Sobolev spaces and embedding theorems

Tomasz Dlotko, Silesian University, Poland

Contents
1. Introductory remarks 1
1.1. Domains 1
1.2. Generalized derivatives 2
1.3. Lp spaces 3
2. Sobolev spaces 5
2.1. Definition of the Sobolev spaces 5
2.2. Dense subsets and approximation in Sobolev spaces 6
3. Embeddings of Sobolev spaces 7
3.1. Continuous embeddings of Sobolev spaces 7
3.2. Compact embeddings of Sobolev spaces 9
4. Applications of Sobolev spaces 10
4.1. Closedness of differential operators in Sobolev spaces 11
4.2. The Lax-Milgram lemma 12
5. References 15
 1

1. Introductory remarks

In this initial part of the lecture an auxiliary material needed in the main body will be presented. The
following notions will be discussed:
• Classes of domains with boundaries satisfying cone condition, Lipschitz condition or of the class C k .
Also an extension property allowing to restrict most proofs to the case of the whole of Rn will be
discussed.
• Generalized (Sobolev, weak) derivatives. Their properties, comparison with distributional deriva-
tives.
• Basic properties of Lp spaces and the space L1loc . Inequalities.
• Poincaré inequality and interpolation inequality.
The theory of Sobolev spaces has been originated by Russian mathematician S.L. Sobolev around 1938
[SO]. These spaces were not introduced for some theoretical purposes, but for the need of the theory of
partial differential equations. They are closely connected with the theory of distributions, since elements of
such spaces are special classes of distributions.
In order to discuss the theory of Sobolev spaces we shall start with some simple basic notions that are
necessary for introducing and studying these spaces. The first object that we need to discuss is the domain
in Rn and the possible classes of the domains that are considered in the theory of Sobolev spaces. This is
important, since elements of such spaces are functions defined on the domains in Rn with, say, real values
(or complex values).

1.1. Domains. By a domain in Rn we understand an open set in n−dimensional real Euclidean space Rn .
There are several classes of domains, described in terms of the ”smoothness” of their boundary ∂Ω considered
in this theory. We will concentrate our description on the case of bounded domains Ω, when the definitions
are simpler pointing e.g. to [AD] for the case of not necessarily bounded Ω.
The three classes of domains are most often considered;
• Domains Ω ⊂ Rn having the cone property.
• Domains having the local Lipschitz property.
• Domains having the C m − regularity property.
We start with the description of the first property. Given a point x ∈ Rn an open ball B1 with center in
x and an open ball B2 not containing x the set Cx = B1 ∩ {x + λ(y − x) : y ∈ B2 , λ > 0} is called a finite
cone in Rn having vertex at x. By x + C0 = x + y : y ∈ C0 we will denote the finite cone with vertex in x
obtained by parallel translation of a finite cone C0 with vertex at 0.
Definition 1. Ω has the cone property if there exists a finite cone C such that each point x ∈ Ω is the vertex
of a finite cone Cx contained in Ω and congruent to C. (That means, Cx is obtained from C by a rigid
motion).
Definition 2. Ω has a locally Lipschitz boundary, if each point x on the boundary of Ω has a neighbourhood
Ux such that bdry Ω ∩ Ux is the graph of a Lipschitz continuous function.
Definition 3. A bounded domain Ω ⊂ Rn is called a domain of class C k , 0 ≤ k ≤ ∞, provided that the
following conditions are satisfied:
• There is a finite open cover of the boundary ∂Ω,
N
X
∂Ω ⊂ Vr (1)
r=1
¯ and ∆ is a
such that the intersection ∂Ω ∩ Vr can be described as: xn = gr (x0 ) where gr ∈ C k (∆)
cube {|xj | < a, j = 1, · · · , n − 1}.
2

• There is real b > 0, such that:

{x : gr (x0 ) − b < xn < gr (x0 ), x0 ∈ ∆} ⊂ Rn \ Ω̄ (2)
also
{x : gr (x0 ) < xn < gr (x0 ) + b, x0 ∈ ∆} ⊂ Ω. (3)

With the exception of the cone property all the other properties require Ω to lie on one side of its boundary.
It can be shown, that the following connections between these conditions are valid:

 m
 C − regularity property (m ≥ 1) ⇒
local Lipschitz property ⇒ (4)


cone property.

Most of the important properties of the Sobolev spaces require only the assumption that ∂Ω satisfies the
cone condition. One more property of the domains is very important in the considerations. This is the, so
called, extension property. This property will be described below.
We set the definition first:
Definition 4. Let Ω be a domain in Rn . For given numbers m and p a linear operator E mapping W m,p (Ω)
into W m,p (Rn ) is called an (m, p) -extension operator for Ω provided that:
(
Eu(x) = u(x) a.e. in Ω,
(5)
∀u∈W m,p (Ω) ∃K(m,p) kEukW m,p (Rn ) ≤ KkukW m,p (Ω) .

The existence of an (m, p)-extension operator for a domain Ω guarantees that W m,p (Ω) inherits many
nice properties possessed by W m,p (Rn ). If, for example, the embedding W m,p (Rn ) ⊂ Lq (Rn ) is known to
hold, similar property will be true for the spaces over Ω. We will quote below a theorem justifying existence
of such extension operator:
Theorem 1. Let Ω be either a half-space in Rn or a bounded domain in Rn having the C m -regularity
property, then for any positive integer m there exists an extension operator E for Ω.

The proof of this result can be found in [AD], p.84 and it uses the notion of the partition of unity.

1.2. Generalized derivatives. By a space L1loc (Ω) we understand the set of all Lebesgue measurable in
Ω functions having absolute value integrable on each compact subset of the set Ω. By a multi-index α we
understand a vector (α1 , · · · , αn ) having natural components αi . We set |α| = α1 + · · · + αn . We also define
the partial derivative
∂ |α| φ
Dα φ = . (6)
∂xα
1
1
· · · ∂xα
n
n

We are now able to define the notion of weak derivative.

Definition 5. A locally integrable function v (element of L1loc (Ω)) is called the α-th weak derivative of
u ∈ L1loc (Ω), if it satisfies
Z Z
|α|
φvdx = (−1)|α| uDα φdx for all φ ∈ C0 (Ω). (7)
Ω Ω

Note that the weak derivative is uniquely determined up to the set of measure zero. We call a function
weakly differentiable if all its partial derivatives of first order exist and k times weakly differentiable if all
its weak derivatives exist for orders up to and including k. Denote the linear space of k times weakly
differentiable functions by W k (Ω). Clearly C k (Ω) ⊂ W k (Ω). The concept of weak derivatives then extends
the concept of the classical derivatives.
 3

There is no time here to introduce the (more general) notion of the distributional derivative. The weak
derivative is a special kind of the distributional derivative.

1.3. Lp spaces. Let Ω be a bounded domain in Rn . By a measurable function we shall mean an equivalent
class of measurable functions on Ω which differ only on a subset of measure zero. The supremum and infimum
of a measurable function will be understood as the essential supremum or essential infimum respectively.
Let p ≥ 1, by Lp (Ω) we denote the Banach space consisting of all measurable functions on Ω thats
p−powers are integrable. The norm in Lp (Ω) is defined by:
µZ ¶ p1
p
kukLp (Ω) = |u| dx . (8)
Ω

When p = ∞, we set:
kukL∞ (Ω) = esssup|u|. (9)

There are a few elementary inequalities of the importance for the future that we shall quote now. The
first is called Young inequality:
ap bq
ab ≤ + , (10)
p q
which holds for positive reals a, b, p, q that satisfy additionally
1 1
+ =1 (11)
p q
(the, so called, condition of Hölder conjugacy). With the same exponents as above the Hölder inequality is
satisfied:
Z
uvdx ≤ kukLp (Ω) kvkLq (Ω) . (12)
Ω
It is a consequence of the Young’s inequality. As a consequence of the Hölder inequality the interpolation
inequality for Lp spaces is satisfied:
kukLq (Ω) ≤ kukλLp (Ω) kuk1−λ
Lr (Ω) , (13)
1 λ (1−λ)
valid for u ∈ Lr (Ω) with p ≤ q ≤ r and q = p + r .
The further and more involved inequality we can face is the, so called, Poincaré inequality. Several versions
of it are known, so we can quote only one of the best known among them.
Lemma 1. Let Ω be a domain with continuous boundary. Let 1 ≤ p < ∞. Then for any u ∈ W 1,p (Ω):
Z ¯ Z ¯p Z Xn ¯ ¯
¯ ¯ ¯ ∂u ¯p
¯u(x) − 1 ¯
u(y)dy ¯ dx ≤ c ¯ ¯
¯ |Ω| ¯ ∂xi ¯ dx. (14)
Ω Ω Ω i=1

Proof. The proof is equivalent with showing that:

Z Z X n ¯ ¯
¯ ∂u ¯p
|u(x)|p dx ≤ c ¯ ¯ dx, (15)
¯ ¯
Ω Ω i=1 ∂xi

for the W 1,p (Ω) functions having zero main value.

Assume, at contrary, that there exists a sequence {uk } ⊂ W 1,p (Ω), such that kuk kW 1,p (Ω) = 1 and
Z Z X n ¯ ¯
¯ ∂uk ¯p
p
1 = kuk kW 1,p (Ω) ≥ p
|uk | dx > k ¯ ¯ dx. (16)
¯ ¯
Ω Ω i=1 ∂xi
4

Equivalently, this can be written as:

Z X n ¯ ¯
¯ ∂uk ¯p 1
¯ ¯
¯ ∂xi ¯ dx < k . (17)
Ω i=1

Using the result of compactness of the embedding W 1,p (Ω) ⊂ Lp (Ω) (that can be proved later), we are thus
able to find a subsequence {ukl } convergent in Lp (Ω) to u. Thanks to (17) we have also, ukl → u in W 1,p (Ω).
∂u
But this same estimate shows also, that ∂x i
= 0 a.e. in Ω. It can be shown, that u = const. in Ω. Because of
the zero mean (the property preserved from the sequence) u must be equal zero (a.e. in Ω). This contradicts
the property kukW 1,p (Ω) = 1. ¤

Similar estimate, even beter known, is valid for the functions equal zero on ∂Ω, more precisely belonging
to the space W01,p (Ω). For more general versions of the Poincaé inequality consult [ZI], p.182.
Finally we present an elementary proof of a version of the interpolation inequality. We have:
Theorem 2. Let u ∈ W0k,p (Ω). Then, for any ε > 0, 0 < |β| < k
|β|
kDβ ukLp (Ω) ≤ εkukW k,p (Ω) + Cε |β|−k kukLp (Ω) , (18)
where C = C(k).

Proof. We establish (18) for the case |β| = 1, k = 2. The rest can be obtained through a suitable induction
argument.
Assume at first that u ∈ C02 (R) and consider an interval (a, b) of the length b − a = ε. For x0 ∈
(a, a + 3ε ), x00 ∈ (b − 3ε , b), by the mean value theorem,
¯ ¯
¯ u(x0 ) − u(x00 ) ¯ 3
|u0 (x̄)| = ¯¯ ¯ ≤ (|u(x0 )| + |u(x00 )|),
¯ ε (19)
x0 − x00
Rx
for some x̄ ∈ (a, b). Consequently, for any x ∈ (a, b), u0 (x) = u0 (x̄) + x̄ u”(s)ds, so that
Z b
3
|u0 (x)| ≤ (|u(x0 )| + |u(x00 )|) + |u00 |dx. (20)
ε a

Integrating the above with respect to x0 and x00 over the intervals (a, a + 3ε ), (b − 3ε , b) respectively, we get
ÃZ ! Z
a+ 3²
² 0 3 0 0 ² ² b
|u (x)| ≤ |u(x )|dx + |u(x”)| + |u”|dx,
3 ² a 3 3 a

Z a+ 3² Z b Z b
¡ ² ¢2 ¡ ² ¢2
3 |u0 (x)| ≤ |u(x0 )|dx0 + |u(x”)|dx” + 3 |u”|dx
a b− 3² a
Z b Z
¡ ² ¢2 b
≤2 |u(x)|dx + 3 |u”|dx
a a

giving
Z b Z b
18
|u0 (x)| ≤ |u00 |dx + |u|dx. (21)
a ε2 a

By Hölder inequality,
( Z Z )
b b
0 p p−1 p−1 00 p (18)p p
|u (x)| ≤ 2 ε |u | dx + p+1 |u| dx . (22)
a ε a
 5

Integrating with respect to x over (a, b), we obtain:

Z b ( Z µ ¶p Z )
b b
0 p p−1 18
|u (x)| dx ≤ 2 εp 00 p
|u | + p
|u| dx . (23)
a a ε a

Dividing R into intervals of length ε, adding such inequalities, we obtain:

Z ½ Z µ ¶p Z ¾
18
|u0 |p dx ≤ 2p−1 εp |u00 |p dx + |u|p dx (24)
R R ε R
Which is the desired result in the one-dimensional case. To extend to the higher dimensions fix i, 1 ≤ i ≤ n
and apply the above estimate to u ∈ C02 (Ω) regarded as a function of xi only. Integrating successively over
the remaining variables we obtain
Z ½ Z ¾
p p−1 p p Cp
|Di u| dx1 · · · dxn ≤ 2 ε |Dii u| dx1 · · · dxn + kukLp (Ω) , (25)
Rn Rn ε
so that
C
kDi ukLp (Ω) ≤ εkDii ukLp (Ω) + kukLp (Ω) , (26)
ε
with C = 36. ¤

We call [AD], [G-T], [TA] for further versions of such theorems. We quote here a version formulated in [G-T],
p. 173:
Theorem 3. Let Ω be a C 2 domain in Rn and u ∈ W k,p (Ω). Then, for any ε > 0, 0 < |β| < k,
|β|
kDβ ukLp (Ω) ≤ εkukW k,p (Ω) + Cε |β|−k kukLp (Ω) , (27)
where C = C(k, Ω).

2. Sobolev spaces

The main body of the lecture will be given in this part, including:
• Definitions of Sobolev spaces and their basic properties.
• Dense subsets and approximation by smooth functions of elements of Sobolev spaces.

2.1. Definition of the Sobolev spaces. Define a functional kukW m,p (Ω) , where m is a nonnegative integer
and 1 ≤ p ≤ ∞, as follows:
 nP o p1
 α p
kukW m,p (Ω) = kD uk , if 1 ≤ p < ∞,
0≤|α|≤m Lp (Ω) (28)
 kuk m,∞ = max kDα uk ∞ .
W (Ω) 0≤|α|≤m L (Ω)

We are now able to define the Sobolev spaces as:

W m,p (Ω) = {u ∈ Lp (Ω) : Dα u ∈ Lp (Ω), 0 ≤ |α| ≤ m} . (29)
Further, we set,
W0m,p (Ω) = the closure of C0∞ (Ω) in W m,p (Ω). (30)
m,p 0,p p
Equipped with the above defined norms, W (Ω) are called Sobolev spaces. Clearly also, W (Ω) = L (Ω),
and for 1 ≤ p < ∞, W00,p = Lp (Ω). Moreover, for any m,
W0m,p (Ω) ⊂ W m,p (Ω) ⊂ Lp (Ω). (31)
m,p
The spaces W (Ω) has been introduced by S.L. Sobolev [SO], now there are many similar, but not neces-
sarily coinciding spaces known in the literature.
6

m,p
Several important properties of the Sobolev spaces are most easily obtained
P by regarding W (Ω) as
p
a closed subspace of the Cartesian product of the spaces L (Ω). Let N = 0≤|α|≤m 1 be the number of
QN
multi-indices α satisfying 0 ≤ |α| ≤ m. For 1 ≤ p ≤ m, let LpN = j=1 Lp (Ω), the norm of u = (u1 , · · · , uN )
in LpN being given by:
  p1
N
X
kukLpN =  kuj kpLp (Ω)  , (32)
j=1

for 1 ≤ p < ∞ and, for p = ∞,

kukL∞
N
= max1≤j≤N kuj kL∞ (Ω) . (33)
p m,p
Known properties of the spaces L (Ω) allow, through this characterization, to show that the spaces W (Ω)
are Banach spaces, they are separable when 1 ≤ p < ∞ also reflexive and uniformly convex when 1 < p < ∞.
When p = 2 then W 2,p (Ω) are Hilbert spaces equipped with the scalar product:
X
(u, v)m = (Dα u, Dα v), (34)
0≤|α|≤m
R
where (u, v) = Ω
u(x)v̄(x)dx is the scalar product in L2 (Ω).

2.2. Dense subsets and approximation in Sobolev spaces. Typically dense subsets of the Sobolev
spaces are constructed using the idea of approximation with mollifiers.
R Let ρ be a non-negative C ∞ (Rn )
function that vanishes outside a unit ball. We also assume that ρ(x)dx = 1. Such a function is called a
mollifier. An example of the mollifier is given by:
µ ¶
1
ρ(x) = c exp , when |x| < 1, (35)
|x|2 − 1
extended with zero outside a unit ball. The constant c is choosen just to normalize the integral of ρ to one.
For any element u ∈ L1loc (Rn ) and h > 0 the regularization of u is defined as:
Z µ ¶
x−y
uh (x) = h−n ρ u(y)dy, (36)
Ω h
when h < dist(x, ∂Ω). Thanks to the properties of the convolutions, uh ∈ C ∞ (Ω0 ) in any subdomain Ω0
strictly included in the domain Ω. What we need to see now is that the functions uh nicely approximate u
as h tends to 0 (in various possible norms). We have:
Lemma 2. For u ∈ C 0 (Ω), the corresponding uh converges to u uniformly on any subdomain Ω0 strictly
included in Ω.

Proof. Changing the variable, we obtain:

Z µ ¶ Z
x−y
uh (x) = h−n ρ u(y)dy = ρ(z)u(x − hz)dz. (37)
|x−y|≤h h |z|≤1

Hence, for a strictly included subdomain Ω0 ⊂ Ω with 2h < dist(Ω0 , ∂Ω),

Z
supΩ0 |u − uh | ≤ supx∈Ω0 ρ(z)|u(x) − u(x − hz)|dz
|z|<1 (38)
≤ supx∈Ω0 sup|z|≤1 |u(x) − u(x − hz)|.
Now, the uniform continuity of u over the set
Bh (Ω0 ) = {x : dist(x, Ω0 ) < h}, (39)
0
shows the uh tends to u uniformly over Ω . ¤
 7

A similar easy proof shows that when u ∈ Lp (Ω) then its mollifier uh converges to u in Lp (Ω). For the
proof see [G-T], p.148.
We can state now a generalizatin of this lemma concerning approximation in the W m,p (Ω) sense (see
[AD], Chapter III). We have:
Lemma 3. Let u ∈ W m,p (Ω) and 1 ≤ p < ∞. Then, for any strict subdomain Ω0 ⊂ Ω holds; uh → u in
W m,p (Ω0 ).

Proof. Define ρh = h−n ρ( hx ). Let ε < dist(Ω0 , ∂Ω) and h < ε. For any φ ∈ C0∞ (Ω0 )
Z Z Z
ρh ∗ u(x)Dα φ(x)dx = ū(x − y)ρh (y)Dα φ(x)dxdy
Ω0 Rn Rn
Z Z
|α|
= (−1) Dxα ū(x − y)ρh (y)φ(x)dxdy (40)
Rn R n
Z
= (−1)|α| ρh ∗ Dα u(x)φ(x)dx,
Ω0

where ū denotes the zero extension of u outside Ω. Therefore, Dα ρh ∗ u = ρh ∗ Dα u, in the sense of

distributions in Ω0 . Since Dα u ∈ Lp (Ω) for 0 ≤ |α| ≤ m, by the previous lemma
limh→0 kDα ρh ∗ u − Dα ukLp (Ω0 ) = limh→0 kρh ∗ Dα u − Dα ukLp (Ω0 ) = 0. (41)
Thus limh→0 kρh ∗ u − ukW m,p (Ω0 ) = 0. ¤

For more informations concerning dense subsets of Sobolev spaces consult [AD], Chapter III.

3. Embeddings of Sobolev spaces

The importance of Sobolev spaces lies in their connections with the spaces of continuous and uniformly
continuous functions. This is expressed in embedding theorems:
• Continuous embeddings between Sobolev spaces and the spaces of continuous and Hölder continuous
functions.
• Compact embeddings of Sobolev spaces. Rellich-Kondrachov theorem.

3.1. Continuous embeddings of Sobolev spaces. We start this section with the proof of the, classical
today, embedding result which is due to S.L. Sobolev. Recall first a generalization of the Hölder inequality
which is due to E. Gagliardo (see [NE 1]).
Lemma 4. Let C = (−1, 1)n , C 0 = (−1, 1)n−1 and let
fi (x0 ) = fi (x1 , , x2 , · · · , xi−1 , xi+1 , · · · , xn ) ∈ Ln−1 (C 0 ).
Put fi (x) = fi (x0 ) in C. Then:
Z Y
n n µZ
Y ¶ n−1
1

n−1 0
|fi |dx ≤ |fi | dx . (42)
C i=1 i=1 C0

We will not include the proof here.

We now have:
1 1
Theorem 4. Let Ω be a Lipschitz domain. Let 1 ≤ p < n and q = p − n1 . Then W 1,p (Ω) ⊂ Lq (Ω), i.e. the
identity mapping from W 1,p (Ω) to Lq (Ω) is bounded.
8

Proof. It can be shown, that suffices to prove the theorem when Ω is a cube. Consider first p > 1. Let
u ∈ W 1,p (C) and C be the cube from the previous lemma. There exists a sequence {uk } ⊂ C 1 (C̄) convergent
to u in W 1,p (C) and uk (x0 , 1) = 0. Consider a power:
np−p
|uk (x)| n−p .
This function is absolutely continuous on all the parallels to the axis and almost everywhere on such parallel:
¯ ´¯¯ np − p ¯ ¯
¯ ∂ ³ np−p np−n ¯ ∂u ¯
k¯
¯ |u | ¯ = |u | n−p ¯
¯ ∂xi ¯ . (43)
n−p
¯ ∂xi k ¯ n−p
k

So, we have next:

Z 1 ¯ ¯
np−p np − n np−p ¯ ∂uk ¯
max |uk (x)| ≤ |uk | ¯ ¯
¯ ∂xi ¯ dxi . (44)
n−p n−p
−1≤xi ≤1 n−p 0

Hence
Z
np−p
max |uk (x)| n−p dx0
C0 −1≤xi ≤1
µZ ¶ p−1 µZ ¯ ¯ ¶ p1 (45)
np − p np p ¯ ∂uk ¯p
≤ |uk | n−p dx ¯ ¯ dx .
n−p ¯ ¯
C C ∂xi

It thus follows from the previous lemma, that:

Z Z ÃY
n
!
np p
|uk (x)| n−p dx ≤ max |uk (x)| n−p dx
C C −1≤xi ≤1
i=1
n µZ
Y ¶ n−1
1
np−p
≤ max |uk (x)| n−p dx0 (46)
i=1 C 0 −1≤xi ≤1
µ ¶ n−1
n µZ ¶ np−n
np − p np np−p n
n−1
≤ |uk | n−p dx kuk kW 1,p (C) ,
n−p C

therefore
µZ ¶ n−p
np np
np − p
|uk (x)| n−p dx ≤ kuk kW 1,p (C) , (47)
C n−p
and, as k → ∞, we get the result for p > 1. If p = 1, it sufficies to pass with p to 1. The proof is
completed. ¤

Consider a subspace C k,µ (Ω̄) of C k (Ω), consisting of all such functions, whose k-th partial derivatives are
µ-Hölder continuous. The norm in this space is introduced through the formula:
X |Dα u(x) − Dα u(y)|
kukC k,µ (Ω̄) = kukC k (Ω) + supx6=y . (48)
|x − y|µ
|α|=k

The following continuous embedding of the Sobolev spaces is now valid:

n
Theorem 5. Let Ω be a domain with Lipschitz boundary. Let p > n and µ = 1 − p. Then W 1,p (Ω) ⊂
C 0,µ (Ω).

For the proof see e.g. [NE], p. 28.

Let’s only quote a general theorem, the two above embeddings are only special cases of which.
 9

Proposition 1. If the domain Ω in Rn has the cone property, the following embeddings are continuous:
np
W j+m,p (Ω) ⊂ W j,q (Ω), when p ≤ q ≤ . (49)
n − mp
If Ω is a Lipschitz domain, then:
n
W j+m,p (Ω) ⊂ C j,λ (Ω̄), for 0 < λ ≤ m − . (50)
p

For the proof of the above, see [AD], p. 97.

3.2. Compact embeddings of Sobolev spaces. In the applications of Sobolev spaces to partial differ-
ential equations it is often important to know whether the embeddings of Sobolev spaces are compact. We
present below an example of such a result, quoting here also a different proof in [NE], p. 29. We start with
two lemmas.
Lemma 5. Let Q be a cube {0 ≤ xi ≤ σ} ⊂ Rn . Let u be a real function in C 1 (Q). Then:
µZ ¶2
2 1 n
kukL2 (Q) ≤ n u(x)dx + σ 2 kuk2H 1 (Q) . (51)
σ Q 2

Proof. For any x, y ∈ Q,

Z x1
∂
u(x1 , · · · , xn ) − u(y1 , · · · , yn ) = u(ξ1 , x2 , · · · , xn )dξ1
y1 ∂ξ1
Z x2 Z xn (52)
∂ ∂
+ u(y1 , ξ2 , · · · , xn )dξ2 + · · · + u(y1 , · · · , yn−1 , ξn )dξn .
y2 ∂ξ2 yn ∂ξn

Taking squares and using the Schwarz inequality,

Z σµ ¶2
2 2 ∂
u (x) + u (y) − 2u(x)u(y) ≤ nσ u(ξ1 , x2 , · · · , xn ) dξ1
0 ∂ξi
Z σµ ¶2 Z σµ ¶2 (53)
∂ ∂
+ nσ u(y1 , ξ2 , · · · , xn ) dξ2 + · · · + nσ u(y1 , · · · , yn−1 , ξn ) dξn .
0 ∂ξ2 0 ∂ξn
Integrating with respect to x1 , · · · , xn , y1 , · · · , yn , we get:
Z µZ ¶2 Xn Z µ ¶2
∂u(x)
2σ n u2 (x)dx − 2 u(x)dx ≤ nσ n+2 dx, (54)
Q Q i=1 Q
∂xi

implying (51). ¤

We have next the following Friedrich’s Inequality:

Lemma 6. For any ε > 0 there exists an integer M > 0 and a real-valued functions w1 , · · · , wM in L2 (Ω),
the domain Ω bounded, that kwj kL2 (Ω) = 1, and for any real-valued function u ∈ H01 (Ω):
M
X
kuk2L2 (Ω) ≤ ²kuk2H 1 (Ω) + (u, wj )2 . (55)
j=1

Proof. It suffices to prove the lemma for u ∈ C01 (Ω), the validity for all u ∈ H01 (Ω) then follow by completion.
Extend u outside Ω with 0 and let Q be a cube containing Ω̄ having edges parallel to the coordinate axes.
Let σ0 be the length of each edge. Divide Q into cubes by introducing hyperplanes xi = ymi , so that
¡ ¢2
ym+1,i − ym,i = σ for all m, i; σ is such that σσ0 is an integer. Denote by Q1 , · · · , QM (M = σσ0 ) the cubes
thus obtained.
10

By the previous lemma,

ÃZ !2
¡ ¢2 1 n 2
kukL2 (Qj ) ≤ n udx + σ kuk2H 1 (Qj ) . (56)
σ Qj 2
Summing over j and introducing the functions:
½ −n
σ 2 in Qj
wj =
0 outside Qj
nσ 2
we get the result provided σ is such that 2 ≤ ε. ¤

We are now able to formulate and prove the famous Rellich’s Theorem:
Theorem 6. Let {um } be a sequence of functions in H01 (Ω) (Ω bounded) such that kum kH 1 (Ω) ≤ const. < ∞.
There exists a subsequence {um0 } convergent in L2 (Ω).

Proof. Let h = 1, 2, · · · , and take for any ε = h1 a finite sequence wjh (j = 1, · · · , M (h)) as in the Friedrich’s
Inequality. Let {um1 } be a subsequence of {um } such that {(um1 , wj1 )L2 } are convergent sequences for
j = 1, 2, · · · , M (1). Similarly we define {um,h+1 }, inductively, to be a subsequence of {umh } such that
{(um,h+1 , wj,h+1 )L2 } are convergent sequences for j = 1, 2, · · · , M (h+1). Let {um0 } be the diagonal sequence
{umm }. Then (um0 , wjh )L2 is convergent as m0 → ∞ , for any h, j.
1
Given ε = h choose m00 such that
M (h)
X
(um0 − uk0 , wjk )2L2 < ε if m0 ≥ m00 , k 0 ≥ m00 . (57)
j=1

Then, by the Friedrich’s Inequality,

2
kum0 − uk0 k2L2 ≤ ε + ε (kum0 kH 1 + kuk0 kH 1 ) ≤ (1 + 4K 2 )ε, (58)
where K = supkum kH 1 . Since ε can be made arbitrarily small, the proof is completed. ¤

Following [AD] we will now quote the general result concerning compact embeddings of Sobolev spaces.
We have:
Proposition 2. Let Ω be a bounded domain in Rn having the cone property. Then the following embeddings
are compact:
n n
W j+m,p (Ω) ⊂ W j,q (Ω) if 0 < n − mp and j + m − ≥ j − , (59)
p q
also,
W j+m,p (Ω) ⊂ C j (Ω̄) if mp > n. (60)

We send to [AD] for the proofs.

4. Applications of Sobolev spaces

Sobolev spaces were introduced mostly for the use of the theory of partial differential equations. Dif-
ferential operators are often closable in such spaces. The advantage of using Sobolev spaces will be briefly
discussed in this part of the lecture.
Of course, Sobolev spaces being examples of Banach or, sometimes, Hilbert spaces are interesting object for
themselves. But their importance is connected with the fact that the theory of partial differential equations
can be, and even most easily, developed just in such a spaces. The reason is because partial differential
operators are very well situated in Sobolev spaces.
 11

Simultaneously, the spaces of continuous (or of class C k ) functions is not very suitable for the studies of
partial differential equations. Instead we need to consider Hölder continuous or Hölder continuous together
with the derivatives; C k,µ classes of functions. These are the other, competitive with the Sobolev spaces,
classes of functions in which p.d.e. can be studied. Why spaces of (only) continuous functions are not very
suitable? The answer is connected with the following observation ([NI], Chapt. 2.5).
Namely, for every k ∈ N the Laplace operator
∆ : C k+2 → C k , u = 0 on ∂Ω, (61)
k 0
is continuous, but its image is not closed in C (Ω̄). In particular, for continuous right hand side f ∈ C (Ω̄)
the solution of the equation:
∆u = f, u = 0 on ∂Ω, (62)
in general must not be a C 2 (Ω̄) function. Similar situation we face for other elliptic operators.

4.1. Closedness of differential operators in Sobolev spaces. Besides the property, that elliptic opera-
tors are closed (or, closable) in the Sobolev spaces, another important property of such spaces are their nice
and reach connections with other function spaces expressed in the previously discussed embedding theorems.
Working with solutions of partial differential equations lying in smooth Sobolev spaces (when the parameters
m, p describing the space are large) it is very easy to check that these solutions are directly elements of the,
say, C k .
We shall describe below the problem of closedness of differential operators in Sobolev spaces more carefully.
We start with the definition:
Definition 6. Let T be a linear operator from a linear vector space X into a linear vector space Y , having
a domain DT . The graph GT of T is the set of points (x, T x) in X × Y , where x ∈ DT . If GT is closed in
the Cartesian product X × Y , we say that T is a closed operator.

We next have:
Definition 7. Let T be a linear operator from a linear vector space X into a linear vector space Y , having
domain DT . A linear operator S from X into Y is called an extension of T if DT ⊂ DS and T x = Sx for
all x ∈ DT .

If, for a linear operator T there is an operator S which is closed, linear and extends T and, for any operator
S 0 having these two properties S 0 needs to be an extension of S, then S is called the closure of T . It will be
denoted by T̄ .
The following is a well known equivalent condition for T to have a closure:
Lemma 7. Let T be a linear operator from a linear subspace DT of a Banach space X into a Banach space
Y . T has a closure T̄ if and only if the following condition is satisfied:
xn ∈ DT , xn → 0, T xn → y imply y = 0. (63)

We will not prove it here.

Consider now a partial differential operator with constant coefficients aα :
X
P (D) = aα Dα . (64)
|α|≤m

It can be considered as an operator from C ∞ into itself, or as an operator from C k (Ω) into C k−m (Ω), for
any m ≤ k. But we prefer to define it as:
X
[P (D)u](x) = aα Dα u(x), (65)
|α|≤m
12

on the linear subspace Ĉ m (Ω) consisting of all the functions in C m (Ω) that have a finite norm:
  p1
 X Z 
kukW m,p (Ω) = |Dα u|p dx . (66)
 Ω 
|α|≤m

It is often desirable to extend P (D) to a closed linear operator in Lp (Ω), (1 ≤ p < ∞). We shall prove, this
is possible.
Theorem 7. The operator P (D) from Lp (Ω) p ∈ [1, ∞) into itself with a domain Ĉ m (Ω) has a closure.

Proof. In view of the necessary condition above it suffices to show that if uk ∈ Ĉ m (Ω), kuk kW 0,p (Ω) →
0, kP (D)uk − vkW 0,p (Ω) → 0, then v = 0. Let φ ∈ C0∞ (Ω). Integration by parts gives:
Z Z
P (D)uk φdx = uk P (−D)φdx. (67)
Ω Ω
R
As k → ∞, the integrals on the right converge to zero whereas the integrals on the left converge to Ω vφdx.
Therefore,
Z
vφdx = 0 (68)
Ω
for all φ ∈ C0∞ (Ω). Thus, by density, we conclude that v = 0. The proof is completed. ¤

This result show that the Sobolev spaces are natural for the studies of differential operators. One can also
choose the spaces of Hölder continuous functions as has been done in [FR].

4.2. The Lax-Milgram lemma. Another advantage of the Sobolev spaces is connected with the fact that
the study of elliptic p.d.e. becames very simple and elegant within these spaces. We quote here one such very
known result. It justifies weak solvability of elliptic equations in Hilbert approach. The following abstract
result known as the Lax-Milgram lemma is used in this proof.
Assume that Ω is a bounded domain in Rn . Then we have
Theorem 8. Let B[x, y] be a bilinear form (that is, linear in x and antilinear in y) in a Hilbert space H
with norm k k and scalar product ( , ) and let B[x, y] be bounded;
|B[x, y]| ≤ const.kxkkyk for x, y ∈ H. (69)
Suppose further that
|B[x, x]| ≥ ckxk2 for all x ∈ H (70)
for some positive constant c. Then every bounded linear functional F (x) in H can be represented in the
form:
F (x) = B[x, v] = B[w, x] (71)
for some elements v, w ∈ H that are uniquely determined by F .

Proof. For fixed v, B[x, v] is a bounded linear functional. Hence there exists a unique y such that:
B[x, v] = (x, y). (72)
Set y = Av. This A is a bounded linear operator in H. Since:
ckvk2 ≤ |B[v, v]| ≤ |(v, x)| ≤ kvkkyk, (73)
or kvk ≤ kykc ;A has a bounded inverse. Hence the range of A, R(A) is a closed linear subspace of H. We
claim that R(A) = H. Indeed, if R(A) 6= H then there is an element z 6= 0 orthogonal to R(A); (z, Av) = 0
for all v ∈ H. This implies that B[z, v] = (z, Av) = 0. Taking v = z, we get B[z, z] = 0. Hence, by (70)
z = 0. A contradiction.
 13

Consider now the functional F (x). By the theorem of Riesz (in Hilbert spaces) there is an element b in
H such that F (x) = (x, b) for x ∈ H. Since R(A) = H, there is an element v such that Av = b. Hence
F (x) = (x, b) = (x, Av) = B[x, v], (74)

and v is uniquely determined. Indeed, if B[x, v 0 ] = F (x) for some v 0 ∈ H and for all x ∈ H, then B[x, v−v 0 ] =
0 for all x. Taking x = v − v 0 and using (70) we get v = v 0 .
The representation F (x) = B[w, x] follows by applying the previous result to the bilinear form B[y, x]. ¤

Consider now a differential operator of order 2m in a bounded domain Ω

X
Lu = aα (x)Dα u. (75)
|α|≤2m

having the coefficients aα in C |α|−m for m < |α| ≤ 2m; then it can be rewritten in the divergence form:
X
Lu = (−1)|ρ| Dρ (aρσ (x)Dσ u). (76)
0≤|ρ|,|σ|≤m

The condition of strong ellipticity has now the form:

 
 X 
Re ξ ρ aρσ (x)ξ σ ≥ c0 |ξ|2m (c0 > 0), (77)
 
|µ|=|σ|=m

for any real vector ξ. We associate with L the bilinear form:

X
B[u, v] = (aρσ Dρ v, Dσ u). (78)
0≤|ρ|,|σ|≤m

Under the following assumptions:

• L is strongly elliptic in Ω with a module of strong ellipticity c0 independent of x in Ω,

• The coefficients aρσ are bounded in Ω by a constant c1 ,
• The principal coefficients of L have modulus of continuity c2 (t);
|aρσ (x) − aρσ (y)| ≤ c2 (|x − y|) if |ρ| = |σ| = m, x ∈ Ω, y ∈ Ω, (79)

and c2 (t) → 0 if t → 0,

we have the following theorem:

Theorem 9. Let L satisfies the assumptions above and let the associated bilinear form B[φ, u] satisfies
ReB[φ, φ] ≥ ckφk2H m (Ω) for all φ ∈ H0m (Ω), (80)
with some c > 0. Then there exists a unique solution of the Dirichlet problem
(
Lu = f in Ω,
(81)
u = 0 on ∂Ω.

Proof. Evidently the bilinear form B[φ, u] and the functional F (ψ) = (ψ, f ) satisfy the assumptions of the
Lax-Milgram lemma with H = H0m (Ω). There exists thus a unique u ∈ H0m that
(ψ, f ) = B[ψ, u] for all ψ ∈ H0m , (82)
and this proves the theorem, since such a u is the generalized solution to the above Dirichlet problem. ¤
14

We leave here the discussion of the further possible applications of the Sobolev spaces in the theory of partial
differential equations. Of course there are much more important points that can be discussed there.
There is a huge literature devoted to the theory of Sobolev spaces. Usually in any larger monograph
devoted to partial differential equations some elements of this theory can be found. There are also some
monographs devoted specially to Sobolev Spaces like [AD], [MR], [MA]. A short list of the references is
presented in order. Of course, in presentation of the lecture only a small part of the material given in
monographs below has been given.
 15

5. References

[AD] R.A. Adams, Sobolev Spaces, Academic Press, New York, 1975.
[EV] L.C. Evans, Partial Differential Equations, AMS, Providence, RI, 1998.
[FR] A. Friedman, Partial Differential Equations, Holt, Rinehart and Winston, New York, 1969.
[FR 1] A. Friedman, Foundations of Modern Analysis, Dover Publications, Inc., New York, 1982.
[G-T] D. Gilbarg, N. Trudinger, Elliptic Partial Differential Equations of Second Order, Springer-Verlag,
Berlin, 1983.
[K-S] D. Kinderlehrer, G. Stampacchia, An Introduction to Variational Inequalities and Their Applications,
Academic Press, New York, 1980.
[MR] J.T. Marti, Introduction to Sobolev Spaces and Finite Element Solution of Elliptic Boundary Value
Problems, Academic Press, London, 1986.
[MA] V.G. Maz’ja, Sobolev Spaces, Springer-Verlag, Berlin, 1985.
[MO] C. B. Morrey, Jr., Multiple Integrals in the Calculus of Variations, Springer-Verlag, New York, 1966.
[NE] J. Nečas, Introduction to the Theory of Nonlinear Elliptic Equations, Teubner, Leipzig, 1983.
[NE 1] J. Nečas, Les Méthodes Directes en Théorie des Équations Elliptiques, Academia, Prague, 1967.
[NI] L. Nirenberg, Topics in Nonlinear Functional Analysis, Courant Institute, New York, 1974.
[SO] S.L. Sobolev, On a theorem of functional analysis, Mat. Sb. 4 (46) 1938, 39-68 (translated into English
in 1963).
[TA] H. Tanabe, Functional Analytic Methods for Partial Differential Equations, Dekker, New York, 1997.
[TR] H. Triebel, Interpolation Theory, Function Spaces, Differential Operators, Veb Deutscher, Berlin, 1978.
[ZI] W.P. Ziemer, Weakly Differentiable Functions, Sobolev Spaces and Functions of Bounded Variation,
Springer-Verlag, New York, 1989.