Professional Documents
Culture Documents
Sobolev Spaces
Sobolev Spaces
Sobolev Spaces
Sobolev spaces
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
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
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).
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
Du = (D 1 u, D 2 u . . . , D n u)
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.
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
and
1, 0 < x < 1,
v( x) =
0, 1 É x < 2.
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 ukW 1,∞ (Ω) = ess sup | u| + ess sup |Du|.
Ω Ω
We may also consider equivalent norms
à !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)
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)
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)
¯
if n − 1 − α > 0.
CHAPTER 1. SOBOLEV SPACES 7
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 < p . On the
n− p
other hand, |Du| ∈ L p (B(0, 1), if − p(α + 1) + n > 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 < p .
Reason.
∞ 1 °
°| x − r i |−α ° 1,p
X °
k ukW 1,p (B(0,1)) É i W (B(0,1))
i =1 2
∞ 1 °
°| x|−α ° 1,p
X °
É i W (B(0,1))
i =1 2
° −α °
= °| x| °W 1,p (B(0,1)) < ∞. ■
Note that 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.
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).
D α (λ u + µv) = λD α u + µD α v,
where à !
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.
CHAPTER 1. SOBOLEV SPACES 9
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
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
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 .
CHAPTER 1. SOBOLEV SPACES 10
Indeed, the Cauchy sequence condition implies this, but the converse is not true
(exercise).
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
à !1
p
k D α u + D α v k L p (Ω )
X p
k u + vkW k,p (Ω) =
|α|É k
à !1
X ¡ α p
k D u k L p (Ω ) + k D α v k L p (Ω )
¢p
É
|α|É k
à !1 à !1
p p
kD α ukL p (Ω) kD α vkL p (Ω)
X p X p
É +
|α|É k |α|É k
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
¯ ¯ ¯ ¯
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
¯ ¯
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).
〈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)
CHAPTER 1. SOBOLEV SPACES 12
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)
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.
¡ 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
CHAPTER 1. SOBOLEV SPACES 14
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
D α f ε = D α φε ∗ f in Ωε
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
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)
= | 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
CHAPTER 1. SOBOLEV SPACES 16
(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.
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
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
Observe that the sum is only over four indices in a neighbourhood of a given
point. ■
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.
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).
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,
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
as i → ∞. ■
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 ).
Reason.
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
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.
CHAPTER 1. SOBOLEV SPACES 22
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
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
¯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
CHAPTER 1. SOBOLEV SPACES 23
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.
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
1 t
( f ε )0 ( t) = ( t2 + ε2 )−1/2 2 t = p for every t ∈ R,
2 t + ε2
2
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
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)}.
Reason.
1 1
max{ u, v} = ( u + v + | u − v|) and min{ u, v} = ( u + v − | u − v|). ■
2 2
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 − 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
we have
ˆ ˆ
lim |Du − Du λ | p dx = lim |Du − Du λ | p dx = 0.
λ→∞ Ω Ω λ→∞ ■
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 .
CHAPTER 1. SOBOLEV SPACES 26
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
THE MORAL : The L p -norm is lower semicontinuous with respect to the weak
convergence.
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 → ∞.
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
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
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 → ∞.
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
Remark 1.35. Mazur’s lemma can be used to give a proof for (1.3) (exercise).
f i = ( u i , Du i ) ∈ L p (Ω; Rn+1 )
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
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
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
CHAPTER 1. SOBOLEV SPACES 30
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).
1 i − 1 3i − 2
k u i k2L2 ((0,2)) = + 2 = , kDu i k2L2 ((0,2)) = 1,
3 i2 i 3 i2
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
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 .
CHAPTER 1. SOBOLEV SPACES 31
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).
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
Remark 1.42. Theorem 1.36 and Theorem 1.41 do not hold when p = 1 (exercise).
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 → ∞.
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 )
CHAPTER 1. SOBOLEV SPACES 33
and
0
ψ = |B( x, r )|−1 χB( x,r) e ∈ L p (Ω; Rm ).
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
u( x + he j ) − u( x)
D hj u( x) = , j = 1, . . . , n,
h
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
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
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
CHAPTER 1. SOBOLEV SPACES 35
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
(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
CHAPTER 1. SOBOLEV SPACES 36
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
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
then
m
X
| u( yi ) − u( x i )| < ε.
i =1
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).
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.
CHAPTER 1. SOBOLEV SPACES 38
(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,
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 à 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
CHAPTER 1. SOBOLEV SPACES 39
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).
where
¯ˆ t ¯
( g0k+1 − g0k )( s) ds¯¯
¯ ¯
| g k+1 ( t) − g k ( t)| = ¯¯
ˆ −∞
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 ■
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
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.
CHAPTER 1. SOBOLEV SPACES 41
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
1− np + nq
k ukL q (Rn ) É cλ kDukL p (Rn ) .
THE MORAL : There is only one possible exponent q for which inequality (2.1)
may hold true for all compactly supported smooth functions.
Observe that
(1) p∗ > p,
(2) If p → n−, then p∗ → ∞ and
n
(3) If p = 1, then p∗ = n−1 .
Sobolev proved the following theorem in the case p > 1 and Nirenberg and
Gagliardo in the case p = 1.
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
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
v = | u |γ ,
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
∗ ∗ ∗
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 ).
CHAPTER 2. SOBOLEV INEQUALITIES 46
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).
CHAPTER 2. SOBOLEV INEQUALITIES 47
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
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 < ∞.
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
CHAPTER 2. SOBOLEV INEQUALITIES 48
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).
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.
CHAPTER 2. SOBOLEV INEQUALITIES 49
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.
n £ n £
Ω⊂ y j − diam(Ω), y j + diam(Ω) =
Y ¤ Y ¤
a j, b j ,
j =1 j =1
1,p
(2) The case u ∈ W0 (Ω) follows by approximation (exercise). ä
Q = [a 1 , b 1 ] × . . . × [a n , b n ], b1 − a1 = . . . = b n − a n
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
Same notation is used for integral averages over other sets as well.
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 )
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
à 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
È !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
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
CHAPTER 2. SOBOLEV INEQUALITIES 52
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
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.
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 )
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.
(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
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).
CHAPTER 2. SOBOLEV INEQUALITIES 54
(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.
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 1 < p < n, inequality (2.2) follows from (2.3) and Hölder’s inequality on the
left-hand side.
CHAPTER 2. SOBOLEV INEQUALITIES 55
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
| u( x) − u( y)| É c| x − y|α
CHAPTER 2. SOBOLEV INEQUALITIES 56
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 -
¡ ¢
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 )
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
¯ ¯
CHAPTER 2. SOBOLEV INEQUALITIES 57
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
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 ) .
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
CHAPTER 2. SOBOLEV INEQUALITIES 58
u( x) = lim u( x i ).
i →∞
1− np
| u( z) − u( y)| É c| z − y| kDukL p (Q ( x,l ))
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.
| u( y) − u( x) − L( x − y)|
lim = 0.
y→ x | x − y|
CHAPTER 2. SOBOLEV INEQUALITIES 59
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 )
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 )
as y → x. ä
| f ( x) − f ( y)| É L| x − y|
CHAPTER 2. SOBOLEV INEQUALITIES 60
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
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|
CHAPTER 2. SOBOLEV INEQUALITIES 61
kD −j h ukL∞ (Rn ) É L
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
A direct combination of Theorem 2.23 and Theorem 2.21 gives a proof for
Rademacher’s theorem.
CHAPTER 2. SOBOLEV INEQUALITIES 62
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.
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).
CHAPTER 2. SOBOLEV INEQUALITIES 64
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)
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
1,2
for every v ∈ W 1,2 (Ω) with v − g ∈ W0 (Ω).
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
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.
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
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.
CHAPTER 2. SOBOLEV INEQUALITIES 68
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)
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 ) → ∞.
where B( x, r ) = { y ∈ Rn : | y − x| < r } is the open ball with the radius r > 0 and the
center x ∈ Rn .
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.
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.
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.
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|
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
THE MORAL : This gives a useful pointwise bound for a compactly supported
smooth function in terms of the Riesz potential of the gradient.
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.
= 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
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 )
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 )
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 .
| 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
È ! n
n−1
=c |Du j ( x)| dx
A j−1
È ! n
n−1
0 1− j 1− j
Éc ϕ (2 | u( x)| − 1)2 |Du( x)| dx
A j−1
È ! n
n−1
n
− j n−
= c2 1 |Du( x)| dx .
A j−1
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 78
In the last equality we used the fact that the sets A j , j ∈ Z, are pairwise disjoint.ä
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 )
for every z ∈ B( x, r ).
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|
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 →∞
= cI 1 (|Du|χB( x,r) )( z)
É crM (|Du|χB( x,r) )( z),
for every z ∈ B( x, r ) \ ( N1 ∪ N2 ).
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 81
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 3.17 (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 .
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 82
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 )
¯{ x ∈ E : f ( x) = 0}¯ Ê |E | .
¯ ¯
2
Then for every a ∈ R and λ > 0, we have
λ
| f ( x) − a| Ê f ( x) − | a| > .
2
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
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 83
É |{ x ∈ E : f ( x) = 0}|
λ ¯¯
¯½ ¾¯
¯
É ¯¯ x ∈ E : | f ( x) − a| > .
2 ¯
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
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 .
¯{ 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}.
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 84
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,
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
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 )
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 )
.
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 85
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 )
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 ).
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 ,
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 86
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 )
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.
Remarks 3.23:
(1) A bounded and connected open set Ω ⊂ Rn satisfies the interior cone condi-
tion, if there exists a bounded cone
THE MORAL : The Sobolev-Poincaré inequality holds for many other sets than
balls as well.
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
Since ×
uBi = u ( y) d y → u ( x )
Bi
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
| u ( x ) − u Ω | É | u ( x ) − u B 0 | + | u B 0 − u Ω |,
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
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 .
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|
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 )
° °
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 ∞.
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
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 ).
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
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
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
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
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 92
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)
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 )
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
for every x, y ∈ Rn \ E .
THE MORAL : This is a pointwise inequality for a function without the gradi-
ent.
É cr β f β#,r ( x).
É cr β f β#,2r ( 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 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 )
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
for every x, y ∈ Rn \ E .
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 95
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 )
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
1− np
| u( x) − u( y)| É c( n, p)kDukL p (Rn ) | x − y|
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#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 )
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)
Proof. Define f ∗ : Rn → R,
f ∗ ( x) = inf f (a) + L| x − a| : a ∈ A .
© ª
f ( b ) − f ( a ) É | f ( b ) − f ( a )| É L | b − a | ,
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| .
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 < p < ∞. Then for every ε > 0 there
exists a Lipschitz continuous function u ε : Rn → R such that
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
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 99
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
λ ¯¯
¯½ ¾¯
n n
|{ x ∈ R M f ( x) > λ}| É ¯ x ∈ R M f 1 ( x) >
¯
: ¯ :
2 ¯
as λ → ∞.
To prove the corresponding estimate for the gradients, we note that
{ x ∈ Rn : u( x) 6= u λ ( x)} ⊂ Ω \ E λ ,
∞
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 )
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 .
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 101
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
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 102
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
| u h ( y) − u( y)| = | u( y + h) − u( y)| É L| h|
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 < p < ∞. If u ∈ W 1,p (Rn ), then Mu ∈ W 1,p (Rn ). Moreover,
there exists c = c( n, p) such that
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
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
almost everywhere in Rn .
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,
CHAPTER 3. MAXIMAL FUNCTION APPROACH TO SOBOLEV SPACES 104
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
We define functions vk : Rn → R, k = 1, 2, . . . , by
vk ( x) = max (| u| ∗ χr m )( x).
1É mÉ k
|D j vk | É max ¯D j (| u| ∗ χr m )¯
¯ ¯
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,
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
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
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
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
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 )
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 )
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.
n
Ei É ∞i =1 cap p (E i ) whenever E i ⊂ R , i = 1, 2, . . . .
¡ S∞ ¢ P
(3) cap p i =1
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.
vk = max u i , k = 1, 2, . . . .
1É i É k
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
(2) To see the inequality in the other direction, let ε > 0 and take u ∈ A (E )
such that
p
k uk É cap p (E ) + ε.
W 1,p (Rn )
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.
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.
(2) Lemma 4.8 implies that cap p ({ x}) = 0 for every x ∈ Rn when 1 < p < n.
Reason.
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
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 111
We have shown that a point has zero capacity when 1 < p É n. By countable
subadditivity all countable sets have zero capacity as well. Next we show that a
point has positive capacity when p > n.
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 )
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
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 112
H s (E ) = lim H δs (E ) = sup H δs (E ).
δ →0 δ>0
inf s : H s (E ) = 0 = sup s : H s (E ) = ∞ .
© ª © ª
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.
This shows that all Lebesgue points of | f | belong to the complement of E and thus
|E | = 0.
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 113
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
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
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 .
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.
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
(2) Claim:
×
lim sup u d y = ∞ for every x ∈ E. (4.1)
r →0 B( x,r )
i= j i= j
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 116
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)
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.
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.
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.
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 117
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 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
à !
∞
\ ∞
\
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
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.
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
p
k uk É cap p (G \ E ) + ε.
W 1,p (Rn )
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)}
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 119
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.
Theorem 4.25. Assume that u ∈ W 1,p (Rn ), 1 < p < ∞. Then there exists c =
c( n, p) such that
c p
cap p { x ∈ Rn : Mu( x) > λ} É p k ukW 1,p (Rn ) .
¡ ¢
λ
for every λ > 0.
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 120
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∗ .
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 )
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 121
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 )
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
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
Now we are ready for a version of the Lebesgue differentiation theorem for
Sobolev functions.
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 122
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 )
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, . . . .
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 )
É M ( u i − u)( x) É 2− i ,
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
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 )
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 )
(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
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
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}.
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 ) .
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
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.
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.
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 )
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 )
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 < p É n and u ∉ W0 (Ω) for p > n.
CHAPTER 4. POINTWISE BEHAVIOUR OF SOBOLEV FUNCTIONS 128
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.
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
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 )
THE END
Bibliography
[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.
[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.
[14] G. Leoni, A first course in Sobolev spaces, 2nd edition, AMS 2017.
131
BIBLIOGRAPHY 132