Sobolev Spaces

JUHA KINNUNEN
Sobolev spaces
Department of Mathematics, Aalto University

2021
Contents
1 SOBOLEV SPACES 1
1.1 Weak derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Properties of weak derivatives . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Completeness of Sobolev spaces . . . . . . . . . . . . . . . . . . . . 9
1.5 Hilbert space structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Approximation by smooth functions . . . . . . . . . . . . . . . . . . . 12
1.7 Local approximation in Sobolev spaces . . . . . . . . . . . . . . . . . 16
1.8 Global approximation in Sobolev spaces . . . . . . . . . . . . . . . . 17
1.9 Sobolev spaces with zero boundary values . . . . . . . . . . . . . . . 18
1.10 Chain rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.11 Truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.12 Weak convergence methods for Sobolev spaces . . . . . . . . . . . 25
1.13 Difference quotients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.14 Absolute continuity on lines . . . . . . . . . . . . . . . . . . . . . . . . 36
2 SOBOLEV INEQUALITIES 42
2.1 Gagliardo-Nirenberg-Sobolev inequality . . . . . . . . . . . . . . . . 43
2.2 Sobolev-Poincaré inequalities . . . . . . . . . . . . . . . . . . . . . . . 49
2.3 Morrey’s inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.4 Lipschitz functions and W 1,∞ . . . . . . . . . . . . . . . . . . . . . . . . 59
2.5 Summary of the Sobolev embeddings . . . . . . . . . . . . . . . . . . 62
2.6 Direct methods in the calculus of variations . . . . . . . . . . . . . . 62
3 MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 70

3.1 Representation formulas and Riesz potentials . . . . . . . . . . . . . 71
3.2 Sobolev-Poincaré inequalities . . . . . . . . . . . . . . . . . . . . . . . 78
3.3 Sobolev inequalities on domains . . . . . . . . . . . . . . . . . . . . . 86
3.4 A maximal function characterization of Sobolev spaces . . . . . . . 89
3.5 Pointwise estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.6 Approximation by Lipschitz functions . . . . . . . . . . . . . . . . . . . 97
3.7 Maximal operator on Sobolev spaces . . . . . . . . . . . . . . . . . . 102
CONTENTS ii
4 POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 106

4.1 Sobolev capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.2 Capacity and measure . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.3 Quasicontinuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.4 Lebesgue points of Sobolev functions . . . . . . . . . . . . . . . . . . 119
4.5 Sobolev spaces with zero boundary values . . . . . . . . . . . . . . . 124
Sobolev spaces
1
In this chapter we begin our study of Sobolev spaces. The Sobolev space is a
vector space of functions that have weak derivatives. Motivation for studying
these spaces is that solutions of partial differential equations, when they exist,
belong naturally to Sobolev spaces.
1.1 Weak derivatives

Notation. Let Ω ⊂ Rn be open, f : Ω → R and k = 1, 2, . . .. Then we use the following
notations:
C (Ω) = { f : f continuous in Ω}
supp f = { x ∈ Ω : f ( x) 6= 0} = the support of f
C 0 (Ω) = { f ∈ C (Ω) : supp f is a compact subset of Ω}
C k (Ω) = { f ∈ C (Ω) : f is k times continuously diferentiable}
C 0k (Ω) = C k (Ω) ∩ C 0 (Ω)
∞
C∞ = C k (Ω) = smooth functions
\
k=1
C 0∞ (Ω) = C ∞ (Ω) ∩ C 0 (Ω)

= compactly supported smooth functions
= test functions
W A R N I N G : In general, supp f * Ω.
Examples 1.1:
(1) Let u : B(0, 1) → R, u( x) = 1 − | x|. Then supp u = B(0, 1).
1
CHAPTER 1. SOBOLEV SPACES 2
(2) Let f : R → R be 
 x2 , x Ê 0,
f ( x) =
− x2 , x < 0.
Now f ∈ C 1 (R) \ C 2 (R) although the graph looks smooth.
(3) Let us define ϕ : Rn → R,
 1
 e |x|2 −1 , x ∈ B(0, 1),
ϕ( x) =
0, x ∈ Rn \ B(0, 1).
Now ϕ ∈ C 0∞ (Rn ) and supp ϕ = B(0, 1) (exercise).
Let us start with a motivation for definition of weak derivatives. Let Ω ⊂ Rn

be open, u ∈ C 1 (Ω) and ϕ ∈ C 0∞ (Ω). Integration by parts gives
ˆ ˆ
Çϕ Çu
u dx = − ϕ dx.
Ω Çx j Ω Çx j
There is no boundary term, since ϕ has a compact support in Ω and thus vanishes
near ÇΩ.
Let then u ∈ C k (Ω), k = 1, 2, . . ., and let α = (α1 , α2 , . . . , αn ) ∈ Nn (we use the
convention that 0 ∈ N) be a multi-index such that the order of multi-index |α| =
α1 + . . . + αn is at most k. We denote
Ç|α| u Çα1 Çαn
Dα u = α α = α ... α u.
Ç x1 1 . . . Ç xn n Ç x1 1 Ç xn n
T HE MORAL : A coordinate of a multi-index indicates how many times a

function is differentiated with respect to the corresponding variable. The order of
a multi-index tells the total number of differentiations.
Successive integration by parts gives

ˆ ˆ
uD α ϕ dx = (−1)|α| D α uϕ dx.
Ω Ω
Notice that the left-hand side makes sense even under the assumption u ∈ L1loc (Ω).
Definition 1.2. Assume that u ∈ L1loc (Ω) and let α ∈ Nn be a multi-index. Then
v ∈ L1loc (Ω) is the αth weak partial derivative of u, written D α u = v, if
ˆ ˆ
α |α|
u D ϕ dx = (−1) vϕ dx
Ω Ω
for every test function ϕ ∈ C 0∞ (Ω). We denote D 0 u = D (0,...,0) = u. If |α| = 1, then
Du = (D 1 u, D 2 u . . . , D n u)
is the weak gradient of u. Here

Çu
D ju = = D (0,...,1,...,0) u, j = 1, . . . , n,
Çx j
(the j th component is 1).
T HE MORAL : Classical derivatives are defined as pointwise limits of differ-

ence quotients, but the weak derivatives are defined as a functions satisfying the
integration by parts formula. Observe, that changing the function on a set of
measure zero does not affect its weak derivatives.
W A R N I N G We use the same notation for the weak and classical derivatives. It
should be clear from the context which interpretation is used.
Remarks 1.3:
(1) If u ∈ C k (Ω), then the classical partial derivatives up to order k are also
the corresponding weak derivatives of u. In this sense, weak derivatives
generalize classical derivatives.
(2) If u = 0 almost everywhere in an open set, then D α u = 0 almost everywhere
in the same set.
Lemma 1.4. A weak αth partial derivative of u, if it exists, is uniquely defined

up to a set of measure zero.
Proof. Assume that v, ve ∈ L1loc (Ω) are both weak αth partial derivatives of u, that
is, ˆ ˆ ˆ
α |α| |α|
uD ϕ dx = (−1) vϕ dx = (−1) veϕ dx
Ω Ω Ω
for every ϕ ∈ C 0∞ (Ω). This implies that
ˆ
(v − ve)ϕ dx = 0 for every ϕ ∈ C 0∞ (Ω). (1.1)
Ω
Claim: v = ve almost everywhere in Ω.
Reason. Let Ω0 b Ω (i.e. Ω0 is open and Ω0 is a compact subset of Ω). The

space C 0∞ (Ω0 ) is dense in L p (Ω0 ) (we shall return to this later). There exists a
sequence of functions ϕ i ∈ C 0∞ (Ω0 ) such that |ϕ i | É 2 in Ω0 and ϕ i → sgn(v − ve)
almost everywhere in Ω0 as i → ∞. Here sgn is the signum function.
Identity (1.1) and the dominated convergence theorem, with the majorant
v|) ∈ L1 (Ω0 ), give
|(v − ve) ϕ i | É 2(|v| + |e
ˆ ˆ
0 = lim (v − ve)ϕ i dx = lim (v − ve)ϕ i dx
i →∞ Ω0 Ω0 i →∞
ˆ ˆ
= (v − ve) sgn(v − ve) dx = |v − ve| dx
Ω0 Ω0
This implies that v = ve almost everywhere in Ω0 for every Ω0 b Ω. Thus v = ve

almost everywhere in Ω. ■
From the proof we obtain a very useful corollary.

Corollary 1.5 (Fundamental lemma of the calculus of variations). If f ∈ L1loc (Ω)

satisfies ˆ
f ϕ dx = 0
Ω
for every ϕ ∈ C 0∞ (Ω), then f = 0 almost everywhere in Ω.
T HE MORAL : This is an integral way to say that a function is zero almost

everywhere.
Example 1.6. Let n = 1 and Ω = (0, 2). Consider


 x, 0 < x < 1,
u ( x) =
1, 1 É x < 2,
and 
1, 0 < x < 1,
v( x) =
0, 1 É x < 2.
We claim that u0 = v in the weak sense. To see this, we show that

ˆ 2 ˆ 2
uϕ0 dx = − vϕ dx
0 0
for every ϕ ∈ C 0∞ ((0, 2)).
Reason. An integration by parts and the fundamental theorem of calculus give

ˆ 2 ˆ 1 ˆ 2
u( x)ϕ0 ( x) dx = xϕ0 ( x) dx + ϕ0 ( x) dx
0 0 1
¯1 ˆ 1
= xϕ( x)¯¯ − ϕ( x) dx + ϕ(2) −ϕ(1)
¯
0
| {z 0}
|{z}
=0
=ϕ(1)
ˆ 1 ˆ 2
=− ϕ( x) dx = − vϕ( x) dx
0 0
for every ϕ ∈ C 0∞ ((0, 2)). ■
1.2 Sobolev spaces

Definition 1.7. Assume that Ω is an open subset of Rn . The Sobolev space
W k,p (Ω) consists of functions u ∈ L p (Ω) such that for every multi-index α with
|α| É k, the weak derivative D α u exists and D α u ∈ L p (Ω). Thus
W k,p (Ω) = { u ∈ L p (Ω) : D α u ∈ L p (Ω), |α| É k}.

If u ∈ W k,p (Ω), we define its norm

Ã ˆ !1
p
α p
X
k ukW k,p (Ω) = |D u| dx , 1 É p < ∞,
|α|É k Ω
and
ess sup |D α u|.
X
k ukW k,∞ (Ω) =
|α|É k Ω
0
Notice that D u = D (0,...,0)
u = u. Assume that Ω0 is an open subset of Ω. We say
that Ω0 is compactly contained in Ω, denoted Ω0 b Ω, if Ω0 is a compact subset of
k,p
Ω. A function u ∈ Wloc (Ω), if u ∈ W k,p (Ω0 ) for every Ω0 b Ω.
THE MORAL : Thus Sobolev space W k,p (Ω) consists of functions in L p (Ω) that
have weak partial derivatives up to order k and they belong to L p (Ω).
Remarks 1.8:
(1) As in L p spaces we identify W k,p functions which are equal almost every-
where.
(2) There are several ways to define a norm on W k,p (Ω). The norm k · kW k,p (Ω)
is equivalent, for example, with the norm
kD α ukL p (Ω) , 1 É p É ∞.
X
|α|É k
and k · kW k,∞ (Ω) is also equivalent with
max kD α ukL∞ (Ω) .

|α|É k
(3) For k = 1 we use the norm

³ ´1
p p p
k ukW 1,p (Ω) = k ukL p (Ω) + kDukL p (Ω)
µˆ ˆ ¶1
p
p p
= | u| dx + |Du| dx , 1 É p < ∞,
Ω Ω
and
k ukW 1,∞ (Ω) = ess sup | u| + ess sup |Du|.
Ω Ω
We may also consider equivalent norms
Ã !1
n ° p
p °D j u° p p
X °
k ukW 1,p (Ω) = k u k L p (Ω ) + L (Ω )
,
j =1
n °
X °
k ukW 1,p (Ω) = k ukL p (Ω) + °D j u°
L p (Ω ) ,
j =1
and
k ukW 1,p (Ω) = k ukL p (Ω) + kDukL p (Ω)
when 1 É p < ∞ and

© ª
k ukW 1,∞ (Ω) = max k ukL∞ (Ω) , kD1 ukL∞ (Ω) , . . . , kD n ukL∞ (Ω) .
Example 1.9. Let u : B(0, 1) → [0, ∞], u( x) = | x|−α , α > 0. Clearly u ∈ C ∞ (B(0, 1) \
{0}), but u is unbounded in any neighbourhood of the origin.
We start by showing that u has a weak derivative in the entire unit ball. When
x 6= 0 , we have
Çu xj xj
( x) = −α| x|−α−1 = −α , j = 1, . . . , n.
Çx j | x| | x|α+2
Thus
x
Du( x) = −α .
| x|α+2
Gauss’ theorem gives
ˆ ˆ
D j ( uϕ) dx = uϕν j dS,
B(0,1)\B(0,ε) Ç(B(0,1)\B(0,ε))
where ν = (ν1 , . . . , νn ) is the outward pointing unit (|ν| = 1) normal of the boundary
and ϕ ∈ C 0∞ (B(0, 1)). As ϕ = 0 on ÇB(0, 1), this can be written as
ˆ ˆ ˆ
D j uϕ dx + uD j ϕ dx = uϕν j dS.
B(0,1)\B(0,ε) B(0,1)\B(0,ε) ÇB(0,ε)
By rearranging terms, we obtain

ˆ ˆ ˆ
uD j ϕ dx = − D j uϕ dx + uϕν j dS. (1.2)
B(0,1)\B(0,ε) B(0,1)\B(0,ε) ÇB(0,ε)
Let us estimate the last term on the right-hand side. Since ν( x) = − | xx| , we have
xj
ν j ( x) = − | x| , when x ∈ ÇB(0, ε). Thus
¯ˆ ¯ ˆ
uϕν j dS ¯¯ É kϕkL∞ (B(0,1)) ε−α dS
¯ ¯
¯
ÇB(0,ε) ÇB(0,ε)
¯
= kϕkL∞ (B(0,1)) ωn−1 εn−1−α → 0 as ε → 0,
if n − 1 − α > 0. Here ωn−1 = H n−1 (ÇB(0, 1)) is the ( n − 1)-dimensional measure of

the sphere ÇB(0, 1).
Next we study integrability of D j u. We need this information in order to be
able to use the dominated convergence theorem. A straightforward computation
gives
ˆ ˆ ˆ
|Du| dx = α | x|−α−1 dx
¯ ¯
¯D j u¯ dx É
B(0,1) B(0,1) B(0,1)
ˆ 1ˆ ˆ 1
−α−1
=α | x| dS dr = αωn−1 r −α−1+n−1 dr
0 ÇB(0,r ) 0
ˆ 1 ¯1
αωn−1
= αωn−1 r n−α−2 dr = r n−α−1 ¯¯ < ∞,
¯
0 n−α−1 0
if n − 1 − α > 0.
The following argument shows that D j u is a weak derivative of u also in a

neighbourhood of the origin. By the dominated convergence theorem
ˆ ˆ ³ ´
uD j ϕ dx = lim uD j ϕχB(0,1)\B(0,ε) dx
B(0,1) B(0,1) ε→0
ˆ
= lim uD j ϕ dx
ε→0 B(0,1)\B(0,ε)
ˆ ˆ
= − lim D j uϕ dx + lim uϕν j dS
ε→0 B(0,1)\B(0,ε) ε→0 ÇB(0,ε)
ˆ
=− lim D j uϕχB(0,1)\B(0,ε) dx
B(0,1) ε→0
ˆ
=− D j uϕ dx.
B(0,1)
Here we used the dominated convergence theorem twice: First to the function
uD j ϕχB(0,1)\B(0,ε) ,
which is dominated by | u|kD ϕk∞ ∈ L1 (B(0, 1)), and then to the function
D j uϕχB(0,1)\B(0,ε) ,
which is dominated by |Du|kϕk∞ ∈ L1 (B(0, 1)). We also used (1.2) and the fact that
the last term there converges to zero as ε → 0.
Now we have proved that u has a weak derivative in the unit ball. We note
n
that u ∈ L p (B(0, 1)) if and only if − pα + n > 0, or equivalently, α 0, or equivalently, α < p . Thus
n− p
u ∈ W 1,p (B(0, 1)) if and only if α < p .
Let ( r i ) be a countable and dense subset of B(0, 1) and define u : B(0, 1) → [0, ∞],
∞ 1
| x − r i |−α .
X
u ( x) = i
i =1 2
n− p
Then u ∈ W 1,p (B(0, 1)) if α 0, then u is unbounded in every open subset of B(0, 1) and not
differentiable in the classical sense in a dense subset.
THE MORAL : Functions in W 1,p , 1 É p < n, n Ê 2, may be unbounded in every

open subset.
Example 1.10. Observe, that u( x) = | x|−α , α > 0, does not belong to W 1,n (B(0, 1).
However, there are unbounded functions in W 1,n , n Ê 2. Let u : B(0, 1) → R,
 ³ ³ ´´
log log 1 + 1 , x = 6 0,
| x|
u ( x) =
0, x = 0.
Then u ∈ W 1,n (B(0, 1)) when n Ê 2, but u ∉ L∞ (B(0, 1)). This can be used to
construct a function in W 1,n (B(0, 1) that is unbounded in every open subset of
B(0, 1) (exercise).
T HE MORAL : Functions in W 1,p , 1 É p É n, n Ê 2, are not continuous. Later

we shall see, that every W 1,p function with p > n coincides with a continuous
function almost everywhere.
Example 1.11. The function u : B(0, 1) → R,


1, xn > 0,
u( x) = u( x1 , . . . , xn ) =
0, xn < 0,
does not belong to W 1,p (B(0, 1) for any 1 É p É ∞ (exercise).
1.3 Properties of weak derivatives

The following general properties of weak derivatives follow rather directly from
the definition.
Lemma 1.12. Assume that u, v ∈ W k,p (Ω) and |α| É k. Then
(1) D α u ∈ W k−|α|,p (Ω),

(2) D β (D α u) = D α (D β u) for all multi-indices α, β with |α| + |β| É k,
(3) for every λ, µ ∈ R, λ u + µv ∈ W k,p (Ω) and
D α (λ u + µv) = λD α u + µD α v,
(4) if Ω0 ⊂ Ω is open, then u ∈ W k,p (Ω0 ),

(5) (Leibniz’s formula) if η ∈ C 0∞ (Ω), then η u ∈ W k,p (Ω) and
Ã !
α
X α β
D (η u ) = D η D α−β u,
βÉα β
where Ã !
α α!
= , α! = α1 ! . . . αn !
β β!(α − β)!
and β É α means that β j É α j for every j = 1, . . . , n.
THE MORAL : Weak derivatives have the same properties as classical deriva-
tives of smooth functions.
Proof. (1) Follows directly from the definition of weak derivatives. See also (2).
(2) Let ϕ ∈ C 0∞ (Ω). Then D β ϕ ∈ C 0∞ (Ω). Therefore
ˆ ˆ
|β| β α
(−1) D (D u)ϕ dx = D α uD β ϕ dx
Ω Ω
ˆ
= (−1)|α| uD α+β ϕ dx
Ω
ˆ
|α| |α+β|
= (−1) (−1) D α+β uϕ dx
Ω
for all test functions ϕ ∈ C 0∞ (Ω). Notice that
|α| + |α + β| =α1 + . . . + αn + (α1 + β1 ) + . . . + (αn + βn )

=2(α1 + . . . + αn ) + β1 + . . . + βn
=2|α| + |β|.
As 2|α| is an even number, the estimate above, together with the uniqueness
results Lemma 1.4 and Corollary 1.5, implies that D β (D α u) = D α+β u.
(3) and (4) Clear.
(5) First we consider the case |α| = 1. Let ϕ ∈ C 0∞ (Ω). By Leibniz’s rule for
differentiable functions and the definition of weak derivative
ˆ ˆ
α
η uD ϕ dx = ( uD α (ηϕ) − u(D α η)ϕ) dx
Ω Ω
ˆ
=− (ηD α u + uD α η)ϕ dx
Ω
for all ϕ ∈ C 0∞ (Ω). The case |α| > 1 follows by induction (exercise). ä
1.4 Completeness of Sobolev spaces

One of the most useful properties of Sobolev spaces is that they are complete.
Thus Sobolev spaces are closed under limits of Cauchy sequences.
A sequence ( u i ) of functions u i ∈ W k,p (Ω), i = 1, 2, . . ., converges in W k,p (Ω) to
a function u ∈ W k,p (Ω), if for every ε > 0 there exists i ε such that
k u i − ukW k,p (Ω) < ε when i Ê iε.
Equivalently,
lim k u i − ukW k,p (Ω) = 0.
i →∞
A sequence ( u i ) is a Cauchy sequence in W k,p (Ω), if for every ε > 0 there exists
i ε such that
k u i − u j kW k,p (Ω) < ε when i, j Ê i ε .
W A R N I N G : This is not the same condition as
k u i+1 − u i kW k,p (Ω) < ε when i Ê iε.
Indeed, the Cauchy sequence condition implies this, but the converse is not true
(exercise).
Theorem 1.13 (Completeness). The Sobolev space W k,p (Ω), 1 É p É ∞, k =

1, 2, . . ., is a Banach space.
T HE MORAL : The spaces C k (Ω), k = 1, 2, . . . , are not complete with respect

to the Sobolev norm, but Sobolev spaces are. This is important in existence
arguments for PDEs.
Proof. Step 1: k · kW k,p (Ω) is a norm.
Reason. (1) k ukW k,p (Ω) = 0 ⇐⇒ u = 0 almost everywhere in Ω.

=⇒ k ukW k,p (Ω) = 0 implies k ukL p (Ω) = 0, which implies that u = 0 almost every-
where in Ω.
⇐= u = 0 almost everywhere in Ω implies
ˆ ˆ
D α uϕ dx = (−1)|α| uD α ϕ dx = 0
Ω Ω
for all ϕ ∈ C 0∞ (Ω). This together with Corollary 1.5 implies that D α u = 0 almost
everywhere in Ω for all α, |α| É k.
(2) kλ ukW k,p (Ω) = |λ|k ukW k,p (Ω) , λ ∈ R. Clear.
(3) The triangle inequality for 1 É p < ∞ follows from the elementary inequal-
ity (a + b)α É aα + bα , a, b Ê 0, 0 < α É 1, and Minkowski’s inequality, since
Ã !1
p
k D α u + D α v k L p (Ω )
X p
k u + vkW k,p (Ω) =
|α|É k
Ã !1
X ¡ α p
k D u k L p (Ω ) + k D α v k L p (Ω )
¢p
É
|α|É k
Ã !1 Ã !1
p p
kD α ukL p (Ω) kD α vkL p (Ω)
X p X p
É +
|α|É k |α|É k
= k ukW k,p (Ω) + kvkW k,p (Ω) . ■
Step 2: Let ( u i ) be a Cauchy sequence in W k,p (Ω). As
kD α u i − D α u j kL p (Ω) É k u i − u j kW k,p (Ω) , |α| É k,
it follows that (D α u i ) is a Cauchy sequence in L p (Ω), |α| É k. The completeness

of L p (Ω) implies that there exists u α ∈ L p (Ω) such that D α u i → u α in L p (Ω). In
particular, u i → u (0,...,0) = u in L p (Ω).
Step 3: We show that D α u = u α , |α| É k. We would like to argue

ˆ ˆ
α
uD ϕ dx = lim u i D α ϕ dx
Ω i →∞ Ω
ˆ
= lim (−1)|α| D α u i ϕ dx
i →∞ Ω
ˆ
|α|
= (−1) u α ϕ dx
Ω
for every ϕ ∈ C 0∞ (Ω). On the second line we used the definition of the weak
derivative. Next we show how to conclude the first and last equalities above.
1 < p < ∞ Let ϕ ∈ C 0∞ (Ω). By Hölder’s inequality we have
¯ˆ ˆ ¯ ¯ˆ ¯
¯ u i D α ϕ dx − α
ϕ ¯ = ¯ ( u i − u)D α ϕ dx¯
¯ ¯ ¯ ¯
uD dx
Ω Ω Ω
¯ ¯ ¯ ¯
É k u i − ukL p (Ω) kD α ϕkL p0 (Ω) → 0
and consequently we obtain the first inequality above. The last inequality follows
in the same way, since
¯ˆ ˆ ¯
¯ D α u i ϕ dx − ϕ ¯ É kD α u i − u α kL p (Ω) kϕk p0 → 0.
¯ ¯
u α dx L (Ω )
Ω Ω
¯ ¯
p = 1, p = ∞ A similar argument as above (exercise).

This means that the weak derivatives D α u exist and D α u = u α , |α| É k. As
we also know that D α u i → u α = D α u, |α| É k, we conclude that k u i − ukW k,p (Ω) → 0.
Thus u i → u in W k,p (Ω). ä
Remark 1.14. W k,p (Ω), 1 É p < ∞ is separable. In the case k = 1 consider the
mapping u 7→ ( u, Du) from W 1,p (Ω) to L p (Ω) × L p (Ω)n and recall that a subset of a
separable space is separable. However, W 1,∞ (Ω) is not separable (exercise).
1.5 Hilbert space structure

The space W k,2 (Ω) is a Hilbert space with the inner product
〈D α u, D α v〉L2 (Ω) ,
X
〈 u, v〉W k,2 (Ω) =
|α|É k
where ˆ
〈D α u, D α v〉L2 (Ω) = D α uD α v dx.
Ω
Observe that
1
k ukW k,2 (Ω) = 〈 u, u〉 2 .
W k,2 (Ω)
1.6 Approximation by smooth functions

This section deals with the question whether every function in a Sobolev space
can be approximated by a smooth function.
Define φ ∈ C 0∞ (Rn ) by
 1
 c e |x|2 −1 , | x| < 1,
φ( x ) =
0, | x| Ê 1,
where c > 0 is chosen so that

ˆ
φ( x) dx = 1.
Rn
For ε > 0, set

1 ³x´
φ φε ( x ) =
.
εn ε
The function φ is called the standard mollifier or Friedrich’s mollifier. Observe
that φε Ê 0, supp φε = B(0, ε) and
ˆ ˆ ³x´ ˆ ˆ
1 1 n
φε ( x) dx = n φ dx = n φ( y)ε d y = φ( x) dx = 1
R n ε R n ε ε Rn Rn
for all ε > 0. Here we used the change of variable y = εx , dx = εn d y.
Notation. If Ω ⊂ Rn is open with ÇΩ 6= ;, we write
Ωε = { x ∈ Ω : dist( x, ÇΩ) > ε}, ε > 0.
If f ∈ L1loc (Ω), we obtain its standard convolution mollification f ε : Ωε → [−∞, ∞],

ˆ
f ε ( x) = ( f ∗ φε )( x) = f ( y)φε ( x − y) d y.
Ω ■
T HE MORAL : Since the convolution is a weighted integral average of f over

the ball B( x, ε) for every x, instead of Ω it is well defined only in Ωε . If Ω = Rn , we
do not have this problem.
Remarks 1.15:
(1) For every x ∈ Ωε , we have
ˆ ˆ
f ε ( x) = f ( y)φε ( x − y) d y = f ( y)φε ( x − y) d y.
Ω B( x,ε)
(2) By a change of variables z = x − y we have

ˆ ˆ
f ( y)φε ( x − y) d y = f ( x − z)φε ( z) dz
Ω Ω
(3) For every x ∈ Ωε , we have

¯ˆ ¯ ˆ
f ( y)φε ( x − y) d y¯¯ É kφε k∞
¯ ¯
| f ε ( x)| É ¯¯ | f ( y)| d y < ∞.
B( x,ε) B( x,ε)
(4) If f ∈ C 0 (Ω), then f ε ∈ C 0 (Ωε ), whenever

1
0 < ε < ε0 = dist(supp f , ÇΩ).
2
Reason. If x ∈ Ωε s.t. dist( x, supp f ) > ε0 (in particular, for every x ∈ Ωε \

Ωε0 ) then B( x, ε) ∩ supp f = ;, which implies that f ε ( x) = 0. ■
Lemma 1.16 (Properties of mollifiers).

(1) f ε ∈ C ∞ (Ωε ).
(2) f ε → f almost everywhere as ε → 0.
(3) If f ∈ C (Ω), then f ε → f uniformly in every Ω0 b Ω.
p
(4) If f ∈ L loc (Ω), 1 É p < ∞, then f ε → f in L p (Ω0 ) for every Ω0 b Ω.
W A R N I N G : Assertion (4) does not hold for p = ∞, since there are functions
in L∞ (Ω) that are not continuous.
Proof. (1) Let x ∈ Ωε , j = 1, . . . , n, e j = (0, . . . , 1, . . . , 0) (the j th component is 1). Let

h 0 > 0 such that B( x, h 0 ) ⊂ Ωε and let h ∈ R, | h| < h 0 . Then
ˆ
f ε ( x + he j ) − f ε ( x) 1 1 x + he j − y ³ x − y ´¸
· µ ¶
= n φ −φ f ( y) d y.
h ε B( x+he j ,ε)∪B( x,ε) h ε ε
Let Ω0 = B( x, h 0 + ε). Then Ω0 b Ω and B( x + he j , ε) ∪ B( x, ε) ⊂ Ω0 .

Claim:
1 x + he j − y ³ x − y ´¸ 1 Çφ ³ x − y ´
· µ ¶
φ −φ → for every y ∈ Ω0 as h → 0.
h ε ε ε Çx j ε
¡ x− y ¢
Reason. Let ψ( x) = φ ε
. Then
Çψ 1 Çφ ³ x − y ´
( x) = , j = 1, . . . , n
Çx j ε Çx j ε
and ˆ ˆ
h h
Ç
ψ( x + he j ) − ψ( x) = (ψ( x + te j )) dt = D ψ( x + te j ) · e j dt.
0 Çt 0 ■
Thus
ˆ | h|
|ψ( x + he j ) − ψ( x)| É |D ψ( x + te j ) · e j | dt
0
ˆ ¯
1 |h| ¯¯ x + te j − y ¯
µ ¶¯
É D φ ¯ dt
ε 0 ¯ ε ¯
| h|
É k D φ k L ∞ (R n ) .
ε
This estimate shows that we can use the Lebesgue dominated convergence theorem
(on the third row) to obtain
Ç fε f ε ( x + he j ) − f ε ( x)
( x) = lim
Çx j h →0 h
ˆ
1 1 x + he j − y ³ x − y ´¸
· µ ¶
= lim n φ −φ f ( y) d y
h →0 ε Ω0 h ε ε
ˆ
1 1 Çφ ³ x − y ´
= n f ( y) d y
ε Ω0 ε Ç x j ε
ˆ
Çφε
= ( x − y) f ( y) d y
Ω Çx j
0
Çφε
µ ¶
= ∗ f ( x ).
Çx j
A similar argument shows that D α f ε exists and
D α f ε = D α φε ∗ f in Ωε
for every multi-index α.

´
(2) Recall that B( x,ε) φε ( x − y) d y = 1. Therefore we have
¯ˆ ˆ ¯
φε ( x − y) f ( y) d y − f ( x) φε ( x − y) d y¯¯
¯ ¯
| f ε ( x) − f ( x)| = ¯¯
B( x,ε) B( x,ε)
¯ˆ ¯
φε ( x − y)( f ( y) − f ( x)) d y¯¯
¯ ¯
= ¯¯
B( x,ε)
ˆ ³ x− y´
1
É n φ | f ( y) − f ( x)| d y
ε B( x,ε) ε
ˆ
1
É Ω n kφ k L ∞ (R n ) | f ( y) − f ( x ) | d y → 0
|B( x, ε)| B( x,ε)
for almost every x ∈ Ω as ε → 0. Here Ωn = |B(0, 1)| and the last convergence
follows from the Lebesgue’s differentiation theorem.
(3) Let Ω0 b Ω00 b Ω, 0 < ε < dist(Ω0 , ÇΩ00 ), and x ∈ Ω0 . Because Ω00 is compact
and f ∈ C (Ω), f is uniformly continuous in Ω00 , that is, for every ε0 > 0 there exists
δ > 0 such that
| f ( x) − f ( y)| < ε0 for every x, y ∈ Ω00 with | x − y| < δ.
By combining this with an estimate from the proof of (ii), we conclude that
ˆ
1
| f ε ( x) − f ( x)| É Ωn kφkL∞ (Rn ) | f ( y) − f ( x)| d y < Ωn kφkL∞ (Rn ) ε0
|B( x, ε)| B( x,ε)
for every x ∈ Ω0 if ε < δ.

(4) Let Ω0 b Ω00 b Ω.
Claim: ˆ ˆ
| f ε | p dx É | f | p dx
Ω0 Ω00
whenever 0 < ε < dist(Ω0 , ÇΩ00 ) and 0 < ε < dist(Ω00 , ÇΩ).
Reason. Take x ∈ Ω0 . Hölder’s inequality implies

¯ˆ ¯
φε ( x − y) f ( y) d y¯¯
¯ ¯
| f ε ( x)| = ¯
¯
B( x,ε)
ˆ
1− 1p 1
É φ ε ( x − y) φε ( x − y) p | f ( y)| d y
B( x,ε)
µˆ ¶ 10 µˆ ¶1
p p
p
É φ ε ( x − y) d y φε ( x − y)| f ( y)| d y
B( x,ε) B( x,ε)
By raising the previous estimate to power p and by integrating over Ω0 , we obtain

ˆ ˆ ˆ
p
| f ε ( x)| dx É φε ( x − y)| f ( y)| p d y dx
Ω0 Ω0 B( x,ε)
ˆ ˆ
= φε ( x − y)| f ( y)| p dx d y
Ω00 Ω0
ˆ ˆ
= | f ( y)| p φε ( x − y) dx d y
ˆΩ Ω0
00
= | f ( y)| p d y.
Ω00
Here we used Fubini’s theorem and once more the fact that the integral of φε is
one. ■
Since C (Ω00 ) is dense in L p (Ω00 ). Therefore for every ε0 > 0 there exists g ∈ C (Ω00 )
such that
µˆ ¶1
p ε0
| f − g| p dx É .
Ω00 3
By (3), we have g ε → g uniformly in Ω0 as ε → 0. Thus
µˆ ¶1
p
p ¯ ¯ 1 ε0
| g ε − g| dx É sup | g ε − g| ¯Ω0 ¯ p < ,
Ω00 Ω0 3
when ε > 0 is small enough. Now we use Minkowski’s inequality and the previous
claim to conclude that
µˆ ¶1 µˆ ¶1
p p
| f ε − f | p dx É | f ε − g ε | p dx
Ω0 Ω0
µˆ ¶1 µˆ ¶1
p p
p p
+ | g ε − g| dx + | g − f | dx
Ω0 Ω0
µˆ ¶1 µˆ ¶1
p p
p p
É2 | g − f | dx + | g ε − g| dx
Ω00 Ω0
0 0
ε ε
É2 + = ε0 .
3 3
Thus f ε → f in L p (Ω0 ) as ε → 0. ä
1.7 Local approximation in Sobolev spaces

Next we show that the convolution approximation converges locally in Sobolev
spaces.
Theorem 1.17. Let u ∈ W k,p (Ω), 1 É p < ∞. then
(1) D α u ε = D α u ∗ φε in Ωε and
(2) u ε → u in W k,p (Ω0 ) for every Ω0 b Ω.
T HE MORAL : Smooth functions are dense in local Sobolev spaces. Thus

every Sobolev function can be locally approximated with a smooth function in the
Sobolev norm.
Proof. (1) Fix x ∈ Ωε . Then
D α u ε ( x) = D α ( u ∗ φε )( x) = ( u ∗ D α φε )( x)
ˆ
= D αx φε ( x − y) u( y) d y
Ω
ˆ
|α|
= (−1) D αy (φε ( x − y)) u( y) d y.
Ω
Here we first used the proof of Lemma 1.16 (1) and then the fact that
Ç ³ ³ x − y ´´ Ç ³ ³ y − x ´´ Ç ³ ³ x − y ´´
φ =− φ =− φ .
Çx j ε Çx j ε Ç yj ε
For every x ∈ Ωε , the function ϕ( y) = φε ( x − y) belongs to C 0∞ (Ω). Therefore

ˆ ˆ
D αy (φε ( x − y)) u( y) d y = (−1)|α| D α u( y)φε ( x − y) d y.
Ω Ω
By combining the above facts, we see that

ˆ
D α u ε ( x) = (−1)|α|+|α| D α u( y)φε ( x − y) d y = (D α u ∗ φε )( x).
Ω
Notice that (−1)|α|+|α| = 1.

(2) Let Ω0 b Ω, and choose ε > 0 s.t. Ω0 ⊂ Ωε . By (i) we know that D α u ε =
D α u ∗ φε in Ω0 , |α| É k. By Lemma 1.16, we have D α u ε → D α u in L p (Ω0 ) as ε → 0,
|α| É k. Consequently
Ã !1
p
α α p
X
k u ε − ukW k,p (Ω0 ) = kD u ε − D u k L p (Ω 0 ) → 0.
|α|É k ä
1.8 Global approximation in Sobolev spaces

The next result shows that the convolution approximation converges also globally
in Sobolev spaces.
Theorem 1.18 (Meyers-Serrin). If u ∈ W k,p (Ω), 1 É p < ∞, then there exist

functions u i ∈ C ∞ (Ω) ∩ W k,p (Ω) such that u i → u in W k,p (Ω).
T HE MORAL : Smooth functions are dense in Sobolev spaces. Thus every

Sobolev function can be approximated with a smooth function in the Sobolev norm.
In particular, this holds true for the function with a dense infinity set in Example
1.9.
Proof. Let Ω0 = ; and
1
½ ¾
Ω i = x ∈ Ω : dist( x, ÇΩ) > ∩ B(0, i ), i = 1, 2, . . . .
i
Then ∞
Ω= Ωi and Ω1 b Ω2 b . . . b Ω.
[
i =1
Claim: There exist η i ∈ C 0∞ (Ω i+2 \ Ω i−1 ), i = 1, 2, . . ., such that 0 É η i É 1 and

∞
x ∈ Ω.
X
η i ( x) = 1 for every
i =1
This is a partition of unity subordinate to the covering {Ω i }.
Reason. By using the distance function and convolution approximation we can

e i ∈ C 0∞ (Ω i+2 \Ω i−1 ) such that 0 É η
construct η e i = 1 in Ω i+1 \Ω i (exercise).
e i É 1 and η
Then we define
η
e i ( x)
η i ( x ) = P∞ , i = 1, 2, . . . .
j =1 η e j ( x)
Observe that the sum is only over four indices in a neighbourhood of a given
point. ■
Now by Lemma 1.12 (5), η i u ∈ W k,p (Ω) and
supp(η i u) ⊂ Ω i+2 \ Ω i−1 .
Let ε > 0. Choose ε i > 0 so small that
supp(φε i ∗ (η i u)) ⊂ Ω i+2 \ Ω i−1
(see Remark 1.15 (4)) and

ε
kφε i ∗ (η i u) − η i ukW k,p (Ω) < , i = 1, 2, . . . .
2i
By Theorem 1.17 (2), this is possible. Define

∞
X
v= φε i ∗ (η i u).
i =1
This function belongs to C ∞ (Ω), since in a neighbourhood of any point x ∈ Ω, there

are at most finitely many nonzero terms in the sum. Moreover,
° °
°X ∞ X∞ °
kv − ukW k,p (Ω) = ° φε i ∗ (η i u) − η i u°
° °
° i=1 i =1 (Ω )
° k,p
W
∞ °
X
°φε ∗ (η i u) − η i u° k,p
°
É i W (Ω )
i =1
X∞ ε
É = ε. ä
i =1 2i
Remarks 1.19:
(1) The Meyers-Serrin theorem 1.18 gives the following characterization for
the Sobolev spaces W k,p (Ω), 1 É p < ∞: u ∈ W k,p (Ω) if and only if there
exist functions u i ∈ C ∞ (Ω) ∩ W k,p (Ω), i = 1, 2, . . . , such that u i → u in
W k,p (Ω) as i → ∞. In other words, W k,p (Ω) is the completion of C ∞ (Ω) in
the Sobolev norm.
Reason. =⇒ Theorem 1.18.

⇐= Theorem 1.13. ■
(2) The Meyers-Serrin theorem 1.18 is false for p = ∞. Indeed, if u i ∈ C ∞ (Ω) ∩

W 1,∞ (Ω) such that u i → u in W 1,∞ (Ω), then u ∈ C 1 (Ω) (exercise). Thus
special care is required when we consider approximations in W 1,∞ (Ω).
(3) Let Ω0 b Ω. The proof of Theorem 1.17 and Theorem 1.18 shows that for
every ε > 0 there exists v ∈ C 0∞ (Ω) such that kv − ukW 1,p (Ω0 ) < ε.
(4) The proof of Theorem 1.18 shows that not only C ∞ (Ω) but also C 0∞ (Ω) is
dense in L p (Ω), 1 É p < ∞.
1.9 Sobolev spaces with zero boundary

values
In this section we study definitions and properties of first order Sobolev spaces
with zero boundary values in an open subset of Rn . A similar theory can be
developed for higher order Sobolev spaces as well. Recall that, by Theorem 1.18,
the Sobolev space W 1,p (Ω) can be characterized as the completion of C ∞ (Ω) with
respect to the Sobolev norm when 1 É p < ∞.
Definition 1.20. Let 1 É p < ∞. The Sobolev space with zero boundary values
1,p
W0 (Ω) is the completion of C 0∞ (Ω) with respect to the Sobolev norm. Thus
1,p
u ∈ W0 (Ω) if and only if there exist functions u i ∈ C 0∞ (Ω), i = 1, 2, . . . , such that
1,p
u i → u in W 1,p (Ω) as i → ∞. The space W0 (Ω) is endowed with the norm of
W 1,p (Ω).
TH E M O R A L : The only difference compared to W 1,p (Ω) is that functions in

1,p
W0 (Ω) can be approximated by C 0∞ (Ω) functions instead of C ∞ (Ω) functions,
that is,
1,p
W 1,p (Ω) = C ∞ (Ω) and W0 (Ω) = C 0∞ (Ω),
where the completions are taken with respect to the Sobolev norm. A function
1,p
in W0 (Ω) has zero boundary values in Sobolev’s sense. We may say that u, v ∈
1,p
W 1,p (Ω) have the same boundary values in Sobolev’s sense, if u − v ∈ W0 (Ω). This
is useful, for example, in Dirichlet problems for PDEs.
1,p
WA RNING : Roughly speaking a function in W 1,p (Ω) belongs to W0 (Ω), if it
vanishes on the boundary. This is a delicate issue, since the function does not
have to be zero pointwise on the boundary. We shall return to this question later.
1,p
Remark 1.21. W0 (Ω) is a closed subspace of W 1,p (Ω) and thus complete (exer-
cise).
Remarks 1.22:
1,p
(1) Clearly C 0∞ (Ω) ⊂ W0 (Ω) ⊂ W 1,p (Ω) ⊂ L p (Ω).
1,p
(2) If u ∈ W0 (Ω), then the zero extension u
e : Rn → [−∞, ∞],

 u( x), x ∈ Ω,
u
e ( x) =
0, x ∈ R n \ Ω,
belongs to W 1,p (Rn ) (exercise).
Lemma 1.23. If u ∈ W 1,p (Ω) and supp u is a compact subset of Ω, then u ∈

1,p
W0 (Ω).
Proof. Let η ∈ C 0∞ (Ω) be a cutoff function such that η = 1 on the support of u.
Claim: If u i ∈ C ∞ (Ω), i = 1, 2, . . . , such that u i → u in W 1,p (Ω), then η u i ∈ C 0∞ (Ω)

converges to η u = u in W 1,p (Ω).
Reason. We observe that

³ ´1
p p p
kη u i − η ukW 1,p (Ω) = kη u i − η ukL p (Ω) + kD (η u i − η u)kL p (Ω)
É kη u i − η ukL p (Ω) + kD (η u i − η u)kL p (Ω) ,

where
µˆ ¶1
p
kη u i − η ukL p (Ω) = |η u i − η u| p dx
Ω
µˆ ¶1
p
= |η| p | u i − u| p dx
Ω
µˆ ¶1
p
p
É kηk L ∞ (Ω ) | u i − u| dx →0
Ω
and by Lemma 1.12 (5)

µˆ ¶1
p
p
kD (η u i − η u)kL p (Ω) = |D (η u i − η u)| dx
Ω
µˆ ¶1
p
= |( u i − u)D η + (Du i − Du)η| p dx
Ω
µˆ ¶1 µˆ ¶1
p p
É |( u i − u)D η| p dx + |(Du i − Du)η| p dx
Ω Ω
µˆ ¶1
p
É kD ηkL∞ (Ω) | u i − u| p dx
Ω
µˆ ¶1
p
+ kηkL∞ (Ω) |Du i − Du| p dx →0
Ω
as i → ∞. ■
Since η u i ∈ C 0∞ (Ω), i = 1, 2, . . . , and η u i → u in W 1,p (Ω), we conclude that

1,p
u ∈ W0 (Ω). ä
1,p
Since W0 (Ω) ⊂ W 1,p (Ω), functions in these spaces have similar general prop-
erties and they will not be repeated here. Thus we shall focus on properties that
are typical for Sobolev spaces with zero boundary values.
1,p
Lemma 1.24. W 1,p (Rn ) = W0 (Rn ) with 1 É p < ∞.
T HE MORAL : The standard Sobolev space and the Sobolev space with zero
boundary value coincide in the whole space.
1,p
WA RNING : W 1,p (B(0, 1)) 6= W0 (B(0, 1)), 1 É p < ∞. Thus the spaces are not
same in general.
Proof. Assume that u ∈ W 1,p (Rn ). Let η k ∈ C 0∞ (B(0, k + 1)) such that η k = 1 on
1,p
B(0, k), 0 É η k É 1 and |D η k | É c. Lemma 1.23 implies uη k ∈ W0 (Rn ).
Claim: uη k → u in W 1,p (Rn ) as k → ∞.

Reason.
k u − uη k kW 1,p (Rn ) É k u − uη k kL p (Rn ) + kD ( u − uη k )kL p (Rn )

µˆ ¶ 1 µˆ ¶1
p p
= | u(1 − η k )| p dx + |D ( u(1 − η k ))| p dx
Rn Rn
µˆ ¶1 µˆ ¶1
p p
p p
= | u(1 − η k )| dx + |(1 − η k )Du − uD η k )| dx
Rn Rn
µˆ ¶1 µˆ ¶1
p p
É | u(1 − η k )| p dx + |(1 − η k )Du| p dx
Rn Rn
µˆ ¶1
p
+ | uD η k | p dx .
Rn
We note that limk→∞ u(1 − η k ) = 0 almost everywhere and | u(1 − η k )| p É | u| p ∈

L1 (Rn ) will do as an integrable majorant. The dominated convergence theorem
gives
µˆ ¶1
p
| u(1 − η k )| p dx → 0.
Rn
A similar argument shows that
µˆ ¶1
p
p
|(1 − η k )Du| dx →0
Rn
as k → ∞. Moreover, by the dominated convergence theorem

µˆ ¶1 µˆ ¶1
p p
| uD η k | p dx Éc | u| p dx
Rn B(0,k+1)\B(0,k)
µˆ ¶1
p
=c | u| p χB(0,k+1)\B(0,k) dx →0
Rn
as k → ∞. Here | u| p χB(0,k+1)\B(0,k) É | u| p ∈ L1 (Rn ) will do as an integrable majo-

rant. ■
1,p 1,p
Since uη k ∈ W0 (Rn ), i = 1, 2, . . . , uη k → u in W 1,p (Rn ) as k → ∞ and W0 (Ω)
1,p
is complete, we conclude that u ∈ W0 (Ω). ä
1.10 Chain rule

We shall prove some useful results for the first order Sobolev spaces W 1,p (Ω),
1 É p < ∞.
Lemma 1.25 (Chain rule). If u ∈ W 1,p (Ω) and f ∈ C 1 (R) such that f 0 ∈ L∞ (R)
and f (0) = 0, then f ◦ u ∈ W 1,p (Ω) and
D j ( f ◦ u) = f 0 ( u)D j u, j = 1, 2, . . . , n
almost everywhere in Ω.
Proof. By Theorem 1.18, there exist a sequence of functions u i ∈ C ∞ (Ω) ∩ W 1,p (Ω),
. . , such that u i → u in ˆW 1,p (Ω) as i → ∞. Let ϕ ∈ C 0∞ (Ω).
i = 1, 2, . ˆ
Claim: ( f ◦ u)D j ϕ dx = lim f ( u i )D j ϕ dx.
Ω i →∞ Ω
Reason. 1 < p < ∞ By Hölder’s inequality

¯ˆ ˆ ¯ ˆ
ϕ ϕ | f ( u) − f ( u i )||D ϕ| dx
¯ ¯
¯ f ( u ) D j dx − f ( u i ) D j dx ¯É
Ω Ω Ω
¯ ¯
µˆ ¶ 1 µˆ ¶ 10
p 0 p
É | f ( u) − f ( u i )| p dx |D ϕ| p dx
Ω Ω
µˆ ¶ 1 µˆ ¶ 10
p p
0 p p0
É k f k∞ | u − u i | dx |D ϕ| dx → 0.
Ω Ω
On the last row, we used the fact that

¯ˆ u ¯
f 0 ( t) dt¯¯ É k f 0 k∞ | u − u i |.
¯ ¯
| f ( u) − f ( u i )| = ¯¯
ui ■
Finally, the convergence to zero follows, because the first and the last term are
bounded and u i → u in L p (Ω).
p = 1, p = ∞ A similar argument as above (exercise).
Next, we use the claim above, integration by parts for smooth functions and
the chain rule for smooth functions to obtain
ˆ ˆ
( f ◦ u)D j ϕ dx = lim f ( u i ) D j ϕ dx
Ω i →∞ Ω
ˆ
= − lim D j ( f ( u i ))ϕ dx
i →∞ Ω
ˆ
= − lim f 0 ( u i )D j u i ϕ dx
i →∞ Ω
ˆ
=− f 0 ( u)D j uϕ dx
ˆΩ
=− ( f 0 ◦ u)D j uϕ dx, j = 1, . . . , n,
Ω
for every ϕ ∈ C 0∞ (Ω). We leave it as an exercise to show the fourth inequality in

the display above.
Finally, we need to show that f ( u) and f 0 ( u) ÇÇxu are in L p (Ω). Since
j
¯ˆ u ¯
| f 0 ( t) dt¯¯ É k f 0 k∞ | u|,
¯ ¯
| f ( u)| = | f ( u) − f (0)| = ¯¯
0
we have
µˆ ¶1/ p µˆ ¶1
p
p 0 p
| f ( u)| dx É k f k∞ | u| dx < ∞,
Ω Ω
and similarly,
µˆ ¶1 µˆ ¶1
p p
¯ f ( u)D j u¯ p dx p
¯ 0 ¯ 0
É k f k∞ |Du| dx < ∞.
Ω Ω ä
1.11 Truncation
The truncation property is an important property of first order Sobolev spaces,
which means that we can cut the functions at certain level and the truncated
function is still in the same Sobolev space. Higher order Sobolev spaces do not
enjoy this property, see Example 1.6.
Theorem 1.26. If u ∈ W 1,p (Ω), then u+ = max{ u, 0} ∈ W 1,p (Ω), u− = − min{ u, 0} ∈

W 1,p (Ω), | u| ∈ W 1,p (Ω) and

Du almost everywhere in { x ∈ Ω : u( x) > 0},
Du+ =
0 almost everywhere in { x ∈ Ω : u( x) É 0},

0 almost everywhere in { x ∈ Ω : u( x) Ê 0},
Du− =
−Du almost everywhere in { x ∈ Ω : u( x) < 0},
and 


 Du almost everywhere in { x ∈ Ω : u( x) > 0},

D | u| = 0 almost everywhere in { x ∈ Ω : u( x) = 0},


−Du almost everywhere in { x ∈ Ω : u( x) < 0}.

T HE MORAL : In contrast with C 1 , the Sobolev space W 1,p are closed under
taking absolute values.
p
Proof. Let ε > 0 and let f ε : R → R, f ε ( t) = t2 + ε2 − ε. The function f ε has the
1
following properties: f ε ∈ C (R), f ε (0) = 0
lim f ε ( t) = | t| for every t ∈ R,

ε→0
1 t
( f ε )0 ( t) = ( t2 + ε2 )−1/2 2 t = p for every t ∈ R,
2 t + ε2
2
and k( f ε )0 k∞ É 1 for every ε > 0. From Lemma 1.25, we conclude that f ε ◦ u ∈

W 1,p (Ω) and
ˆ ˆ
( f ε ◦ u)D j ϕ dx = − ( f ε )0 ( u)D j uϕ dx, j = 1, . . . , n,
Ω Ω
for every ϕ ∈ C 0∞ (Ω). We note that




 1, t > 0,

0
lim ( f ε ) ( t) = 0, t = 0,
ε→0 


−1, t < 0,
and consequently
ˆ ˆ
| u|D j ϕ dx = lim ( f ε ◦ u)D j ϕ dx
Ω ε→0 Ω
ˆ
= − lim ( f ε )0 ( u)D j uϕ dx
ε→0 Ω
ˆ
=− D j | u|ϕ dx, j = 1, . . . , n,
Ω
for every ϕ ∈ C 0∞ (Ω), where D j | u| is as in the statement of the theorem. We leave

it as an exercise to prove that the first equality in the display above holds.
The other claims follow from formulas
1 1
u+ = ( u + | u|) and u− = (| u| − u).
2 2 ä
Remarks 1.27:
(1) If u, v ∈ W 1,p (Ω), then max{ u, v} ∈ W 1,p (Ω) and min{ u, v} ∈ W 1,p (Ω). More-
over,

Du almost everywhere in { x ∈ Ω : u( x) Ê v( x)},
D max{ u, v} =
Dv almost everywhere in { x ∈ Ω : u( x) É v( x)},
and

Du almost everywhere in { x ∈ Ω : u( x) É v( x)},
D min{ u, v} =
Dv almost everywhere in { x ∈ Ω : u( x) Ê v( x)}.
1,p 1,p 1,p

If u, v ∈ W0 (Ω), then max{ u, v} ∈ W0 (Ω) and min{ u, v} ∈ W0 (Ω) (exer-
cise).
Reason.
1 1
max{ u, v} = ( u + v + | u − v|) and min{ u, v} = ( u + v − | u − v|). ■
2 2
(2) If u ∈ W 1,p (Ω) and λ ∈ R, then Du = 0 almost everywhere in { x ∈ Ω : u( x) =

λ} (exercise).
1,p
(3) If u ∈ W 1,p (Ω) and λ ∈ R, then min{ u, λ} ∈ Wloc (Ω) and

Du almost everywhere in { x ∈ Ω : u( x) < λ},
D min{ u, λ} =
0 almost everywhere in { x ∈ Ω : u( x) Ê λ}.
A similar claim also holds for max{ u, λ}. This implies that a function
u ∈ W 1,p (Ω) can be approximated by the truncated functions
u λ = max{−λ, min{ u, λ}}


λ almost everywhere in { x ∈ Ω : u( x) Ê λ},



= u almost everywhere in { x ∈ Ω : −λ < u( x) < λ},


−λ almost everywhere in { x ∈ Ω : u( x) É −λ},

in W 1,p (Ω). (Here λ > 0.)
Reason. By applying the dominated convergence theorem to
| u − u λ | p É 2 p (| u| p + | u λ | p ) É 2 p+1 | u| p ∈ L1 (Ω),
we have ˆ ˆ
lim | u − u λ | p dx = lim | u − u λ | p dx = 0,
λ→∞ Ω Ω λ→∞
and by applying the dominated convergence theorem to
|Du − Du λ | p É |Du| p ∈ L1 (Ω),
we have
ˆ ˆ
lim |Du − Du λ | p dx = lim |Du − Du λ | p dx = 0.
λ→∞ Ω Ω λ→∞ ■
T HE MORAL : Bounded W 1,p functions are dense in W 1,p .
1.12 Weak convergence methods for So-

bolev spaces
Let 1 É p < ∞ and let Ω ⊂ Rn be an open set. Recall that L p (Ω; Rm ) is the space of
Rm -valued p-integrable functions f : Ω → Rm with m ∈ N. This section discusses
weak convergence techniques for L p (Ω; Rm ) even though most of the results hold
for more general Banach spaces as well.
Definition 1.28. Let 1 < p < ∞ and m ∈ N, and let Ω ⊂ Rn be an open set. A
sequence ( f i ) i∈N of functions in L p (Ω; Rm ) converges weakly in L p (Ω; Rm ) to a
function f ∈ L p (Ω; Rm ), if
ˆ ˆ
lim f i · g dx = f · g dx
i →∞ Ω Ω
0 p
for every g ∈ L p (Ω; Rm ) with p0 = p−1 .
Next we show that weakly convergent sequences are bounded and that the L p
norm is lower semicontinuous with respect to the weak convergence.
Lemma 1.29. Let 1 < p < ∞ and m ∈ N, and let Ω ⊂ Rn be an open set. If a
sequence ( f i ) i∈N converges to f weakly in L p (Ω; Rm ), then ( f i ) i∈N is bounded in
L p (Ω; Rm ). Moreover, we have
k f kL p (Ω;Rm ) É lim inf k f i kL p (Ω;Rm ) . (1.3)

i →∞
Proof. The claim

sup k f i kL p (Ω;Rm ) < ∞.
i
follows from the uniform boundedness principle or the closed graph theorem. In
0
order to prove (1.3), let g ∈ L p (Ω; Rm ) with k gkL p0 (Ω;Rm ) = 1 and
ˆ
k f kL p (Ω;Rm ) = f ( x) · g( x) dx.
Ω
The definition of weak convergence, Cauchy–Schwarz’s and Hölder’s inequalities
imply
ˆ ˆ
k f k L p (Ω ; R m ) = f ( x) · g( x) dx = lim f i ( x) · g( x) dx
Ω i →∞ Ω
ˆ
É lim inf | f i ( x)|| g( x)| dx É lim infk f i kL p (Ω;Rm ) k gkL p0 (Ω;Rm )
i →∞ Ω i →∞
= lim infk f i kL p (Ω;Rm ) . ä

i →∞
THE MORAL : The L p -norm is lower semicontinuous with respect to the weak
convergence.
A bounded sequence in L p (Ω; Rm ) need not have a convergent subsequence.

However, the following result shows that it always has a weakly convergent
subsequence if 1 < p < ∞. This will be important in our applications of weak
convergence. The following result holds, since L p (Ω; Rm ) is reflexive and separable
when 1 < p < ∞. Theorem 1.30 does not hold for p = 1. This can be seen by
considering the standard mollifier that approximates the Dirac’s delta.
Theorem 1.30. Let 1 < p < ∞ and m ∈ N, and let Ω ⊂ Rn be an open set. Assume
that ( f i ) i∈N is a bounded sequence in L p (Ω; Rm ). There exists a subsequence
( f i k )k∈N and a function f ∈ L p (Ω; Rm ) such that f i k → f weakly in L p (Ω; Rm ) as
k → ∞.
T HE MORAL : This shows that L p with 1 < p < ∞ is weakly sequentially

compact, that is, every bounded sequence in L p has a weakly converging subse-
quence. One of the most useful applications of weak convergence is in compactness
arguments. A bounded sequence in L p does not need to have any convergent sub-
sequence with convergence interpreted in the standard L p sense. However, there
exists a weakly converging subsequence.
Remark 1.31. Theorem 1.30 is equivalent to the fact that L p spaces are reflexive
for 1 < p < ∞.
Weak convergence is often too weak mode of convergence and we need tools
to upgrade it to stronger modes of convergence. We begin with the following
result, which is related to Lemma 1.29. The next result holds, since L p (Ω; Rm ) is
a uniformly convex Banach space.
Lemma 1.32. Let 1 < p < ∞ and m ∈ N, and let Ω ⊂ Rn be an open set. Assume
that a sequence ( f i ) i∈N converges to f weakly in L p (Ω; Rm ) and
lim sup k f i kL p (Ω;Rm ) É k f kL p (Ω;Rm ) . (1.4)

i →∞
Then f i → f in L p (Ω; Rm ) as i → ∞.
Observe that, under the assumptions in Lemma 1.32, by (1.3) and (1.4) we
have
k f kL p (Ω;Rm ) É lim inf k f i kL p (Ω;Rm ) É lim sup k f i kL p (Ω;Rm ) É k f kL p (Ω;Rm ) ,

i →∞ i →∞
which implies
lim k f i kL p (Ω;Rm ) = k f kL p (Ω;Rm ) .
i →∞
This means that the limit exists with an equality in (1.4).
Next we discuss another method to upgrade weak convergence to strong
convergence. Mazur’s lemma below asserts that a convex and closed subspace of a
reflexive Banach space is weakly closed.
Theorem 1.33 (Mazur’s lemma). Assume that X is a normed space and that
x i → x weakly in X as i → ∞. Then there exists a sequence of convex combinations
Pm Pm
x i = j=ii a i, j x j , with a i, j Ê 0 and j=ii a i, j = 1, such that e
e x i → x in the norm of X
as i → ∞.
T HE MORAL : For every weakly converging sequence, there is a sequence of

convex combinations that converges strongly. Thus weak convergence is upgraded
to strong convergence for a sequence of convex combinations. Observe that some
of the coefficients a i may be zero so that the convex combination is essentially for
a subsequence.
Remark 1.34. Since L p (Ω; Rm ) is a uniformly convex Banach space, the Banach–
Saks theorem which asserts that a weakly convergent sequence has a subsequence
whose arithmetic means converge in the norm. Let 1 < p < ∞ and m ∈ N, and
let Ω ⊂ Rn be an open set. Assume that a sequence ( f i ) i∈N converges to f weakly
in L p (Ω; Rm ) as i → ∞. Then there exists a subsequence ( f i k )k∈N for which the
arithmetic mean 1k kj=1 f i j converges to f in L p (Ω; Rm ) as k → ∞. The advantage
P
of the Banach–Saks theorem compared to Mazur’s lemma is that we can work

with the arithmetic means instead of more general convex combinations
Remark 1.35. Mazur’s lemma can be used to give a proof for (1.3) (exercise).
Theorem 1.36. Let 1 < p < ∞. Assume that ( u i ) is a bounded sequence in

W 1,p (Ω). There exists a subsequence ( u i k ) and u ∈ W 1,p (Ω) such that u i k → u
weakly in L p (Ω) and Du i k → Du weakly in L p (Ω) as k → ∞. Moreover, if
1,p 1,p
u i ∈ W0 (Ω), i = 1, 2 . . . , then u ∈ W0 (Ω).
Proof. (1) Assume that u ∈ W 1,p (Ω). Denote
f i = ( u i , Du i ) ∈ L p (Ω; Rn+1 )
for every i ∈ N. Then ( f i ) i∈N is a bounded sequence in L p (Ω; Rn+1 ). By Theorem

1.30, there exists a subsequence ( f i k )k∈N that converges weakly to some f in
L p (Ω; Rn+1 ) as k → ∞. Consider f = ( u, v) with u ∈ L p (Ω) and v = (v1 , . . . , vn ) ∈
L p (Ω; Rn ). We show that u ∈ W 1,p (Ω) and that ( u i k , Du i k ) converges to ( u, Du)
weakly in L p (Ω; Rn+1 ) as k → ∞. It suffices to prove that the weak gradient Du
exists in Ω and that v = Du.
By using test functions of the form ( g 1 , 0, . . . , 0) or (0, g 2 , . . . , g n+1 ) in the defini-
tion of weak convergence, we conclude that u i k → u weakly in L p (Ω) and Du i k → v
weakly in L p (Ω; Rn ) as k → ∞. For ϕ ∈ C 0∞ (Ω) and j = 1, . . . , n, we have
ˆ ˆ
uD j ϕ dx = lim u i k D j ϕ dx
Ω k→∞ Ω
ˆ
= − lim D j u i k ϕ dx.
k→∞ Ω
On the other hand, since Du i k → v weakly in L p (Ω; Rn ), by using the test function
0
(0, . . . , ϕ, . . . , 0) ∈ L p (Ω; Rn ), where ϕ is in the j th position, we have
ˆ ˆ
lim D j u i k ϕ dx = v j ϕ dx.
k→∞ Ω Ω
This implies ˆ ˆ
uD j ϕ dx = − v j ϕ dx
Ω Ω
for every ϕ ∈ C 0∞ (Ω) and thus Du = v in Ω. This shows that the weak partial
derivatives D j u, j = 1, . . . , n, exist and belong to L p (Ω). It follows that u ∈ W 1,p (Ω).
1,p
(2) For the second claim, we assume that u i ∈ W0 (Ω) for every i ∈ N and that
the sequence
( f i k )k∈N = (( u i k , Du i k ))k∈N
converges weakly to f = ( u, Du) in L p (Ω; Rn+1 ). By Theorem 1.33, there exists a

sequence of convex combinations
m
Xk m
Xk
hk = a k, j f i j = a k, j ( u i j , Du i j )
j=k j=k
that converges to f = ( u, Du) in L p (Ω; Rn+1 ) as k → ∞. This implies

m
Xk m
Xk
a k, j u i j → u and a k, j Du i j → Du
j=k j=k
in L p (Ω) as k → ∞ and thus

m
Xk
a k, j u i j → u
j=k
in W 1,p (Ω) as k → ∞. Moreover,

m
Xk 1,p
a k, j u i j ∈ W0 (Ω)
j=k
1,p
for every k ∈ N. Since W0 (Ω) is a closed subspace of W 1,p (Ω), it follows that
1,p
u ∈ W0 (Ω). ä
Remarks 1.37:
(1) Theorem 1.36 is equivalent to the fact that W 1,p spaces are reflexive for
1 < p < ∞.
(2) Since u i k → u weakly in L p (Ω) and Du i k → Du weakly in L p (Ω) as k → ∞
in Theorem 1.36, by Lemma 1.29 we have
k ukW 1,p (Ω) É lim inf k u i k kW 1,p (Ω) .

k→∞
Thus the W 1,p -norm is lower semicontinuous with respect to the weak
convergence in W 1,p .
(3) Another way to see that W 1,p spaces are reflexive for 1 < p < ∞ is to
recall that a closed subspace of a reflexive space is reflexive. Thus it is
enough to find an isomorphism between W 1,p (Ω) and a closed subspace of
L p (Ω, Rn+1 ) = L p (Ω, Rn ) × · · · × L p (Ω, Rn ). The mapping u 7→ ( u, Du) will do
1,p
for this purpose. This holds true for W0 (Ω) as well. This approach can be
used to characterize elements in the dual space by the Riesz representation
theorem, see [3, p. 62–65], [14, Section 11.4], [17, Section 4.3].
Example 1.38. The Sobolev space is not compact in the sense that every bounded
sequence ( u i ) in W 1,p (Ω) has a converging subsequence ( u i k ) and u ∈ W 1,p (Ω) such
that u i k → u in W 1,p (Ω). For i = 1, 2, . . . , consider u i : (0, 2) → R,



0, 0 < x É 1,

u i ( x) = ( x − 1) i, 1 É x É 1 + 1i , (1.5)


1 + 1i < x < 2.

1,
Then u i ∈ W 1,1 ((0, 2)) and k u i kW 1,1 ((0,2)) É 2 for every i = 1, 2, . . . . However, there
does not exist a subsequence that converges in W 1,1 ((0, 2)). To conclude this,
assume that there exists a subsequence ( u i k ) that converges in W 1,1 ((0, 2)). In
particular, the subsequence ( u i k ) converges in L1 ((0, 2)) and the limit function
u ∈ L1 ((0, 2)) is 
0, 0 < x É 1,
u ( x) =
1, 1 < x < 2.
However, u ∉ W 1,1 ((0, 2)). This example also shows that Theorem 1.36 does not
hold true in the case p = 1 (exercise).
Example 1.39. For i = 1, 2, . . . , consider u i : (0, 2) → R,




 0, 0 < x É 1,
 p
u i ( x) = ( x − 1) i, 1 É x É 1 + 1i


 p1 ,

1 + 1i < x < 2.
i
Then u i ∈ W 1,2 ((0, 2)),
1 i − 1 3i − 2
k u i k2L2 ((0,2)) = + 2 = , kDu i k2L2 ((0,2)) = 1,
3 i2 i 3 i2
for every i = 1, 2, . . . and u i → u weakly in W 1,2 (Ω) as i → ∞, where u = 0 (exercise).

Clearly
0 = k u||W 1,2 ((0,2)) < 1 É lim inf k u i kW 1,2 ((0,2)) .
i →∞
This shows that norm is only lower semicontinous in the weak topology but not
continuous. Observe carefully that u i 9 u in W 1,2 ((0, 2)), since since
lim kDu i k2L2 ((0,2)) = 1 6= 0.

i →∞
Remarks 1.40:
(1) An open set Ω ⊂ Rn is an extension domain for W 1,p (Ω), if there exists a
bounded linear operator L : W 1,p (Ω) → W 1,p (Ω) such that Lu = u on Ω for
every u ∈ W 1,p (Ω). It can be shown that open sets with Lipschitz boundary
are extension domains, see [8, Section 4.4] and [14, Section 13.1]. Observe
1,p
that every open set Ω ⊂ Rn is an extension domain for W0 (Ω), since we
may take the zero extension to Rn \ Ω.
(2) Extension domains have certain compactness results that are useful, for
np
example, in the existence theory for PDEs. Let 1 É p < n and p∗ = n− p .
Assume that Ω ⊂ Rn is an extension domain with finite measure. The
Kondrachov-Rellich compactness theorem asserts that the embedding
W 1,p (Ω) → L q (Ω) is compact. This means that for every bounded sequence
( u i ) in W 1,p (Ω), 1 É p < n, there exists a subsequence ( u i k ) and u ∈ W 1,p (Ω)
such that u i k → u in L q (Ω) as k → ∞ for every 1 É q < p∗ . For the proof,
1,p
see [8, Section 4.6] and [14, Theorem 12.18]. Moreover, if u i ∈ W0 (Ω),
1,p
i = 1, 2 . . . , then u ∈ W0 (Ω) for every open set Ω ⊂ Rn .
Theorem 1.41. Let 1 < p < ∞ and let Ω ⊂ Rn be an open set. Assume that ( u i ) i∈N
is a bounded sequence in W 1,p (Ω) such that u i → u weakly in L p (Ω) as i → ∞ or
that u i → u almost everywhere in Ω as i → ∞. Then u ∈ W 1,p (Ω), u i → u weakly
1,p
in L p (Ω), and Du i → Du weakly in L p (Ω; Rn ) as i → ∞. Moreover, if u i ∈ W0 (Ω)
1,p
for every i ∈ N, then u ∈ W0 (Ω).
T HE MORAL : In order to show that u ∈ W 1,p (Ω) it is enough to construct

functions u i ∈ W 1,p (Ω), i = 1, 2, . . . , such that u i → u almost everywhere in Ω as
i → ∞ and sup i k u i kW 1,p (Ω) < ∞.
Proof. (1) It suffices to prove that u ∈ W 1,p (Ω) and that ( u i , Du i ) → ( u, Du) weakly
in L p (Ω; Rn+1 ) as i → ∞. We prove the latter claim by showing that each subse-
quence ( u i k )k∈N has a further subsequence, also denoted by ( u i k )k∈N , such that
( u i k , Du i k ) → ( u, Du) (1.6)
p0
weakly in L p (Ω; Rn+1 ) as k → ∞. To see this let g ∈ L (Ω; Rn+1 ). Then
ˆ ˆ
ai = ( u i , Du i ) · g dx → ( u, Du) · g dx = a
Ω Ω
as i → ∞, since otherwise the definition of convergent real-valued sequences

implies that (a i ) i∈N has a subsequence whose all subsequences fail to converge to
a. This is a contradiction with respect to (1.6) when tested with g.
Let ( u i k )k∈N be a subsequence of ( u i ) i∈N . By Theorem 1.36, there exists a
subsequence, also denoted by ( u i k )k∈N , and a function v ∈ W 1,p (Ω) such that
( u i k , Du i k ) → (v, Dv) weakly in L p (Ω; Rn+1 ) as k → ∞. It suffices to show that
u = v almost everywhere. If u i → u weakly in L p (Ω), then u i k → u weakly in
L p (Ω) and u = v almost everywhere by the uniqueness of weak limits. Hence we
may assume that u i → u almost everywhere in Ω as i → ∞.
Since u i → u almost everywhere in Ω as i → ∞, by Theorem 1.33, there exists
a sequence of convex combinations
m
Xk
hk = a k, j ( u i j , Du i j )
j=k
that converges to (v, Dv) in L p (Ω; Rn+1 ) as k → ∞. In particular,

m
Xk
h k,1 = a k, j u i j
j=k
converges to v in L p (Ω) as k → ∞, and therefore some subsequence of ( h k,1 )k∈N

converges to v almost everywhere in Ω. On the other hand, by the assumptions,
m
Xk
lim h k,1 = lim a k, j u i j = u
k→∞ k→∞ j = k
almost everywhere in Ω. This shows that u = v almost everywhere in Ω, from

which we conclude that u ∈ W 1,p (Ω) and that (1.6) holds.
1,p 1,p
(2) If u i ∈ W0 (Ω) for every i ∈ N, then u ∈ W0 (Ω) by the first part of the
proof and Theorem 1.36. ä
Remark 1.42. Theorem 1.36 and Theorem 1.41 do not hold when p = 1 (exercise).
In order to demonstrate weak convergence techniques, we prove continuity of

the operator u 7→ | u| in Sobolev spaces W 1,p (Ω). This operator is easily shown to
be bounded in Sobolev spaces, but continuity is not a consequence of boundedness,
since the operator is not linear.
Theorem 1.43. Let 1 < p < ∞ and let Ω ⊂ Rn be an open set. Assume that ( u i ) i∈N
is a sequence in W 1,p (Ω) that converges to u in W 1,p (Ω). Then | u i | → | u| in W 1,p (Ω)
as i → ∞.
Proof. Since u i → u in L p (Ω) and
|| u i ( x)| − | u( x)|| É | u i ( x) − u( x)|
for every x ∈ Ω, we obtain | u i | → | u| in L p (Ω) as i → ∞.

Next we discuss convergence of the gradients. Since ( u i ) i∈N converges in
W 1,p (Ω), it ss a bounded sequence in W 1,p (Ω). Theorem 1.26 gives |D | u i |( x)| =
|Du i ( x)| and |D | u|( x)| = |Du( x)| for almost every x ∈ Ω. It follows that (| u i |) i∈N
is a bounded sequence in W 1,p (Ω). Since | u i | → | u| weakly in L p (Ω) as i → ∞,
Theorem 1.41 implies D | u i | → D | u| weakly in L p (Ω; Rn ) as i → ∞.
To upgrade weak convergence to strong convergence, we note that
ˆ ˆ
p
lim |D | u i |( x)| dx = lim |Du i ( x)| p dx
i →∞ Ω i →∞ Ω
ˆ ˆ
= |Du( x)| p dx = |D | u|( x)| p dx.
Ω Ω
Lemma 1.32 implies that D | u i | converges to D | u| in L p (Ω; Rn ), as i → ∞. This

shows that | u i | → | u| in W 1,p (Ω) as i → ∞. ä
As a final result in this section we show that pointwise uniform bounds are
preserved under weak convergence.
Theorem 1.44. Let 1 < p < ∞ and m ∈ N, and let Ω ⊂ Rn be an open set. Assume
that sequences ( f i ) i∈N and ( g i ) i∈N are such that f i converges to f weakly in
L p (Ω; Rm ) and g i converges to g weakly in L p (Ω) as i → ∞. If | f i ( x)| É g i ( x) for
almost every x ∈ Ω, then | f ( x)| É g( x) for almost every x ∈ Ω.
Proof. Let x ∈ Ω be a Lebesgue point of g and all the components of f , and fix
´
0 < r < d ( x, ÇΩ). Assume that B( x,r ) f ( y) d y 6= 0. Denote
¯× ¯−1 ×
e=¯ f ( y) d y¯ f ( y) d y ∈ R m
¯ ¯
B( x,r ) B( x,r )
and
0
ψ = |B( x, r )|−1 χB( x,r) e ∈ L p (Ω; Rm ).
By Cauchy–Schwarz’s inequality and the assumptions, we have

¯× ¯ × ˆ
f ( y) d y¯ = e · f ( y) d y = f ( y) · ψ( y) d y
¯ ¯
¯
B( x,r ) B( x,r ) Ω
ˆ ˆ
= lim f i ( y) · ψ( y) d y É lim inf | f i ( y)||ψ( y)| d y
i →∞ Ω i →∞ Ω
ˆ ×
É lim inf g i ( y)|ψ( y)| d y = g( y) d y.
i →∞ Ω B( x,r )
This implies
¯× ¯ ×
f ( y) d y¯ É g( y) d y,
¯ ¯
¯
B( x,r ) B( x,r )
´
which clearly holds also if B( x,r ) f ( y) d y = 0. Since almost every point x ∈ Ω is a
Lebesgue point of g and all components of f and the claim follows by taking r → 0
on both sides of the previous estimate. ä
1.13 Difference quotients

In this section we give a characterization of W 1,p , 1 < p < ∞, in terms of difference
quotients. This approach is useful in regularilty theory for PDEs. Moreover, this
characterization does not involve derivatives.
Definition 1.45. Let u ∈ L1loc (Ω) and Ω0 b Ω. The j th difference quotient is
u( x + he j ) − u( x)
D hj u( x) = , j = 1, . . . , n,
h
for x ∈ Ω0 and h ∈ R such that 0 < | h| < dist(Ω0 , ÇΩ). We denote
D h u = (D 1h u, . . . , D nh u).
THE MORAL : Note that the definition of the difference quotient makes sense
at every x ∈ Ω whenever 0 < | h| < dist( x, ÇΩ). If Ω = Rn , then the definition makes
sense for every h 6= 0.
Theorem 1.46.
(1) Assume u ∈ W 1,p (Ω), 1 É p < ∞. Then for every Ω0 b Ω, we have
kD h ukL p (Ω0 ) É ckDukL p (Ω)
for some constant c = c( n, p) and all 0 < | h| < dist(Ω0 , ÇΩ).

(2) If u ∈ L p (Ω0 ), 1 < p < ∞, and there is a constant c such that
k D h u k L p (Ω 0 ) É c
whenever 0 < | h| < dist(Ω0 , ÇΩ), then u ∈ W 1,p (Ω0 ) and kDukL p (Ω0 ) É c.
(3) Let 1 < p < ∞, and assume that u ∈ L p (Rn ) and that there exists a constant
C such that
kD h ukL p (Rn ;Rn ) É c
for every h 6= 0. Then the weak derivative Du with respect to Rn exists,

u ∈ W 1,p (Rn ) and kDukL p (Rn ;Rn ) É c.
THE MORAL : Pointwise derivatives are defined as limit of difference quotients

and Sobolev spaces can be characterized by integrated difference quotients.
W A R N I N G : Claim (2) does not hold for p = 1 (exercise).
Proof. (1) First assume that u ∈ C ∞ (Ω) ∩ W 1,p (Ω). Then

ˆ h
Ç
u( x + he j ) − u( x) = ( u( x + te j )) dt
0 Çt
ˆ h
= Du( x + te j ) · e j dt
0
ˆ h
Çu
= ( x + te j ) dt, j = 1, . . . , n,
0 Çx j
for all x ∈ Ω0 , 0 < | h| < dist(Ω0 , ÇΩ). By Hölder’s inequality

¯ u( x + he j ) − u( x) ¯
¯ ¯
|D hj u( x)| = ¯¯ ¯
h ¯
ˆ | h| ¯
¯ Çu
¯
1 ¯
É ¯ ( x + te j )¯¯ dt
| h| −|h| ¯ Ç x j
Ãˆ ¯ p !1/ p
| h| ¯¯
1 Çu ¯ 1− 1
É ¯ ( x + te j )¯¯ dt |2 h | p ,
| h| −|h| ¯ Ç x j
which implies
ˆ | h| ¯¯ ¯p
2 p−1 ¯ Ç u ( x + te j )¯ dt
|D hj u( x)| p É
¯
| h| ¯ Çx ¯
−| h| j
Next we integrate over Ω and switch the order of integration by Fubini’s theorem
0
to conclude
ˆ ˆ ˆ | h| ¯¯¯p
2 p−1 ¯ Ç u ( x + te j )¯ dt dx
|D hj u( x)| p dx É
¯
| h| ¯ Çx
Ω0 Ω −| h|
0
¯
j
ˆ ˆ
2 p−1 |h|
¯p
¯ Çu
¯
¯
= ¯ ( x + te j ¯ dx dt
)¯
| h| −|h| Ω0 Ç x j
¯
ˆ ¯ ¯p
¯ Çu
É 2p
¯
¯
¯ Çx ( x ) ¯ dx.
Ω
¯
j
The last inequality follows from the fact that, for 0 < | h| < dist(Ω0 , ÇΩ) and | t| É | h|,
we have ˆ ¯ ¯p ˆ ¯p
¯ Çu ¯ Çu
¯
¯ ¯
¯ ( x + te )
j ¯
¯ dx É ¯ ( x ) ¯ dx.
0 Çx
¯ Çx
Ω Ω
¯ ¯
j j
Using the elementary inequality (a 1 + · · · + a n )α É nα (aα1 + · · · + aαn ), a i Ê 0, α > 0,
we obtain
ˆ ˆ Ã n
!p
2
ˆ n
p
h p
|D hj u( x)|2 |D hj u( x)| p dx
X X
|D u( x)| dx = dx É n 2
Ω0 Ω0 j =1 Ω0 j =1
n
ˆ n
ˆ ¯ ¯p
p p ¯ Çu
|D hj u( x)| p dx É 2 p n
X X ¯
=n 2 2 ¯
¯ Çx ( x )¯ dx
j =1 Ω0 j =1 Ω
¯
j
ˆ
p
É 2p n 1+ 2
|Du( x)| p dx
Ω
The general case u ∈ W 1,p

(Ω) follows by an approximation, see Theorem 1.18.
Let u i ∈ C ∞ (Ω) ∩ W 1,p (Ω), i ∈ N, such that u i → u in W 1,p (Ω) as i → ∞. By
passing to a subsequence, if necessary, we may also assume that u i → u pointwise
almost everywhere in Ω as i → ∞. Assume that 0 < | h| < dist(Ω0 , ÇΩ). Then
D h u i ( x) → D h u( x) for almost every x ∈ Ω0 as i → ∞. By Fatou’s lemma and
assumption we obtain
ˆ ˆ
|D h u( x)| p dx É lim inf |D h u i ( x)| p dx
Ω0 i →∞ Ω0
ˆ
É C ( n) lim inf |Du i ( x)| p dx
i →∞ Ω
ˆ
= C ( n) |Du( x)| p dx.
Ω
(2) Let ϕ ∈ C 0∞ (Ω0 ). Then by a change of variables we see that, for 0 < | h| <
dist(supp ϕ, ÇΩ0 ), we have
ˆ ˆ
ϕ( x + he j ) − ϕ( x) u( x − he j ) − u( x)
u ( x) dx = − ϕ( x) dx, j = 1, . . . , n.
Ω0 h Ω0 −h
This shows that
ˆ ˆ
uD hj ϕ dx = − (D −j h u)ϕ dx, j = 1, . . . , n.
Ω0 Ω
By assumption
sup kD −j h ukL p (Ω0 ) É c < ∞.
0<| h|<dist(Ω0 ,ÇΩ)
Since 1 < p < ∞, by Theorem 1.30 there exists f ∈ L p (Ω0 ; Rn ) and a sequence
( h i ) i∈N converging to zero such that D −h i u → f weakly in L p (Ω0 ; Rn ) as i → ∞.
This implies
ˆ ˆ ˆ
Çϕ
µ ¶
h h
u dx = u lim D i ϕ dx = lim uD j i ϕ dx
Ω0 Çx j Ω0 h →0 j
i h i →0 Ω 0
ˆ ˆ
−h i
= − lim (D j u)ϕ dx = − f j ϕ dx
h i →0 Ω 0 Ω0
for every ϕ ∈ C 0∞ (Ω0 ). Here the second equality follows from the dominated conver-
gence theorem and the last equality is weak convergence the weak convergence
tested with g = (0, . . . , ϕ, . . . , 0), where ϕ is in the j th position. It follows that
Du = f in the weak sense in Ω0 and thus u ∈ W 1,p (Ω0 ). By (1.3),
kDukL p (Ω0 ;Rn ) = k f kL p (Ω0 ;Rn ) É lim infkD −h i ukL p (Ω0 ;Rn ) É c.
i →∞
(3) Let Ω i = B(0, 2 i ) and Ω0i = B(0, i ) for every i ∈ N. Assertion (2) and the
assumption imply that u i = u|Ω i , i ∈ N, has a weak derivative Du i in Ω0i and
kDu i kL p (Ω0 ;Rn ) É c. Since Du i+1 = Du i almost everywhere in Ω0i , we see that the
i
limit
f ( x) = lim χΩ0 ( x)Du i ( x)
i →∞ i
exists for almost every x ∈ R . The weak derivative of u with respect to Rn

n
coincides with f ∈ L1loc (Rn ; Rn ) and Fatou’s lemma implies
³ˆ ´ 1p
kDukL p (Rn ;Rn ) = k f kL p (Rn ;Rn ) = lim |χΩ0 Du i | p dx
i
Rn i →∞
³ˆ ´ 1p
É lim inf |Du i | p dx É c.
i →∞ Ω0i
From this it also follows that u ∈ W 1,p (Rn ). ä
1.14 Absolute continuity on lines

In this section we relate weak derivatives to classical derivatives and give a
characterization W 1,p in terms of absolute continuity on lines.
Recall that a function u : [a, b] → R is absolutely continuous, if for every ε > 0,
there exists δ > 0 such that if a = x1 < y1 É x2 < y2 É . . . É xm < ym = b is a partition
of [a, b] with
m
X
( yi − x i ) < δ,
i =1
then
m
X
| u( yi ) − u( x i )| < ε.
i =1
Absolute continuity can be characterized in terms of the fundamental theorem of

calculus.
Theorem 1.47. A function u : [a, b] → R is absolutely continuous if and only if

there exists a function g ∈ L1 ((a, b)) such that
ˆ x
u ( x) = u ( a) + g( t) dt.
a
By the Lebesgue differentiation theorem g = u0 almost everywhere in (a, b).

T HE MORAL : Absolutely continuous functions are precisely those functions

for which the fundamental theorem of calculus holds true.
Examples 1.48:
(1) Every Lipchitz continuous function u : [a, b] → R is absolutely continuous.
(2) The Cantor function u is continuous in [0, 1] and differentiable almost
everywhere in (0, 1), but not absolutely continuous in [0, 1].
Reason.
ˆ 1
u(1) = 1 6= 0 = u(0) + u0 ( t) dt.
0
| {z }
=0 ■
The next result relates weak partial derivatives with the classical partial
derivatives.
1,p
Theorem 1.49 (Nikodym, ACL characterization). Assume that u ∈ Wloc (Ω),
1 É p É ∞ and let Ω0 b Ω. Then there exists u∗ : Ω → [−∞, ∞] such that u∗ = u
almost everywhere in Ω and u∗ is absolutely continuous on ( n − 1)-dimensional
Lebesgue measure almost every line segments in Ω0 that are parallel to the
coordinate axes and the classical partial derivatives of u∗ coincide with the weak
p
partial derivatives of u almost everywhere in Ω. Conversely, if u ∈ L loc (Ω) and
p 1,p
there exists u∗ as above such that D i u∗ ∈ L loc (Ω), i = 1, . . . , n, then u ∈ Wloc (Ω).
T HE MORAL : This is a very useful characterization of W 1,p , since many

claims for weak derivatives can be reduced to the one-dimensional claims for
absolute continuous functions. In addition, this gives a practical tool to show that
a function belongs to a Sobolev space.
Remarks 1.50:
(1) The ACL characterization can be used to give a simple proof of Example
1.9 (exercise).
(2) In the one-dimensional case we obtain the following characterization:
u ∈ W 1,p ((a, b)), 1 É p É ∞, if u can be redefined on a set of measure zero
in such a way that u ∈ L p ((a, b)) and u is absolutely continuous on every
compact subinterval of (a, b) and the classical derivative exists and belongs
to u ∈ L p ((a, b)). Moreover, the classical derivative equals to the weak
derivative almost everywhere.
(3) A function u ∈ W 1,p (Ω) has a representative that has classical partial
derivatives almost everywhere. However, this does not give any informa-
tion concerning the total differentiability of the function. See Theorem
2.21.
(4) The ACL characterization can be used to give a simple proof of the Leibniz
rule. If u ∈ W 1,p (Ω) ∩ L∞ (Ω) and v ∈ W 1,p (Ω) ∩ L∞ (Ω), then uv ∈ W 1,p (Ω)
and
D j ( uv) = vD j u + uD j v, j = 1, . . . , n,
almost everywhere in Ω (exercise), compare to Lemma 1.12 (5).

(5) The ACL characterization can be used to give a simple proof for Lemma
1.25 and Theorem 1.26. The claim that if u, v ∈ W 1,p (Ω), then max{ u, v} ∈
W 1,p (Ω) and min{ u, v} ∈ W 1,p (Ω) follows also in a similar way (exercise).
(6) The ACL characterization can be used to show that if Ω is connected,
1,p
u ∈ Wloc (Ω) and Du = 0 almost everywhere in Ω, then u is a constant
almost everywhere in Ω (exercise).
Proof. Since the claims are local, we may assume that Ω = Rn and that u has a
compact support.
=⇒ Let u i = u ε i , i = 1, 2, . . . , be a sequence of standard convolution approxima-
tions of u such that supp u i ⊂ B(0, R ) for every i = 1, 2, . . . and
1
k u i − ukW 1,1 (Rn ) < , i = 1, 2, . . .
2i
By Lemma 1.16 (2), the sequence of convolution approximations converges point-
wise almost everywhere and thus the limit lim i→∞ u i ( x) exists for every x ∈ Rn \ E
for some E ⊂ Rn with |E | = 0. We define

 lim u i ( x), x ∈ Rn \ E,
∗
u ( x) = i→∞
0, x ∈ E.

We fix a standard base vector in Rn and, without loss of generality, we may assume
that it is (0, . . . , 0, 1). Let
ˆ Ã n ¯¯ Ç u ¯!
X i +1 Ç u i ¯¯
f i ( x1 , . . . , xn−1 ) = | u i+1 − u i | + ¯ − ( x1 , . . . , xn ) dxn
R j =1 Ç x j Çx j ¯
¯
and ∞
X
f ( x1 , . . . , xn−1 ) = f i ( x1 , . . . , xn−1 ).
i =1
By the monotone convergence theorem and Fubini’s theorem
ˆ ˆ ∞
X
f dx1 . . . dxn−1 = f i dx1 . . . dxn−1
Rn−1 Rn−1 i =1
∞
ˆ
X
= f i dx1 . . . dxn−1
i =1 Rn−1
ˆ µ
∞ ¯ Ç u i+1 Ç u i ¯
¯ ¯¶
X
= | u i+1 − u i | + ¯
¯ − ¯ dx
i =1 Rn Çx j Çx j ¯
X∞ 1
< < ∞.
i =1 2i
This shows that f ∈ L1 (Rn−1 ) and thus f < ∞ ( n − 1)-almost everywhere in Rn−1 .
x = ( x1 , . . . , xn−1 ) ∈ Rn−1 such that f ( b
Let b x) < ∞. Denote
g i ( t) = u i ( b
x, t) and g ( t) = u ∗ ( b
x, t).
Claim: ( g i ) is a Cauchy sequence in C (R).
Reason. Note that

iX
−1
g i = g1 + ( g k+1 − g k ), i = 1, 2, . . . ,
k=1
where
¯ˆ t ¯
( g0k+1 − g0k )( s) ds¯¯
¯ ¯
| g k+1 ( t) − g k ( t)| = ¯¯
ˆ −∞
É | g0k+1 ( s) − g0k ( s)| ds

R
ˆ ¯
¯ Ç u k+1 Çu k
¯
¯
É ¯¯ (bx, s) − x, s)¯¯ ds É f k ( b
(b x).
R Ç xn Ç xn
Thus
¯ ¯
¯ iX
−1 ¯ iX−1
¯ ( g k+1 ( t) − g k ( t))¯ É |g ( t) − g k ( t)|
¯ ¯
¯k=1 ¯ k=1 k+1
∞
X
É f k (b
x) = f ( b
x) < ∞
k=1
for every t ∈ R. This implies that ( g i ) is a Cauchy sequence in C (R). Since C (R) is
complete, there exists g ∈ C (R) such that g i → g uniformly in R. It follows that
x} × R ⊂ R n \ E .
{b ■
Claim: ( g0i ) is a Cauchy sequence in L1 (R).
Reason.
ˆ ¯¯ i−1 ¯
¯ iX
−1
ˆ
¯X
¯ ( g k+1 ( t) − g k ( t))¯ dt É | g k+1 ( t) − g k ( t)| dt
¯
R ¯ k=1 ¯ k=1 R
iX
−1
É f k (b
x) É f ( b
x) < ∞.
k=1
This implies that ( g i ) is a Cauchy sequence in L1 (R). Since L1 (R) is complete,

there exists ge ∈ L1 (R) such that g0i → ge in L1 (R) as i → ∞ . ■
Claim: g is absolutely continuous in R.

Reason.
ˆ t ˆ t
g( t) = lim g i ( t) = lim g0i ( s) ds = ge0 ( s) ds
i →∞ i →∞ −∞ −∞
This implies that g is absolutely continuous in R and g0 = ge almost everywhere in

R. ■
Claim: ge is the weak derivative of g.
Reason. Let ϕ ∈ C 0∞ (R). Then

ˆ ˆ ˆ ˆ
0 0
gϕ dt = lim g i ϕ dt = − lim g0i ϕ dt = − geϕ dt. ■
R i →∞ R i →∞ R R
Thus for every ϕ ∈ C 0∞ (Rn ) we have

ˆ ˆ
Çϕ Ç u∗
u∗ ( b
x, x n ) x, xn ) dxn = −
(b x, xn )ϕ( b
(b x, xn ) dxn
R Ç xn R Ç xn
and by Fubini’s theorem
ˆ ˆ
Çϕ Ç u∗
u dx = − ϕ dx.
Rn Ç xn Rn Ç xn
This shows that u∗ has the classical partial derivatives almost everywhere in Rn
and that they coincide with the weak partial derivatives of u almost everywhere
in Rn .
⇐= Assume that u has a representative u∗ as in the statement of the theo-
rem. For every ϕ ∈ C 0∞ (Rn ), the function u∗ ϕ has the same absolute continuity
properties as u∗ . By the fundamental theorem of calculus
ˆ
Ç( u∗ ϕ)
x, t) dt = 0
(b
R Ç xn
x ∈ Rn−1 . Thus
for ( n − 1)-almost every b
ˆ ˆ
Çϕ Ç u∗
u∗ ( b
x, t) x, t) dt = −
(b x, t)ϕ( b
(b x, t) dt
R Ç xn R Ç xn
and by Fubini’s theorem
ˆ ˆ
Çϕ Ç u∗
u∗ dx = − ϕ dx.
Rn Ç xn Rn Ç xn
Ç u∗
Since u∗ = u almost everywhere in Rn , we see that Ç xn
is the nth weak partial
Ç u∗
derivative of u. The same argument applies to all other partial derivatives Çx j
,
j = 1, . . . , n as well. ä
x
Example 1.51. The radial projection u : B(0, 1) → ÇB(0, 1), u( x) = | x| is discontinu-
xj
ous at the origin. However, the coordinate functions | x| , j = 1, . . . , n, are absolutely
continuous on almost every lines. Moreover,
xi x j
µ
xj
¶ δ i j | x| − | x|
Di = ∈ L p (B(0, 1))
| x| | x |2
whenever 1 É p < n. Here 
1, i = j,
δi j =
0, i 6= j,
is the Kronecker symbol. By the ACL characterization the coordinate functions of
u belong to W 1,p (B(0, 1)) whenever 1 É p < n.
Remark 1.52. We say that a closed set E ⊂ Ω to be removable for W 1,p (Ω), if
|E | = 0 and W 1,p (Ω \ E ) = W 1,p (Ω) in the sense that every function in W 1,p (Ω \ E )
can be approximated by the restrictions of functions in C ∞ (Ω). Theorem 1.49
implies the following removability theorem for W 1,p Ω): if H n−1 (E ) = 0, then E
is removable for W 1,p (Ω). Observe, that if H n−1 (E ) = 0, then E is contained in a
measure zero set of lines in a fixed direction (equivalently the projection of E onto
a hyperplane also has H n−1 -measure zero).
This result is quite sharp. For example, let Ω = B(0, 1) and E = { x ∈ B(0, 1) :
x2 = 0}. Then 0 < H n−1 (E ) < ∞, but E is not removable since, using Theorem 1.49
again, it is easy to see that the function which is 1 on the upper half-plane and 0
on the lower half-plane does not belong to W 1,p Ω). With a little more work we can
show that E 0 = E ∩ B(0, 12 ) is not removable for W 1,p (B(0, 1)).
Sobolev inequalities
2
The term Sobolev inequalities refers to a variety of inequalities involving functions
and their derivatives. As an example, we consider an inequality of the form
µˆ ¶1 µˆ ¶1
q p
| u| q dx Éc |Du| p dx (2.1)
Rn Rn
for every u ∈ C 0∞ (Rn ), where constant 0 < c < ∞ and exponent 1 É q < ∞ are
independent of u. By density of smooth functions in Sobolev spaces, see Theorem
1.18, we may conclude that (2.1) holds for functions in W 1,p (Rn ) as well. Let u ∈
C 0∞ (Rn ), u 6≡ 0, 1 É p < n and consider u λ ( x) = u(λ x) with λ > 0. Since u ∈ C 0∞ (Rn ),
it follows that (2.1) holds true for every u λ with λ > 0 with c and q independent of
λ. Thus
µˆ ¶1 µˆ ¶1
q p
q p
| u λ | dx Éc |Du λ | dx
Rn Rn
1
for every λ > 0. By a change of variables y = λ x, dx = λn
d y, we see that
ˆ ˆ ˆ ˆ
q q 1 1 q
| u λ ( x)| dx = | u(λ x)| dx = | u( y)| n d y = n | u( x)| q dx
R n R n R n λ λ Rn
and
ˆ ˆ
|Du λ ( x)| p dx = λ p |Du(λ x)| p dx
Rn Rn
ˆ
λp
= |Du( y)| p d y
λn Rn
ˆ
λp
= |Du( x)| p dx.
λn Rn
Thus
µˆ ¶1 µˆ ¶1
1 q λ p
n | u| q dx Éc n |Du| p dx
λ q Rn λ p Rn
42
CHAPTER 2. SOBOLEV INEQUALITIES 43
for every λ > 0, and equivalently,
1− np + nq
k ukL q (Rn ) É cλ kDukL p (Rn ) .
Since this inequality has to be independent of λ, we have

n n np
1− + =0 ⇐⇒ q= .
p q n− p
THE MORAL : There is only one possible exponent q for which inequality (2.1)
may hold true for all compactly supported smooth functions.
For 1 É p < n, the Sobolev conjugate exponent of p is

np
p∗ = .
n− p
Observe that
(1) p∗ > p,
(2) If p → n−, then p∗ → ∞ and
n
(3) If p = 1, then p∗ = n−1 .
2.1 Gagliardo-Nirenberg-Sobolev inequal-

ity
The following generalized Hölder’s inequality will be useful for us.
1 1
Lemma 2.1. Let 1 É p 1 , . . . , p k É ∞ with p1 +···+ pk = 1 and assume f i ∈ L p i (Ω),
i = 1, . . . , k. Then
ˆ k
Y
| f 1 . . . f k | dx É k f i k L p i (Ω ) .
Ω i =1
Proof. Induction and Hölder’s inequality (exercise). ä
Sobolev proved the following theorem in the case p > 1 and Nirenberg and
Gagliardo in the case p = 1.
Theorem 2.2 (Gagliardo-Nirenberg-Sobolev). Let 1 É p < n. There exists

c = c( n, p) such that
µˆ ¶ 1 µˆ ¶1
∗ p∗ p
| u| p dx Éc |Du| p dx
Rn Rn
for every u ∈ W 1,p (Rn ).

T H E M O R A L : The Sobolev-Gagliardo-Nirenberg inequality implies that

∗
W 1,p (Rn ) ⊂ L p (Rn ), when 1 É p < n. More precisely, W 1,p (Rn ) is continuously
∗
imbedded in L p (Rn ), when 1 É p < n. This is the Sobolev embedding theorem for
1 É p < n.
Proof. (1) We start by proving the estimate for u ∈ C 0∞ (Rn ). By the fundamental
theorem of calculus
ˆ xj
Çu
u ( x1 , . . . , x j , . . . , x n ) = ( x1 , . . . , t j , . . . , xn ) dt j , j = 1, . . . , n.
−∞ Çx j
This implies that

ˆ
| u( x)| É |Du( x1 , . . . , t j , . . . , xn )| dt j , j = 1, . . . , n.
R
By taking product of the previous estimate for each j = 1, . . . , n, we obtain

n µˆ ¶ 1
n Y n−1
| u( x)| n−1 É |Du( x1 , . . . , t j , . . . , xn )| dt j .
j =1 R
We integrate with respect to x1 and then we use generalized Hölder’s inequality

for the product of ( n − 1) terms to obtain
ˆ µˆ ¶ 1 ˆ n µˆ ¶ 1
n n−1 Y n−1
| u| n−1 dx1 É |Du| dt 1 |Du| dt j dx1
R R R j =2 R
µˆ ¶ 1
n µˆ ˆ ¶ 1
n−1 Y n−1
É |Du| dt 1 |Du| dx1 dt j .
R j =2 R R
Next we integrate with respect to x2 and use again generalized Hölder’s inequality
ˆ ˆ ˆ "µˆ ¶ 1 n µˆ ˆ ¶ 1 #
n n−1 Y n−1
| u| n − 1 dx1 dx2 É |Du| dt 1 |Du| dx1 dt j dx2
R R R R j =2 R R
µˆ ˆ ¶ 1
n−1
= |Du| dx1 dt 2
R R
ˆ "µˆ ¶ 1
n µˆ ˆ ¶ 1 #
n−1 Y n−1
· |Du| dt 1 |Du| dx1 dt j dx2
R R j =3 R R
µˆ ˆ ¶ 1
n−1
É |Du| dx1 dt 2
R R
µˆ ˆ ¶ 1
n µˆ ˆ ˆ ¶ 1
n−1 Y n−1
· |Du| dt 1 dx2 |Du| dx1 dx2 dt j .
R R j =3 R R R
Then we integrate with respect to x3 , . . . , xn and obtain

ˆ n µˆ ˆ ¶ 1
n Y n−1
| u| n−1 dx É ... |Du| dx1 . . . dt j . . . dxn
Rn j =1 R R
µˆ ¶ n
n−1
= |Du| dx .
Rn
This is the required inequality for p = 1.

If 1 1 is to be chosen later. Since γ > 1, we have v ∈ C 1 (Rn ). Hölder’s

inequality implies
µˆ ¶ n−1 ˆ
n n
| u|γ n−1 dx É |D (| u|γ )| dx
Rn Rn
ˆ
=γ | u|γ−1 |Du| dx
Rn
µˆ ¶ p−1 µˆ ¶1
p p p
(γ−1) p−1 p
Éγ | u| dx |Du| dx .
Rn Rn
Now we choose γ so that | u| has the same power on both sides. Thus
γn p p( n − 1)
= (γ − 1) ⇐⇒ γ= .
n−1 p−1 n− p
This gives
γn p( n − 1) n pn
= = = p∗
n−1 n− p n−1 n− p
and consequently
µˆ ¶ 1 µˆ ¶1
∗ p∗ p
| u| p dx Éγ |Du| p dx .
Rn Rn
This proves the claim for u ∈ C 0∞ (Rn ).

(2) Assume then that u ∈ W 1,p (Rn ). By Lemma 1.24 we have W 1,p (Rn ) =
1,p
W0 (Rn ). Thus there exist u i ∈ C 0∞ (Rn ), i = 1, 2, . . . , such that k u i − ukW 1,p (Rn ) →
0 as i → ∞. In particular k u i − ukL p (Rn ) → 0, as i → ∞. Thus there exists a
subsequence ( u i ) such that u i → u almost everywhere in Rn and u i → u in L p (Rn ).
∗
Claim: ( u i ) is a Cauchy sequence in L p (Rn ).
Reason. Since u i − u j ∈ C 0∞ (Rn ), we use the Sobolev-Gagliardo-Nirenberg inequal-

ity for compactly supported smooth functions and Minkowski’s inequality to
conclude that
k u i − u j kL p∗ (Rn ) É ckDu i − Du j kL p (Rn )

¡ ¢
É c kDu i − DukL p (Rn ) + kDu − Du j kL p (Rn ) → 0. ■
∗ ∗ ∗
Since L p (Rn ) is complete, there exists v ∈ L p (Rn ) such that u i → v in L p (Rn ) as
i → ∞.
∗
Since u i → u almost everywhere in Rn and u i → v in L p (Rn ), we have u = v
∗ ∗
almost everywhere in Rn . This implies that u i → u in L p (Rn ) and that u ∈ L p (Rn ).
Now we can apply Minkowski’s inequality and the Sobolev-Gagliardo-Nirenberg

inequality for compactly supported smooth functions to conclude that
k ukL p∗ (Rn ) É k u − u i kL p∗ (Rn ) + k u i kL p∗ (Rn )

É k u − u i kL p∗ (Rn ) + ckDu i kL p (Rn )
¡ ¢
É k u − u i kL p∗ (Rn ) + c kDu i − DukL p (Rn ) + kDukL p (Rn )
→ ckDukL p (Rn ) ,
∗
since u i → u in L p (Rn ) and Du i → Du in L p (Rn ). This completes the proof. ä
Remarks 2.3:
(1) The Gagliardo-Nirenberg-Sobolev inequality shows that if u ∈ W 1,p (Rn )with
∗
1 É p < n, then u ∈ L p (Rn ) ∩ L p (Rn ), with p∗ > p.
(2) The Gagliardo-Nirenberg-Sobolev inequality shows that if u ∈ W 1,p (Rn )
with 1 É p < n and Du = 0 almost everywhere in Rn , then u = 0 almost
everywhere in Rn .
(3) The Sobolev-Gagliardo-Nirenberg inequality holds for Sobolev spaces with
zero boundary values in open subsets of Rn by considering the zero exten-
sions. There exists c = c( n, p) > 0 such that
µˆ ¶ 1 µˆ ¶1
p∗ p
p∗ p
| u| dx Éc |Du| dx
Ω Ω
1,p
for every u ∈ W0 (Ω), 1 É p < n. If |Ω| < ∞, by Hölder’s inequality
µˆ ¶1 µˆ ¶ 1
q p∗ 1− p1∗
q p∗
| u| dx É | u| dx |Ω |
Ω Ω
µˆ ¶1
1− p1∗ p
É c |Ω | |Du| p dx
Ω
∗
whenever 1 É q É p . Thus for sets with finite measure all exponents below
the Sobolev exponent will do.
1,p p∗
(4) The Sobolev-Gagliardo-Nirenberg inequality shows that Wloc (Rn ) ⊂ L loc (Rn ).
To see this, let Ω b Rn and choose a cutoff function η ∈ C 0∞ (Rn ) such that
1,p
η = 1 in Ω. Then η u ∈ W0 (Rn ) = W 1,p (Rn ) and η u = u in Ω and
k ukL p∗ (Ω) É kη ukL p∗ (Rn ) É ckD (η u)kL p (Rn ) < ∞.
(5) The Sobolev-Gagliardo-Nirenberg inequality holds for higher order Sobolev

n np
spaces as well. Let k ∈ N, 1 É p < k and p∗ = n− k p . There exists c =
c( n, p, k) such that
µˆ ¶ 1 µˆ ¶1
p∗ p
p∗ k p
| u| dx Éc |D u| dx
Rn Rn
for every u ∈ W k,p (Ω). Here |D k u|2 is the sum of squares of all kth order
partial derivatives of u (exercise).
The Sobolev–Gagliardo–Nirenberg inequality has the following consequences.
Corollary 2.4. Let 1 É p < n and let Ω ⊂ Rn be an open set. Assume that
1,p
u ∈ W0 (Ω) is such that |Du| = 0 almost everywhere in Ω. Then u = 0 almost
everywhere in Ω.
Proof. Extend u as zero outside Ω. Then then have |Du| = 0 almost everywhere
in Rn . Theorem 2.2 implies
k ukL p∗ (Rn ) É ckDukL p (Rn ) = 0.
It follows that u = 0 almost everywhere in Rn , and thus almost everywhere in Ω.ä

np
Since p∗ = n− p → ∞ as p → n−, one might expect that W 1,n (Ω) would be
continuously embedded in L∞ (Ω). This is false for n > 1. Let Ω = B(0, 1) ⊂ Rn . The
function
1
µ µ ¶¶
u( x) = log log 1 +
| x|
belongs to W 1,n (Ω) but not L∞ (Ω) (exercise).
The following result is a version of the Sobolev inequality for the full range
1 É p < ∞.
Corollary 2.5. Let 1 É p < ∞, let Ω ⊂ Rn be an open set with |Ω| < ∞, and assume
1,p np
that u ∈ W0 (Ω). Let 1 É q É p∗ = n− p , for 1 É p < n, and 1 É q < ∞ for n É p < ∞.
There exists a constant c = c( n, p, q) such that
³ˆ ´ 1q 1 1 1 ³ˆ ´ 1p
q
| u| dx É c | Ω| n− p+q |Du| p dx .
Ω Ω
1,p
T HE MORAL : Let Ω ⊂ Rn be an open set with |Ω| < ∞. If u ∈ W0 (Ω) with
p Ê n, then u ∈ L q (Ω) for every q with 1 É q < ∞.
Proof. Extend u as zero outside Ω. Then Du( x) = 0 for almost every x ∈ Ω c .

Assume first that 1 É p < n. Hölder’s inequality and Theorem 2.2 imply
³ˆ ´ 1q 1 ³ˆ np ´ nnp
−p
− 1p + 1q
| u| q dx É |Ω| n | u| n− p dx
Ω Ω
1 1 1 ³ˆ ´ 1p
É c( n, p)|Ω| n− p+q |Du| p dx .
Ω
n pe
Assume then that n É p < ∞. If q > p, choose 1 < pe < n satisfying q = n− pe . By
the first part of the proof and Hölder’s inequality, we obtain
³ˆ ´ 1q 1 ³ˆ ´ 1pe
− 1pe + 1q
| u| q dx É c( n, p, q)|Ω| n |Du| pe dx
Ω Ω
1 1 1 ³ˆ ´ 1p
É c( n, p, q)|Ω| n− p+q |Du| p dx .
Ω
Finally, if q É p, the claim follows from the previous case for some qe > q and
Hölder’s inequality on the left-hand side. ä
Remark 2.6. Let 1 É p < n and let Ω ⊂ Rn be an open set with |Ω| < ∞. The proof
of Corollary 2.5 shows that the Sobolev inequality
³ˆ np ´ nnp
−p ³ˆ ´ 1p
| u( x)| n− p dx É c( n, p) |Du( x)| p dx
Ω Ω
1,p
holds for every u ∈ W0 (Ω).
Remark 2.7. When p = 1 the Sobolev-Gagliardo-Nirenberg inequality is related

to the isoperimetric inequality. Let Ω ⊂ Rn be a bounded domain with smooth
boundary and set



1, x ∈ Ω ,

u ε ( x) = 1 − dist(εx,Ω) , 0 < dist( x, Ω) < ε,


0, dist( x, Ω) Ê ε.

Note that u can been considered as an approximation of the characteristic function

of Ω. The Lipschitz constant of x 7→ dist( x, Ω) is one so that the Lipschitz constant
of u ε is ε−1 and thus this function belongs to W 1,1 (Rn ), for example, by the ACL
characterization, see Theorem 1.49, we have

 1 , 0 < dist( x, Ω) < ε,
|Du ε ( x)| É ε
0, otherwise.
The Sobolev-Gagliardo-Nirenberg inequality with p = 1 gives

µˆ ¶ n−1 µˆ ¶ n−1
n−1 n n n n
|Ω| n = |uε | n−1 dx É |uε | n−1 dx
ˆΩ ˆ Rn
1
Éc |Du ε | dx É c dx
Rn {0<dist( x,Ω)<ε} ε
|{ x ∈ Rn : 0 < dist( x, Ω) < ε}|
=c → cH n−1 (ÇΩ)
ε
This implies
n
|Ω| n−1 É cH n−1 (ÇΩ),
which is an isoperimetric inequality with the same constant c as in the Sobolev-

Gagliardo-Nirenberg inequality. According to the classical isoperimetric inequality,
if Ω ⊂ Rn is a bounded domain with smooth boundary, then
n−1 −1
|Ω | n É n−1 Ωn n H n−1 (ÇΩ),
where H n−1 (ÇΩ) stands for the ( n − 1)-dimensional Hausdorff measure of the
boundary ÇΩ. The isoperimetric inequality is equivalent with the statement that
among all smooth bounded domains with fixed volume, balls have the least surface
area.
Conversely, the Sobolev-Gagliardo-Nirenberg inequality can be proved by the

isoperimetric inequality, but we shall not consider this argument here. From
these considerations it is relatively obvious that the best constant in the Sobolev-
−1
Gagliardo-Nirenberg when p = 1 should be the isoperimetric constant n−1 Ωn n .
This also gives a geometric motivation for the Sobolev exponent in the case p = 1.
2.2 Sobolev-Poincaré inequalities

We begin with a Poincaré inequality for Sobolev functions with zero boundary
values in open subsets.
Theorem 2.8 (Poincaré). Assume that Ω ⊂ Rn is bounded and 1 É p < ∞. Then

there is a constant c = c( p) such that
ˆ ˆ
| u| p dx É c diam(Ω) p |Du| p dx
Ω Ω
1,p
for every u ∈ W0 (Ω).
1,p
T HE MORAL : This shows that W0 (Ω) ⊂ L p (Ω) when 1 É p < ∞, if Ω ⊂ Rn
is bounded. The main difference compared to the Gagliardo-Nirenberg-Sobolev
inequality is that this applies for the whole range 1 É p < ∞ without the Sobolev
exponent.
Remark 2.9. The Poincaré inequality above also shows that if Du = 0 almost
everywhere, then u = 0 almost everywhere. For this it is essential that the
function belongs to the Sobolev space with zero boundary values.
Proof. (1) First assume that u ∈ C 0∞ (Ω). Let y = ( y1 , . . . , yn ) ∈ Ω. Then
n £ n £
Ω⊂ y j − diam(Ω), y j + diam(Ω) =
Y ¤ Y ¤
a j, b j ,
j =1 j =1
where a j = y j − diam(Ω) and b j = y j + diam(Ω), j = 1, . . . , n. As the proof of Theorem

2.2, we obtain
ˆ bj
| u( x)| É |Du( x1 , . . . , t j , . . . , xn )| dt j
aj
Ãˆ !1
bj p
1− 1p p
É (2 diam(Ω)) |Du( x1 , . . . , t j , . . . , xn )| dt j , j = 1, . . . , n.
aj
The second inequality follows from Hölder’s inequality. Thus

ˆ ˆ b1 ˆ bn
| u( x)| p dx = ... | u( x)| p dx1 . . . dxn
Ω a1 an
ˆ b1 ˆ b n ˆ b1
p−1
É (2 diam(Ω)) ... |Du( t 1 , x2 , . . . , xn )| p dt 1 dx1 . . . dxn
a1 an a1
ˆ b1 ˆ bn
É (2 diam(Ω)) p ... |Du( t 1 , x2 , . . . , xn )| p dt 1 . . . dxn
a1 an
ˆ
p
= (2 diam(Ω)) |Du( x)| p dx.
Ω
1,p
(2) The case u ∈ W0 (Ω) follows by approximation (exercise). ä
The Gagliardo-Nirenberg-Sobolev inequality in Theorem 2.2 and Poincaré’s

inequality in Theorem 2.8 do not hold for functions u ∈ W 1,p (Ω), at least when
Ω ⊂ Rn is an open set |Ω| < ∞, since nonzero constant functions give obvious
counterexamples. However, there are several ways to obtain appropriate local
estimates also in this case.
Next we consider estimates in the case when Ω is a cube. Later we consider
similar estimates for balls. The set
Q = [a 1 , b 1 ] × . . . × [a n , b n ], b1 − a1 = . . . = b n − a n
is a cube in Rn . The side length of Q is
l (Q ) = b 1 − a 1 = b j − a j , j = 1, . . . , n,
and n o
Q ( x, l ) = y ∈ Rn : | y j − x j | É 2l , j = 1, . . . , n
is the cube with center x and sidelength l . Clearly,

p
|Q ( x, l )| = l n and diam(Q ( x, l )) = nl
The integral average of f ∈ L1loc (Rn ) over cube Q ( x, l ) is denoted by

ˆ
1
×
f Q ( x,l ) = f dy = f ( y) d y.
Q ( x,l ) |Q ( x, l )| Q ( x,l )
Same notation is used for integral averages over other sets as well.
Theorem 2.10 (Poincaré inequality on cubes). Let Ω be an open subset of

1,p
Rn . Assume that u ∈ Wloc (Ω) with 1 É p < ∞. Then there is c = c( n, p) such that
µ× ¶1 µ× ¶1
p p
| u − uQ ( x,l ) | p d y É cl |Du| p d y
Q ( x,l ) Q ( x,l )
for every cube Q ( x, l ) b Ω.

T HE MORAL : The Poincaré inequality shows that if the gradient is small in

a cube, then the mean oscillation of the function is small in the same cube. In
particular, if the gradient is zero, then the function is constant.
Proof. (1) First assume that u ∈ C ∞ (Ω). Let z, y ∈ Q = Q ( x, l ) = [a 1 , b 1 ] × · · · ×

[a n , b n ]. Then
| u( z) − u( y)| É | u( z) − u( z1 , . . . , z n−1 , yn )| + . . . + | u( z1 , y2 , . . . , yn ) − u( y)|

n
ˆ bj
X
É |Du( z1 , . . . , z j−1 , t, y j+1 , . . . , yn )| dt
j =1 a j
p
By Hölder’s inequality and the elementary inequality (a 1 + · · · + a n ) p É n p (a 1 +
p
· · · + a n ),a i Ê 0, we obtain
Ã ˆ !p
n bj
p
X
| u( z) − u( y)| É |Du( z1 , . . . , z j−1 , t, y j+1 , . . . , yn )| dt
j =1 a j
 Ãˆ !1 p
n bj p
X p 1− 1p 
É |Du( z1 , . . . , z j−1 , t, y j+1 , . . . , yn )| dt (b j − a j )
j =1 aj
n
ˆ bj
p p−1
|Du( z1 , . . . , z j−1 , t, y j+1 , . . . , yn )| p dt.
X
Én l
j =1 a j
By Hölder’s inequality and Fubini’s theorem

ˆ ˆ ¯× ¯p
p
¯ ¯
| u( z) − uQ | dz = ¯ ( u( z) − u( y)) d y¯ dz
¯ ¯
Q Q Q
ˆ µ× ¶p ˆ ×
É | u( z) − u( y)| d y dz É | u( z) − u( y)| p dz d y
Q Q Q Q
ˆ ˆ ˆ bj
n p l p−1 X
n
É |Du( z1 , . . . , z j−1 , t, y j+1 , . . . , yn )| p dt d y dz
|Q | j =1 Q Q aj
ˆ ˆ
n p l p−1 X
n
É (b j − a j ) |Du( z)| p dz dw
|Q | j=1 Q Q
ˆ
p+1 p
Én l |Du( z)| p dz.
Q
1,p
(2) The case u ∈ Wloc (Ω) follows by approximation. There exist u i ∈ C ∞ (Rn ),
i ∈ N, satisfying u i → u in W 1,p (Q ) as i → ∞. By passing to a subsequence,
if necessary, we may in addition assume that u i → u almost everywhere in Q .
Moreover, it follows from Hölder’s inequality and the L p convergence that
× ³× ´ 1p
|( u i )Q − uQ | É | u i ( x) − u( x)| dx É | u i ( x) − u( x)| p dx →0
Q Q
and thus ( u i )Q → uQ as i → ∞. Fatou’s lemma and the first part of the proof for
u i ∈ C ∞ (Ω) give
³× ´ 1p ³× ´ 1p
| u − uQ | p dx É lim inf | u i − ( u i )Q | p dx
Q i →∞ Q
³× ´ 1p
É lim inf c( n, p, q) l |Du i | p dx
i →∞ Q
³× ´ 1p
É c( n, p, q) l |Du| p dx ,
Q
and the proof is complete. ä
Theorem 2.11 (Sobolev–Poincaré inequality on cubes). Let Ω be an open

1,p
subset of Rn . Assume that u ∈ Wloc (Ω) with 1 É p < n. Then there is c = c( n, p)
such that
µ× ¶ 1∗ µ× ¶1
p p
p∗ p
| u − uQ ( x,l ) | dy É cl |Du| d y
Q ( x,l ) Q ( x,2 l )
for every cube Q ( x, 2 l ) b Ω.
1,p p∗
THE MORAL : The Sobolev-Poincaré inequality shows that Wloc (Rn ) ⊂ L loc (Rn ),
when 1 É p < n. This is a stronger version of the Poincaré inequality on cubes in
which we have the Sobolev exponent on the left-hand side.
Proof. Let η ∈ C 0∞ (Rn ) be a cutoff function such that
0 É η É 1, |D η| É cl , supp η ⊂ Q ( x, 2 l ) and η = 1 in Q ( x, l ).
Notice that the constant c = c( n) does not depend on the cube. Then ( u − uQ ( x,l ) )η ∈
W 1,p (Rn ) and by the Gagliardo-Nirenberg-Sobolev inequality, see Theorem 2.2,
and the Leibniz rule, see Theorem 1.12 (5), we have
µˆ ¶ 1∗ µˆ ¶ 1∗
∗ p ∗ p
| u − uQ ( x,l ) | p d y É |( u − uQ ( x,l ) )η| p d y
Q ( x,l ) Rn
µˆ ¶1
p
¯D ( u − uQ ( x,l ) )η ¯ p d y
¯ £ ¤¯
Éc
Rn
µˆ ¶1 µˆ ¶1
p p
Éc η p |Du| p d y + c |D η| p | u − uQ ( x,l ) | p d y
Rn Rn
µˆ ¶1 µˆ ¶1
p
p c p
p
Éc |Du| d y + | u − uQ ( x,l ) | d y .
Q ( x,2 l ) l Q ( x,2 l )
By the Poincaré inequality on cubes, see Theorem 2.10, we obtain

µˆ ¶1
p
| u − uQ ( x,l ) | p d y
Q ( x,2 l )
µˆ ¶ 1 µˆ ¶1
p p
É | u − uQ ( x,2l ) | p d y + | uQ ( x,2l ) − uQ ( x,l ) | p d y
Q ( x,2 l ) Q ( x,2 l )
µˆ ¶1
p 1
É cl |Du| p d y + | uQ ( x,2l ) − uQ ( x,l ) ||Q ( x, 2 l )| p .
Q ( x,2 l )
By Hölder’s inequality and Poincaré inequality on cubes, see Theorem 2.10, we

have
×
1 n
| uQ ( x,2l ) − uQ ( x,l ) ||Q ( x, 2 l )| p É (2 l ) p | u − uQ ( x,2l ) | d y
Q ( x,l )
¶1
|Q ( x, 2 l )|
µ×
n p
p
É (2 l ) p | u − uQ ( x,2l ) | d y
|Q ( x, l )| Q ( x,2l )
µˆ ¶1
p
É cl |Du| p d y .
Q ( x,2 l )
By collecting the estimates above we obtain

µˆ ¶ 1∗ µˆ ¶1
p p
p∗ p
| u − uQ ( x,l ) | dy Éc |Du| d y .
Q ( x,l ) Q ( x,2 l ) ä
Remark 2.12. The Sobolev-Poincaré inequality also holds in the form

µ× ¶ 1∗ µ× ¶1
∗ p p
| u − uQ ( x,l ) | p d y É cl |Du| p d y .
Q ( x,l ) Q ( x,l )
Observe that there is the same cube on both sides. We shall return to this question
later.
Remark 2.13. The Sobolev–Poincaré inequality in Theorem 2.11 holds with the
same cubes on the both sides and it holds also for p = 1. We shall not consider
these versions here.
Remark 2.14. In this remark we consider the case p = n.
(1) As Example 1.10 shows, functions in W 1,n (Rn ) are not necessarily bounded.
(2) Assume that u ∈ W 1,n (Rn ). The Poincaré inequality implies that
× µ× ¶1
n
| u( y) − uQ | d y É | u ( y) − u Q | n d y
Q Q
µ× ¶1
n
É cl |Du( y)|n d y
Q
É ckDukL n (Rn ) < ∞
for every cube Q where c = c( n). Thus if u ∈ W 1,n (Rn ), then u is of bounded
mean oscillation, denoted by u ∈ BMO(Rn ), and
×
k uk∗ = sup | u( y) − uQ | d y É ckDukL n (Rn ) ,
Q ⊂Rn Q
where c = c( n).
(3) Assume that u ∈ W 1,n (Rn ). The John-Nirenberg inequality for BMO func-
tions gives
c 1 γk uk∗
×
eγ|u( x)−uQ | dx É +1
Q c 2 − γk uk∗
n c2
for every cube Q in R with 0 < γ < k u k∗ , where c 1 = c 1 ( n) and c 2 = c 2 ( n).
c2
By choosing γ = 2k uk∗ , we obtain
× | u ( x )− u Q | × | u( x)− uQ |
c
e kDuk n dx É ec k u k∗ dx É c
Q Q
p
for every cube Q in Rn . In particular, this implies that u ∈ L loc (Rn ) for
every power p, with 1 É p < ∞. This is the Sobolev embedding theorem in
the borderline case when p = n.
In fact, there is a stronger result called Trudinger’s inequality, which states
that for small enough c > 0, we have
µ ¶ n
× | u( x)− uQ | n−1
c kDuk n
e dx É c
Q
for every cube Q in Rn , n Ê 2, but we shall not discuss this issue here
p
T HE MORAL : W 1,n (Rn ) ⊂ L loc (Rn ) for every p, with 1 É p < ∞. This is the
Sobolev embedding theorem in the borderline case when p = n.
The next theorem gives a general Sobolev–Poincaré inequality for Sobolev

functions.
Theorem 2.15. Let 1 < p < ∞, let Ω ⊂ Rn be an open set, and assume that
1,p np
u ∈ Wloc (Ω). Let 1 É q É p∗ = n− p , for 1 < p < n, and 1 É q < ∞ for n É p < ∞.
There exists a constant c = c( n, p, q) such that
³× ´ 1q ³× ´ 1p
| u − uQ ( x,l ) | q d y É cl |Du| p d y (2.2)
Q ( x,l ) Q ( x,2 l )
for every cube Q ( x, 2 l ) b Ω.
Proof. By Theorem 2.11, for 1 < p < n, we obtain

³× ´ nnp
−p ˆ ´ nnp
−p
np n− p ³ np
−
| u − uQ ( x,l ) | n− p d y = c( n, p) l p | u − uQ ( x,l ) | n− p d y
Q ( x,l ) Q ( x,l )
ˆ ´ 1p
n³
1− p
É c( n, p) l |Du| p d y (2.3)
Q ( x,2 l )
³× ´ 1p
= c( n, p) l |Du| p d y .
Q ( x,2 l )
For 1 p Ê n,

n pe
there exists 1 < pe < n such that q = n− pe , and (2.2) follows from (2.3) with exponent
pe and an application of Hölder’s inequality on the right-hand side. For q É p,
the claim follows from the previous case and Hölder’s inequality on the left-hand
side. ä
The next remark shows that it is possible to obtain a Poincaré inequality

on cubes without the integral average also for functions that do not have zero
boundary values. However, the functions have to vanish in a large subset.
Remark 2.16. Assume u ∈ W 1,p (Rn ) and u = 0 in a set A ⊂ Q ( x, l ) = Q satisfying
| A | Ê γ|Q | for some 0 < γ É 1.
This means that u = 0 in a large portion of Q . By the Poincaré inequality there

exists c = c( n.p) such that
µ× ¶ 1 µ× ¶ 1 µ× ¶1
p p p
p p p
| u| d y É | u − uQ | d y + | uQ | d y
Q Q Q
µ× ¶1
p
É cl |Du| p d y + | uQ |,
Q
where
¯× ¯ ×
| uQ | = ¯¯ u( y) d y¯¯ É χQ \ A ( y)| u( y)| d y
¯ ¯
Q Q
¶1− 1 µ× ¶1
|Q \ A |
µ
p p
p
É | u( y)| d y
|Q | Q
µ× ¶1
p
1− 1p
É (1 − c) | u( y)| p d y .
Q
1− 1p
Since 0 É (1 − c) < 1, we may absorb the integral average to the left hand side
and obtain
µ× ¶1 µ× ¶1
p p
1− 1p
(1 − (1 − c) ) | u| p d y É cl |Du| p d y .
Q Q
It follows that there exists c = c( n, p, γ) such that
µ× ¶1 µ× ¶1
p p
| u| p d y É cl |Du| p d y .
Q Q
A similar argument can be done with the Sobolev-Poincaré inequality on cubes

(exercise).
2.3 Morrey’s inequality

Let A ⊂ Rn . A function u : A → R is Hölder continuous with exponent 0 < α É 1, if
there exists a constant c such that
| u( x) − u( y)| É c| x − y|α
for every x, y ∈ A . We define the space C 0,α ( A ) to be the space of all bounded
functions that are Hölder continuous with exponent α with the norm
| u( x) − u( y)|
k ukC α ( A ) = sup | u( x)| + sup . (2.4)
x∈ A x,y∈ A,x6= y | x − y| α
Remarks 2.17:
(1) Every function that is Hölder continuous with exponent α > 1 in the whole
space is constant (exercise).
(2) There are Hölder continuous functions that are not differentiable at any
point. Thus Hölder continuity does not imply any differentiability proper-
ties.
(3) C 0,α ( A ) is a Banach space with the norm defined above (exercise).
(4) Every Hölder continuous function on A ⊂ Rn can be extended to a Hölder
continuous function on Rn with the same exponent and same constant.
Moreover, if A is bounded, we may assume that the Hölder continuous
extension to Rn is bounded (exercise).
The next result shows that every function in W 1,p (Rn ) with p > n has a 1 − np -
¡ ¢
Hölder continuous representative up to a set of measure zero.
Theorem 2.18 (Morrey). Assume that u ∈ W 1,p (Rn ) with p > n. Then there is
c = c( n, p) > 0 such that
1− np
| u( z) − u( y)| É c| z − y| kDukL p (Rn )
for almost every z, y ∈ Rn .
Proof. (1) Assume first that u ∈ C ∞ (Rn ) ∩ W 1,p (Rn ). Let z, y ∈ Q ( x, l ). Then
ˆ 1 ˆ 1
Ç¡ ¢
u( z) − u( y) = u( tz + (1 − t) y) dt = Du( tz + (1 − t) y) · ( z − y) dt
0 Çt 0
and
¯× ¯
¯ ¯
| u( y) − uQ ( x,l ) | = ¯
¯ ( u( z) − u( y)) dz¯¯
Q ( x,l )
¯×
¯ ˆ 1 ¯
¯
=¯ Du( tz + (1 − t) y) · ( z − y) dt dz¯
¯ ¯
¯ Q ( x,l ) 0 ¯
n 1
ˆ ˆ 1¯
¯ Ç u ( tz + (1 − t) y)¯ | z j − y j | dt dz
¯ ¯
X ¯
É n ¯ Çx
j =1 l Q ( x,l ) 0 j
¯
n
ˆ 1ˆ
¯ Çu
¯ ¯
X 1 ¯
É ¯ ( tz + (1 − t ) y )¯ dz dt
Q ( x,l ) Ç x j
n − 1
j =1 l
¯ ¯
0
n
ˆ 1 ˆ
¯ Çu
¯ ¯
X 1 1 ¯
= ¯ ( w )¯ dw dt.
Q ( tx+(1− t) y,tl ) Ç x j
n − 1 n
j =1 l 0 t
¯ ¯
Here we used the fact that | z j − y j | É l , Fubini’s theorem and finally the change of
variables w = tz + (1 − t) y ⇐⇒ z = 1t (w − (1 − t) y), dz = 1
tn dw. By Hölder’s inequality
n
ˆ 1 ˆ
¯ Çu
¯ ¯
X 1 1 ¯
¯ ( w )¯ dw dt
tn Q ( tx+(1− t) y,tl ) Ç x j
n −1
j =1 l
¯ ¯
0
n
ˆ 1 µˆ ¯p ¶1
¯ Çu
¯
X 1 1 ¯ p 1
É ¯ (w)¯ dw |Q ( tx + (1 − t) y, tl )| p0 dt
¯
j =1 l
n −1
0 t n
Q ( tx+(1− t) y,tl )
¯ Ç x j
n(1− 1p ) ˆ 1
1 n(1− p )
l t
É nkDukL p (Q ( x,l )) dt (Q ( tx + (1 − t) y, tl ) ⊂ Q ( x, l )))
l n−1 0 tn
np 1− np
= l kDukL p (Q ( x,l )) .
p−n
Thus
| u( z) − u( y)| É | u( z) − uQ ( x,l ) | + | uQ ( x,l ) − u( y)|

np 1− np
É2 l kDukL p (Q ( x,l )) (2.5)
p−n
for every z, y ∈ Q ( x, l ).
For every z, y ∈ Rn , there exists a cube Q ( x, l ) 3 z, y such that l = | z − y|. For
z+ y
example, we may choose x = 2 . Thus
1− np 1− np
| u( z) − u( y)| É c| z − y| kDukL p (Q ( x,l )) É c| z − y| kDukL p (Rn )
for every z, y ∈ Rn .
(2) Assume then that u ∈ W 1,p (Rn ). Let u ε be the standard mollification of u.
Then
1− np
| u ε ( z) − u ε ( y)| É c| z − y| kDu ε kL p (Rn ) .
Now by Lemma 1.16 (2) and by Theorem 1.17, we obtain
1− np
| u( z) − u( y)| É c| z − y| kDukL p (Rn ) .
when z and y are Lebesgue points of u. The claim follows from the fact that almost
every point of a locally integrable function is a Lebesgue point. ä
Remarks 2.19:
(1) Morrey’s inequality implies that u can be extended uniquely to Rn as a
Hölder continuous function u such that
1− np
| u( x) − u( y)| É c| x − y| kDukL p (Rn ) for all x, y ∈ Rn .
Reason. Let N be a set of zero measure such that Morrey’s inequality holds
for all points in Rn \ N . Now for any x ∈ Rn , choose a sequence of points
( x i ) such that x i ∈ Rn \ N , i = 1, 2 . . . , and x i → x as i → ∞. By Morrey’s

inequality ( u( x i )) is a Cauchy sequence in R and thus we can define
u( x) = lim u( x i ).
i →∞
Now it is easy to check that u satisfies Morrey’s inequality in every pair of

points by considering sequences of points in Rn \ N converging to the pair
of points. ■
(2) If u ∈ W 1,p (Rn ) with p > n, then u is essentially bounded.
Reason. Let y ∈ Q ( x, 1). Then Morrey’s and Hölder’s inequality imply
| u( z)| É | u( z) − uQ ( x,1) | + | uQ ( x,1) |

× ˆ
É | u( z) − u( y)| d y + | u( y)| d y
Q ( x,1) Q ( x,1)
µˆ ¶1
p
p
É ckDuk L p (R n ) + | u( y)| d y
Q ( x,1)
É ck ukW 1,p (Rn )
for almost every z ∈ Rn . Thus k ukL∞ (Rn ) É ck ukW 1,p (Rn ) . ■
This implies that
k uk 0, 1− n É ck ukW 1,p (Rn ) , c = c( n, p),

C p (R n )
where u is the Hölder continuous representative of u. Hence W 1,p (Rn ) is

0,1− np
continuously embedded in C (Rn ), when p > n.
(3) The proof of Theorem 2.18, see (2.5), shows that if Ω is an open subset of
1,p
Rn and u ∈ Wloc (Ω), p > n, then there is c = c( n, p) such that
1− np
| u( z) − u( y)| É c| z − y| kDukL p (Q ( x,l ))
for every z, y ∈ Q ( x, l ), Q ( x, l ) b Ω. This is a local version of Morrey’s

inequality.
0,1− np
THE MORAL : W 1,p (Rn ) ⊂ C (Rn ), when p > n. More precisely, W 1,p (Rn ) is
0,1− np
continuously embedded in C (Rn ), when p > n. This is the Sobolev embedding
theorem for p > n.
Definition 2.20. A function u : Rn → R is differentiable at x ∈ Rn if there exists a

linear mapping L : Rn → R such that
| u( y) − u( x) − L( x − y)|
lim = 0.
y→ x | x − y|
If such a linear mapping L exists at x, it is unique and we denote L = Du( x) and

call Du( x) the derivative of u at x. If the derivative Du exists, it is unique and
satisfies
Du( y − x) = Du( x) · ( y − x)
for every y ∈ Rn , where

Çu Çu
µ ¶
Du( x) = ( x), . . . , ( x)
Ç x1 Ç xn
is the pointwise gradient of u at x.
1,p
Theorem 2.21. If u ∈ Wloc (Rn ), n < p É ∞, then u is differentiable almost every-
where and its derivative equals its weak derivative almost everywhere.
T HE MORAL : By the ACL characterization, see Theorem 1.49, we know

that every function in W 1,p , 1 É p É ∞ has classical partial derivatives almost
everywhere. If p > n, then every function in W 1,p is also differentiable almost
everywhere.
1,∞ 1,p
Proof. Since Wloc (Rn ) ⊂ Wloc (Rn ), we may assume n < p < ∞. By the Lebesgue
differentiation theorem
×
lim |Du( z) − Du( x)| p dz = 0
l →0 Q ( x,l )
for almost every x ∈ Rn . Let x be such a point and denote
v( y) = u( y) − u( x) − Du( x) · ( y − x),
1,p
where y ∈ Q ( x, l ). Observe that v ∈ Wloc (Rn ) with n < p < ∞. By (2.5) in the proof
of Morrey’s inequality, there is c = c( n, p) such that
µ× ¶1
p
|v( y) − v( x)| É cl |Dv( z)| p dz
Q ( x,l )
for almost every y ∈ Q ( x, l ), where l = | x − y|. Since v( x) = 0 and Dv( z) = Du( z) −

Du( x), we obtain
¶1
| u( y) − u( x) − Du( x) · ( y − x)|
µ×
p
p
Éc |Du( z) − Du( x)| dz →0
| y − x| Q ( x,l )
as y → x. ä
2.4 Lipschitz functions and W 1,∞

Let A ⊂ Rn and 0 É L < ∞. A function f : A → R is called Lipschitz continuous with
constant L, or an L-Lipschitz function, if
| f ( x) − f ( y)| É L| x − y|
for every x, y ∈ Rn . Observe that a function is Lipschitz continuous if it is Hölder

continuous with exponent one. Moreover, C 0,1 ( A ) is the space of all bounded
Lipschitz continuous functions with the norm (2.4).
Examples 2.22:
(1) For every y ∈ Rn the function x 7→ | x − y| is Lipschitz continuous with
constant one. Note that this function is not smooth.
(2) For every nonempty set A ⊂ Rn the function x 7→ dist( x, A ) is Lipschitz
continuous with constant one. Note that this function is not smooth when
A 6= Rn (exercise).
The next theorem describes the relation between Lipschitz functions and
Sobolev functions.
Theorem 2.23. A function u ∈ L1loc (Rn ) has a representative that is bounded and
Lipschitz continuous if and only if u ∈ W 1,∞ (Rn ).
THE M O R A L : The Sobolev embedding theorem for p > n shows that W 1,p (R n ) ⊂
0,1− np
C (Rn ). In the limiting case p = ∞ we have W 1,∞ (Rn ) = C 0,1 (Rn ). This is the
Sobolev embedding theorem for p = ∞.
1,p
Proof. ⇐= Assume that u ∈ W 1,∞ (Rn ). Then u ∈ L∞ (Rn ) and u ∈ Wloc (Rn ) for
every p > n and thus by Remark 2.19 we may assume that u is a bounded
continuous function. Moreover, we may assume that the support of u is compact.
By Lemma 1.16 (3) and by Theorem 1.17, the standard mollification u ε ∈ C 0∞ (Rn )
for every ε > 0, u ε → u uniformly in Rn as ε → 0 and
kDu ε kL∞ (Rn ) É kDukL∞ (Rn )
for every ε > 0. Thus

¯ˆ ¯
¯ 1 ¯
| u ε ( x) − u ε ( y)| = ¯ Du ε ( tx + (1 − t) y) · ( x − y) dt¯
¯ ¯
¯ 0 ¯
É kDu ε kL∞ (Rn ) | x − y|

É kDukL∞ (Rn ) | x − y|
for every x, y ∈ Rn . By letting ε → 0, we obtain
| u( x) − u( y)| É kDukL∞ (Rn ) | x − y|
for every x, y ∈ Rn .
=⇒ Assume that u is Lipschitz continuous. Then there exists L such that
| u( x) − u( y)| É L| x − y|
for every x, y ∈ Rn . This implies that

¯ u( x − he j ) − u( x) ¯
¯ ¯
−h
| D j u ( x )| = ¯
¯ ¯ÉL
h ¯
for every x ∈ Rn and h 6= 0. This means that
kD −j h ukL∞ (Rn ) É L
for every h 6= 0 and thus

1 1
kD −j h ukL2 (Ω) É kD −j h ukL∞ (Rn ) |Ω| 2 É L|Ω| 2 ,
where Ω ⊂ Rn is bounded and open.

As in the proof of Theorem 1.46, by Theorem 1.30, there exists g ∈ L2 (Ω0 ; Rn )
and a sequence ( h i ) i∈N converging to zero such that D −h i u → g weakly in L p (Ω0 ; Rn )
as i → ∞. This implies
ˆ ˆ ˆ
Çϕ
µ ¶
h h
u dx = u lim D i ϕ dx = lim uD j i ϕ dx
Ω Çx j Ω h →0 j
i h i →0 Ω 0
ˆ ˆ
−h
= − lim (D j i u)ϕ dx = − g j ϕ dx
h i →0 Ω Ω
for every ϕ ∈ C 0∞ (Ω0 ). It follows that Du = g in the weak sense in Ω and thus
u ∈ W 1,2 (Ω).
Claim: D j u ∈ L∞ (Ω), j = 1, . . . , n,
−h i
Reason. Let f i = D j u, i = 1, 2 . . . . Since f i → D j u weakly in L2 (Ω) as i → ∞, by
Mazur’s lemma as in the proof of Theorem 1.36, there exists a sequence convex
combinations such that
mi
X
fei = a i,k f k → D j u
k= i
in L p (Ω) as i → ∞. Observe that

° °
°Xmi ° mi ° °
X ° −h
k fei kL∞ (Ω) = ° a i,k f k ° É a i,k °D j k u( x)° É L.
° ° °
° k= i L ∞ (Ω )
L∞ (Ω) k= i
°
Since there exists a subsequence that converges almost everywhere, we conclude

that
¯ ¯
¯D j u( x)¯ É L, j = 1, . . . , n,
for almost every x ∈ Ω. ■
This shows that Du ∈ L∞ (Ω), with kDukL∞ (Ω) É L. As u is bounded, this

implies u ∈ W 1,∞ (Ω) for all bounded subsets Ω ⊂ Rn . Since the norm does not
depend on Ω, we conclude that u ∈ W 1,∞ (Rn ). ä
A direct combination of Theorem 2.23 and Theorem 2.21 gives a proof for
Rademacher’s theorem.
Corollary 2.24 (Rademacher). Let f : Rn → R be locally Lipschitz continuous.

Then f is differentiable almost everywhere.
W A R N I N G : For an open subset Ω of Rn , Morrey’s inequality and the charac-

terization of Lipschitz continuous functions holds only locally, that is, W 1,p (Ω) ⊂
0,1− np 0,1
C loc (Ω), when p > n and W 1,∞ (Ω) ⊂ C loc (Ω).
Example 2.25. Let
Ω = { x ∈ R2 : 1 < | x| < 2} \ {( x1 , 0) ∈ R2 : x1 Ê 1} ⊂ R2 .
Then we can construct functions such that u ∈ W 1,∞ (Ω), but u 6∈ C 0,α (Ω), for
example, by defining u( x) = θ , where θ is the argument of x in polar coordinates
with 0 < θ < 2π. Then u ∈ W 1,∞ (Ω), but u is not Lipschitz continuous in Ω.
However, it is locally Lipschitz continuous in Ω.
2.5 Summary of the Sobolev embeddings

We summarize the results related to Sobolev embeddings below. Assume that Ω is
an open subset of Rn .
∗ 1,p p∗ np
1 É p < n W 1,p (Rn ) ⊂ L p (Rn ), Wloc (Ω) ⊂ L loc (Ω), p∗ = n− p (Theorem 2.2
and Theorem 2.11).
1,n p
p = n W 1,n (Rn ) ⊂ BMO(Rn ), Wloc (Ω) ⊂ L loc (Ω) for every p, with 1 É p < ∞
(Remark 2.14 (3)).
0,1− np 1,p 0,1− np
n < p < ∞ W 1,p (Rn ) ⊂ C (Rn ), Wloc (Ω) ⊂ C loc (Ω) (Theorem 2.18).
1,∞ 0, 1
p = ∞ W 1,∞ (Rn ) = C 0,1 (Rn ), Wloc (Ω) = C loc (Ω) (Theorem 2.23).
2.6 Direct methods in the calculus of vari-

ations
Sobolev space methods are important in existence results for PDEs. Assume that
Ω ⊂ Rn is a bounded open set. Consider the Dirichlet problem

∆ u = 0 in Ω,
 u = g on ÇΩ.
Let u ∈ C 2 (Ω) be a classical solution to the Laplace equation

n Ç2 u
∆u =
X
=0
j =1 Ç x2j
and let ϕ ∈ C 0∞ (Ω). An integration by parts gives

ˆ ˆ ˆ n Ç2 u
ϕ∆ u dx =
X
0= ϕ div Du dx = ϕ dx
Ω Ω Ω j =1 Ç x2j
ˆ ˆ ˆ
n
X Ç2 u n
X Ç u Çϕ
= ϕ dx = dx = Du · D ϕ dx
j =1 Ω Ç x2j j =1 Ω Çx j Çx j Ω
for every ϕ ∈ C 0∞ (Ω). Conversely, if u ∈ C 2 (Ω) and

ˆ
Du · D ϕ dx = 0 for every ϕ ∈ C 0∞ (Ω),
Ω
then by the computation above

ˆ
ϕ∆ u dx = 0 for every ϕ ∈ C 0∞ (Ω).
Ω
By the fundamental lemma in the calculus of variations, see Corollary 1.5, we

conclude ∆ u = 0 in Ω.
THE MORAL : Assume u ∈ C 2 (Ω). Then ∆ u = 0 in Ω if and only if

ˆ
Du · D ϕ dx = 0 for every ϕ ∈ C 0∞ (Ω).
Ω
This gives a motivation to the definition below.
Definition 2.26. A function u ∈ W 1,2 (Ω) is a weak solution to ∆ u = 0 in Ω, if

ˆ
Du · D ϕ dx = 0
Ω
for every ϕ ∈ C 0∞ (Ω).
THE MORAL : There are second order derivatives in the definition of a classical
solution to the Laplace equation, but in the definition above is enough to assume
that only first order weak derivatives exist.
The next lemma shows that, in the definition of a weak solution, the class of
test functions can be taken to be the Sobolev space with zero boundary values.
Lemma 2.27. If u ∈ W 1,2 (Ω) is a weak solution to the Laplace equation, then
ˆ
Du · Dv dx = 0
Ω
1,2
for every v ∈ W0 (Ω).
Proof. Let v i ∈ C 0∞ (Ω), i = 1, 2, . . . , be such that v i → v in W 1,2 (Ω). Then by the

Cauchy-Schwarz inequality and Hölder’s inequality, we have
¯ˆ ˆ ¯ ¯ˆ ¯
¯ ¯ ¯ ¯
¯ Du · Dv dx − Du · Dv i dx¯ = ¯ Du · (Dv − Dv i ) dx¯¯
¯ ¯
Ω Ω Ω
¯
ˆ
É |Du||Dv − Dv i | dx
Ω
µˆ ¶ 1 µˆ ¶1
2 2
É |Du|2 dx |Dv − Dv i |2 dx →0
Ω Ω
as i → ∞. Thus ˆ ˆ
Du · Dv dx = lim Du · Dv i dx = 0.
Ω i →∞ Ω ä
Remark 2.28. Assume that Ω ⊂ Rn is bounded and g ∈ W 1,2 (Ω). If there exists a
weak solution u ∈ W 1,2 (Ω) to the Dirichlet problem

∆ u = 0 in Ω,
 u − g ∈ W 1,2 (Ω),
0
then the solution is unique. Observe that the boundary values are taken in the
Sobolev sense.
1,2
Reason. Let u 1 ∈ W 1,2 (Ω), with u 1 − g ∈ W0 (Ω), and u 2 ∈ W 1,2 (Ω), with u 2 − g ∈
1,2
W0 (Ω), be solutions to the Dirichlet problem above. By Lemma 2.27
ˆ ˆ
Du 1 · Dv dx = 0 and Du 2 · Dv dx = 0
Ω Ω
1,2
for every v ∈ W0 (Ω) and thus
ˆ
1,p
(Du 1 − Du 2 ) · Dv dx = 0 for every v ∈ W0 (Ω).
Ω
Since
1,2
u 1 − u 2 = ( u 1 − g) − ( u 2 − g) ∈ W0 (Ω),
| {z } | {z }
1, 2 1, 2
∈W0 (Ω) ∈W0 (Ω)
we may choose v = u 1 − u 2 and conclude

ˆ ˆ
2
|Du 1 − Du 2 | dx = (Du 1 − Du 2 ) · (Du 1 − Du 2 ) dx = 0.
Ω Ω
This implies Du 1 − Du 2 = 0 almost everywhere in Ω. By the Poincaré inequality,

see Theorem 2.8, we have
ˆ ˆ
| u 1 − u 2 |2 dx É c diam(Ω)2 |Du 1 − Du 2 |2 dx = 0.
Ω Ω
This implies u 1 − u 2 = 0 ⇐⇒ u 1 = u 2 almost everywhere in Ω. This is a PDE

proof of uniqueness and in the proof of Theorem 2.31 we shall see a variational
argument for the same result. ■
Next we consider a variational approach to the Dirichlet problem for the

Laplace equation.
Definition 2.29. Assume that g ∈ W 1,2 (Ω). A function u ∈ W 1,2 (Ω) with u − g ∈
1,2
W0 (Ω) is a minimizer of the variational integral
ˆ
I ( u) = |Du|2 dx
Ω
with boundary values g, if

ˆ ˆ
|Du|2 dx É |Dv|2 dx
Ω Ω
1,2
for every v ∈ W 1,2 (Ω) with v − g ∈ W0 (Ω).
T HE MORAL : A minimizer u minimizes the variational integral I ( u) in the

class of functions with given boundary values, that is,
ˆ ½ˆ ¾
1,2
|Du|2 dx = inf |Dv|2 dx : v ∈ W 1,2 (Ω), v − g ∈ W0 (Ω) .
Ω Ω
If there is a minimizer, then infimum can be replaced by minimum.
1,2
Theorem 2.30. Assume that g ∈ W 1,2 (Ω) and u ∈ W 1,2 (Ω) with u − g ∈ W0 (Ω).
Then ˆ ½ˆ ¾
1,p
|Du|2 dx = inf |Dv|2 dx : v ∈ W 1,2 (Ω), v − g ∈ W0 (Ω)
Ω Ω
if and only if u is a weak solution to the Dirichlet problem

∆ u = 0 in Ω,
 u − g ∈ W 1,2 (Ω).
0
T HE MORAL : A function is a weak solution to the Dirichlet problem if and

only if it is a minimizer of the corresponding variational integral with the given
boundary values in the Sobolev sense.
Proof. =⇒ Assume that u ∈ W 1,2 (Ω) is a minimizer with boundary values g ∈

W 1,2 (Ω). We use the method of variations by Lagrange. Let ϕ ∈ C 0∞ (Ω) and ε ∈ R.
1,2
Then ( u + εϕ) − g ∈ W0 (Ω) and
ˆ ˆ
|D ( u + εϕ)|2 dx = (Du + εD ϕ) · (Du + εD ϕ) dx
Ω
ˆΩ ˆ ˆ
2 2
= |Du| dx + 2ε Du · D ϕ dx + ε |D ϕ|2 dx
Ω Ω Ω
= i (ε).
Since u is a minimizer, i (ε) has minimum at ε = 0, which implies that i 0 (0) = 0.

Clearly ˆ ˆ
i 0 (ε) = 2 Du · D ϕ dx + 2ε |D ϕ|2 dx
Ω Ω
and thus ˆ
0
i (0) = 2 Du · D ϕ dx = 0.
Ω
This shows that ˆ
Du · D ϕ dx = 0
Ω
for every ϕ ∈ C 0∞ (Ω).
1,2
⇐= Assume that u ∈ W 1,2 (Ω) is a weak solution to ∆ u = 0 with u − g ∈ W0 (Ω)
1,2
and let v ∈ W 1,2 (Ω) with v − g ∈ W0 (Ω). Then
ˆ ˆ
|Dv|2 dx = |D (v − u) + Du|2 dx
Ω Ω
ˆ
= (D (v − u) + Du) · (D (v − u) + Du) dx
Ω
ˆ ˆ ˆ
= |D (v − u)|2 dx + 2 D (v − u) · Du dx + |Du|2 dx.
Ω Ω Ω
Since
1,2
v − u = (v − g) − ( u − g) ∈ W0 (Ω),
| {z } | {z }
1,2 1, 2
∈W0 (Ω) ∈W0 (Ω)
by Lemma 2.27 we have ˆ

Du · D (v − u) dx = 0
Ω
and thus
ˆ ˆ ˆ ˆ
|Dv|2 dx = |D (v − u)|2 dx + |Du|2 dx Ê |Du|2 dx
Ω Ω Ω Ω
1,2
for every v ∈ W 1,2 (Ω) with v − g ∈ W0 (Ω). Thus u is a minimizer. ä
Next we give an existence proof using the direct methods in the calculus
variations. This means that, instead of the PDE, the argument uses the variational
integral.
Theorem 2.31. Assume that Ω is a bounded open subset of Rn . Then for every
1,2
g ∈ W 1,2 (Ω) there exists a unique minimizer u ∈ W 1,2 (Ω) with u − g ∈ W0 (Ω),
which satisfies
ˆ ½ˆ ¾
1,2
|Du|2 dx = inf |Dv|2 dx : v ∈ W 1,2 (Ω), v − g ∈ W0 (Ω) .
Ω Ω
T HE MORAL : The Dirichlet problem for the Laplace equation has a unique
solution with Sobolev boundary values in any bounded open set.
WA RNING : It is not clear whether the solution to the variational problem

attains the boundary values pointwise.
Proof. (1) Since I ( u) Ê 0, in particular, it is bounded from below in W 1,2 (Ω) and
1,2
since u is a minimizer, g ∈ W 1,2 (Ω) and g − g = 0 ∈ W0 (Ω), we note that
½ˆ ¾ ˆ
2 1,2 1,2
0 É m = inf |Du| dx : u ∈ W (Ω), u − g ∈ W0 (Ω) É |D g|2 dx < ∞.
Ω Ω
The definition of infimum then implies that there exists a minimizing sequence
1,2
u i ∈ W 1,2 (Ω) with u i − g ∈ W0 (Ω), i = 1, 2, . . ., such that
ˆ
lim |Du i |2 dx = m.
i →∞ Ω
The existence of the limit implies the sequence ( I ( u i )) is bounded. Thus there
exists a constant M < ∞ such that
ˆ
I (u i ) = |Du i |2 dx É M for every i = 1, 2, . . . ,
Ω
(2) By the Poincaré inequality, see Theorem 2.8, we obtain

ˆ ˆ
| u i − g|2 dx + |D ( u i − g)|2 dx
Ω Ω
ˆ ˆ
2 2
É c diam(Ω) |D ( u i − g)| dx + |D ( u i − g)|2 dx
Ω Ω
ˆ
É ( c diam(Ω)2 + 1) |Du i − D g|2 dx
µ Ωˆ ˆ ¶
2
É ( c diam(Ω) + 1) 2 |Du i |2 dx + 2 |D g|2 dx
µ Ωˆ ¶ Ω
É c(diam(Ω)2 + 1) M + |D g|2 dx < ∞
Ω
1,2
for every i = 1, 2, . . . This shows that ( u i − g) is a bounded sequence in W0 (Ω).
1,2
(3) By reflexivity of W0 (Ω), see Theorem 1.36, there is a subsequence ( u i k − g)
1,2
and a function u ∈ W 1,2 (Ω), with u − g ∈ W0 (Ω), such that u i k → u weakly in L2 (Ω)
Çu i k Çu
and Çx j
→ Çx j
, j = 1, . . . , n, weakly in L2 (Ω) as k → ∞. By lower semicontinuity of
L2 -norm with respect to weak convergence, see (1.3), we have
ˆ ˆ ˆ
2 2
|Du| dx É lim inf |Du i k | dx = lim |Du i |2 dx.
Ω k→∞ Ω i →∞ Ω
1, 2
Since u ∈ W 1,2 (Ω), with u − g ∈ W0 (Ω), we have
ˆ ˆ
mÉ |Du|2 dx É lim |Du i |2 dx = m
Ω i →∞ Ω
which implies ˆ
|Du|2 dx = m.
Ω
Thus u is a minimizer.
1,2
(4) To show uniqueness, let u 1 ∈ W 1,2 (Ω), with u 1 − g ∈ W0 (Ω) and u 2 ∈
1,2
W 1,2 (Ω), with u 2 − g ∈ W0 (Ω) be minimizers of I ( u) with the same boundary
function g ∈ W 1,2 (Ω). Assume that u 1 6= u 2 , that is, |{ x ∈ Ω : u 1 ( x) 6= u 2 ( x)}| > 0. By
the Poincaré inequality, see Theorem 2.8, we have
ˆ ˆ
2 2
0< | u 1 − u 2 | dx É c diam(Ω) |Du 1 − Du 2 |2 dx
Ω Ω
u1 +u2
and thus |{ x ∈ Ω : Du 1 ( x) 6= Du 2 ( x)}| > 0. Let v = 2 . Then v ∈ W 1,2 (Ω) and
1 1 1,2
v− g = ( u 1 − g) + ( u 2 − g) ∈ W0 (Ω).
2 | {z } 2 | {z }
1, 2 1, 2
∈W0 (Ω) ∈W0 (Ω)
By strict convexity of ξ 7→ |ξ|2 we conclude that
1 1
|Dv|2 < |Du 1 |2 + |Du 2 |2 on { x ∈ Ω : Du 1 ( x) 6= Du 2 ( x)}.
2 2
Since |{ x ∈ Ω : Du 1 ( x) 6= Du 2 ( x)}| > 0 and using the fact that both u 1 and u 2 are
minimizers, we obtain
ˆ ˆ ˆ
1 1 1 1
|Dv|2 dx < |Du 1 |2 dx + |Du 2 |2 dx = m + m = m.
Ω 2 Ω 2 Ω 2 2
Thus I (v) < m. This is a contradiction with the fact that u 1 and u 2 are minimiz-
ers. ä
Remarks 2.32:
(1) This approach generalizes to other variational integrals as well. Indeed,
the proof above is based on the following steps:
(a) Choose a minimizing sequence.
(b) Use coercivity
k u i kW 1,2 (Ω) → ∞ =⇒ I ( u i ) → ∞.
to show that the minimizing sequence is bounded in the Sobolev space.

(c) Use reflexivity to show that there is a weakly converging subsequence.
(d) Use lower semicontinuity of the variational integral to show that the
limit is a minimizer.
(e) Use strict convexity of the variational integral to show uniqueness.
(2) If we consider C 2 (Ω) instead of W 1,2 (Ω) in the Dirichlet problem above,
then we end up having the following problems. If we equip C 2 (Ω) with the
supremum norm
k ukC 2 (Ω) = k ukL∞ (Ω) + kDukL∞ (Ω) + kD 2 ukL∞ (Ω) ,
where D 2 u is the Hessian matrix of second order partial derivatives, then

the variational integral is not coercive nor the space is reflexive. Indeed,
when n Ê 2 it is possible to construct a sequence of functions for which

the supremum tends to infinity, but the L2 norm of the gradients tends to
zero. The variationall integral is not coersive even when n = 1. If we try to
obtain coercivity and reflexivity in C 2 (Ω) by changing norm to k ukW 1,2 (Ω)
then we lose completeness, since the limit functions are not necessarily
in C 2 (Ω). The Sobolev space seems to have all desirable properties for
existence of solutions to PDEs.
Maximal function approach to
3
Sobolev spaces
We recall the definition of the maximal function.
Definition 3.1. The centered Hardy-Littlewood maximal function M f : Rn →

[0, ∞] of f ∈ L1loc (Rn ) is
ˆ
1
M f ( x) = sup | f ( y)| d y,
r >0 | B ( x, r )| B( x,r )
where B( x, r ) = { y ∈ Rn : | y − x| < r } is the open ball with the radius r > 0 and the
center x ∈ Rn .
T HE MORAL : The maximal function gives the maximal integral average of

the absolute value of the function on balls centered at a point.
Note that the Lebesgue differentiation theorem implies

ˆ
1
| f ( x)| = lim | f ( y)| d y
r →0 |B( x, r )| B( x,r )
ˆ
1
É sup | f ( y)| d y = M f ( x)
r >0 |B( x, r )| B( x,r )
for almost every x ∈ Rn .

We are interested in behaviour of the maximal operator in L p -spaces and begin
with a relatively obvious result.
Lemma 3.2. If f ∈ L∞ (Rn ), then M f ∈ L∞ (Rn ) and k M f kL∞ (Rn ) É k f kL∞ (Rn ) .
THE MORAL : If the original function is essentially bounded, then the maximal
function is essentially bounded and thus finite almost everywhere. Intuitively this
70
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 71
is clear, since the integral averages cannot be larger than the essential supremum
of the function. Another way to state this is that M : L∞ (Rn ) → L∞ (Rn ) is a
bounded operator.
Proof. For every x ∈ Rn and r > 0 we have

ˆ
1 1
| f ( y)| d y É k f kL∞ (Rn ) |B( x, r )| = k f kL∞ (Rn ) .
|B( x, r )| B( x,r ) |B( x, r )|
By taking supremum over r > 0, we have M f ( x) É k f kL∞ (Rn ) for every x ∈ Rn and
thus k M f kL∞ (Rn ) É k f kL∞ (Rn ) . ä
The following maximal function theorem was first proved by Hardy and Little-
wood in the one-dimensional case and by Wiener in higher dimensions.
Theorem 3.3 (Hardy-Littlewood-Wiener).

(1) If f ∈ L1 (Rn ), there exists c = c( n) such that
c
|{ x ∈ Rn : M f ( x) > λ}| É k f k L 1 (R n ) for every λ > 0. (3.1)
λ
(2) If f ∈ L p (Rn ), 1 < p É ∞, then M f ∈ L p (Rn ) and there exists c = c( n, p) such
that
k M f k L p (R n ) É c k f k L p (R n ) . (3.2)
T HE MORAL : The first assertion states that the Hardy-Littlewood max-

imal operator maps L1 (Rn ) to weak L1 (Rn ) and the second claim shows that
M : L p (Rn ) → L p (Rn ) is a bounded operator for p > 1.
WA RNING : f ∈ L1 (Rn ) does not imply that M f ∈ L1 (Rn ) and thus the Hardy-
Littlewood maximal operator is not bounded in L1 (Rn ). In this case we only have
the weak type estimate.
3.1 Representation formulas and Riesz po-

tentials
We begin with considering the one-dimensional case. If u ∈ C 01 (R), there exists an
interval [a, b] ⊂ R such that u( x) = 0 for every x ∈ R \ [a, b]. By the fundamental
theorem of calculus,
ˆ x ˆ x
0
u ( x) = u ( a) + u ( y) d y = u0 ( y) d y, (3.3)
a −∞
since u(a) = 0. On the other hand,

ˆ b ˆ ∞
0
0 = u ( b ) = u ( x) + u ( y) d y = u( x) + u0 ( y) d y,
x x
so that ˆ ∞
u ( x) = − u0 ( y) d y. (3.4)
x
Equalities (3.3) and (3.4) imply
ˆ x ˆ ∞
2 u ( x) = u 0 ( y) d y − u0 ( y) d y
x
ˆ−∞
x ˆ ∞
u0 ( y)( x − y) u0 ( y)( x − y)
= dy+ dy
| x − y| x | x − y|
ˆ−∞ 0
u ( y)( x − y)
= d y,
R | x − y|
from which it follows that

ˆ
1 u0 ( y)( x − y)
u ( x) = d y for every x ∈ R.
2 R | x − y|
Next we extend the fundamental theorem of calculus to Rn .
Lemma 3.4 (Representation formula). If u ∈ C 01 (Rn ), then

ˆ
1 Du( y) · ( x − y)
u ( x) = dy for every x ∈ Rn ,
ωn−1 Rn | x − y| n
where ωn−1 = nΩn is the ( n − 1)-dimensional measure of ÇB(0, 1).
T HE MORAL : This is a representation formula for a compactly supported

continuously differentiable function in terms of its gradient. A function can be
integrated back from its derivative using this formula.
Proof. If x ∈ Rn and e ∈ ÇB(0, 1), by the fundamental theorem of calculus

ˆ ∞ ˆ ∞
Ç
u ( x) = − ( u( x + te)) dt = − Du( x + te) · e dt.
0 Çt 0
By the Fubini theorem

ˆ
ωn−1 u( x) = u( x) 1 dS ( e)
ÇB(0,1)
ˆ ˆ ∞
=− Du( x + te) · e dt dS ( e)
ÇB(0,1) 0
ˆ ∞ˆ
=− Du( x + te) · e dS ( e) dt (Fubini)
0 ÇB(0,1)
ˆ ∞ˆ
y 1
=− Du( x + y) · dS ( y) dt
0 ÇB(0,t) t t n−1
( y = te, dS ( e) = t1−n dS ( y))
ˆ ∞ˆ
y
=− Du( x + y) · dS ( y) dt
0 ÇB(0,t) | y| n
ˆ
Du( x + y) · y
=− dy (polar coordinates)
| y|n
ˆR
n
Du( z) · ( z − x)
=− dz ( z = x + y, d y = dz)
| z − x| n
ˆ R
n
Du( y) · ( x − y)
= d y. ä
R n | x − y| n
Remark 3.5. By the Cauchy-Schwarz inequality and Lemma 3.4, we have

¯ ˆ ¯
¯ 1 Du( y) · ( x − y) ¯
| u( x)| = ¯¯ d y¯
ωn−1 Rn | x − y| n ¯
ˆ
1 |Du( y)|| x − y|
É dy
ωn−1 Rn | x − y| n
ˆ
1 |Du( y)|
= dy
ωn−1 Rn | x − y|n−1
1
= I 1 (|Du|)( x),
ωn−1
where I α f , 0 < α < n, is the Riesz potential

ˆ
f ( y)
I α f ( x) = d y.
Rn | x − y|n−α
THE MORAL : This gives a useful pointwise bound for a compactly supported
smooth function in terms of the Riesz potential of the gradient.
Remark 3.6. A similar estimate holds almost everywhere if u ∈ W 1,p (Rn ) or u ∈

1,p
W0 (Ω) (exercise).
We begin with a technical lemma for the Riesz potential for α = 1.
Lemma 3.7. If E ⊂ Rn is a measurable set with |E | < ∞, then

ˆ
1 1
d y É c ( n )| E | n .
E | x − y|n−1
Proof. Let B = B( x, r ) be a ball with |B| = |E |. Then |E \ B| = |B \ E | and thus

ˆ ˆ
1 1 1 1
n
d y É |E \ B| n−1 = |B \ E | n−1 É d y.
E \B | x − y| −1 r r B\E | x − y|n−1
This implies
ˆ ˆ ˆ
1 1 1
dy = dy+ dy
| x − y|n−1 | x − y| n −1 | x − y|n−1
E
ˆ E \B ˆ E ∩B
1 1
É n−1
dy+ dy
| x − y | | x − y|n−1
ˆ B\E E ∩B
1
= n−1
dy
B | x − y|
1 1
= c( n) r = c( n)|B| n = c( n)||E | n . ä
Lemma 3.8. Assume that |Ω| < ∞ and 1 É p < ∞. Then

1
k I 1 (| f |χΩ )kL p (Ω) É c( n, p)|Ω| n k f kL p (Ω) .
T HE MORAL : If |Ω| < ∞, then I 1 : L p (Ω) → L p (Ω) is a bounded operator for

1 É p < ∞.
Proof. If p > 1, Hölder’s inequality and Lemma 3.7 give

ˆ ˆ
| f ( y)| | f ( y)| 1
n−1
dy = 1 1
dy
Ω | x − y| Ω | x − y| p ( n−1)
| x − y| p0
( n−1)
µˆ ¶ 1 µˆ ¶ 10
| f ( y)| p p 1 p
É n − 1
dy n −1
d y
Ω | x − y| Ω | x − y|
µˆ p ¶1
1 | f ( y)| p
É c |Ω| np 0
n−1
dy
Ω | x − y |
µˆ ¶1
p−1 | f ( y)| p p
= c|Ω| np n −1
d y .
Ω | x − y|
For p = 1, the inequality above is clear. Thus by Fubini’s theorem and Lemma 3.7,
we have
ˆ ˆ ˆ
p−1 | f ( y)| p
| I 1 (| f |χΩ )( x)| p dx É c|Ω| n
n−1
d y dx
Ω Ω Ω | x − y|
ˆ
p−1 1
É c | Ω| n | Ω| n | f ( y)| p d y. ä
Ω
Next we show that the Riesz potential can be bounded by the Hardy-Littlewood
maximal function. We shall do this for the general α although α = 1 will be most
important for us.
Lemma 3.9. If 0 < α < n, there exists c = c( n, α), such that

ˆ
| f ( y)|
n−α
d y É cr α M f ( x)
B( x,r ) | x − y |
for every x ∈ Rn and r > 0.
Proof. Let x ∈ Rn and denote A i = B( x, r 2− i ), i = 0, 1, 2, . . . . Then

ˆ ∞
ˆ
| f ( y)| X | f ( y)|
n −α
d y = n−α
dy
B( x,r ) | x − y| i =0 A i \ A i+1 | x − y|
∞ µ r ¶α− n
ˆ
X
É i +1
| f ( y)| d y
i =0 2 Ai
∞ µ 1 ¶α− n µ r ¶α 1 µ r ¶− n
ˆ
= Ωn
X
| f ( y)| d y
i =0 2 2 i Ωn 2 i Ai
∞ µ 1 ¶α− n µ r ¶α 1
ˆ
= Ωn
X
| f ( y)| d y
i =0 2 2i | A i | A i
∞ µ 1 ¶i
É cM f ( x) r α
X
α
i =0 2
= cr α M f ( x). ä
Theorem 3.10 (Sobolev inequality for Riesz potentials). Assume that α >
0, p > 1 and α p < n. Then there exists c = c( n, p, α), such that for every f ∈ L p (Rn )
we have
pn
k I α f k L p ∗ (R n ) É c k f k L p (R n ) , p∗ = .
n − αp
THE MORAL : This is the Sobolev inequality for the Riesz potentials. Observe
that of α = 1, then p∗ is the Sobolev conjugate of p.
Proof. If f = 0 almost everywhere, the claim is clear. Thus we may assume that
f 6= 0 on a set of positive measure and thus M f > 0 everywhere. By Hölder’s
inequality
ˆ µˆ ¶ 1 µˆ ¶ 10
| f ( y)| p
p
(α− n) p0 p
dy É | f ( y)| d y | x − y| dy ,
R \B( x,r )
n | x − y|n−α Rn \B( x,r ) Rn \B( x,r )
where
ˆ ˆ ∞ˆ
(α− n) p0 0
| x − y| dy = | x − y|(α−n) p dS ( y) d ρ
Rn \B( x,r ) r ÇB( x,ρ )
ˆ ∞ ˆ
(α− n) p0
= ρ 1 dS ( y) d ρ
r ÇB( x,ρ )
| {z }
=ωn−1 ρ n−1
ˆ ∞
0
= ωn−1 ρ (α−n) p +n−1 d ρ
r
ωn−1 0
= r n−(n−α) p .
( n − α) p0 − n
The exponent can be written in the form

p αp − n
n − ( n − α) p0 = n − ( n − α) = ,
p−1 p−1
and thus ˆ
| f ( y)| α− n
n −α
d y É cr p k f kL p (Rn ) .
Rn \B( x,r ) | x − y|
Lemma 3.9 implies
ˆ
| f ( y)|
| I α f ( x)| É dy
n | x − y| n−α
ˆR ˆ
| f ( y)| | f ( y)|
= n−α
dy+ dy
B( x,r ) | x − y | R \B( x,r )
n | x − y|n−α
α− n
³ ´
É c r α M f ( x ) + r p k f k L p (R n ) .
By choosing
¶− p
M f ( x)
µ
n
r= ,
k f k L p (R n )
we obtain αp
αp
| I α f ( x)| É cM f ( x)1− n k f kLnp (Rn ) . (3.5)
np
By raising both sides to the power p∗ = n−α p , we have
αp
∗ p∗
| I α f ( x)| p É cM f ( x) p k f kLnp (Rn )
The maximal function theorem, see (3.2), implies

ˆ αp ∗
ˆ
p∗ n p
| I α f ( x)| d y É ck f k L p (R n )
( M f ( x)) p dx
Rn Rn
αp
p∗ p
= ck f kLnp (Rn ) k M f kL p (Rn )
αp
p∗ p
É ck f kLnp (Rn ) k f kL p (Rn )
and thus αp p
n + p∗
k I α f k L p ∗ (R n ) É c k f k L p (R n ) = c k f k L p (R n ) . ä
Remark 3.11. From the proof of the previous theorem we also obtain a weak type
estimate when p = 1. Indeed, by (3.5) with p = 1, there exists c = c( n, α) such that
α α
| I α f ( x)| É cM f ( x)1− n k f k n1
L (R n )
and thus the maximal function theorem with p = 1, see (3.1), implies
¯n n−α α o¯
|{ x ∈ Rn : | I α f ( x)| > t}| É ¯ x ∈ Rn : cM f ( x) n k f k n1 n > t ¯
¯ ¯
L (R )
n −α· n ¯
¯n o¯
É ¯ x ∈ Rn : M f ( x) > ct n−α k f k 1n nn−α ¯
¯
L (R )
n α
É ct− n−α k f k n−1 α k f k L 1 (R n )
L (R n )
n n
− n− n−α
= ct α kf k
L 1 (R n )
for every t > 0. This also implies that

n n
|{ x ∈ Rn : | I α f ( x)| Ê t}| É ct− n−α k f k n−1 α
L (R n )
for every t > 0.

This gives a second proof for the Sobolev-Gagliardo-Nirenberg inequality, see

Theorem 2.2.
Corollary 3.12 (Sobolev-Gagliardo-Nirenberg inequality). If 1 É p < n, there

exists a constant c = c( n, p) such that
np
k ukL p∗ (Rn ) É ckDukL p (Rn ) , p∗ = ,
n− p
for every u ∈ C 01 (Rn ).
THE MORAL : The Sobolev-Gagliardo-Nirenberg inequality is a consequence

of pointwise estimates for the function in terms of the Riesz potential of the
gradient and the Sobolev inequality for the Riesz potentials.
Proof. 1 < p < ∞ By Remark 3.5

1
| u( x)| É I 1 (|Du|)( x) for every x ∈ Rn ,
ωn−1
Thus Theorem 3.10 with α = 1 gives
k ukL p∗ (Rn ) É ck I 1 (|Du|)kL p∗ (Rn ) É ckDukL p (Rn ) .
p = 1 Let
A j = { x ∈ Rn : 2 j < | u( x)| É 2 j+1 }, j ∈ Z,
and let ϕ : R → R, ϕ( t) = max{0, min{ t, 1}}, be an auxiliary function. For j ∈ Z define
u j : Rn → [0, 1],

0, | u( x)| É 2 j−1 ,



u j ( x) = ϕ(21− j | u( x)| − 1) = 21− j | u( x)| − 1, 2 j−1 < | u( x)| É 2 j ,


1, | u( x)| > 2 j .

Lemma 1.25 implies u j ∈ W 1,1 (Rn ), j ∈ Z. Observe that Du j = 0 almost everywhere

in Rn \ A j−1 , j ∈ Z. Then
| A j | É |{ x ∈ Rn : | u( x)| > 2 j }|
= |{ x ∈ Rn : u j ( x) = 1}| (| u( x)| > 2 j =⇒ 21− j | u( x)| − 1 > 1)
É ¯{ x ∈ Rn : I 1 (|Du j |)( x) Ê ωn−1 }¯
¯ ¯
(Remark 3.5)
µˆ ¶ n
n−1
Éc |Du j ( x)| dx (Remark 3.11)
Rn
Ãˆ ! n
n−1
=c |Du j ( x)| dx
A j−1
Ãˆ ! n
n−1
0 1− j 1− j
Éc ϕ (2 | u( x)| − 1)2 |Du( x)| dx
A j−1
Ãˆ ! n
n−1
n
− j n−
= c2 1 |Du( x)| dx .
A j−1
By summing over j ∈ Z, we obtain

ˆ ˆ
n X n
| u( x)| n−1 dx = | u( x)| n−1 dx
Rn j ∈Z A j
n
2( j+1) n−1 | A j |
X
É
j ∈Z
Ãˆ ! n
X n−1
Éc |Du( x)| dx
j ∈Z A j−1
Ã ˆ ! n
n−1
X
Éc |Du( x)| dx
j ∈Z A j−1
µˆ ¶ n
n−1
=c |Du( x)| dx .
Rn
In the last equality we used the fact that the sets A j , j ∈ Z, are pairwise disjoint.ä
Remark 3.13. The Sobolev-Gagliardo-Nirenberg inequality for u ∈ W 1,p (Rn ) fol-

lows from Corollary 3.12 by using the fact that C 01 (Rn ) is dense in W 1,p (Rn ),
1 É p < n.
3.2 Sobolev-Poincaré inequalities

Next we consider Sobolev-Poincaré inequalities in balls, compare with Theorem
2.10 and Theorem 2.11 for the corresponding estimates over cubes.
First we study the one-dimensional case. Assume that u ∈ C 1 (R) and let
y, z ∈ B( x, r ) = ( x − r, x + r ). By the fundamental theorem of calculus
ˆ y
u( z) − u( y) = u0 ( t) dt.
z
Thus ˆ y ˆ x+ r ˆ
| u( z) − u( y)| É | u0 ( t)| dt É | u0 ( t)| dt = | u0 ( t)| dt
z x− r B( x,r )
and
¯ × ¯
¯ ¯
| u( z) − u B( x,r) | = ¯¯ u( z) − u( y) d y¯¯
B( x,r )
¯× × ¯
¯ ¯
=¯ ¯ u( z) d y − u( y) d y¯¯
B( x,r ) B( x,r )
× ˆ
É | u( z) − u( y)| d y É | u0 ( y)| d y.
B( x,r ) B( x,r )
This is a pointwise estimate of the oscillation of the function. Next we generalize

this to Rn .
Lemma 3.14. Let u ∈ C 1 (Rn ) and B( x, r ) ⊂ Rn . There exists c = c( n) such that

ˆ
¯ ¯
¯ u( z) − u B( x,r) ¯ É c |Du( y)|
dy
B( x,r ) | z − y|n−1
for every z ∈ B( x, r ).
T HE MORAL : This is a pointwise estimate for the oscillation of the function

in terms of the Riesz potential of the gradient.
Proof. For any y, z ∈ B( x, r ), we have

ˆ 1 ˆ 1
Ç
u ( y) − u ( z ) = ( u( t y + (1 − t) z)) dt = Du( t y + (1 − t) z) · ( y − z) dt.
0 Çt 0
By the Cauchy-Schwarz inequality

ˆ 1
| u( y) − u( z)| É | y − z| |Du( t y + (1 − t) z)| dt.
0
Let ρ > 0. In the next display, we make a change of variables
1
w = t y + (1 − t) z ⇐⇒ y = (w − (1 − t) z), dS ( y) = t1−n dS (w).
t
³ ´n−1
Then we have |w − z| = t| y − z| and t n−1 = | z−ρw| , where ρ = | y − z|. We arrive at
ˆ
| u( y) − u( z)| dS ( y)
B( x,r )∩ÇB( z,ρ )
ˆ 1ˆ
Éρ |Du( t y + (1 − t) z)| dS ( y) dt
0 B( x,r )∩ÇB( z,ρ )
ˆ 1 ˆ
1
=ρ |Du(w)| dS (w) dt
0 t n−1 B( x,r )∩ÇB( z,tρ )
ˆ 1ˆ
n |Du(w)|
=ρ n−1
dS (w) dt
0 B( x,r )∩ÇB( z,tρ ) | z − w|
ˆ ρˆ
|Du(w)| 1
= ρ n−1 dS (w) ds ( s = tρ , dt = ds)
0 B( x,r )∩ÇB( z,s) | z − w|
n−1 ρ
ˆ
|Du(w)|
= ρ n−1 n−1
dw. (polar coordinates)
B( x,r )∩B( z,ρ ) | z − w|
Since B( x, r ) ⊂ B( z, 2 r ), an integration in polar coordinates gives

×
¯ ¯
¯ u( z) − u B( x,r) ¯ É | u( z) − u( y)| d y
B( x,r )
ˆ 2r ˆ
1
= | u( y) − u( z)| dS ( y) d ρ
|B( x, r )| 0 B( x,r )∩ÇB( z,ρ )
ˆ 2r ˆ
1 n−1 |Du( y)|
É ρ n−1
d y dρ
|B( x, r )| 0 B( x,r )∩B( z,ρ ) | z − y|
ˆ 2r ˆ
1 |Du( y)|
É ρ n−1 d ρ dy
|B( x, r )| 0 B( x,r ) | z − y|n−1
ˆ
|Du( y)|
= c ( n) d y. ä
B( x,r ) | z − y|n−1
Remarks 3.15:
(1) Assume that u ∈ C 1 (Rn ). By Lemma 3.14 and Lemma 3.9, we have
ˆ
|Du( y)|
| u( z) − u B( x,r) | É c dy
B( x,r ) | z − y|n−1
= cI 1 (|Du|χB( x,r) )( z)
ˆ
|Du( y)|χB( x,r) ( y)
Éc dy
B( z,2 r ) | z − y|n−1
É crM (|Du|χB( x,r) )( z),
for every z ∈ B( x, r ).
Next we show that the corresponding inequalities hold true almost every-
1,p
where if u ∈ Wloc (Rn ), 1 É p < ∞. Since C ∞ (B( x, r )) is dense in W 1,p (B( x, r )),
there exists a sequence u i ∈ C ∞ (B( x, r )), i = 1, 2, . . . , such that u i → u in
W 1,p (B( x, r )) as i → ∞. By passing to a subsequence, if necessary, we
obtain an exceptional set N1 ⊂ Rn with | N1 | = 0 such that
lim u i ( z) = u( z) < ∞
i →∞
for every z ∈ B( x, r ) \ N1 . By linearity of the Riesz potential and by Lemma

3.8, we have
° I 1 (|Du i |χB( x,r) ) − I 1 (|Du|χB( x,r) )° p

° °
L (B( x,r ))
= I 1 ((|Du i | − |Du|)χB( x,r) )°L p (B( x,r))

° °
°
1° °
É c|B( x, r )| n °|Du i | − |Du|°L p (B( x,r)) ,
which implies that
I 1 (|Du i |χB( x,r) ) → I 1 (|Du|χB( x,r) ) in L p (B( x, r )) as i → ∞.
By passing to a subsequence, if necessary, we obtain an exceptional set

N2 ⊂ B( x, r ) with | N2 | = 0 such that
lim I 1 (|Du i |χB( x,r) )( z) = I 1 (|Du|χB( x,r) )( z) < ∞

i →∞
for every z ∈ B( x, r ) \ N2 . Thus
| u( z) − u B( x,r) | = lim | u i ( z) − ( u i )B( x,r) |

i →∞
É c lim I 1 (|Du i |χB( x,r) )( z)

i →∞
= cI 1 (|Du|χB( x,r) )( z)
É crM (|Du|χB( x,r) )( z),
for every z ∈ B( x, r ) \ ( N1 ∪ N2 ).
(2) By Lemma 3.14 and (3.5), we have

ˆ
|Du( y)|
| u( z) − u B( x,r) | É c dy
B( x,r ) | z − y|n−1
= cI 1 (|Du| χB( x,r) )( z)
p° °p
É cM (|Du| χB( x,r) )( z)1− n °|Du| χB( x,r) °Ln p (Rn )
for every z ∈ B( x, r ). The corresponding inequalities hold true almost

1,p
everywhere if u ∈ Wloc (Rn ), 1 É p < ∞.
This gives a proof for the Sobolev-Poincaré inequality on balls, see Theorem
2.11 for the correspoding statement for cubes. Maximal function arguments can
be used for cubes as well.
Theorem 3.16 (Sobolev-Poincaré inequality on balls). Assume that u ∈ W 1,p (Rn )

and let 1 < p < n. There exists c = c( n, p) such that
µ× ¶ 1 µ× ¶1
p∗ p
p∗ p
| u − u B( x,r) | dy É cr |Du| d y
B( x,r ) B( x,r )
n
for every B( x, r ) ⊂ R .
THE MORAL : The Sobolev-Poincaré inequality is a consequence of pointwise

estimates for the oscillation of the function in terms of the Riesz potential of the
gradient and the Sobolev inequality for the Riesz potentials.
Proof. By Remark 3.15, we have
| u( y) − u B( x,r) | É cI 1 (|Du| χB( x,r) )( y)
for almost every y ∈ B( x, r ). Thus Theorem 3.10 implies

µˆ ¶ 1∗ µˆ ¶ 1∗
∗ p ∗ p
| u − u B( x,r) | p d y Éc I 1 (|Du|χB( x,r) ) p d y
B( x,r ) Rn
µˆ ¶1
p
Éc (|Du|χB( x,r) ) p d y
Rn
µˆ ¶1
p
p
=c |Du| d y . ä
B( x,r )
A similar argument can be used to prove a counterpart of Theorem 2.10 as

well.
Theorem 3.17 (Poincaré inequality on balls). Assume that u ∈ W 1,p (Rn )and
let 1 < p < ∞. There exists c = c( n, p) such that
µ× ¶1 µ× ¶1
p p
| u − u B( x,r) | p d y É cr |Du| p d y
B( x,r ) B( x,r )
for every B( x, r ) ⊂ Rn .
Proof. By Remark 3.15, we have
| u( y) − u B( x,r) | É crM (|Du|χB( x,r) )( y)
for almost every y ∈ B( x, r ). The maximal function theorem with p > 1, see (3.2),
implies
ˆ ˆ
p p
| u − u B( x,r) | d y É cr M (|Du|χB( x,r) ) p d y
B( x,r ) Rn
ˆ
p
É cr (|Du|χB( x,r) ) p d y
Rn
ˆ
= cr p |Du| p d y. ä
B( x,r )
The maximal function approach to Sobolev-Poincaré inequalities is more in-

volved in the case p = 1, since then we only have a weak type estimate. However,
it is possible to consider that case as well, but this requires a different proof. We
begin with two rather technical lemmas.
Lemma 3.18. Assume that E ⊂ Rn is a measurable set and that f : E → [0, ∞] is

a measurable function for which
¯{ x ∈ E : f ( x) = 0}¯ Ê |E | .
¯ ¯
2
Then for every a ∈ R and λ > 0, we have
¯{ x ∈ E : f ( x) > λ}¯ É ¯ x ∈ E : | f ( x) − a| > λ ¯ .

¯½ ¾¯
¯ ¯ ¯ ¯
¯ 2 ¯
Proof. Assume first that |a| É λ2 . If x ∈ E with f ( x) > λ, then
λ
| f ( x) − a| Ê f ( x) − | a| > .
2
Thus { x ∈ E : f ( x) > λ} ⊂ x ∈ E : | f ( x) − a| > λ2 and

© ª
¯{ x ∈ E : f ( x) > λ}¯ É ¯ x ∈ E : | f ( x) − a| > λ ¯ .

¯½ ¾¯
¯ ¯ ¯ ¯
¯ 2 ¯
Assume then that |a| > λ2 . If x ∈ E with f ( x) = 0, then
λ
| f ( x) − a| = | a| > .
2
Thus
λ
½ ¾
{ x ∈ E : f ( x) = 0} ⊂ x ∈ E : f ( x) > .
2
If |E | = ∞, then by assumption
¯{ x ∈ E : f ( x) = 0}¯ Ê |E | = ∞
¯ ¯
2
and thus ¯ x ∈ E : f ( x) Ê λ2 ¯ = ∞. On the other hand, if |E | < ∞, then

¯© ª¯
¯{ x ∈ E : f ( x) > λ}¯ É |E | − ¯{ x ∈ E : f ( x) = 0}¯

¯ ¯ ¯ ¯
É |{ x ∈ E : f ( x) = 0}|
λ ¯¯
¯½ ¾¯
¯
É ¯¯ x ∈ E : | f ( x) − a| > .
2 ¯
This completes the proof. ä
Lemma 3.19. Assume that u ∈ C 0,1 (Rn ), that is, u is a bounded Lipschitz contin-
uous function in Rn , and let B( x, r ) be a ball in Rn . Then there exists λ0 ∈ R for
which
¯{ y ∈ B( x, r ) : u( y) Ê λ0 }¯ Ê |B( x, r )| ¯{ y ∈ B( x, r ) : u( y) É λ0 }¯ Ê |B( x, r )| .
¯ ¯ ¯ ¯
and
2 2
Proof. Denote E λ = { y ∈ B( x, r ) : u( y) Ê λ}, λ ∈ R, and set
|B( x, r )|
½ ¾
λ0 = sup λ ∈ R : |E λ | Ê .
2
Note that |λ0 | É k ukL∞ (Rn ) < ∞. Thus there exists an increasing sequence of real
numbers (λ i ) such that λ i → λ0 and
|B( x, r )|
|E λ i | Ê for every i = 1, 2, . . . .
2
T∞
Since E λ0 = i =1
E λ i , E λ1 ⊃ E λ2 ⊃ . . . and |E λ i | É |B( x, r )| < ∞, we conclude that
|B( x, r )|
|E λ0 | = lim |E λ i | Ê .
i →∞ 2
This shows that
¯{ y ∈ B( x, r ) : u( y) Ê λ0 }¯ Ê |B( x, r )| .
¯ ¯
2
A similar argument shows the other claim (exercise). ä
The next result is Theorem 3.16 with p = 1.

1,1
Theorem 3.20. Assume that u ∈ Wloc (Rn ). There exists c = c( n) such that
µ× ¶ n−1 ×
n n
| u − u B( x,r) | n−1 dy É cr |Du| d y
B( x,r ) B( x,r )
for every B( x, r ) ⊂ Rn .
Proof. By Lemma 3.19 there is a number λ0 ∈ R for which
¯{ y ∈ B( x, r ) : u( y) Ê λ0 }¯ Ê |B( x, r )| ¯{ y ∈ B( x, r ) : u( y) É λ0 }¯ Ê |B( x, r )| .
¯ ¯ ¯ ¯
and
2 2
Denote
v+ = max{ u − λ0 , 0} and v− = − min{ u − λ0 , 0}.
1,1
Both of these functions belong to Wloc (Rn ). For the rest of the proof v Ê 0 denotes
either v+ or v− . All statements are valid in both cases.
Let
A j = { y ∈ B( x, r ) : 2 j < v( y) É 2 j+1 }, j ∈ Z,
and let ϕ : R → R, ϕ( t) = max{0, min{ t, 1}}, be an auxiliary function. We define

v j : B( x, r ) → [0, 1],
v j ( y) = ϕ(21− j v( y) − 1), j ∈ Z.
Lemma 1.25 implies v j ∈ W 1,1 (B( x, r )), j ∈ Z. By Remark 3.15 (2) with p = 1, we
have
n ° 1
|v j ( y) − (v j )B( x,r) | n−1 É cM (|Dv j |χB( x,r) )( y)°|Dv j |χB( x,r) ° n−1 1
°
L (R n )
.
1
Lemma 3.18 with λ = 2 and a = (v j )B( x,r) gives
| A j | É |{ y ∈ B( x, r ) : v( y) > 2 j }|
¯½
1 ¯¯
¾¯
¯
:
É ¯ y ∈ B( x, r ) v j ( y) >
¯
2 ¯
¯½
1 ¯¯
¾¯
¯
É ¯¯ y ∈ B( x, r ) : |v j ( y) − (v j )B( x,r) | >
4 ¯
¯½ ¾¯
° 1
É ¯¯ y ∈ Rn : M (|Dv j |χB( x,r) )( y) Ê c°|Dv j |χB( x,r) ° 1−1 n
¯ ° ¯
¯.
L (R ) ¯
n
The last term can be estimated using the weak type estimate for the maximal
function, see (3.1), and the fact that
|Dv j | = 21− j |Dv|χ A j−1
almost everywhere in B( x, r ). Thus we arrive at

¯n ° 1 o¯
¯ y ∈ Rn : M (|Dv j |χB( x,r) )( y) Ê c°|Dv j |χB( x,r) °L1−1 n(Rn ) ¯
¯ ° ¯
1
ˆ
É ck|Dv j |χB( x,r) k n−1 1 |Dv j ( y)|χB( x,r) ( y) d y
L (R n )
Rn
° n
= c°|Dv j |χB( x,r) ° n−1 1 n
°
L (R )
jn °
− n−1 °
° n
É c2 |Dv|χ A j−1 ∩B( x,r) ° n−1 1 .
L (R n )
Combining the above estimates for | A j |, we obtain

ˆ ˆ
n X n X ( j +1) n
v( y) n−1 d y = v( y) n−1 d y = 2 n−1 |A j|
B( x,r ) j ∈Z A j j ∈Z
( j +1) n jn ° ° n
2− n−1 °|Dv|χ A j−1 ∩B( x,r) °Ln−1 (1Rn )
X
Éc 2 n−1
j ∈Z
° ° n
°X ° n−1
É c ° |Dv|χ A j−1 ∩B( x,r) °
° °
° j∈Z ° 1
L (R n )
° n
É c °|Du|χB( x,r) ° n−1 1
°
L (R n )
.
Since | u − λ0 | = v+ + v− , we obtain
µ× ¶ n−1 µ× ¶ n−1
n n n n
| u − u B( x,r) | n−1 dy É2 | u − λ0 | n−1 dy
B( x,r ) B( x,r )
µ× ¶ n−1 µ× ¶ n−1
n n n n
É2 v+ ( y) n−1 d y +2 v− ( y) n−1 d y
B( x,r ) B( x,r )
É c °|Du|χB( x,r) °L1 (Rn )

° °
ˆ
=c |Du( y)| d y. ä
B( x,r )
THE MORAL : The proof shows that in this case a weak type estimate implies
a strong type estimate. Observe carefully, that this does not hold in general. The
reason why this works here is that we consider gradients, which have the property
that they vanish on the set where the function itself is constant.
Next we give a maximal function proof for Morrey’s inequality, see Theorem
2.18 and Remark 2.19 (3).
Theorem 3.21 (Morrey’s inequality). Assume that u ∈ C 1 (Rn ) and let n < p <
∞. There exists c = c( n, p) such that
µ× ¶1
p
p
| u( y) − u( z)| É cr |Du| dw
B( x,r )
n
for every B( x, r ) ⊂ R and y, z ∈ B( x, r ).
Proof. By Lemma 3.14
| u( y) − u( z)| É | u( y) − u B( x,r) | + | u B( x,r) − u( z)|

ˆ ˆ
|Du(w)| |Du(w)|
Éc dw + c dw
B( x,r ) | y − w|n−1 B( x,r ) | z − w|n−1
for every y, z ∈ B( x, r ). Hölder’s inequality gives

ˆ µˆ ¶ 1 µˆ ¶ 10
|Du(w)| p
p
(1− n) p0 p
n−1
dw É |Du| dw | y − w| dw ,
B( x,r ) | y − w| B( x,r ) B( x,r )
where
ˆ ˆ
(1− n) p0 0
| y − w| dw É | y − w|(1−n) p dw
B( x,r ) B( y,2 r )
ˆ 2r ˆ
0
= ρ (1−n) p dS (w) d ρ
0 ÇB( y,ρ )
ˆ 2r
0 0
= ωn−1 ρ (1−n) p +n−1 d ρ = cr n−(n−1) p .
0
Since
( n − ( n − 1) p0 ) p10 = 1 − np ,
we have
ˆ µˆ ¶1
|Du(w)| 1− n p
n −1
dw É cr p |Du| p dw .
B( x,r ) | y − w| B( x,r )
The same argument applies to the other integral as well, so that

µˆ ¶1
p
1− np p
| u( y) − u( z)| É cr |Du| dw . ä
B( x,r )
3.3 Sobolev inequalities on domains

In this section we study open sets Ω ⊂ Rn for which the Sobolev-Poincaré inequality
µˆ ¶ 1∗ µˆ ¶1
p∗
p
p
p np
|u − uΩ | dy É c( p, n, Ω) |Du| d y , 1 É p < n, p∗ = ,
Ω Ω n− p
holds true for all u ∈ C (Ω). We already know that this inequality holds if Ω
∞
is a ball, but are there other sets for which it holds true as well? We begin by
introducing an appropriate class of domains.
Definition 3.22. A bounded open set Ω ⊂ Rn is a John domain, if there is c J Ê 1

and a point x0 ∈ Ω so that every point x ∈ Ω can be joined to x0 by a path γ : [0, 1] →
Ω such that γ(0) = x, γ(1) = x0 and
dist(γ( t), ÇΩ) Ê c−J1 | x − γ( t)|
for every t ∈ [0, 1].
T HE MORAL : In a John domain every point can be connected to the distin-

guished point with a curve that is relatively far from the boundary.
Remarks 3.23:
(1) A bounded and connected open set Ω ⊂ Rn satisfies the interior cone condi-
tion, if there exists a bounded cone
C = { x ∈ Rn : x12 + · · · + x2n−1 É ax2 , 0 É xn É b}
such that every point of Ω is a vertex of a cone congruent to C and entirely

contained in Ω. Every domain with interior cone condition is a John
domain (exercise). Roughly speaking the main difference between the
interior cone condition and a John domain is that rigid cones are replaced
by twisted cones.
(2) The collection of John domains is relatively large. For example, a domain
whose boundary is von Koch snowflake is a John domain.
Theorem 3.24. If Ω ⊂ Rn is a John domain and 1 É p < n, then

µˆ ¶ 1 µˆ ¶1
p∗ p
p∗ p
|u − uΩ | dx É c( p, n, c J ) |Du| dx , 1 < p < n,
Ω Ω
for all u ∈ C ∞ (Ω).
THE MORAL : The Sobolev-Poincaré inequality holds for many other sets than
balls as well.
WA RNING : A rooms and passages example shows that the Sobolev-Poincaré

inequality does not hold for all sets.
Proof. Let x0 ∈ Ω be the distinguished point in the John domain. Denote B0 =

B( x0 , r 0 ), r 0 = 41 dist( x0 , ÇΩ). We show that there is a constant M = M ( c J , n) such
that for every x ∈ Ω there is a chain of balls B i = B( x i , r i ) ⊂ Ω, i = 1, 2, . . . , with the
properties
(1) |B i ∪ B i+1 | É M |B i ∩ B i+1 |, i = 1, 2, . . . ,

(2) dist( x, B i ) É Mr i , r i → 0, x i → x as i → ∞ and
(3) no point of Ω belongs to more than M balls B i .
To construct the chain, first assume that x is far from x0 , say x ∈ Ω \ B( x, 2 r 0 ). Let
γ be a John path that connects x to x0 . All balls on the chain are centered on
γ. We construct the balls recursively. We have already defined B0 . Assume that
B0 , . . . , B i have been constructed. Starting from the center x i of B i we move along
γ towards x until we leave B i for the last time. Let x i+1 be the point on γ where
this happens and define
1
B i+1 = B( x i+1 , r i+1 ), r i+1 = | x − x i+1 |.
4c J
By construction B i ⊂ Ω. Property (1) and dist( x, B i ) É Mr i in (2) follow from

the fact that the consecutive balls have comparable radii and that the radii are
comparable to the distances of the centers of the balls to x.
To prove (3) assume that y ∈ B i 1 ∩ · · · ∩ B i k . Observe that the radii of B i j ,
j = 1, . . . , k, are comparable to | x − y|. By construction, if i j < i m , the center of B i m
does not belong to B i j . This implies that the distances between the centers of B i j
are comparable to | x − y|. The number of points in Rn with pairwise comparable
distances is bounded, that is, if z1 , . . . , z m ∈ Rn satisfy
r
< dist( z i , z j ) < cr for i 6= j,
c
then m É N = N ( c, n). Thus k is bounded by a constant depending only on n and
c J . This implies (3). Property (3) implies , r i → 0, x i → x as i → ∞.
The case x ∈ B( x, 2 r 0 ) is left as an exercise.
Since ×
uBi = u ( y) d y → u ( x )
Bi
for every x ∈ Ω as i → ∞, we obtain

∞
X
| u ( x) − u B 0 | É | u B i − u B i+1 |
i =0
X∞ ¡ ¢
É | u B i − u B i ∩B i+1 | + | u B i ∩B i+1 − u B i+1 |
i =0
X∞ µ |B i |
×
|B i+1 |
× ¶
É |u − uB i | d y + | u − u B i+1 | d y
i =0 |B i ∩ B i +1 | B i |B i ∩ B i+1 | B i+1
∞ ×
X
Éc | u − u B i | d y (property (1))
i =0 B i
X∞ ×
Éc ri |Du| d y (Poincaré inequality, see Theorem 3.17)
i =0 Bi
∞
ˆ
X |Du|
=c d y.
i =0 B i r ni −1
Property (2) implies | x − y| É cr i for every y ∈ B i and
1 c
É for every y ∈ Bi.
r ni −1 | x − y|n−1
Thus ∞
ˆ ˆ
X |Du( y)| |Du( y)|
| u ( x) − u B 0 | É c dy É c d y.
i =0 B i | x − y|n−1 Ω | x − y|
n−1
The last inequality follows from (3). We observe that
| u ( x ) − u Ω | É | u ( x ) − u B 0 | + | u B 0 − u Ω |,
where by Lemma 3.7 we have

ˆ
1
| u B0 − u Ω | É | u( x) − u B0 | dx
| Ω| Ω
ˆ ˆ
1 |Du( y)|
Éc dx d y
|Ω| Ω Ω | x − y|n−1
ˆ µˆ
1 1
¶
=c |Du( y)| dx dy
|Ω | Ω Ω | x − y|
n−1
ˆ
−1+ n1
É c |Ω| |Du( y)| d y.
Ω
By the John condition we have

1
c|Ω| n Ê dist( x0 , ÇΩ) Ê c−J1 | x − x0 |
and by taking supremum over x ∈ Ω we obtain

1
diam Ω É c( n, c J )|Ω| n
and thus
n−1 c
|Ω |− n É for every y ∈ Ω.
| x − y|n−1
This implies ˆ
|Du( y)|
| u B0 − u Ω | É c dy
Ω | x − y|n−1
and thus ˆ
|Du( y)|
| u ( x) − u Ω | É c d y = cI 1 (|Du|χΩ )( x).
Ω | x − y|n−1
Theorem 3.10 implies
µˆ ¶ 1 µˆ ¶ 1
p∗ p∗
p∗ p∗
| u ( x) − u Ω | dx Éc | I 1 (|Du( x)|χΩ ( x))| dx
B( x,r ) Rn
µˆ ¶1
p
p
Éc (|Du( x)|χΩ ( x)) dx
Rn
µˆ ¶1
p
=c |Du( x)| p dx . ä
Ω
3.4 A maximal function characterization

of Sobolev spaces
Similar argument as in the proof of Sobolev-Poincaré inequality gives the following
pointwise estimate.
Theorem 3.25. Assume that u ∈ C 1 (Rn ). There exists a constant c = c( n) > 0 such
that
| u( x) − u( y)| É c| x − y|( M |Du|( x) + M |Du|( y))
for every x, y ∈ Rn .
Proof. Let x, y ∈ Rn . Then x, y ∈ B( x, 2| x − y|) and B( x, 2| x − y|) ⊂ B( y, 4| x − y|). By

Remark 3.15 we obtain
| u( x) − u( y)| É | u( x) − u B( x,2| x− y|) | + | u B( x,2| x− y|) − u( y)|

É c| x − y|( M |Du|( x) + M |Du|( y)). ä
Remarks 3.26:
(1) If |Du| ∈ L p (Rn ), 1 < p É ∞, then by (3.2) we have M |Du| ∈ L p (Rn ).
(2) If |Du| ∈ L1 (Rn ), then by (3.1) we have M |Du| < ∞ almost everywhere.
(3) If |Du| ∈ L∞ (Rn ), then M |Du| É k M |Du|kL∞ (Rn ) É kDukL∞ (Rn ) everywhere.
Thus
| u( x) − u( y)| É ckDukL∞ (Rn ) | x − y|
for every x, y ∈ Rn . In other words, u is Lipschitz continuous.

Theorem 3.27. Assume that u ∈ W 1,p (Rn ), 1 < p < ∞. There exists c = c( n) and a
set N ⊂ Rn with | N | = 0 such that
for every x, y ∈ Rn \ N .
Proof. C 0∞ (Rn ) is dense in W 1,p (Rn ) by Lemma 1.24. Thus there exists a sequence
u i ∈ C 0∞ (Rn ), i = 1, 2, . . . , such that u i → u in W 1,p (Rn ) as i → ∞. By passing to a
subsequence, if necessary, we obtain an exceptional set N1 ⊂ Rn with | N1 | = 0 such
that
lim u i ( x) = u( x) < ∞
i →∞
for every x ∈ Rn \ N1 . By the sublinearity of the maximal operator and the maximal
function theorem
° ° ° °
° M |Du i | − M |Du|° p n É ° M (|Du i | − |Du|)° p n
L (R ) L (R )
° °
É c |Du i | − |Du|| L p (Rn )
° °
É ckDu i − DukL p (Rn )
which implies that M |Du i | → M |Du| in L p (Rn ) as i → ∞. By passing to a sub-

sequence, if necessary, we obtain an exceptional set N2 ⊂ Rn with | N2 | = 0 such
that
lim M |Du i |( x) = M |Du|( x) < ∞
i →∞
for every x ∈ Rn \ N2 . By Theorem 3.25
| u( x) − u( y)| = lim | u i ( x) − u i ( y)|

i →∞
É c| x − y| lim ( M |Du i |( x) + M |Du i |( y))

i →∞
É c| x − y|( M |Du|( x) + M |Du|( y))
for every x ∈ Rn \ ( N1 ∪ N2 ). ä
Remark 3.28. Compare the proof above to Remark 3.15, which shows that the
result holds for u ∈ W 1,p (Rn ), 1 É p É ∞.
The following definition motivated by Theorem 3.25.
Definition 3.29. Assume that 1 < p < ∞ and let u ∈ L p (Rn ). For a measurable
function g : R n → [0, ∞] we denote g ∈ D ( u) if there exists an exceptional set
N ⊂ Rn such that | N | = 0 and
| u( x) − u( y)| É | x − y|( g( x) + g( y)) (3.6)
for every x, y ∈ Rn \ N . We say that u ∈ L p (Rn ) belongs to the Hajłasz-Sobolev

space M 1,p (Rn ), if there exists g ∈ L p (Rn ) with g ∈ D ( u). This space is endowed
with the norm
k uk M 1,p (Rn ) = k ukL p (Rn ) + inf k gkL p (Rn ) .
g∈D ( u)
THE MORAL : The space M 1,p (Rn ) is defined through the pointwise inequality
(3.6).
Theorem 3.30. Assume that 1 < p < ∞. Then M 1,p (Rn ) = W 1,p (Rn ) and the asso-
ciate norms are equivalent, that is, there exists c such that
1
k ukW 1,p (Rn ) É k uk M 1,p (Rn ) É ck ukW 1,p (Rn )
c
for every measurable function u that belongs to M 1,p (Rn ) = W 1,p (Rn ).
THE MORAL : This is a pointwise characterization of Sobolev spaces. This can

be used as a definition of the first order Sobolev spaces on metric measure spaces.
Proof. ⊃ Assume that u ∈ W 1,p (Rn ). By Theorem 3.27 there exists c = c( n) and a
set N ⊂ Rn with | N | = 0 such that
for every x, y ∈ Rn \ N . Thus g = cM |Du| ∈ D ( u) ∩ L p (Rn ) and by the maximal

function theorem
k uk M 1,p (Rn ) = k ukL p (Rn ) + inf k gkL p (Rn )

g∈D ( u)
É k ukL p (Rn ) + k cM |Du|kL p (Rn )

É k ukL p (Rn ) + ckDukL p (Rn )
É ck ukW 1,p (Rn ) ,
where c = c( n, p).
⊂ Assume then that u ∈ M 1,p (Rn ). Then u ∈ L p (Rn ) and there exists g ∈ L p (Rn )
with g ∈ D ( u). Then
| u( x + h) − u( x)| É | h|( g( x + h) + g( x))
for almost every x, h ∈ Rn and thus

ˆ ˆ
p p
| u( x + h) − u( x)| dx É | h| ( g( x + h) + g( x)) p dx
Rn Rn
ˆ
É 2 p | h| p ( g( x + h) p + g( x) p ) dx
Rn
p+1 p
É2 k g k L p (R n ) | h | p .
By the characterization of the Sobolev space with the integrated difference quo-
tients, see Theorem 1.46, we conclude u ∈ W 1,p (Rn ) and
k ukW 1,p (Rn ) É ck ukL p (Rn ) + ck gkL p (Rn ) .
The inequality k ukW 1,p (Rn ) É ck uk M 1,p (Rn ) follows by taking infimum over all g ∈
D ( u) ∩ L p (Rn ). ä
Remark 3.31. The pointwise characterization of Sobolev spaces in Theorem 3.30 is

very useful in studying properties of Sobolev spaces. For example, if u ∈ M 1,p (Rn )
and g ∈ D ( u) ∩ L p (Rn ), then by the triangle inequality
¯ ¯
¯| u( x)| − | u( y)|¯ É | u( x) − u( y)| É | x − y|( g( x) + g( y))
Thus g ∈ D (| u|) ∩ L p (Rn ) and consequently | u| ∈ M 1,p (Rn ).
The pointwise characterization of Sobolev spaces in Theorem 3.30 can be used

to show a similar result as Theorem 1.41.
Lemma 3.32. The function u belongs to W 1,p (Rn ) if and only if u ∈ L p (Rn ) and
there are functions u i ∈ L p (Rn ), i = 1, 2, . . . , such that u i → u almost everywhere
and g i ∈ D ( u i ) ∩ L p (Rn ) such that g i → g almost everywhere for some g ∈ L p (Rn ).
Proof. If u ∈ W 1,p (Rn ), then the claim of the lemma is clear. To see the converse,
suppose that u, g ∈ L p (Rn ), g i ∈ D ( u i ) ∩ L p (Rn ) and u i → u almost everywhere and
g i → g almost everywhere. Then
¡ ¢
| u i ( x) − u i ( y)| É | x − y| g i ( x) + g i ( y) (3.7)
for all x, y ∈ Rn \ F i with |F i | = 0, i = 1, 2, . . . Let A ⊂ Rn be such that u i ( x) → u( x)

and g i ( x) → g( x) for all x ∈ Rn \ A and | A | = 0. Write F = A ∪ ∞
S
F . Then |F | = 0.
i =1 i
Let x, y ∈ Rn \ F , x 6= y. From (3.7) we obtain
¡ ¢
| u( x) − u( y)| É | x − y| g( x) + g( y)
and thus g ∈ D ( u) ∩ L p (Rn ). This completes the proof. ä
3.5 Pointwise estimates

In this section we revisit pointwise inequalities for Sobolev functions.
Definition 3.33. Let 0 < β < ∞ and R > 0. The fractional sharp maximal function
of a locally integrable function f is defined by
×
f β#,R ( x) = sup r −β | f − f B( x,r) | d y,
0< r <R B( x,r )
If R = ∞ we simply write f β# ( x).
T HE MORAL : The fractional sharp maximal function controls the mean

oscillation of the function instead of the average of the function as in the Hardy-
Littlewood maximal function.
Next we prove a more general pointwise inequality than in Theorem 3.27.

Lemma 3.34. Suppose that f is locally integrable and let 0 < β < ∞. Then there
is c = c(β, n) and a set E with |E | = 0 such that
| f ( x) − f ( y)| É c| x − y|β f β#,4| x− y| ( x) + f β#,4| x− y| ( y)

¡ ¢
(3.8)
for every x, y ∈ Rn \ E .
THE MORAL : This is a pointwise inequality for a function without the gradi-
ent.
Proof. Let E be the complement of the set of Lebesgue points of f . By Lebesgue’s

theorem |E | = 0. Fix x ∈ Rn \ E , 0 < r < ∞ and denote B i = B( x, 2− i r ), i = 0, 1, . . .
Then
∞
X
| f ( x) − f B( x,r) | É | f B i+1 − f B i |
i =0
X∞ |B i |
×
É | f − fBi | d y
i =0 |B i +1 | B i
∞ ×
É c (2− i r )β (2− i r )−β
X
| f − fBi | d y
i =0 Bi
É cr β f β#,r ( x).
Let y ∈ B( x, r ) \ E . Then B( x, r ) ⊂ B( y, 2 r ) and we obtain
| f ( y) − f B( x,r) | É | f ( y) − f B( y,2r) | + | f B( y,2r) − f B( x,r) |

×
É cr β f β#,2r ( y) + | f − f B( y,2r) | dz
B( x,r )
×
É cr β f β#,2r ( y) + c | f − f B( y,2r) | dz
B( y,2 r )
É cr β f β#,2r ( y).
Let x, y ∈ Rn \ E , x 6= y and r = 2| x − y|. Then x, y ∈ B( x, r ) and hence
| f ( x) − f ( y)| É | f ( x) − f B( x,r) | + | f ( y) − f B( x,r) |

É c| x − y|β f β#,4| x− y| ( x) + f β#,4| x− y| ( y) .
¡ ¢
Remark 3.35. Lemma 3.34 gives a Campanato type characterization for Hölder
continuity. Let 0 < β É 1 and assume that f ∈ L1loc (Rn ) with f β# ∈ L∞ (Rn ). In other
words, there exists a constant M < ∞ such that
×
r −β | f − f B( x,r) | d y É M
B( x,r )
for every ball B( x, r ) ⊂ Rn . By Lemma 3.34, there exists a set E ⊂ Rn with |E | = 0

such that
| f ( x) − f ( y)| É c( n, β)| x − y|β f β# ( x) + f β# ( y)
¡ ¢
for every x, y ∈ Rn \ E . This implies that
| u( x) − u( y)| É c( n, β)k f β# kL∞ (Rn ) | x − y|β ,
for every x, y ∈ Rn \ E with |E | = 0. In other words, if f β# ∈ L∞ (Rn ), then f can be

redefined on a set of measure zero so that the function is Hölder continuous in Rn
with exponent β. On the other hand, if f ∈ C 0,β (Rn ), then
¯ × ¯
¯ ¯
| f ( y) − f B( x,r) | = ¯¯ f ( y) − f ( z) dz¯¯
B( x,r )
×
É | f ( y) − f ( z)| dz É cr β
B( x,r )
for every y ∈ B( x, r ). Thus

×
f β#,R ( x) = sup r −β | f ( y) − f B( x,r) | d y É c
0< r <R B( x,r )
for every x ∈ Rn and this implies that f β# ∈ L∞ (Rn ). Thus f can be redefined on a
set of measure zero so that the function is Hölder continuous with exponent β if
and only if f β# ∈ L∞ (Rn ). In the limiting case β = 0 we obtain the space of bounded
mean oscillation BMO (Rn ), which consists of functions f ∈ L1loc (Rn ) satisfying
f 0# ∈ L∞ (Rn ).
Definition 3.36. Let 0 É α < n and R > 0. The fractional maximal function of
f ∈ L1loc (Rn ) is
×
Mα,R f ( x) = sup r α | f | d y,
0< r <R B( x,r )
For R = ∞, we write Mα,∞ = Mα . If α = 0, we obtain the Hardy–Littlewood
maximal function and we write M0 = M .
1,1
If u ∈ Wloc (Rn ), then by the Poincaré inequality with p = 1, see Theorem 3.20,
there is c = c( n) such that
× ×
| u − u B( x,r) | d y É cr |Du| d y
B( x,r ) B( x,r )
for every ball B( x, r ) ⊂ Rn . It follows that

× ×
α−1 α
r | u − u B( x,r) | d y É cr |Du| d y
B( x,r ) B( x,r )
and consequently
u#1−α,R ( x) É cMα,R |Du|( x)
for every x ∈ Rn and R > 0. Thus we have proved the following useful inequality.
1,1
Corollary 3.37. Let u ∈ Wloc (Rn ) and 0 É α < 1. Then there is c = c( n, α) and a
set E ⊂ Rn with |E | = 0 such that
| u( x) − u( y)| É c| x − y|1−α Mα,4| x− y| |Du|( x) + Mα,4| x− y| |Du|( y)

¡ ¢
for every x, y ∈ Rn \ E .
Remark 3.38. Corollary 3.37 gives a Morrey type condition for Hölder conti-
nuity. Compare to Remark 3.35, where Hölder continuity was characterized
1,1
by a Campanato approach. Let 0 É α < 1 and assume that u ∈ Wloc (Rn ) with
Mα |∇ u| ∈ L∞ (Rn ). In other words, there exists a constant M < ∞ such that
×
rα |Du| d y É M
B( x,r )
for every ball B( x, r ) ⊂ Rn . By Corollary 3.37, there exists E ⊂ Rn with |E | = 0 such

that
| u( x) − u( y)| É c( n, α)| x − y|1−α Mα |Du|( x) + Mα |Du|( y)
¡ ¢
(3.9)
for every x, y ∈ Rn \ E . This implies that
| u( x) − u( y)| É c( n, α)k Mα |Du|kL∞ (Rn ) | x − y|1−α ,
for every x, y ∈ Rn \ E with |E | = 0. In other words, if Mα |∇ u| ∈ L∞ (Rn ) then u can

be redefined on a set of measure zero so that the function is Hölder continuous in
Rn with exponent 1 − α. This shows that u is Hölder continuous with the exponent
1 − α, after a possible redefinition on a set of measure zero.
Remark 3.39. From (3.9) we recover Morrey’s inequality in Theorem 2.18. To see
this, assume that u ∈ W 1,p (Rn ) with n < p < ∞. By Hölder’s inequality we have
¢1
M n |Du|( x) É c( n) M n |Du| p ( x) p É c( n)kDukL p (Rn ) < ∞,
¡
p
for every x ∈ Rn . Thus (3.9), with α = np , implies
1− np
| u( x) − u( y)| É c( n, p)kDukL p (Rn ) | x − y|
for every x, y ∈ Rn \ E with |E | = 0. This shows that u is Hölder continuous with

the exponent 1 − np after a possible redefinition on a set of measure zero.
The next result shows that this gives a characterization of W 1,p (Rn ) for 1 <
p É ∞.
Theorem 3.40. Let 1 < p < ∞. Then the following four conditions are equivalent.
(1) u ∈ W 1,p (Rn ).

(2) u ∈ L p (Rn ) and there is g ∈ L p (Rn ), g Ê 0, such that
| u( x) − u( y)| É | x − y|( g( x) + g( y))
for every x, y ∈ Rn \ E with |E | = 0.

(3) u ∈ L p (Rn ) and there is g ∈ L p (Rn ), g Ê 0, such that the Poincaré inequality
× ×
| u − u B( x,r) | d y É c r gdy
B( x,r ) B( x,r )
holds for every x ∈ Rn and r > 0.

(4) u ∈ L p (Rn ) and u#1 ∈ L p (Rn ).
Proof. (1) We have already seen that (1) implies (2).

(2) To prove that (2) implies (3), we integrate the pointwise inequality twice
over the ball B( x, r ). After the first integration we obtain
¯ × ¯
¯ ¯
| u( y) − u B( x,r) | = ¯¯ u( y) − u( z) dz¯¯
B( x,r )
×
É | u( y) − u( z)| dz
B( x,r )
µ × ¶
É 2 r g( y) + g( z) dz
B( x,r )
from which we have

× µ× × ¶
| u( y) − u B( x,r) | d y É 2 r g( y) d y + g( z) dz
B( x,r ) B( x,r ) B( x,r )
×
É 4r g( y) d y.
B( x,r )
(3) To show that (3) implies (4) we observe that
1
× ×
u#1 ( x) = sup | u − u B( x,r) | d y É c sup g d y = cM g( x).
r >0 r B( x,r ) r >0 B( x,r )
(4) Then we show that (4) implies (1). By Lemma 3.34
| u( x) − u( y)| É c| x − y|( u#1 ( x) + u#1 ( y))
for every x, y ∈ Rn \ E with |E | = 0. If we denote g = cu#1 , then g ∈ L p (Rn ) and
| u( x) − u( y)| É | x − y|( g( x) + g( y))
for every x, y ∈ Rn \ E with |E | = 0. Then we use the characterization of Sobolev

spaces W 1,p (Rn ), 1 < p < ∞, with integrated difference quotients, see Theorem
1.46. Let h ∈ Rn . Then
| u h ( x) − u( x)| = | u( x + h) − u( x)| É | h|( g h ( x) + g( x)),
from which we conclude that
k u h − ukL p (Rn ) É | h|(k g h kL p (Rn ) + k gkL p (Rn ) ) = 2| h|k gkL p (Rn ) .
The claim follows from this. ä
Remark 3.41. It can be shown that u ∈ W 1,1 (Rn ) if and only if u ∈ L1 (Rn ) and there
is a nonnegative function g ∈ L1 (Rn ) and σ Ê 1 such that
¡ ¢
| u( x) − u( y)| É | x − y| Mσ| x− y| g( x) + Mσ| x− y| g( y)
for every x, y ∈ Rn \ E with |E | = 0. Moreover, if this inequality holds, then |Du| É

c( n, σ) g almost everywhere.
3.6 Approximation by Lipschitz functions

Smooth functions in C ∞ (Ω) and C 0∞ (Ω) are often used as canonical test functions
in mathematical analysis. However, in many occasions smooth functions can be
replaced by a more flexible class of Lipschitz functions. One highly useful property
of Lipschitz functions, not shared by the smooth functions, is that the pointwise
minimum and maximum over L-Lipschitz functions are still L-Lipschitz. The
same is in fact true also for pointwise infimum and supremum of L-Lipschitz
functions, if these are finite at a single point. In particular, it follows that if
u : A → R is an L-Lipschitz function, then the truncations max{ u, c} and min{ u, c}
with c ∈ R are L-Lipschitz.
Theorem 3.42 (McShane). Assume that A ⊂ Rn , 0 É L < ∞ and that f : A → R

is an L-Lipschitz function. There exists an L-Lipschitz function f ∗ : Rn → R such
that f ∗ ( x) = f ( x) for every x ∈ A .
T HE MORAL : Every Lipschitz continuous function defined on a subset A of

Rn can be extended as a Lipschitz continuous function to the whole Rn .
Proof. Define f ∗ : Rn → R,
f ∗ ( x) = inf f (a) + L| x − a| : a ∈ A .
© ª
We claim that f ∗ ( b) = f ( b) for every b ∈ A . To see this we observe that
f ( b ) − f ( a ) É | f ( b ) − f ( a )| É L | b − a | ,
which implies f ( b) É f (a) + L| b − a| for every a ∈ A . By taking infimum over a ∈ A

we obtain f ( b) É f ∗ ( b). On the other hand, by the definition f ∗ ( b) É f ( b) for every
b ∈ A . Thus f ∗ ( b) = f ( b) for every b ∈ A .
Then we claim that f ∗ is L-Lipschitz in Rn . Let x, y ∈ Rn . Then
f ∗ ( x) = inf f (a) + L| x − a| : a ∈ A
© ª
© ª
É inf f (a) + L(| y − a| + | x − y|) : a ∈ A
© ª
É inf f (a) + L| y − a| : a ∈ A + L| x − y|
= f ∗ ( y) + L | x − y| .
By switching the roles of x and y, we arrive at f ∗ ( y) É f ∗ ( x) + L| x − y|. This implies

that −L| x − y| É f ∗ ( x) − f ∗ ( y) É L| x − y|. ä
Remark 3.43. The function f ∗ : Rn → R,

© ª
f ∗ ( x) = sup f (a) − L| x − a| : a ∈ A .
is an L-Lipschitz extension of f as well. We can see, that f ∗ is the largest L-

Lipschitz extension of f in the sense that if g : Rn → R is L-Lipschitz and g| A = f ,
then g É f ∗ . Correspondingly, the function f ∗ is the smallest L-Lipschitz extension
of f .
Since C 0∞ (Rn ) is dense in W 1,p (Rn ), also compactly supported Lipschitz func-
tions are dense in W 1,p (Rn ). By Theorem 3.27, we give a quantitative density
result for Lipschitz functions in W 1,p (Rn ). The main difference of the following
result to the standard mollification approximation u ε → u as ε → 0 is that the
value of the function is not changed in a good set { x ∈ Rn : u ε ( x) = u( x)} and there
is an estimate for the measure of the bad set { x ∈ Rn : u ε ( x) 6= u( x)}.
Theorem 3.44. Assume that u ∈ W 1,p (Rn ), 1 0 there
exists a Lipschitz continuous function u ε : Rn → R such that
(1) |{ x ∈ Rn : u ε ( x) 6= u( x)}| < ε and

(2) k u − u ε kW 1,p (Rn ) < ε.
Proof. Let E λ = { x ∈ Rn : M |Du|( x) É λ}, λ > 0. We show that u is cλ-Lipschitz in

E λ . By Theorem 3.27
| u( x) − u( y)| É c| x − y|( M |Du|( x) + M |Du|( y)) É cλ| x − y|
for almost every x, y ∈ E λ . The McShane extension theorem allows us to find a

cλ-Lipschitz extension vλ : Rn → R. We truncate vλ and obtain a 2 cλ-Lipschitz
function
u λ = max{−λ, min{vλ , λ}}.
Observe that | u λ | É λ in Rn and u λ = u almost everywhere in E λ .

(1) We consider measure of the set
Rn \ E λ = { x ∈ Rn : M |Du|( x) > λ}.
There exits c = c( n, p) such that

ˆ
λ p |{ x ∈ Rn : M |Du|( x)| > λ}| É c |Du( x)| p dx → 0
{ x∈Rn :|Du( x)|> λ2 }
as λ → ∞, since Du ∈ L p (Rn ). This follows by choosing f = |Du| in the following

general fact for the Hardy-Littlewood maximal function.
Claim: If f ∈ L p (Rn ), then there exists c = c( n, p) such that
ˆ
c
|{ x ∈ Rn : M f ( x) > λ}| É | f ( x)| p dx, λ > 0.
λ { x∈Rn :| f ( x)|> λ2 }
Reason. Let f = f 1 + f 2 , where f 1 = f χ{| f |> λ } and f 2 = f χ{| f |É λ } . Then

2 2
ˆ ˆ µ ¶1− p
λ p
| f 1 ( x)| dx = | f 1 ( x)| p | f 1 ( x)|1− p dx É k f kL p (Rn ) < ∞.
Rn { x∈Rn :| f ( x)|> λ } 2
2
λ
This shows that f 1 ∈ L1 (Rn ). On the other hand, | f 2 ( x)| É 2 for every x ∈ Rn ,
λ
which implies k f 2 kL∞ (Rn ) É 2 and f 2 ∈ L∞ (Rn ). Thus every L p function can be
1
represented as a sum of an L function and an L∞ function. By Lemma 3.2, we
have
λ
k M f 2 kL∞ (Rn ) É k f 2 kL∞ (Rn ) É
2
From this we conclude using sublinearity of the maximal operator that
λ
M f ( x) = M ( f 1 + f 2 )( x) É M f 1 ( x) + M f 2 ( x) É M f 1 ( x) +
2
for every x ∈ Rn and thus M f ( x) > λ implies M f 1 ( x) > λ2 . It follows that
λ ¯¯
¯½ ¾¯
n n
|{ x ∈ R M f ( x) > λ}| É ¯ x ∈ R M f 1 ( x) >
¯
: ¯ :
2 ¯
for every λ > 0.

p = 1 By the maximal function theorem on L1 (Rn ), see (3.1), we have
ˆ
¯ x ∈ R n : M f 1 ( x) > λ ¯ É c k f 1 k 1 n = c
¯½ ¾¯
¯ ¯
L (R ) | f ( x)| dx
¯ 2 ¯ λ λ { x∈Rn :| f ( x)|> λ }
2
for every λ > 0.

1 1, see (3.2), we have
¾¯ µ ¶ p ˆ ˆ
¯ x ∈ R n : M f 1 ( x) > λ ¯ É 2
¯½
p c
| f 1 ( x)| p dx
¯ ¯
( M f 1 ( x )) dx É
¯ 2 ¯ λ Rn λ p Rn
ˆ
c
= p | f ( x)| dx
λ { x∈Rn :| f ( x)|> λ }
2
for every λ > 0. ■
Thus we conclude that
λ p |Rn \ E λ | É λ p |{ x ∈ Rn : M |Du|( x)| > λ}|

ˆ
Éc |Du( x)| p dx
{ x∈Rn :|Du( x)|> λ2 }
and consequently λ p |Rn \ E λ | → 0 and |Rn \ E λ | → 0 as λ → ∞.

(2) Next we prove an estimate for k u − u ε kW 1,p (Rn ) . Since u λ = u in E λ and
| u λ | É λ in Rn , we have
ˆ
p
k u λ − u k L p (R n ) = | u λ − u| p dx
Rn \E λ
µˆ ˆ ¶
É 2p | u λ | p dx + | u| p dx
Rn \E Rn \E λ
µ λ
ˆ ¶
p p n
É 2 λ |R \ E λ | + | u| p dx → 0
Rn \E λ
as λ → ∞.
To prove the corresponding estimate for the gradients, we note that
D ( u λ − u) = χRn \E λ D ( u λ − u) = χRn \E λ Du λ − χRn \E λ Du
almost everywhere. Recall that u λ is cλ-Lipschitz and thus |Du λ | É cλ almost

everywhere.
ˆ
p
kD ( u λ − u)kL p (Rn ) = |D ( u λ − u)| p dx
Rn \E λ
µˆ ˆ ¶
É 2p |Du λ | p dx + |Du| p dx
Rn \E Rn \E λ
µ λ
ˆ ¶
p p n
É 2 (2 cλ) |R \ E λ | + |Du| p dx → 0
Rn \E λ
as λ → ∞. Thus k u − u λ kW 1,p (Rn ) → 0 as λ → ∞. Observe that
{ x ∈ Rn : u( x) 6= u λ ( x)} ⊂ Ω \ E λ ,
with |Rn \ E λ | → 0 as λ → ∞. This proves the claims. ä
Remark 3.45. Let E λ = { x ∈ Rn : M |Du|( x) É λ}, λ > 0. Let Q i , i = 1, 2, . . . be a

Whitney decomposition of Rn \ E λ with the following properties: each Q i is open,
cubes Q i , i = 1, 2, . . ., are disjoint, Rn \ E λ = ∪∞ Q , 4Q i ⊂ Rn \ E λ , i = 1, 2, . . .,
i =1 i
∞
X
χ2Q i É N < ∞,
i =1
and
c 1 dist(Q i , E λ ) É diam(Q i ) É c 2 dist(Q i , E λ )
for some constants c 1 and c 2 .

Then we construct a partition of unity associated with the covering 2Q i ,
i = 1, 2, . . . This can be done in two steps. First, let ϕ i ∈ C 0∞ (2Q i ) be such that
0 É ϕ i É 1, ϕ i = 1 in Q i and
c
|D ϕ i | É ,
diam(Q i )
for i = 1, 2, . . . Then we define
ϕ i ( x)
φ i ( x ) = P∞
j =1 ϕ j ( x)
for every i = 1, 2, . . .. Observe that the sum is over finitely many terms only since
ϕ i ∈ C 0∞ (2Q i ) and the cubes 2Q i , i = 1, 2, . . ., are of bounded overlap. The functions
φ i have the property
∞
X
φ i ( x) = χRn \E λ ( x)
i =1
for every x ∈ Rn .
Then we define the function u λ by


 u( x), x ∈ E ,
λ
u λ ( x) = P
 ∞ φ ( x) u ,
i =1 i 2Q i x ∈ Rn \ E λ .
The function u λ is a Whitney type extension of u|E λ to the set Rn \ E λ .

First we claim that
k u λ kW 1,p (Rn \E λ ) É ck ukW 1,p (Rn \E λ ) . (3.10)
Since the cubes 2Q i , i = 1, 2, . . ., are of bounded overlap, we have

ˆ ˆ ¯X∞ ¯p ∞
ˆ
p
| u 2Q i | p dx
X
| u λ | dx = φ i ( x) u 2Q i ¯ dx É c
¯ ¯
¯
Rn \E λ Rn \E λ i =1 i =1 2Q i
∞ × ˆ
| u| p dx É c | u| p dx.
X
Éc |2Q i |
i =1 2Q i Rn \E λ
Then we consider an estimate for the gradient. We recall that

∞
Φ( x ) =
X
φ i ( x) = 1
i =1
for every x ∈ Rn \ E λ . Since the cubes 2Q i , i = 1, 2, . . ., are of bounded overlap, we

see that Φ ∈ C ∞ (Rn \ E λ ) and
∞
D j Φ( x ) =
X
D j φ i ( x) = 0, j = 1, 2, . . . , n,
i =1
for every x ∈ Rn \ E λ . Hence we obtain

¯X∞ ¯ ¯X ∞ ¯
|D j u λ ( x)| = ¯ D j φ i ( x) u 2Q i ¯ = ¯ D j φ i ( x)( u( x) − u 2Q i )¯
¯ ¯ ¯ ¯
i =1 i =1
∞
−1
X
Éc diam(Q i ) | u( x) − u 2Q i |χ2Q i ( x)
i =1
and consequently
∞
|D j u λ ( x)| p É c diam(Q i )− p | u( x) − u 2Q i | p χ2Q i ( x).
X
i =1
Here we again used the fact that the cubes 2Q i , i = 1, 2, . . ., are of bounded overlap.
This implies that for every j = 1, 2, . . . , n,
ˆ ˆ ∞
³X ´
p
|D j u λ | dx É c diam(Q i )− p | u − u 2Q i | p χ2Q i dx
Rn \E λ Rn \E λ i =1
∞
ˆ
diam(Q i )− p | u − u 2Q i | p dx
X
É
i =1 2Q i
∞
ˆ ˆ
|Du| p dx É c |Du| p dx.
X
Éc
i =1 2Q i Rn \E λ
Then we show that u λ ∈ W 1,p (Rn ). We know that u λ ∈ W 1,p (Rn \ E λ ) and that it
is Lipschitz continuous in Rn (exercise). Moreover u ∈ W 1,p (Rn ) and u = u λ in E λ
by (i). This implies that w = u − u λ ∈ W 1,p (Rn \ E λ ). and that w = 0 in E λ . By the
ACL property, u is absolutely continuous on almost every line segment parallel to
the coordinate axes. Take any such line. Now w is absolutely continuous on the
part of the line segment which intersects Rn \ E λ . On the other hand w = 0 in the
complement of E λ . Hence the continuity of w in the line segment implies that w
is absolutely continuous on the whole line segment.
We have
k u − u λ kW 1,p (Rn ) = k u − u λ kW 1,p (Rn \E λ )

É k ukW 1,p (Rn \E λ ) + k u λ kW 1,p (Rn \E λ )
É ck ukW 1,p (Rn \E λ ) .
3.7 Maximal operator on Sobolev spaces

Assume that u is Lipschitz continuous with constant L, that is
| u h ( y) − u( y)| = | u( y + h) − u( y)| É L| h|
for every y, h ∈ Rn , where we denote u h ( y) = u( y + h). Since the maximal function

commutes with translations and the maximal operator is sublinear, we have
|( Mu)h ( x) − Mu( x)| = | M ( u h )( x) − Mu( x)|

É M ( u h − u)( x)
ˆ
1
= sup | u h ( y) − u( y)| d y
r >0 |B( x, r )| B( x,r )
É L | h |.
This means that the maximal function is Lipschitz continuous with the same
constant as the original function if Mu is not identically infinity. Observe, that
this proof applies to Hölder continuous functions as well.
Next we show that the Hardy-Littlewood maximal operator is bounded in
Sobolev spaces.
Theorem 3.46. Let 1 1, is a bounded operator. Thus

the maximal operator is not only bounded on L p (Rn ) but also on W 1,p (Rn ) for p > 1.
Proof. The proof is based on the characterization of W 1,p (Rn ) by integrated dif-
ference quotients, see Theorem 1.46. By the maximal function theorem with
1 < p < ∞, see (3.2), we have Mu ∈ L p (Rn ) and
k( Mu)h − MukL p (Rn ) = k M ( u h ) − MukL p (Rn )

É k M ( u h − u)kL p (Rn )
É c k u h − u k L p (R n )
É ckDukL p (Rn ) | h|
for every h ∈ Rn . Theorem 1.46 gives Mu ∈ W 1,p (Rn ) with
kDMukL p (Rn ) É ckDukL p (Rn ) .
Thus by the maximal function theorem

³ ´1
p p p
k MukW 1,p (Rn ) = k MukL p (Rn ) + kDMukL p (Rn )
É k MukL p (Rn ) + kDMukL p (Rn )

¡ ¢
É c k ukL p (Rn ) + kDukL p (Rn )
É ck ukW 1,p (Rn ) . ä
A more careful analysis gives a pointwise estimate for the partial derivatives.
Theorem 3.47. Let 1 < p < ∞. If u ∈ W 1,p (Rn ), then Mu ∈ W 1,p (Rn ) and
|D j Mu| É M (D j u), j = 1, 2, . . . , n, (3.12)
almost everywhere in Rn .
THE MORAL : Differentiation commutes with a linear operator. The sublinear

maximal operator semicommutes with differentiation.
Proof. If χB(0,r) is the characteristic function of B(0, r ) and

χB(0,r)
χr = ,
|B(0, r )|
then
ˆ ˆ
1 1
| u( y)| d y = | u( x − y)| d y
|B( x, r )| B( x,r ) |B(0, r )| B(0,r )
ˆ
1
= χB(0,r) | u( x − y)| d y
|B(0, r )| Rn
= (| u| ∗ χr )( x),
where ∗ denotes the convolution. Now | u| ∗ χr ∈ W 1,p (Rn ) and by Theorem 1.17 (1)
D j (| u | ∗ χ r ) = χ r ∗ D j | u | , j = 1, 2, . . . , n,
almost everywhere in Rn .
Let r m , m = 1, 2, . . . , be an enumeration of positive rationals. Since u is locally
integrable, we may restrict ourselves to the positive rational radii in the definition
of the maximal function. Hence
Mu( x) = sup(| u| ∗ χr m )( x).

m
We define functions vk : Rn → R, k = 1, 2, . . . , by
vk ( x) = max (| u| ∗ χr m )( x).
1É mÉ k
Now (vk ) is an increasing sequence of functions in W 1,p (Rn ), which converges to

Mu pointwise and
|D j vk | É max ¯D j (| u| ∗ χr m )¯
¯ ¯
1É mÉ k
= max ¯χr m ∗ D j | u|¯

¯ ¯
1É mÉ k
É M ( D j | u | ) = M ( D j u ), j = 1, 2, . . . , n,
almost everywhere in Rn . Here we also used Remark 1.27 and the fact that by
Theorem 1.26
¯ ¯
¯ D j | u |¯ = | D j u |, j = 1, 2, . . . , n,
almost everywhere. Thus

n
X n
X
kDvk kL p (Rn ) É k D j v k k L p (R n ) É k M (D j u)kL p (Rn )
j =1 j =1
and the maximal function theorem implies

n
X
kvk kW 1,p (Rn ) É k MukL p (Rn ) + k M (D j u)kL p (Rn )
j =1
n
X
É c k u k L p (R n ) + c k D j u k L p (R n ) É c < ∞
j =1
for every k = 1, 2, . . . Hence (vk ) is a bounded sequence in W 1,p (Rn ) which converges
to Mu pointwise. Theorem 1.41 implies Mu ∈ W 1,p (Rn ), vk → Mu weakly in L p (Rn )
and D j vk → D j Mu weakly in L p (Rn ) as k → ∞.
Next we prove the pointwise estimate for the gradient. By Mazur’s lemma, see
Theorem 1.33, there is a sequence of convex combinations such that
m
Xk
wk = a k,l D j vl → D j Mu, j = 1, . . . , n,
l =k
in L p (Rn ) as k → ∞. There is a subsequence of (wk ) which converges almost

everywhere to D j Mu. Thus we have
m
Xk ¯ ¯ m Xk
|wk | É a k,l ¯D j vl ¯ É a k,l M (D j u) = M (D j u)
l =k l =k
for every l = 1, 2, . . . and finally
|D j Mu| = lim |wk | É M (D j u), j = 1, . . . , n,

k→∞
almost everywhere in Rn . This completes the proof. ä
Remarks 3.48:
(1) Estimate (3.11) also follows from (3.12). To see this, we may use the
maximal function theorem, see (3.2), and (3.12) to obtain
k MukW 1,p (Rn ) É k MukL p (Rn ) + kDMukL p (Rn )

É ck ukL p (Rn ) + k M |Du|kL p (Rn )
É ck ukW 1,p (Rn ) ,
where c is the constant in (3.2).

(2) If u ∈ W 1,∞ (Rn ), then a slight modification of our proof shows that Mu
belongs to W 1,∞ (Rn ). Moreover,
k MukW 1,∞ (Rn ) = k MukL∞ (Rn ) + kDMukL∞ (Rn )

É k ukL∞ (Rn ) + k M |Du|kL∞ (Rn )
É k ukW 1,∞ (Rn ) .
Hence in this case the maximal operator is bounded with constant one.
Recall, that after a redefinition on a set of measure zero u ∈ W 1,∞ (Rn ) is a
bounded and Lipschitz continuous function, see Theorem 2.23.
Pointwise behaviour of Sobolev
4
functions
In this chapter we study fine properties of Sobolev functions. By definition, Sobolev

functions are defined only up to Lebesgue measure zero and thus it is not always
clear how to use their pointwise properties to give meaning, for example, to
boundary values.
4.1 Sobolev capacity

Capacities are needed to understand pointwise behavior of Sobolev functions.
They also play an important role in studies of solutions of partial differential
equations.
Definition 4.1. For 1 < p < ∞, the Sobolev p-capacity of a set E ⊂ Rn is defined
by
p
cap p (E ) = inf k ukW 1,p (Rn )
u∈A (E )
³ ´
p p
= inf k ukL p (Rn ) + kDukL p (Rn )
u∈A (E )
ˆ
¡ p
| u| + |Du| p dx,
¢
= inf
u∈A (E ) Rn
where A (E ) = u ∈ W 1,p (Rn ) : u Ê 1 on a neighbourhood of E . If A (E ) = ;, we set

© ª
cap p (E ) = ∞. Functions in A (E ) are called admissible functions for E .
T HE MORAL : Capacity measures the size of exceptional sets for Sobolev

functions. Lebesgue measure is the natural measure for functions in L p (Rn ) and
the Sobolev p-capacity is the natural outer measure for functions in W 1,p (Rn ).
106
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 107
Remark 4.2. In the definition of capacity we can restrict ourselves to the admissi-
ble functions u for which 0 É u É 1. Thus
p
cap p (E ) = inf k uk
W 1,p (Rn )
,
u∈A 0 (E )
where A 0 (E ) = u ∈ W 1,p (Rn ) : 0 É u É 1, u = 1 on a neighbourhood of E .

© ª
Reason. (1) Since A 0 (E ) ⊂ A (E ), we have
p
cap p (E ) É inf k uk
W 1,p (Rn )
.
u∈A 0 (E )
(2) For the reverse inequality, let ε > 0 and let u ∈ A (E ) such that
p
k uk É cap p (E ) + ε.
W 1,p (Rn )
Then v = max{0, min{ u, 1}} ∈ A 0 (E ), |v| É | u| and by Remark 1.27 we have |Dv| É
|Du| almost everywhere. Thus
p p p
inf k uk É k vk É k uk É cap p (E ) + ε
u∈A 0 (E ) W 1,p (Rn ) W 1,p (Rn ) W 1,p (Rn )
and by letting ε → 0 we obtain
p
inf k uk É cap p (E ).
u∈A 0 (E ) W 1,p (Rn )
Remarks 4.3:
(1) There are several alternative definitions for capacity and, in general, it
does not matter which one we choose. For example, when 1 < p < n, we
may consider the definition
ˆ
cap p (E ) = inf |Du| p dx,
Rn
where the infimum is taken over all u ∈ L p∗ (Rn ) with |Du| ∈ L p (Rn ), u Ê 0
and u Ê 1 on a neighbourhood of E . Some estimates and arguments may
become more transparent with this definition, but we stick to our original
definition.
(2) The definition of Sobolev capacity applies also for p = 1, but we shall not
discuss this case here.
The Sobolev p-capacity enjoys many desirable properties, one of the most
important of which says that it is an outer measure.
Theorem 4.4. The Sobolev p-capacity is an outer measure, that is,
(1) cap p (;) = 0,

(2) if E 1 ⊂ E 2 , then cap p (E 1 ) É cap p (E 2 ) and
n
Ei É ∞i =1 cap p (E i ) whenever E i ⊂ R , i = 1, 2, . . . .
¡ S∞ ¢ P
(3) cap p i =1
T HE MORAL : Capacity is an outer measure, but measure theory is useless

since there are very few measurable sets.
Proof. (1) Clearly cap p (;) = 0.

(2) A (E 2 ) ⊂ A (E 1 ) implies cap p (E 1 ) É cap p (E 2 ) .
(3) Let ε > 0. We may assume that ∞ i =1 cap p (E i ) < ∞. Choose u i ∈ A (E i ) so
P
that
p
ku i k É cap p (E i ) + ε2− i , i = 1, 2, . . . .
W 1,p (Rn )
S∞
Claim: v = sup i u i is admissible for i =1
Ei.
Reason. First we show that v ∈ W 1,p (Rn ). Let
vk = max u i , k = 1, 2, . . . .
1É i É k
Then (vk ) is an increasing sequence such that vk → v pointwise as k → ∞. More-

over
|vk | = | max u i | É | sup u i | = |v|, k = 1, 2, . . . ,
1É i É k i
and by Remark 1.27
|Dvk | É max |Du i | É sup |Du i |, k = 1, 2, . . . .

1É i É k i
We show that (vk ) is a a bounded sequence in W 1,p (Rn ). To conclude this, we

observe that
ˆ ˆ
p
kvk k = |vk | p dx + |Dvk | p dx
W 1,p (Rn )
Rn Rn
ˆ ˆ
p
É sup | u i | dx + sup |Du i | p dx
Rn i Rn i
ˆ ∞
ˆ ∞
| u i | p dx + |Du i | p dx
X X
É
Rn i =1 Rn i =1
∞ µˆ ˆ ¶
p p
X
= | u i | dx + |Du i | dx
i =1 Rn Rn
∞
(cap p (E i ) + ε2− i )
X
É
i =1
X∞
É cap p (E i ) + ε < ∞, k = 1, 2, . . . .
i =1
Since vk → v almost everywhere, by weak compactness of Sobolev spaces, see

Theorem 1.41, we conclude that v ∈ W 1,p (Rn ). Since u i ∈ A (E i ), there exists an
open set O i ⊃ E i such that u i Ê 1 on O i for every i = 1, 2, . . . . It follows that
v = sup i u i Ê 1 on ∞ O , which is a neighbourhood of ∞
S S
i =1 i
E .
i =1 i ■
We conclude that
∞ ∞ ∞
p X p X
cap p (E i ) + ε.
¡[ ¢
cap p E i É kvkW 1,p (Rn ) É k u i k 1,p n É
W (R )
i =1 i =1 i =1
The claim follows by letting ε → 0. ä
Remark 4.5. The Sobolev p-capacity is outer regular, that is,
cap p (E ) = inf{cap p (O ) : E ⊂ O, O open}.
Reason. (1) By monotonicity,
cap p (E ) É inf{cap p (O ) : E ⊂ O, O open}.
(2) To see the inequality in the other direction, let ε > 0 and take u ∈ A (E )
such that
p
W 1,p (Rn )
Since u ∈ A (E ) there is an open set O containing E such that u Ê 1 on O , which

implies
p
cap p (O ) É k ukW 1,p (Rn ) É cap p (E ) + ε.
The claim follows by letting ε → 0. ■
THE MORAL : The capacity of a set is completely determined by the capacities

of open sets containing the set. The same applies to the Lebesgue outer measure.
4.2 Capacity and measure

We are mainly interested in sets of vanishing capacity, since they are in some
sense exceptional sets in the theory Sobolev spaces. Our first result is rather
immediate.
Lemma 4.6. |E | É cap p (E ) for every E ⊂ Rn .
T HE MORAL : Sets of capacity zero are of measure zero. Thus capacity is a

finer measure than Lebesgue measure.
Proof. If cap p (E ) = ∞, there is nothing to prove. Thus we may assume that

cap p (E ) < ∞. Let ε > 0 and take u ∈ A (E ) such that
p
W 1,p (Rn )
There is an open O ⊃ E such that u Ê 1 in O and thus

ˆ
p p
| E | É |O | É | u| p dx É k ukL p (Rn ) É k uk É cap p (E ) + ε.
W 1,p (Rn )
O
The claim follows by letting ε → 0. ä

Remark 4.7. Lemma 4.6 shows that cap p (B( x, r )) > 0 for every x ∈ Rn , r > 0. This
implies that capacity is nontrivial in the sense that every nonempty open set has
positive capacity.
Lemma 4.8. Let x ∈ Rn and 0 < r É 1. Then there exists c = c( n, p) such that
cap p (B( x, r )) É cr n− p
THE MORAL : For the Lebesgue measure of a ball we have |B( x, r )| É cr n , but
for the Sobolev capacity of a ball we have cap p (B( x, r )) É cr n− p . Thus the natural
scaling dimension for capacity is n − p. Observe, that the dimension for capacity
is smaller than n − 1.
Proof. Define a cutoff function




 1, y ∈ B( x, r ),

u( y) = 2 − | y−r x| , y ∈ B( x, 2 r ) \ B( x, r ),


0, y ∈ Rn \ B( x, 2 r ).

Observe that 0 É u É 1, u is 1r -Lipschitz and |Du| É 1

r almost everywhere. Thus
u ∈ A (B( x, r )) and
ˆ ˆ
cap p (B( x, r )) É | u( y)| p d y + |Du( y)| p d y
B( x,2 r ) B( x,2 r )
É (1 + r − p )|B( x, 2 r )| É ( r − p + r − p )|B( x, 2 r )|
= 2 r − p |B( x, 2 r )| = cr n− p ,
with c = c( n, p) ä
Remarks 4.9:
(1) Lemma 4.8 shows that every bounded set has finite capacity. Thus there
are plenty of sets with finite capacity.
Reason. Assume that E ⊂ Rn is bounded. Then E ⊂ B(0, r ) for some r ,

0 < r É 1, and
cap p (E ) É cap p (B(0, r )) É cr n− p < ∞. ■
(2) Lemma 4.8 implies that cap p ({ x}) = 0 for every x ∈ Rn when 1 < p < n.
Reason.
cap p ({ x}) É cap p (B( x, r )) É cr n− p , 0 < r É 1.
The claim follows by letting r → 0. ■
Remark 4.10. Let x ∈ Rn and 0 < r É 12 . Then there exists c = c( n) such that
1 1−n
µ ¶
capn (B( x, r )) É c log .
r
Reason. Use the test function

¡
1 −1
log | x−1 y| , y ∈ B( x, 1) \ B( x, r ),
¢
 log r



u( y) = 1, y ∈ B( x, r ),


0, y ∈ Rn \ B( x, 1).

This implies that cap p ({ x}) = 0 for every x ∈ Rn when p = n (exercise). ■
We have shown that a point has zero capacity when 1 n.
Lemma 4.11. If p > n, then cap p ({ x}) > 0 for every x ∈ Rn .
THE MORAL : For p > n every set containing at least one point has a positive
capacity. Thus there are no nontrivial sets of capacity zero. In practice this means
that capacity is a useful tool only when p É n.
Proof. Let x ∈ Rn and assume that u ∈ A ({ x}). Then there exists 0 < r É 1 such that
u( y) Ê 1 on B( x, r ). Take a cutoff function η ∈ C 0∞ (B( x, 2)) such that 0 É η É 1, η = 1
in B( x, r ) and |D η| É 2. By Morrey’s inequality, see Theorem 2.18, there exists
c = c( n, p) > 0 such that
1− np
|(η u)( y) − (η u)( z)| É c| y − z| kD (η u)kL p (Rn )
for almost every y, z ∈ Rn . Choose y ∈ B( x, r ) and z ∈ B( x, 4) \ B( x, 2) so that

(η u)( y) Ê 1 and (η u)( z) = 0. Then 1 É | y − z| É 5 and thus
ˆ
p
|D (η u)( y)| p d y = kD (η u)kL p (Rn )
B( x,2)
Ê c | y − z|n− p |(η u)( y) − (η u)( z)| p Ê c > 0.

| {z } | {z }
Ê5n− p Ê1
On the other hand

ˆ µˆ ˆ ¶
p p p p
|D (η u)( y)| d y É 2 |D η( y) u( y)| d y + |η( y)Du( y)| d y
B( x,2) B( x,2) B( x,2)
ˆ ˆ
 
= 2p  |D η( y)| p | u( y)| p d y + |η( y)| p |Du( y)| p d y

B( x,2) | {z } B( x,2) | {z }
É2 p É1
µˆ ˆ ¶
p p p
É4 | u( y)| d y + |Du( y)| d y
B( x,2) B( x,2)
p
É 4 p k uk .
W 1,p (Ω)
p
This shows that there exists c = c( n, p) > 0 such that k ukW 1,p (Ω) Ê c > 0 for every
u ∈ A ({ x}) and thus cap p ({ x}) Ê c > 0. ä
In order to study the connection between capacity and measure, we need to

consider lower dimensional measures than the Lebesgue measure. We recall the
definition of Hausdorff measures.
Definition 4.12. Let E ⊂ Rn and s Ê 0. For 0 < δ É ∞ we set

( )
∞ ∞
s s
H δ (E ) = inf
X
B ( x i , r i ), r i É δ .
[
ri : E ⊂
i =1 i =1
The (spherical) s-Hausdorff measure of E is
H s (E ) = lim H δs (E ) = sup H δs (E ).
δ →0 δ>0
The Hausdorff dimension of E is
inf s : H s (E ) = 0 = sup s : H s (E ) = ∞ .
© ª © ª
THE MORAL : The Hausdorff measure is the natural s-dimensional measure

up to scaling and the Hausdorff dimension is the measure theoretic dimension of
the set. Observe that the dimension can be any nonnegative real number less or
equal than the dimension of the space.
We begin by proving a useful measure theoretic lemma. In the proof we need

some tools from measure and integration theory and real analysis.
Lemma 4.13. Assume that 0 < s < n, f ∈ L1loc (Rn ) and let
ˆ
1
½ ¾
n
E = x ∈ R : lim sup s |f |dy > 0 .
r →0 r B( x,r )
Then H s (E ) = 0.
T HE MORAL : Roughly speaking the lemma above says that the set where
a locally integrable function blows up rapidly is of the corresponding Hausdorff
measure zero.
Proof. (1) Assume first that f ∈ L1 (Rn ).

(2) By the Lebesgue differentiation theorem
×
lim | f ( y)| d y = | f ( x)| < ∞,
r →0 B( x,r )
for almost every x ∈ Rn . If x is a Lebesgue point of | f |, then

ˆ
1
×
lim sup s | f ( y)| d y = c lim sup r n−s | f ( y)| d y = 0.
r →0 r B( x,r ) r →0 B( x,r )
This shows that all Lebesgue points of | f | belong to the complement of E and thus
|E | = 0.
(3) Let ε > 0 and

ˆ
1
½ ¾
n
E ε = x ∈ R : lim sup |f |dy > ε .
r →0 rs B( x,r )
Since E ε ⊂ E and |E | = 0, we have |E ε | = 0.
Claim: H s (E ε ) = 0 for every ε > 0.
Reason. Let 0 < δ < 1. For every x ∈ E ε there exists a r x with 0 < r x É δ such that
ˆ
1
| f | d y > ε.
r sx B( x,r x )
By the Vitali covering theorem, there exists a subfamily of countably many pair-
wise disjoint balls B( x i , r i ), i = 1, 2, . . . , such that
∞
[
Eε ⊂ B( x i , 5 r i ).
i =1
This gives
ˆ ˆ
∞ 5s X
∞ 5s
H 5sδ (E ε ) É s
X
(5 r i ) É |f |dy = | f | d y.
i =1 ε i=1 B( x i ,r i ) ε S∞
B( x i ,r i )
i =1
By disjointness of the balls

¯ ¯
∞ ∞
¯∞
¯[ ¯ X
r ni
X
¯ B( x, r i )¯ = |B( x i , r i )| = c
¯
¯ i=1 ¯ i=1 i =1
∞ rn
ˆ
i
X
Éc s |f |dy
i =1 ε r i B( x i ,r i )
n− s ˆ
δ
Éc |f |dy → 0 as δ → 0.
ε Rn
By absolute continuity of integral

ˆ
S∞
|f |dy → 0 as δ → 0.
i =1
B( x i ,r i )
Thus ˆ
s 5s
H (E ε ) = lim H 5sδ (E ε ) É lim | f | d y = 0.
δ →0 ε δ →0 S∞
B( x i ,r i )
i =1
This shows that H s (E ε ) = 0 for every ε > 0. ■
(4) By subadditivity of the Hausdorff measure

Ã !
∞ ∞
s s
H (E ) = H H s (E 1 ) = 0.
[ X
E1 É
k k
k=1 k=1
This shows that H s (E ) = 0.

(5) Assume then that f ∈ L1loc (Rn ). Then

ˆ
1
µ½ ¾¶
H s (E ) = H s x ∈ Rn : lim sup s |f |dy > 0
r →0 r B( x,r )
Ã
∞
ˆ ¾!
1
½
s n
=H x ∈ R lim sup s | f χB(0,k) | d y > 0
[
:
k=1 r →0 r B( x,r )
∞
ˆ
1
µ½ ¾¶
H s x ∈ Rn : lim sup s
X
É | f χB(0,k) | d y > 0 = 0. ä
k=1 r →0 r B( x,r )
Next we compare capacity to the Hausdorff measure.
Theorem 4.14. Assume that 1 < p < n. Then there exists c = c( n, p) such that
cap p (E ) É cH n− p (E ) for every E ⊂ Rn .
THE MORAL : Capacity is smaller than ( n− p)-dimensional Hausdorff measure.

In particular, H n− p (E ) = 0 implies cap p (E ) = 0.
Proof. Let B( x i , r i ), i = 1, 2, . . . , be any covering of E such that the radii satisfy

r i É δ. Subadditivity implies
∞ ∞
X X n− p
cap p (E ) É cap p (B( x i , r i )) É c ri .
i =1 i =1
By taking the infimum over all coverings by such balls and observing that H δs (E ) É
H s (E ) we obtain
n− p
cap p (E ) É cH δ (E ) É cH n− p (E ). ä
We next consider the converse of the previous theorem. We prove that sets of
p-capacity zero have Hausdorff dimension at most n − p.
Theorem 4.15. Assume that 1 n − p.
Proof. (1) Let E ⊂ Rn be such that cap p (E ) = 0. Then for every i = 1, 2, . . . , there is
p
u i ∈ A 0 (E ) such that k u i kW 1,p (Rn ) É 2− i . Define u = ∞
P
i =1 u i .
Claim: u ∈ A (E ).
Pk
Reason. Let vk = k = 1, 2, . . . . Then vk ∈ W 1,p (Rn ) and
i =1 u i ,
° °
°X k ° X k
kvk kW 1,p (Rn ) = ° u i ° É k u i kW 1,p (Rn )
° °
° i=1 ° 1,p n i =1
W (R )
∞ ∞
X X − pi
É k u i kW 1,p (Rn ) É 2 < ∞.
i =1 i =1
Thus (vk ) is a bounded sequence in W 1,p (Rn ). Since 0 É u i É 1, we observe that

(vk ) is an increasing sequence and thus vk → u almost everywhere. Theorem 1.41
implies u ∈ W 1,p (Rn ). Moreover, u Ê 1 almost everywhere on a neighbourhood of E
which shows that u ∈ A (E ). ■
(2) Claim:
×
lim sup u d y = ∞ for every x ∈ E. (4.1)
r →0 B( x,r )
Reason. Let m ∈ N and x ∈ E . Then for r > 0 small enough B( x, r ) is contained in

an intersection of open sets O i , i = 1, . . . , m, with the property that u i = 1 almost
everywhere on O i . This implies that u = ∞
P
i =1 u i Ê m almost everywhere in B( x, r )
and thus ×
u d y Ê m.
B( x,r )
This proves the claim. ■
T HE MORAL : This gives a method to construct a function that blows up on

any set of zero capacity.
(3) Claim: If s > n − p, then

ˆ
1
lim sup |Du| p d y = ∞ for every x ∈ E.
r →0 rs B( x,r )
Reason. Let x ∈ E and, for a contradiction, assume that

ˆ
1
lim sup s |Du| p d y < ∞.
r →0 r B( x,r )
Then there exists c < ∞ such that

ˆ
1
lim sup |Du| p d y É c.
r →0 rs B( x,r )
The we choose R > 0 so small that

ˆ
|Du| p d y É cr s
B( x,r )
for every 0 < r É R . Denote B i = B( x, 2− i R ), i = 1, 2, . . . . Then by Hölder’s inequality

and the Poincaré inequality, see Theorem 3.17, we have
×
| u B i+1 − u B i | É |u − uB i | d y
B i+1
|B i |
×
É |u − uB i | d y
|B i+1 | B i
µ× ¶1
p
p
Éc |u − uB i | d y
Bi
µ× ¶1
p
−i p
É c2 R |Du| d y
Bi
p − n+ s
−i
É c(2 R ) p .
For k > j , we obtain

kX
−1 kX
−1 p − n+ s
|uBk − uB j | É | u B i+1 − u B i | É c (2− i R ) p
i= j i= j
and thus ( u B i ) is a Cauchy sequence when s > n − p. This contradicts (4.1) and
thus the claim holds true. ■
(4) Thus
ˆ
1
½ ¾
E ⊂ x ∈ Rn : s |Du| p d y = ∞
r B( x,r)
ˆ
1
½ ¾
⊂ x ∈ Rn : s |Du| p d y > 0 .
r B( x,r)
Lemma 4.13 implies

ˆ
1
µ½ ¾¶
H s (E ) É H s x ∈ Rn : |Du| p d y > 0 = 0.
rs B( x,r )
s
This shows that H (E ) = 0 whenever n − p < s < n. The claim follows from this,
since H s (E ) = 0 implies H t (E ) = 0 for every t Ê s. ä
Remark 4.16. It can be shown that even H n− p (E ) < ∞, 1 < p < n, implies cap p (E ) =
0.
4.3 Quasicontinuity
In this section we study fine properties of Sobolev functions. It turns out that
Sobolev functions are defined up to a set of capacity zero.
Definition 4.17. We say that a property holds p-quasieverywhere, if it holds

except for a set of p-capacity zero.
THE MORAL : Quasieverywhere is a capacitary version of almost everywhere.
Recall that by Meyers-Serrin theorem 1.18 W 1,p (Rn )∩C (Rn ) is dense in W 1,p (Rn )
for 1 É p < ∞ and, by Theorem 1.13, the Sobolev space W 1,p (Rn ) is complete. The
next result gives a way to find a quasieverywhere converging subsequence.
Theorem 4.18. Assume that u i ∈ W 1,p (Rn ) ∩ C (Rn ), i = 1, 2, . . . , and that ( u i ) is a

Cauchy sequence in W 1,p (Rn ). Then there is a subsequence of ( u i ) that converges
pointwise p-quasieverywhere in Rn . Moreover, the convergence is uniform outside
a set of arbitrarily small p-capacity.
T HE MORAL : This is a Sobolev space version of the result that for every
Cauchy sequence in L p (Rn ), there is a subsequence that converges pointwise
almost everywhere. The claim concerning uniform convergence is a Sobolev space
version of Egorov’s theorem.
Proof. There exists a subsequence of ( u i ), which we still denote by ( u i ), such that

∞
p
2 i p k u i − u i+1 kW 1,p (Rn ) < ∞.
X
i =1
For i = 1, 2, . . . , denote
∞
E i = x ∈ Rn : | u i ( x) − u i+1 ( x)| > 2− i
© ª [
and Fj = Ei.
i= j
By continuity 2 i | u i − u i+1 | ∈ A (E i ) and thus

p
cap p (E i ) É 2 i p k u i − u i+1 kW 1,p (Rn ) .
By subadditivity we obtain
∞ ∞
p
2 i p k u i − u i+1 kW 1,p (Rn ) .
X X
cap p (F j ) É cap p (E i ) É
i= j i= j
Thus
Ã !
∞
\ ∞
\
cap p F j É lim cap p (F j ) ( F j ⊂ F j , F j+1 ⊂ F j , j = 1, 2, . . . )
j =1 j →∞ j =1
∞
p
2 i p k u i − u i+1 kW 1,p (Rn ) = 0.
X
É lim
j →∞ i = j
Here we used the fact that the tail of a convergent series tends to zero. We observe
that ( u i ) converges pointwise in Rn \ ∞
T
F . Moreover,
j =1 j
kX
−1 kX
−1
| u l ( x ) − u k ( x )| É | u i ( x) − u i+1 ( x)| É 2− i É 21−l
i=l i=l
for every x ∈ Rn \ F j for every k > l > j , which shows that the convergence is
uniform in Rn \ F j . ä
Definition 4.19. A function u : Rn → [−∞, ∞] is p-quasicontinuous in Rn if for

every ε > 0 there is a set E such that cap p (E ) < ε and the restriction of u to Rn \ E ,
denoted by u|Rn \E , is continuous.
Remark 4.20. By outer regularity, see Remark 4.5, we may assume that E is open
in the definition above.
The next result shows that a Sobolev function has a quasicontinous represen-
tative.
Corollary 4.21. For each u ∈ W 1,p (Rn ) there is a p-quasicontinuous function

v ∈ W 1,p (Rn ) such that u = v almost everywhere in Rn .
T H E M O R A L : Every L p function is defined almost everywhere, but every

W 1,p function is defined quasieverywhere.
Proof. By Theorem 1.18, for every function u ∈ W 1,p (Rn ), there are functions
u i ∈ W 1,p (Rn ) ∩ C (Rn ), i = 1, 2, . . . , such that u i → u in W 1,p (Rn ) as i → ∞. By
Theorem 4.18 there exists a subsequence that converges uniformly outside a set
of arbitrarily small capacity. Uniform convergence implies continuity of the limit
function and thus the limit function is continuous outside a set of arbitrarily small
p-capacity. This completes the proof. ä
Next we show that the quasicontinuous representative given by Corollary 4.21

is essentially unique. We begin with a useful observation.
Remarks 4.22:
(1) If G ⊂ Rn is open and E ⊂ Rn with |E | = 0, then cap p (G ) = cap p (G \ E ).
Reason. Ê Monotonicity implies cap p (G ) Ê cap p (G \ E ).

É Let ε > 0 and let u ∈ A (G \ E ) be such that
p
k uk É cap p (G \ E ) + ε.
W 1,p (Rn )
Then there exists an open O ⊂ Rn with (G \ E ) ⊂ O and u Ê 1 almost every-

where in O . Since O ∪ G is open G ⊂ (O ∪ G ) and u Ê 1 almost everywhere
in O ∪ (G \ E ), and almost everywhere in O ∪ G since |E | = 0, we have
u ∈ A (G ).
p
cap p (G ) É k ukW 1,p (Rn ) É cap p (G \ E ) + ε.
By letting ε → 0, we obtain cap p (G ) É cap p (G \ E ). ■
(2) For any open G ⊂ Rn we have |G | = 0 ⇐⇒ cap p (G ) = 0.
Reason. =⇒ If |G | = 0, then (1) implies
cap p (G ) = cap p (G \ G ) = cap p (;) = 0.
⇐= If cap p (G ) = 0, then Lemma 4.6 implies |G | É cap p (G ) = 0. ■
WA RNING : It is not true in general that capacity and measure have

the same zero sets.
Theorem 4.23. Assume that u and v are p−quasicontinuous functions on Rn . If

u = v almost everywhere in Rn , then u = v p-quasieverywhere in Rn .
T HE MORAL : Quasicontinuous representatives of Sobolev functions are

unique.
Proof. Let ε > 0 and choose open G ⊂ Rn such that cap p (G ) < ε and that the
restrictions of u and v to Rn \ G are continuous. Thus { x ∈ Rn \ G : u( x) 6= v( x)} is
open in the relative topology on Rn \ G , that is, there exists open U ⊂ Rn with
U \ G = { x ∈ Rn \ G : u( x) 6= v( x)}
and
|U \ G | = |{ x ∈ Rn \ G : u( x) 6= v( x)}| = 0.
Moreover,
{ x ∈ Rn : u( x) 6= v( x)} ⊂ G ∪ { x ∈ Rn \ G : u( x) 6= v( x)} = G ∪ U.
Remark 4.22 (1) with G and E replaced by U ∪ G and U \ G , respectively, implies
cap p ({ x ∈ Rn : u( x) 6= v( x)}) É cap p (G ∪ U ) = cap p (G ) < ε.

Remarks 4.24:
(1) The same proof gives the following local result: Assume that u and v are
p−quasicontinuous on an open set O ⊂ Rn . If u = v almost everywhere in
O , then u = v p-quasieverywhere in O .
(2) Observe that if u and v are p−quasicontinuous and u É v almost ev-
erywhere in an open set O , then max{ u − v, 0} = 0 almost everywhere
in O and max{ u − v, 0} is p−quasicontinuous. Then Theorem 4.23 im-
plies max{ u − v, 0} = 0 p-quasieverywhere in O and consequently u É v
p-quasieverywhere in O .
(3) The previous theorem enables us to define the trace of a Sobolev function
to an arbitrary set. If u ∈ W 1,p (Rn ) and E ⊂ Rn , then the trace of u to E is
the restriction to E of any p−quasicontinuous representative of u. This
definition is useful only if cap p (E ) > 0.
4.4 Lebesgue points of Sobolev functions

By the maximal function theorem with p = 1, see (3.1), there exists c = c( n) such
that
c
|{ x ∈ Rn : M f ( x) > λ}| É
k f kL1 (Rn )
λ
for every λ > 0. By Chebyshev’s inequality and the maximal function theorem
with 1 λ}| É k M f kL p (Rn ) É p k f kL p (Rn )
λ p λ
for every λ > 0. Thus the Hardy-Littlewood maximal function satisfies weak type
estimates with respect to Lebesgue measure for functions in L p (Rn ). Next we
consider capacitary weak type estimates for functions in W 1,p (Rn ).
Theorem 4.25. Assume that u ∈ W 1,p (Rn ), 1 λ} É p k ukW 1,p (Rn ) .
¡ ¢
λ
for every λ > 0.
T HE MORAL : This is a capacitary version of weak type estimates for the

Hardy-Littlewood maximal function.
Proof. Denote E λ = { x ∈ Rn : Mu( x) > λ}. Then E λ is open and by Theorem 3.46
Mu ∈ W 1,p (Rn ). Thus
Mu
∈ A ( E λ ).
λ
Since the maximal operator is bounded on W 1,p (Rn ), see (3.11), we obtain
¡ ¢ ° Mu ° p
° °
1 p c p
° λ ° 1,p n = λ p k MukW 1,p (Rn ) É λ p k ukW 1,p (Rn ) .
cap p E λ É ° °
W (R ) ä
This weak type inequality can be used in studying the pointwise behaviour of
Sobolev functions. We recall that x ∈ Rn is a Lebesgue point for u ∈ L1loc (Rn ) if the
limit ×
u∗ ( x) = lim u ( y) d y
r →0 B( x,r )
exists and ×
lim | u ( y) − u ∗ ( x ) | d y = 0 .
r →0 B( x,r )
The Lebesgue differentiation theorem states that almost all points are Lebesgue
points for a locally integrable function. If a function belongs to W 1,p (Rn ), then
using the capacitary weak type estimate, see Theorem 4.25, we shall prove
that it has Lebesgue points p-quasieverywhere. Moreover, we show that the
p-quasicontinuous representative given by Corollary 4.21 is u∗ .
We begin by proving a measure theoretic result, which is analogous to Lemma

4.13.
Lemma 4.26. Let 1 0 .
r →0 B( x,r )
Then cap p (E ) = 0.
THE MORAL : Roughly speaking the lemma above says that the set where an
p
L function blows up rapidly is of capacity zero. The main difference compared to
Lemma 4.13 is that the size of the set is measured by capacity instead of Hausdorff
measure.
Proof. The argument is similar to the proof of Lemma 4.13, but we reproduce it
here.
(1) By the Lebesgue differentiation theorem
×
lim | f ( y)| p d y = | f ( x)| p < ∞,
r →0 B( x,r )
for almost every x ∈ Rn . If x is a Lebesgue point of | f | p , then

×
p
lim sup r | f ( y)| p d y = 0.
r →0 B( x,r )
This shows that all Lebesgue points of | f | p belong to the complement of E and
thus |E | = 0.
(2) Let ε > 0 and
n × o
n p
E ε = x ∈ R : lim sup r | f |p d y > ε .
r →0 B( x,r )
Since E ε ⊂ E and |E | = 0, we have |E ε | = 0. We show that cap p (E ε ) = 0 for every

ε > 0, then the claim follows by subadditivity. Let 0 < δ < 15 . For every x ∈ E ε there
is r x with 0 < r x É δ such that
×
p
rx | f | p d y > ε.
B( x,r x )
By the Vitali covering theorem, there exists a subfamily of countably many pair-
wise disjoint balls B( x i , r i ), i = 1, 2, . . . , such that
∞
[
Eε ⊂ B( x i , 5 r i ).
i =1
By subadditivity of the capacity and Lemma 4.8 we have

∞ ∞
X X n− p
cap p (E ε ) É cap p (B( x i , 5 r i )) É c ri
i =1 i =1
∞
ˆ ˆ
cX c
É | f |p d y = | f ( y)| p d y.
ε i=1 B( x i ,r i ) ε S∞
B( x i ,r i )
i =1
Here c = c( n, p). Finally we observe that by the disjointness of the balls

¯ ¯
∞ rp
ˆ ˆ
¯∞
¯[ ¯ X ∞
i p δp
| f |p d y → 0
X
¯ B( x i , r i )¯ = |B( x i , r i )| É |f | dy É
¯
i =1 ε B( x i ,r i ) ε Rn
¯ i=1 ¯ i=1
as δ → 0. By absolute continuity of integral

ˆ
S∞
| f |p d y → 0
i =1
B( x i ,r i )
as δ → 0. Thus ˆ
c
cap p (E ε ) É | f |p d y → 0
ε S∞
B( x i ,r i )
i =1
as δ → 0, which implies that cap p (E ε ) = 0 for every ε > 0. ä
Now we are ready for a version of the Lebesgue differentiation theorem for
Sobolev functions.
Theorem 4.27. Assume that u ∈ W 1,p (Rn ) with 1 < p < ∞. Then there exists
E ⊂ Rn such that cap p (E ) = 0 and
×
lim u ( y) d y = u ∗ ( x )
r →0 B( x,r )
exists for every x ∈ Rn \ E . Moreover

×
lim | u ( y) − u ∗ ( x ) | d y = 0
r →0 B( x,r )
for every x ∈ Rn \ E and the function u∗ is the p-quasicontinuous representative

of u.
T HE MORAL : A function in W 1,p (Rn ) with 1 < p < ∞ has Lebesgue points
p-quasieverywhere. Moreover, the p-quasicontinuous representative is obtained
as a limit of integral averages.
Proof. (1) By Theorem 1.18 there exist u i ∈ C ∞ (Rn ) ∩ W 1,p (Rn ) such that
p
ku − u i k É 2− i( p+1) , i = 1, 2, . . . .
W 1,p (Rn )
Denote
E i = { x ∈ Rn : M ( u − u i )( x) > 2− i }, i = 1, 2, . . . .
By Theorem 4.25 there exists c = c( n, p) such that

p
cap p (E i ) É c2 i p k u − u i kW 1,p (Rn ) É c2− i , i = 1, 2, . . . .
Clearly
×
| u i ( x) − u B( x,r) | É | u i ( x) − u( y)| d y
B( x,r )
× ×
É | u i ( x) − u i ( y)| d y + | u i ( y) − u( y)| d y,
B( x,r ) B( x,r )
which implies that
lim sup | u i ( x) − u B( x,r) |

r →0
× ×
É lim sup | u i ( x) − u i ( y)| d y + lim sup | u i ( y) − u( y)| d y
r →0 B( x,r ) r →0 B( x,r )
É M ( u i − u)( x) É 2− i ,
for every x ∈ Rn \ E i . Here we used the fact that

×
lim sup | u i ( x) − u i ( y)| d y = 0, i = 1, 2, . . . ,
r →0 B( x,r )
since u i is continuous and

×
| u i ( y) − u( y)| d y É M ( u i − u)( x) for every r > 0.
B( x,r )
S∞
Let F k = i=k
E i , k = 1, 2, . . . . Then by the subadditivity of capacity we have
∞ ∞
2− i .
X X
cap p (F k ) É cap p (E i ) É c
i=k i=k
If x ∈ Rn \ F k and i, j Ê k, then
| u i ( x) − u j ( x)| É lim sup | u i ( x) − u B( x,r) | + lim sup | u B( x,r) − u j ( x)|

r →0 r →0
−i −j
É2 +2 .
Thus ( u i ) converges uniformly in Rn \ F k to a continuous function v in Rn \ F k .

Furthermore
lim sup |v( x) − u B( x,r) | É |v( x) − u i ( x)| + lim sup | u i ( x) − u B( x,r) |

r →0 r →0
−i
É |v( x) − u i ( x)| + 2
for every x ∈ Rn \ F k . The right-hand side of the previous inequality tends to zero
as i → ∞. Thus
lim sup |v( x) − u B( x,r) | = 0
r →0
and consequently ×
v( x) = lim u( y) d y = u∗ ( x)
r →0 B( x,r )
for every x ∈ Rn \ F k . Define F = ∞

T
F .
k=1 k
Then
∞
2− i = 0
X
cap p (F ) É lim cap p (F k ) É c lim
k→∞ k→∞ i = k
and ×
lim u ( y) d y = u ∗ ( x )
r →0 B( x,r )
exists for every x ∈ Rn \ F . This completes the proof of the first claim.
(2) To prove the second claim, consider
½ × ¾
E = x ∈ Rn : lim sup r p |Du( y)| p d y > 0 .
r →0 B( x,r )
Lemma 4.26 shows that cap p (E ) = 0. By the Poincaré inequality, see Theorem
3.17, we have
× ×
lim | u( y) − u B( x,r) | p d y É c lim r p |Du( y)| p d y = 0
r →0 B( x,r ) r →0 B( x,r )
for every x ∈ Rn \ E . We conclude that

×
lim | u( y) − u∗ ( x)| d y
r →0 B( x,r )
µ× ¶1
p
∗ p
É lim | u( y) − u ( x)| d y
r →0 B( x,r )
µ× ¶1
p
É lim | u( y) − u B( x,r) | p d y + lim | u B( x,r) − u∗ ( x)| = 0
r →0 B( x,r ) r →0
whenever x ∈ Rn \ (E ∪ F ). Finally we observe that
cap p (E ∪ F ) É cap p (E ) + cap p (F ) = 0.
(3) Let ε > 0 and choose k large enough so that cap p (F k ) < 2ε . Then by the
outer regularity of capacity, see Remark 4.5, there is an open set O containing F k
so that cap p (O ) < ε. Since ( u i ) converges uniformly to u∗ on Rn \ O we conclude
that u∗ | X \O is continuous. Thus u∗ is p-quasicontinuous. ä
4.5 Sobolev spaces with zero boundary

values
In this section we return to Sobolev spaces with zero boundary values started in
Section 1.9. Assume that Ω is an open subset of Rn and 1 É p < ∞. Recall that
1,p
W0 (Ω) with 1 É p < ∞ is the closure of C 0∞ (Ω) with respect to the Sobolev norm,
see Definition 1.20. Using pointwise properties of Sobolev functions we discuss
1,p
the definition of W0 (Ω).
1,p
The first result is a W0 (Ω) version of Corollary 4.21 which states that for
every u ∈ W 1,p (Rn ) there is a p-quasicontinuous function v ∈ W 1,p (Rn ) such that
u = v almost everywhere in Rn .
1,p
Theorem 4.28. If u ∈ W0 (Ω), there exists a p-quasicontinuous function v ∈
W 1,p (Rn ) such that u = v almost everywhere in Ω and v = 0 p-quasieverywhere in
R n \ Ω.
THE MORAL : Quasicontinuous functions in Sobolev spaces with zero bound-

ary values are zero quasieverywhere in the complement.
1,p
Proof. Since u ∈ W0 (Ω), there exist u i ∈ C 0∞ (Ω), i = 1, 2, . . . , such that u i → u in
W 1,p (Ω) as i → ∞. Since ( u i ) is a Cauchy sequence in W 1,p (Rn ), by Theorem 4.18
it has a subsequence of ( u i ) that converges pointwise p-quasieverywhere in Rn
to a function v ∈ W 1,p (Rn ). Moreover, the convergence is uniform outside a set
of arbitrary small p-capacity and, as in Corollary 4.21, the limit function v is
p-quasicontinuous. ä
Theorem 4.29. If u ∈ W 1,p (Rn ) is p-quasicontinuous and u = 0 p-quasieverywhere

1,p
in Rn \ Ω, then u ∈ W0 (Ω).
T HE MORAL : Quasicontinuous functions in a Sobolev space on the whole

space which are zero quasieverywhere in the complement belong to the Sobolev
space with zero boundary values. In particular, continuous functions in a Sobolev
space on the whole space which are zero everywhere in the complement belong to
the Sobolev space with zero boundary values.
Proof. (1) We show that u can be approximated by W 1,p (Rn ) functions with com-
pact support in Ω. If we can construct such a sequence for u + = max{ u, 0}, then we
can do it for u − = − min{ u, 0}, and we obtain the result for u = u + − u − . Thus we
may assume that u Ê 0. By Theorem 1.24 we may assume that u has a compact
support in Rn and by considering truncations min{ u, λ}, λ > 0, we may assume
that u is bounded (exercise).
(2) Let δ > 0 and let O ⊂ Rn be an open set such that cap p (O ) < δ and the
restriction of u to Rn \ O is continuous. Denote
E = { x ∈ Rn \ Ω : u( x) 6= 0}.
By assumption cap p (E ) = 0. Let v ∈ A 0 (O ∪ E ) such that 0 É v É 1 and

p
k vk < δ,
W 1,p (Rn )
see Remark 4.2. Then v = 1 in an open set G containing O ∪ E . Define
u ε ( x) = max{ u( x) − ε, 0}, 0 < ε < 1.
Let x ∈ ÇΩ \ G . Since u( x) = 0 and the restriction of u to Rn \ G is continuous, there

exists r x > 0 such that u ε = 0 in B( x, r x ) \ G . Thus (1 − v) u ε = 0 in B( x, r x ) ∪ G for
every x ∈ ÇΩ \ G . This shows that (1 − v) u ε is zero in a neighbourhood of Rn \ Ω,
which implies that (1 − v) u ε is compactly supported in Ω. Lemma 1.23 implies
1,p
(1 − v) u ε ∈ W0 (Ω). We show that this kind of functions converge to u in W 1,p (Rn ).
(3) Since

u − ε in { x ∈ Rn : u( x) Ê ε},
uε =
0 in { x ∈ Rn : u( x) É ε},
by Remark 1.27 we have

Du almost everywhere in { x ∈ Rn : u( x) Ê ε},
Du ε =
0 almost everywhere in { x ∈ Rn : u( x) É ε}.
Thus
k u − (1 − v) u ε kW 1,p (Rn ) É k u − u ε kW 1,p (Rn ) + kvu ε kW 1,p (Rn ) .
Using the facts that u − u ε É ε and supp( u − u ε ) ⊂ supp u, we obtain
k u − u ε kW 1,p (Rn ) É k u − u ε kL p (Rn ) + kDu − Du ε kL p (Rn )

É εkχsupp u kL p (Rn ) + kχ{0<uÉε} DukL p (Rn ) → 0
as ε → 0. Observe that, by the dominated convergence theorem, we have

µ ˆ ¶1
p
p
lim kχ{0<uÉε} DukL p (Rn ) = lim χ{0<uÉε} |Du| dx
ε→0 ε→0 Rn
µˆ ¶1
p
= lim χ{0<uÉε} |Du| p dx = 0,
Rn ε→0
where χ{0<uÉε} |Du| p É |Du| p ∈ L1 (Rn ) may be used as an integrable majorant. On

the other hand,
kvu ε kW 1,p (Rn ) É kvu ε kL p (Rn ) + kD (vu ε )kL p (Rn )

É kvu ε kL p (Rn ) + k u ε DvkL p (Rn ) + kvDu ε kL p (Rn )
É k uvkL p (Rn ) + k uDvkL p (Rn ) + kvDu ε kL p (Rn )
É k ukL∞ (Rn ) kvkL p (Rn ) + k ukL∞ (Rn ) kDvkL p (Rn ) + kvDukL p (Rn )
É 2k ukL∞ (Rn ) kvkW 1,p (Rn ) + kvDukL p (Rn )
1
É 2δ p k ukL∞ (Rn ) + kvDukL p (Rn ) .
Since v = vδ → 0 in L p (Rn ) as δ → 0, there is a subsequence (δ i ) for which v i =

vδ i → 0 almost everywhere as i → ∞. By the dominated convergence theorem, we
have
µ ˆ ¶1
p
lim kv i DukL p (Rn ) = lim |v i | p |Du| p dx
i →∞ i →∞ Rn
µˆ ¶1
p
p p
= ( lim |v i | )|Du| dx = 0,
Rn i →∞
where |v i | p |Du| p É |Du| p , so that |Du| p ∈ L1 (Rn ) may be used as an integrable

majorant. Thus we conclude that
µ 1 ¶
p
lim kv i u ε kW 1,p (Rn ) É lim 2δ i k ukL∞ (Rn ) + kv i DukL p (Rn ) = 0.
i →∞ i →∞
Thus
k u − (1 − v i ) u ε kW 1,p (Rn ) → 0
as ε → 0 and i → ∞. Since
1,p
(1 − v i ) u ε ∈ W0 (Ω) and (1 − v i ) u ε → u in W 1,p (Rn )
1,p
as ε → 0 and i → ∞, we conclude that u ∈ W0 (Ω). ä
Remark 4.30. If u ∈ W 1,p (Rn ) is continuous and zero everywhere in Rn \ Ω, then

1,p
u ∈ W0 (Ω).
We obtain a very useful characterization of Sobolev spaces with zero boundary

values on an arbitrary open set by combining the last two theorems.
1,p
Corollary 4.31. u ∈ W0 (Ω) if and only if there exists a p-quasicontinuous
function u∗ ∈ W 1,p (Rn ) such that u∗ = u almost everywhere in Ω and u = 0 p-
quasieverywhere in Rn \ Ω.
THE MORAL : Quasicontinuous functions in Sobolev spaces with zero bound-

ary values are precisely functions in the Sobolev space on the whole space which
are zero quasieverywhere in the complement. This result can be used to show that
a given function belongs to the Sobolev space with zero boundary values without
constructing an approximating sequence of compactly supported smooth functions.
There is also a characterization of Sobolev spaces with zero boundary values

using Lebesgue points for Sobolev functions.
Theorem 4.32. Assume that Ω ⊂ Rn is an open set and u ∈ W 1,p (Rn ) with 1 < p <
1,p
∞. Then u ∈ W0 (Ω) if and only if
×
lim u( y) d y = 0
r →0 B( x,r )
for p-quasievery x ∈ Rn \ Ω.
T HE MORAL : A function in the Sobolev space on the whole space belongs to

the Sobolev space with zero boundary values if and only if the limit of integral
averages is zero quasieverywhere in the complement.
1,p
Proof. =⇒ If u ∈ W0 (Ω), then by Theorem 4.28 there exists a p-quasicontinuous
function u∗ ∈ W 1,p (Rn ) such that u∗ = u almost everywhere in Ω and u∗ = 0
p-quasieverywhere in Rn \ Ω. Theorem 4.27 shows that the limit
×
u∗ ( x) = lim u ( y) d y
r →0 B( x,r )
exists p-quasieverywhere and that the function u∗ is a p-quasicontinuous repre-

sentative of u. This shows that
×
lim u ( y) d y = u ∗ ( x ) = 0
r →0 B( x,r )
for p-quasievery x ∈ Rn \ Ω.
⇐= Assume then that u ∈ W 1,p (Rn ) and
×
lim u ( y) d y = 0
r →0 B( x,r )
for p-quasievery x ∈ Rn \ Ω. Theorem 4.27 shows that the limit

×
∗
u ( x) = lim u( y) d y
r →0 B( x,r )
exists p-quasieverywhere and that the function u∗ is a p-quasicontinuous repre-

sentative of u. We conclude that u∗ ( x) = 0 for p-quasievery x ∈ Rn \ Ω. ä
1,p
Example 4.33. Let Ω = B(0, 1) \ {0} and u : Ω → R, u( x) = 1 − | x|. Then u ∈ W0 (Ω)
1,p
for 1 n.
THE MORAL : A function that belongs to the Sobolev space with zero boundary
values does not have to be zero at every point of the boundary.
Remark 4.34. Theorem 4.32 gives a practical tool to show that a function belongs
to a Sobolev space with zero boundary values. For example, the following claims
follow from Theorem 4.32 (exercise).
1,p
(1) Assume that u ∈ W 1,p (Ω) has a compact support, then u ∈ W0 (Ω).
1,p 1,p
(2) Assume that u ∈ W0 (Ω). Then | u| ∈ W0 (Ω).
1,p
(3) Assume that u ∈ W0 (Ω). If v ∈ W 1,p (Ω) and 0 É v É u almost everywhere
1,p
in Ω, then v ∈ W0 (Ω).
1,p
(4) Assume that u ∈ W0 (Ω). If v ∈ W 1,p (Ω) and |v| É | u| almost everywhere
1,p
in Ω \ K , where K is a compact subset of Ω, then v ∈ W0 (Ω).
Let E ⊂ Ω be a relatively closed set, that is, there exists a closed F ⊂ Rn such
1,p 1,p
that E = Ω ∩ F , with |E | = 0. It is clear that W0 (Ω \ E ) ⊂ W0 (Ω). By
1,p 1,p
W0 (Ω \ E ) = W0 (Ω)
1,p
we mean that every u ∈ W0 (Ω) can be approximated by functions in C 0∞ (Ω \ E )
1,p
or in W0 (Ω \ E ).
1,p
Theorem 4.35. Assume that E is a relatively closed subset of Ω. Then W0 (Ω) =
1,p
W0 (Ω \ E ) if and only if cap p (E ) = 0.
Proof. ⇐= Assume cap p (E ) = 0. Lemma 4.6 implies |E | = 0 so that it is reasonable

1,p 1,p
to ask whether W0 (Ω) = W0 (Ω \ E ) when we consider functions defined up to a
set of measure zero.
1,p 1,p
It is clear that W0 (Ω \ E ) ⊂ W0 (Ω). To see reverse inclusion, let u i ∈ C 0∞ (Ω),
i = 1, 2, . . . , be such that u i → u in W 1,p (Ω) as i → ∞. Since cap p (E ) = 0 there are
v j ∈ A 0 (E ), j = 1, 2, . . . , be such that kv j kW 1,p (Rn ) → 0 as j → ∞. Then (1 − v j ) u i ∈
W 1,p (Ω) and, since v j = 1 in a neighbourhood of E , supp(1 − v j ) u i is a compact
1,p
subset of Ω \ E for every i, j = 1, 2, . . . . Lemma 1.23 implies (1 − v j ) u i ∈ W0 (Ω \ E ),
i, j = 1, 2, . . . .
Moreover, we have
k u − (1 − v j ) u i kW 1,p (Ω) É k u − u i kW 1,p (Ω) + kv j u i kW 1,p (Ω) ,
where k u − u i kW 1,p (Ω) → 0 as i → ∞ and
kv j u i kW 1,p (Ω) É kv j u i kL p (Ω) + kD (v j u i )kL p (Ω)

É k u i kL∞ (Ω) kv j kL p (Ω) + kv j Du i kL p (Ω) + k u i Dv j kL p (Ω)
É k u i kL∞ (Ω) kv j kL p (Ω) + kv j Du i kL p (Ω) + k u i kL∞ (Ω) kDv j kL p (Ω)
É 2k u i kL∞ (Ω) kv j kW 1,p (Ω) + kv j Du i kL p (Ω) .
Since v j → 0 in L p (Ω) as j → ∞, there is a subsequence, still denoted by (v j ),

for which v j → 0 almost everywhere as j → ∞. By the dominated convergence
theorem, we have
µ ˆ ¶1
p
lim kv j Du i kL p (Ω) = lim |v j | p |Du i | p dx
j →∞ j →∞ Ω
µˆ ¶1
p
p p
= ( lim |v j | )|Du i | dx = 0.
Ω j →∞
Observe that |v j | p |Du i | p É |Du i | p for j = 1, 2, . . . , so that |Du i | p ∈ L1 (Ω) may be

used as an integrable majorant. Thus
k u − (1 − v j ) u i kW 1,p (Ω) → 0 as i, j → ∞.
Since
1,p
(1 − v j ) u i ∈ W0 (Ω \ E ) and (1 − v j ) u i → u in W 1,p (Ω \ E )
1,p
as i, j → ∞, we conclude that u ∈ W0 (Ω \ E ).
=⇒ Let x0 ∈ Ω and let i 0 ∈ N be large enough that
1
dist( x0 , Rn \ Ω) > .
i0
Define
1
½ ¾
Ω i = x ∈ Ω : dist( x, Rn \ Ω) > ∩ B( x0 , i ), i = i 0 , i 0 + 1, . . . .
i
Observe that Ω i b Ω i+1 b · · · b Ω and Ω = ∞ Ω i . Let u i : Rn → R,

S
i= i 0
u i ( x) = dist( x, Rn \ Ω2 i ).
1,p 1
Then u i is Lipschitz continuous, u i ∈ W0 (Ω) and u i ( x) Ê 2i for every x ∈ E ∩ Ω i ,
1,p 1,p 1,p
i = 1, 2, . . . . Since W0 (Ω) = W0 (Ω \ E ) we have u i ∈ W0 (Ω \ E ), i = 1, 2, . . . .
Fix i and let v j ∈ C 0∞ (Ω \ E ), j = 1, 2, . . . , such that v j → u i in W 1,p (Ω \ E ) as
j → ∞. Since 3 i ( u i − v j ) Ê 1 in a neigbourhood of E ∩ Ω i ,
p
cap p (E ∩ Ω i ) É k3 i ( u i − v j )kW 1,p (Ω\E )
p
= (3 i ) p k u i − v j k →0 as j → ∞.
W 1,p (Ω\E )
Thus cap p (E ∩ Ω i ) = 0, i = 1, 2 . . . , and by subadditivity

Ã !
∞ ∞
(E ∩ Ω i ) É cap p (E ∩ Ω i ) = 0.
[ X
cap p (E ) = cap p
i =1 i =1 ä
THE END
Bibliography
[1] D. Aalto and J. Kinnunen, Maximal functions in Sobolev spaces. Sobolev

spaces in mathematics. I, 25–67, Int. Math. Ser. (N. Y.), 8, Springer, New
York, 2009.
[2] D.R. Adams and L.I. Hedberg, Function spaces and potential theory,
Springer 1996.
[3] R.A. Adams and J.J.F. Fournier, Sobolev spaces (second edition), Aca-
demic Press 2003.
[4] L. Diening, P. Harjulehto, P. Hästö and M. Ružička, Lebesgue and

Sobolev spaces with variable exponents, Springer 2011.
[5] A. Björn and J. Björn, Nonlinear potential theory on metric spaces,

European Mathematical Society (EMS) 2011.
[6] E. Di Nezza, G. Palatucci and E. Valdinoci, Hitchhiker’s guide to the

fractional Sobolev spaces, Bull. Sci. Math. 136 (2012), no. 5, 521–573.
[7] L.C. Evans, Partial differential equations, 2nd edition, AMS 2010.
[8] L.C. Evans and R.F. Gariepy, Measure theory and fine properties of
functions, CRC Press 1992.
[9] P. Hajłasz, Sobolev inequalities, truncation method, and John domains.

Papers on analysis, 109–126, Rep. Univ. Jyväskylä Dep. Math. Stat., 83,
2001.
[10] L. Diening, P. Harjulehto, P. Hästö, Peter and M. Ruzicka, Lebesgue and

Sobolev spaces with variable exponents, Springer 2011.
[11] J. Heinonen, Lectures on analysis on metric spaces, Springer 2001.
[12] J. Heinonen, T. Kilpeläinen and O. Martio, Nonlinear potential theory

of degenerate elliptic equations. Unabridged republication of the 1993
original. Dover Publications, 2006.
[13] J. Heinonen, P. Koskela, N. Shanmugalingam and J. Tyson, Sobolev

spaces on metric measure spaces. An approach based on upper gradients,
Cambridge University Press 2015.
[14] G. Leoni, A first course in Sobolev spaces, 2nd edition, AMS 2017.
[15] V. Maz’ya, Sobolev spaces with applications to elliptic partial differential

equation, 2nd edition, Springer 2011.
131
BIBLIOGRAPHY 132
[16] E. M. Stein, Singular Integrals and Differentiability Properties of Func-

tions, Princeton University Press, 1970.
[17] W.P. Ziemer, Weak differentiable functions, Springer 1989.

Sobolev Spaces

Uploaded by

Copyright:

Available Formats

You might also like

Sobolev Spaces

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sobolev Spaces

Uploaded by

Copyright:

Available Formats

JUHA KINNUNEN

Department of Mathematics, Aalto University

3 MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 70

4 POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 106

1.1 Weak derivatives

C 0∞ (Ω) = C ∞ (Ω) ∩ C 0 (Ω)

Now ϕ ∈ C 0∞ (Rn ) and supp ϕ = B(0, 1) (exercise).

Let us start with a motivation for definition of weak derivatives. Let Ω ⊂ Rn

T HE MORAL : A coordinate of a multi-index indicates how many times a

Successive integration by parts gives

for every test function ϕ ∈ C 0∞ (Ω). We denote D 0 u = D (0,...,0) = u. If |α| = 1, then

is the weak gradient of u. Here

T HE MORAL : Classical derivatives are defined as pointwise limits of differ-

Lemma 1.4. A weak αth partial derivative of u, if it exists, is uniquely defined

Claim: v = ve almost everywhere in Ω.

Reason. Let Ω0 b Ω (i.e. Ω0 is open and Ω0 is a compact subset of Ω). The

This implies that v = ve almost everywhere in Ω0 for every Ω0 b Ω. Thus v = ve

From the proof we obtain a very useful corollary.

Corollary 1.5 (Fundamental lemma of the calculus of variations). If f ∈ L1loc (Ω)

T HE MORAL : This is an integral way to say that a function is zero almost

Example 1.6. Let n = 1 and Ω = (0, 2). Consider

We claim that u0 = v in the weak sense. To see this, we show that

for every ϕ ∈ C 0∞ ((0, 2)).

Reason. An integration by parts and the fundamental theorem of calculus give

for every ϕ ∈ C 0∞ ((0, 2)). ■

1.2 Sobolev spaces

W k,p (Ω) = { u ∈ L p (Ω) : D α u ∈ L p (Ω), |α| É k}.

If u ∈ W k,p (Ω), we define its norm

and k · kW k,∞ (Ω) is also equivalent with

max kD α ukL∞ (Ω) .

(3) For k = 1 we use the norm

when 1 É p < ∞ and

By rearranging terms, we obtain

= kϕkL∞ (B(0,1)) ωn−1 εn−1−α → 0 as ε → 0,

if n − 1 − α > 0. Here ωn−1 = H n−1 (ÇB(0, 1)) is the ( n − 1)-dimensional measure of

The following argument shows that D j u is a weak derivative of u also in a

THE MORAL : Functions in W 1,p , 1 É p < n, n Ê 2, may be unbounded in every

T HE MORAL : Functions in W 1,p , 1 É p É n, n Ê 2, are not continuous. Later

Example 1.11. The function u : B(0, 1) → R,

does not belong to W 1,p (B(0, 1) for any 1 É p É ∞ (exercise).

1.3 Properties of weak derivatives

Lemma 1.12. Assume that u, v ∈ W k,p (Ω) and |α| É k. Then

(1) D α u ∈ W k−|α|,p (Ω),

(4) if Ω0 ⊂ Ω is open, then u ∈ W k,p (Ω0 ),

for all test functions ϕ ∈ C 0∞ (Ω). Notice that

|α| + |α + β| =α1 + . . . + αn + (α1 + β1 ) + . . . + (αn + βn )

1.4 Completeness of Sobolev spaces

k u i − ukW k,p (Ω) < ε when i Ê iε.

W A R N I N G : This is not the same condition as

k u i+1 − u i kW k,p (Ω) < ε when i Ê iε.

Theorem 1.13 (Completeness). The Sobolev space W k,p (Ω), 1 É p É ∞, k =

T HE MORAL : The spaces C k (Ω), k = 1, 2, . . . , are not complete with respect

Proof. Step 1: k · kW k,p (Ω) is a norm.

Reason. (1) k ukW k,p (Ω) = 0 ⇐⇒ u = 0 almost everywhere in Ω.

= k ukW k,p (Ω) + kvkW k,p (Ω) . ■

Step 2: Let ( u i ) be a Cauchy sequence in W k,p (Ω). As

kD α u i − D α u j kL p (Ω) É k u i − u j kW k,p (Ω) , |α| É k,

it follows that (D α u i ) is a Cauchy sequence in L p (Ω), |α| É k. The completeness