< sup {r'(B(O,
r>0
01/ dv
Jr
< G• r (n-t)Q-n+1 sup TO-11)12/4[B(0, 01-
r >O
4.9. The Case p = 1.
213

This combined with (4.9.5) and (4.9.4) yield the desired result.

We are now prepared for the main result of this section.

4.9.4. Theorem. Let p > 0 be a Radon measure on R" and let l < n,
q > 1. Then if

sup il[B(x1 r)] : x E R n , r > 0 - M < co,


r

there exists C = C(q,n) such that

U.
lfg
04 5 GM 1/g if D 1 uI11

whenever u E Cô (R n ).

Proof. If l = n, then for u E Cô (Rn) and x E R,


n

u(x) < fRn


ID n uldy,

from which the result follows.


If 1 = f < n, the result follows from Lemma 4.9.1.
Next, consider the case l < n, t > 1 and q > n/(n - 1). Since u E
' n/(n_1)
Co (R"), it follows that u E Kt— l (R n ) and therefore u = 91-1 * f,
f E L"/("-1)(Rn) with Ilfllnf(n-1) - IIuIIl-1,n/(n-1) ^^
IIDr-lulln/(n_1).
Thus, Theorem 4.7.2 implies

Ilullq:µ <— CM 1/a IID 1-l ulln,(n-1 ).

Since W e-i lk/n-1 < CIIDrulll by Theorem 2.4.1, our result is established
in this case.
Finally, consider l < n, t > 1, and q < n/(n - 1). We proceed by
induction on t, assuming that the result holds for derivatives of orders up
to and including t - 1. As in (2.4.5),

L. l u(,),,,, ) <C
Rn
it ID
z(_iu (in)1 1)dy dFe(x)
I yI
5 CgIDuI Iul g-l fi * µ dy (by ^bini's theorem)
Rn
< CgllulIn/(1n_ O II IDulfl * µll (by Holder's inequality)
T

where r -1 - 1 - (q - 1) (n - l)n - 1 . By Sobolev's inequality,

11u1r, f(n -i) < CIiD l uqi - . (4.9.6)


214 4. Poincaré Inequalities—A Unified Approach

To estimate * µ HT let m be a measure on R'" defined by m =


(I 1 * µ)T dx and apply the induction hypothesis to obtain

II IDuII^ * IIIr = II IDuI II=;m


< Csup{r (1-1)- nni1D(x, r)rT : r > 0,x E R n } IID 1 uII1
= C sup{r (r--1)~ "IIIj * µjlT ; ^(z T) : r > 0, x E R"} IJD ' uII I •
,

This combined with (4.9.6) and Lemma 4.9.3 establishes the proof. D

Exercises

4.1. Give a proof of Corollary 4.5.3 based on the argument that imme-
diately precedes Section 4.1. You will need the material in Section
2.6.

4.2. The following provides another method that can be used to define
the trace of a Sobolev function u E 14/ 1 'P(S1) on the boundary of a
Lipschitz domain S1,
STEP 1. Assume first that u E C 1 (Si), u > O. For each x o E (3S1 and
with the (n + 1)-cube centered at z o with side length 2r denoted by
C(x o , r), we may assume (after a suitable rotation and relabeling of
coordinate axes) that there exists r > 0 such that C(xo, r) fl an can
be represented as the graph of Lipschitz function f where the unit
exterior normal can be expressed as

(Di f,...,Dn, f,1)


^fl + ID fl 2
H"-a.e. on C(xo, r) n 011. With en+1 = (O,... ,1) and under the as-
sumption that apt u C C(xo, r), appeal to the Gauss-Green theorem
(see Theorem 5.8.7) to conclude that there exists a non-negative con-
stant C, depending only on the Lipschitz constant of f , such that

u dH" < C (uen +1) • v dH"


f n an
= C div(ue n+ t )dr.
n
<C IDuldz.
n
If u assumes both positive and negative values, write lui _ u+ + u -
to obtain

f812
IuI dH n < C n IDul dz.
Exercises

215

STEP 2. With no restriction on spt u, use a partition of unity to


obtain
lul dHn <_ C n (lul + IDuj)dx.
an
STEP 3. Prove that

lulPdHn Ç C (lulp + IDulp)d^


8S2 5

by replacing lui by lulp in the preceding step.


STEP 4. Now under the assumption that u E WLP(11), refer to Ex-
ercise 2.17 to obtain a sequence of smooth functions u,E such that
l^ uk — ulli, p; 1t —' 0 and

fa I
iuk--uelpdH" —+ 0 as k, l oo .

The limiting function u' E I'(011) is called the trace of u.

4.3. Prove that u' obtained in the preceding exercise is equivalent to the
trace obtained in Remark 4.4.5.

4.4. Prove the following Poincaré -type inequality which provides an esti-
mate of the measure of nul > k} in terms of WWDull1. Let u E W"" 1 (B)
where B is an open ball of radius r and suppose IA is a measure of
total mass 1 supported in B n {x : u(x) = 0}. Then, if k > 0,

kl {x : lu(x)l > k }l < Cr IDuI + Cr" (I l * ^) •• 'Dui,


B s
where R 1 is the Riesz kernel (see Section 2-6). Hint: Choose x, y E B
with u(y) = 0 and obtain

iu(x)i <_ iu(x) — u(z)i+ u(z)


whenever z E B. An application of polar coordinates yields

u(s) < Cll r * (xs IDul)(x) + (x$ . 'DOM].

4.5. The technique in the preceding exercise yields yet another proof of
Corollary 4.5.3 which is outlined as follows. Let u E W 1 •p(B(r)) where
B(r) is a ball of radius r and let N = {x : u(x) = 0}. Let coo be a non-
negative smooth function with apt cp contained in the ball of radius
2r and such that cp is identically one on B(r). Select x E B(r) in
accordance with the result of Exercise 3.15 and define h = cp[u(x)—u].
Then, for each y E R" ,

V(v)u(x) = (P(v)ü(sr) + h(Y)-


216 4. Poincaré Inequalities—A Unified Approach

Recall from Theorem 2.6.1 that the operator J : r1(R") —■ W'.p(R")


defined in terms of the Bessel kernel gi by J(f) = g1 * f is an isom-
etry. Therefore, with ti as any non-negative Radon measure µ, an
application of Fubini's theorem yields

f h dA= f91 *(J -


'h)dµ (J -rh)• gt * Adz.

Thus, if is concentrated on B(r) n N and satisfies Ilgl * µll p r < 1, it


follows that

µ[$(r) n N]Iu(z)I <_ it.r'[v(u - u(x)))IIp < CIEç(u -- u(x)II,.


Taking the supremum over all such 1.4 leads to

Bi,p[B(r) n N)lu(x)I" < C r - p Iu(x) — u(y)Ipdy


f (2r)

+ IDulpdg .
B(2r)

Use Exercise 3.15 to estimate

Iu(x) - u(g)Ipdy
B(2r)
in terms of the norm of Du.

4.6. Poincaré's inequality states that if u E W h .p(SZ) and u vanishes on


a set N of positive Bi, p -capacity, then IIuIIp•;n < CIIDuIIP;n, where
C depends on Si and the capacity of N. In the event that more is
known about u, this result can be improved. Using the indirect proof
of Section 4.1, prove that if u E W 1 "p(0) is a harmonic function that
vanishes at some point z o E Si, then there exists C = C(z o , St) such
that
IIuIip•;a < CIIDuII p; ^
4.7. Lemma 4.2.2 is one of many interpolation results involving different
orders of derivatives of a given function. In this and the next exercise,
we will establish another one that has many useful applications. Prove
the following: Let g be a measurable function on R n , and let 0 < ar <
n,O<e<1.Then
1-E .
Ira£(g)(x)I <— C(M9(x)) (!a(191)(z))`,
where Mg is the maximal function of g. Refer to the proof of Theorem
2.8.4 and choose 6 in(2.8.4) as

act +
1.(191)(x)
Mg(x)
Historical Notes 217

4.8. Let f = !.(g), g > O. Prove the interpolation inequality


i
Ili a— /
k },,
IID k f II* < CIlf IlâlcrIlf
where k is any multi-index with 0 < Iki < ar, 1/r = lkI/ap+(1-Ikl)/s,
and p < a < oo. Use the previous exercise to prove

Iliac(9)11r <— CIIgIIP-`IIIQ(19l)I1:


where 1 < p < oo, 1/r = (1 - e)/p + E/s, p < a < oo. Then let
f = I (g) and observe that
Q

IDkf (W)I <— Icr—k(g)(x).

4.9. Prove the following as a consequence of Theorem 4.2.1. Let Sl C R"


be a bounded, connected, extension domain. Suppose rr E LP` (Il),
n 0, p > 1. Prove that there exists C = C(p, v 0) such that ,

iluflp^^ CIIDullp;n
whenever u E W hip(1l), fn uv dx = O.

4.10. When I C R"+' is a Lipschitz domain, Exercise 4.2 shows one way
of defining the trace, u' E I"(00), when u E W'"p(11), p > 1. Note
that

fa. l u•1pdN" < c(Iulp


n
+ IDulp)dz.

Let v E I^ (Oft), u O. Prove that there exists C = C(p, v, fl) such


that
IIuIIP:n < CILDuIIp;^
whenever u E W"p(S?), fafl u'v'dH" = O.

4.11. At the beginning of this chapter an indirect proof of the following


Poincaré inequality is given: if u E W'.p(1l) and u = 0 on a set of
positive measure S, then llullp:n < CllbullP;n. Show that essentially
the same argument will establish the same conclusion if it is only
assumed that u = 0 on a set of positive B1, p -capacity.

Historical Notes
4.1. Lemmas 4.1.3 and 4.1.4 provide the main idea that serves as the key-
stone for the developments in this chapter. They are due to Norman Meyers
(ME4} and many other results in this chapter, such as those in Sections 2,
5, and 6 are taken from this paper. It should be emphasized that Lemma
4.1.3 is an abstract version of the usual indirect proof of the basic Poincaré
inequalities.
218 4. Poincaré Inequalities—A Unified Approach

4.4. In Remark 4.4.5 an approach to the subject of trace theory on the


boundary is indicated which is based on the material in Chapter 3 con-
cerning the property of Sobolev functions being defined everywhere in the
complement of small exceptional sets. Another approach to this subject is
presented in [LM].
In the proof of Theorem 4.4.6 it is not necessary to use the Morse-Sard
theorem if we are willing to use the full strength of the "Boxing Inequality"
[GU] and not the version reflected in Lemma 5.9.3. The inequality in [GU]
states that there is a constant C = C(n) such that any compact set K c R"
can be covered by a sequence of balls {B(r ; )} such that

E rt.1-i < CH" --l (ax).


i=1
This inequality could be used to establish (4.4.6) if K is taken as E t and
by observing that aE t C v -1 (t) since tr is continuous.
4.5. The proof of the Poincaré inequality here is, of course, based on the
material in the previous sections, particularly Theorem 4.2.1. This proof is
contained in [ME4]. There are several other proofs of the Poincaré inequal-
ity including the one in [P] which is especially interesting.
4.7. All of the Sobolev-type inequalities discussed thus far are in terms
of inequalities defined on R". There also are similar inequalities that hold
for functions defined on submanifolds of R". For example, in minimal sur-
face theory, Sobolev inequalities are known to hold for functions defined on
submanifolds where the inequality includes a term involving the mean cur-
vature of the submanifold, cf. [MS]. In case of a minimal surface, the mean
curvature is O. Theorem 4.7.2 is a result of the same ilk in that the left
side involves integration with respect to a measure which can be taken
as a suitable Hausdorff measure restricted to some submanifold. However,
it is different in the respect that the right side of the inequality involves
the li-norm of the gradient relative to Lebesgue measure on R" and not
the norm relative to Hausdorff measure restricted to the submanifold. This
interesting result was proved by David Adams [AD2]. Theorem 4.7.4 states
that measures with suitable growth over all balls are elements of the dual
of W k ip(R"). Thus, Theorem 4.7.2 is closely related to (4.2.1).
Theorem 4.7.5 which yields a characterization of those measures in the
dual of Wk.P(R") is due to Hedberg and Wolff [HW] although the proof we
give is adapted from [AD7].
4.9. Inequality (4.9.1) is due to Meyers-Ziemer NZ] in case q = 1. The
proof for the case 1 < q < n/(n — 1) is taken from [MA3]. This inequality
is also established in Chapter 5 in the setting of BV functions, cf. Theorem
5.12.5. Corollary 4.1.5 is an observation that was communicated to the
author by David Adams. This result when applied to Theorem 4.5.1 yields
Historical Notes 219

more information if Lemma 4.1.4 were used. This is an interesting example


of the critical role played by the sharpness of a constant, in this instance, the
_,(
exponent of Bm k p A) in Theorem 4.5.1. Indeed, in the work of Hedberg
[HE2I, it was essential that the best exponent appear. He gave a different
proof of Theorem 4.5.1.
5

Functions of Bounded
Variation
A function of bounded variation of one variable can be characterized as
an integrable function whose derivative in the sense of distributions is a
signed measure with finite total variation. This chapter is directed to the
multivariate analog of these functions, namely the class of L' functions
whose partial derivatives are measures in the sense of distributions. Just
as absolutely continuous functions form a subclass of BV functions, so it
is that Sobolev functions are contained within the class of BV functions of
several variables. While functions of bounded variation of one variable have
a relatively simple structure that is easy to expose, the multivariate theory
produces a rich and beautiful structure that draws heavily from geometric
measure theory. An interesting and important aspect of the theory is the
analysis of sets whose characteristic functions are BV (called sets of finite
perimeter). These sets have applications in a variety of settings because of
their generality and utility. For example, they include the class of Lipschitz
domains and the fact that the Gauss—Green theorem is valid for them
underscores their usefulness. One of our main objectives is to establish
Poincaré-type inequalities for functions of bounded variation in a context
similar to that developed in Chapter 4 for Sobolev functions. This will
require an analysis of the structure of BV functions including the notion of
trace on the boundary of an open set.

5.1 Definitions
5.1.1. Definition. A function u E L'(0) whose partial derivatives in the
sense of distributions are measures with finite total variation in SZ is called
a function of bounded variation. The class of all such functions will be
denoted by BV (Sl). Thus u E BV (i2) if and only if there are Radon (signed
measures) measures p ] , P 2 , ... A n defined in 12 such that for i 1, 2, ... , n,
,

I Dpi+( 11 ) < oc and

for all coo ECo (S)).


f uD ; cod_fcod i(
5.1.1)

The gradient of u will therefore be a vector valued measure with finite


5.1. Definitions
221

total variation:

- sup{u divvdx : v=(v 1 ,...,v„) E Cp (fZ; R n ),


n
lv(x)l < 1 for x E f2} < oo. (5.1.2)
The divergence of a vector field y is denoted by div y and is defined by
div v = En 1 D i v; . Observe that in (5.1.1) and (5.1.2), the space Cô (11)
may be replaced by CVO). The space BV (0) is endowed with the norm

IIU IIBV = IIuIII,n + (5.1.3)

If u E BV(0) the total variation IlDulI may be regarded as a measure,


for if f is a non-negative real-valued continuous function with compact
support 12, define

IlDu!!( f ) = sup{udiv v dx : i, = (v i , . . . , v,^) E Co (S^; R n ),


c ^

Iv(x)1 < f (x) for x E sl }. (5.1.4)

5.1.2. Remark. In order to see that 'Dull as defined by (5.1.4) is in fact


a measure, an appeal to the Riesz Representation Theorem shows it is
sufficient to prove that 'Pull is a positive linear functional on C0 (l) which
is continuous under monotone convergence. That is, if { f; } is a sequence
of non-negative functions in C0 (0) such that fi j g for g E C0(12), then
rrDull(f ) I^D u^I( g), cf. (F4, Theorem 2.5.51. In order to prove that IlDuli
;

has these properties, let = Du and refer to (5.1.1) to see that p satisfies

u div cpdx = —f ço • dg

where 5o E Co (S2; R"). Therefore, we may write (5.1.4) as

IIDuII( f) = sup{f y • dp : U = (v 1 , ... ,v„) E Co(Si; R n )


lv(x)I < f(x) for x E S2}. (5.1.5)
To show that Pull is additive, let f, g E C0(0) be non-negative functions
and suppose y E Co(f2, R") is such that lvl < f + g. Let h = inf{ f, Ivl} and
define
!v(T)I 0 0
h(x) ^(^)
i ( )I
w(x) =

{
o Iv(x)1= a.
222 5. Functions of Bounded Variation

It is easy to verify that w E C0(f1) and Iv — WI = M — h < g. Therefore,


since Iwl = h < f

1 v•dµ=w•dµ+ (v—w).du
rz 0
f
IlDull(f) + IlDull(9)-
This implies that IlDull(f + g) < IlDuII(f) + IIDuII(g) . The opposite in-
equality is obvious and consequently it follows that IlDull is additive. It
is clearly positively homogeneous. It remains to show that it is contin-
uous under monotone convergence. For this purpose, let f, I g and let
y E Co(SZ, R") be such that M < g. Also, define h; = inf{ f,, M} and

hi(x) 1) Iv(x)I 0
w i (s) = IU(x)I
o Iv(x)I =0.

Note that wi E Co(f2), Iivil = h; < f,, and that Iv — w11 = IVI — h, 1 Q as
i —4 oo. Since Iv — w; I = M — h i < 214 Lebesgue's Dominated Convergence
Theorem implies

f v di.4 = Du - y = lim Du - w, < lim II Dull (hi).


s—.00 I--60o
By taking the supremum of the left side over all such y it follows that
IIDuII(9) _< lim,—.0.0 IIDuII(h , ). Since h i < g for all i = 1, 2, ..., we have
IIDuII(g) = lim,_00 II Dull (h,). This establishes that IIDuII is a non-negative
Radon measure on ft.
We know that the space of absolutely continuous u with u' E L' (R 1 )
is contained within BV(R 1 ). Analogously, in R" we have that a Sobolev
function is also BV . That is, W 1 " 1 (St) C BV (11), for if u E W 1,1 (11), then

I
"

udivv dx i — EDiuvdx
n i`1
and the gradient of u has finite total variation with

II Du II(s^) = f Dudx.

5.2 Elementary Properties of BV Functions


In this section we establish a few results concerning convergence properties
of BV functions. We begin with the following which is almost immediate
from definitions, but yet extremely useful.

5.2.1. Theorem. Let 11 C R" be an open set and u E BV (11) a sequence


5.2. Elementary Properties of BV Functions 223

of functions that converge to a function u in 14 0 ,01). Then

lirn inf ['Dui II (U) ?.IIDuII (U)


3-.00
for every open set U C S2.

Proof. Let v be a vector field such that v E Cp (U; R") and Iv(x)I < I for
T E U. Then

I udivv dx = limu ; divv dx < ZirninfIlDuiIl(U)•


J-9013 U i-60o

The result follows by taking the supremum over all such v. ❑

5.2.2. Remark. Note that the above result does not assert that the limit
function u is an element of BV(SZ). However, if u E L 1 (SZ) and we assume
that
sup{ I,DuiOM) : t = 1,2,. _.} < 00
then u E BV (S2). Indeed, if pp E Co (S2), and Du; is any partial derivative
of u i , then

lim cP Duidx _ — limu i Dcp dx u DSp dx


no ^ ft

and therefore

U,. u DSp da < sup Iol lim inf Wu; IRO) < oo.
{^ 00

Since Cr(l) is dense in the space of continuous functions with compact


support, we have that

Du((p) _ — ( uD,dx
n
is a bounded functional on COI). That is, Du is a measure on O.
Theorem 5.2.1 established the lower semicontinuity of the total variation
of the gradient measure relative to convergence in LL We now will prove
an elementary result that provides upper semicontinuity.

5.2.3. Theorem. Let {u 1 ) E BV(S1) be a sequence such that u i —. u in


LL(S1) and
fi ni 1 ID u, I I( 1l) IlDull(^)
1 00

Then,
lim sup IIDuII(U n n) < IIDuII(U n si)
224 5. Functions of Bounded Variation

whenever U is an open subset of 11.

Proof. Since V = f1 — U is open, it follows from Theorem 5.2.1 that


IlDuII(U) < limnf IlDuill(U)

IIDuII(V) <— liminf


i-0oc
IlDuill(V)•
But,
IlDull(v n s2) + IlDull(V) = = lim IlDuill (n)
..^^

> lim sup II n i2) + liminf II Dud! (V)


i^oo i-00
> limsup [1 Dud! (U n f2) + 1IDu11(V)• ❑
i-+00
In view of the last result and Theorem 5.2.1, the following is immediate.

5.2.4. Corollary. If {u,} E BV(0) is a sequence such that u, -i u in


LL(fl), limi-. 00 IIDu1II(f 1 ) = IlDul1(1), and 11Dull(tU) = co, where U is an
open subset of f1, then
lim Il Duill(U ) = I lDull(U)•
i^oo

5.3 Regularization of BY Functions


Here we collect some results that employ the technique of regularization
introduced in Section 1.6. Thus, for each e > 0, cp, is the regularizing kernel
and o f = u * cp c . From the proof of Theorem 1.6.1, it follows that if U Cc U,
and u E Li c (U), then lluell1;u < IIuII i; n for all sufficiently small e > 0. In
this sense, regularization does not increase the norm. We begin by showing
that a similar statement is valid when the BV norm is considered.

5.3.1. Theorem. Suppose U is an open set with U C S2 and let u E BV (f1).


Then, for all sufficiently small e > 0,

IIuCII$v(u) IIUIIBV(n)-
Proof. In view of Theorem 5.2.1, it suffices to show that IIDuclI(U) <
IIDull(i 2 ) for all sufficiently small e > O. Select v E C0 (U, Rtm) with Id < 1.
Choose rq > 0 such that {x : d(x,U) < rt} C f2. Note that Iv,' < 1 and
spt o f C {x : d(x, U) < q} for all small e > 0. For all such e > 0, Fubini's
Theorem yields

fu u ( div v dx = u ( div v dx ïu
(div v) E dx
R L

fnudi v vEdx < IlDull(u)•


5.3. Regularization of BV Functions 225

The result follows by taking the supremum over all such al. a

5.3.2. Proposition. Let u E BV(n) and f E Co (n). Then fu E BV (1l)


and D(f u) = D f u + f Du in the sense of distributions.
Proof. Let U be an open set such that apt f C U C U C SZ. Then,
u e f E C(U) with D(f u e ) = (D f )u, + f Du e at all points in U. However,
NE — uIl i ; u —► 0 as e —4 0. (Of course, we consider only those e > 0 for which
u 1 (x) is defined for r E U.) In particular, when considered as distributions,
u £ u. That is, u E —+ u in 0'(U) and therefore Du e —4 Du in . (U), (see
Section 1.7). Since f E C(U), it follows that f Du, —► f Du in 2'(U).
Clearly, (D f )u (D f )u in 2'(U). Finally, with the observation that
(

fu, — fu in i'(U) and therefore that D(f u e ) —> D(fu) in 0'(U), the
conclusion readily follows. D

We now proceed to use the technique of regularization to show that


BV functions can be approximated by smooth functions and thus obtain a
result somewhat analogous to Theorem 2.3.2 which states that C°°(11)fl{u :
Iluljk,p;c^ < oo} is dense in Wk'D(1). Of course, it is not possible to obtain
a strict analog of this result for BV functions because a sequence {ui} E
C°°(1) that is fundamental in the BV norm will converge to a function in
W"(SZ). However, we obtain the following approximation result.

5.3.3. Theorem. Let u E BV(f2). Then there exists a sequence fu l l E


C°° (Q) such that
lim
i—.00° i
lu ; — uldx = 0

and
l irn I D u= I I( o) ^
IlDull(^)
^ -•oo

Proof. In view of Theorem 5.2.1, it suffices to show that for every e > 0,
there exists a function v E E C°°(1) such that

1 lu, — v jdr < F and IIDuFII(n) <


f + e. (5.3.1)

Proceeding as in Theorem 2.3.2, let 11i be subdomains of S/ such that


SZ CC S2 i+1 and U,_ o ili = O. Since Mull is a measure we may assume,
;

by renumbering if necessary, that 1lDulI(S1 — n o ) < E. Let Uo = 11, and


U; = for i = 1, 2, .... By Lemma 2.3.1, there is a partition of
unity subordinate to the covering Ui = 52 ;+ , — S2^_ a i = 0, 1, .... Thus,
,

there exist functions L such that f, E Co (Ui), 0 < fi < 1, and E°_° o fi i 1
on h. Let yo, be a regularizer as discussed at the beginning of this section.
Then, for each i there exists ei > 0 such that
spt(( fiu) f , ) C Ui, (5.3.2)
226 5. Functions of Bounded Variation

I(liu)E, — fiuldx < E2 —(i + 1) (5. 3.3)


n

I(uDf,)E, — uD f,Idx < c2 - " 1) . (5.3.4)


n
Define
00

vc = E(ufi ) ^ :

i =o
Clearly, y r E C° (1) and u = EMI ufi . Therefore, from (5.3.3)

I Iv e — uldx <

Reference to Proposition 5.3.2 leads to


E I(ufi)e, ufildx < E.
$-o 11
(5.3.5)

Dv, = E(fiDu)E, +E(uDfi)E,


1 —o i= o
00 00

E(f1Du)E, + ERuD fi)^ — uD fi). :

i=0
+=o
Here we have used the fact that E D f, = 0 on ft. Therefore,
00 00
I I Du E Idx <
i =i f
I(f,Du)E,Idx + ^ I(uD f,) E , — uD f,ldx.
i=o

The last term is less than E by (5.3.4). In order to estimate the first term,
let V E Cp (ft; R") with sup I'iI < 1. Then, with w e *

IfJ ÇO c * (fi Du) dx = f f tI E fid(Du) by Fubini's theorem,

= udiv(t f fi)dx

<— IIDuII(Ui)
since spt rj' fi C U; and 11/),L1 < 1. Taking the supremum over all such 11)
y ields

J I(fiDu),, Idx < IIDuII(Ui),


Therefore, since each x E SZ belongs to at most two of the sets fl„
_ = 0, 1 , . . . .

I
00

IDvEIdx < E IiDuII(U;} + e


i -o
CO

<_ IIDuII(^}) + E IIDuII(Ui) + E


s =1

< Ilb^lll(^r) + 2 IIDuII(ll — fa ) + e


-< IlDu ll(SZ) + 3 e.
5.3. Regularization of BV Functions 227

Since e > 0 is arbitrary, this along with (5.3.5) establishes (5.3.1). ❑

5.3.4. Corollary. Let 1 c R" be a bounded extension domain for W 13 (1).


Then BV(f) n {u : By < 1} is compact in L'(1).

Proof. Let u ; E BV(11) be a sequence of functions with the property that


< 1. By Theorem 5.3.3 there exist functions v, E C°°(I) such that

l v, - u,Idx < i` 1 and


j lDviIdx< 2 .
n
Thus, the sequence {IIvill,,l;n} is bounded. Then, by the Rellich-Kondra-
chov compactness theorem (Theorem 2.5.1), there is a subsequence of {v,}
that converges to a function u in L' (1). Referring to Remark 5.2.2, we
obtain that v E BV(f ). ❑

In Theorem 2.1.4 we found that u E W'.P if and only if u E LP and u has


a representative that is absolutely continuous on almost all line segments
parallel to the coordinate axes and whose partial derivatives belong to L".
We will show that a similar result holds for BV functions.
Since we are concerned with functions for which changes on sets of mea-
sure zero have no effect, it will be necessary to replace the usual notion of
variation of a function by essential variation. If u is defined on the interval
[a, the essential variation of u on la, bJ is defined as

ess Vâ (u) = sup 1E lu(ti) — u(ti- i )f


i= 1

where the supremum is taken over all finite partitions a < to < t i ... tr < b
such that each t i is a point of approximate continuity of u. (See Remarks
3.3.5 and 4.4.5 for discussions relating to approximate continuity.) From Ex-
ercise 5.1, we see that u E BV (a, b) if and only if ess Vâ (u) < oo. Moreover,
ess Vâ (u) = IIDull [(a, b)). We will use this fact in the following theorem. As
in Theorem 2.1.4, if 1 < i < n, we write x = (±,x,) where i E R"' and
we define u;(x,) = u(l x i ). Note that u = depends on the choice di. but for
,

simplicity, this dependence will not be exhibited in the notation. Also, we


consider rectangular cells R of the form R = (a l , b 1 ) x (a2, b 2 ) x • • . x (a,,, b").

5.3.5. Theorem. Let u E LL(R"). Then u E BV oc (R") if and only if

ess V^ ^ (u, )dz < o0


JR
for each rectangular cell R C R" ', each i = 1, 2, ... , n, and a, < b,.
-

Proof. Assume first that u E BV1Qc (R"). For 1 < i < n it will be shown
228 5. Functions of Bounded Variation

that

ess Vb ' (ui ) dx < 00

for each rectangular cell R C Rn - 1 and at < 14. For notational simplicity,
we will drop the dependence on i and take R of the form R = R x [a, b].
Now consider the mollified function u E * u and note that

1 lu E — uldx —+ a as

and
limsup 1Du r ds < oo (Theorem 5.3.1).
E^ 0 R
Consequently, with u,,, (x,) u E (x, x,), it follows that u,-, i u, in L' (a, b)
for H" -1 -a.e. E R. Theorem 5.2.1 implies that l m inf e . o ^^Du^^^[(a, b)] >
`I Du; II [(a, b)] and therefore, from Exercise 5.1,

ess Vb (u ; ) < lim inf eas Vb (u E ,;)


E-0
for E R. Fatou's lemma yields

f ess V (ui)dH"
b -1 (2) < lirninf
E
R
ess Vâ (u^.,,)dHn -1 (x)

= lirn inf
E-6 o R

< hm sup DuIdx < oo.


E -o R

For the other half of the theorem, let u £ Lloc (R") and assume

ess V6 (u, )dH" -1 (i) < 00


R

for each 1 < i < n, a < b, and each rectangular cell R C R" -1 . Choose
cp E G (R), kol < 1, where R R x (a, b) and employ Exercise 5.1 to
obtain
u Dog dx f ess ab (u;)dHn -1 (i) < oo.
RR

This shows that the partial derivatives of u are totally finite measures over
R and therefore that u E B V oc (Rn ). O

5.4 Sets of Finite Perimeter


The Gauss-Green theorem is one of the fundamental results of analysis
and although its proof is well understood for smoothly bounded domains
5.4. Sets of Finite Perimeter 229

or even domains with piece-wise smooth boundary, the formulation of the


result in its ultimate generality requires the notion of an exterior normal to
a set with no smoothness properties in the classical sense. In this section,
we introduce a large class of subsets of R" for which the Gauss-Green
theorem holds. These sets are called sets of finite perimeter and it will be
shown that they possess an exterior normal which is defined in the same
spirit as Lebesgue points of LP-derivatives. The Gauss-Green theorem in
the setting of sets of finite perimeter will be proved in Section 5.8.

5.4.1. Definition. A Borel set E C R" is said to have finite perimeter


in an open set f provided that the characteristic function of E, XE, is a
function of bounded variation in 0. Thus, the partial derivatives of XE are
Radon measures in f and the perimeter of E in 0 is defined as

P(E , n) = IIDXEII(1)•
A set E is said to be of locally finite perimeter if P(E, SZ) < oo for every
bounded open set O. If E is of finite perimeter in R", it is simply called a
set of finite perimeter. From (5.1.4), it follows that

P(E, S?) = supdiv v dx : y = (v 1i . . . , ^n ^ E ^D (^s R^)+ Iv(z)1 Ç 1 •'


E
(5.4.1)
5.4.2. Remark. We will see later that sets with minimally smooth bound-
aries, say Lipschitz domains, are of finite perimeter. In case E is a bounded
open set with C 2 boundary, by a simple application of the Gauss Green
theorem it is easy to see that E is of finite perimeter. For if v E Co (1l; R")
with I1vII < 1, then

divv dx ^ U • v dH" -1 < H (II n 0E) < oo


JE 8E
" - `

where v(x) is the unit exterior normal to E at x. Therefore, by (5.4.1),


P(E, Il) < oo whenever 0 is an open set.
Moreover, it is clear that P(E, S2) = n 8E). Indeed, since E is a
C 2 -domain, there is an open set, U, containing OE such that d(x) d(x, E)
is Cl on U - OE and Dd(x) = (x - (x)) jd(x) where (x) is the unique
point in OE that is nearest to x. Therefore, the unit exterior normal v to
E has an extension t E Co(Rn) such that PI < 1. Hence, if v = rjv with
q E Co (S2), we have,

div v dx - div rjv dx _


LE
This implies

P( E, SZ) > sup rtdHn - 1 n E Co (f), IqI < 1


8E
H" -1 (1-2 f? OE).
230 5. Functions of Bounded Variation

Intuitively, the measure DXE is nothing more than surface measure


(H"' 1 -measure) restricted to the boundary of E, at least if E is a smoothly
bounded set. One of the main results of this chapter is to show that this
idea still remains valid if E is a set of finite perimeter. Of course, since we
are in the setting of measure theory, the topological boundary of E is no
longer the appropriate object of study. Rather, it will be seen that a subset
of the topological boundary, defined in terms of metric density, will carry
the measure DXE.
In Theorem 2.7.4 we observed that the isoperimetric inequality lead to
the Sobolev inequality via the co-area formula. Conversely, in Remark
2.7.5 we indicated that the Sobolev inequality can be used to establish
the isoperimetric inequality. We now return to this idea and place it in
the appropriate context of sets of finite perimeter. We will establish the
classical isoperimetric inequality for sets of finite perimeter and also a local
version, called the relative isoperimetric inequality.

5.4.3. Theorem. Let E C R" be a bounded set of finite perimeter. Then


there is a constant C = C(n) such that
EI(n-1)/n <
CIIDXE II(R") = CP(E). (5.4.2)

Moreover, for each ball B(r) C R",

min {fB(r) n El, IB(r) - E)I} ( " -1)/n < CIIDXEII(B(r)) = CP(E, B(r)).
(5.4.3)
Proof. The inequality (5.4.2) is a special case of the Sobolev inequality for
BV functions since XE is BV. We will give a general treatment of Sobolev-
type inequalities in Section 11. If u E BV (R" ), refer to Theorem 5.3.3 to
find functions u E Co (B") such that
;

lim f lu - ; uldx = 0 ,

1.-yoo

li m Du, 11(R" ) = IlD u ll(R" ).


t-y 00
By passing to a subsequence, we may assume that ui -' u a.e. Then, by
Fatou's lemma and Sobolev's inequality (Theorem 2.4.1),

llull "/(n-^ ) < liminf


--• i
Iluilln/(n-1)
00

< lim CIIDu1II(Rn)


< C`IIDulI(R").
To prove the relative isoperimetric inequality (5.4.3), a similar argument
along with Poincaré's inequality for smooth functions (Theorem 4.4.2),
yields
Ilu - tl(r)1Ift/("-1):B(r) Ç C' llDu ll($(r))
5.4. Sets of Finite Perimeter 231

where u(r) = fÈ(r) u(x)dx and B(r) is any ball in Rn. Now [et u = X E and
obtain
n1(n-1)
n-^) IB(r) n El
lu(x) — dz Ef
B(r) ^(T)In/( (IB(r)
— IB(r)l
n fl/(n1)
IB(r) El
)

+ IB(r) — El.
IB(r)I
If IB(r ) — El > IB(r) n El, then (IB(r) — EIV(IB(r)I) and

CIIDX&II(B(r)) = CIiDuII(B(r)) ? Ilü -- ri(r)Iin/(n-1);$(r)

(
IB(r) n EI(n.-1)/n
^

IB(r)I
lB(r) —El \
(n -1)/n
1 min B(r) n El ^ IB(r) — El
> f
2 IB(r)I IB(01
A similar argument treats the case IBM n Et > IB(r) — El. ❑

We now return to the topic of the co-area formula which was proved
in Theorem 2.7.1 for smooth functions. Simple examples show that (2.7.1)
cannot hold for BV functions (consider a step function). However, a version
is valid if the perimeters of level sets are employed. In the following, we let

Et =S2 n {x : u(x) > t}.

5.4.4. Theorem. Let S2 C Rn be open and u E BV(fI). Then

MuII( 11 ) _ ^f DXEg !I (s^)dt.


R'

Moreover, if u E L' (S2) and Et has finite perimeter zn S2 for almost all t
with
II DX Es II ( 0 )dt < oo,
JR i
then u E BV (1Z).

Proof. We will first proof the second assertion of the theorem. For each
t E R', define a function ft : R n --► R' by

}e if = t 70
—X Rn _E t if t < 0.

Thus,
u(x) = Ît(x)dt, a E R^.
R^
232 5. Functions of Bounded Variation

Now consider a test function cp E Co (S2), such that sup kol < 1. Then

u(x)Sp(x)dx = ft(x)cp(x)dtdx
f „ f^ RI
ft(x)Sp(x)dxdt. (5.4.4)
=1...
Now (5.4.4) remains valid if Sp is replaced by any one of its first partial
derivatives. Also, it is not difficult to see that the mapping t --> I DXE, 11( 2 )
is measurable. Therefore, if coo is taken as cp E Co (SZ; Rn) with snp Icpl < 1,
we have

Du(cp) _ --u • div cp dx = — ft(x)div cp(x)dxdt


f n R ' J^

< D ft (cp)dt < IIDXE,11(12)dt < co. (5.4.5)


Rz R1
However, the sup of (5.4.5) over all such cp equals IIDuII(SZ), which estab-
lishes the second assertion.
In order to prove the opposite inequality under the assumption that
u E BV(1), let {Pk} be a sequence of polyhedral regions invading St and
Lk : Pk --+ R 1 piecewise linear maps such that

lim 1Lk — uldx = 0 (5.4.6)


k-oao Pk

and
lim
k--oo
I D L kIdx = 11Dull(n), (5.4.7)

(see Exercise 5.2). Let

E^ = Pk n (x . Lk(x) > t),


xt = —Et •
From (5.4.6) it follows that there is a countable set S C R' such that for
each j = 1, 2, .. .
limF IX t (x) — Xt (x)Idx = 0 (5.4.8)
k~cio
i
whenever t it S. Thus, for t S, and E > 0, refer to (5.4.1) to find
W E Co (SZ; R”) such that Icpl < 1 and

IIDXE, II(Sa) — f div cp dx < 2 . (5.4.9)


r

Let M = fR „ Idiv cpldx and choose j such that spt cp C P) . Choose k o > j
such that for k > k o ,
E
IXt -- Xt l dx< .
Pt 2M
5.5. The Generalized Exterior Normal 233

For k > !c o ,

f ,
div cp di -
E^
div cp dx
<MP
IXt —Xtldx< e .
^
i
(5.4.10)

Therefore, from (5.4.9) and (5.4.10)

II DXE, 1I (i) < div Sp dx + €


Et
_ IIDXE: II(fl)+ e.
<
Thus, for t S,
IIDXE, II(n) < lku inf IIDXE; II (n)•
Therefore, Fatou's lemma implies

o IIDxE, II(n)dt < lkrn inf Ri DX E .:, I((SI)dt


R
<licn inf H " -1 [L k 1 (t ) n fZ)dt (by Remark 5.4.2)
k —• oo R'

_ lim inf
<
k
Pk
I DLkldx (by (2.7.1))

= (by (5.4.7)). ❑

5.5 The Generalized Exterior Normal


In Remark 5.4.2 we observed that a smoothly bounded set has finite perime-
ter. We now begin the investigation of the converse by determining the reg-
ularity properties possessed by the boundary of a set of finite perimeter.

5.5.1. Definition. Let E be of locally finite perimeter. The reduced bound-


ary of E, 0 - E, consists of all points x E R' for which the following hold:

(i) IIDXEIU[B(x, r)1 > 0 for all r > 0,

(ii) If v,.(x) = -DXE[B(x, r))/IIDXEIj[B(x, r)], then the limit v(x) _


v,, (x) exists with Iv(x)I = 1.
v(x) is called the generalized exterior normal to E at x. We will employ
the notation v(x) = w(x, E) in case there is a possibility of ambiguity. The
notation ô` is used in 0 - E to indicate that the normal to E is pointing in
the direction opposite to the gradient.
Observe that v(x, E) is essentially the Radon-Nikodym derivative of
DXE with respect to IIDXEII. To see this, let p(x) be the vector-valued
234 5. Functions of Bounded Variation

function defined by

DXE[ (x, r)]


,r )^
P(x) - - lim t .
r POIIDXE IIIB(x
-

From the theory of differentiatiou of measures in Chapter 1 (see Remark


1.3.9) this implies that p is the Radon- Nikodym derivative of DXE with
respect to II DXE II and that

DXE(B) = —JB P(1H

for all Borel sets B C Fr. Moreover,

I dive dx = - f v(x) p(x)dIIDXEII

whenever y E Cp(Rn; Rn). Consequently, by (5.1.2), Ip(x)I = 1 for IIDXEII -

a.e. x E R" and therefore, p(x) = v(x, E) for IIDXEII-a.e. x E R". Thus,
we have
DXE(B) = - Y(x, E)4IDXEII(x),
Bna -E
IIDXEII(R" - ô E) = O.
The next lemma is a preliminary versiou of the Gauss-Green theorem.

5.5.2. Lemma. Suppose E is of locally finite perimeter and let f E Co (R" ).


Then, for almost all r > 0,

Al dx = - f d(D1XE) + f (11)vi (V, B( r ))din- I (y)


EnB(r) B (r) En8(B(r))

where B(r) = B(x,r) and v,(y, B(r)) is the i th component of the unit
exterior normal.

Proof. To simplify notation, we will take x = O. From Proposition 5.3.2,


we have that fXE E BV(1 ). Let S be the countable set of r such that
II DI(fxE)II [a(B(r))] O. Select r S and let of be a piecewise linear
function on (0, oo) such that rt e z- 1 on (0,r] and rk =_
0 on (r + e, oo).
Since Di [ f X E] is a measure, we have

f (x)XE(x)Difrje(IxI)1dx = — lie(IxI)d(Ds[fXE])(x)
R"
- Di(fxE)[B(r)]

f(r
ne(I xI )d(Di[f xE])( x).
+i) -B(r)
5.5. The Generalized Exterior Normal 235

Therefore
1f
J D(r+)_B(r)
f (x)XE(z) dx
1x1

= - Di(fXE)[B(r)] - vk(IxI)d(D [fXEJ)(x)•


f(r+c)-B(r)
Since r S and IrJI < 1, the integral on the right converges to 0 as e J. O.
By the co-area formula (Theorem 2.7.3), the integral on the left can be
expressed as

1
^
JB(r+c)—B(r) I
f (z)XE(x) x dx
1
f ?+C
+ dHn1(x)dt.
Éf(x)
n8(B(r)) I
I I
Therefore

f (x) ^ dHn-1(x)
fBfr+c)—B(r) f (x)XE(x) 2t dZ ^ -
Ii L8(B(r)) ^

which implies

f (x) Î I dH n-1 (x)


= Di (f XE)[B(r)1
fn8(B(r))
for almost all r > O. Moreover, from Proposition 5.3.2,

D If XEJ(B(r)) = (Dif)XE(B(r)) + f DiXE(B(r))


D i f (z)dx + f d(DtXE). ❑

f f1B(r) B (r)

5.5.3. Corollary. If E has finite perimeter in ii, then for almost all r > 0
with B(r) C SZ,

PE fl 8(0,f2) < P(E, B(r)) + H' 1 [E n a(B(r))].

E
Proof. Choose Co (S1, R") with lvi < 1 and let r > 0 be a number for
which the preceding lemma holds. Then

div v dx - - v • d(DXE)
f f1B(r) B(r)

+ v(x) , 1.(xf B(r))dHn-


(x)
J fZâ(B(r))
Ç IIDXEII(B(r)) + H" '[E - n ô(B(r))j.
236 5. Functions of Bounded Variation

Taking the supremum over all such u establishes the result. ❑

Remark. Equality actually holds in the above corollary, but this is not
needed in the immediate sequel.
The next lemma will be needed later when we begin to investigate bound-
ary regularity of sets of finite perimeter.

5.5.4. Lemma. Let E be a set with locally finite perimeter. Then, for each
z € 0 - E, there is a positive constant C = C(n) such that for all sufficiently
small r > 0,
r' IB(x, r) n E > C, (5.5.1)
r -n IB(s, r) - El ? C, (5.5.2)
C S ri
-n IIDXEII(B(x, r)) 5C - l . (5.5.3)

Proof. To simplify the notation, we may assume that ï = O. Since 0 E


8 E, there is a positive constant C = C(n) such that
-

li/T( 0 )I = IDXE($(r))IIIIDxEII[B(r)J ? C (5.5.4)

for all small r > O. For almost all r > 0, it follows from Lemma 5.5.2 that

s
DXE(B(r)) - ldHn`1 (s)
fEn5(.(R), Ix
and therefore
DXE(B(r))I < H" -1 [E n O(B(r))J.
Consequently, (5.5.4) implies

IIDXE 1(B(r)) < C -1 1P


-L
[E n ô(B(r))) < C-1rn-1. (5.5.5)

Note that (5.5.5) holds for all small values of r since the left side is a
left-continuous function of r. This establishes the upper bound in (5.5.3).
To establish (5.5.1), recall from Corollary 5.5.3 and (5.5.5) that for almost
all r > 0,

P(E n B(r)) < P(E, B(r)) + Nn-1[E n 8($(r)))


and
P (E,B ( r )) < C-1Hn-1[E n 0(B(r))]•

Thus, an application of the isoperimetric inequality (Theorem 5.4.3) and


the previous two inequalities lead to

IE n B(r)I ( n -1) in < CP(E n B(r)) < CH" -1 (E n 8(B(r))1,


5.6. Tangential Properties of the Reduced Boundary 237

for some constant C = C(n). Let h(r) = 1E n B(r)I and observe that the
co-area formula (Theorem 2.7.3) yields
T

h(r) = ID(Ix[)Idx = Hn -1 [E n a(B(t)))dt.


E rZB(r) o

Hence, h i (r) > Ch(r) ( " -1 )/" and therefore that h(r) (1 /" )-1 h'(r)
n(h l /"(r))' > C. This implies h(r) 1 /n > Cr, thus establishing (5.5.1).
Note that (5.5.1) implies (5.5.2) since P(E) = P(Rn - E) and v E
VRP E
The lower bound in (5.5.3) follows immediately from (5.5.1), (5.5.2), and
the relative isoperimetric inequality (Theorem 5.4.3)

- EI ("-1)/"
IIDXEII(B(r)) > Crain IB(r) n EI 'BO-)
(

r n - I rn rn

5.6 Tangential Properties of the Reduced


Boundary and the Measure-Theoretic Normal
Now that we have introduced the definition of the unit exterior normal to a
set of finite perimeter, we ask whether the existence of the exterior normal
implies some type of regularity of the boundary. In order for the theory
to run parallel to the classical development, the hyperplane orthogonal to
the generalized normal in some sense should be tangent to the reduced
boundary (see Definition 5.5.1). Although it cannot be expected that this
plane is tangent in the usual sense, it will be shown that it is so in the
measure-theoretic sense.
For this purpose, we will employ a "blow-up" technique which views the
local behavior of a set at a point by examining a sequence of dilations of the
set at the point. Specifically, let E be a set of locally finite perimeter and
suppose for notational simplicity that 0 E rÎ - E. For each c > 0, consider
the dilation T(x) = x/E and let EE = Te (E). Note that XE, = XE 0 T'
and that the scaling of DXEI is of order n - 1. That is,

DXE c [B(r/e)} = 6 1- "DXE[B(r)] for r > 0


(5.6.1)
IIDXE^ ii[B(r/e)1 - e 1-n IIDXEII[B(r)) for r > O.
The proof of the second equation, for example, can be obtained by choos-
ing a sequence (u 1 ) E CO° [B(r)) such that u ; XE in L 1 [B(r)] and
fB(r) ID%
Idx -. IIDXEII [B(r)] (Theorem 5.3.3). However,

IDu1,eIdx = E 1-" 'Dui PS


B(r/e) f (r)
238 5. Functions of Bounded Variation

where u;, e = u ; o TE 1 , Then, u; E -► XE o TT 1 = XE, in L 1 IB(E/r)I and by


Theorem 5.2.1,

lim inf W us e ldx > IIDXE, II[B(r/e)].


'~°° J (*/e)
Hence,
l-n
IIDXEE II[B(r/e)] < e II D XEI [B( r )] .
The reverse inequality is obtained by a similar argument involving a se-
quence of smooth function approximating XE a T' .

5.6.1. Definition. For x E 8 - E, let r(x) denote the (n -1)-plane orthog-


onal to v(x, E), the generalized exterior normal to E at x. Also, define the
half-spaces
N +(x) = {y : v(x) • (y - x) > 01
H (x) = {y : v(x) . (y - x) < 0).
-

5.6.2. Theorem. If E is of locally finite perimeter and 0 E 8 - E, then

XE, —► XH- in LÎoc(R') as e 0 .

and
IIDXOI(U) -, IIDXH- II(U)
whenever U is a bounded open set with 11 11-1 [(8U) fl 7r(0)1 = O.

Proof. Without loss of generality, we may assume that the exterior normal
to E at 0 is directed along the x n -axis so that v n (0) = 1 and v i (0) = ... =
Vn -1(0) = O. It is sufficient to show that for each sequence {e } -+ 0, there ;

is a subsequence (which we denote by the full sequence) such that

IXE^ , - XH- Idx --; 0 and IIDXE., 11(U) IIDXH- 11(U) (5.6.2)

as e; O.
From (5.6.1) and (5.5.3) we obtain for each r > 0,
- 1 = C-t r n-1 1
IIDXE. II[B(r)] =
e l-
nIID^ E II[B(er)) < C -l e l-n( er )n
(5.6.3)
and
e t-n e l-n (er') n-1 = Crn-1 (5.6.4)
IIDXE, II[B(r)] - IIDXEIIIB(er)) ^ C

for all sufficiently


ffi small e > O. Thus, for each B(r), Crn -1 <_
IIXEc II ev(s(r ^}

< C -1 rn -1 for all sufficiently small e > 0. Therefore we may invoke the
compactness of BV functions (Corollary 5.3.4) and a diagonalization pro-
cess to conclude that XE 1i —+ XA in LL(Rn). For each bounded open
5.6. Tangential Properties of the Reduced Boundary 239

set S 2, XEc. --0 XA in 2'01) (in the sense of distributions) and therefore
-

DXE L DXA in 9'(S3). Note that DXA 0 from (5.6.4). Moreover


DXE C, -i DXA weakly in the sense of Radon measures an d therefore, for
all but countably many r > 0,

DXE1, [B(r)) --; DXA[B(r)1. (5.6.5)

From (5.6.1), and the definition of the generalized exterior normal,

lim DiXEC[B(r)#/IIDXEr II[B(r)1


E^0

= Jim DIXE[B(Er)1/I1 DXEII[B(ir)1 = o , i = 1,2, ... , n - 1, (5.6.6)


e-s0
whereas
li ô DnX E, [B(r)1/IIDXE, 1I [B(r)j = -1. (5.6.7)

Thus, from (5.6.7) and (5.6.3),

lira IIDXE, I1[B(r)1 = - lim D n X E. , [B(r)1 = -D n XA[B(r)1. (5.6.8)


i^oo ■-•oo
From the lower semicontinuity of the total variation measure (Theorem
5.2.1) we obtain

l =m ôô f II DXE^, II [$(r)] >- II DXA II [B(r)1

and therefore IIDXAII[B(r)1 < -D„X A [B(r)) from (5.6.8). Since the oppo-
site inequality is always true, we conclude that

IIDXAIIIB(r)1 = - D„XA[B(r)] (5.6.9)

for all r > 0_ Therefore, by Theorem 1.3.8 and Remark 1.3.9,

IIDXAII(B(r)) = DnXA(B(r)) = v,(x,A)d1IDXAII(x)-


B (r)
-

This implies that v. (x, A) = 1 for IIDXAlI-a.e. x and thus that v v (x, A) = 0
for IIDXA II a.e. x, i = 1, 2,... ,n - 1. Consequently, we conclude that the
-

measures D i XA are identically zero, i = 1, 2, ... ,n - 1. Hence, XA depends


only on x,^ and is a non-increasing function of that variable. Let

a = sup{x n : XA (z) = 1}.


Since DXA # 0, we know that A oo. The proof will be completed by
showing that A = 0. If A < 0, we would have B(r) C Rn - A for r < Al
I
and since XE, —■ XA in LL(R"),

0 = 18(r) n Al — s lim IEe , n B(r)I = i11m a nI B(re ; ) n El


= lim r"(rei) 'IB(re;) n El
-
240 5. Functions of Bounded Variation

which contradicts (5.5.1). A similar contradiction is reached if A > O. There-


fore, A = H and by (5.6.8) and (5.6.9),
-

lim
t-400 I1DxE„ II [B(r)1 = IIDrIxN - 1I[B(r)}

for all but countably many r > 0. If U is an open set with U C B(r) for
such an r > 0 and

II DX H - II(aU) = H n-1 [ir(0) n au] o,


then Corollary 5.2.4 implies IIDXE.,11(U) IIDXH- 11(U).
We now will explore the sense in which the hyperplane 7r(x) introduced in
Definition 5.6.1 is tangent to a E at r. For this, we introduce the following.
-

5.6.3. Definition. Let y E Rn with Ivl = 1. For r E Rn and e > 0, let


C(x, E , v) = R" n { y = I(y - x) - vi > Ely - xI}.
In what follows, it will be clear from the context that both r and v are
fixed and therefore, we will simply write C(e) = C(x, e, w).
C(e) is a cone with vertex at x whose major axis is parallel to the vector
v. If M were a smooth hypersurface with v normal to M at x, then for
each e> 0
C(e) n M n B(x, r) = 0 (5.6.10)
for all r > 0 sufficiently small. When M is replaced by 0 E, Theorem 5.6.5
-

below yields an approximation to (5.6.10).


Before we begin the proof of Theorem 5.6.5, we introduce another con-
cept for the exterior normal to a set. This one states, roughly, that a unit
vector n is normal to a set E at a point r if E lies completely on one (the
appropriate) side of the hyperplane orthogonal to n, in the sense of metric
density. The precise definition is as follows.

5.6.4. Definition. Let E C B n be a Lebesgue measurable set. A unit


vector n is called the measure-theoretic normal to E at x if

limr "IB(x,r)n {y:(y—ac)-n<O,y0E}I = 0


-

r-•0

and
lirnr'IB(x,r) n {y: (y —x)-n> 0, y E E}I =
The treasure-theoretic normal to E at x will be denoted by n(x, E) and we
define
a'E = {x : n(x, E) exists).
5.6. Tangential Properties of the Reduced Boundary 241

The following result proves that the measure theoretic normal exists
whenever the generalized exterior normal does. Thus, 0 E C â' E. -

5.6.5. Theorem. Let E be a set with locally finite perimeter. Suppose


0 E 8 - E. Let y be the generalized exterior normal to E at 0 and 7r(0) the
hyperplane orthogonal to v. Then,

Ern r "-IIIDXEII[G(e) n B(r)} = 0, (5.fî.11)


r ^o

lim r - nIE n H+ n B(r)I = 0, and (5.6.12)


r—.0

lim r -n I(B(r) -- E) n H -I - 0. ( 5. 6.13)


r —.0

Proof. Again we use the "blow-up" technique that was employed to obtain
(5.6.1). Thus, let Tr (x) = x/r and recall that

IIDXErII[ 3 ( 1 )] = r 1- "IIDXEII[D(r)}.
Note that Tr [C(E) n B(r)) = C(e) n B(1). Therefore
l-
1I 1 XErII1'(e) n B(1)) = r "I1DxEII[G(e) n B(r)),
and by Theorem 5.6.2,

T1—"IIDXEII[C(E) n B(r)J -. Hn 1 [C(e) n B(1) n 7r(0)) = O.


This proves (5.6.11).
Similarly,
r-nIE n B(r) n H + I = 'ET n B(1) n H + I
and since XE,. —. XH- in LL(Rn) (Theorem 5.6.2),

li mIEr nB(1)nH+I=1H - n B(1) nH+1=0.

This establishes (5.6.12) and (5.6.13) is treated similarly. D

The following is an easy consequence of the relative isoperimetric in-


equality and complements (5.5.3).

5.6.6. Lemma. There exists a constant C = C(n) such that

DEII
Ix Bxr( )1 > C
liminf
r—.o rn-1

whenever x E a'E.

Proof. Recall from Definition 5.6.4 that if x E 9'E, then


11mo r - n1B(x, r) n E n H+(z)1 = 0
242 5. Functions of Bounded Variation

and
tinâ r-NB(x,r) -- El n H (x)1 = ü
-

where H+ (x) and H (z) are the half-spaces determined by the exterior
-

normal, n(x, E). Since B(x, r)nH (x) = ([B(x, r)-EjnH (x))u(B(x, r)n
- -

E n H (x)), the last equality implies that


-

IB(r) n Ej 1B(x, r) n E n H- (x)1 1


lim inf I > lim = 2- .
r-00 I B (r r~0 IB(Z, r)I
Similarly,
IB(r) - El 1
lim i nf >
r I B(r)I - 2
and consequently,

lirn
IB(r) n EI
_ r-P0
hill
- El _1
r^0
IB(r)I IB(r)I 2-

The result now follows from the relative isoperimetric inequality (5.4.3). ❑

This result allows us to make our first comparison of the measures IIDXEII
and H n- ' restricted to û'E.

5.6.7. Theorem. There is a positive constant C such that if E is a set


with locally finite perimeter, and B C as is a Bord set, then

H" '(B) < CIIDXEII(B)•


-

Proof. For each x E B we obtain from Lemma 5.6.6 that


liminf
r-4)
II DXEI ^[B(x, r)j > C
r
Our conclusion thus follows from Lemma 3.2.1.

5.6.8. Corollary. If E is a set with locally finite perimeter, then

(0 '* E - ô E) - O.
H " (ô - (5.6.14)

Moreover, IIDX E II and the restriction of Hn ' to i'E have the same null
-

sets.

Proof. From the discussion in Definition 5.5.1, we have that IIDXEII(R" -

8 E) = 0 and therefore IIDXEII (a' E - 0 E) = O. Thus, (5.6.14) follows


- -

from the previous theorem. Moreover, if B C S E with H" -1 (B) = 0, then


-

IIDXEII (B) = 0 because of the second inequality in (5.5.3). This establishes


the second assertion.

5.7. Rectifiability of the Reduced Boundary 243

5.7 Rectifiability of the Reduced Boundary


Thus far, we have shown that the measure-theoretic normal to a set E
of locally finite perimeter exists whenever the generalized exterior normal
exists (Theorem 5.6.5). Moreover, (5.6.11) states that the measure IIDXEII
has no mass inside the cone C(e), at least in the sense of measure density.
This indicates that the reduced boundary may have some appealing tan-
gential properties. Indeed, it will be shown that H" -1 -almost all of a - E
can be decomposed into countably many sets each of which is contained
within some Cl manifold of dimension (n - 1).

5.7.1. Definition. A set A C R'n is called countably (n - 1)-rectiflcable if


A C AD U [U°_0 fi(R n- ' )] where Fr' (A 0 ) = 0, and each fi : R" -1 - R" is
Lipschitz, i = 1, 2, .... Because a Lipschitz map defined on an arbitrary set
in Rn -1 can be extended to a[] of R"'' (Theorem 3.6.2), countable (n-1)-
`
rectifiability is equivalent to the statement that there exist sets Ei C R n-
and Lipschitz maps fi : Ei --► Rn such that A C Ao U (U_ ° (E)].
i1fé s

The next result is an easy consequence of Rademacher's theorem and


Theorem 3.6.2, concerning the approximation of Lipschitz functions.

5.7.2. Lemma. A set A C Rn -1 is countably (n - 1)-rectifiable if and


_
only if A C 1421 0 A, where Hn -1 (A 0 ) = 0, and each A,, i > 1, is an
(n - 1)-dimensional embedded C L submanifold of R n .

Proof. Obviously, only one direction requires proof. For this purpose, for
each Lipschitz function f, in the Definition 5.7.1, we may use Theorem
3.10.5 to find C' functions g i ,, j = 1, 2, ..., such that

00

f,(R" -1 ) C [Li gi,.i(Rn-`)


j=1

where 11" 1 (NO = O. Let Co denote the critical set of g, ^


Cla = R" -1 n {y : Jgi,j(y) = 0),
where Jg;,; (y) denotes the Jacobian of g ij at y. By an elementary area
formula, see [F4, Theorem 3.2.31, Hn -1 [gi,,(C, ,j)I = 0 and therefore the
set
(e. 00
Ao -u Ni U U g' / (C',1) = 0
1,3 =1
,

1 =1
has zero H" -1 measure.
For each y E R" -r - Ci,1 an application of the implicit function the-
orem ensures the existence of an open set Ui.j (y) containing y such that
244 5. Functions of Bounded Variation

g,d I U1,i(y) is univalent and that g,,i(U,,,(y)) is an (n - 1)-dimensional C 1


submanifold of R". Clearly, there exists a sequence of points y 1 , y2, ... in
Rn —1 — C,, i such that U l U=d (yk) D R n —1 — C;2 and

1J 9j(U(Yk)) ^ g. ^(Rn-1 - C=,,i).


00

k-1

Therefore, for each i,

U gi,; (Uo (yk))


00

f,(R n ' `) - AD C
;, k = 1

from which the result follows. ❑

5.7.3. Theorem. If E C R n is of locally finite perimeter, then S E is -

countably (n - 1)-rectifiable.

Proof. Clearly, in view of Corollary 5.5.3, we can reduce the argument to


the case of E with finite perimeter. Now recall from the proof of Lemma
5.6.6, that if x E ô- E, then

lim r - n IB(x, r) E n H +(x)I


r—•0
=0
and
lim r - nl[B(x,r) - n H (x)I = O.
r—*0

Since B(x, r) n H (x) = ([B(x, r) - E] n


- (x)) U (B(x, r) n E n H (x)),
-

the last equality implies that

IB(x, r) n E n H (x)1 1
lirn
r—*o
l B(x, r) I _ -

2.
Therefore, with the aid of Egoroff's theorem, for each 0 < e < 1 and
each positive integer i, there is a measurable set F, C 8 E and a positive -

number r, > 0 such that IIDXEII[(a E) - F,) < 1/(2i) and

IE n H+(x) n B(x, r)I (5.7.1)


< ()" IB(x , r)I

IE n H (x) n B(x, > 4IB(x,r)I (5.7.2)


-

whenever x E F, and r < r,. Furthermore, by Lusin's theorem, there is a


compact set M, C F, such that IIDXEII[F, - M,] < 1/(2i) and the restriction
of v(., E) to M, n 0 E is uniformly continuous. Since H n— ' restricted to
-

0 E is absolutely continuous with respect to IIDXEII (Theorem 5.6.7), our


-

conclusion will follow if we can show that each M, is countable (n - 1)-


rectifiable.

5.7. Rectifiability of the Reduced Boundary 245

We will first prove that for each x E M;,

C(z,e,v(x, E)) n M, n B(x,r0 = (5.7.3)

where C(x, e, v(x, E)) is the cone introduced in Definition 5.6.3. Thus, we
will show that I v(x) . (x — y)I < elx — y1 whenever x, y E M, and Ix — yl <
(1/2)r;. If this were not true, first consider the consequences of v(x) • (y —

x) > efx — yI. Since the projection of the vector y — x onto v(x) satisfies
Iproj,, ( ) (y — x)j > Elx — y1, it would follow that B(y,elx — y1) C H+ (x).
Also, since E < 1,
B(y, eix — yl) c B(x, 2 Ix — yl )
and therefore

B(y, Ejx — y1) c 11+ (x) n B(x, 2 1x — yl)• (5.7.4)

However, since 2Iz — yl < r,, it follows from (5.7.1) and (5.7.2) that

I n H + (x) n B(x , 2 Ix — vI)I < r }n


IB(x, 2Ix — yl)Î
4 l2 /
< (5.7.5)
4 enIB(0,1x — 11I)I
and

1E n B(y, eIx — yl)I > IE n B(y, Eli — vl) n H (y)I

? 41B(y, E1x — 3/I)I

= 4 E n 1$( 0 , Ix — y1)1• (5.7.6)

T hus, from (5.7.4), a contradiction is reached because

4 E n I$( 0 , lx — yI)I < IE n B(u,Elx - yI)I


< IE n H + (r) n B(x, 21z — yl)I < 4 e n 1$( 0 , IT — yl)I•

A similar contradiction is reached if v(x) • (y — x) < —Elx — vI and thus,


(5.7.3) is established.
We will now proceed to show that each M, is countably (n-1)-rectifiable.
In fact, we will show that M. is finitely (n — 1)-rectifiable. First, recall that
M; is compact and that v(•, E) is uniformly continuous on M. It will be
shown that for each x o E M, there exists a t > 0 such that M ; n B(xo, t)
is the image of a set A C R n-1 under a Lipschitz map. For this purpose,
assume for notational simplicity that "(x o , E) = v(ro) is the nth basis
vector (0,0,—. , 1). Let ir(xo) be the hyperplane orthogonal to v(x0) and

246 5. Functions of Bounded Variation

let p : M1 -o r(x 0 ) denote the orthogonal projection of M, into 1r(xo). The

=
conclusion will be established by showing that p is univalent on B(x o , t)nM1
and that pr' Ip[B (xo, t) n M ] is Lipschitz.
To see that p is univalent, assume the contrary and suppose that y, z E
Mi are points near x o with Iz-yI < IN and p(y) = p(z). Let u = z-y/Iz-yI
and note that Iv(x o ) uI = 1. Since y is continuous, it would follow that
Iv(y) • uI > e if y were sufficiently close to x o . However, (5.7.3) implies that
Iv(y) • uI < e, a contradiction. Thus, there exists 0 < t < 1r, such that p is
univalent on B(x o , t) n Mi .
Let L be the inverse of p restricted to p[B(x o , t) n M,1 and let y, z
p[B(xo, t) n Md. Then

IL(z) L(y)I _ IL(z)- L(y)I


-

I z - yI (I L(z) --- L(y)I 2 - Iproj,(z0)[L(z) - L(y))I2)1/2


1
IPro:6(. 0 (L(x)-L(1ry)I 2 \
(1
Using again the continuity of y, the last expression is close to
1
1/2
(5.7.7)
1 IProiy( v) IL(r)-L(y)J12 }
L(01 3
IL(x) —

provided that y is close to xo. by (5.7.3), (5.7.7) is bounded above by


1/(1 e 2 ) 1 / 2 , which proves that L is Lipschitz in some neighborhood of r o .
-

Since M, is compact, this proves that M, is finitely (n - 1)-rectifiable. ❑

The following is an immediate consequence of Lemma 5.7.2 and the pre-


vious result.

5.7.4. Corollary. If E C R" is of locally finite perimeter, then


OD

c7 Ec UM;UN
-

t=1
where H" ' (N) = 0 and each Mi is an (n - 1)-dimensional embedded C'

submanifoid of R".

5.8 The Gauss-Green Theorem


In this section it will be shown that the Gauss- Green formula is valid on
sets of locally finite perimeter. The two main ingredients in the formulation
of this result are the boundary of a set and the exterior normal. Since we
are in the setting of sets of finite perimeter, it should not be surprising
5.8. The Gauss-Green Theorem 247

that the boundary of a set will be taken as the reduced boundary and the
exterior normal as the measure-theoretic exterior normal.
In Definition 5.6.4, we introduced the notion of the measure-theoretic
exterior normal and demonstrated (Theorem 5.6.5) that

a -E C E. (5.8.1)

Moreover, from (5.6.14),

H'(a'E - 0 E)= 0. - (5.8.2)

One of the main objectives of this section is to strengthen this result by


showing that if B C a' E, then

H n-1 (B) = IIDXEII(B). (5.8.3)

This is a crucial result needed for the proof of the Gauss-Green theorem.
5.8.1. Theorem. If E C Bn has locally finite perimeter, then

Hn-:(B) = IIDXEII(B)
whenever B C a' E is a Borel set.

Proof. If r E a E, it follows from Theorem 5.6.3 that


-

r l-n
IIDXEII [ B ( x ► r)] = IIDXET 1)] -0 IIDXH lI[$(xt 1)] -

= H [B(x, 1) fl ir(x)]
n-1

= a(n - 1)

where ir(x) is the hyperplane orthogonal to v(x, E). Therefore,

lim IIDXEII[B(n r)], =1 , xEa - E. (5.8.4)


r^0 a(n - 1)r

Since H n (a'E - E) = IIDXEII[a`E —19 4 = 0 (Corollary 5.6.9) we -

may assume that B C a E and B C U1__, Mi , where each Mi is an (n- 1)-


-

manifold of class C l (Corollary 5.7.4). Fix i and let p = Hn -1 I Mi. Since


A is smooth,
Bz r
= 1, x E B fl Mi,
,

hm }^ 1
r^0 (Y(f9 - 1)T

and therefore, by (5.8.4),

Bxr ,
li. D = 1, x E B f1 M;.
II X EIE( ( B(,) r
)I
248 5. Functions of Bounded Variation

By the Besicovitch Differentiation Theorem (Theorem 1.3.8 and Remark


1.3.9),
Hn-1(B n
= p(B) = IIDXEII(B n Mi ).
The result easily follows from this. ❑

We now are able to establish the Gauss-Green theorem in the context of


sets of finite perimeter.

5.8.2. Theorem. Let E be a set with locally finite perimeter. Then,

1 div V dx = 8• E
n(t, E) • V (x)dHnr 1 (x)

whenever V E Co (Rn; Rn).

Proof. Choose a ball B(r) containing spt V. Then

div V dx = — V . d(DXE) (from Lemma 5.5.2)


JE
= V(s) • v(x, E)dIIDXE II (from Definition 5.5.1)
fa_ E
=V (X) •• n(x, E)dH n-! (x) (by the preceding theorem).
a• E

5.8.3. Remark. The Gauss-Green theorem is one of the basic results in


analysis and therefore, the above result alone emphasizes the importance
of sets of finite perimeter. Therefore, a question of critical importance is
how large is the class of sets of finite perimeter. The definition alone does
not allow easy identification of such sets. However, it is not difficult to see
that a Lipschitz domain, SZ, is a set with locally finite perimeter. An outline
of the proof will be given here while details are left as an exercise, for the
reader. We may assume that ft is locally of the form

fl={(w,y). 0 <y< g(w)}

where g is a non-negative Lipschitz function defined on an open cube


Q C Rn -1 . Since g admits a Lipschitz extension (Theorem (3.6.2) we may
assume that g is defined on R" .
-1 Let gE be a mollifier of g (Section 1.6)
and recall that
^Dgejd^ < f IDg l dz (5.8.5)
4 Q
for all c > 0. Each set

52E ={ (w,3/):0 y<9E(w), wEQ}


5.9. Pointwise Behavior of BV Functions 249

is obviously of finite perimeter because the classical Gauss Green theo-


rem applies to it (see Remark 5.4.2). Let X E denote Xç and observe that
llDX E Il(Rn) = H n-1 (ô1l ). Since

I + iDgE lZdx = H n-1 1{(w, t/) = 9E(w)+ w E Q}],

it follows from (5.8.5) that IIDX f 11(Rn) < C where C is some constant
independent of E. We may apply the compactness property of BV functions
(Corollary 5.3.4) to conclude that X is BV in R n , thus showing that it is
locally of finite perimeter. Moreover, Rademacher's theorem on the almost
everywhere total differentiability of Lipschitz functions (Theorem 2.2.1)
implies that the measure-theoretic normal is H n- t -almost everywhere given
by
(Dg(w),1)
n(z, f2) r

(5.8.6)
1% 1 + IDg(w)1 2
where x = (w, g(w)).
We conclude this section by stating without proof a useful characteriza-
tion of sets of finite perimeter. This will be stated in terms of the measure-
theoretic boundary.

5.8.4. Definition. If E C R n is a Lebesgue measurable set, the measure-


theoretic boundary of E is defined by

8,E = {z. D(E,x) >0} n fx.D(Rn—E,z)> 0}.

If we agree to call the measure-theoretic interior (exterior) of E all points


x for which D(E, x) =1 (D(E, x) = 0), then 8,E consists of those points
that are in neither the measure-theoretic interior nor exterior of E. See
Exercise 5.3 for more on this subject. In Lemma 5.9.5, we shall see that
0'E and ô, E differ by at most a set of Hn - l -measure O.

5.8.5. Theorem. Let E C Rn be Lebesgue measurable. Then E has locally


finite perimeter if and only if

Hn 1(K n3 M E) < o0
-

for every compact set K C R".

The reader is referred to [F4, Theorem 4.5.111 for the proof.

5.9 Pointwise Behavior of BV Functions


We now begin a treatment for BV functions analogous to that developed for
Sobolev functions in the first three sections of Chapter 3. It will be shown

250 5. Functions of Bounded Variation

that a 13V function can be defined by means of its Lebesgue points every-
where except for a set of Hn - '-measure zero and a set that is analogous
to the set of jump discontinuities in R 1 .
In the definition below, the following notation will be used:

A t = {x : u(x) > t},

B t = {x:u(x)<t},
n )I
15(E, x) = lim sup IR B(x'
-

r-.o IB(x,r)I
and
E n B(x, r)
D(E, x) = urn i nf I
r~O IB(x , r)I •
In case the upper and lower limits are equal, we denote their common value
by D(E, z). Note that the sets A t and B e are defined up to sets of Lebesgue
measure zero.

5.9.1. Definition. If u is a Lebesgue measurable function defined on R",


the upper (lower) approximate limit of u at a point x is defined by

ap lirn sup u(y) = inf {t : D(A e , x) = 0}


y-•x

(ap Lim inf u(y) = sup{t : D(B E , x) = 0}).


3+-•x
We speak of the approximate limit of u at r in case

aplimsupu(y) = apliminfu(y)•
v-x v-x

u is said to be approximately continuous at x if

ap lim u(y) = u(x).

5.9.2. Remark. If u is defined on an open set S1, reference to the definitions


imply that u is approximately continuous at r if for every open set U
containing u(x),
D[71-' (U) n 1,xj = 1.
An equivalent and rather appealing formulation is the one used in Remark
3.3.5. It is as follows: u is approximately continuous at x if there exists a
Lebesgue measurable set E containing z such that D(E, x) = 1 and u1 E
is continuous at x. It is clear that this formulation implies the previous
one. To see the validity of the opposite direction, let r 1 > r 2 > r 3 > ... be
positive numbers tending to zero such that

< I $(2, r )I, for r < rk.


1 11
B(x, r) n y : Iu(y) - u(x)1 >
k
5.9. Pointwise Behavior of BV Functions 251

Define

E Rn -- U{B(x , rk) - B(z, rk +1)} n v: lu(y) - u(z)I > -


k =1

Clearly, u I E is continuous at x. In order to complete the assertion, we will


show that 1)(É, x) = O. For this purpose, choose e > 0 and let J be such
that Er j z < E. Let r be such that 0 < r < rj and let K > J be the
integer such that rK+1 < r < rK. Then,
00
l( R" - E) f1 B(x, r)1 < E 103(x, rk) - B(x, rx +l )}
k=K

n fy : 1u(v) -10)1 >


k
< + °°
18(s, r)1 I B(x, rk )1
2k
+l k=K
IB(x, r)I + -
°°` I$(x, r)1
- 2K 2k
k=K+1
00
< IB(x ) r)I Z.,
k^ K
2
I B(x, r)IE ^

which yields the desired result since e is arbitrary.


One of the main results of this section is that a BV function can be
defined in terms of its approximate limits If" -1 -almost everywhere. For
this, the following is needed.

5.9.3. Lemma. Let n > 1 and 0 < T < 1/2. Suppose E is a Lebesgue
measurable set such that D(E,x) > r whenever x E E. Then there exists a
constant C = C(r, n) and a sequence of closed balls B(xi, ri) with x; E E
such that
00

E C 1J B(zer)
: =1
and
E(r,)"-1 < CIIDxEII R"]. [

i =1

Proof. For each x E R n , the continuous function


IB(r, r) n E1
f(r) _
IB(x,r)I
252 5. Functions of Bounded Variation

assumes the value r for some r s > 0 because it exceeds this value for some
possibly different r and approaches zero as r -. oo. Since r < 1/2, the
relative isoperimetric inequality, (5.4.3), implies

[ro(n)rx1(n-1)/n < CIIDXEII[B(x rs)]. ,

Now apply Theorem 1.3.1 to the family of all such balls B(r, r = ) to obtain a
sequence of disjoint balls B(x„ ri) such that U 1 B(r , 5r ; ) D E. Therefore,
CC 00
ITa(n)](11 -1)/n ^`(5r,)n-1 < 5(n -1)G*G^` IIDXEII{B(x, ri}]
L
i=^1 i=^1
< 501-1)CIIDxEIIERn].

In addition to (5.5.3) concerning the (n - 1)-density of the measure


IIDXEII, we will need the following.

5.9.4. Lemma. Let E C Rn be a set with locally finite perimeter. Then,


for Hn-1-almost every x € Rn — as E,

ÎIDXEII{B(x,r)1
l i rnsu p a ( rt - 1) r n-1
r—)
_ O.
Proof. For each positive number a let

A = (R n — c3'E) fl x:lim sup 1IDXEIIIB(x^


cY ( n - 1 ) r
r)l >a .
n-i
r-40

It follows from Lemma 3.2.1 that

IIDXEH{(A) > C)Hn -1 (A).

Therefore Hn -1 (A) = 0 since IIDXEII(A) = 0, thus establishing the con-


clusion of the Lemma. ❑

This leads directly to the next result which is needed to discuss the
points of approximate continuity of BV functions. Recall the definition of
the measure-theoretic boundary, 0.E, Definition 5.8.4. The next result,
along with (5.8.2) shows that all of the boundaries associated with a set
of finite perimeter, 0 - E, 0' E, and 0,E, are the same except for a set of
Hn -1 -measure zero.

5.9.5. Lemma. Let E C R'1 be a set with locally finite perimeter.. Then
a•E c a E and lin -1 (0M E - a* E) = O.

Proof. It follows immediately from Definition 5.6.4 that a' E C â M E. In


order to prove the second assertion, consider a point z E ° M E such that
5.9. Pointwise Behavior of 13V Functions 253

D(E, z) > b and D(Rn - E, z) > b where 0 < b < 1/2 and define a
continuous function f by

I E n B(z , r)I 1 - IB(z,r) _ E I


f (r) -
IB(z, r)I IB(z, r)I
Thus,
lim sup f (r) = D(E, z) > b
r --• p
and
liminf f(r) =1 -D(E,z)<1 -b,
r ---•0
with b < 1 - b. Hence, there are arbitrarily small r > 0 such that b <
f (r) < 1- b and for all such r, the relative isoperimetric inequality, (5.4.3),
implies
rbct(n)rn}(n-1) /n
< Ci IIDxEillB(z+ r)l'
Thus,
limsup IIDXEII
[B(z, r)i >
cx(n - )rn-1 0,
r-•
and reference to Lemma 5.9.4 now establishes the conclusion.

In the next theorem, it is shown that a BV function is approximately


Hn_
continuous at all points except for a set of i -measure zero and a count-
ably (n - 1)-rectifiable set E which, roughly speaking, includes the points
at which u has a jump discontinuity (in the sense of approximate limits).
H"-
It is also shown that at '-almost all points of E, u has one-sided ap-
proximate limits. Later, these results will be refined and stated in terms of
integral averages.
Recall from Definition 5.9.1 that At _ {x : u(x) > t}.

5.9.6. Theorem. Let u E BV (Rn). If

4x) = ap lim sup u(y),


I,—s

A(s) apliminfu(y),
y^ a

and
E = R" n {z : A(x) < µ(x)},
then
(i) E is countably (n - 1)-rectifiable,

(ii) -co < A(x) S p(x) < oo for 11' 1 -almost all z E Rn ,

(iii) for 11n -1 -almost all z E E, there is a unit vector y such that n(z, A,)
v whenever .\(z) < s < µ(z).
254 5. Functions of Bounded Variation

(iv) For all z as in (iii), w ith -oo < A(x) < p(x) < oo, let H (z) - {y - :

(y - z) • v < 0 } and H+ (z) = {y : (y - z) • y > 0 }. Then, there a re


Lebesgue measurable sets E and E+ such that
-

IE_
n H (z) n B(z,r)I = 1- lim I+
-
E n H+ (x) n B(z, r }I
lim
r^0 IH —(z) n B(,Z, T)I r—PD 111+(z) n B(z, Ty

and

urn u(x) = p(z), Jim u(x) = a(z).


zEE'rzH - (a) zEE+f11i+(a)

Proof. Applying Theorems 5.4.4 and 5.7.3, there exists a countable dense
subset Q of R 1 such that H(A t ) < oo and 8* A i is countably (n - 1)-
rectifiable whenever t E Q. From Remark 5.9.2 we see that
Hn -1 [{U(B M A t
- a . At) . t E Q}] = 0.

It follows immediately from definitions that

{x : A(x) < t < p(x)} C A i for t E R I , (5.9.1)

and therefore E C {Ua M A t : tEQ},Hn` 1 [E- {Uô'A t : tEQ]]= O. This


proves that E is (n - 1)-countably rectifiable.
Let I = {x : A(x) = -oo} U {x : p(x) = oo}. We will show that
H" - `(I) = O. For this purpose it will be sufficient to assume that u has
compact support. First, we will prove that H" -1 [{x : A(x) = oo}] = O. Let
L t = {z : A(x) > t} and note that D(L t , x) = 1 whenever x E L t . Now
apply Lemma 5.9.3 to conclude that there is a sequence of halls {B(r)}
whose union contains L t such that
00

E Ti -1 <- GIIDXL, II•


1=1

Since u has compact support, we may assume that diam B(r,) < a, for
some positive number a. Therefore, Theorem 5.4.4 implies

Hâ -1 1{x:A(x)=oo})=Hâ -l [{nL 1 :tE R'}l


<Cl^rni fIIDXLIII(R") =o.

From this it easily follows that Hn -1 [{x : A(x) = oo f] = O. A similar proof


yields H" - ' [{x : p(x) _ -oo}] = O. Thus, the set {x : p(x) - A(x) } is well-
defined for H" -1 -a.e. x and the proof of (ii) will be concluded by showing
that Hn -1 [{x : p(x) - A(x) +oo}] = O. Since E is countably (n - 1)-
rectifiable, it is a-finite with respect to Hn -1 restricted to E. Therefore,
5.10. The Tr ace of a BV Fu nction 255

we may apply Lemma 1.5.1 to obtain


00

1 (1,4 - A)dHn- i _ 0
00
H n `' [{x : A(s) < t < µ(x) }idt

< Hn-1(0A., A t )dt (by 5.9.1)


0
00
< li n-1 (a'A t )dt (by Lemma 5.9.5)
o
00
_< CIIDXA, Ildt (by Theorem 5.6.7)
0
< CIIDuII(Rn) (by Theorem 5.4.4)
< oo, since spt u is compact.

We will prove that (iii) holds at each point

zE E--{U(a.4 A t -0* A t ) :t E Q}.

If t E Q with A(z) < t < µ(z), then z E a m A t and therefore z € O A t . Con-


sequently, n(z, A t ) exists. But is must be shown that n(z, A t ) = n(z, A 8 )
whenever A(z) < 9 < i(z). It follows from the definition of the measure
theoretic exterior normal (Definition 5.6.4) that

D(A t , z) = 1/2 = D(A., z). (5.9.2)

If s < t, then A. D A t and therefore D(A, - A t , z) = 0, which implies that


n(z, A t ) = n(z, A,).
For the proof of the first assertion of (iv), let z E E-- I and choose e > 0
such that A(z) < µ(z) - e < A(z). Observe that D(A, (z ) €, z) = 0 while

lim
IA y (,) -E fl H(z) n B(z, r)I = 1
-

r—.o IH—(z) fl 11(z, r)I '

from (5.9.2). Therefore

u-ri^(z) -- e, µ(z) + e] n H - (z) n B(z,r)I -1


hill
r-0 IH-(z) n B (z,r)I
By an argument similar to that in Remark 5.9.2, this implies that there
is a set E with the desired properties. The second assertion is proved
-

similarly. 0

5.10 The Trace of a BV Function


For a given set SZ C R" with suitably regular boundary and u E BV(n),
we will show that it is possible to assign values to u at Hn -1 -almost all
256 5. Functions of Bounded Variation

points of OA even though u, when considered as a member of L 1 (it), is


defined only as an element of an equivalence class of functions. Recall that
two measurable functions are called equivalent if they differ at most on
a set of Lebesgue measure zero. The difficulty with defining the trace of
a function on the boundary is that an may have zero Lebesgue measure,
precisely where the function may be undefined. The theory requires further
development in order for this difficulty to be circumvented. The approach
we use for this is as follows. For a certain class of domains S2 C Rn (called
admissible domains below), if u E BV(A) is extended to all of R" by
defining u E 0 on R" - A, then an easy application of th e co-area formula
shows that u E BV(R"). By means of Theorem 5.9.6 we then are able
to define u H" -1 -almost everywhere including E, the set of approximate
jump discontinuities.

5.10.1. Definition. A bounded domain SZ of finite perimeter is said to be


admissible if the following two conditions are satisfied:
(i) H n-1 (ÔA - ôMA) = 0,

(ii) There is a constant M = M(11) and for each x E au there is a ball


B(x, r) with
H "-1 [(a M E) n (aMmj < MH"
! pM E) n q (5.10.1)
-

whenever E C iZ fl B(x, r) is a measurable set.


5.10.2. Remark. It is not difficult to see that a Lipschitz domain is ad-
missible. For this purpose, we may assume that SZ is of the form

A ={(w, y ):OÇ y Ç 9(w)}

where g is a non-negative Lipschitz function defined on an open ball B C


R n-1 . From Remark 5.8.3 we know that A is a set of finite perimeter. Let
E C A be a measurable set and we may as well assume that H" -1 (S2 fl
a m E) < oo for otherwise (5.10.1) is trivially satisfied. Since 0„E = (a M En
ail) U (A n O M E) and H" -1 (011) < oo, we conclude from Theorem 5.8.5
that E has finite perimeter. Hence, we may apply the Gauss Green theorem
(Theorem 5.8.2) with the constant vector field V = (0, 0, ... ,1) and (5.8.6)
to obtain

V • n(x, SZ)r1Hn -1 (x) + V • n(x, E)dH"-1 (3 • ) = U.


(b• E)n(Bn) (8 • E)nS2

Therefore, if a is the Lipschitz constant of g, we hav e


Hn-1i(a.E) n
1
+l al
2
(a^)j < H "-1 ((a ' r) n
5.10. The Tr ace of a BV F u nction 257

and reference to Lemma 5.9.5 establishes the desired conclusion.

5.10.3. Definition. Whenever u is a real valued Lebesgue measurable


function defined on an open S et S2, we denote by u ü the extension of u to
Rn:
u(r) x E S2
uo ( x ) = Q
{

x E .Rn - S2_
Observe that uo is merely a measurable function and is therefore de-
fined only almost everywhere. Later in the development, we will consider
u E BV(S2) where S2 is an admissible domain, and then we will be able
to define uo everywhere except for an H" -1 -null set. If S2 is a smoothly
bounded domain and u E BV (0), it is intuitively clear that u o E BV (R" )
because the variation of u o is greater than that of u by only the amount
contributed by H" - ' (Oft). The next result makes this precise in the context
of admissible domains.

5.10.4. Lemma. If 11 is an admissible domain and u E BV(1l), then


uo E BV(Rn) and Iluollsv(R.) < CIlullsv(a) where C = C(f2).

Proof. It suffices to show that u0 is BV lu a neighborhood of each point of


an because an is compact. If we write u in terms of its positive and negative
parts, u = u+ - u - , it follows from Theorem 5.3.5 that u E BV (S2) if and
only if u+(0), It 01) E BV(0). Therefore, we may as well assume that u is
-

non-negative. For each x E Oft, let B(x, r) be the ball provided by condition
(ii) of Definition 5.10.1. Let cP be a smooth function supported by B(x, r)
such that 0 < < 1 and cp 1 on B(x, r/2). Clearly, youo E By (S2) and
Theorem 5.4.4 and Lemma 5.9.5 implies
on

I lin -1 1S2f1 a M A r ldt = IIH(cPuo)ll(SZ) < co (5.10.2)

where A t = {x : Spu o (x) > t). Since !A t -fflB(x, r)I = 0 for t > 0, (5.10.1)
and (5.10.2) imply

1
00
Hn -
1 (am A t )dt < GIID(0uo)II(S 2 ) < oo.

Hence, by Theorem 5.4.4, Spun E BV (R") with

IlDuoil[$(r, */ 2 )l11 <_ IID(puo)II(Rn) < cIID(taüo ) II(n)-


However, by (5.1.2),

IInOvuo)II(SZ) = sup o(pdiv V dr : V E Co (SZ; R"), IVI < 1


J nu
258 5. Functions of Bounded Variation

and

upcp div V dx = u div(^pV)dx — D Dv • V dx.


n nu
Therefore

IIDuoIIf B(x, r/2)]II <_ CIID(vuo)II(n) < CIIDuIl(n) + c(r)IIuII I S2 ;

s r + c(r)]IIuIIev(n) •

This is sufficient to establish the result because OS/ is compact. 0

We now are able to define the trace of u on the boundary of an admissible


domain.

5.10.5. Definition. If 0 is an admissible domain and u E BV (11), the


trace, u', of u on an is defined by

IL' (x) = puo (x) + Auo (x)

where µ,yo (x) and A uo (x) are the upper and lower approximate limits of u o
as discussed in Definition 5.9.1 and Theorem 5.9.6.

5.10.6. Remark. We will analyze some basic properties of the trace in light
of Theorem 5.9.6. Let E = {x : A uo (x) < Puo (z) ), A t = {x : u o (x) > t },
and select a point x o € E fl WI where (iii) of Theorem 5.9.6 applies. Thus,
there is a unit vector u such that

n(x a , A t ) = 4r whenever A u° (x o ) < t < puo (xo).

We would like to conclude that

y = ±n(x 0 , SZ).

(5.10.3)

For this purpose, note that 0 E [A uo (zo), A t" (z 0 )] and A t C 0 for t > O.
If t > 0 and u $ ±n(z o , SZ), then simple geometric considerations yield

B(xa, 01I >


limsup I(At _n O.
f'o IB (x o, ) I

This is impossible since A t C O. On the other hand, if t < 0 and y $


±n(xo, SZ), then

lim sup I(Bt — n) n B(xo, r)I > 0,


f--00 I B(xo, r)I

an impossibility since Bt = {x : uo (x) < t } C D. Hence, (5.10.3) is estab-


lished.

5.10. The Trace of a BV Function 259

Also, observe that

if y - n(x o , Si), then A 0 (x0) = 0. (5.10.4)


For, if a uo (r 0 ) < 0 there would exist t < 0 such that Au° (To) < t < ilup (xa).
Because y = n(x o , A i ), it follows that

D(A t n{x: (x—x o )•v >0} zo) = O. ,

But t < 0 implies (R" - 12) - A i l - O. This, along with the fact that
u = n(x o , SZ) yields

D(A 1 n {x : (z - x 0 ) • v > 0}, z 0 ) > 1/2,

a contradiction. Therefore, ALO (x o ) < O.


On the other hand, if A 0 (xo) > 0, there would exist t > 0 such that
D(A i , x o ) = 1. This would imply that

D(A t C1 {z:(x - z o )•v>0},x o )'1/2

which is impossible since A t -S2 = 0 and

D(SZ n {z: (z-z0)•v>0},xo)= 0.

Thus, (5.10.4) follows and a similar argument shows that

if y = -n(xo, f2), then pun (x 0 ) = O.

( 5. 10. 5)

Later, in Section 5.12, after certain Poincaré-type inequalities have been


established for BV functions we will he able to show that if f2 is admissible
and u E BV (f2), then

Iint
r-60 fB ( r)flO
iu(Y) - u` (i)i"/("-1)dy = 0 (5.10.6)

for 11" '-almost all x E BSZ.


-

We conclude this section with a result that ensures the integrability of


u` over an.

5.10.7. Theorem. If f2 is an admissible domain, there is a constant M =


M(I1) such that
1
"-

8^ n
whenever u E BV(f2).

Proof. Since by definition, if A„. + p ue , it suffices to establish the


inequality for the non-negative function p = A u ., the case involving A
being treated in a similar manner. As in the proof of Lemma 5.10.4, we
260 5. Functions of Bounded Variation

need only consider the case when it is replaced by (pp, where coo is a smooth
function with 0 < cp < 1, cp = 1 on B(x, r/2), and spt u C B(x, r), where
B(x, r) is a ball satisfying the condition (5.10.1).
First, with A t = {x : (pµ(x) > t), observe that A t n 0,,,11 c (â M A t ) n
(0,,4 0), for t > 0. Indeed, let x E A t n OM fl and suppose x Om A t . Then
either D(A t , x) = I or D(A t , x) = 0. In the first case D(f , x) = 1 since
(A t - SZ = 0 for t > 0. Hence, x a xi s 2, a contradiction. In the second
-

case, a contradiction again is reached since the definition of cpi..t(x) implies


z ¢ A t . Therefore, we have

H "-' (At n âMn1 < Hn -1 [(aMAt) n (akin)i, t > 0. (5.10.7)

Thus, we have

µ dH n-1 < dHn-i


f (a,r%2)naM L2 8(x,r)f18M ^3
00
< Hn-t (At n 0,, 4 3-1) (by Lemma 1.5.1)
0
00
H"-1[(aMAt) (by 5.10.7)
n (8M f)jdt
o
< M Hn-`[(a,,,At) n f2)dt (by 5.10.1)
< MIID(cpp)11(SZ) (by Theorem 5.4.4)
< Mlllullsv(n)•
Since A u almost everywhere, the last inequality follows as in the proof
of Lemma 5.10.4. O

5.11 Poincaré -Type Inequalities for BV Functions


In this section we prove the main inequality (Theorem 5.11.1) from which
essentially all Poincaré-type inequalities for 13V functions will follow. This
result is analogous to Theorem 4.2.1 which was established in the context
of Sobolev spaces. In accordance with the previous section, throughout we
will adopt the following conventions concerning the point-wise definition of
BV functions. If u E BV (R") set

u(x) = 2 [au (x) + µu (z)] (5.11.1)

at any point where the right side is defined. From Theorem 5.9.6(iî), we
know that this occurs at H" - t-almost all s E R" . If u E BV(SZ), St admis-
sible, then we know by Lemma 5.10.4 that n o E BV (R") and therefore up
is defined H" -1 -a.e. on R". Thus, we may define u on II in ternis of up as
5.11. Poincaré-Type Inequalities for BV Functions 261

follows:
_ uo(x) x E S2 (5.11.2)
u( x ) r 1 2u 0 (x) x E 80.
At first glance, it may appear strange to define u(x) = 2u.0(x) on OS),
but reference to (5.10.4) and (5.10.5) shows that this definition implies the
intuitively satisfying fact that at H' -a.e. x E an, either u(x) = puo (x)
or u(x) = A uo (x). Consequently, u is a Borel function defined H" - -a.e. on
Si Note that we have for H" -1 -a.e. x,
au(x) + bu(x) = (au + bu)(x)
whenever a, b E R 1 .

5.11.1. Theorem. Let Si be a connected, admissible domain and suppose


u E BV(SZ). If T E [BV (S2)]' and T(Xn) = 1, then

11u - T(u)IIn/(n-1);52 <_ CIITII IIDuII(i 1 ), (5.11.3)


where 11TII denotes the norm of T as an element of [BV(S2)J', and C
C(Sl, n).

Proof. It suffices to show that

Ilu — T (u )I I 1;0 <_ cI I TII IIDu11(n ) , (5.11.4)


for if we set f = u - T(u), then by Sobolev's inequality and (5.11.4),

IIIIIn/(n--1) s1 = IIfOIin/(n-1);R" < C,IIfollav(Rn) <_ CII f IIBv(Z).


;

The last inequality follows from Lemma 5.10.9. Also, note that the Sobolev
inequality holds for f o because it holds for the regularizers of fo, whose 13V

_
norms converge to IIfolIBy(R^) by Corollary 5.2.4. Thus, in view of the fact
^
that 11TII > ICI - ` (5. 11 . 3 ) follows from (5.11.4).
To prove (5.11.4), it is sufficient to assume fSZ u(x)dx = 0 since the
inequality is unchanged by adding a constant to u. With this assumption,
(5.11.4) will follow if we can show

Ilulll;^ < CIIDuII(IP) (5.11.5)


because

Ilu - 7'(u)Il1sz IIuIl1;i1 + ISiI 1I 7'II IIuhIBV(s1)


< (1 + ICI IITIII(IIull1;n + IIDuII(i1))
<_ 21n1 IITII(IlulIi , ^1 + IlDull(S))•
If (5.11.5) were not true for some constant C, there would exist a sequence
uk E BV (S2) such that

I uk(x)dx = 0, Iluk 1, and IlDuk1l(11) --> 0. (5.11.6)


262 5. Functions of Bounded Variation

For each uk, form the extension uk, o by setting u = 0 on Rn — S2. From
Lemma 5.10.4, it follows that the sequence {Iluk,ollBv(R^)} is bounded
and therefore an application of Corollary 5.3.4 implies that there exist
u E BV(Rn) and a subsequence of (uk,o} (which will still be denoted by
the full sequence) such that uk, o u in L i (0). Therefore l^ulli st = 1 ;

from (5.11.6). But (5.11.6) also show that llDull(11) = 0 and therefore u
constant on 11 since SZ is connected. Consequently, u - 0 which contradicts
Ilulli;a = 1. Thus, (5.11.5) and therefore (5.11.4) is established. Q

5.11.2. Corollary. Let 12 be a connected, admissible domain. Let c and


M be numbers such that

lao + aicl < MIlao + a[ull av(n)


for all a a , a i E R 2 . Then there exists C = C(I1) such that

Hu — 1);n Ç CM llDull(Rn ).

(5.11.7)

Proof. Define a linear map To on the subspace of BV (S2) generated by


X s-1 and u by T0(x0) = 1 and To (u) = c. From the hypotheses, the norm
of To is bounded by M and therefore an application of the Hahn-Banach
theorem provides an extension, T, to BV (S2) with the same norm. Now
apply Theorem 5.11.1 to obtain the desired result.

5.12 Inequalities Involving Capacity


We now will investigate the role that capacity plays in Sobolev-type in-
equalities by considering the implications of Theorem 5.11.1. Recall that
the BV (S2) is endowed with the norm

IIullBv(n) = liuill;fl + IuDull(Il).


However, for notational convenience, we will henceforth treat BV(Rn) sep-
arately and its norm will be given by

HullBv(R°) = llDull(R n )-
In Section 2.6, Bessel capacity was introduced, developed, and subse-
quently applied to the theory of Sobolev spaces. Because of the irreflexivity
of L i , it was necessary to restrict our attention top > 1. The case p = 1
is naturally associated with BV functions and the capacity in this case is
defined as follows:

7(E) = : v E Y(Rn), E C int {v > 1 }},


where
y(Rn) = Ln/(n-l)(Rn) n BV(Rn).
5.12. Inequalities Involving Capacity 263

Note that W1,1(Rn) c Y(R") and by regularization (cf. Theorem 5.3.1),


that
RILL' (n-1) <_ CI1DuII(Rn ) (5.12.1)
for u E Y(R"). A simple regularization argument also yields that in case
E is compact,
7(E) = inf{IIDvIj I : v E Co (R"), E C int{v > 1 }}. (5.12.2)

5.12.1. Lemma. If E C R" is a Suslin set, then

y(E) = sup{y(K) : K C E, K compact}.

Proof. Referring to Theorem 2.6.8, we see that it is only necessary to


show that y is left continuous on arbitrary sets since right continuity on
compact sets follows directly from (5.12.2). Thus, it suffices to prove that
if E 1 C E2 C ... are subsets of Rn, then
^

y (U E; — lirn
i =1 i-.00

For this purpose, suppose

a = limy(Ei)< oo, and e>O.


•-I00

Choose non-negative v i E (R") so that

Ei C int{x : v i (3) > 1) and II Dv ; II(R" ) < 7(E; ) +e2',

and let h i = sup{v1, v 2 , ... , v;}. Note that h, E Y(R") and

h i sup{h 3 _ 1 , vj }, Ei_1 C int{x ?: 1).

Therefore, letting I j = inf {hj _ i i v3 ), it follows from Theorem 5.4.4 (which


remains valid for functions in Y(R")) and Lemma 5.9.5 that

II Dh1II (R ) +'r( E;-1 ) <_ IIDh1II(R ) + IIDI,^I(R^)


"
"

+ co
lin -1 (a m {h j > t })dt
1_m

f
+oo
+ H n 1(a%,{1j > t })dt.
^

It is an easy matter to verify that

a x, {h i > t} U 9 M {Ij > t} C aM {hj _ 1 > t} U a {v j > t}

8,K {hj > t} n 8M {I1 > t} c r3 M {hj _1 > t} n aM{vj > t}.
264 5. Functions of Bounded Variation

Consequently,
Hn 1 (O m {hj > t}) + Hn
-
- 1 (â,,, {Ij > t})

< H" -1 (0,M {hj_1 > t}) + H" -1 (0, 4 {vi > t})
and therefore,
+oa
H n- I (aM{hj-, > t })dt
Ilbhjll(R n ) +7(Ej-1) <_
f00

+
^ f
+m
)
H n- 1 (ô,^{v > t })dt

-
IIDhj - 1II(R " ) + IIDU,II(R " )
< IIDhj-1II(R n ) +7(E,) + E2 -? .
It follows by induction that

IlDhall(R n ) < 7(Ej) + E e2 - s.


i =1

Therefore, letting w = timj._. œ hi, (5.12.1) implies

IIwII("+1)/n = ?^ ^ Ph/ I (n+1)/n tim ^pCIID►t,II (R") < ^(a + E _ },

whereas the proof of Theorem 5.2.1 implies

ODwII(R") limiuf IlDh,II(R") < oo.


1 ^ o0

Thus, w E Y(R") and


00

7 u E, < IIDwII(R") < 1 m e llbh, ll(R") < a+ e. El


1 ,-i

In addition to the properties above, we will also need the following.

5.12.2. Lemma. If A C R" is compact, then

7(A) = inf{P(U) : A C U,U open and lUI < oo}. (5.12.3)

Proof. Let 71(A) denote the right side of (5.12.3).


7(A) < 7 1 (A): Choose rt > 0 and let A C U where U is bounded, open
and
P(U) < y 1 (A) + rj.
5.12. Inequalities Involving Capacity 265

Let X = Xu and X E Xu * ço,, where cp, is a regularizer. Then X E > 1 on


A for all sufficiently small e > 0 and

7(A) < IDX,ldx (by (5.12.2))


R"
< IJ DXu II( R") (by Theorem 5.3.1)
< + rt.

This establishes the desired inequality since n is arbitrary.


71(A) < 7(A): If q > 0, (5.12.2) yields u E C(R) such that u > 1 on
A and
IDuldx < -y(A) + rt.

By the co-area formula,


co
IDuldx = fin-1 u -1 (t)ldt
Rn 0 l
l
H" -1 [u -1 (t)ldt
o
>
Hn-1(u—' (to)]
-1
for some 0 < t o < 1. Since a{u > t o } C u (to), with the help of Lemma
5.9.5, it follows that

71(A) < P({u > to}) < H" -1 [i9{u > toil < H n- [ 11 I(t o )]
1 -

< y(A) + q. fl

We now are able to characterize the null sets of 7 in terms of Hn -1 .

5.12.3. Lemma. If E C R" is a Suslin set, then

ry(E) = 0 if and only if H" - '(E) = O.

Proof. The sufficiency is immediate from the definition of H" -1 and the
fact that 7[B(r)J = Cr''.- In fact, by a scaling argument involving x -+ rx,
it follows that 7[B(r)] = y[l3(1)]r" -1 .
To establish necessity, Lemma 5.12.1 along with the inner regularity of
H" 1 1F4, Corollary 2.10.481 shows that it is sufficient to prove that if
-

A C R" is compact with 7(A) = 0, then lift -1 (A) = O. For e > 0, the
previous lemma implies the existence of an open set U ❑ A such that
P(U) < e. Lemma 5.9.3 provides a sequence of closed balls {B (r ; ) } such
that U i'° 1 B(r,) DU ❑ A and
CCO

>2r1 ^ CP(U) < C. ❑


266 5. Functions of Bounded Variation

We proceed with the following result which provides some information


concerning the composition of [BV(R`t))' .

5.12.4. Theorem. Let II be a positive Radon measure on R n . The following


four statements are equivalent.
(I) H n— I (A) = 0 implies that µ(A) 0 for all Borel seta A C Rn and
that there is a constant M such that I f udpl 5 MII uIlev(Rn) for all
u E BV(R").

(ii) There is a constant M such that p(A) < MP(A) for all Borel seta
ACR" with IAI<oo.

(iii) There is a constant M1 such that t(A) < M j 7(A) for all Borel sets
ACR".
_
(iv) There is a constant Mi such that p[B(x,r)) < Mar" -1 whenever
x ERn and rER' .
The ratios of the smallest constants M, M 1 , and Mf r have upper bounds
depending only on n.
Proof. By taking u = XA, (ii) clearly follows from (i) since

IIuIIBV(R *) = IIDuUI(R n ) = P(A).


For the implication (ii) = (iii), consider a compact set K and observe
that from the regularity of p and (ii),

p(K) = inf {µ(U) : K C U, U open and IUI < oo}


< Minf{P(U) : K C U, U open and IUI < co).
Lemma 5.12.2 yields p(K) < M7(K). The inner regularity of p and Lemma
(5.12.1) give (iii).
Crn
Since 7[B(r)) -1 , (iii) implies (iv).
Clearly, (iv) implies that p vanishes on sets of Hn - l-measure zero. Con-
sequently, if u E BV(Rn), our convention (5.11.1) implies that u is defined
p-a.e. If u is also non-negative, we obtain from the co-area formula
00

IIDuII(R n ) = IIull$v(R.) = f P(A t )dt (5.12.4)

where A t = {x : u(x) > t). In particular, this implies that for a.e. t, A t has
finite perimeter. For all such t, define

Ft = A t n {x : D(A t ,x) > 1/21.


For x E A t , the upper approximate limit of u at x is greater than t (see
5.11.1), and therefore
A t - Ft C {x : 0 < D(A t , x) < 1/21. (5.12.5)
5.12. Inequalities Involving Capacity 267

Therefore, A t -- Ft C aM A t . In fact, A t Ft C a„, r A t ()silt because


- -

x E ô' A t implies that D(A t , x) = 1/2. Therefore, H" -1 (At - Ft ) 0


by Lemma 5.9.5. Thus, we may apply Lemma 5.9.3 to F t and obtain a
sequence of balls (B(r; )} such that Ft C Us_ 1 B(r i ) and
00

E=1
i
,
7- 71- 1 < CP(Ft )-

Therefore (iv) yields


µ(Ft) < CMZ P(Ft ). (5.12.6)
Now, P(A t ) = P(Ft ) and since p vanishes on sets of N" '-measure zero,
-

we have µ(A t ) = p,(FF ). Thus, Lemma 1.5.1, (5.12.6), and (5.12.4) imply
00

f udp =p(A t )dt < CM2IIuIIBV(R")-


0

If u is not non-negative, apply the above arguments to IuI to obtain (i). ❑

5.12.5. Remark. A positive Radon measure p satisfying one of the con-


ditions of Theorem 5.12.4 can be identified with an element of [BV(R")j•
and M can be chosen as its norm.
Suppose SZ is an admissible domain and p a positive measure such that
spt p C SZ. In addition, if p E [BV (1 ))', then there exists a constant
C C(1, p) such that

udp < CIIuIIBV(p) < CIIuIIBv(R„)


whenever u E BV(Rn). Thus, p E [BV(R")]' and Theorem 5.12.4 applies.
On the other hand, if spt p C R and one of the conditions of Theorem 5.12.4
holds, then p E (BV(St))' because of Lemma 5.10.4. Thus, for measures p
supported by SZ, p E [BV (0)]* if and only if one of the conditions of
Theorem 5.12.4 holds and in this case there is a constant C = C(11) such
that

C`'IIIPII[BV(R")]• < 1114 11[13V (On' <


_ CI' pII[BV (Rn )]'. (5.12.7)
For the applications that follow, it will be necessary to have yet another
formulation for the capacity 7.

5.12.6. Lemma. If A C R" is a Suslin set, then


i(A) = sup{p(A)} (5.12.8)
where the supremum is taken over the set of positive Radon measures p E
[BV(R0)]* with IIPIIiBV(R")l• < 1.

Proof. Because of the inner regularity of -y (Lemma 5.12.1) it suffices to


consider the case when A is compact. Referring to the Minimax theorem
268 5. Functions of Bounded Variation

stated in Section 2.6, let X denote the set of all positive Radon measures
p with spt p C A and p(Rn) = 1. Let Y be the set of all non-negative
functions f E Co (Rn) such that IIDf II► < 1. From the Minimax Theorem,
we have
sup inf f f dµ = inf sup f f dt.t.
f EY PEX pEX fEy

It is easily seen that the left side is equal to the reciprocal of 7(A) whereas
the right side is the reciprocal of the right side of (5.12.8).

One of the most frequently used Poincaré-type inequalities is

IIullp;n < C(U)IIDüllp;n


where p > 1, u E W 1 P(11), and An u(x)dx = O. Inequalities of this type were
treated from a general perspective in Section 4.2. In the next theorem, we
will obtain a Poincaré-type inequality for BV functions normalized so that
their integral with respect to a measure in [BV(Q)j• is zero. That is, the
measures under consideration are those with the property that p[B(x, r)] <
Mr" -1 for all balls B(x, r). For example, this includes Lebesgue measure
restricted to a bounded domain or (n — 1)-Hausdorff measure restricted to
a compact smooth hypersurface in R" .

5.12.7. Theorem. Let SZ be a connected admissible domain in R" and let


p be a non-trivial positive Radon measure such that apt µ C U and for some
constant M > Q that
14B(x, r)] < Mr"'
for all balls B(x,r) in R". Then, there exists a constant C = C(1) such
that for each u E BV(U),

II^ - u(0IIn /(n --1);11 G ^ IIDuII(n)


µ( )

where û(µ) _ f u(z) dp,(x).


71s

Proof. Theorem 5.12.4 states that u E [MTV)] * and because S2 is admis-


sible, (5.12.7) shows that may be regarded as an element of IBV(11)j`.
Therefore, Theorem 5.11.1 is applicable and the result follows immedi-
ately.

This leads directly to the Poincaré inequality for BV functions.

5.12.8. Corollary. Let SZ be a connected, admissible domain and let A C Si


be a Suslin set with H" -1 (A) > O. Then for u E BV(S2) with the property
5.12. Inequalities Involving Capacity 269

that u(x) = 0 for H" -1 -almost all x E A, there exists a constant C = C(f2)
such that
Ilulln/(n-1);n
7( A)
^IDuII (^)•

Proof. From Lemma 5.12.6 we find that A supports a positive Radon mea-
sure µ E [BV(1)1' such that µ(A) > 2 -1 7(A) and IIIiIIiBv(R.)i• < 1. Thus,
for any u E BV (S2) with the property in the statement of the corollary,
f u dp = O. Our conclusion now follows from the preceding theorem. ❑

We now consider inequalities involving the median of a function rather


than the mean. The definition of the median is given below.

5.12.9. Definition. If u E BV (S2) an d a positive Radon measure in


[BV (S2)]', we define med(u, p) as the set of real numbers t such that

µ[S2 n {x : u(x) > t}] < 2µ(SZ)

µin {x : u(x) < t}] <

It is easily seen that med(u, A) is a non-empty compact interval and that


if a s and al are constants, then
med(a o + al u, p) = ao + a 1 med(u, it) . (5.12.9)
If c E med(u, p), then µ(A c ) > ε(S2) where A, = SZ n {x : u(x) > c}].
Consequently,
;p(n) _< cµ( k ) 5jlu(x)Id(x).
Similarly, if c Ç 0, then µ(B c ) > ;µ(S1) where Bc = 11 n {x : ta(x) < c }]
and
-- µ(^ < -cµ(i3 c ) <J -
u(x)d1i(x) < Iu(x)d(x).
BC f
Therefore,
Ici ^ ? I uldu ^ (5.12.10)
u(^) n
and (5.12.9) thus implies

lao + aid <_ 2 llPIIEBV(n)l• Ilao + a l ullav ( o)•


µ(n)
The following is now a direct consequence of Corollary 5.11.2.

5.12.10. Theorem. Let 11 be a connected admissible domain in R" and let


µ be a positive Radon rneasure such that apt p C and for some constant
M>0 that
µ[$(z, r)] < Mrn -1
270 5. Functions of Bounded Variation

for all balls B(x, r) zn R". Then there exists a constant C = C(1l) such
that for u E BV(1l) and c E med(u,p),

I u — c l l n r ( n- 1 0 C M I D uI I(0).
p( n )
5.12.11. Corollary. Let SZ be a connected, admissible domain and let A
and B be disjoint Suslin subsets of fl of positive H" -1 -measure. Then
there exists a constant C = C(f) such that for each u E BV(11) with u > 0
Hn -1 -a.e. on A and u < 0 Hn -1 -a.e. on B,

IIuIIn/(n-1);f2 Ç C[7(A) -1 + 7(B)-IIIDuII(0).


Proof. Lemma 5.12.6 yields measures p and Y supported by A and B
respectively such that

7(A) > p(A) > 7(A), IIpII[BV (R")). < 1,

7(B) > v(B) ? 7(8), II^II i BV R" )3 . < 1.(

Define
A = p(A)v + v(B)p
and observe that 0 E med(u, A) for u E BV (S1). Since A(i) - 2p(A)v(B)
and IIAIIIBv(Rn)]. < p(A) + v(B), the conclusion follows from Theorems
5.12.4 and 5.12.10. ❑

5.13 Generalizations to the Case p > 1


Since a BV function u is defined H" -1 -almost everywhere by means of
(5.11.1), an obvious question arises whether the results of the previous
section can be extended by replacing hull"/(n-1);n that appears on the left
side of the inequalities by the appropriate LP-norm of u defined relative to a
measure that is absolutely continuous with respect to H". We will show
that this can be accomplished by establishing Poincaré-type inequalities
that involve Ilull n / ( n _1) ^ where A is a positive measure that satisfies one
, ,

of the conditions of Theorem 5.12.2.

5.13.1. Theorem. Let A be a positive Radon measure on R". The following


two conditions are equivalent:
(i) H" -1 (A) = 0 implies A(A) = 0 for all Borel sets A and for 1 < p <
n/(n - 1), there exists a constant C = C(p, n, A) such that

Il^lln.a <_ CIIDuiIR "1


for all u E BV (R" ).

5.13. Generalizations to the Case p > 1 271

(ii) There is a constant C 1 such that

A[B(x, r 11 < Cp r F(n-1)

for all balls B(x,r).


The ratios of the smallest constants C and Cl have upper bounds de-
pending only on n.

Proof. The case p = 1 is covered by Theorem 5.12.2, so we may assume


that p > 1. Suppose (ii) holds and let f € f7' (a), f > O. Then, by Holder's
inequality,

I /p'
f (x)da < f F'da A[B(x, r)] 1 / n
B (s,r)
1,
C ill f II F• ,Arn-
Thus, the measure f a defined by

f a(E) = f (x)da(x)
E

satisfies condition (ii) with p = 1. Therefore, by Theorem 5.12.2,

If uf da < MClllfllpf)u'IIBv(R")

for all u E BV (R'1 From the definition of Hausdorff measure, it is clear


).

that hf n-1 (A) = 0 implies a(A) = 0. Thus, (i) is established.


Now assume that (i) holds. For each Borei set A C Rn and each e > 0,
reference to Lemma 5.12.1 supplies an open set U D A such that P(U) <
ry(A) + e. Therefore, from (i) with u = XA,
4
A(A) 1/p < A(U) 1 /11 < CI1DXuII {R n )
=CP(U)
< C7(A) + Ce.

In view of the fact that 7[B(x, r)] = Cr"', (ii) is established.

With the help of the preceding theorem, results analogous to those of


the Section 5.12 are easily obtained. For example, we have the following.

5.13.2. Theorem. Let Ii be a connected, admissible domain in R n . Let p


and ) be positive Radon measures supported by Sl such that

14[B(x, r)l Ç C1rn-1,


272 5. Functions of Bounded Variation

a[B(z, r)] < C2-P( n-i) 7 1 < p < n/(n — 1),


for all balls B(x,r). Then, there exists a constant M = M(1) such that

IJu — (i)Ilp.a 5 M
C
Cs 2
IlDull(o)
u(1)
for all u E BV (12).

Proof. From Theorem 5.13.1 and Lemma 5.10.3 we have

llullp.a;n = lluollp,a
^
C ll nai l e v(Rn )

^ CIIu l HV( ^^.

Applying this inequality to u — û(p), we obtain

Ilu — u(1z)llp,a;s1 < Culu — fi(p)IIBV(n)


< CIlu — fi(1t)IIt;i1 + cllDull(0)
< CllDull(1),

Other results analogous to those in the preceding section are established


in a similar way and are stated without proof.

5.13.3. Theorem. If c E med(u, µ), then

Ilu — cll p ,a < M GiCs IlDull (n).


^(^)

Also,
If u(x) = 0 on A where A is a Stalin set of positive 1-1'-measure, then

Ilullp,a < M CA llDull(fZ).

5.14 The Trace Defined in Terms of Integral


Averages
For u E BV (I0), recall the following facts established in Theorem 5.9.6:
(i) E = {x : \(x) < µ(x)} is countably (n — 1)-rectifiable,

(ii) —oo < A(x) < p(x) < oo for fin I-almost all
- x E Rn ,
(iii) For Hn '-almost every z E E, there exists a unit vector y such that
-

n(z, A.) = y whenever a(z) < s < µ(z).


5.14. The Trace Defined in Terms of Integral Averages 273

Although our convention (5.11.1) of setting u(x) = z[A,(x) + µu(x))


allows a meaningful pointwise definition of u at fin - '-almost all points,
the simple example of u as the characteristic function of a ball shows that
it is not possible to define u in terms of its Lebesgue points Hn -1 -almost
everywhere. This is merely one of the ways that the BV theory differs from
the Sobolev theory developed in Chapters 3 and 4. However, in this section
we will show that a slightly weaker result holds:

lim u(y)dy = 1A„(37) + Pu(x)J


B(x,r) 2
1
for Hn - '-almost all x E SZ. If S2 is admissible, then a similar result will be
shown to hold for the trace u`, the only difference being that the ball B(x, r)
in the above expression will be replaced by B(x, r) n SZ. Briefly stated then,
a BV function can be defined pointwise H" - '-almost everywhere on SZ as
the limit of its integral averages.

5.14.1. Remark. If u E BV(R) is bounded, then it is easily seen that


(with convention (5.11.1) in force),

lim !u(r) - u(xo)rdx = 0 (5.14.1)


r^0
B(zo,r)
for x o E and that for all z E E for which (iii) above holds,

lim lu(s) - A u (z) I° dz = 0, (5.14.2)


r--+0 9+(sir)

li m °
Iu(z) — A u (z)I dz = 0, (5.14.3)
r ^ 0
_(x,r)

where
B + (z,r) = B(z, r) n (y: (y - z )•v > 0),
B - (z,r)= B(z,r) n {y: (y z) < 0),
and

lb verify (5.14.1), use Remark 5.9.2 to conclude that there is a Lebesgue


measurable set A such that /3(A, x0) = 1 and
lim = u(xo).
x ^xo
xEA
Then

lim r - " lu(x) - u(x o )I°dx = lim r - n lu(x) - u(x o )rdx


r^o f (zo,r)
7-00 B(xfo,r)nA

+ li O r " J
— °
Ili(r) — 4L(x0)I GIX.
B(xo ,r)f1A
274 5. Functions of Bounded Variation

The first term tends to 0 by the continuity of u I A. The second term also
tends to 0 because u is bounded and D(A, x o ) = 0, where A = R" - A.
Now consider (5.14.3), the proof of (5.14.2) being similar. from the def-
inition of (h(z), we have that D(A = , z) = 0 for each t > p(z). From (iii)
above, D(A, fl {y : (y - z) • v < O}) = 0 for A(z) < s < µ(z). Thus, if e > 0,
t- s<e, and s<µ(z)<t,we have

r—+0
r—.0
-n
f8- (z,r)
I u(x) - A u (z)I° dx < lim sup r -n e a
r—+0 fa' (z,r)fl(A, —Ai)

IB-(z,r) n AtU)4,)J (

+ sup lu(x) -„(z)]° 'limsup .


zE R^ 1--00 Tn

The last term is zero since u is bounded and therefore the conclusion follows
since e is arbitrary.
Our task now is to prove (5.14.1), (5.14.2), and (5.14.3) without the
assumption that u is bounded. For this we need the following lemma.

5.14.2. Lemma. If u E BV(Rn) and A (x o ) = µ,,(x 0 ), then there is a


constant C = C(n) such that
i/Q
limsup Iu(x) - u(xo)I ° dx < Climsupr l- °MujI[B(xo,r)].
•—•0 iB(=o,r) T •0

Proof. For each r > 0, consider the median of u in B(x o , r),

t r = inf f t : IB(x o , r) n {x : u(x) > t}I < IB(x o , r)1},

and apply Theorem 5.12.10 and Minkowski's inequality to conclude

limsup Iu(x) - u(x o )I°dx


r•—■0 8(zo,r)

< C lim sup r i- n IIbuII[B(x o , r)1 + C'I tr - u(so)i.


r-00

Moreover, t r -. u(x 0 ) since a u (x o ) = µ 4 (x 0 ).

5.14.3. Theorem. If u E BV(R"), then (5.14.1) holds for tfn '-almost -

all x o E R" - E, whereas (5.14.2) and (5.14.3) hold for H' -almost all
x a E E.

Proof. For each positive integer i, let


i if u(x) > i
ui (x) = u(z) if fu(x)1 < i
-i if u(x) < -i,
5.14. The Trace Defined in Terms of Integral Averages 275

Wt = {x : -i < Au(x) < pu(x) < i },


and observe that
Hn-1 (5.14.4)
wi) =a
1=1

by Theorem 5.9.6(1).
For each e > 0, let

r-*Q
Z = xo : lim sup
; {
IID(u - u;)Ii[B(xo r )] < F
c((n - 1)rn-1
,

_ (5.14.5)

and refer to Lemma 3.2.1 to obtain


EHn-1(U
- Z1) s IID(u - us)11(U) (5.14.6)

whenever U C R't is open. By Theorem 5.4.4,


00
I I Du(ü - % )II(U) <_ P[{x : (u - u s )(x) > s }, Ulds
0

+ P[{x : (u - u;) ( x ) < s }, U]da


00
00

= P[{ x : u(x) > i + 8}, U]d9


Q —
o
+ P[{x : u(x) < -i + s }, U]ds
0 0
- P[{x : u(x) > s }, U)ds. (5.14.7)
1.1>i

If U is bounded, then

P[{x : u(x) > a }, U]ds = IIDuII(U) < 00,

and therefore the last integral in (5.14.7) tends to zero as i --0 oo. Hence,
we obtain from (5.14.6) that Hn 1 (U Z ) -. 0 as i -. oc. Then,
- - ;

[00 00

H -1
"

) =1i -j
n (Rn -
Z; ) = 0,

and
Hn— r
(5.14.8)
j =1i =j
276 5. Functions of Bounded Variation

To prove (5.14.1), it suffices by (5.14.4) to consider xo E (U°__ 1 W;) — E.


Because u — u ; has 0 as an approximate limit at each such point 1 0 of W;,
it follows from Lemma 5.14.2 that
1/s

lirn ju(x) — u, (s) ^ ° dz


r^ 0 B(so,r)

< C lim supr l- nIlD(u — ui)111B(xo, r)1. (5.14.9)


r-60
From (5.14.8) we may as well assume that zo E n71 1 U i Z.. For i suffi-
ciently large, reference to Remark 5.14.1 yields
tro

urn 114(X) — u(x)l ° dx = 0.


r
-^ o (4(20,0

From (5.14.9), there exists i sufficiently large such that


va
lien iu(x) — u i (x)rdx < Ce.
r—.0 IB(zo,r)

Consequently, (5.14.1) follows for unbounded u E BV (Rn) by Minkowski's


inequality and the fact that e is arbitrary.
Essentially the same argument establishes (5.14.2) and (5.14.3). D

As an immediate consequence of the above result, we obtain the follow-


ing.

5.14.4. Theorem. If iZ C Rn is open and u E BV (0), then

Hm l u(x) — u(xo)I nl(n-1) dx = 0


r•p B(xo,r)

for H" -1 -almost all x o E SZ — E, and

lirn u(x)dx = u(x o )


r—.0 B(xo.r)

for Hn -1 -almost all x o E fi. If SI is admissible, then the trace te satisfies

11în lu(x) — u'(x o )Inl ( n --1) dx = 0


r—.00 B(s o ,r)Ilrt

for 11' 1 -almost every x o E OR.

Proof. The statements concerning the integral averages of u follow imme-


diately from (5.14.2) and (5.14.3). Also, referring to Remark 5.10.6 leads
to the last part of the theorem. ❑
Exercises 277

Exercises

5.1. Suppose u is a function of a single variable and that u E BV(a, b),


a < b. Prove that IIDuiR(a, b) = ess V'(u). Hint: Use regularizers of u.

5.2. Suppose Sl C Rn is an open set and u E BV (f1). Prove that there


exists a sequence of polyhedral regions {Pk} invading Sl and piecewise
linear maps Lk : Pk —+ R l such that

li M IL k —ul=o,

klit^ IDL k I = II^u^I(^)•


Pe
Hint: By Theorem 5.3.3, it suffices to consider the case when u E
CO°42) n BV(fl). Let P1 C P2 C ... C f2 be polyhedral regions such
that IS1 — Pk1 - O. Choose each Pk as a simplicial complex, so that
it is composed of n-dimensional simplices. Since IjDuII i o < oo, we
;

may choose k so as to make fil_p IDul arbitrarily small. Moreover,


we may assume without loss of generality that each simplex, a, in
Pk has its diameter small enough to ensure that the oscillation of lui
and 'Dui over o is small, uniformly with respect to all a. Suppose a
is spanned by the unit vectors v l , 1) 2 , ... , v n . Define the linear map
Lk on a so that it agrees with u at the n + 1 vertices of a. Clearly,
IILkJI^,Pi IIuIIi,n. To see that the L 4 -norm of the gradients also
converge, note that

'Dui = IDu(p)I Ioi


s
for some p E a. On each of the edges of a determined by the vectors v^,
i = 1, 2, ... , n, there is a point p, such that [Du(p i ) — DL k (p.)] • v, = 0
(by the Mean Value theorem). But DLk(p,) = DLk(p) since Lk is
linear and IDu(p i )I is close to IDu(p)I because of the small oscillation
of IDuI over a. Therefore, I Du(p)—DLk(p)I is small and consequently,
Ja
f
IDLk I is close to a 'Dui, uniformly wiLh respect to all a.

5.3. if E C R" is a measurable set, let us say that E is open in the density
topology if D(E, x) = 1 for each s E E. Prove that the sets open in
this sense actually produce a topology. In order to show this, it must
be established that an uncountable union of density open sets is open;
in particular, it must be shown that it is measurable. Hint: Use the
Vitali covering theorem. If we agree to call the exterior of E all points
x such that D(E, x) = 0, we see that am E is the boundary of E i n
the density topology.
278 5. Functions of Bounded Variation

5.4. In the setting of metric spaces, there is an inequality, called the Eden-
berg Inequality, that vaguely has the form of the co-area formula. It
states that if X is a separable metric space and u : X -' R" is a
Lipschitz map, then for any E C X and all integers 0 < k < n
• — k )a(n) k n
f"
n
H -k (E n u- 1 (y))dHk (y) < 2 na(n a(n) H (E).

Here, L is the Lipschitz constant of u and f • denotes the upper


Lebesgue integral. Also, Hin denotes m-dimensional Hausdorff mea-
sure which has a meaningful definition in a metric space.
STEP 1. By the definition of H", for every integer s > 0 there exists
a countable covering of E in X by sets E . i = 1, 2, ..., such that
; ,

diamE;. <11s and


00
Hn ( E) > ^n ) ^ (d i am E;,. )" — .
8
i-1
Hence,
o0
a tà —k
H" -k (E n u '(y)} <( "_k li rn i nf
- ) [diam (E;,„ n u-1( y )jn--k
2 ^oo
i=1
.

STEP 2. Consider the characteristic function of the set u(A), X(u(A)),


where u(A) denotes the closure of u(A). Then

Idiam (A n u - 1( y ))]" - k < (diam A)" —k X(u(A), y).


Hence, from Step 1,
00

K r" (E {l u -1 (y)) < a( " kk) liminf E diarn E;, a ) n—k X(u(E), y).
2 •^ 00 i=i
(

STEP 3. Apply Fatou's lemma (which is valid with the upper integral)
to obtain
a(n — k)