Editorial Board
J.H. Ewing F.W. Gehring P.R. Halmos
Graduate Texts in Mathematics
Weakly Differentiable
Functions
Sobolev Spaces and Functions of
Bounded Variation
SpringerVerlag
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
987654321
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 nonlinear
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 198788 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 spacesthe 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
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
Multiindices
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 nballs 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 LPSpaces 18
Integration of a function via its distribution
function
Young's inequality
Holder's and Jensen's inequality
1.6 Regularization 21
LPspaces 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
Nonincreasing rearrangement of a function
Elementary properties of rearranged functions
Lorentz spaces
O'Neil's inequality, for rearranged functions
Equivalence of LPnorm and (p, p)norm
Hardy's inequality
Inclusion relations of Lorentz spaces
Exercises 37
Historical Notes 39
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 LusinType 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
5.8 The GaussGreen Theorem 246
The equivalence of the restriction of Hausdorff
measure to the measuretheoretic boundary
and if DXEII
The GaussGreen 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
measuretheoretic boundary
5.11 SobolevType 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 selfcontained. In particular, most of the in
formation contained in this chapter will be wellknown 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
d(z, E) = inf{Iz – yI . y E E }.
It is a simple exercise (see Exercise 1.I) to show that
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 ntuple of nonnegative
integers, a is called a multiindex and the length of a is
n
^a = xa1 l • x2
2 •' •
^an
n
D° =D,'.. D = a x Q'...a x
1 nQR •
The gradient of a realvalued function u is denoted by
C k (SZ) = {u : °
RL, Dau E C (St), p< lai < k },
1.2 Measures on Rn
For the definition of Lebesgue outer measure, we consider closed ndimen
sional intervals
I = {x .a;< x < , i1,...,n}
;
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
i=1
Moreover, if E1 C E2 C ... are measurable, then
00
U;_i
Ei = ili rn'Ei I ^0o
(1.2.5)
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 oalgebra, A, con
tains a sufficiently rich supply of sets to be useful. This possible objection
is met by the following result.
A U B C U Ik where
k=I
1.2. Measures on R" 5
00
Because d(A, B) > 0, there exists disjoint open sets U and V such that
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
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 nonmeasurable. Let
1
T =An x. 1 <d(x,C)< (1.2.13)
+1
;
A— Ci(A—C)U (ÛT; )
(1.2.14)
^=1
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,
^
j_
00
[! 2 ilai b:l —
1+1a,— b,)
p:Rn xN—.R"
then there exists E C E which is maximal In the sense that it is not a subset
of any other member of E.
u.F';.
00
.F= CI
i 1
E {UB:BE.r}c{UB:BEF.F`}
xE{UB.BE .F^^`}.
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 nballs {B(a, h(s)) : s E S} is disjointed. Let
Then
whenever s E 52 and
or from (1.3.3)
Ac U{uB:B E B{}.
i1
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
(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)
To complete Step II, it remains to estimate the number of points in the set
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
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
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
2tl
1 , ... , 8 M=
t=1
2i
14 
1
00 00
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.
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
which implies
Thus,
p(A{UB: B EBk, UBk 2 })<(11/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 .
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
f(y) CLAW
B (s,r)
denotes
,IB(x, r)]1 f (y) dµ01)•
B(z,r)
]i m
r^d i
f (y) dp(y) = f (x)
B(= r)
+ 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
Hence,
The results above show that the RadonNikodym 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)
r0 v(B(x, r)J
,
[ (
; 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.
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 11 7 (E) agrees with intuitive
notions of "ydimensional area" when E is a wellbehaved set. For example,
it can be shown that H" agrees with the usual definition of ndimensional
area on an ndimensional 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.
"
Therefore,
Ç HE (U) <
1/:(E)
1
E H£ (B.) < E 2na(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 ndimensional 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
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
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)
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 kdimensional 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)
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.
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
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
take p' = co and IIvIIp ;n = esse sup M. Holder's inequality can be ex
,
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
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
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 nonnegative, realvalued function in Co (R") with the property
that
cp(x)dt = 1, spt cp C B(0, 1), (1.6.1)
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 multiindex a.
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
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,
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 uv llp+llU vE lip +IlvE tic lip <_ 317.
Hence u, u in LP(Rn) as e O. ❑
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.
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
T (cp) = f (p'(z)dx
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
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
D; Di T = D • D 1 T.
27
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 nonincreasing 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,
which implies that f'(t) < g(t). But g(t) < fi(t) and therefore the proof
is complete. ❑
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). ❑
I^f Îi p = ( f 0o
r t1/Pf(t)lPdt/1
)
)
1/13
28 L Preliminaries
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
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
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 nonincreasing). 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
and
h+'(t) < a rg'* (t).
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 .
and
02111 <_ IIfllillgalli < a
a
a9(s)ds. (1.8.13)
= 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
Let
f(z) = fi(z)
i=t ‘00

where
0 if If (z)1 < yit
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).

The series on the right is an infinite Riemann sum for the integral
a f (y)dy,
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(Yi09.6(af(Y81))(Yi — y+1)
=t
1
f' (t)
af(y)9 ' (CI f (y))dy•
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
Eafohog"(af(y._,))(vsv,_,)
i =1
00
`
Eui00'( u i1)(f'(us)  1 y(uti 1))
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)
tg .• (t) f■(t) _
^
Ag ..
(A)f'(A) E
1=1
^
^
1
= r
ui +1g " (ui +1) {f (?ti+ I ) — f (%)1
'
.
1=1
+ „,+ .
Ef (ua) [
u,
g' (Odd
1
1
i
Ç Eth+oscui+otf.(u,i+i)  f'(ui)1
i =1
+ E f'(ui)g'(ui)[ui+1 ^ ui]•
i=1
^
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
^ [
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
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
and
00
+ (f(u) — f'(u))g"(u)du
t
^
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
i
co cc
F^x)Pxp r ldx < (= f (t)ptFrldt
l r r)
a
f
0o Z
f (t)dt s r ldx
0
f
(
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.
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
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))r4(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.
R°
1.4. Prove that C"(11) is a Banach space with the norm defined in Sec

tion 1.1.
a^
D'( fT)_^ Dg f D^ AT —
1.7. Lemma 1.8.13 shows that if 1 < p < oo and 1 < q < r < oo, then
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 G6set G D A such that
HI (A) = H " ( G ).
38 1. Preliminaries
1.12. Using only basic information, prove that the class of simple functions
is dense in the Lorentz space L(p, q).
[u _ û niJ =0.
t1
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
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
u(x + h) — u(x)
uh(x) = h
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 multiindex 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
(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 nonoverlapping 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
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 sidelength 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 kdimensional measure could be
defined for integer values of k. kdimensional 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 [SCI1] 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 nonincreasing 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 1.8.71.8.9,
were first proved in [O]. The reader may consult LCA21, [CA31, 1LP], [PE]
for further developments in this area.
2
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.
Ilullk.p;sl = f E IDa^;lp^
1 j <k
a
(2.1.4)
E IID°ullp,n• (2.1.5)
IQ!<k
li ô NE  u I'Ik.p ;n , = 0
44 2. Sobolev Spaces and Their Basic Properties
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,
= (1) 1'an
1 Dÿ (P (L=1) u(y)dy
n a
= en
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^
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
f iD 1 codx = — f D,rtSp dz
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
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 ,
 — uD cp dx.
;
R "
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.
o
H(t)= t^
Ot < 0 —
for j = 1, 2, then there exists C = C(p, M) such that for any nonnegative
v E Cp (R")
IlDaH(u)Ilp < ClID2vIl,
for 1 < p < n/2 and any multiindex 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 nonnegative function in
Q
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 It21H(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
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
_
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
T: .
That is, for some constant M, we assume that both T and T` 1 satisfy,
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
"
50
Note that
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
=  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.
"
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
Q(x, v, t)  1(x +
hi) — f (x) D (x) • v.
t
For v, v' E Sn 1 and t > 0 note that
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
I
formula
f o T J2` dx = f dx (2.2.9)
T(E)
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
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
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
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,
;
54
If E is open, let
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
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  ,li1) and
00
E f 1 on a (2.3.1)
i^l
2.4. Sobolev Inequalities 55
u(x) = E fl (x)u(x),
; 1
and consequently,
^
i =1
The conclusion follows from the Monotone Convergence theorem. ❑
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 nball with its equatorial (n1)plane deleted. The function u defined
by u  1 on the top halfball and u —I on the bottom halfball 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.
IlUllnp/(np);fl C^IDuI^P,^,.
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/n1
o0
/n1 < 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/(n1)
< (f ID l u tt , z2 7 . . . , z„ ) Idt
00 n 1/(n])
I Diu #^ i
dx1
 oo ^2 (f:va
00 1/(ri1)
< 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/(n1) <
n
L .
fID(lur)lidx
< q luI¢—'IDuldx
R^
ll l u r ^ 1 l6'11Dut!p,
q■/71'
that
Iu(x)I < IDu(r)ldII 1 (r) (2A.4)
a=
where A x is any ray whose endpoint 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 endpoint
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
1/p
w(n  1 )1/445 IIDull') (f Ix  1/1 (1n)p' dy , (2.4.6)
to
and because Isptu  B(x, R)I = IB(x, R)  spt ul, it therefore follows that
I x  y l(1n)p' dy <
Y1 (1n)p 'dy.
apt u E1(z,R) f (r,R)—spt u
I' 
Consequently,
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 nball. But
a(n)Rn = lspt ul and therefore
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
k— ^
where
qk—Y = np/(n — kp + p)•
An application of (2.4.12) to y = D i u, 1 < j Ç n, yields
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
k1
P=(y) _E 1i D ° u(z)(y  x)a
ial = o Cr.
and
1
( y) _ kE
Rz ar
lai=k
^
[f0.okiDo.(0  t)x + ty)dt (y  x)a.
k1
%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
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
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 RellichKondrachov Compactness Theorem 61
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.
where A 1 (0) is the ray whose endpoint 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(1n)P dy (2.4.18)
13
But,
1 /p'
U, yl(ln)Fldy
r
= l7'la(n))1/P'rlnip
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. ❑
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.
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
_
IDue(x)I < Iu(x — y)I ID(Pe(y)Idy
B{ox)
€—n—i
sup{ID'(y)I : y E R" } IIuIIi
< en1 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
_
1
< fmomo jDu o #y(t) •' '^ ' (t)IcpE(rJ)dtdy
<^ IC l IDuIdr^e.
Thus, B is contained within an rneighborhood 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
where
A = 1 /g"(nrkp)/np.
1  ( n  kp)/r^p
Then, by Theorem (2.4.2)
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
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.
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:
A') = (21)n/2 a+z f(1/)dv. (2.6.1)
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)
IrIQn)
+o
if 0 < a < n. Thus, it follows for some constants C 1 and C2, that
g a (x) < C1 ae
IxIn 
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
u= 9Q*f
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
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
(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 ),
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 nonnegative function fi E IP(Rn) such that
i =t i1 i 1
with
Ba,p(Ej) < e2 3 .
Hence, ga * f; ► g a * f uniformly on R"  U? _ 1 E3 , where Ba,, p (U °_ 1 E) ) _<
?
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.
gi  i1 E f f
3= i
9Q * f(x) = f lirn
^oo
g o, * gi(x)
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
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.
—• 00 —*CO
Proof. For the proof of (2.6.14) assume that the limit is finite and let f,
be a nonnegative function in L" (Rh) such that ga * fi > 1 on Ei with
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
(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 ;
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.
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
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 .
sup
R^
g^ *f du: f?o, VG ^
l
J1
and thus obtain
F( f, v) = f g * f dv
Q (2.6.22)
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
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
,
and therefore
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)
for x E B(2r). From (2.6.3) and (2.6.4), there exists C = C(a, p, n) such
that
and therefore
However,
Hence,
J R. [C3rcif ()}'dy = C
2P r n°' p 11f11p.
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
< C(r)r°
which, by (2.6.28), implies
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 fxIan if r/t < Ix' < 1. (2.6.30)
B(r)
If lx1 > 1, then (2.6.3) yields
(2.6.31)
B a , p (E) > C
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
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 righthand side
involves Hausdorff 0dimensional 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
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 MorseSand Theorem [MSE1], [SA], states that a realvalued 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 1dimensional 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.

H"1[u1(^) 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 MorseSard the
orem referred to above. In preparation for the proof, we first require the
following lemma.
jdiv
sup dx : cp E C (R";R"), supp] <_ 1 = H" 1 [8U].
i
{
1 divSp dxj,(x).
!c r^(x)dH"
U
I (x)
sup div (09 dx :çP E CQ (R"; R"), suP iWi < 1 < Hit1 (8U).
u
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 n1 (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 L1(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
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.

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
Du^dx =
R" N R"
In order to prove the opposite inequality, let Lk: R" e R 1 be piecewise
linear maps such that
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
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
lini[u1(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 ,
H "1[u1(t)] <L
t d1 dx+
— (P• v dHn 1 + e
8E=
< fInt[Lk1 (t)1 + e.
Thus, for t ft S,
 k ^ oo R"
_'Duldx. CI
R"
I, ^
f(x)dHm(=) =
Y
n f 1 (31))dH k (g)
g(x).If
(x)dHm (x) =
J f '(!I)
g(x)dHmk(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.
= f`(t)dt
0
< n 1 a(n) l in Nn1[13t1dt.
f0 00
The coarea formula, Theorem 2.7.1, shows that the last integral equals
l DuldX,
R^
thus establishing the theorem.
(P 1 )/ v
5_ n l a(n) E/n q unP/(nP)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) =7r1/2n1/2
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
Fh(x) = 1  min[dK(x), h} • h 1
I h
Hr [dxi (t))dt
0
= Hn  ' [di' (t h ))
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 HardyLittlewoodWiener maximal theorem.
EIBkI?5  nlAtl
k =1
and therefore, from (2.8.1),
DO
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
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 tplIAeIdt
0
00
< p2 • 5" 9 2 I f(d^ dt
o flifi>02}
p2p • 5" tv2 f{ifi>e} If ldx dt
o
00
= p 2P • 5" tP2p({1f Î > t})dt
0
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:
(ii)
If(y)Idy < C 613
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 
,
I If(^)Id y
(z,6) Ix —
yln a = ^ À(=
ko
vo
,
6 an
^ ^)
;^ Ix
If(y)Id^
yIn a
k0 1
^ C 6a M(f)(z)+
where a(n) denotes the volume of the unit nball. 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 ).
I f we choose
d=
II/11p ) ,
( M(f)(x)\Din
YI (an) 1/4
Ix  If M I = ( Ix  1/1 (an)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(ao)rdEl
• Ix  If(y)^pdv •
n n
fa (raf )Qdx [f
n c
ix  ^I (a
t

n)r
If (y)IPdzdy
Thus,
* Ip = fQ+A ,
a + Li < n,
elegant proof of this fact which follows easily from the estimate discussed
in Remark 2.8.5.
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
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
90
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.
^ un1
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
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
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
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
1
a$
u(s) = e'z
" .11 û(092 (âi1 R)dy + 4u(^)rl(^/R)dl!
=
`^ p
R max( 2 ,IIUII46 ).
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
LR(x) = R "L(Rx).
2.9. Limiting Cases of the Sobolev Inequality 93
IILRIfn, ;
CrIp
and therefore
)lip'
1/P
+C e—C' !TIP' dx
11<lx1<co
because
IILRII'1= IILIIi = C < oo.
94 2. Sobolev Spaces and Their Basic Properties
=( 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
I KR( x )l
I 77(yIR)1dy
R"—B(Q,eR) M
00
< C r"^edr
ER
< CR "
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)
where
LR(x) = RT L(Rx), L(y) = 9 9 (y)•
We can rewrite (2.9.13) as
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
Pclz11x1(tn" Iyllrdx
I9t(x  y)  9t(x)IQ'dx < C
96
Cf ILR(9)t 19I 6 dy
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. ❑
— 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
[d1r 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.
98
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.
Ia * f E L(r,q)
and
Since
f = 1
i
(j ' co(s)ds ) d(1  log t)—p+'
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
1 if f (z) > 1
f i (x) = If (x) if — 1 < f (x) < 1
—1 if f(x) < —1
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
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
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 `
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
.
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
(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)
1
00
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) ;
(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,

Proof. Consider part (1) first. Because u' (t) is nonincreasing we have that
/(r1)
l u'(t) I*' < K(1 + I logtl)I(t)I
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 /
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 *

fcPdx = f gcodx
{ suPf Q * f(2)
EE
Ba,p(E) > C
u E Cfl (R"))
With the aid of Theorem 2.6.1, prove that there exists a constant
C = C(p, n) such that
Lk: R" . RI
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 nonincreasing,
'
otu(t) dH "' .
IDur l
>u (t)
[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
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 >
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
(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
KC B(r i ) C B(2r,) C U
i = l i =1
and 00
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 .
IDukI dx < C
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 "
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'
where
v(x, r) = v(y)dy.
B(s,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
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.
2.19. 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
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 biLipschitzian 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 bimeasurable 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 Kquasiconformal mapping, then
uoTE W1 (R").
Historical Notes 109
for all cp E W 1 '( 1l) n L°°(1). Then there exist u E W "P(Il) and a subse
,
Havin and Maz'ya NM], and Resetnjak [RES). Most of the material in
Section 2.6 has been adopted from IME1).
2.7. The coarea 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/(n1)
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) =
np { 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
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 kthorder
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 LPnorm. 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 LPnorm 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
17sense 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 Lusintype approximation for
Sobolev functions.
(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 selfcontained 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 HardyLittlewood 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.
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
(3.1.2)
J (sp,r) u(v)div v (v)dv = Du(g)
B (xD,r)
• (y  ro)p(Iy = xoI)dy.
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 17 1 11 (014 5
Rn
171+1
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 GaussGreen
theorem,
j Rn
JR
div V;(y)lu(y)Idy = rV(y)' D(IuI)(3►)^dy.
n
L
1 (1/2)(tn2) 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
3.1.4. Theorem. Let k be a positive integer such that kp < n, p > 1, let 0
be a nonempty 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.
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.
u(y)dy = n
f ( s, i ) [Du(y) • (y — x)1dy
^ n
u(y)dy —
f (x,l} f (x,6)
 1 an
[Du(3I) •• (t! — x))dy
a(=,6)
— 1 ly — xr [Du(y)  (y — x)]dy•
nf <ly—sl<1
n (3.1.8)
x1 1 "iDu(y)idy (3.1.9)
Iv — < 00 .
B (s,l)
116 3. Pointwise Behavior of Sobolev Functions
which, by (3.1.6) and (3.1.7), is finite. Thus (3.1.9) still holds when k > 1.
By (3.1.9)
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 nonnegative CO° functions on R", with compact
supports, and spt (p i C 1 for all j such that the interiors of the sets
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. ❑
3.2.1. 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
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,
2
BEY' BEYY•
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
BEYY•
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. ❑
Proof. The result is obvious for i = 0, so choose 0 < a < n. For each
positive integer i let
p[B(x'r)] > ii
Ai = R" fl x : lx, < i, lirn sup
}
r—.0 r°
3. Pointwise Behavior of Sobolev Functions
118
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 ;
3.2.3. Corollary. Suppose u E LP(R n ), 1 < p < oo, and let 0 < a < n. If
E is defined by
r— 0
,
then H°(E) = O.
n
p(A) 1u1 p dx.
A
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
u E W k 'p(R n ), (3.3.1)
then
bak 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 kA .
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
f n(x,i)
ly
 xla  k+ l  niDu(y)Idy < oc. (3.3.6)
ba k Iv _
xl l nlDu(y)Idy ■ 0, (3.3.7)
f <lyxj<1
6Akn 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),
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(tk)p u(j)  u = . r l p dv ► 0
4 (m,r) (3.3.12)
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 '
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(tk)pn
J(x,r)
ju(y) f ux,rIPdy < Cirif(k1AP_"
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
1/p
+ l(Diu)x,r l j [ dv } , (3.3.20)
6( x,r)
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 12 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)
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. ❑
3.3.5. 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
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/(nkP)
IAI GIBk.P(A)1
U
i1
Ai n B(x , r i )
Q
is thin at x 0 .
Then,
1
]
Theorem. Let xo E R", p > 1, and suppose u E W' '(R") has the property
that
"(p1)
dr
J: '
r p—" IDuI
B(xo,r) r
dx < oc.
T../(2r) = u dx.
B(x0,2r)
3.3. Lebesgue Points for Sobolev Functions 125
1 )p
B1.1,1A(xo e) B(xo, r)) < C(2e
, ^Dvr^pd^
B(so,2r)
A= , 2
^
1
is thin at x0 . Clearly
lim u(x) = u(X0)
s oso
—
sER" —A
for B 1 , p q.e. xp E R" cf. 1ME31. Therefore, with Theorem 3.3.3, we obtain
the following.
for H" — pa.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.
exists for all x E R" — E and for each multiindex ar with 0 < led < rn.
Thus, for all such x, we are able to define the Taylor polynomial em) in
the usual way:
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
[
L(z ) r) l
(
u(y) p=m t)(b)^pd^1
< Tm
Iai=m
a! o ^ t
(^  m
1/p
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
„` ^
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 righthand 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
cPE(x)dx = 1, (3.4.9)
Rn
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. LPDerivatives for Sobolev Functions 129
so 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
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<ly2I<tr B(y,e)
and by (3.4.8)
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
Bkm,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.
as r j 0, for all laf = m for all z E R", except for a set F with
Bk_mp(F) = O.
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. ❑
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,
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.
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
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)
1Rz(0Ipdy < Mr k .
(IB(z,r)
B
1 1B
ISz(y)ldy <
(z,r) /p
(z,r) B(x,r)
FSz(y)Ipdy < Cr', 0 < r < oo.
at most k 1, we have
—
r° lLx(y)Idy =
8(x,1)
I^x(y)Idy
B (a ,r)
< Cr k + Cr ki , 0 < r < oo.
ci)£ * P=P
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 mtuples (y} whose compo
nents are indexed by multiindices 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
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,
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
Q(z) = E bQz ar
o<iaI<k
it follows that
çp(z)Q(z)dz Q(0).
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 Lwconclusion 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.
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).
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
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
Let 0(x) = H°(S M ) < C(n) and let rj : RI , 10,11 be of class C°° with
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
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 11 0(x)N(a)h(x) 1 faM, for x E U  A. ❑
where
1 /p
0
= cp6(x) * (D P2•)(z) + f Rp(x , y)R1 . (y)db (3.6.4)
and consequently
f (s•,Kd(x))
IRx• (y)Idy < M[Kd(x)} n+k (3.6.7)
3.6. An ImoVersion of the Whitney Extension Theorem 139
(3.6.8)
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
where
IR0(4, x')I C(a, k)M1x'  x l lk  lal.
By Taylor's theorem for polynomials, it follows that
k11091 1
D^ Ps • (x) _ E — D°+aPs. ( x +)( x  x')a.
a!
k 1 1091
DRP=• ( x ) _ LrDfi+aP=` (x • ) + ^+S(zi, x+)1(x — x' a )
L4
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)
Ix  al <Ixv1+lyal S 31x_yl.
Thus, utilizing (3.6.12),
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,
2d(x o ) > d(x) > d(3I)  ix  Ili > d(11) > ix  1/1.
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
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
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
= 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)
lu(y)jpdy = o(r°) as r 1 O.
( (r.r)
3.7. An Observation on Differentiation 143
U ={x:d(x,A)< 1}.
lien Iu(y)Idy = 0
(x,r)
r~0 B
{UB(s, 5h(s)) : s € S} J U  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'
1 /p l/p
Ix — vI? Ix — sI —
Is — yI>d(s)  5h(s)= 5 h(s).
1. <
dx
C
— yln +a
Ix
°°
i 5^(a)i
< C(a)h(s) — °.
—^—
r'dr
144 3. Pointwise Behavior of Sobolev Functions
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
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
lu(y)Idy
Ix — É/1 " +a < o0
R"
B(x,r lul
)
< Cr a .
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,
From Theorem 3.6.2 it follows that there exists an open set U D E and
riE C k '' 1 (U) such that
(fB(x,r)
in — i,IP < Cr k
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,
1/p
^
(wt * Ps)(x + h) + D°cp t (x + h — y)Rs (y)dy. (3.9.2)
148 3. Pointwise Behavior of Sobolev Functions
and therefore D'P1 (y) = D' P1 (x) for all y E R if (al = k. Hence, "
thus establishing (ii) if Pal = k. However, if 0 < l < k, then u E t t .P(x) and
the associated polynomial is
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.
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
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)
( )
= 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 nonnegative
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
— N(y)dy•
Thus, the distribution T + N has the property that
(T+N)(7i)> 0 for E 2(f1), ip > 0.
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,
^ 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
IxxoI<t
for almost all r o E Sl. This implies that
for all measurable E C Sl. Since p(E) > v(E) it follows that the measure
p  N = T is nonnegative. ❑
tES
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 nonnegative
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
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 nonnegative measure in S2. Thus, for E Cr 1l), lb >15,
T(?i) — f f idx ^ O.
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 nonnegative measure. Similar reasoning applied to the func
tion —u implies that Da (—u) — (—u a ) is a nonnegative measure or equiv
alently, that Dau — u a is a nonpositive 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.
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 Wk1,p0 1) 
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") .
Et = f x : Mp ,Ru(x) > t}
and choose r E Et . For notational convenience, we will assume that x = 0
3.10. A LusinType 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)
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
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
—
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
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
Ç E Bk,p(B)
BEf
5. E (50 nkp (by Theorem 2.6.13)
B(x,r)EF
<^ Mf (y) p dY
EB
BEf
Bk ,p lEej < Bk, p [{UB : B E gill + Bk, p [{UB : B E gall < tF Ilf lip ,
lu(v) u(x)(Pdy . 0
B(x,r)
uniformly on Rf2  U as r J. 0.
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
and therefore,
158
+^ _
and therefore, by the preceding theorem and the definition of capacity, we
obtain a constant C = C(k, p, n) such that
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
Let
00
B(x,r)
Iu(y)  U(X)14 0
if 1271 < T
r r (s) =
an otherwise.
3.11. The Main Approximation 159
Then
lu(y)Idy = r r * IuI(x)
• 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 nonnegative 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.
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.
3.11.1. 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 nonempty
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
where K(x, t) denotes the closed cube with center x and sidelength 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
Proof. For A E (0,1), let Ka denote the set of all closed cubes of the form
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
where
a
W ( x) = j^ ^[( z  Qik)/24 (3.1 1.5)
k=1
Now, for r E P, and any multiindex a with 0 <10/1 < t, we have
where A l depends only on t and n. Hence for almost all x E P; and any
multiindex 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
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)
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
so that by (3.11.10)
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
so that Iz! < 2141, IxI An < Alzla^" and (3.11.12) holds.
,
An
ë n IxI dx 5_ &n IxlAndx = Cb an
B(:,6) 13(0,46)
Ôn f
Since we have shown that
for all z E R n , the general result follows by a change of variables; that is,
replace z by z  y. ❑
so that
and by (3.11.15),
> y(k)t —n IU fl B(b, t)I.
Hence by (3.11.13),
rCa
Iy  bl kn ^(^)db ? a.
7(k)JRh1+
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
3.11.5. Lemrna. Let p > 1, k a nonnegative 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 nonempty 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 ,
E
o <t<1/2 in 2

F= aU  E.
Then, for each x E F there exists t E (0,1/2] such that
It follows from (3.11.18) and (3.11.21) that y has the required proper
ties. D
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
Rtm,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
so that by (3.11.24)
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 nonempty compact sets, such
that
C; C Int C‘ .4. 1 (3.11.27)
for i a nonnegative 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
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)
so
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.

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 ) — Cc1 • ( 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)
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
lirn
r ^0 8(z,r)
1u(y) — u(x)Idy = o
Ak(x) = u(y)dy
B(s,)
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.
le m
r D
Iu(y)  u(x o )jdy = O.
B(zo,r)
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
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
4/t(r) _ JKr U
utlF + ' Du t I)dH n1 .
1/p
iu t (z)I < Mr 11 Iut[ P dH " 1
KT
1/p
1
+ Mr' IDut1pdHn
Kr
/P  Mr113S001(r)1/P
< Mr13Spt(r)1
< [M 213 + Mlazo (t) 1 /2P ,
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
a < y < b, lb  al <17 1 , lkl < lb al}' > (1  E)lb  aI.
Exercises 173
(xo+h,yo+ k) E E}.
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
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.
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 LIsense 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
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.
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
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 seminorm 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
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
< 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,
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 mk 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,
4%0= E T(Pa(Du))x«
«1Sk
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 multiindex a. Consequently, by Taylor's theorem for polynomials,
In particular,
D'P( x ) = aaa !
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
1171 •T(1)r +1
C(l, p, iHI) IITI I (4.1.8)
a&T(1)1+t 1)
a!T(1) ( )
If lal _ k1, 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)
^. k+ 1
< C'(k,p, I^I) II l#
T (Z)
In general, if lai = k — i , k > ti, we have
_
< C'(k,p, If2I)IITII
T (1) k+1 
Proceeding in this way, we find that
k+1
fill
^
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 mk .p(1l)) . is a non
negative nontrivial 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,
L(u) = E p ( Pa( Du )) xa
iai<k
C = C(k, p, M).
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)
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.
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
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
^ 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).
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
_
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.
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° 1m
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 multiindices 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). ❑
T = E 1 101I7'va ( ) (4.3.4)
1a1= 0
4.3. The Dual of Wm'p(SZ) 187
T = y + E (1)I0ID°`ya (4.3.5)
Ian=m
la1= f
IT(92i ) — 4.02)1 = E vaDacp i — ya Da vjt
101.4)
m
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
T = v + E (1)Ia!Dckva
v, v a E (S/) •
The dual space (Wô 'P(f2))' is denoted by Wm,p'(A).
188 4. Poincaré Inequalities—A Unified Approach
_ 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),
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

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
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
Proof. Let
=u ds and Q = u dx
A B
and define T E ( W i .p(Sl))' by
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
(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 .
i1
i =k +1
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".
where Et = {x: v(x) > t). By the MorseSard 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
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 itmeasure O. Consequently, by Lebesgue's dominated
convergence theorem,
f V c dL —+ fv dµ.
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. ❑
and
f u dHn 1 S C ^llüll p • in + 11Dullp,nl
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
u =9mk * f
where f E I1(R n ) and IJ f II
suppose that p is a nonnegative measure with the properties that
Now suppose
spt p C 1 and
9mk*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
Thus, p E (Wmkip(f2))`.
4.5. Poincaré Inequalities 195
119mk * <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 <_ CIIgmk * pIIp' <_ C.
The result now follows from Theorem 4.2.1 and Corollary 4.1.5. ❑
II^IIk,P;^ C(Bra_k,p(N)) 
LIP IID k+l ullm(k+l),p;n
and
k+1
IIuIIp•,O < C (Bink,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. ❑
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 nonnegative Radon measures com
pactly supported in N such that
9m(kj) * E L P. (R " ).
Consider all functionals of the form
where Ia l < k— j. We will verify that all such T are elements of (Wm P (0) r.
,
Daudµ = f D°lidµ.
Da ii = 9m,a1 * g
4.6. Another Version of Poincaré's Inequality 197
DaU dp = f 9mIoI * tz • g dx
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
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 InequalitiesA Unified Approach
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.
IIuIIk,p;n CllD
k+1 fllm (k+1),p;t2
1 B ^
1Ip_ +
Po Pi
1  1 B B
I4 + go Pi ^
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
where aq/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).
discuss further what other measures play an important role in the inequality
of Theorem 4.7.2.
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
jl < (n — k) R l/p
I1111 M'1p
(fR.ILL[B(x,r)]dx) rkn1+(a1p)dr.
7(k) o
(4.7.3)
We now will evaluate
Ftt[B(x, r)]dx•
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 (^)
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
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 righthand side attains a minimum. An ele
mentary calculation shows that
1la
R =
N(At)
M ►
202 4. Poincaré Inequalities— A Unified Approach
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/ _ 1B + 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
for every Bore] set E C Rn, then there is a constant M such that
U .
Iu.(x) I A. dHa(x) <— Cilfllp (4.7.7)
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 leftside 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 nonnegative 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.
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
Rn
(fk * p) • ( Ik * 11)1/(pl)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
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,
,
R*
p ^ dt
< C r
p
t+[B(x, t)^
tn —k t
dx.
—
R"
and therefore,
00
kn 2r dr
9k * f (x) ? C f (y)dy r e
o B(s,r) r
a, 1/(P1)
dr
^^ ( r n 9k(y  z)d►l(z) dy Tkn e 2r—
0 BfV.r1 T
(
r
)1/(p1)
>
C ^, n rk ne e  2r dT
ardF^(z) dy Tk n
r
o B(y,r)
i/(PI)
> C fo w F^rB(^J, T)1 e2P
,r dr
 Tn  kp r
This implies
tic,
R" o Tnkp r
>c l^(^1 i/(p1)  2p,
' dg y)
^ e r ^r
T^Bnkpr)^
r
L
,^I l 4B(Yr)] 1 /(P 1) dr
2 (
? Ce
P — du(ll)• D
T R" r
where T E (Wm k, P( 11) )' and L : WmiP(SZ) ' Pk(Rn) is the associated

T(Pa(D)u)Xa
L(u) =
1
Ek
that
L(u) = E f P° (D)udp = 0. (4.8.1)
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 nonnegative measure on
R n such that p E (Wm  k , P(fl))*, µ(Rn) # 0 and
then
IlUIIk,p;i2 < CIIDk+1uIlm—(k+l),p;S2
and
CIID lc+l ullm_(k+1) PP
I
JM
AnR D"u dH a = 0 for all 0 < < k
and
Hullap';ct < CIID k+l ulfm— (k +l),pm
where C = C(k, p, rn, M A , SZ).
surprising that the coarea 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.
B ( x,r
Il rq( _l) ) 1!
sup : x E R", r> 0 < M
1/q
uQdµ CM1/61IIDuII 1 (4.9.1)
R"
whenever u E Co (R").
f r<dµf4Et)dt
_F ( 4 .9 . 2 )
<rnE
i=1
< C H 1 [{u 1 (t)}n
R"
implies
1/q'
J IttIgdµ < CM 1
14
Ilsllg';^•
for all g E Lq (u), g > O. However, by the Riesz Representation theorem,
^n
I^uld^
!
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
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
BE.* i=1
Er n.
00
< M5m C m n
i=1
< M5 m c mn lUl .
4.9.3. Lemma. Lei > 0 be a Radon measure on R", t < n, 1 < q <
(n1 +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(in)Qµ[B(x, r)j : x E R", r > 0}.
1/r
f B(0.2r) f
_ (n1} 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 N1)r
dx
+ 1)r
B(O r)
, 
B(y r) Ix 
,
YIN
The first integral can be estimated by
dx < Crnr(n1)
I (n1)r
fB(//,r) I x ` 11I
and the second integral is dominated by r nr(nI) Here we have used the
fact that (n  1)r < n. Thus,
dx < Crni(n1)
,
41
(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 (nl) Q n dt and write
This combined with (4.9.5) and (4.9.4) yield the desired result.
4.9.4. Theorem. Let p > 0 be a Radon measure on R" and let l < n,
q > 1. Then if
U.
lfg
04 5 GM 1/g if D 1 uI11
whenever u E Cô (R n ).
Since W ei lk/n1 < 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 gl fi * µ dy (by ^bini's theorem)
Rn
< CgllulIn/(1n_ O II IDulfl * µll (by Holder's inequality)
T
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
f812
IuI dH n < C n IDul dz.
Exercises
215
fa I
iukuelpdH" —+ 0 as k, l oo .
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,
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" ,
+ IDulpdg .
B(2r)
Iu(x)  u(g)Ipdy
B(2r)
in terms of the norm of Du.
act +
1.(191)(x)
Mg(x)
Historical Notes 217
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
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
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
Poincarétype 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,
,
total variation:
has these properties, let = Du and refer to (5.1.1) to see that p satisfies
u div cpdx = —f ço • dg
{
o Iv(x)1= a.
222 5. Functions of Bounded Variation
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
I
"
udivv dx i — EDiuvdx
n i`1
and the gradient of u has finite total variation with
II Du II(s^) = f Dudx.
Proof. Let v be a vector field such that v E Cp (U; R") and Iv(x)I < I for
T E U. Then
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
and therefore
U,. u DSp da < sup Iol lim inf Wu; IRO) < oo.
{^ 00
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.
Then,
lim sup IIDuII(U n n) < IIDuII(U n si)
224 5. Functions of Bounded Variation
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
The result follows by taking the supremum over all such al. a
fu, — fu in i'(U) and therefore that D(f u e ) —> D(fu) in 0'(U), the
conclusion readily follows. D
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
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
vc = E(ufi ) ^ :
i =o
Clearly, y r E C° (1) and u = EMI ufi . Therefore, from (5.3.3)
I Iv e — uldx <
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 *
= 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
I
00
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
,
Proof. Assume first that u E BV1Qc (R"). For 1 < i < n it will be shown
228 5. Functions of Bounded Variation
that
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,
f ess V (ui)dH"
b 1 (2) < lirninf
E
R
ess Vâ (u^.,,)dHn 1 (x)
= lirn inf
E6 o R
For the other half of the theorem, let u £ Lloc (R") and assume
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
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
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
where u(r) = fÈ(r) u(x)dx and B(r) is any ball in Rn. Now [et u = X E and
obtain
n1(n1)
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
(
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 coarea 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
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
and
lim
koo
I D L kIdx = 11Dull(n), (5.4.7)
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)
_ lim inf
<
k
Pk
I DLkldx (by (2.7.1))
= (by (5.4.7)). ❑
function defined by
a.e. x E R" and therefore, p(x) = v(x, E) for IIDXEIIa.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 GaussGreen theorem.
where B(r) = B(x,r) and v,(y, B(r)) is the i th component of the unit
exterior normal.
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
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) ^ dHn1(x)
fBfr+c)—B(r) f (x)XE(x) 2t dZ ^ 
Ii L8(B(r)) ^
which implies
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,
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)
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)
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
Note that (5.5.5) holds for all small values of r since the left side is a
leftcontinuous 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,
for some constant C = C(n). Let h(r) = 1E n B(r)I and observe that the
coarea formula (Theorem 2.7.3) yields
T
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
and
IIDXOI(U) , IIDXH II(U)
whenever U is a bounded open set with 11 111 [(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 ;
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 = Ct r n1 1
IIDXE. II[B(r)] =
e l
nIID^ E II[B(er)) < C l e ln( er )n
(5.6.3)
and
e tn e ln (er') n1 = Crn1 (5.6.4)
IIDXE, II[B(r)]  IIDXEIIIB(er)) ^ C
< 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

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
This implies that v. (x, A) = 1 for IIDXAlIa.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

lim
t400 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
r•0
and
lirnr'IB(x,r) n {y: (y —x)n> 0, y E E}I =
The treasuretheoretic 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. 
Proof. Again we use the "blowup" 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,
DEII
Ix Bxr( )1 > C
liminf
r—.o rn1
whenever x E a'E.
and
tinâ rNB(x,r)  El n H (x)1 = ü

where H+ (x) and H (z) are the halfspaces determined by the exterior

normal, n(x, E). Since B(x, r)nH (x) = ([B(x, r)EjnH (x))u(B(x, r)n
 
lirn
IB(r) n EI
_ rP0
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.
(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. 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
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
k1
f,(R n ' `)  AD C
;, k = 1
countably (n  1)rectifiable.
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 
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
However, since 2Iz — yl < r,, it follows from (5.7.1) and (5.7.2) that
=
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 IzyI < IN and p(y) = p(z). Let u = zy/IzyI
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
c7 Ec UM;UN

t=1
where H" ' (N) = 0 and each Mi is an (n  1)dimensional embedded C'
—
submanifoid of R".
that the boundary of a set will be taken as the reduced boundary and the
exterior normal as the measuretheoretic exterior normal.
In Definition 5.6.4, we introduced the notion of the measuretheoretic
exterior normal and demonstrated (Theorem 5.6.5) that
a E C E. (5.8.1)
This is a crucial result needed for the proof of the GaussGreen 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.
r ln
IIDXEII [ B ( x ► r)] = IIDXET 1)] 0 IIDXH lI[$(xt 1)] 
= H [B(x, 1) fl ir(x)]
n1
= a(n  1)
hm }^ 1
r^0 (Y(f9  1)T
Bxr ,
li. D = 1, x E B f1 M;.
II X EIE( ( B(,) r
)I
248 5. Functions of Bounded Variation
1 div V dx = 8• E
n(t, E) • V (x)dHnr 1 (x)
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 measuretheoretic 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.
Hn 1(K n3 M E) < o0

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:
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.
aplimsupu(y) = apliminfu(y)•
vx vx
Define
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
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
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,)n1 < 5(n 1)G*G^` IIDXEII{B(x, ri}]
L
i=^1 i=^1
< 5011)CIIDxEIIERn].
ÎIDXEII{B(x,r)1
l i rnsu p a ( rt  1) r n1
r—)
_ O.
Proof. For each positive number a let
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 measuretheoretic 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.
D(E, z) > b and D(Rn  E, z) > b where 0 < b < 1/2 and define a
continuous function f by
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  :
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
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.
Since u has compact support, we may assume that diam B(r,) < a, for
some positive number a. Therefore, Theorem 5.4.4 implies
1 (1,4  A)dHn i _ 0
00
H n `' [{x : A(s) < t < µ(x) }idt
lim
IA y (,) E fl H(z) n B(z, r)I = 1

similarly. 0
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.
nonnegative. 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
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.
and
s r + c(r)]IIuIIev(n) •
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
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
But t < 0 implies (R"  12)  A i l  O. This, along with the fact that
u = n(x o , SZ) yields
( 5. 10. 5)
Iint
r60 fB ( r)flO
iu(Y)  u` (i)i"/("1)dy = 0 (5.10.6)
8^ n
whenever u E BV(f2).
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

Thus, we have
at any point where the right side is defined. From Theorem 5.9.6(iî), we
know that this occurs at H"  talmost 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. PoincaréType 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 .
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
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
Hu — 1);n Ç CM llDull(Rn ).
(5.11.7)
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:
y (U E; — lirn
i =1 i.00
+ co
lin 1 (a m {h j > t })dt
1_m
f
+oo
+ H n 1(a%,{1j > t })dt.
^
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(Ej1) <_
f00
+
^ f
+m
)
H n 1 (ô,^{v > t })dt

IIDhj  1II(R " ) + IIDU,II(R " )
< IIDhj1II(R n ) +7(E,) + E2 ? .
It follows by induction that
71(A) < P({u > to}) < H" 1 [i9{u > toil < H n [ 11 I(t o )]
1 
< y(A) + q. fl
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
(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
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
E=1
i
,
7 71 1 < CP(Ft )
we have µ(A t ) = p,(FF ). Thus, Lemma 1.5.1, (5.12.6), and (5.12.4) imply
00
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 nonnegative
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).
that u(x) = 0 for H" 1 almost all x E A, there exists a constant C = C(f2)
such that
Ilulln/(n1);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. ❑
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,
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. ❑
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
If uf da < MClllfllpf)u'IIBv(R")
IJu — (i)Ilp.a 5 M
C
Cs 2
IlDull(o)
u(1)
for all u E BV (12).
llullp.a;n = lluollp,a
^
C ll nai l e v(Rn )
Also,
If u(x) = 0 on A where A is a Stalin set of positive 11'measure, then
(ii) —oo < A(x) < p(x) < oo for fin Ialmost all
 x E Rn ,
(iii) For Hn 'almost every z E E, there exists a unit vector y such that

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
+ 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 (
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.
all x o E R"  E, whereas (5.14.2) and (5.14.3) hold for H' almost all
x a E E.
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)rn1
,
_ (5.14.5)
If U is bounded, then
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
Exercises
li M IL k —ul=o,
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 coarea 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).
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)
Гораздо больше, чем просто документы.
Откройте для себя все, что может предложить Scribd, включая книги и аудиокниги от крупных издательств.
Отменить можно в любой момент.