3

Pointwise Behavior of
Sobolev Functions
In this chapter the pointwise behavior of Sobolev functions is investigated.
Since the definition of a function U E Wk,P(r!) requires that the kth-order
distributional derivatives of u belong to V(r!), 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 Wk,P(R n ), then u has derivatives of order k when computed in the
metric induced by the V-norm. That is, it will be shown for all points
x in the complement of some exceptional set, there is a polynomial Px of
degree k such that the V-norm of the integral average of the remainder
lu - Pxl over a ball B(x, r) is o(rk). Of course, if u were of class ok, then
the V-norm could be replaced by the sup norm.
We will also investigate to what extent the converse of this statement is
true. To this end, it will be shown that if u has derivatives of order k in the
LP-sense at all points in an open set r!, and if the derivatives are in V(r!),
then u E Wk,P(r!). 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 Of., e::; k. The approximants will have the
property that they are close to u in the Sobolev norm and that they agree
pointwise with u on large sets. That is, the sets on which they do not agree
will have small capacity, thus establishing a Lusin-type approximation for
Sobolev functions.

3.1 Limits of Integral Averages of

Sobolev Functions
In this and the next two sections, it will be shown that a Sobolev function
u E Wk,P(r!) can be defined everywhere, except for a set of capacity zero,
in terms of its integral averages. This result is analogous to the one that

W. P. Ziemer, Weakly Differentiable Functions

© Springer Science+Business Media New York 1989
3.1. Limits of Integral Averages of Sobolev Functions 113

holds for integrable functions, namely, if u E L1, then

1 lu(y) - u(x)ldy --+ 0 as r --+ 0

h(x,r)

for almost all x E Rn. Since our result deals with Sobolev functions, the
proof obviously will require knowledge of the behavior of the partial deriva-
tives of u. The development we present here is neither the most efficient nor
elegant. These qualities have been sacrificed in order give a presentation
that is essentially self-contained and clearly demonstrates the critical role
played by the gradient of u in order to establish the main result, Theorem
3.3.3. Later, in Section 3.10, we will return to the subject of Lebesgue points
and prove a result (Theorem 3.10.2) that extends Theorem 3.3.3. Its proof
will employ the representation of Sobolev functions as Bessel potentials
(Theorem 2.6.1) and the Hardy-Littlewood maximal theorem (Theorem
2.8.2).
In this first section, it will be shown that the limit of integral averages
of Sobolev functions exist at all points except possibly for a set of capacity
zero. We begin by proving a lemma that relates the integral average of u
over two concentric balls in terms of the integral of the gradient.

3.1.1. Lemma. Let u E W 1 ,P[B(xo,r)], p ~ 1, where Xo ERn and r > o.

Let 0 < fJ < r. Then

r- n [ u(y)dy-fJ-n [ u(y)dy = .!.r-n [ [Du(y)·(y-xo)]dy

iB(xo,r) iB(xo,6) n iB(xo,r)

- .!.fJ- n [ [Du(y) . (y - xo)]dy

n iB(xo,6)

- .!. [ Iy - xol-n[Du(y) . (y - xo)]dy. (3.1.1)

n i B(xo,r)-B(xo,6)

Proof. Define JL on R1 by

Define a vector field V by V(y) = JL(ly - xol)(y - xo). Since u is the

strong limit of smooth functions defined on B(xo, r) (Theorem 2.3.2), an
application of the Gauss-Green theorem implies

[ u(y)divV(y)dy=- [ Du(y)·(y-xo)JL(ly-xol)dy. (3.1.2)

iB(xo,r) iB(xo,r)

An easy calculation of div V establishes equation (3.1.1). o

114 3. Pointwise Behavior of Sobolev Functions

3.1.2. Lemma. Let f be a positive real number such that fp < n, p 2: 1,

and let u E W1,P(R n ). Then

(3.1.3)

for all x ERn.

Proof. (i) We suppose first of all that u vanishes outside a bounded set.
Let x E R n and for each positive integer j, define a Coo vector field Yj on
Rn by
1 ] (1/2)(£-n)
Yj(y)= [ y+IY_X I2 (y-x).

Since lui E W1,P(R n ) (by Corollary 2.1.8), lui is therefore the strong limit
of smooth functions with compact support. Therefore, by the Gauss-Green
theorem,

{ divYj(y)lu(y)ldy = - ( Yj(y). D(lul)(y)]dy.

JRn JRn
Moreover, since ID(lul)l = IDul a.e.,

{ divYj(y)lu(y)ldy ~ ( Yj(y)IDu(y)ldy. (3.1.4)

JRn JRn
By calculating the divergence on the left-hand side of (3.1.4) one obtains

Ln D+ Iy - xl f 2
1 2
/ )(t-n-2) [fly - xl 2 + y] lu(y)ldy
~
( [1
JRn y+ Iy - x 2 1
](1/2)(l-n)
Iy - xIIDu(y)ldy.

The inequality (3.1.3) now follows, in this case, when j ----> 00.
(ii) The general case. Let TJ be a Coo function on R, such that 0 ~ TJ ~ 1,
TJ(t) = 1 when t ~ 1 and TJ(t) = 0 when t 2: 2. Define
Uj(Y) = u(Y)TJ(j-llyl)
for y E Rn. By applying (i) to Uj and then letting j ----> 00, one can verify
(3.1.3) in the general case. 0

3.1.3. Lemma. Let £ be a positive real number and k a positive integer

such that (k+£-l)p < n. Then there exists a constant C = C(n,k,f) such
that
3.1. Limits of Integral Averages of Sobolev Functions 115

This follows from Lemma 3.1.2 by mathematical induction. 0

We are now in a position to prove the main theorem of this section

concerning the existence of integral averages of Sobolev functions.

3.1.4. Theorem. Let k be a positive integer such that kp < n, p > 1, let n
be a non-empty open subset of R n and let u E Wk,p(n). Then there exists
a subset E of n, such that

and
lim 1 u(y)dy (3.1.5)
6->0+ h(x,6)

exists for all x E n- E.

Proof. (i) We suppose first of all that n= Rn. Define

g(y) = L IDOu(y)1 (3.1.6)

lol=k

for y E Rn. Then 9 E LP (Rn). Let E be the set of all those points x of R n
for which
(Ik * g)(x) = 00. (3.1.7)
Then, from the definition of Riesz capacity (Definition 2.6.2),

and therefore from (2.6.7),

Bk,p(E) = O.

Consider x E R n '" E. By (3.1.1),

f u(y)dy - 8- n f u(y)dy =.!:. f [Du(y) . (y - x)]dy

1 B(x,l) 1 B(x,6) n 1 B(x,l)

- .!:.8- n f [Du(y) . (y - x)]dy

n lB(x,6)

- .!:. f Iy - xrn[Du(y) . (y - x)]dy. (3.1.8)

n 16<ly-xl<1
When k = 1, it follows from (3.1.6) and (3.1.7) that

f Iy - xI1-nIDu(y)ldy < 00. (3.1.9)

lB(x,l)
116 3. Pointwise Behavior of Sobolev Functions

When k > 1, it follows from Lemma 3.1.3 with l = 1 and k -1 substituted

for k, that

which, by (3.1.6) and (3.1. 7), is finite. Thus (3.1.9) still holds when k > 1.
By (3.1.9)

lim [ Iy - xl-n[Du(y) . (y - x)]dy (3.1.10)

6-+0+ J6<ly-xl<1

exists. It also follows from (3.1.9) that

lim [ Iy - xI1-nIDu(y)ldy = 0,
6-+0+ J B(x,6)

hence
6- n [ [Du(y) . (y - x)]dy --+ 0 (3.1.11)
JB(x,6)
as 6 --+ 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 n be an open set of Rn. There exists an in-
creasing sequence {'Pj} of non-negative Coo functions on Rn, with compact
supports, and spt 'Pj C n for all j such that the interiors of the sets
{x: x E R n and 'Pj(x) = I}
tend to n as j --+ 00. Define

Uj(x) = {o'Pj(X) . u(x) xE n

x¢n.
By applying (i) to each of the functions Uj, one can easily prove the theorem
in this case. 0

3.2 Densities of Measures

Here some basic results concerning the densities of arbitrary measures are
established that will be used later in the development of Lebesgue points
for Sobolev functions.

3.2.1. Lemma. Let JL ~ 0 be a Radon measure on Rn. Let 0 < A< 00

and 0 < 0: ::; n. Suppose for an arbitrary Borel set A C R n that
.
I1m JL[B(x, r)] \
sup > 1\
r-+O rOt
3.2. Densities of Measures 117

for each x E A. Then there is a constant C = C(o, n) such that

Jl(A) ~ C>.HO«A).

Proof. Assume Jl(A) < 00 and choose e > o. Let U :) A be an open set
with Jl(U) < 00. Let 9 be the family of all closed balls B(x, r) C U such
that
x E A, 0 < r < e/2, Jl[B(x, r)] > >..
rO<
Clearly, 9 covers A finely and thus, by Corollary 1.3.3, there is a disjoint
subfamily F C 9 such that
A C [U{B: B E F*}] U [U{B: B E F - F*}]
whenever F* is a finite subfamily of F. Thus, by Definition 1.4.1,

H~(A) ~ C L
BE:F"
(8(:)) 0< + C50< L
BE:F -:F"
C(:)) 0<

where 8(B) denotes the diameter of the ball B. Since F C 9 and F is

disjoint, we have

L (8(:)) 0< ~ C>.-l L Jl(B)

BE:F BE:F
~ C>.-l Jl (U) < 00.
Since
C50< L
BE:F-:F"
C(:)) 0<

can be made arbitrarily small with an appropriate choice of F* , we conclude

H~e(A) ~ C>.-l Jl (U).

Since Jl is a Radon measure, we have that Jl(A) = inf{Jl(U) : U :) A, U

open}. Thus, letting e -+ 0, we obtain the desired result. 0

3.2.2. Lemma. Let Jl ~ 0 be a Radon measure on Rn that is absolutely

continuous with respect to Lebesgue measure. Let

A = R n n {x: lim sup Jl[B(~,r)] > O}.

r-+O r
Then, HO«A) = 0 whenever 0 ~ 0 < n.
Proof. The result is obvious for 0 = 0, so choose 0 < 0 < n. For each
positive integer i let

Ai = R n n {x: Ixl < i,limsup Jl[B(~,r)] > C 1 }

r-+O r
118 3. Pointwise Behavior of Sobolev Functions

and conclude from the preceding lemma that

(3.2.1)

Since Ai is bounded, JL(Ai) < 00. Therefore HO(Ai) < 00 from (3.2.1).
Since a < n, Hn(Ai) = 0 and therefore IAil = 0 from Theorem 1.4.2. The
absolute continuity of JL implies JL(Ai) = 0 and consequently HO(Ai) = 0
from (3.2.1). But A = U~lAi' and the result follows. 0

3.2.3. Corollary. Suppose u E V(R n ), 1 5:. p < 00, and let 0 5:. a < n. If
E is defined by

E = {x: limsupr-o {
r--+O } B(x,r)
lu(x)IPdx > o} ,
then HO(E) = O.
Proof. This follows directly from Lemma 3.2.2 by defining a measure JL as

3.3 Lebesgue Points for Sobolev Functions

We will now prove the principal result of the first three sections (Theorem
3.3.3) which is concerned with the existence of Lebesgue points for Sobolev
functions. We will show that if u E wk,p(Rn), then

lim 1 lu(y) - u(x)IPdx = 0

r--+O h(x,r)

for Bk,p-q.e. x E Rn. 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 W1,P(R n ), 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 lu - pi E Wl~'~(Rn) for each real number p and then
apply Theorem 3.1.4 to conclude that

lim 1 lu(y) - pldy

r-+oh(x,r)

exists for B1,p-q.e. x E Rn. 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 Wk,P(Rn ) since it is not true in general that lui E Wk,P(R n ),
3.3. Lebesgue Points for Sobolev Functions 119

cf. Remark 2.1.10.

3.3.1. Lemma. Let k be a non-negative integer and A,p real numbers such
that p > 1, kp < n, and k < A < nip. If

(3.3.1 )

then
8>.-k 1 u(y)dy --+ 0 (3.3.2)
h(x,15)

as 8 --+ 0+, for all x E R n except for a set E with B>.,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 Bk,p-null set. This lemma offers a slight
variation in that the integral averages when multiplied by the factor 8>'-k
converge to 0 on a larger set, the complement of a B>.,p-null set. At some
points of this larger set, the integral averages may converge to infinity, but
at a rate no faster than 8k ->'.

Proof of Lemma 3.3.1. (i) Suppose k = O. It follows from Corollary 3.2.3

that
8p>.-n [ lu(y)IPdy --+ 0 (3.3.3)
JB(x,15)

as 8 --+ 0+, for all x E R n except for a set E with Hn->'P(E) = O. From
the definition of Hausdorff measure, for e > 0 there is a countable number
of sets {Ei} such that E C UE i and ~(diamEi)n->.p < e. Each Ei is
contained in a ball Bi ofradius ri where ri = diamEi . Therefore, with the
aid of Theorem 2.6.13,
00 00

B>.,p(E) ::; 2: B>.,p(Ei ) ::; C 2: r~->'P ::; Ce.

i=l i=l

Since e is arbitrary, we have that B>.,p(E) = O.

Now consider x E Rn - E. From Holder's inequality, there is a constant
C = C(n,p) such that

8>'-n I[ U(Y)dyl ::; C [8>.p-n [ lu(y)IPdY] lip (3.3.4)

} B(x,15) } B(x,15)

(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

1 Iy -
R
n xl>.-n [2:
lal=k
IDaU(Y)I] dy = 00. (3.3.5)
120 3. Pointwise Behavior of Sobolev Functions

Then R>.,p(E) = 0 and therefore B>.,p(E) = 0 by (2.6.8).

Consider x E Rn - E. When k = 1, it follows from (3.3.5) that

r
iB(x,l)
Iy - xl>.-k+1-nIDu(y)ldy < 00. (3.3.6)

When k > 1, we replace f by). - k + 1 and k by k - 1 in Lemma 3.1.3 and

again derive (3.3.6) from (3.3.5). For x E R n - E, we now show that

O>.-k r Iy - xll-nIDu(y)ldy -+ 0, (3.3.7)

i8<ly-xl<1
as 0 ! O. Let r E (0,1) be arbitrary. Clearly

O>.-k 1r<ly-xl<1
Iy - xll-nIDu(y)ldy -+ 0 (3.3.8)

as 0 ! O. When 0 < 0 < r we have

O>.-k r Iy - xll-nIDu(y)ldy::; r Iy - xl>.-k+1-nIDu(y)ldy.

i8<ly-xl<r i B(x,r)
(3.3.9)
It follows from (3.3.6) that the right-hand side of (3.3.9) approaches zero
as r ! O. (3.3.7) now follows from (3.3.8) and (3.3.9).
Clearly,
r
iB(x,l)
Iy - xIlDu(y)ldy < 00. (3.3.10)

Since). - k - n < 0, it follows that

o>.-k-n r
i B(x,8)
Iy - xIIDu(y)ldy::; r
i B(x,8)
Iy - xl>.-k+1- n IDu(y)ldy,

so that by (3.3.6),

o>.-k-n r
i B(x,8)
Iy - xIIDu(y)ldy -+ 0 (3.3.11)

as 0 ! O. By putting r = 1 in Lemma 3.1.1, one can obtain (3.3.2) from

Lemma 3.1.1, (3.3.10), (3.3.11), and (3.3.7). 0

3.3.2. Theorem. Let f, k be integers such that k :::: 1, 0 ::; k ::; f and
fp < n, p > 1. Let u E wk,p(Rn) and for each x E Rn and r > 0 put

ux,r = 1 u(y)dy.
Ts(x,r)

Then
r(i-k)p 1 lu(y) - ux,rlPdy -+ 0 (3.3.12)
Ts(x,r)
3.3. Lebesgue Points for Sobolev Functions 121

as r ! 0, for all x E Rn except for a set E with Bl,p(E) = O.

Proof. We proceed by induction on k. Suppose to begin with that k = o.

It follows from Corollary 3.2.3 that

(3.3.13)

for all x E R n , except for a set E' with Hn-lp(E') = 0 and therefore

Bl,p(E') = O. (3.3.14)

We now have, for x E Rn - E',

r(l-(n/p» [ r
JB(x,r)
Iu(y) - ux,rIPdy]l/P :::; r(l-(n/p)) [ r
JB(x,r)
IU(Y)lPdy]l/P

r
l/P
+ r(l-(n/p)) IUx,r I [ dy ] (3.3.15)
JB(x,r)
But by Lemma 3.3.1,
(3.3.16)
as r ! 0, for all x E R n except for a set E" with Bl,p(E") = O. (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,p(R n ). Let U E Wk,P(R n ). By the Poincare inequality,
which we shall prove in a more general setting later in Chapter 4 (for
example, see Theorem 4.4.2),

r(l-k)p-n r
JB(x,r)
Iu(y) - ux,rlPdy :::; Cr[l-(k-l)]p-n r
JB(x,r)
IDu(y)IPdy,
(3.3.17)
for all x E R n , where C depends only on n. By the induction assumption,
there exists a set F', with
Bl,p(F') = 0 (3.3.18)
and
r[l-(k-l)]p-n r
JB(x,r)
IDiU(Y) - (Diu)x,rIPdy -+ 0 (3.3.19)

as r ! 0, for all x E Rn - F'. But

122 3. Pointwise Behavior of Sobolev Functions

+ I(DiU)x,rl [ l f
JB(x,r)
dy
l/P
, (3.3.20)

and by Lemma 3.3.1

rl-(k-l)(Dou)
Z X,T
-+ 0 (3.3.21)
as r ! 0, for all x E Rn, except for a set F" with

Bl,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. D

3.3.3. Theorem. Let k be a positive integer such that kp < n, let n be an

open set of R n and let u E Wk,p(n). Then

1 lu(y) - u(x)IPdy -+ 0 (3.3.22)

h(x,r)
as r ! 0, for all x E n, except for a set E with Bk,p(E) = o.
Proof. (i) When n = Rn, (3.3.22) follows from Theorem 3.3.2 and Theorem
3.1.4.
(ii) When n is arbitrary, the theorem can be derived from (i) as in the
proof of Theorem 3.1.4. D

3.3.4. Corollary. Let k be a positive integer such that kp < n, let n be an

open set of Rn and let u E Wk,p(n). Then

lim 1 lu(y)IPdy exists and = lu(x)IP (3.3.23)

r---+O+ h(x,r)
for all x E n, except for a set E with Bk,p(E) = o.
3.3.5. Remark. Theorem 3.3.3 states that on the average, the oscillation
of u at x is approximately equal to u(x) at Bk,p-q.e. x E n. 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 Xo if there exists a measurable set A such
that
lim IB(xo, r) n AI = 1 (3.3.24)
T---+O IB(xo, r)1
and u is continuous at Xo relative to A. It is not difficult to show that if
u has a Lebesgue point at Xo then u is approximately continuous at Xo.
A proof of this is given in Remark 4.4.5. Thus, in particular, Theorem
3.3.3 implies that u E W1,P(R n ) is approximately continuous at B1,p-q.e.
x ERn.
3.3. Lebesgue Points for Sobolev Functions 123

Approximate continuity is a concept from measure theory. A similar con-

cept taken from potential theory is fine continuity and is defined in terms
of thin sets. A set A c R n is said to be thin at Xo relative to the capacity
Bk,p if
Ii [Bk'P[A n B(xo, r)l] i/(p-i} dr < 00.
(3.3.25)
10 Bk,p[B(xo, r)l r
A function u is finely continuous at Xo if there exists a set A that is thin
at Xo and
lim u(x) = u(xo).
Z--+Xo
x~A

It follows from standard arguments in potential theory that A can be taken

as a measurable set. In the case of the capacity, B i ,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

for any set A c Rn , it follows that (3.3.25) implies

lim IB(xo, r) n (Rn - A)I = 1,

r--+O IB(xo,r)1
and therefore fine continuity implies approximate continuity.
We now will show that the approximate continuity property of Sobolev
functions can be replaced by fine continuity. First, we need the following
lemma.

Lemma. If {Ai} is a sequence of sets each of which is thin at Xo, then

there exists a sequence of real numbers {ri} such that

UAi n B(xo, ri)

00

i=i

is thin at xo.

Proof. Because Ai is thin at Xo, it follows that there exists a sequence

{ri} -+ 0 such that

Bk ,p[Ai n B(xo, rdl -+

0
as .
Z -+ 00.
Bk,p[B(xo, rill
We may assume the ri to have been chosen so that

(i [Bk'P[Ai n B(xo, r)l] i/(p-i) dr < T(i+i}.

10 Bk,p[B(xo, r)l r
124 3. Pointwise Behavior of Sobolev Functions

Then,

r [Bk,P[A i n B(xo, r) n B(xo, ri )]] l/(p-l) dr

1

10 Bk,p[B(xo, r)] r
(i [Bk,p[A i n B(xo, r) n B(xo, r i )]] l/(p-l) dr
10 Bk,p[B(xo, r)] r
+ r1 [Bk,p[A i n B(xo, r) n B(xo, r i )]] l/(p-l) dr

11 [
1ri Bk,p[B(xo, r)] r
T(i+ 1) + Bk [A- n B(xo r.)]l/(P-l) 1 ] l/(p-l) dr
< ,p" • ri C r n-kp r
< 2 -(i+l) + B k,p [A i n B( Xo, ri )]l/(P-l)C1 [1 -
1
Bk,p[B(xo, ri)J1/(p-l)
]
< 2- i for ri sufficiently small.
Since capacity is countably subadditive, the result easily follows. 0

For ease of exposition, we now restrict our attention to u E W 1 ,p(Rn).

Again, we see the important role played by the growth of the gradient in
order to obtain some regularity at a given point.

Theorem. Let Xo ERn, p > 1, and suppose u E w 1 ,p(Rn) has the property
that
r 1 [
r p- n r IDulPdx
jl/(P-l)
dr < 00.
10 1B(xo,r) r
Suppose also that
lim 1 u(y)dy = U(Xo).
r->O Tn(xo,r)

Then u is finely continuous at Xo.

Proof. For each c > 0, let

A(xo, c) = R n n {x: lu(x) - u(xo)1 > c}.

For r > 0, let

Vr(x) = CPr(x)[u(x) - u(2r)]
where CPr is a smooth function such that CPr == 1 on B(xo, r), spt CPr C
B(xo,2r), IDCPrl :::; Cr- 1 and where

u(2r) = 1 udx.
Tn(xo,2r)
3.3. Lebesgue Points for Sobolev Functions 125

Because of the assumption u(2r) ---+ u(xo) as r --+ 0, note that for all
sufficiently small r,

Vr(x) ~ c:/2 for x E A(xo, c:) n B(xo, r).

Therefore, by appealing to Exercise 2.8, which allows BI,p to be expressed
in terms of a variational integral, there exists C = C(p, n) such that

BI,p[A(xo, c:) n B(xo, r)) :::; C(2c:- I )P [ IDvrlPdx

lB(xo,2r)

:::; C(2c:- I )p [ IDulPdx

lB(xo,2r)

+ (C2c:- I r- l )p [ lu - u(2r)IPdx. (3.3.26)

lB(xo,2r)

An application of Poincare's inequality (cf. Theorem 4.4.2) yields

[ lu - u(2r)IPdx :::; Cr P [ IDulPdx

lB(xo,2r) lB(xo,2r)

and therefore (3.3.26) can be written as

BI,p[A(xo, ~~~ B(xo, r)) :::; Cc:-IrP-n [ IDulPdx,

r lB(xo,2r)

which directly implies that A(xo,C:) is thin at Xo. Now let C:j be a sequence
tending to O. By the preceding lemma, there is a decreasing sequence rj --+ 0
such that 00

A= U[A(xo,C:j)nB(xo,rj))
j=l

is thin at Xo. Clearly

lim
3:-+Xo
u(x) = u(xo)
xERn-A

and the theorem is established. 0

It can be shown that

I [ jl/(P-I)
[ rp-n [ IDulPdx dr < 00 (3.3.27)
1o lB(xo,r) r

for BI,p-q.e. Xo E R n cf. [ME3). Therefore, with Theorem 3.3.3, we obtain

the following.

Corollary. If u E WI,P(R n ) then u finely continuous at all points except

for a set of BI,p capacity zero.
126 3. Pointwise Behavior of Sobolev Functions

Observe that Corollary 3.2.3 implies

limsuprp - n [ IDulPdx = 0 (3.3.28)

r-+O JB(xo,r)
for Hn-p-a.e. Xo E Rn. Although (3.3.27) implies (3.3.28) for each Xo, the
exceptional set for the former is larger than that for the latter.

3.4 V-Derivatives for Sobolev Functions

In the previous three sections, the continuity properties of Sobolev func-
tions were explored through an investigation of Lebesgue points and fine
continuity. We now proceed to analyze their differentiability properties. We
begin by proving that Sobolev functions can be expanded in a finite Tay-
lor series such that for all points in the complement of an exceptional set,
the integral average of the remainder term tends to 0, (Theorem 3.4.2). In
keeping with the spirit of this subject, it will be seen that the exceptional
set has zero capacity. Observe that Theorem 3.3.3 provides the first step
in this direction if we interpret the associated polynomial as one of degree
o and the remainder at x as lu(y) - u(x)l.
When k, m are integers such that 0 ~ m ~ k, (k - m)p < nand
u E wk,p(Rn), it follows from Theorem 3.1.4 that there exists a subset E
of Rn such that
Bk-m,p(E) = 0 (3.4.1)
and
lim 1 DOu(y)dy (3.4.2)
r-+O+ h(x,r)

exists for all x E R n - E and for each multi-index a with 0 ~ lal ~ m.

Thus, for all such x, we are able to define the Taylor polynomial p~m) in
the usual way:

(3.4.3)

(Recall the notation introduced in Section 1.1.) Observe that when u is a

em function on Rn , Taylor's theorem can be expressed in the form
u(y) = p~m-1)(y) + m L ~! [1 (1- t)m-1
1

lol=m 0
. D Ou[(l - t)x + tyjdtj (y - x)O. (3.4.4)

3.4.1. Theorem. Let 1 ~ m ~ k and suppose (k - m)p < n. Let u E

3.4. V-Derivatives for Sobolev Functions 127

Wk,P(R n ) and E be the set described in (3.4.1) and (3.4.2). Then

[( lu(y) - p~m)(y)IPdy j 1/P

::; rm L ~
1
{ (1 - t)m-1
i B(x,r) lo:l=m a. io
[
. en (
i B(x,tr)
IDO:u(y) - DO:u(x)IPdy j 1/P
dt (3.4.5)

and

[1B(x,r)
lu(y) - p~m-l}(y)IPdy j 1/P
::; rm L ~
lo:l=m a.
1 0
1
(1 - t)m-1

(3.4.6)

for all x E Rn except for a set E' :J E with Bk-m,p(E') = o.

Proof. (i) Suppose first of all that u is a em function on Rn. Let x E Rn,
r > 0 and put B = B(x, r). Let cp be a function of V' (B) with Ilcpllpl ::; 1
where p' is the conjugate of p. By (3.4.3) and (3.4.4),

iB( [u(y) - p~m)(y)lcp(y)dy = L ~ (1- t)m-1

a.
lo:l=m

. [In {DO:u((1- t)x + ty) - DO:u(x)}(y - X)o:cp(Y)dY] dt.

Hence, by Holder's inequality,

1
1
B
[u(y) - p~m)(Y)lcp(Y)dYI ::; rm L ~
lo:l=m a.
11
0
(1 - t)m-1

. [lnIDO:U((1- t)x + ty) _ Do:u(x)IPdY] lip dt.

By making the substitution z = x + t(y - x) in the right-hand side and
then taking the supremum over all cp, one obtains (3.4.5).
The inequality (3.4.6) can be derived similarly.
(ii) Now let u be an arbitrary function of Wk,P(R n ). By Theorem 3.3.3,
there exists a set E' :J E, with Bk-m,p(E') = 0 such that

(3.4.7)
128 3. Pointwise Behavior of Sobolev Functions

when 0 ~ lal ~ m and x E Rn - E'.

Consider x E Rn - E'. There exists a constant M (depending on x), such
that
(3.4.8)

for allial = m and all C > O. Let {!Pel be a sequence of regularizers as

discussed in Section 1.6. Thus, !Pe E C8"(Rn ),

(3.4.9)

spt !Pe C B(O,£) and

sup !Pe(x) ~ c£-n (3.4.10)
xERn
for all £ (where C depends only on n), while
(3.4.11)

as £ ! 0, for 0 ~ lal ~ m and x E Rn - E'. Put Ue = !Pe * u. Each Ue E

coo(Rn)nWk,p(Rn ). Let us denote by (3.4.5)e and (3.4.6)e the inequalities
(3.4.5) and (3.4.6) with U replaced by U e . Since U e is smooth we know
that (3.4.5)e and (3.4.6)e are valid. By (3.4.11) and Fatou's lemma, the
lower-limit as £ ! 0 of the left-hand side of (3.4.5)e and (3.4.6)e is greater
than or equal to the left-hand side of (3.4.5) and (3.4.6). The result of the
theorem will thus follow from Theorem 1.6.1(ii) and Lebesgue's Dominated
Convergence theorem when we show for each a with lal = m and r > 0
fixed, that the following function of t, 0 ~ t ~ 1, is bounded; that is,

rn [ IDaue(y)IPdy ~ Mr n (3.4.12)
iB(x,tr)

where M is independent of £.
We now proceed to establish (3.4.12). For any measurable subset E of
Rn, we have (when lal = m)

P
lIDaue(y)IPdY = llLn !Pe(Y - Z)Dau(z)dzI dy,

hence by (3.4.10)

[ IDaue(y)IPdy ~ CP£-(np) [ [[ IDaU(Z)ldZ]P dy. (3.4.13)

iE iE iB(y,e)

Thus, when p > 1, we have by Holder's inequality

3.4. V-Derivatives for Sobolev Functions 129

so that

[ IDClue(y)IPdy ~ Cc:- n [[ IDClu(z)IPdzdy, (3.4.14)

JE JEJB(y,e)

where C depends only on nand p. When p = 1, (3.4.14) follows from

(3.4.13).
When tr ~ 3c:, we let E be the ball with center x and radius tr. Since
B(y, c:) C B(x, 4c:) when y E B(x, tr), (3.4.14) implies that

[ IDClue(y)IPdy ~ Cc:- n [ [ IDClu(z)IPdzdy.

J B(x,tr) J B(x,tr) J B(x,4e)

It now follows from (3.4.8) and (3.4.12) holds in the case where tr ~ 3c:.
When tr > 3c:, we have

and a double application of (3.4.14) yields

[ IDClue(y)IPdy ~ Cc:- n [ [ IDClu(z)IPdzdy

J B(x,tr) J B(x,3e) J B(x,4e)

+ Cc:- n [ [IDClu(z)IPdzdy
J3e'5:.1y-xl<tr J B(y,e)

and by (3.4.8)

~ c'tnrn + Cc:- n [ [IDClu(w + y)IPdwdy

J3e~ly-xl<tr J B(O,e)
~ c'tnrn + Cc:- n [ [[ IDClU(y)IPdY] dw
J B(O,e) J2e~ly-xl<tr+e
so that
[ IDClue(y)IPdy ~ c"tnrn.
JB(x,tr)

Thus (3.4.12) is established. o

3.4.2. Theorem. Let 0 ~ m ~ k and suppose (k - m)p < n. Let u E
Wk,P(R n ). Then,

as r ! 0, for all x E Rn, except for a set F with

Bk-m,p(F) = O.
130 3. Pointwise Behavior of Sobolev Functions

This is the main result of this section. In particular, it states that the
integral average over a ball of radius r of the remainder term involving the
formal Taylor polynomial of degree k tends to 0 as r --+ 0 at a speed greater
than rk at almost every point. If a Taylor polynomial of smaller degree is
considered, the integral average tends to 0 at perhaps a slower speed, but
on a larger set.

Proof of Theorem 3.4.2. When m = 0, the theorem reduces to Theorem

3.3.3. Suppose m > O. By Theorem 3.3.3,

1 IDau(y) - Dau(x)IPdy --+ 0 (3.4.15)

h(x,r)

as r ! 0, for all lal = m for all x E R n , except for a set F with

Bk-m,p(F) = O.
Consider x E R n - F and an a with lal = m. Define

(3.4.16)

for r > O. By (3.4.15), 7J(r) --+ 0 as r ! 0, hence

11 (1- t)m-17J(tr)dt --+ 0 (3.4.17)

r ! O. The required result now follows from (3.4.16), (3.4.17), and Theorem
3.4.1. 0

3.5 Properties of V-Derivatives

In this section we consider arbitrary functions that possess formal Taylor se-
ries expansions and investigate their relationship with those functions that
have Taylor series expansions in the metric of £P, such as those discussed
in the previous section.

3.5.1. Definition. Let E eRn. A bounded function u defined on E

belongs to Tk(E), k > 0, if there is a positive number M and for each
x E E there is a polynomial Px (.) of degree less than k of the form

( ) _ ""' ua(x) (
Pxy-~--,-y-x )a , (uo =u) (3.5.1)
a.
lal~O

whose coefficients Ua satisfy

lua(x)1 :5 M for x E E, 0:5 lal < k,
3.5. Properties of V-Derivatives 131

and
Ua(y) = D a Px(Y) + Ra(x, y)
whenever x, y E E and where Ra(x, y) :S Ciy - xl k - 1al , 0 :S 10:1 < k.
The class t k (E) is defined as all functions U on E such that for each
x E E there is a polynomial Px (.) of degree less than or equal to k of the
form (3.5.1) such that for 0 :S 10:1 :S k,

whenever x, y E E with IRa(x, y)1 :S ely - xl k - 1al and

lim Ra(x, y) = 0
y-+x Iy - xl k - 1al
uniformly on E.
As a mnemonic, Tk(E) and tk(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 0" or "little 0," respectively.

3.5.2. Remark. Clearly, if U E Tk(E) then Ua is locally Lipschitz on E,

o :S 10:1 < k. If E is an open set, note that the derivatives Dau exist on E,
o :S 10:1 < k, and that
Dau(x) = D a Px(x) = ua(x) for x E E.

Since IDa Px(x)1 :S M for x E E, it follows that U E Wl~~l'P(E) for every

p 2: 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 ek(E), then U would have an expansion as in Definition 3.5.1
with
L 1
,Dau(x)(y - x)a.
0:.
O~lal9

e
Moreover, if u E k (Rn) and E c Rn, then the restriction of u to E,
ulE, belongs to tk(F) for each compact set FeE. One of the reasons for
identifying the class t k (E) is that it applies dire~tly to the Whitney exten-
sion theorem [WH], which we state here without proof. We will provide a
different version in Section 3.6.

3.5.3. Whitney Extension Theorem. Let E c R n be compact. If u E

e
t k (E), k > 0 an integer, then there exists u E k (Rn) such that for 0 :S
1,81 :S k
D{3u(x) = D{3 Px(x) for all x E E.
In view of this result, it follows that u E tk(E) if and only if u is the
restriction to E of a function of class ek(Rn).
132 3. Pointwise Behavior of Sobolev Functions

We now introduce another class of functions similar to those introduced

in Definition 3.5.1 but different in the respect that the remainder term
is required to have suitable decay relative to the V-norm instead of the
Loo-norm. The motivation for this definition is provided by the results
established in Section 4 concerning Taylor expansions for Sobolev functions.

3.5.4. Definition. For 1 :5 p :5 00, k a non-negative integer, and x E

Rn, Tk,P(x) will denote those functions u E V for which there exists a
polynomial PxO of degree less than k and a constant M = M(x, u) such
that for 0 < r < 00
)
~P
( 1 lu(y) - Px(y)IPdy :5 Mrk. (3.5.2)
h(x,r)
When p = 00, the left side of (3.5.2) is interpreted to mean esssupyEB(x,r)
lu(y) - Px(y)l. Tk,p(x) is a Banach space if for each u E Tk,P(x) the norm
of u, lIuIlTk,P(x), is defined as the sum of lIull p , the absolute value of the
coefficients of Px , and the smallest value of M in (3.5.2).

3.5.5. Definition. A function u E Tk,P(x) belongs to tk,P(x) if there is a

polynomial of degree less than or equal to k such that

1 lu(y) -
( h(x,r) Px(Y)IPdY) lip = o(rk) as r -+ O. (3.5.3)

Note that if u E Tk,P(x) the polynomial Px is uniquely determined. To

see this write

where
1 IRx(y)IPdY) lip :5 Mrk.
( h(x,r)
If Px were not uniquely determined, we would have u(y) = Qx(Y) + Rx(Y),
where Rx satisfies an integral inequality similar to that of Rx.
Let Sx(Y) = Px(Y) - Qx(Y). In order to show that Sx == 0, first note that

1
h(x,r)
ISx(y)ldY:5 (1 h(x,r)
ISx(Y)IPdY) lip :5 Crk, 0 < r < 00.

Now let Lx be the sum of terms of Sx of lowest order and let Mx = Sx - Lx.
Thus, Lx has the property that for each A E Rl, Lx (Ay + x) = Aa Lx (y + x),
where a is an integer, 0 < a :5 k - 1. Since Mx is a polynomial of degree
at most k - 1, we have

1 IMx(y)ldy :5 Cr k- l , 0 < r < 00.

h(x,r)
3.5. Properties of V-Derivatives 133

It follows from the inequality ILx(Y)1 :::; ISx(Y)1 + IMx(Y)1 that

ra 1 ILx(y)ldy =
h(X.l)

=1 ILx(y)ldy
h(x.r)
:::; Crk + Cr k - 1 , 0 < r < 00.
This is impossible for all small r > 0 if a < k - 1 and Lx is non-zero. If
a = k - 1, then Mx == 0 and the term Cr k - 1 above can be replaced by O.
A similar argument holds in case u E tk·p(x).
Obviously, tk(E) C tk·p(x) and Tk(E) C Tk·p(x) whenever x E E and
p :::: 1. We now consider the question of the reverse inclusion. For this
purpose, we first need the following lemma.

3.5.6. Lemma. Let k be a non-negative integer. Then there exists cp E

C~(Rn) with sptcp C {Ixl :::; I} such that for every polynomial P on Rn
of degree :::; k and every e > 0,

CPe * 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 m-tuples {yo:} whose compo-
nents are indexed by multi-indices a = (all a2,"" an) with 0 :::; lal :::; k.
The number m is determined by k and n. Define a linear map T: V -+ W
by

thus,

where 0 :::; lal :::; k and xO: = Xfl X~2 ••• x~n.
Note that vector space, range T, has the property that range T = W for
if not, there would exist a vector, a = {ao:} orthogonal to range T. That
is,
~ ao:yo: = 0 whenever y = {yo:} E range T.

This implies,
134 3. Pointwise Behavior of Sobolev Functions

Select 1/ E V such that 1/ > 0 in {x : Ixl < I}. Now define 'Ij; by 'Ij; = ~ aa xa 1/
and note that 'Ij; E V. Therefore,

which implies ~ aaxa = 0 whenever Ixl < 1. But this implies that all m
numbers aa = 0, a contradiction. Thus, range T = W. In particular, this
implies there is cp E V such that

r
JRn
cp(x)dx = 1,

Since any polynomial Q of degree no greater than k is of the form

Q(z) = L baz a ,
O~lal~k

it follows that
r
JRn
cp(z)Q(z)dz = Q(O).

Given a polynomial P = P(x) as in the statement of the lemma, let z =

(x - y)/c and set Q(z) = P(x - cz) to obtain the desired result. D

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 V-
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 Loo-conclusion from a V-hypothesis. A critical role is played by the
existence of a smoothing kernel cp that leaves all polynomials of a given
degree invariant under the action of convolution.

3.5.7. Theorem. Let E C Rn be closed and suppose u E Tk,P(x), 1 ::; P ::;

00, k > 0, with IluIITk,p(x) ::; M for all x E E. Then u E Tk(E). Also, if E
is compact and if u E tk,P(x) for all x E E with (3.5.3) holding uniformly
on E, then u E tk(E).

In view of Whitney's Extension theorem (Theorem 3.5.3), note that a

function satisfying the second part of the theorem is necessarily the restric-
tion of a function of class ck(Rn). In the next section, we will investigate
Whitney's theorem in the context of V.

Proof of Theorem 3.5.7. Let cp E C(f(Rn) be the function obtained in

Lemma 3.5.6 such that
CPc * P(x) = P(x) (3.5.4)
3.5. Properties of V-Derivatives 135

whenever P is a polynomial of degree less than k, e > 0, and x E Rn. Note

that (3.5.4) implies

(3.5.5)

Since U E Tk,P(xo) for all Xo E E, we have for Xo and x E E,

U(y) = PXo (y) + R(xo, y) (3.5.6)

and
u(y) = Px(Y) + R(x, y) (3.5.7)

(i
where

B{x',r)
IR(x*,y)IPdy )'~ :::; Mrk, (3.5.8)

with x* either Xo or x. Now let e = Ix - xol and for 0 :::; 1{31 < k consider
1= Df3cpg * u(x).
For each fixed Z ERn, define Rz as Rz(x) = R(z, x) whenever x ERn.
From (3.5.6) and (3.5.5) it follows that

1= Df3cpg * Pxo(x) + Df3cpg * Rxo(x)

= Df3 PXo (x) + Df3cpg * Rxo (x).

Similarly, using (3.5.7) and (3.5.5), we have

1= Df3 Px(x) + Df3cpg * Rx(x)

= Uf3(x) + Df3cpg * Rx(x).
Therefore,

Df3px (x) = Df3pxo (x) + [Df3cpg * (Rxo - Rx)](x)

= Df3 Pxo(x)

+ J e-{n+ 1f3I )Df3cp [(X~y)] [R(xo,y)-R(x,y)]dy.

Because cp == 0 on Ixl > 1, the last integral is taken over B(x,e). Since
B(x, e) C B(xo, 2e), the integral is dominated by

C [1
JB{xo,2g)
IR(xo, y)ldy + 1
h{x,g)
IR(x, Y)ldY] e- 1f31 (3.5.9)

where C depends on an upper bound for IDf3cpl. Jensen's inequality and

(3.5.8) implies that (3.5.9) is bounded by CMe k- 1f31 , thus proving U E
Tk(E).
136 3. Pointwise Behavior of Sobolev Functions

A similar proof establishes the second assertion of the theorem. Indeed,

as before we obtain

Df3 Px(x) = Df3 PXQ (x) + [Df3<pe * (RxQ - Rx)](x)

= Df3pxQ (x)

+ J e-(n+If31l Df3<p [(X ~ y)] [R(xo, y) - R(x, y))dy

~C [1
k(XQ,2e)
IR(xo, y)ldy + 1
k(x,e)
IR(x, Y)ldY] e- 1f31 .

Since (3.5.3) is assumed to hold uniformly on E, for 'r/ > 0 arbitrary, the
last expression is dominated by 'r/e k- 1f31 = 'r/lx - xol k- 1f31 provided Ix - xol
is sufficiently small. The compactness of E is used in this case to ensure
that IRf3(x, y)1 ~ Clx - ylk-If31 whenever x, y E E. 0

3.6 An V-Version of the Whitney Extension

Theorem
We now return to the Whitney Extension Theorem (Theorem 3.5.3) that
was stated without proof in the previous section. It states that for a com-
pact set E C R n , a function u is an element of t k (E) if and only if it is the
restriction to E of a function of class ck(Rn). The result we establish here,
which was first proved in [CZ), is slightly stronger in that the full strength
of the hypothesis u E tk(E) is not required. Instead, our hypothesis requires
that u E tk,P(x) for all x E E with (3.5.3) holding uniformly on E.
We begin by proving a lemma that establishes the existence of a smooth
function which is comparable to the distance function to an arbitrary closed
set.

3.6.1. Lemma. Let A c Rn be closed and for x E R n let d(x) = d(x,A)

denote the distance from x to A. Let U = {x : d(x) < 1}. Then there is a
function 8 E COO(U - A) and a positive number M = M(n) such that

M-1d(x) ~ 8(x) ~ Md(x), x E U - A,

ID<:t8(x)1 ~ C(a)d(x)l-I<:tl, x E U - A, lal ;::: o.

Proof. Let h(x) = 2~d(x), x E U - A, and consider a cover of U - A by
closed balls {B(x, h(x))}, with center x and radius h(x), x E U - A. From
Theorem 1.3.1 there is a countable set S C U - A such that {B(s, h(s)) :
s E S} is disjointed and

Rn - A ~ {UB(s, 5h(s)) : s E S} ~ U - A.
3.6. An V-Version of the Whitney Extension Theorem 137

With a = (3 = 10 and ,\ = 210' we infer from Lemma 1.3.4 that

1
:3 ::; h(x)/h(s) ::; 3 for s E Sx. (3.6.1 )

Let (}(x) = HO(Sx) ::; C(n) and let TJ : R1 ---+ [0,1] be of class Coo with
TJ(t) = 1 for t ::; 1, TJ(t) = 0 for t ~ 2.
Now define 1j; E coo(Rn) by 1j;(x) = 1](lxi) and Vs E Coo(U) by

vs(x) = h(s)1j; [(:h(S~)] for s E S, x E U.

Note that spt Vs C B(s, 10h(s)), Vs == h(s) on B(s, 5h(s)) and from (3.6.1)
that
IDavs(x)1 ::; h(s)N(a)[5h(s)r 1al
::; 5-laI3Ial-1 N(a)h(x)Ha l for s E Sx,

where N(a) is a bound for IDf31j;I, 1(31 ::; lal. Now define

8(x) = L vs(x) = L vs(x) for x E U.

sES sESx

Clearly,
d~~) = h~) ::; 8(x) ::; 3{}(x)h(x) = 230{}(X)d(X)
and

The following is only a prelude to the V-version of the Whitney exten-

sion theorem, although its proof supplies all of the necessary ingredients.
Its hypothesis only invokes information pertaining to the spaces Tk,P(x)
(bounded difference quotients) and not the spaces tk,P(x) (differentiabil-
ity). In particular, the theorem states that if u is Lipschitz on A (the case
when k = 1) then u can be extended to a Lipschitz function on an open
set containing A. This fact is also contained in the statement of Theorem
3.5.7.

3.6.2. Theorem. Let A C R n be closed and let U = {x : d(x, A) < I}.

If u E V(U), 1 ::; P ::; 00, and there is a positive constant M such that
IluIITk,p(x) ::; M for all x E A, where k is a non-negative integer, then there
exists U E C k- 1,1(U) such that Df3u (x) = Df3 Px(x) for x E A, 0 ::; 1(31 < k.

Proof. Let 8 denote the function determined in Lemma 3.6.1. Define u = u

on A and for x E U - A let

u(x) = 'P15(x) * u(x) (3.6.2)

138 3. Pointwise Behavior of Sobolev Functions

where <P is the function determined by Lemma 3.5.6 and where

Thus, u is defined at x as the convolution of <P8(x) and u evaluated at x.

Because both <P and 8 are of class Coo it is easily verified that u E Coo (U-
A). For x E U, let x* be a point in A such that Ix - x*1 = d(x) = d(x, A).
Because U E Tk,P(x) we may write

(3.6.3)

where

By substituting this expression into (3.6.2), we obtain

D{3u(x) = D{3 [<P8(x) * Px' (x)] + D{3 [<P8(x) * Rx' (x)]

= f
(D{3<P8(x) * Px' (x) + R{3(x, y)Rx' (y)dy

= <P8(x) * (D{3Px')(x) +f R{3(x,y)Rx·(y)dy (3.6.4)

where R{3(x,y) = D{3{8(x)-n<p[(x - y)8(X)-1]}. Applying Lemma 3.5.6 to

the first term on the right side of (3.6.4) we obtain

(3.6.5)

We wish to estimate the remainder term in (3.6.5) which requires an anal-

ysis of R{3(x, y). It can be shown that

and consequently

If R{3(x, y)R x' (Y)dyl ::; C(f3)d(x)-n- I{31 C

JS(x,8)
IRx·(y)ldy. (3.6.6)

Because 8(x) is comparable to d(x) (Lemma 3.6.1) and Ix - x*1 = d(x), it

follows that B(x,8(x)) C B(x*, Kd(x)) for some K > o. Therefore from
(3.6.3) and Holder's inequality,

[ IRx.(y)ldy::; M[Kd(x)]n+k (3.6.7)

JB(x.,Kd(x))
3.6. An V-Version of the Whitney Extension Theorem 139

which along with (3.6.5) and (3.6.6) implies,

Df3 u(x) - Df3px .(x) = Sf3(x*, x) (3.6.8)

where
ISf3(x*,x)1 ~ C(,B,k)Mlx-x*l k -If3I.
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*l. We now proceed to establish the estimate for
arbitrary x* E A.
By assumption liuIiTk,p(x) ~ M for all x E A. Therefore, we may apply
Theorem 3.5.7 to conclude that u E Tk(A). Thus, if xt E A,

Px • (x*) = u(x*)

and

(3.6.9)

where
IRa(xr, x*)1 ~ C(a, k)Mlx* - xrl k - 1al .
By Taylor's theorem for polynomials, it follows that
k-l-If31
Df3px .(x)= " .!..Df3+aPx .(x*)(x-x*)a.
L..J a!
lal=O

Thus, by (3.6.9) and Taylor's theorem,

By Taylor's theorem, it follows that

Therefore, since

Ix - x*1 ~ Ix - xrl and Ix· - xrl ~ Ix* - xl + Ix - xrl ~ 21x - xii,

140 3. Pointwise Behavior of Sobolev Functions

(3.6.10) becomes (after some algebraic simplification)

{k-l)-If31 ~

Df3 Px• (x) - L ~!Df3+0 Pxi (x~)(x - X~)'l = O(lx - x~lk-If3I).

101=0
It follows from (3.6.8), that
(k-l)-If31 1
Df3 u(x) - L ,Df3+oPXi(x~)(x - xi)°
Q.
= O(lx - x~lk-If3I)
101=0
or
(3.6.11)
Thus, (3.6.11) holds whenever xi E A and x E U - A and Theorem 3.5.7
implies that it also holds with Df3 u(x) replaced by uf3(x) whenever x E
A. This implies that Df3 u is a continuous extension of uf3 and that this
extension has a Taylor series expansion about each point in A. Since u E
COO(U - A) it now follows that U E Ck-l(U).
In order to prove that u E Ck-l,l(U) it suffices to show that Df3 u is
Lipschitz, 1.81 = k -1. We know from (3.6.11) that if a E A, and 1.81 = k-1
IDf3 u(x) - Df3 u(a) I :s C(k)Mlx - al (3.6.12)
for x E U. Therefore, it is necessary to consider only the case x, Y E U - A.
First suppose Ix - YI 2: ~d(y) and let a E A be such that d(y) = la - YI·
Then, la - yl :s 21x - yl and
Ix - al :s Ix - yl + Iy - al :s 31x - YI·
Thus, utilizing (3.6.12),
IDf3 u(x) - Df3 u(y) I :s IDf3 u(x) - Df3 u(a) I + IDf3 u(y) - Df3 u(a) I
:s IDf3u(x) - Df3 u(a) I + IDf3 u(y) - Df3 u(a) I
:s C(k)M[lx - al + Iy - all
:s 5C(k)Mlx - YI·
Finally, suppose Ix - yl < ~d(y) and d(y) = la - yl. Using (3.6.5) with
1.81 = k - 1 and the Mean Value theorem, we have

IDf3 u(x) - Df3 u(y) I = J R(a, z) [Rf3(x, z) - Rf3(y, z)l dz

:s Ix - yl J IDxRf3(xo, z)IIR(a, z)ldz (3.6.13)

where Xo is a point on the line segment joining x and y. Now spt Rf3(xo, y) c
B(xo,8(xo)) and 8(xo) :s Cd(xo). Thus,
IDxRf3(xo, z)1 :s C(.8)d(xo)-n-k,
3.6. An V-Version of the Whitney Extension Theorem 141

(1,81 =k - 1). Therefore, (3.6.13) implies

ID.Bu(x) - D.Bu(y) I :5 C(,B)lx - yld(xo)-n-k r

JB(xo,Gd(xo))
IR(a,z)ldz.
(3.6.14)
Since Lip(d) = 1, we have
1
2d(xo) ;::: d(x) ;::: d(y) -Ix - yl > "2d(y) > Ix - yl·

If z E U, Iz - xol :5 Cd(xo), then

Iz - al :5 Iz - xol + Ixo - al
:5 Cd(xo) + Ixo - al
:5 Cd(xo) + Ixo - yl + d(y)
:5 Cd(xo) + Ix - yl + d(y)
:5 Cd(xo) + d(xo) + 2d(xo).
That is,
B(xo, Cd(xo)) C B(a, (C + 3)d(xo)).
Therefore, reference to (3.6.7) implies

r
JB(xo ,Gd(xo))
IR(a, z)ldz :5 C[d(xo)t+k

and this, along with (3.6.14) completes the proof. o

This proof leads directly to the following which is the Whitney extension
theorem in the context of tk,P(x) spaces.

3.6.3. Theorem. Let A C R n be closed and let U = {x : d(x, A) < 1}. If

u E V(U), 1 :5 p :5 00, and u E tk,P(x) for all x E A with (3.5.3) holding
uniformly on A, then there exists u E Ck(U) such that D.Bu(x) = D.Bpx(x)
for x E A, 0 :5 I.BI :5 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
IS.B(x*, x)1 :5 o(lx - x* Ik-I.BI)
and
IRa(xi,x*)1 :5 o(lx* - xil k- 1al ),
thus allowing (3.6.11) to be replaced by
D.Bu(x) - D.BPxr(x) = o(lx - xilk-I.BI).
The remainder of the argument proceeds as before. o
142 3. Pointwise Behavior of Sobolev Functions

3.7 An Observation on Differentiation

We address the technicality of showing that IluIITk,P(x) is a measurable
function of x and then establish a result in differentiation theory that will
be needed later in the sequel.

3.7.1. Lemma. Let u E Tk,P(x) for all x in a measurable set E. Then,

IluIITk,P(x) is a measurable function of x.

Proof. Recall that the norm IluIITk,P(x) is the sum of the numbers Ilullp,
IDOPx(x)l, 0 ~ lal ~ k -1, and the pth root of

supr- kp 1 lu(y) - Px(y)IPdy.

r>O h(x,r)

Also recall that DO Px(x) = uo(x). To show that DO Px(x) is measurable

in x consider the function cp of Lemma 3.5.6 and define

uc(x) = CPc * u(x).

If we write u(y) - Px(Y) = Rx(Y), then

DOuc(x) = DO(cpc * Px)(x) + DO(cpc * Rx)(x)

= DO Px(x) + J E-(n+joj) DOcp [~] Rx(x - y)dy.

The above integral is dominated by

CE-(n+joj) r IRx(x - y)ldy ~ CE-(n+joj)Ek+ n

iB(x,c)
= CE k - joj --+ 0 as E --+ O.

This shows that DO Px(x) is the limit of smooth functions DOue(x) for all
x E E, and is therefore measurable. The remainder of the proof is easy to
establish. 0

3.7.2. Lemma. Let u E V(Rn), 1 ~ p < 00, be such that for some C,
a > 0 and all r > 0,
lip
(
1 lu(y)IPdy )
~ Cra,
h(x,r)

for all x in a measurable set E C Rn. Then, for almost all x E E,

lip
(
1 lu(y)IPdy ) = o(r a ) as r 1 O.
h(x,r)
3.7. An Observation on Differentiation 143

Proof. Without loss of generality we may assume that E is bounded and

that u has compact support. Given c > 0, let ACE be a closed set such
that IE - AI < c. Let U be the open set defined by

U = {x: d(x, A) < I}.

It will suffice to establish the conclusion for almost all x E A.
First, observe that the hypotheses imply that

lim 1 lu(y)ldy = 0
r-+O h(x,r)

for x E A and therefore, u = 0 almost everywhere on A.

Let h(x) = lod(x,A). Recall from Theorem 1.3.1 that there is a count-
able set S C U - A such that {B(s, h(s)) : s E S} is disjointed and

{UB(s,5h(s)) : s E S} ::J U - A.

Therefore, since u = 0 almost everywhere on A,

[[ lu(y)1 dydx < [ [ lu(y)1 dydx

JA J u Ix - yln+a - JA JU-A Ix - yln+a
:5 [L [ lu(Y)I:!a dx
JA SESJB(S,Sh(s» Ix - yl

""1
= L..J
sES B(s,Sh(s»
lu(y)1 1Ix -
A
dx + dy.
yin a
(3.7.1)

Let Xs E A be such that Is - xsl = d(s, A) = d(s). Hence, B(s,5h(s)) C

B(xs, Is - xsi + 5h(s)) and Is - xsl = d(s) = 10h(s). By Jensen's inequality
and the hypothesis of the lemma

lip ( ) lip
lui :5 ( lul P ) :5 C lul P :5 Ch(s)a.
t(S,Sh(S» t(S,Sh(S» t(X.,lSh(S»
(3.7.2)
Now for x E A, y E B(s, 5h(s)), we have

Ix - yl ~ Ix - sl-Is - yl ~ d(s) - 5h(s) = 5h(s).

Hence, for y E B(s,5h(s)) we estimate by spherical coordinates with origin
at y,

[ dx < C [00 r-a-ldr

J A Ix - yln+a - J1Sh(s)1
:5 C(a)h(s)-a.
144 3. Pointwise Behavior of Sobolev Functions

This, along with (3.7.2) yields

rr
JA JB(s,5h(s)) Ix - yl
lu(y~
+
a dydx ~ C(a)h(st.

Since {B(s, h(s)) : s E S} is disjointed, it follows from (3.7.1) that

rr
JA Ju Ix -
lu(y)1 dydx
yln+a
~ C(a) L h(s)n < 00
sES

r lu(y)ldy
and therefore,
< 00
Ju Ix -yln+a
for almost every x E A. Clearly,

r
JRn_u Ix -
lu(y)ldy <
yln+a
00

for all x E A, and therefore

r :-,1_u(=y.:...;.)ld-=y~
JRn Ix - yln+a
< 00

for almost all x E A.

An analysis of the argument shows that this was established by using
only the fact that
1 lui ~ Cra.
h(x,r)

If we apply the above argument with v = lulP , our hypothesis becomes

1 Iv(y)ldy ~ Crap
h(x,r)

for all x E E and therefore

00 > r Iv(y)ldy _ r lu(y)IPdy
JRn Iy - xl n+ap - JRn Iy - xl n+ap
for almost all x E E. But, for all such x, and for e > 0,

r
JB(x,r)
Ilu(y):PdY < e for all small r > o.
Y - X n+pa

That is,
lip
(
1 lu(y)IPdy ) < era
h(x,r)

for all small r. o

3.8. Rademacher's Theorem in the V-Context 145

3.8 Rademacher's Theorem in the V-Context

Recall the fundamental result of Rademacher which states that a Lipschitz
function defined on Rn has a total differential at almost all points (Theorem
2.2.1). We rephrase this result in terms of the present setting by replacing
the hypothesis that u is Lipschitz by u E Tk,P(x) for all x in some set E.
If k = 1 and p = 00, this yields the usual Rademacher hypothesis. The
conclusion we will establish is that u E tk,p (x) for almost all x E E.

3.8.1. Theorem. Let u E Tk,P(x) for all x E E, where E c Rn is measur-

able, k a non-negative integer and 1 < p < 00. Then u E tk,P(x) for almost
all x E E.

Proof. By Lemma 3.7.1 and Lusin's theorem we may assume that E is

compact and that IluIITk.P(x) :::; M for all x E E. Since u E Tk,P(x) for
x E E, we may write u(y) = Px(Y) + Rx(Y) where Px is a polynomial of
degree less than k and where

l' > O. (3.8.1)

From Theorem 3.6.2 it follows that there exists an open set U :::> E and
u E Ck-I,I(U)
such that

D{3u(x) = D{3 Px(x), 0:::; 1,81 < k. (3.8.2)

Because u is of class Ck-l,l it follows from Theorem 2.1.4 that u E Wl~~(Rn)

and therefore we may apply Theorem 3.4.2. Thus, for almost all x E Rn,
there is a polynomial Qx of degree at most k such that u(y) = Qx(y)+Rx(Y)
where

1 IRx(Y)IPdY) lip =
( h(x,r) o(rk) as 1'10. (3.8.3)

Because u E C k - l (R n ), the argument following Definition 3.5.5 implies

that
D{3u(x) = D{3Qx(x), 0:::; 1,81 < k. (3.8.4)
Therefore, in view of (3.8.1), (3.8.2), and (3.8.3)

(1 lu _ U1P ) liP:::; Crk

h(x,r)
for almost all x E E. Appealing to Lemma 3.7.2 we have
lip
(
1 lu - ul P) = o(rk) as l' 10
h(x,r)
146 3. Pointwise Behavior of Sobolev Functions

for almost all x E E. Consequently, for all such x,

lip
(
1 lu(y) - Qx(y)IPdy ) < (1 lu(y) _ U(Y)IPdY) lip
h(x,r) JB(x,r)

+ (1 h(x,r)
Iu(y) _ Qx(y)IPdY) lip

::; o(rk) as r ----> 0,

thus establishing the result. o

3.9 The Implications of Pointwise Differentiability

We have seen in Section 4 of this chapter that Sobolev functions possess
LP-derivatives almost everywhere. This runs parallel to the classical result
that an absolutely continuous function f on the real line is differentiable
almost everywhere. Of course, the converse is false. However, if it is assumed
that f' exists everywhere and that 1f'1 is integrable, then f is absolutely
continuous (Exercise 3.16). It is natural, therefore, to inquire whether this
result has a counterpart in the multivariate V theory. It will be shown
that this question has an affirmative answer. Indeed, we will establish that
if a function has an V derivative everywhere except for a small exceptional
set, and if the coefficients of the associated Taylor polynomial are in V,
then the function is in a Sobolev space.
We begin the investigation by asking the following question. Suppose
u E LP(Rn) has V-derivatives at x ERn; that is, suppose u E tk,P(x)
where k is a positive integer. Then, is it possible to relate the distributional
derivatives of u (which always exist) to the V-derivatives of u? The first
step in this direction is given by the following lemma. First, recall that
u E tk,P(x) if there is a polynomial Px of degree k such that

as r ----> 0, (3.9.1)

and u E Tk,P(x) if there is a polynomial P x of degree less than k and a

number M > 0 such that

0< r < 00.

3.9.1. Lemma. Suppose u E V(R n ), p 2: 1.

3.9. The Implications of Pointwise Differentiability 147

(i) If u E Tk,p(x), then

lim inf CPt

t--->O
* DQu(y) > -00,
with Ix - yl < t, and where DQ u denotes the distributional derivative
of u, 0 ::; lal ::; k;
(ii) If u E tk,P(x), then

lim sup CPt * DQu(y) = DQ Px(x),

t--->O

with Ix - yl < t, 0 ::; lal ::; k.

The function CPt above is a mollifier as described in Section 1.6. Since
CPt E C(f(Rn), its convolution with a distribution T is again a smooth
function. Moreover, for small t and Iy - xl ::; t, the quantity CPt * T(y) gives
an approximate description of the behavior of T in a neighborhood of x.
Indeed, if T is a function, then

lim sup CPt * T(y) = T(x)

t--->o
ly-xl9
whenever x is a Lebesgue point for T. This will be established in the proof of
Lemma 3.9.3. Very roughly then, the statement in (ii) of the above lemma
states that, on the average, the behavior of the distribution DQ u near x
is reflected in the value of the coefficient, DQ Px(x), of the Taylor series
expansion.
Let F(y, t) = CPt * u(y). F is thus a function defined on a subset of R n+l,
namely R n x (0,00) and is smooth in y. The lower and upper limits stated
in (i) and (ii) above can be interpreted as non-tangential approach in Rn+l
of (y, t) to the point (x, 0) whch is located on the hyperplane t = O.

Proof of Lemma 3.9.1. Proof of (ii). Let

u(y) = Px(Y) + Rx(Y) and Ft(Y) = F(y, t).

Then

Therefore

DQ Ft(x + h) = JDQcpt(x + h - y)u(y)dy

= JDQcpt(x + h - y)Px(y)dy + J DQcpt(x + h - y)Rx(y)dy

= (cpt *DQPx)(x+h) + JDQcpt(x+h-y)Rx(y)dy. (3.9.2)

148 3. Pointwise Behavior of Sobolev Functions

There is a constant C = C(ID'PI) such that

ID°'Pt(x + h - y)1 ~ CC n - k

for lal = k. Consequently, for h E B(O, t), it follows that

J
I D°'Pt(X + h - Y)Rx(Y)dyl ~ CC n - k r
lB(x+h,t)
IRx(Y)ldy

~ CC n - k r
1B(x,2t)
IRx(y)ldy ---> 0 as t 1 0,

by (3.9.1). Writing Px in terms of its Taylor series, we have

and therefore DOPx(y) = DOPx(x) for all y E R n if lal = k. Hence,

'Pt * DO Px(x + h) = DO Px(x), and reference to (3.9.2) yields

lim sup 'Pt * DOu(x + h) = DO Px(x), 0~ Ihl ~ t, (3.9.3)

t!O

thus establishing (ii) if lal = k. However, if 0 ~ £ ~ k, then u E tl,P(x) and

the associated polynomial is

Thus, applying (3.9.3) to this case leads to the proof of (ii).

The proof of (i) is similar and perhaps simpler. The only difference is
that because P x is of degree at most k - 1, we have

if lal = k. The integral is estimated as before and its absolute value is seen
to be bounded for all t > 0, thus establishing (i). 0

The next two lemmas, along with the preceding one, will lead to the
main result, Theorem 3.9.4.

3.9.2. Lemma. Let T be a distribution and suppose for all x in an open

set n c Rn that

lim inf 'Pt

t-->O
* T(y) > -00, Ix - yl ~ t,
tES
3.9. The Implications of Pointwise Differentiability 149

where S c (0, 00) is a countable set having 0 as its only limit point. Let C
be a closed set such that C n 0 =I- O. Then there exist N > 0 and an open
set 0 1 C 0 with enO l =I- 0 such that 'Pt * T(y) ;::: -N > -00 whenever
Iy - xl ::; t, x E enO l , t E S.
Proof. Let
F*(x) = inf{'Pt * T(y) : Ix - yl ::; t, t E S}.
Then, F. (x) > - 00 for x E 0 since 0 is the only limit point of S and it is
easy to verify that F. is upper semicontinuous. Thus, the sets

cnOn{x:F.(x);:::-i}, i=1,2, ... ,

are closed relative to C n 0 and their union is C n O. Since C n 0 is of the

second category in itself, the Baire Category theorem implies that one of
these sets has a non-empty interior relative to C n O. D

One of the fundamental results in distribution theory is that a non-

negative distribution is a measure. The following lemma provides a gener-
alization of this fact.

3.9.3. Lemma. Let 8 > 0, N > 0, and suppose S is as in Lemma 3.9.2. If

T is a distribution in an open set 0 such that

'Pt *T(x) > -N > -00 for x E 0, t E Sn (0,8)

and
limsup 'Pt * T(x) ;::: 0 for almost all Xo E 0,
tLO
Ix-xQ l::St
then T is a non-negative measure in O.

Proof. Let 'l/J E .9'(0), 'l/J ;::: 0, and recall from Section 1.7, that the convo-
lution 'Pt * T is a smooth function defined by

'Pt * T(x) = T(Tx(Pt)

where (Pt(y) = 'Pt( -y) and Tx<Pt(Y) = <Pt(Y - x). Then,

T('l/J * <pt} = T * ({; * 'Pt)(O)

= (T * 'Pt) * {;(O)
J
= T * 'Pt( -y){;(y)dy

= J* T 'Pt(Y)'l/J(y)dy.
150 3. Pointwise Behavior of Sobolev Functions

Now 'I/J * CPt --+ 'I/J in 9(n) as t --+ 0+. Moreover, since 'I/J is non-negative
and CPt * T(x) ;?: -N for x E nand t E S n (0,8), it follows with the help
of Fatou's lemma, that
T('I/J) = lim T('I/J * CPt)
tiO
tES

;?: liminf [ T
t!O
tES
in * CPt (y)'I/J(y)dy
;?: [liminfT * CPt (y)'I/J(y)dy
in tiO

k
tES

;?: -N 'I/J(y)dy.

Thus, the distribution T +N has the property that

(T + N)('I/J) ;?: 0 for 'I/J E 9(n), 'I/J;?: o.
That is, T + N is a non-negative measure on n, call it JL. Let JL = v + a
where v is absolutely continuous with respect to Lebesgue measure and a
is singular. Clearly, n is the union of a countable number of sets of finite
v measure. Thus, by the Radon-Nikodym theorem, there exists f E Ll(n)
such that
v(E) = l f(x)dx
for every measurable set E C n. Since T +N = JL, it follows that

CPt * T(x) +N = CPt * (T + N)(x) = CPt * JL(x)

= k CPt(x - y)f(y)dy + k CPt (x - y)da(y), (3.9.4)

for x E n. Because a is a singular measure, a result from classical differen-

tiation theory states that
lim a[B(xo, t)] = 0
t--+OIB(xo, t)1
for almost all Xo En, cf. [SA, Lemma 7.1]. Therefore, at all such Xo with
Ix - xol ~ t,

! cpt(x - y)da(y) = [
i B(x,t)
CPt(x - y)da(y)

~ [ CPt (x - y)da(y)
i B(xo,2t)
< Gil II a[(B(xo,2t)]
- IB(xo, t)1
cP 00

--+ 0 as t --+ 0+ with Ix - xol < t.

3.9. The Implications of Pointwise Differentiability 151

To treat the other term in (3.9.4), recall that f has a Lebesgue point at
almost all Xo E O. That is,

1 If(y) - f(xo)ldy -t 0 as r -t 0+.

h(xo,r)

Therefore,

r
JB~~
f(Y)IPt(X - y)dy - f(xo) = r
JB~~
[f(y) - f(xo)]lPt(X - y)dy

:::; CIIIPlioo 1 If(y) - f(xo)ldy -t 0 as t -t 0, Ix - xol :::; t.

h(xo,2t)
Consequently,
N:::; lim sup IPt * T(x) + N = f(xo)
tLO
Ix-xol:::=;t
for almost all Xo E O. This implies that

v(E) ?= NIEI
for all measurable E c O. Since J.L(E) ?= v(E) it follows that the measure
J.L - N = T is non-negative. 0

3.9.4. Theorem. Let T be a distribution in an open set 0 C Rn and let

f E Ltoc(O). Assume

limsuPIPt *T(y)?= f(x), Ix - yl:::; t,

tLO
for almost all x E 0, and

liminf IPt * T(y) > -00, Ix - yl :::; t,

tLO
tES

for all x E O. Then T - f is a non-negative measure in O.

Proof. We first assume that f == O. Lemma 3.9.2 implies that every open
subset of 0 contains an open subset 0' such that for some N > 0, IPt*T(x) >
-N for x EO', t E S. Lemma 3.9.3 implies that T is a measure in 0'.
Let 0 1 be the union of all open sets 0' cO such that T is a non-negative
measure on 0'. From Remark 1.7.2 we know that T is a measure in 0 1 .
We wish to show that 0 1 = O. Suppose not. Applying Lemma 3.9.2 with
C = R n - 0 1 , there is an open set 0' C 0 such that IPt * T(x) ?= -N
for y E C no', Ix - yl :::; t, and t E S. Let O2 = 0 1 U 0' and note that
O2 - 0 1 = C no'. Let
152 3. Pointwise Behavior of Sobolev Functions

for some e > o. Take e sufficiently small so that 0 3 n (Rn - 0 1 ) # 0.

Consider CPt * T(x) for x E 0 3 and t < e. Now T is a non-negative measure
in 0 1 . Therefore, if d(x,Rn - 0 1 ) > t, CPt * T(x) 2: O. On the other hand,
if d(x, R n - 0 1 ) ::::; t, there exists y E Rn - 0 1 such that Ix - yl : : ; t.
Since B(x, t) C02, it follows that y E O2 - 0 1 = C n 0'. Consequently,
cpt*T(x) 2: -N. Hence, cpt*T(x) is bounded below for x E 0 3 , t E Sn(O, e),
and thus T is a measure in 0 3 by Lemma 3.9.3. But 0 3 n (Rn - ot) # 0
thus contradicting the definition of 0 1 .
For the case 1# 0, for each N > 0 define

N, l(x)2:N
IN(X) = { I(x), -N::::; I(x) ::::; N
-N, I(x)::::;-N

and let R be the distribution defined by R = T - IN. Clearly R satisfies

the same conditions as did T when I was assumed to be identically zero.
Therefore, R is a non-negative measure in O. Thus, for 1/J E CO'(O), 1/J 2: 0,

R(1/J) = T(1/J) - j IN1/Jdx 2: o.

Letting N --+ 00, we have that

T(1/J) - j 11/J dx 2: o.
That is, T - I is a non-negative measure in O. o
Now that Theorem 3.9.4 is established, we are in a position to consider
the implications of a function U with the property that U E Tk,P(x) for
every x E 0, where 0 is an open subset of Rn. From Theorem 3.8.1 we
have that U E tk,P(x) for almost all x E O. Moreover, in view of Lemma
3.9.1 (ii), it follows that whenever U E tk,P(x),

lim sup CPt * DQu(y) = DQPx(x), Ix - yl : : ; t,

t!O

for 0 ::::; lal : : ; k. For convenience of notation, let uQ(x) = DOl Px(x), and
assume U E V(O). Then Theorem 3.9.4 implies that the distribution
Q

DQ u - U is a non-negative measure. Similar reasoning applied to the func-

Q

tion -u implies that DQ( -u) - (-uQ) is a non-negative measure or equiv-

alently, that DQ u - U is a non-positive measure. Thus, we conclude that
Q

DQ u = U almost everywhere in O. That is, the distributional derivatives

Q

of u are functions in V(O). In summary, we have the following result.

3.9.5. Theorem. Let 1 ::::; p < 00 and let k be a non-negative integer. II

U E Tk,P(x) lor every x E 0 and the V-derivatives, U Q' belong to V(O),
3.10. A Lusin-Type Approximation for Sobolev Functions 153

o ~ lal ~ k, then u E Wk,p(n).

Clearly, the hypothesis that the V-derivatives belong to V(n) is neces-

sary. On the other hand, we will be able to strengthen the result slightly
by not requiring that u E Tk,P(x) for all x E n. The following allows an
exceptional set.

3.9.6. Corollary. Let KeRn be compact and let n = R n - K. Suppose

Hn-I[7ri(K)] = 0 where the 7ri: R n --+ Rn-l, i = I,2, ... ,n, are n inde-
pendent orthogonal projections. Assume u E Tk,P(x) for all x E n and that
uo E V(n), 0 ~ lal ~ k. Then u E Wk,P(R n ).
Proof. Assume initially that the projections 7ri are given by

7ri(X) = (XI, ... ,X2, ... ,Xn )

where (Xl, ... , Xi,"" xn) denotes the (n -I)-tuple with the xi-component
deleted. Theorem 3.9.5 implies that U E Wk,p(n). In view of the assumption
on K, reference to Theorem 2.1.4 shows that u E wl,p(Rn) since u has a
representative that is absolutely continuous on almost all lines parallel to
the coordinate axes. Now consider DOu, lal = 1. Since DOu E Wk-l,p(n)
a similar argument shows that DOu E WI,P(R n ) and therefore that u E
W 2 ,p(Rn). Proceeding inductively, we have that u E Wk,P(R n ).
Recall from Theorem 2.2.2 that u E Wk,P(R n ) remains in the space
Wk,P(R n ) when subjected to a linear, non-singular change of coordinates.
Thus, the initial restriction on the projections 7ri is not necessary and the
proof is complete. 0

In the special case of k = 1, it is possible to obtain a similar result

that does not require the exceptional set K to be compact. We state the
following [BAZ, Theorem 4.5], without proof.

3.9.7. Theorem. Let KeRn be a Borel set and suppose Hn-I[7ri(K)] = 0

where the 7ri : R n --+ Rn-l, i = 1,2, ... , n, are n independent orthogonal
projections. Let n = R n - K and assume u E Lfoc(n) has the property that
its partial derivatives exist at each point of n and that they are in Lfoc(n).
Then u E Wl~'~(Rn).

3.10 A Lusin-Type Approximation for Sobolev

Functions
Lusin's Theorem states that a measurable function on a compact inter-
val agrees with a continuous function except perhaps for a closed set of
arbitrarily small measure. By analogy, it seems plausible that a Sobolev
154 3. Pointwise Behavior of Sobolev Functions

function u E Wk,p(n) should agree with a function of class Ck(n) except

for a set of small measure. Moreover, if the requirement concerning the
degree of smoothness is lessened, perhaps it could be expected that there
is a larger set on which there is agreement. That is, one could hope that u
agrees with a function of class Ci(n), 0 ~ £ < k, except for a set of small
Bk_i,p-capacity. Finally, because Sobolev functions can be approximated
in norm by functions of class Ck(n), it is also plausible that the Lusin-type
approximant could be chosen arbitrarily close to u in norm. The purpose
of this and the next section is to show that all of this is possible.
In this section, we begin by showing that if u E Wk,P(R n ), then u agrees
with a function, v, of class C i on the complement of an open set of ar-
bitrarily small Bk_i,p-capacity. In the next section, it will also be shown
that Ilu - vlli,p can be made small. The outline of the proof of the exis-
tence of v is as follows. If u E Wk,P(R n ) and 0 ~ £ ~ k, then Theorem
3.4.2 implies that u E ti,P(x) for all x except for a set of Bk_i,p-capacity O.
This means that the remainder terms tends to 0 (with appropriate speed)
at Bk-i,p-q.e. x E Rn. We have already established that if a function u
has an V-derivative of order £ at all points of a closed set A (that is, if
u E ti,P(x) for each x E A) and if the remainder term tends to 0 in £P uni-
formly on A, then there exists a function v E ci(Rn) which agrees with u
on A (Theorem 3.6.3). Thus, to establish our result, we need to strengthen
Theorem 3.4.2 by showing that the remainder tends uniformly to 0 on the
complement of sets of arbitrarily small capacity. This will be accomplished
in Theorem 3.10.4 below.
In the following, we will adopt the notation
lip

Mp,RU(X) = sup (
1 lu(y)IPdy )
O<r<R JB(x,r)

whenever u E V(Rn), 1 < p < 00, and 0 < R < 00.

3.10.1. Theorem. If 1 < p < 00 and k is a non-negative integer such that
kp ~ n, then there is a constant C = C(k,p, n) such that

(3.10.1)

Proof. We use Theorem 2.6.1 to represent u as u = gk*f where f E V(Rn)

and Ilullk,p rv Ilfllp' Thus, it is sufficient to establish (3.10.1) with Ilullk,p
replaced by Ilfllp. Since lui ~ gk * If I, we may assume f ~ O. Let

E t = {x : Mp,RU(X) > t}

and choose x E E t . For notational convenience, we will assume that x = 0

3.10. A Lusin-Type Approximation for Sobolev Functions 155

and denote B(O, r) = B(r). Thus, there exists 0 < r < R < 1 such that

1 lu(y)IPdy > t P

r
h(r)
or
t(r) (Ln gk(y - w)f(w)dw dy > t P.

Utilizing the simple inequality (a + b)P :::; 2P- 1 (a P + bP) whenever a, b ~ 0,

it therefore follows that either

1 (( gk(y - W)f(W)dW)P dy > 21 - pt p (3.10.2)

h(r) J1wl<5,2r
or
1 (( gk(y - W)f(W)dW)P dy > 2 1 - pt p. (3.10.3)
h(r) Jwl>2r
1

If y E B(r), then from Lemma 2.8.3(i) and the fact that gk :::; Ch, (2.6.3),
we obtain

{ gk(y - w)f(w)dw:::; C ( I f(~~_kdW

J1wl<2r J1y -wl9r Y- W

:::; CrkMf(y),

where C = C(k, n). Thus, in case (3.10.2) holds, we have

t P :::; Cr kp 1 Mf(y)Pdy (3.10.4)

h(r)

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

We will now establish the estimate

{ gk(y - w)f(w)dw :::; C inf ( gk(y - w)f(w)dw (3.10.5)

J1wl>2r YEB(r) J1wl>2r
for all y E B(r). Recall that r < 1. Now if y and ware such that Iyl < r <
2r :::; Iwi :::; 2, we have

~Iwi ~ Iwi + Iyl ~ Iw - yl ~ Iwl-lyl ~ Iwl- I~I.

Consequently, if Yl and Y2 are any two points of B(r), refer to (2.6.3) and
the inequality preceding it to conclude that for some constant C = C(k, n)

(3.10.6)
156 3. Pointwise Behavior of Sobolev Functions

If Iwi > 2 and Y E B(r), then ~Iwi > Iwi + 1 ;::: Iw - yl ;::: Iwl - Iyl ;:::
Iwl - 1 > Iwl/2. Therefore, in this case we also have
(3.10.7)

Our desired estimate (3.10.5) follows from (3.10.6) and (3.10.7). Thus, in
case (3.10.3) holds, there is a constant C = C(k,p,n) such that

tP < C inf
YEB{r)
(rJ 1w l>2r
9k(W - Y)f(W)dW)P

::; C inf (9k * f(y))P.

YEB{r)

To summarize the results of our efforts thus far, for each x E E t there
exists 0 < r < 1 such that either

tp::;Cr kP1 Mf(y)Pdy (3.10.8)

k{x,r)

or
t::;C inf 9k*f(y). (3.10.9)
YEB{x,r)

Let 91 be the family of all closed balls for which (3.10.8) holds. By Theorem
1.3.1, there exists a disjoint subfamily F such that

Bk,p[{UB : B E 9d] ::; Bk,p[{UB : B E F}]

::; L Bk,p(B)
BEF

::; C L (5rt- kp (by Theorem 2.6.13)

B{x,r)EF

::; ~ L
BEF JB
rMf(y)Pdy
C
::; tP Ilfll~ (by Theorem 2.8.2). (3.10.10)

Let 92 be the family of closed balls for which (3.10.9) holds, then the def-
inition of Bessel capacity implies that Bk,p[{UB : B E 92}] ::; (C/tP)lIfll~.
Thus

which establishes our result. o

We now have the necessary information to prove that integral averages
of Sobolev functions can be made uniformly small on the complement of
3.10. A Lusin-Type Approximation for Sobolev Functions 157

sets of small capacity. This result provides an alternate proof of Theorem

3.3.3, as promised in the introduction to Section 1 of this chapter.

3.10.2. Theorem. Let 1 < p < 00 and k be a non-negative integer such

that kp ~ n. If u E wk,p(Rn), then for every E > 0 there exists an open
set U c R n with Bk,p(U) < E such that

1 ju(y) - u(x)JPdy -t 0
h(x,r)
uniformly on Rn - U as r 1 o.

Proof. With the result of Theorem 3.3.3 in mind, we define

Aru(x) = 1 ju(y) - u(x)JPdy

h(x,r)
for x E R n and r > O. Select e such that 0 < e < 1. Since u E wk,p(Rn),
there exists 9 E C~ (Rn) such that

jju - gjjt,p < gP+1 /2.

Set h =u - g. Then

Aru(x) ~ 2P- 1 [A rg(x) + Arh(x)],

Arh(x) ~2 P- 1 (1 h(x,r)
jh(y)jPdy + jh(X)jP) ,

and therefore,

where C = C (p). Consequently, for each x E Rn,

sup Aru(x) ~C[ sup Arg(x) + Mp,Rjhj(x) + jh(X)JP] .

O<r<R O<r<R
Since 9 has compact support, it is uniformly continuous on R n and therefore
there exists 0 < R < 1 such that

sup CArg(x) < e

O<r<R
whenever x E Rn. Therefore,

{ x: sup Aru(x) >

O<r<R
3e} c {x : CMp,Rjhj(x) > n U {x: Cjh(x)JP > n
158 3. Pointwise Behavior of Sobolev Functions

c {x: CMp,Rlhl(x) > e} u {x: (Cr 1)1/Plh(x)1 > I}.

Since h E Wk,P(Rn ), by Theorem 2.6.1 we can write h = gk * f, where
Ilflip '" lihllk,p' Now
{x: (Cr 1 )1/Plh(x)1 > I} C {x: (Cr 1 )1/Pgk * Ifl(x) > I}
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
ei = C- 1e2- i to obtain 0 < Ri < 1 such that

Bk,p [{x: sup Aru(x) > 3ei }] < eTi.

O<r<R;
Let
u= u
i=1
{x: sup Aru(x) >
O<r<R;
3ei}
to establish the conclusion of the theorem. o
3.10.3. Remark. If we are willing to accept a slightly weaker conclusion in
Theorem 3.10.2, the proof becomes less complicated. That is, if we require
only that
1 lu(y) - u(x)ldy ~0
Tn(x,r)

uniformly on R n - U as r ! 0, rather than

1 lu(y) - u(x)IPdy ~ 0,
Tn(x,r)

then an inspection of the proof reveals that it is only necessary to show

C
Bk,p[{X : Mu(x) > t}] ~ tP Il uI11,p'

To prove this, let u = gk * f, where Ilfllp '" Ilulik,p and define

r (x) = {r~ if Ixl ~ r
r 0 otherwise.
3.11. The Main Approximation 159

Then

1 lu(y)ldy = rr * lul(x)
h(x,r)
~ rr * (gk * Ifl)(x)
~ gk * Mlfl(x),
which implies Mu ~ gk * Mlfl. From the definition of capacity,
Bk,p[{X: Mu(x) 2: t}] ~ Bk,p[{X : gk * Mlfl(x) 2: t}]
~ rpllM Iflll~
~ crpllfll~, by Theorem 2.8.2,
~ crPliullt,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 l, k be non-negative integers such that l ~ k and

(k -l)p < n. Let u E Wk,P(R n ). Then, for each e > 0, there exists an open
set U with Bk-l,p(U) < e such that

r- l [1
h(x,r)
lu(y) - P~l)(y)IPdYlI/P ____ 0
uniformly on R n - U as r 1 o.
Finally, as a direct consequence of Theorems 3.10.4 and 3.6.3, we have
the following.

3.10.5. Theorem. Let l, k be non-negative integers such that l ~ k and

(k -l)p < n. Let u E Wk,P(Rn ) and e > O. Then there exists an open set
U c R n and a C t function v on R n , such that
Bk-l,p(U) <e
and
Dav(x) = Dau(x)
for all x E R n - U and 0 ~ lal ~ l.

3.11 The Main Approximation

We conclude the approximation procedure by proving that the smooth
function v obtained in the previous theorem can be modified so as to be
close to u in norm.
160 3. Pointwise Behavior of Sobolev Functions

In addition to some preliminary lemmas, we will need the following ver-

sion of the Poincare inequality which will be proved in Theorem 4.5.1.

3.11.1. Theorem. Let a E (0,1), l a positive integer, and 1 :S p < 00.

Then there exists a constant C = C (a, l, p, n) such that for every non-empty
bounded convex subset n of R n with diameter p and every u E wl,p(n)
for which
In n {x: u(x) = O}I ~ alnl,
we have the inequality

3.11.2. Lemma. Let l be a positive integer and let u be a function Wl,P(R n )

which vanishes outside a bounded open set U. Let fl, a E (0,1) and let

E = BUn {x: inf

O<t~8
IK(x,t) n (Rn -
tn
U)I >
-
a} (3.11.1)

where K(x, t) denotes the closed cube with center x and side-length t. Let
m be a positive integer such that m :S l and let c > O. Then there exists a
function v E wm,p(Rn ) and an open set V such that
(i) IIu - vllm,p < C;
(ii) E c V and vex) = 0 when x E V U (Rn - U).

Proof. For A E (0,1]' let K),. denote the set of all closed cubes of the form

where i l , i 2 , ... ,in are arbitrary integers. Let A :S ~fl and let

be those cubes of K). that intersect E. Let ai be the center of Ki and let

Pi = K(ai' 4A).
Let ( be a Coo function on R n , such that 0 :S ( :S 1, ((x) = 0, when
x E K(O, 1) and ((x) = 1 when x rf. K(O, 3/2). Define

v).(x) = u(x) }] ( (x ;/i ) (3.11.2)

for x ERn. Clearly v).(x) = 0 when d(x, E) :S ~A, so that, for any A, we
can define v by v = v). and find an open set V satisfying (ii).
3.11. The Main Approximation 161

We keep i fixed for the moment and estimate

lIu - VA lIe,p;pi' (3.11.3)

We observe that there exists a constant 1', depending only on n and such
that at most l' of the cubes Pj intersect Pi (including Pi). Denote these by

where s :::; 1'. Then, for x E Pi

VA(x) = u(x)w(x), (3.11.4)

where
II ([(x - ajk)/2,\j.
8

w(x) = (3.11.5)
k=l
Now, for x E Pi and any multi-index a with 0 :::; lal :::; i, we have

where Al depends only on i and n. Hence for almost all x E Pi and any
multi-index "y with 0 :::; I"YI :::; i, we have

11'1
ID'YVA(x)1 :::; A2 L: ,\r-I'YI L: IDf3u(x)l, (3.11.6)
r=O 1f3I=r

where A2 depends only on i and n.

Let y be a point where Ki intersects E. Clearly, there is a subcube Qi of
Pi with center y and edge length 3'\. By (3.11.1), u and hence its derivatives
are zero on a subset Z of Qi with

, (3.11.7)

By applying the Poincare inequality to the interior of the convex set I{ we

obtain, when 1.81 < i,

Li IDf3 u(x)IPdx:::; A3,\p(e-If3IlI'f:eLi IDeu(x)IPdx (3.11.8)

where A3 = A3(i, a,p, n). But, with a suitable constant A3, (3.11.8) will
still hold when 1.81 = i. By (3.11.6) and (3.11.8) (since ,\ :::; 1)

Li ID'YVA(x)IPdx :::; A4 'f:e Li IDeu(x)IPdx (3.11.9)

1
162 3. Pointwise Behavior of Sobolev Functions

for 0::; hi::; t, where A4 = A 4(t,p,a,n). Let

r
X>. = UP i.
i=l

Then

Ix>. ID'Yv>.(x)IPdx ::; A41~i ~ hi IDeu(x)IPdx

for 0 ::; hi ::; t. But each point of X>. belongs to at most T of the cubes Pi,
hence

(3.11.10)

for 0 ::; hi ::; t. Now

Ilu - v>.I!~,p ::; 2P L

O::;I'Y19
[1x>.
ID'Yv>.(x)IPdx + 1x>.
ID'Yu(x)IPdX],

so that by (3.11.10)

Ilu - v>.II~,p ::; A5 L 1

O::;hl::;i x>.nu
ID'Yu(x)IPdx (3.11.11)

where A5 = A 5(t,p,a,n). But

X>. n U c un {x: d(x, aU) < 2vnA}.

Hence I(X>. n U)I -+ 0 as A 1 O. Therefore by (3.11.11)

Ilu - V>.l!i,p -+ 0

as A -+ 0+.
The required function v is now obtained by putting v = v>., with suffi-
ciently small A. 0

3.11.3. Lemma. Let 0 ::; A ::; n. Then there exists a constant C = C(A, n)
such that
1 Ix - yl>.-ndx ::; Cly - zl>.-n, (3.11.12)
k(z,t5)

for all y, z E Rn and all 8 > O.

Proof. We first show that there exists a constant C, such that (3.11.12)
holds when y = 0 and z is arbitrary.
3.11. The Main Approximation 163

When Izl 2: 38, we have

1
Izl::; Ixl + Iz - xl < Ixl + 8::; Ixl + 31zl ,
so that Izl ::; ~Ixl, Ixl A- n ::; AlzIA-n, and (3.11.12) holds.
When Izl < 38,

8- n r
1B(z,8)
IxlA-ndx ::; 8- n r
1B(0,48)
IxlA-ndx = C8 A- n.

Hence, it is clear that (3.11.12) holds with y = O.

Since we have shown that

8- n r
lB(z,8)
IxlA-ndx ::; ClzIA-n,

for all z E Rn, the general result follows by a change of variables; that is,
replace z by z - y. D

Throughout the remainder of this section, it will be more convenient

to employ the Riesz capacity, Rk,p, rather than the Bessel capacity, Bk,p-
This will have no significant effect on the main result, Theorem 3.11.6. See
Remark 3.11.7.

3.11.4. Lemma. Let k be a non-negative real number such that kp < n.

Let U be a bounded non-empty open subset of R n and F a subset of au
with the property that for each x E F, there is atE (0,1) for which

U-,-n_B,----:-(x:-:-'t..:...;.)I >
,--I (J
(3.11.13)
IB(x, t)1 -
where (J E (0,1). Then there exists a constant C = C(n,p, k) such that
(3.11.14)

Proof. Let (J, U, and F be as described above. The cases k = 0 and k > 0
are treated separately.
(i) We consider first the case where k > O. Let 'Ij; be a non-negative
function in V(Rn) with the property that

'},tk) Ln Ix - Ylk-n'lj;(y)dy 2: 1 (3.11.15)

for all x E U. Let C 1 be the constant of Lemma 3.11.3. It can be assumed

that C 1 2: 1. Consider a point bE F and let t be such that (3.11.13) holds
for x = b. By Lemma 3.11.3,

C 1 1y - bl k- n 2: 1 Ix - ylk-ndx
Tn(b,t)
164 3. Pointwise Behavior of Sobolev Functions

so that

and by (3.11.15),

Hence by (3.11.13),

for all x E F, and therefore,

Thus
Rk,p(F) :::; (~2) p Rk,p(U).
The required inequality now follows.
(ii) Now let k = 0, so that Rk,p becomes Lebesgue measure. Let 13 be
the collection of all closed balls B with center in F and radius between 0
and 1 such that
lunBI (3.11.16)
IBI ~a.
Hence, by Theorem 1.3.1, there exists sequence {Br}, Br E 13, such that
Br n BB = 0 when r i= sand

U Br .
00

Fe
r=l
Thus 00 00

IFI:::; L IBrl = 5n L IBrl

r=l r=l
and by (3.11.16)

L IU n Brl :::; 5 a- lUl·

00

< 5n a- l n l

r=l
Since a < 1, the required inequality follows. o
3.11. The Main Approximation 165

3.11.5. Lemma. Let p > 1, k a non-negative real number such that kp < n
and £ a positive integer. There exists a constant C = C(n,p, k, £) such that
for each bounded non-empty open subset U of Rn, each u E Wt,p (Rn)
which vanishes outside U and every 10 > 0 there exists a Coo function v on
R n with the properties
(i) Ilu - vllt,p < 10,
(ii) Rk,p(Spt v) s:; CRk,p(U) and
(iii) sptv c V = Rn n {x: d(x, U) < e}.
Proof. Let U, u, and 10 be as described above. Since U =1= 0, it follows that
Rk,p(U) > O. Let

E=8Un (x: inf IB(x,t)-U)1 > ~}.

l O<t~1/2 tn - 2
(3.11.17)

Then E is closed. By Lemma 3.11.2 there exists a function Vo E wt,p(Rn)

and an open set Vo such that
1
Ilu - vollt,p < '2 10 , (3.11.18)

E c Vo and vo(x) = 0 when x E Vo U (Rn - U). Set

F=8U-E.

Then, for each x E F there exists t E (0,1/2] such that

IU n B(x, t)1 > (1

'---:=---'----,--'-'- (3.11.19)
IB(x, t)1 -
where (1 = 1 - 1/(2o:(n)). Let C 1 be the constant appearing in Lemma
3.11.4. Then
1
Rk,p(U U F) s:; '2CRk,p(U), (3.11.20)

where C = 2CW- p . Let

B = R n n {x: vo(x) =1= O}.

Then B C U U F and hence Rk,p(B) s:; ~CRk,p(U), so that there exists an
open set W with B c Wand

By applying a suitable mollifier to Vo we can obtain a Coo function v with

spt v C V n Wand
(3.11.21)
166 3. Pointwise Behavior of Sobolev Functions

It follows from (3.11.18) and (3.11.21) that v has the required proper-
t~s. 0

We are now in a position to prove the main theorem.

3.11.6. Theorem. Let e, m be positive integers with m ~ e, (e - m)p < n

and let n be a non-empty open subset of Rn. Then, for u E Wl,p(n) and
each c > 0, there exists a c m function v on n such that if

F = n n {x: u(x) ¥- v(x)},

then
Rl-m,p(F) < c and Ilu - vllm,p < c.
Proof. It can be assumed that the set A = n n {x : u( x) ¥- O} is not empty.
Initially, it will be assumed that n = Rn and A bounded. We will show
that there exists a C m function v on R n satisfying the conclusion of the
theorem and that spt v is contained in the set V = R n n {x : d( x, A) < c}.
Let C be the constant of Lemma 3.11.5. Let u be defined by its values
at Lebesgue points everywhere on n except for a set E with Bl,p(E) =
Rl,p(E) = O. By Theorem 3.10.5 there exists an open set U of R n and a
C m function h on R n , such that U :J E,

(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 e- m for k and u - h for u in Lemma 3.11.5, we obtain a Coo
function rp on Rn such that

IIu - h - rpllm,p < c, (3.11.23)

Rl-m,p(Spt rp) ~ CRl-m,p(U), (3.11.24)

and
spt rp c V. (3.11.25)
Put v = h + rp. Then the second part of the theorem follows from (3.11.23).
Clearly,

FeRn n [{x: h(x) ¥- u(x)} u spt rp] C U U spt rp, (3.11.26)

so that by (3.11.24)

Rl-m,p(F) ~ (1 + C)Rl-m,p(U).
3.11. The Main Approximation 167

Thus, the first part of the conclusion follows from (3.11.22). Since spt hand
spt!p are both contained in V, it follows that spt v C V.
We now consider the general case when n is an arbitrary open subset of
Rn. Let {Cj } ~o be an infinite sequence of non-empty compact sets, such
that
Ci C lnt CHI (3.11.27)
for i a non-negative integer and

(3.11.28)

Put C- 1 = 0. For each i 2: 0, let !Pi be a Coo function on R n such that

o ::; !Pi
::; 1,
Ci C int{x : !Pi(X) = I}, (3.11.29)
and
(3.11.30)
Put
'lj;o = !Po and 'lj;i = !Pi - !Pi-l (3.11.31)
when i 2: 1. Then each 'lj;i is Coo on R n with compact support and

(3.11.32)

Hence, for each x E n, 'lj;i(X) i- 0 for at most two values of i. Therefore

00

(3.11.33)

for all x E n. For each i = 0,1,2, ... define

u.(X) = {U(X)'Ij;i(X) when x E n (3.11.34)

, 0 when x ~ n.
By the conclusion of our theorem proved under the assumption that n =
Rn, there exists for each i 2: 0 a cm function Vi on R n with compact
support such that
C
Ilui - villm,p < 2H1 ' (3.11.35)

and
(3.11.36)

where

Moreover,
(3.11.37)
168 3. Pointwise Behavior of Sobolev Functions

For each x E n, there are at most two values of i for which Vi(X) f O.
Hence we can define
L Vi(X)
00

V(x) =
i=O

for x E n. It is easily seen that F C U~OFi' hence

Rl-m,p(F) :::; c.

Also 00

Ilu - vllm,p :::; L Ilui - villm,p < c. o

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:::; C[Rk,p+(Rk,p)n/(n-k p)] for kp < n, cf. [A5].
Therefore, the Riesz capacity in the previous theorem can be replaced by
Bessel capacity.

Exercises
3.1. Prove that the statement

lim 1 u(y)dy = U(X)

r-->O h(x,r)

for B1,p-q.e. x E R n and any u E W1,P(R n ) implies the apparently

stronger statement

lim 1 lu(y) - u(x)ldy = 0

r-->O h(x,r)

for B1,p-q.e. x E Rn. See the beginning of Section 3.3.

3.2. It was proved in Theorem 2.1.4 that a function u E w1,p(Rn) has

a representative that is absolutely continuous on almost all line seg-
ments parallel to the coordinate axes. If a restriction is placed on p,
more information can be obtained. For example, if it is assumed that
p < n - 1, then u is continuous on almost all hyperplanes parallel
to the coordinate planes. To prove this, refer to Theorem 3.10.2 to
conclude that there is a sequence of integral averages

which, for each c > 0, converges uniformly to u on the complement

of an open set Ue: whose B1,p-capacity is less than c. Hence E =
Exercises 169

ng>oug is a set of B1,p-capacity O. It follows from Theorem 2.16.6

(or Exercise 2.16) that the projection of E onto a coordinate axis
has linear measure O. Note that u is continuous on 11"-1 (t), t ¢ 1I"(E)
where 11" denotes the projection. Corresponding results for p > n - k,
k an integer, can be easily stated and proved.

3.3. At the beginning of Section 3.9, an example is given which shows that
u need not be bounded when u E w1,n[B(O, r)], r < 1. This example
can be easily modified to make the pathology even more striking. Let
u(x) = 10glog(1/lxl) for smalllxl and otherwise defined so that u is
positive, smooth and has compact support. Now let
00

v(x) = LTku(x - rk)

k=l

where {rkl is dense in Rn. Then v E w1,n(Rn) and is unbounded in

a neighborhood of each point.

3.4. Use (2.4.18) to show that if u E Wl~'%(Rn), p > n, then u is classically

differentiable almost everywhere.

3.5. Verify that lIuIITk,P(x), which is discussed in Definition 3.5.4, is in fact

a norm.

3.6. If u E W1,P(R n ), the classical Lebesgue point theorem states that

lim 1 lu(x) - u(xo)ldx = 0

r-+O h(xo,r)

for a.e. Xo. Of course, u E L1(Rn) is sufficient to establish this result.

Since u E W1,P(R n ), this result can be improved to the extent that
(*) holds for B1,p-q.e. Xo E Rn (Theorem 3.3.3). Give an example
that shows this result is optimal. That is, show that in general it is
necessary to omit a B1,p-null set for the validity of (*).

3.7. Prove that (3.3.22) can be improved by replacing p by p* = np/(n-

kp).

3.8. A measurable function u is said to have a Lebesgue point at Xo if

lim 1 lu(y) - u(xo)ldy = o.

r-+O h(xo,r)

A closely related concept is that of approximate continuity. A mea-

surable function u is said to be approximately continuous at Xo if
there exists a measurable set E with density 1 at Xo such that u is
continuous at Xo relative to E. Show that if u has a Lebesgue point
170 3. Pointwise Behavior of Sobolev Functions

at xo, then u is approximately continuous at Xo. See Remark 4.4.5.

Show that the converse is true if u is bounded and that it is false
without this assumption.
Another definition of approximate continuity is the following. u IS
approximately continuous at Xo if for every c: > 0, the set

Ac = {x: lu(x) - u(xo)1 2: c:}

has density 0 at Xo. Ac is said to have density 0 at Xo if

Prove that the two definitions of approximate continuity are equiva-

lent.

3.9. The definition of an approximate total differential is analogous to

that of approximate continuity. If u is a real valued function defined
on a subset of Rn, we say that a linear function L : R n - t RI is an
approximate differential of u at Xo if for every c: > 0 the set

A _ { . lu(x) - u(xo) - L(x - xo) }

c- x. Ix -Xo I 2:c:

has density 0 at Xo. Prove the analog of Exercise 3.8; show that if u
is an element of tl,I(XO), then u has an approximate total differential
at Xo.

3.10. The definition of an approximate total differential given in Exercise

3.9 implies that the difference quotient

lu(x) - u(xo) - L(x - xo)

Ix -xol
approaches 0 as x - t Xo through a set E whose density at Xo is
1. In some applications, it is necessary to have more information
concerning the set E. For example, if u E WI,P(R n ), p > n - 1, then
it can be shown that u has a regular approximate total differential at
almost all points Xo. The definition of this is the same as that for an
approximate total differential, except that the set E is required to be
the union of boundaries of concentric cubes centered at Xo. To prove
this, consider

u Xo
(t
,z
) = u(xo + tz)t - u(xo) _ L( )
z ,

and define
Exercises 171

where C is a cube centered at Xo. Since Xo is fixed throughout the

argument, let Ut(z) = u Xo (t, z).
STEP 1. For each Xo and each cube C with Xo as center, observe that
Ut E WI,P( C) for all sufficiently small t > O. With

prove that O!Xo (t) -+ 0 as t -+ 0 for almost all Xo.

STEP 2. Show that U has a regular approximate differential at all Xo
that satisfy the conclusion of Step 1 and for which Du(xo) exists. For
this purpose, let L(z) = Du(xo) . z. Since Ut E WI,P(C), it follows
that Ut E WI,P(Kr ) for almost all r > 0 where Kr is the boundary of
a cube of side length 2r. Moreover, from Exercise 3.2, we know that
Ut is continuous on all such K r . Let

Let E t = [1/2,1] n {r: 'Pt(r) < O!xo(t)I/2} and conclude that

1([1/2,1]- Et)1 :::; O!xo(t)1/2.

STEP 3. Use the Sobolev inequality to prove that for z E K r , r E E t ,

and f3 = (n - 1)/p

IUt(z)1 :::; Mr- f3 ([r lutIPdHn-l) lip

+ Mr l - f3 ([r IDUtIPdHn-l) lip

:::; Mr- f3 'Pt(r)I/P + Mr l - f3 'Pt(r)l/p

:::; [M2f3 + M]O!xo(t)I/2 p,

where M = M(p, n).

STEP 4. Thus, for z E Kr and r E E t ,

rxo(t· r) = r- l sup{lut(z)1 : z E K r }
:::; 2[M2f3 + 1]· O!xo(t)1/2P.

STEP 5. For each positive integer i, let ti = 2- i and let Eti be the
associated set as in Step 2. Set A = U~l Eti and note that 0 is a
point of right density for A (Step 1) and that rxo(t) -+ 0 as t -+ 0,
tEA.
172 3. Pointwise Behavior of Sobolev Functions

3.11. Prove that u defined in (3.6.2) belongs to COO(U - A).

3.12. Give an example which shows that the uniformity condition in the
second part of the statement in Theorem 3.5.7 is necessary.

3.13. In this and the next exercise, it will be shown that a function with
minimal differentiability hypotheses agrees with a C 1 function on a
set oflarge 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 closed and
that u is its partial derivatives are continuous on E.
STEP 2. For each (x, y) E E consider the differences

~(x, y; h, k) = lu(x + h, y + k) - u(x, y) - hD 1u(x, y) - kD 2u(x, y)1

~l(X, y; h) = lu(x + h, y) - u(x, y) - hD 1u(x, y)1

~2(X, y; k) = lu(x, y + k) - u(x, y) - kD 2u(x, y)1
where D1 = a/ax and D2 = a/ay. Choose positive numbers E, T.
Using the information in Step 1, prove that there exists a > 0 such
that the set ACE consisting of all points (x, y) with the property
that
l{x+h:~l(X,y;h)<Th, (x+h,Y)EE,

a::; x ::; b, Ib - al < a, Ihl < Ib - al}1 2: (1- E)lb - al

satisfies IE - AI < E. Perhaps the following informal description of A
will be helpful. For fixed (x, y), let us agree to call a point (x + h, y)
"good" if ~l(x,y;h) ::; Th and (x + h,y) E E. The set A consists
of those points (x, y) with the property that if Ix is any interval
parallel to the x-axis containing x whose length is less than a, then
the relative measure of the set of good points in Ix is large.
STEP 3. Now repeat the analysis of Step 2 with E replaced with A to
obtain a positive number 0'1 < a and a closed set B C A, IA - BI < E
which consists of all points (x, y) with the property that

I{y + k : ~2(X, y; k) ::; Tk, (x, y + k) E A,

a::; y ::; b, Ib - al < 0'1, Ikl < Ib - al}1 2: (1 - E)lb - al·

Exercises 173

STEP 4. Let (12 < (11 be such that

ID 1u(x + h2' y + k2) - D1U(X + hI, y + k 1 )1 < r

for any 2 points (X+h2' y+k2), (X+h1' y+k1) in E with Ih2-h11 < (12,
Ik2 - k11 < (11·
STEP 5. Choose (xo, Yo) E B and let R = [aI, bI] x [a2' b2] be any
rectangle containing (xo, Yo) whose diameter is less than (12 < (11 < (1.
Let

and for each (Yo + k)

(xo + h,yo + k) E E}.

Now for any (h, k) such that Yo + k E E2 and Xo +h E E 1(yo + k),

we have (xo + h, Yo + k) E En R and therefore

~(xo, Yo; h, k) :::; ~l(XO, Yo; h) + ~2(XO, Yo; k)

+ IhIID1u(xo,yo + k) - D1u(xo,yo)1
:::; r(lhl + Ikl).
From this conclude that

IB n R n {(xo + h, Yo + k) : ~(xo, Yo; h, k) :::; 2r(lhl + Iklnl

2: (1- c)2(b 1 - a1)(b 2 - a2) = (1 - c)2IRI·
STEP 6. Take R to be a square with (xo, Yo) as center and appeal to
Exercise 3.8 to reach the desired conclusion.

3.14. We continue to outline the proof that a function whose partial deriva-
tives exist almost everywhere agrees with a C 1 function on a set of
large measure. Let u be a real valued function defined on a measurable
set E c R n , and for each positive number M and x E E let

A(x, M) = En {y : lu(y) - u(x)1 :::; Mly - xl}.

If A(x, M) has density 1 at x, u is said to be of approximate linear
distortion at x. Our objective is to show that if u is of approximate
linear distortion at each point E, then there exists sets Ek such that
E = U~lEk and u is Lipschitzian on each of the sets Ek.
174 3. Pointwise Behavior of Sobolev Functions

STEP 1. If Xl and X2 are any two points of R n , then

is a positive number less than 1 which is independent of the choice of

Xl and X2'

STEP 2. For each positive integer k, let Ek be the set of those points
X E E such that lu(x)1 ::; k and that if r is any number such that
0< r ::; 11k, then

IA(x, k) n B(x, r)1 > 1 _ D:.

IB(x,r)1
Prove that E = Uk:lEk.
STEP 3. In order to show that u is Lipschitz on Ek, choose any two
points Xl,X2 E Ek. If IX2 - xII> 11k, then

IU(X2) - u(xdl ::; 2k21x2 - xII.

Thus, assume that

Let
Al = A(Xl' k) n B(Xl' IX2 - XII),
A2 = A(X2' k) n B(X2' IX2 - XII).
Prove that IAI n A21 > O. If x* E Al n A2, show that
lu(x*) - u(xi)1 ::; klx* - Xii, i = 1,2
Ix* - xii::; IX2 - XII, i = 1,2.
Now conclude that

STEP 4. If u has partial derivatives almost everywhere, appeal to the

previous exercise to conclude that u is of approximate linear distor-
tion at almost every point. Now refer to Theorem 3.11.6 to find a Cl
function that agrees with u on a set of arbitrarily large measure.

3.15. Suppose u E wl,p(Rn). Prove that for Bl,p-q.e. X ERn, U is abso-

lutely continuous on almost every ray .Ax whose endpoint is x.

3.16. Let f be a measurable function defined on [0,1] having the property

that I' exists everywhere on [0,1] and that 11'1 is integrable. Prove
that f is an absolutely continuous function.
Historical Notes 175

Historical Notes
3.1. The idea that an integrable function has a representative that can be
expressed as the limit of integral averages originates with Lebesgue [LE2].
The set of points for which the limit of integral averages does not exist (the
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], [ARS1], [FU], [FL], [GI].
However, optimal results for the exceptional sets in terms of capacity were
obtained in [FZ], [BAZ], [ME2], [CFR]. The development in this section is
taken from [MIZ].
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 2.10.9] 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] , [ME3]. 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 V-sense were first studied
in depth by Calderon and Zygmund [CZ]. They also proved Theorem 3.4.2
where the exceptional set was obtained as a set of Lebesgue measure zero.
The proof of the theorem with the exceptional set expressed in terms of
capacity appears in [BAZ], [ME2], and [CFR].
3.5. The spaces Tk(E), tk(E), Tk,P(x), and tk,P(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 t~(xo) for all Xo E Q, then f E bu(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 3.6.3), observe that the latter is more general in the
176 3. Pointwise Behavior of Sobolev Functions

sense that the remainder term of the function u in question is required to

approach zero only in LP and not in L OO • On the other hand, the function
is required to be defined only on the set E in Whitney's theorem, while the
condition u E tk,P(x) in Theorem 3.5.3 implies that u E V[B(x, r)] (for all
small r > 0) thus requiring u to be defined in a neighborhood of x.
3.9. Theorem 3.9.4 and the preceding lemmas are due to Calderon [CA4]j
the remaining results are from [BAZ].
3.10. The main result ofthis section is Theorem 3.10.2 which easily implies
that Sobolev functions are uniformly differentiable in V on the complement
of sets of small capacity. The proof is due to Lars Hedberg. Observe that
the proof of Theorem 3.10.2 becomes simpler if we are willing to accept a
subsequence {rij} such that

t B(x,rij)
lu(y) - u(x)IPdy - 0

uniformly on Rn-u where Bk,p(U) < €. This can be proved by the methods
of Lemma 2.6.4. However, this result would not be strong enough to apply
Theorem 3.6.3, thus not making it possible to establish the approximation
result in Theorem 3.10.5.
3.11. These results appear in [MIZ]. The main theorem (Theorem 3.11.6)
is analogous to an interesting result proved by J .H. Michael [MI] in the
setting of area theory. He proved that a measurable function f defined on
a closed cube Q C Rn can be approximated by a Lipschitz function 9 such
that
I{x : f(x) i g(xHI < €
and IAU, Q) - A(g, Q)I < € where AU, Q) denotes the Lebesgue area of f
on Q. Theorem 3.11.6 was first proved by Liu [LI] in the case m = f.