Lectures on

Sobolev Spaces

S. Kesavan
The Institute of Mathematical Sciences,
Chennai.
e-mail: [email protected]
2
1 Distributions
In this section we will, very briefly, recall concepts from the theory of dis-
tributions that we will need in the sequel. For details, see Kesavan [1],
Chapter 1.
Throughout these lectures, we will be working with an open set Ω ⊂ RN .
Let us briefly motivate our study of distributions and Sobolev spaces.
One of the important partial differential equations that we often study is the
Laplace’s equation:
−∆u = f in Ω
together with some appropriate boundary condition. It turns out that in
elasticity and structural engineering, the importance of the solution u stems
from the fact that it minimizes, amongst ‘admissible functions’ v, the energy
functional Z Z
1 2
J(v) = |∇v| dx − f v dx.
2 Ω Ω
Now it can happen, in many applications, that the function f is not
continuous but just, say, in L2 (Ω). Then, for the second term in the above
expression for J to make sense, it follows that v must also belong to L2 (Ω).
The first term in J will make sense if all the first partial derivatives of u, i.e.
∂u
∂xi
, 1 ≤ i ≤ N are all in L2 (Ω) as well.
But, when we are dealing with functions in L2 (Ω), what do we mean by its
derivatives? This is where we need to generalize the notion of a function and
its derivatives and interpret partial differential equations in the new set-up.
The framework for this comes from the theory of distributions. Just as we
can think of a real number as a linear operator on R acting by multiplication,
we can consider certain functions as linear operators on some special space
of functions and then generalize this.
Definition 1.1 Let f : Ω → R be a continuous function. Its support,
denoted supp(f ), is the closure of the set where f is non-zero. The function
is said to be of compact support in Ω if the support is a compact set
contained inside Ω. 

Definition 1.2 The space of test functions in Ω, denoted D(Ω), is the

space of all C ∞ functions defined on Ω which have compact support in Ω. 

The space D(Ω) is a very rich space and can be made into a locally convex
topological vector space.
Definition 1.3 The dual of D(Ω) is called the space of distributions,
denoted D0 (Ω), on Ω and its elements are called distributions on Ω. 
Example
R 1.1 A function u : Ω → R is said to be locally integrable
if K |u| dx < ∞ for every compact subset K of Ω. A locally integrable
function defines a distribution in the following way: if ϕ ∈ D(Ω), then
Z
u(ϕ) = uϕ dx
Ω

3
where we use the same symbol u for the function as well as the distribution
it generates, for, if u and v generate the same distribution, then it can be
shown that u = v a.e. Now every smooth function and every function in any
of the spaces Lp (Ω), for 1 ≤ p ≤ ∞, are all locally integrable and so they all
can be considered as distributions. 

Example 1.2 Define

T (ϕ) = ϕ(0)
for all ϕ ∈ D(R). This defines a distribution on R and it can be shown that
it cannot be obtained from any locally integrable function in the sense of
Example 1.1 above. This is called the Dirac distribution supported at the
origin and is denoted usually by the symbol δ.

Example 1.3 Let µ be a measure on Ω such that µ(K) < ∞ for every
compact subset K contained in Ω. Then it defines a distribution given by
Z
Tµ (ϕ) = ϕ dµ
Ω

for every ϕ ∈ D(Ω). The Dirac distribution mentioned above is just the
distribution generated by the Dirac measure supported at the origin.

Example 1.4 Define the distribution T on R by

T (ϕ) = ϕ0 (0)

for every ϕ ∈ D(R). This distribution does not fall into any of the categories
of the preceding examples and is an entirely new object. It is called the
dipole distribution. 

Assume that u : R → R is a smooth function. Then, so is its derivative u0 ,

and both of them define distributions as in Example 1.1. Now, if ϕ ∈ D(R),
then, by integration by parts, we have
Z Z
u ϕ dx = − uϕ0 dx.
0
R R

Similarly, Z Z Z
00 0 0
u ϕ dx = − u ϕ dx = uϕ00 dx
R R R
and so on. We can use this to define the derivative of any distribution T on
R as follows: if T is a distribution on R, then, for any positive integer k,
k
define the distribution ddxTk by

dk T
k
(ϕ) = (−1)k T (ϕ(k) )
dx
k
for any ϕ ∈ D(R), where ϕ(k) = ddxϕk . We can do this on any open subset
Ω ⊂ RN , for any space dimension N . To describe it we first establish some

4
useful notation.

Notation
A multi-index α is an N -tuple of non-negative integers. Thus,

α = (α1 , · · · , αN )

where the αi are all non-negative integers. We define

N
X
|α| = αi ;
i=1

x α
= xα1 1 xα2 2 · · · xαNN
for x = (x1 , x2 , · · · , xn ) ∈ RN ;

∂ |α|
Dα = .
∂xα1 1 · · · ∂xαNN

Definition 1.4 If T ∈ D0 (Ω) is a distribution on an open set Ω ⊂ RN , and

if α is any multi-index, we can define the distribution Dα T by

Dα T (ϕ) = (−1)|α| T (Dα ϕ)

for all ϕ ∈ D(Ω). 

Thus, by this trick of transfering the burden of differentiation onto elements

of D(Ω), every distribution becomes infinitely differentiable.

dδ
Example 1.5 The dipole distribution (cf. Example 1.4) is nothing but − dx ,
where δ is the Dirac distribution (cf. Example 1.2). 

Example 1.6 Consider the Heaviside function H : R → R, defined by


1, ifx > 0,
H(x) =
0, ifx ≤ 0.

This is locally integrable and so defines a distribution. If ϕ ∈ D(R), then

Z ∞ Z ∞
dH 0
(ϕ) = − H(x)ϕ (x) dx = − ϕ0 (x) dx = ϕ(0) = δ(ϕ).
dx −∞ 0

Thus, eventhough H is differentiable almost everywhere in the classical sense

with the classical derivative being equal to zero a.e. (which is trivially a
locally integrable function), the distribution derivative is not zero, but the
Dirac distribution. 

This leads us to the following question: if a locally integrable function u

admits a classical derivative (a.e.), denoted u0 , which is also locally integrable,
when can we say that the classical and distributional derivatives of u are the

5
same? The answer lies in the integration by parts formula. If, for every
ϕ ∈ D(Ω) we have Z Z
u ϕ dx = − uϕ0 dx,
0

then the classical derivative is also the distributional derivative. This obvi-
ously happens when u is a smooth function. It is also true for absolutely
continuous functions, cf. Kesavan [1].
In this context, we can also ask the following question. We know that
if u is a differentiable function on R such that u0 ≡ 0, then u is a constant
function. Does the same hold for the distribution derivative? That is, if T is
a distribution so that dTdx
is the zero distribution, then, is T the distribution
generated by a constant function? In other words, does there exist a constant
c ∈ R such that, for every ϕ ∈ D(R), we have
Z
T (ϕ) = c ϕ dx?
R

We claim that this is indeed the case. If dT

dx
= 0, then for every ϕ ∈ D(R), we
0
have T (ϕ ) = 0. Thus we need to know when an arbitrary member of D(R)
can be expressed an the derivative of another member of the same space.
0
Lemma 1.1 Let R ϕ ∈ D(R). Then, there exists ψ ∈ D(R) such that ψ = ϕ,
if, and only if, R ϕ dx = 0.
Proof: If ϕ = ψ 0 with ψ of compact support, then clearly R ϕ dx = 0.
R

Conversely, let supp(ϕ) ⊂ [−a, a] ⊂ R. Define


0, ifx ≤ −a,
ψ(x) = Rx
−a
ϕ(t) dt, ifx > −a.

Then, clearly ψ is infinitely differentiable and, by the given condition, has

support in [−a, a] as well. Its derivative is obviously ϕ. 
R
Now, let ϕ0 ∈ D(R) be chosen such that R ϕ0 (t) dt = 1. Then, given any
ϕ ∈ D(R), define
Z 
ϕ1 = ϕ − ϕ dx ϕ0 ∈ D(R).
R

Then the integral of ϕ1 over R vanishes. So ϕ1 = ψ 0 for some ψ ∈ D(R). If

dT
dx
= 0, then, by definition, we have that T (ψ 0 ) = 0 and so we get
Z 
T (ϕ) = ϕ dx T (ϕ0 )
R

which establishes our claim with c = T (ϕ0 ). Note that whatever the function
ϕ0 we may choose in D(R) whose integral is unity, the value of T (ϕ0 ) is the
same (why?) so that c is well-defined.
We conclude this section by mentioning two very useful collections of
smooth functions.

6
 Consider the following function defined on RN .
(
− 1 2
ρ(x) = e 1−|x| , if|x| < 1,
0, if|x| ≥ 1,

where |x| is the euclidean length of the vector x ∈ RN . It can be shown that
this is a function in D(RN ), with support in B(0; 1), the ball centred at the
origin and of unit radius. Let
Z
k = ρ(x) dx.
RN

Definition 1.5 The family of mollifiers {ρε }ε>0 is the collection of func-
tions defined by
ρε (x) = k −1 ρ(x/ε). 
Then it is easy to see that ρε ∈ D(RN ) with support B(0; ε), the ball centered
at the origin and of radius ε and is such that
Z Z
ρε (x) dx = ρε (x) dx = 1
RN B(0;ε)

for all ε > 0.

∞
Definition 1.6 Let {Ui }m i=1 be open sets whose union is U . Then a C -
partition of unity, subordinate to the collection {Ui }m i=1 is a collection
∞
{ψi }mi=1 of C functions defined on U such that
(i) supp(ψi ) ⊂ Ui , for all 1 ≤ i ≤ m;
(ii) 0P ≤ ψi (x) ≤ 1, for all x ∈ U and for all 1 ≤ i ≤ m;
(iii) m i=1 ψi (x) = 1, for all x ∈ U. 

2 Sobolev Spaces
Let Ω ⊂ RN be an open set and let ∂Ω denote its boundary.
Definition 2.1 Let m be a positive integer and let 1 ≤ p ≤ ∞. The Sobolev
Space W m,p (Ω) is defined by

W m,p (Ω) = {u ∈ Lp (Ω) | Dα u ∈ Lp (Ω), for all |α| ≤ m}. 

The space W m,p (Ω) is a vector space contained in Lp (Ω) and we endow it
with the norm k.km,p,Ω defined as follows.
  p1
X
kukm,p,Ω =  kDα ukpLp (Ω) 
|α|≤m

if 1 ≤ p < ∞ and
kukm,∞,Ω = max kDα ukL∞ (Ω) .
|α|≤m

7
 Notations and conventions
• When p = 2, we will write H m (Ω) instead of W m,2 (Ω). The corre-
sponding norm k.km,2,Ω will be written as k.km.,Ω and it is generated by
the inner-product
X Z
(u, v)m,Ω = Dα uDα v dx.
|α|≤m Ω

• We define a semi-norm om W m,p (Ω) by

  p1
X
|u|m,p,Ω =  kDα ukpLp (Ω) 
|α|=m

when 1 ≤ p < ∞ and

|u|m,∞,Ω = max kDα ukL∞ (Ω) .

|α|=m

• For consistency, we define W 0,p (Ω) to be Lp (Ω). Henceforth the norm

in Lp (Ω) will be denoted as |.|0,p,Ω when p 6= 2 and by |.|0,Ω when p = 2.
It follows from the definition of these spaces that the mapping
 
∂u ∂u
u 7→ u, ,···,
∂x1 ∂xN

is an isometry from W 1,p (Ω) onto a subspace of (Lp (Ω))N +1 .

Theorem 2.1 The space W 1,p (Ω) is complete. It is reflexive if 1 < p < ∞
and separable if 1 ≤ p < ∞. In particular, H 1 (Ω) is a separable Hilbert
space.
Proof: Let {un } be a Cauchy sequence in W 1,p (Ω). Then {un }, { ∂u n
∂xi
}, 1 ≤
p
i ≤ N are all Cauchy sequences in L (Ω). Thus, there exist functions u and
vi , 1 ≤ i ≤ N in Lp (Ω) such that un → u and ∂un
∂xi
→ vi , 1 ≤ i ≤ N in Lp (Ω).
Now, let ϕ ∈ D(Ω). Then
Z Z
∂un ∂ϕ
ϕ dx = − un dx
Ω ∂xi Ω ∂xi
by the definition of the distribution derivative. Since D(Ω) ⊂ Lq (Ω) for any
1 ≤ q ≤ ∞, in particular, it is true when q is the conjugate exponent of p.
Hence, we can pass to the limit in the above relation to get
Z Z
∂ϕ
vi ϕ dx = − u dx
Ω Ω ∂xi

for each 1 ≤ i ≤ N . This shows that

∂u
= vi ∈ Lp (Ω), 1 ≤ i ≤ N
∂xi

8
and so it follows that u ∈ W 1,p (Ω) and that un → u in W 1,p (Ω). This proves
the completeness of the spaces W 1,p (Ω).
The image of W 1,p (Ω) in (Lp (Ω))N +1 via the isometry described above
is thus a closed subspace and so it inherits the reflexivity and separability
properties of (Lp (Ω))N +1 . This completes the proof of the theorem. 

Remark 2.1 Let {un } be a sequence in W 1,p (Ω), 1 < p < ∞. Let un → u
in Lp (Ω) and let { ∂u
∂xi
n
} be bounded in Lp (Ω) for all 1 ≤ i ≤ N . Then, since
Lp (Ω) is reflexive, it follows that there exists a subsequence {unk } such that

∂unk
* vi
∂xi
weakly in Lp (Ω). Thus if ϕ ∈ D(Ω), then we have
Z Z
∂unk ∂ϕ
ϕ dx = − unk dx
Ω ∂xi Ω ∂xi
from which we deduce that
Z Z
∂ϕ
vi ϕ dx = − u dx
Ω Ω ∂xi

∂u
so that it follows that u ∈ W 1,p (Ω) and that ∂xi
= vi . This is a very useful
1
observation. Since L (Ω) is separable, the same idea will also work for the
case p = ∞. 

Remark 2.2 Theorem 2.1 is true for all spaces W m,p (Ω). 

Definition 2.2 The closure of the subspace D(Ω) in W m,p (Ω) is the closed
subspace W0m,p (Ω) .

We will see later that, in general, W0m,p (Ω) is a proper closed subspace
of W m,p (Ω). We will, however, see in the next section that the two spaces
coincide when Ω = RN .

3 The case Ω = RN .
Let p = 2 and consider the case H m (RN ). If u belongs to this space, then it
follows that u, Dα u ∈ L2 (RN ), for all |α| ≤ m. Now, for a square integrable
function on RN , we can define its Fourier transform, which will also be square
integrable on RN , and by the Plancherel theorem, the L2 -norms of both the
function and its Fourier transform will be the same. Further, we also know
that
D
d α u(ξ) = (2πi)|α| ξ α u
b(ξ)
for ξ ∈ RN . Thus it follows that if u ∈ H m (RN ), then u
b(.) and ξ 7→
α 2 N
ξ ub(ξ), |α| ≤ m are all in L (R ), and conversely.

9
 We can see easily that the same powers of ξ occur both in (1 + |ξ|2 )m
and in the sum |α|≤m |ξ α |2 and so we have the existence of two constants
P
M1 > 0 and M2 > 0, which depend only on m and N , such that
X
M1 (1 + |ξ|2 )m ≤ |ξ α |2 ≤ M2 (1 + |ξ|2 )m .
|α|≤m

Consequently, we can write

m
H m (Ω) = {u ∈ L2 (RN ) | (1 + |ξ|2 ) 2 u
b(ξ) ∈ L2 (RN )}.

By the Plancherel theorem, it also follows that the norm defined by the
following relation is equivalent to the norm in H m (RN ) and we will denote
it by the same symbol:
Z
2
kukm,RN = (1 + |ξ|2 )m |b
u(ξ)|2 dξ.
RN

Lemma 3.1 Let {ρε }ε>0 be the family of mollifiers.

(i) If u : RN → R is continuous, then ρε ∗ u → u pointwise, as ε → 0.
(ii) If u : RN → R is continuous with compact support, then ρε ∗ u → u
uniformly, as ε → 0.
(iii) If u ∈ Lp (RN ), then ρε ∗ u → u in Lp (RN ), for 1 ≤ p < ∞.
Proof: (i) Let x ∈ RN . Then, given η > 0, there exists δ > 0 such that for
all |y| < δ, we have |u(x − y) − u(x)| < η. Thus, if ε < δ, we have
Z Z
|ρε ∗ u(x) − u(x)| ≤ |u(x − y) − u(x)|ρε (y) dy < η ρε (y) dy = η.
|y|≤ε |y|≤ε

This proves the first statement.

(ii) If supp(u) = K which is compact, then

supp(ρε ∗ u) ⊂ K + B(0; ε)

which is compact and is contained within a fixed compact set, say, K +B(0; 1)
if we restrict ε to be less than or equal to unity. On this compact set, u is
uniformly continuous and the δ corresponding to η in the previous step is now
independent of the point x and so the pointwise convergence is now uniform.
(iii) From the step (ii) above it is immediate that if u is continuous with
compact support in RN , then ρε ∗ u converges to u in Lp (RN ) as well, since
the entire family is supported in a single compact set and the convergence
there is uniform. Now, we know that continuous functions with compact
support are dense in Lp (RN ) for 1 ≤ p < ∞ and so given u ∈ Lp (RN ), we
can find, for every η > 0, a continuous function g with compact support such
that
η
|u − g|0,p,RN < .
3
Then, for ε sufficiently small, we have
η
|ρε ∗ g − g|0,p,RN < .
3
10
Then

|ρε ∗ u − u|0,p,RN ≤ |u − g|0,p,RN + |g − ρε ∗ g|0,p,RN + |ρε ∗ (g − u)|0,p,RN .

But by Young’s inequality

η
|ρε ∗ (g − u)|0,p,RN ≤ |ρε |0,1,RN |g − u|0,p,RN <
3
since the integral of ρε is unity and the result now follows immediately. 
Theorem 3.1 For 1 ≤ p < ∞, we have

W0m,p (RN ) = W m,p (RN ).

Proof: We will prove it for m = 1.

Step 1. Let {ρε }ε>0 be the family of mollifiers. Then, if u ∈ W 1,p (RN ), it
follows from the preceding lemma that ρε ∗ u → u in Lp (RN ). Also, since
Dα (ρε ∗ u) = ρε ∗ Dα u, for any multi-index α, it follows, again from the
preceding lemma, that ρε ∗ u → u in W 1,p (RN ) as well. Further, notice that
ρε ∗ u ∈ C ∞ (RN ), by the properties of convolutions.

Step 2. Let ζ ∈ D(RN ), be such that 0 ≤ ζ ≤ 1, ζ ≡ 1 on B(0; 1) and

supp(ζ) ⊂ B(0; 2). For every positive integer k, define

ζk (x) = ζ(x/k).

Then ζk ≡ 1 on B(0; k) and supp(ζk ) ⊂ B(0; 2k). Thus ζk ∈ D(RN ). Choose

a sequence εk decreasing to zero. Let uk = ρεk ∗ u. Define

ϕk = ζk uk .

Then ϕk ∈ D(RN ). From the properties of ζ it now follows that ϕk = uk on

B(0; k) and also that |ϕk | ≤ |uk |. Thus,
Z Z
p p p
|uk − ϕk |0,p,RN = |uk − ϕk | dx ≤ 2 |uk |p dx.
|x|>k |x|>k

Now, by the triangle inequality (Minkowski’s inequalty), we have

Z  p1 Z  p1 Z  p1
p p p
|uk | dx ≤ |uk − u| dx + |u| dx .
|x|>k |x|>k |x|>k

The first term on the right hand-side of the above inequlaity can be made as
small as we please for large k since uk → u in Lp (RN ). The second term can
also be made as small as we please for large k, since it represents the tail of
a convergent integral over RN . Thus, ϕk → u in Lp (RN ).

Step 3. Notice that

∂ϕk ∂uk ∂ζk
= ζk + uk , 1 ≤ i ≤ N.
∂xi ∂xi ∂xi

11
Then, since the derivatives of uk converge to those of u in Lp (RN ), we have
∂u
that the first term in the above relation converges to ∂xi
in Lp (RN ), exactly
as in Step 2 above. Further
∂ζk 1 ∂ζ  x 
= .
∂xi k ∂xi k
Since the derivatives of ζ are uniformly bounded, we deduce that
∂ζk
uk → 0
∂xi
in Lp (RN ). Thus, it follows that
∂ϕk ∂u
→
∂xi ∂xi
in Lp (RN ) for all 1 ≤ i ≤ N . Hence it follows that ϕk → u in W 1,p (RN ) and
so u ∈ W01,p (RN ). This completes the proof. 

4 Friedrich’s theorem and applications

In the preceding section we saw that any element of W 1,p (RN ) can be ap-
proximated by functions from D(RN ). We now investigate to what extent we
can approximate functions in W m,p (Ω) by smooth functions when Ω ⊂ RN
is a proper open subset. We begin with a useful technical lemma.
e : RN → R by its extension by zero
Lemma 4.1 Let u : Ω → R. Define u
outside Ω, i.e. 
u(x), ifx ∈ Ω,
u
e(x) =
0, ifx 6∈ Ω.
Then, if u ∈ W 1,p (Ω) and if ψ ∈ D(Ω), we have ψu
f ∈ W 1,p (RN ) and, for all
1 ≤ i ≤ N,  ∼
∂ f ∂u ∂ψ
(ψu) = ψ + u .
∂xi ∂xi ∂xi
Proof: Since the extension by zero maps functions in Lp (Ω) into functions
in Lp (RN ), it is enough to prove the formula for the derivatives of ψu
f given
N
in the statement of the lemma. Let ϕ ∈ D(R ). Then
R ∂ϕ
R ∂ϕ
R  ∂(ψϕ) ∂ψ

RN
ψu ∂xi dx = Ω ψu ∂xi dx = Ω u ∂xi − ϕ ∂xi dx
f
R ∂u R ∂ψ
= − Ω ∂x ψϕ dx − u ϕ dx
Ω ∂xi
R i ∂u ∂ψ
∼
= − RN ψ ∂xi + u ∂x i
ϕ dx

which completes the proof. 

Theorem 4.1 (Friedrichs) Let 1 ≤ p < ∞. Let Ω ⊂ RN be an open set
and let u ∈ W 1,p (Ω). Then, there exists a sequence {un } in D(RN ) such that
un → u in Lp (Ω) and ∂u n
∂xi
→ ∂x∂u
i
in Lp (Ω0 ) for any relatively compact open
set Ω0 of Ω.

12
Proof: Let, as before, u e denote the extension of u by zero outside Ω. Then
if {ρε }ε>0 denotes the family of mollifiers, we know that (cf. Lemma 3.1)
ρε ∗ u
e→u e → u in Lp (Ω). Now, let Ω0 ⊂⊂ Ω. Then
e in Lp (RN ) and so ρε ∗ u
we can find another relatively compact open subset Ω00 such that

Ω0 ⊂⊂ Ω00 ⊂⊂ Ω.

Let ψ ∈ D(Ω) be such that ψ ≡ 1 in Ω00 . Let d = d(∂Ω0 , ∂Ω00 ) > 0. Now,

supp(ρε ∗ ψu
f − ρε ∗ u
e) = supp(ρε ∗ (1 − ψ)e
u)
⊂ B(0; ε) + supp(1 − ψ)

which will be contained in RN \Ω0 if ε < d. Thus

ρε ∗ ψu e in Ω0 .
f = ρε ∗ u

f ∈ W 1,p (RN ) and

Now, ρε ∗ ψu
 ∼
∂(ρε ∗ ψu)
f ∂(ψu)
g ∂ψ ∂u
= ρε ∗ = ρε ∗ u+ ψ
∂xi ∂xi ∂xi ∂xi

which converges, in Lp (RN ) to

 ∼
∂ψ ∂u
u+ ψ .
∂xi ∂xi

In particular, in Lp (Ω0 ), we have

∂(ρε ∗ u
e) ∂(ρε ∗ ψu)
f ∂u
= → .
∂xi ∂xi ∂xi
Thus, for a sequence εk ↓ 0, we have constructed vk = ρεk ∗ u
e such that
vk ∈ C (R ), vk → u in L (Ω) and such that, for all 1 ≤ i ≤ N , ∂v
∞ N p k
∂xi
∂u
→ ∂x i
p 0 0
in L (Ω ), for any relatively compact open subset Ω contained in Ω. Now let
ζk be as in the proof of Theorem 3.1 and set uk = ζk vk which will have the
same convergence properties as the sequence {vk }. 

Thus, given u ∈ W 1,p (Ω), while we can approximate it in Lp (Ω) by

smooth functions, its derivatives can be approximated by the correspond-
ing sequences of derivatives only in Lp of relatively compact subsets. We
then ask ourselves whether it is at all possible to approximate elements in
W 1,p (Ω) by smooth functions.
Definition 4.1 Let Ω ⊂ RN be an open set. A bounded linear operator P :
W 1,p (Ω) → W 1,p (RN ) is said to be an extension operator if the restriction
of P u to Ω is u for every u ∈ W 1,p (Ω). 

Theorem 4.2 If there exists an extension operator on W 1,p (Ω), then given
u ∈ W 1,p (Ω), there exists a sequence {un } in D(RN ) such that un → u in
W 1,p (Ω).

13
Proof: There exists a sequence {un } in D(RN ) converging to P u in W 1,p (Ω),
as proved in Theorem 3.1. Since P u = u in Ω, the result follows immediately.


Corollary 4.1 If there exists an extension operator on W 1,p (Ω), then C ∞ (Ω)
is dense in W 1,p (Ω). 

Example 4.1 Let

RN
+ = {x = (x1 , · · · , xN ) ∈ R
N
| xN > 0}

be the upper half space. Then it admits an extension operator which is

defined in the following way. Set x = (x0 , xN ) where x0 = (x1 , · · · , xN −1 ) ∈
RN −1 . Define
u(x0 , xN )

if xN > 0,
P u(x) = 0
u(x , −xN ) if xN < 0.
It can be shown that P defines an exiension operator from W 1,p (RN
+ ) to
1,p N
W (R ). 

Definition 4.2 Let Ω ⊂ RN be an open subset such that its boundary ∂Ω

is bounded (and hence compact). Let Q be a unit cube in RN with centre at
the origin and with edges parallel to the coordinate axes. Let Q+ = Q ∩ RN +
and let Q0 = Q ∩ RN −1 . The domain Ω is said to be of class C k , where k is
a non-negative integer, if for every x ∈ ∂Ω, we can find a neighbourhood U
of x in RN and a bijective map T : Q → U such that both T and T −1 are
C k -maps on Q and U respectively and the following relations hold:

T (Q+ ) = U ∩ Ω and T (Q0 ) = U ∩ ∂Ω. 

Example 4.2 If Ω ⊂ RN is an open set such that its boundary is bounded

and is of class C 1 , then there exists an extension operator on W 1,p (Ω). .

We now look at some very useful applications of Friedrich’s theorem.

Theorem 4.3 (Chain Rule) Let 1 ≤ p ≤ ∞. Let Ω ⊂ RN be an open subset.

Let G : R → R be a continuously differentiable map such that G(0) = 0.
Assume, further, that there exists a positive constant M such that |G0 (s)| ≤
M for all s ∈ R. Then, if u ∈ W 1,p (Ω), we have G ◦ u ∈ W 1,p (Ω) and,

∂(G ◦ u) ∂u
= (G0 ◦ u) , 1 ≤ i ≤ N,
∂xi ∂xi
where (G ◦ u)(x) = G(u(x)).

Proof: Since G(0) = 0, by the mean value theorem, it follows that |G(s)| ≤
M s, for all s ∈ R. Thus, it follows that both G ◦ u and (G0 ◦ u) ∂x∂u
i
are in
p
L (Ω). Thus it suffics to prove the formula for the derivative of G ◦ u given
in the statement above to prove the theorem.

14
Case 1: Let 1 ≤ p < ∞. Let {un } be a sequence in D(RN ) as in Friedrich’s
theorem. Then, it is clear that G ◦ un → G ◦ u in Lp (Ω) and also that, for a
subsequence that we will henceforth work with and therefore continue to de-
note as {un }, G0 (un (x)) → G0 (u(x)) almost everywhere. Now, let ϕ ∈ D(Ω).
Choose a relatively compact subset Ω0 contained in Ω such that supp(ϕ) ⊂ Ω0 .
Since un is smooth, we have by the classical Green’s theorem (integration by
parts)
Z Z Z
∂ϕ ∂ϕ ∂un
(G ◦ un ) dx = (G ◦ un ) dx = − (G0 ◦ un ) ϕ dx.
Ω ∂xi Ω0 ∂xi Ω0 ∂xi
Then, with the convergences observed above and those guaranteed by Friedrich’s
theorem, we get, on passing to the limit,
Z Z Z
∂ϕ 0 ∂u ∂u
(G ◦ u) dx = − (G ◦ u) ϕ dx = − (G0 ◦ u) ϕ dx
Ω ∂xi Ω0 ∂xi Ω ∂xi
which proves the result.

Case 2: Let p = ∞. If ϕ ∈ D(Ω), choose Ω0 as before. Then u ∈ W 1,∞ (Ω)

implies that u ∈ W 1,∞ (Ω0 ) and hence in W 1,q (Ω0 ) for any 1 ≤ q < ∞ since Ω0
is compact. Thus, the calculations in the preceding case are still valid and
the result follows. 

Remark 4.1 The condition G(0) = 0 was used only to prove that if u ∈
Lp (Ω),then so does G ◦ u. If Ω were bounded, then constant functions are
in Lp (Ω) for all 1 ≤ p ≤ ∞ and the mean value theorem yields |G(u(x))| ≤
|G(0)| + M |u(x)| which shows that G ◦ u ∈ Lp (Ω) and so the condition
G(0) = 0 is no longer necessary in that case. 

Theorem 4.4 Let Ω ⊂ RN be a bounded open set. Let 1 < p < ∞. If

u ∈ W 1,p (Ω), then |u| ∈ W 1,p (Ω) and

∂|u| ∂u
= sgn(u) , 1≤i≤N
∂xi ∂xi
where 
 +1, if u(x) > 0,
sgn(u)(x) = 0, if u(x) = 0,
−1, if u(x) < 0.

√
Proof: Let ε > 0. Let fε (t) = t2 + ε. Then fε ∈ C 1 (R) and
t
fε0 (t) = √
t2 + ε
which shows that |fε0 (t)| ≤ 1. Since Ω is bounded, the chain rule holds (even
though fε (0) 6= 0). Thus, u ∈ W 1,p (Ω) implies that fε ◦ u ∈ W 1,p (Ω) and

∂(fε ◦ u) u ∂u
= √ , 1 ≤ i ≤ N.
∂xi u + ε ∂xi
2

15
 √
Now, |fε (t) − |t| | ≤ ε. Hence fε ◦ u → |u| in Lp (Ω). Further,
 ∂(fε ◦ u) p  ∂u p
Z   Z  
 ∂xi  dx ≤
    dx
Ω ∂xi
 
Ω

which is bounded. Thus, since 1 < p < ∞, we deduce that (cf. Remark 2.1)
|u| ∈ W 1,p (Ω) and that

∂|u| ∂(fε ◦ u)
= lim
∂xi ε→0 ∂xi
in Lp (Ω) if that limit exists. But

u ∂u ∂u
√ → sgn(u)
u2 + ε ∂xi ∂xi

pointwise and since the p-th powers of the absolute values of all these func-
∂u p
tions are bounded by | ∂xi
| , it follows from the dominated convergence theo-
∂u
rem that the L -norms of these functions converge to the Lp -norm of ∂x
p
i
. The
pointwise convergence and the convergence of the norm implies convergence
in Lp and this completes the proof. 

Theorem 4.5 Let Ω ⊂ RN be a bounded open set and let 1 < p < ∞. Let
u ∈ W 1,p (Ω). Then, for any t ∈ R, and for any 1 ≤ i ≤ N , we have that
∂u
∂xi
= 0 almost everywhere on the set {x ∈ Ω | u(x) = t}.

Proof: Let u ≥ 0. Then u = |u|. Thus,

∂u ∂|u| ∂u
sgn(u) = = .
∂xi ∂xi ∂xi
∂u
It then follows that on the set {x ∈ Ω | u(x) = 0}, we have that ∂xi
= 0
almost everywhere. If u ∈ W 1,p (Ω), then u = u+ − u− , where
1 1
u+ = (|u| + u), u− = (|u| − u).
2 2
Then

{x ∈ Ω | u(x) = 0} = {x ∈ Ω | u+ (x) = 0} ∩ {x ∈ Ω | u− (x) = 0}.

+ −
∂u
Then ∂x i
= ∂(u
∂xi
)
− ∂(u
∂xi
)
= 0 almost everywhere on this set. For any t ∈ R,
consider the function u − t. 

Remark 4.2 Notice that the set {x ∈ Ω | u(x) = t} may itself be of measure
zero, in which case the above result gives no new information. However, if
the function takes a constant value on a set of positive measure, then its
derivative vanishes almost everywhere in that set. 

16
Theorem 4.6 (Stampacchia) Let Ω ⊂ RN be a bounded open set and let
1 < p < ∞. Let f : R → R be a Lipschitz continuous function. Then, if
u ∈ W 1,p (Ω), we have f ◦ u ∈ W 1,p (Ω) and if f 0 is continuous except at a
finite number of points {t1 , · · · , tk }, then
 0 ∂u
∂(f ◦ u) def (f ◦ u) ∂x (x), if u(x) 6∈ {t1 , · · · , tk },
(x) = vi = i
∂xi 0, otherwise.
Proof: First of all, it follows from the Lipschitz continuity of f and the
boundedness of Ω that f ◦ u ∈ Lp (Ω). If M is the Lipschitz constant of f ,
then we also have that |f 0 (t)| ≤ M . Then it also follows that vi ∈ Lp (Ω) for
1 ≤ i ≤ N.
Let {ρε }ε>0 denote the family of mollifiers in R. Choose a sequence εn ↓ 0
and set fn = ρεn ∗ f . Then fn ∈ C ∞ (R) and (cf. Lemma 3.1) fn (t) → f (t)
for all t ∈ R. Now,
R 
|fn (t) − fn (t0 )| =  |s|<εn (f (t − s) − f (t0 − s))ρεn (s) ds
 

≤ M |t − t0 | |s|<εn ρεn (s) ds = M |t − t0 |.

R

Thus, it follows that |fn0 (t)| ≤ M for all t ∈ R. We also have that fn ◦u → f ◦u
in Lp (Ω) since for any x ∈ Ω,
Z 
 
|fn (u(x)) − f (u(x))| =  (f (u(x) − s) − f (u(x)))ρεn (s) ds ≤ M εn
|s|<εn

which implies that fn ◦ u → f ◦ u uniformly on Ω and, as Ω is bounded, in

Lp (Ω) as well. Now, fn ◦ u ∈ W 1,p (Ω) and
∂(fn ◦ u) ∂u
= (fn0 ◦ u)
∂xi ∂xi
which is clearly bounded in Lp (Ω). Thus, by Remark 2.1, we do have that
f ◦ u ∈ W 1,p (Ω) and that
∂(f ◦ u) ∂(fn ◦ u)
= lim
∂xi n→∞ ∂xi
in Lp (Ω), if that limit exists.
Now, assume that the derivative of f exists and is continuous except at
a finite number of points {ti }ki=1 . Let Ei = {x ∈ Ω | u(x) = ti }, 1 ≤ i ≤ k.
Set E = ∪ki=1 Ei . Then
∂u ∂u
(fn0 ◦ u) → (f 0 ◦ u)
∂xi ∂xi
∂u
on Ω\E. On E, we have that ∂x i
= 0 almost everywhere by the preceding
0 ∂u
theorem. Thus (fn ◦ u) ∂xi = 0 almost everywhere on E. It follows that
∂fn ◦u
∂xi
→ vi pointwise almost everywhere in Ω. Also
   
(fn ◦ u) ∂u − vi  ≤ 2M  ∂u 
 0   
 ∂xi   ∂xi 

17
and so by the dominated convergence theorem, we deduce that the limit in
Lp (Ω) of ∂((f∂xni◦u) is indeed vi for all 1 ≤ i ≤ N , which completes the proof.

Proposition 4.1 Let u ∈ W 1,p (Ω), 1 ≤ p < ∞. Let K ⊂ Ω be compact. If
u vanishes on Ω\K, then u ∈ W01,p (Ω).
Proof: Choose relatively compact sets Ω00 and Ω0 such that

K ⊂ Ω00 ⊂⊂ Ω0 ⊂⊂ Ω.

Let ψ ∈ D(Ω) such that ψ ≡ 1 on Ω00 and such that supp(ψ) ⊂ Ω0 . Then
ψu = u. Let {un } be a sequence in D(RN ) such that un → u in Lp (Ω) and
∂un ∂u
∂xi
→ ∂x i
in Lp of all relatively compact open subsets of Ω.Then ψun ∈ D(Ω)
and ψun → ψu = u in W 1,p (Ω) since all the supports are contained in
Ω0 ⊂⊂ Ω. This shows that u ∈ W01,p (Ω). 
Proposition 4.2 Let Ω be bounded and let 1 < p < ∞. If f : R → R is a
Lipschitz continuous function such that f (0) = 0 and whose derivative exists
and is continuous except at a finite number of points, then, if u ∈ W01,p (Ω),we
also have f ◦ u ∈ W01,p (Ω).
Proof: Let {un } be a sequence in D(Ω) converging in W 1,p (Ω) to u. Then,
since f is Lipschitz continuous, we have

|f (un (x)) − f (u(x))| ≤ M |un (x) − u(x)|

and it is now immediate that f ◦ un → f ◦ u in Lp (Ω). Further, it is clear

from the formula for the derivatives that f 0 ◦ un is bounded in Lp (Ω). Thus
it follows that, for a subsequence,f ◦ unk * f ◦ u weakly in W 1,p (Ω) (cf.
Remark 2.1). But, since f (0) = 0, it follows that f ◦ un vanishes outside
the support of un , which is compact. Thus, by the preceding proposition,
f ◦ un ∈ W01,p (Ω) and so it follows that f ◦ u ∈ W01,p (Ω) as well (since a closed
subspace is also weakly closed). 
In particular, if u ∈ W01,p (Ω), it follows that the functions |u| ∈ W01,p (Ω)
and so we also have that u+ and u− are in W01,p (Ω). This fact is very useful
when deriving maximum principles for second order elliptic partial differential
equations. Stampacchia’s theorem is not valid in spaces W m,p (Ω) when m > 1
and this will, in some sense, explain why we do not have good maximum
principles for higher order elliptic problems.

5 Poincaré’s inequality
In the previous section, we saw that the existence of an extension operator for
W 1,p (Ω) depended on the nature of the domain. In general, the extension by
zero will not map W 1,p (Ω) into W 1,p (RN ). For example, let Ω = (0, 1) ⊂ R
and consider the function u ≡ 1 in Ω. Then u e ∈ Lp (R) for any 1 ≤ p ≤ ∞.
But
de
u
= δ0 − δ1
dx
18
where δ0 (ϕ) = ϕ(0) and δ1 (ϕ) = ϕ(1) for ϕ ∈ D(R), and this cannot come
from any locally integrable function, as already observed earlier.
We will now show that, irrespective of the nature of the domain, the
extension by zero provides an extension operator from W01,p (Ω) to W 1,p (RN )
for all 1 ≤ p < ∞.
Theorem 5.1 Let 1 ≤ p < ∞. Let u e denote the extension by zero outside Ω
for any function u defined on Ω ⊂ RN . Then, if u ∈ W01,p (Ω), we have that
e ∈ W 1,p (RN ) and that
u

∂e
u ∂u
g
= , 1 ≤ i ≤ N.
∂xi ∂xi
Proof: The extension by zero maps functions in Lp (Ω) into the space Lp (RN ).
Thus it suffices to prove the above formula for the derivatives of u
e.
Since u ∈ W01,p (Ω), let {un } be a sequence in D(Ω) converging to u in
W 1,p (Ω). Let ϕ ∈ D(RN ). Then, by classical integration by parts for smooth
functions, we have
Z Z
∂ϕ ∂un
un dx = − ϕ dx.
Ω ∂xi Ω ∂xi

There is no boundary term since un ∈ D(Ω). Passing to the limit, we see

that Z Z
∂ϕ ∂u
u dx = − ϕ dx.
Ω ∂xi Ω ∂xi
In other words,
Z Z g
∂ϕ ∂u
u
e dx = − ϕ dx
RN ∂xi RN ∂xi
which completes the proof. 

Theorem 5.2 (Poincaré’s inequality) Let Ω ⊂ RN be a bounded open subset.

Let 1 ≤ p < ∞. Then there exists a constant C = C(p, Ω) > 0 such that

|u|0,p,Ω ≤ C|u|1,p,Ω

for all u ∈ W01,p (Ω). In particular, the mapping u 7→ |u|1,p,Ω is a norm on

W01,p (Ω) which is equivalent to the norm kuk1,p,Ω . On H01 (Ω), the bilinear
form
Z X N Z
∂u ∂v
< u, v >1,Ω = dx = ∇u.∇v dx
Ω i=1 ∂xi ∂xi Ω

defines an inner-product yielding the norm |.|1,Ω which is equivalent to the

norm k.k1,Ω .
Proof: Let Ω = (−a, a)N , a > 0. Let u ∈ D(Ω). Assume that 1 < p < ∞.
Then, Z xN
∂u 0
u(x) = (x , t) dt
−a ∂xN

19
where x = (x0 , xN ) and x0 = (x1 , · · · , xN −1 ). Then, by Hölder’s inequality,
 p1
 ∂u 0 p
 
1
|u(x)| ≤  (x , t) dt |xN + a| p0
∂xN
where p0 = p/(p − 1) is the conjugate exponent. Thus,
 ∂u 0 p
Z a 
p
p
|u(x)| ≤ (2a) p0  (x , t) dt
−a ∂xN


which yields, on integrating with respect to x0 ,

 ∂u p
Z Z  
p
0 p 0
|u(x , xN )| dx ≤ (2a) p0
 ∂xN  dx
 
Ω

which in turn yields, on integrating with respect to xN ,

 ∂u p
Z Z  
p
p 0 +1
|u| dx ≤ (2a) p
 ∂xN  dx.
 
Ω Ω

It is easy to deduce this inequality when p = 1 as well. Thus for 1 ≤ p < ∞,

we deduce from this that
|u|0,p,Ω ≤ 2a|u|1,p,Ω
for all u ∈ D(Ω). But since D(Ω) is dense in W01,p (Ω), the above inequality
also holds for all u ∈ W01,p (Ω).
Now if Ω is an arbitrary bounded domain, then we can find a > 0 such
that Ω ⊂ (−a, a)N = Ω. e Let u e denote the extension of u by zero outside Ω.
1,p
Then, if u ∈ W0 (Ω), it follows that u e ∈ W01,p (Ω).
e Thus

|u|0,p,Ω = |e
u|0,p,Ωe ≤ 2a|e
u|1,p,Ωe = 2a|u|1,p,Ω
∂u ∂u
since e
∂xi
= ∂x
f
i
. This completes the proof. 

We now make several important remarks.

• This inequality will be very crucial in the study of Dirichlet boundary
value problems for elliptic partial differential equations.
• It is easy to see that this proof works when Ω is unbounded, but is
bounded in some particular direction, i.e. if Ω is contained in some
infinite strip of finite width. However, it is not true for truly unbounded
regions. In particular, if Ω = RN , consider ζk as in Step 2 of the proof
of Theorem 3.1. Then
∂ζk 1 ∂ζ  x 
= ( .
∂xi k ∂xi k
If p > N , it is then easy to see that |ζk |1,p,RN → 0 as k → ∞ while
1
|ζk |0,p,RN ≥ |B(0; k)| p → ∞
(where |B(0; k)| denotes the Lebesgue measure of the ball B(0; k)) since
ζk ≡ 1 on B(0; k). Thus we cannot have Poincaré’s inequality in such
domains.

20
 • This inequality also shows that if Ω is bounded, then W 1,p (Ω) and
W01,p (Ω) cannot be equal (which was the case for Ω = RN ). For the
constant function u ≡ 1 on Ω is in W 1,p (Ω) and |u|1,p,Ω = 0 but |u|0,p,Ω >
0. Hence the constant function cannot belong to W01,p (Ω).

• If u ∈ W02,p (Ω) where Ω is bounded, then u, ∂x ∂u

i
∈ W01,p (Ω), for all
1 ≤ i ≤ N . Thus, it follows that
   
 ∂u   ∂u 
 
 ∂xi  ≤ C  
0,p,Ω ∂x i

1,p,Ω

which is the same as saying

|u|1,p,Ω ≤ C|u|2,p,Ω .

Thus, for some constant C 0 = C 0 (p, Ω) > 0 we get

|u|0,p,Ω ≤ C 0 |u|2,p,Ω .

In general, there exists a constant C > 0, delpending only on m, p and

Ω, such that
|u|0,p,Ω ≤ C|u|m,p,Ω
for all u ∈ W0m,p (Ω) when m ≥ 1 is an integer and 1 ≤ p < ∞.

• We saw that the constant C in Poincaré’s inequality depended in some

way on the diameter of Ω. This is not the best possible constant. If
p = 2, it turns out that the best constant is connected to ‘the prin-
cipal eigenvalue of the Laplace operator with homogeneous Dirichlet
boundary conditions on ∂Ω’. For instance, if Ω = (0, 1), we get that
the constant is unity from the proof of Poincaré’s inequality. But the
best constant can be shown to be π1 . Thus, for all u ∈ H01 (0, 1), we
have
Z 1  21 Z 1  21
2 1 0 2
|u| dx ≤ |u | dx .
0 π 0

Equality is achieved for the function u = sin πx and constant multiples

thereof.

6 Imbedding theorems
We know that if u ∈ W 1,p (Ω), then, u ∈ Lp (Ω). We now ask the question if
the information that the first order partial derivatives of u are also in Lp (Ω)
will give us more information on the function u, vis-à-vis its smoothness or
integrability with respect to other exponents. In particular we would like to
∗
know if W 1,p (RN ) is continuously imbedded in Lp (RN ), where p 6= p∗ . In
other words, we are looking for an inequality of the form

|u|0,p∗ ,RN ≤ C|u|1,p,RN

21
for all u ∈ W 1,p (RN ), where C > 0 is a constant which depends only on p
and N . A simple analysis will show us when this will be possible, if at all,
and what is the value of p∗ that we should expect. Let λ > 0 be a fixed real
number. Then consider the scaling x 7→ λx which maps RN onto itself. Let
u ∈ W 1,p (RN ). Then if we set uλ (x) = u(λx), it follows that uλ ∈ W 1,p (RN )
as well. We have
∂uλ ∂u
(x) = λ (λx).
∂xi ∂xi
Consequently, by the change of variable formula, we get
 ∂uλ p
Z   Z  p Z  p
p
 ∂u  p−N
 ∂u 

 ∂xi (x) 
 dx = λ 
 ∂xi (λx) 
 dx = λ  ∂xi  dx.
 (x) 
RN R N RN

Thus,
N
|uλ |1,p,RN = λ1− p |u|1,p,RN .
Similarly
N
|uλ |0,p∗ ,RN = λ− p∗ |u|0,p∗ ,RN .
hence, for all λ > 0, we must have
|uλ |1,p,RN N N |u|1,p,RN
0 < C ≤ = λ1− p + p∗ .
|uλ |0,p∗ ,RN |u|0,p∗ ,RN
If 1 − Np + pN∗ is strictly positive, then we let λ → 0 to get a contradiction and
if it is strictly negative, we can let λ → ∞ to get a contradiction. Thus, if
at all we can hope for such an inequality, it follows that we must have that
this number is zero, i.e.
1 1 1
∗
= − .
p p N
For this to be possible, it is necessary that p < N as well. In that case notice
that we also have p∗ > p.
We now split our investigation into three cases, viz., p < N, p = N and
p > N.
Theorem 6.1 (Sobolev’s Inequality) Let 1 ≤ p < N and define p∗ as above.
Then, there exists C = C(p, N ) > 0 such that for all u ∈ W 1,p (RN ),
|u|0,p∗ ,RN ≤ C|u|1,p,RN .
In particular, we have the continuous inclusion
∗
W 1,p (RN ) ,→ Lp (RN ) 
.
The proof of this result is rather technical and we refer the reader to
Kesavan [1].
Corollary 6.1 Let 1 ≤ p < N . Then we have the continuous inclusions
W 1,p (RN ) ,→ Lq (RN )
for all q ∈ [p, p∗ ].

22
Proof: Let p < q < p∗ . Then, there exists α ∈ (0, 1) such that
1 α 1−α
= + .
q p p∗
p p∗
αq N (1−α)q
Then |u| ∈ L (R ) and |u|
αq ∈ L (RN ) if u ∈ W 1,p (RN ).Thus,
(1−α)q

since q = αq + (1 − α)q, we get, by Hölder’s inequality,

|u|0,q,RN ≤ |u|α0,p,RN |u|1−α

0,p∗ ,RN
≤ α|u|0,p,RN + (1 − α)|u|0,p∗ ,RN
≤ α|u|0,p,RN + C(1 − α)|u|1,p,RN
≤ C 0 kuk1,p,RN .

(The inequality in the second line above comes from the generalised AM-GM
inequality.) This completes the proof. 

Corollary 6.2 Let Ω ⊂ RN , and let 1 ≤ p < N . Then, there exists C > 0
such that, for all u ∈ W01,p (Ω),

|u|0,p∗ ,Ω ≤ C|u|1,p,Ω
|u|0,q,Ω ≤ Ckuk1,p,Ω , for all p < q < p∗ .

In particular, for all p ≤ q ≤ p∗ , we have the continuous inclusions

W01,p (Ω) ,→ Lq (Ω).

If Ω = RN 1
+ or if Ω has bounded boundary and is of class C , then we also
have the continuous inclusions

W 1,p (Ω) ,→ Lq (Ω)

for all p ≤ q ≤ p∗ .

Proof: The extension by zero imbeds W01,p (Ω) into W 1,p (RN ) and now it is
easy to see the conclusions for the space W01,p (Ω). If Ω = RN + or if Ω has
1
bounded boundary and is of class C , then there exists an extension operator

P : W 1,p (Ω) → W 1,p (RN )

and again the results follow immediately. 

Let us rewrite the Sobolev inequality for W01,p (Ω) , where Ω is a bounded
open subset of RN in the following form: there exists a constant S > 0 such
that, for all u ∈ W01,p (Ω),

|u|1,p,Ω ≥ S|u|0,p∗ ,Ω .

The best possible constant S(p, N, Ω) is then given by

|u|1,p,Ω
S(p, N, Ω) = inf .
u∈W0 (Ω) |u|0,p∗ ,Ω
1,p

u6=0

23
Now if Ω1 ⊂ Ω2 are two bounded open subsets of RN , then the extension
by zero outside Ω1 maps an element of W01,p (Ω1 ) into W01,p (Ω2 ), keeping
the norms unchanged in their value. It follows from this immediately that
S(p, N, Ω1 ) ≥ S(p, N, Ω2 ). Now let B1 = B(0; r1 ) and B2 = B(0; r2 ) be two
concentric balls centered at the origin. Then the scalings
r2 r1
x 7→ x and x 7→ x
r1 r2

map functions in W01,p (B1 ) into functions in W01,p (B2 ) and vice-versa. It then
follows that
S(p, N, B1 ) = S(p, N, B2 ).
Since the Lebesgue measure is translation invariant, this is also true if the
balls are concentric but are centered elsewhere in RN . Now if Ω is any
bounded open set, we can always find two concentric balls one within Ω and
the other containing Ω. Thus it follows that S(p, N, Ω) is independent of
the domain. In fact for all bounded domains Ω, we have that S(p, N, Ω) =
S(p, N ), where
|u|1,p,RN
S(p, N ) = inf .
u∈W 1,p (RN ) |u|0,p∗ ,RN
u6=0

The value of this best constant has been worked out (independently) by
Aubin and Talenti. When p = 2, the only minimizers of this optimization
problem have been shown to be the functions U (x), Uε (x − x0 ), where ε >
0, x0 ∈ RN and
N −2
U (x) = C(1 + |x|2 )− 2
N −2
Uε (x) = Cε (ε + |x|2 )− 2
where C and Cε are positive normalization constants.
In particular, this shows that for any Ω a bounded open subset of RN ,the
minimization problem stated above can never have a solution. For, if there
were one, then the extension by zero outside Ω would be a minimizer for the
problem in RN as well, but it has been shown that the only minimizers in
RN are the functions U and Uε above and they never vanish anywhere. This
is true for all 1 ≤ p < ∞.

Theorem 6.2 Let p = N . Then we have the continuous inclusions

W 1,N (RN ) ,→ Lq (RN )

for all q ∈ [N, ∞). If Ω = RN + or if Ω is an open subset with bounded

boundary and is of class C 1 , then the above result is true with Ω replacing
RN . If Ω is any open set, we have the continuous inclusions

W01,N (Ω) ,→ Lq (Ω)

for all q ∈ [N, ∞). 

24
 Once again, we refer the reader to Kesavan [1] for the proof.
We do not have the inclusion of any of the spaces mentioned in the pre-
ceding theorem in L∞ of the corresponding domain.

Example 6.1 Let Ω = B(0; 12 ) ⊂ R2 . Let

2
u(x) = log log , x 6= 0.
|x|
Then u ∈ H 1 (Ω), i.e. p = 2 = N , but clearly u 6∈ L∞ (Ω). 

We now turn to the case p > N . To motivate the theorem, let us consider
the case N = 1 and p > 1.

Example 6.2 Let 1 < p < ∞. Let I = (0, 1) and consider u ∈ W 1,p (I).
Then u0 ∈ Lp (I) and so is integrable. Thus the function
Z x
u(x) = u0 (t) dt
0

is absolutely continuous and its derivative (both in the classical and distri-
butional sense) is u0 . Thus, we have that (u − u)0 = 0 and it follows that

u = u + c, a.e.

where c is a constant, since u − u is a constant distribution as was shown in

Section 1. Hence it follows that we can consider u as an absolutely continuous
function on I and so extends continuously to I = [0, 1]. (Recall that elements
of Lp are only equivalence classes of functions; by saying that an element of
W 1,p (I) is continuous, we mean that the equivalence class of any u ∈ W 1,p (I)
has a representative that is absolutely continuous.) Thus, we can write
Z x
u(x) = u(0) + u0 (t) dt, x ∈ I.
0

We then see that

1
|u(0)| ≤ |u(x)| + |u0 |0,p,I |x| p0 ≤ |u(x)| + |u0 |0,p,I

where p0 is the conjugate exponent of p, using Hölder’s inequality. Now

applying the triangle inequality in Lp to the functions |u(x)| and the constant
function |u0 |0,p,I , we get that

|u(0)| ≤ |u|0,p,I + |u0 |0,p,I ≤ Ckuk1,p,I

where C only depends on p. Again,

|u(x)| ≤ |u(0)| + |u0 |0,p,I ≤ Ckuk1,p,I

for any x ∈ I. Thus we have established the continuous inclusion

W 1,p (I) ,→ C(I).

25
In addition, we have, for x, y ∈ I,
Z y
u(x) − u(y) = u0 (t) dt
x

from which we deduce that

1
|u(x) − u(y)| ≤ C|u|1,p,I |x − y| p0

and notice that p10 = 1 − p1 . Thus all the functions in W 1,p (I) are not only
absolutely continuous, but are also Hölder continuous with exponent 1− p1 . 

Theorem 6.3 Let p > N . Then we have the continuous inclusion

W 1,p (RN ) ,→ L∞ (RN )

and there exists a constant C = C(N, p) > 0 such that, for all u ∈ W 1,p (RN ),
and for almost all x, y ∈ RN , we have
N
|u(x) − u(y)| ≤ C|u|1,p,RN |x − y|1− p .

The same conclusions hold when RN is replaced by RN + or by Ω with a bounded

boundary and of class C . The analogous result is also true for W01,p (Ω) when
1

Ω ⊂ RN is any open set. In particular, if Ω is a bounded open set, we can

consider its elements as being Hölder continuous with exponent 1 − Np . 

We refer the reader to Kesavan [1] for the proof.

∂u
Now let u ∈ W 2,p (RN ). Then u and ∂x i
, 1 ≤ i ≤ N are all in W 1,p (RN ).
∗
If p < N , then this implies that u ∈ W 1,p (RN ). Now assume that p∗ < N
as well. This will happen if
1 1 1 1
< ∗ = −
N p p N
N ∗∗
or, in other words, when p < 2
. In this case u ∈ Lp (RN ), where

1 1 1 1 2
∗∗
= ∗− = − .
p p N p N
More genrally, we have the following result.

Theorem 6.4 Let m ≥ 1 be an integer. Let 1 ≤ p < ∞.

(i) If p1 − N
m
> 0, then

1 1 m
W m,p (RN ) ,→ Lq (RN ), = − .
q p N
1 m
(ii) If p
− N
= 0, then

W m,p (RN ) ,→ Lq (RN ), q ∈ [p, ∞).

26
 1 m
(iii) If p
− N
< 0, then

W m,p (RN ) ,→ L∞ (RN ).

In the last case, set k to be the integral part and θ to be the fractional part
of m − Np . Then there exists C > 0 such that for all u ∈ W m,p (RN ), we have

|Dα u|0,∞,RN ≤ Ckukm,p,RN , for all |α| ≤ k

and, for almost all x, y ∈ RN , and for all |α| = k, we have

|Dα u(x) − Dα u(y)| ≤ Ckukm,p,RN |x − y|θ .

In particular we have the continuous inclusion

W m,p (RN ) ,→ C k (RN )

for m > Np . The same results are true when RN is replaced by RN + or by Ω

m m,p
of class C with bounded boundary and for the spaces W0 (Ω) for any open
subset Ω of RN . Thus, if Ω is bounded and sufficiently smooth, we have, for
m > Np ,
W m,p (Ω) ,→ C k (Ω). 

Remark 6.1 If m > Np , and if |α| < k, where k is the integral part of
m − Np , it follows that Dα u is Lipschitz continuous, by virtue of the mean
value theorem and the fact that the highest order derivatives are bounded.


7 Compactness Theorems
In the previous section we saw various continuous inclusions of the Sobolev
spaces in the Lebesgue spaces and spaces of smooth functions. We now
investigate the compactness of these inclusions.
For example consider the case p > N and W 1,p (Ω) where Ω ⊂ RN is
bounded. Then we saw (cf. Theorem 6.3) that the functions in W 1,p (Ω)
are in C(Ω) and that they are also Hölder continuous with exponent 1 − p1 .
Thus, it follows that if B is the unit ball in W 1,p (Ω), then the elements of B
are uniformly bounded and that they are equicontinuous as well. Since Ω is
compact, it then follows that B is relatively compact in C(Ω), by the Ascoli-
Arzela theorem. Consequently the inclusion of W 1,p (Ω) in C(Ω) is compact.
This argument works for all the cases covered in Theorem 6.4 as well, when
p > N.
When p ≤ N , we have continuous inclusions into the Lebesgue spaces.
To examine the compactness of these inclusions, we need an analogue of the
Ascoli-Arzela theorem which describes criteria for the relative compactness
of subsets in the Lebesgue spaces. This comes from the theorem of Fréchet
and Kolmogorov. We omit the proofs and refer the reader to Kesavan [1].

27
Theorem 7.1 (Rellich-Kondrasov) Let Ω ⊂ RN be a bounded open set. Let
1 ≤ p < ∞.Then, the following inclusions are compact:
(i) if p < N, W 1,p (Ω) ,→ Lq (Ω), for all 1 ≤ q < p∗ ;
(ii) if p = N , W 1,N (Ω) ,→ Lq (Ω), for all 1 ≤ q < ∞;
(iii) if p > N , W 1,p (Ω) ,→ C(Ω). 
Example 7.1 We do not have compactness if Ω is unbounded. For example,
consider Ω = R. Let I = (0, 1) and let Ij = (j, j + 1) for j ∈ Z. Let f be a
C 1 function supported in I. Define
fj (x) = f (x − j), j ∈ Z.
Then all the fj ’s are in W 1,p (R) for any 1 ≤ p < ∞ and the sequence {fj } is
clearly bounded in that space. However, since they all have disjoint supports,
we have that if i 6= j, then
1
|fi − fj |0,q,R = 2 q |f |0,q,R .
Thus the sequence {fj } cannot have a convergent subsequence in any Lq (R)
for 1 ≤ q < ∞ and so none of the inclusions
W 1,p (R) ,→ Lq (R)
can be compact. 

Example 7.2 When Ω ⊂ RN is bounded and p < N the inclusion

∗
W 1,p (Ω) ,→ Lp (Ω)
is not compact. For example, when p = 2 < N , assume that this inclusion is
compact. Consider the minimization problem: find u ∈ H01 (Ω) such that
J(u) = min J(v)
1 (Ω)
v∈H0
|v|0,2∗ ,Ω =1

where
J(v) = |v|1,Ω .
Denote the infimum by m ≥ 0. Let {vn } be a minimizing sequence. Then, it
is clearly bounded in H01 (Ω) and so it has a weakly convergent subsequence
{vnk } converging (weakly) to, say, v ∈ H01 (Ω). Since the inclusion
∗
H01 (Ω) ,→ L2 (Ω)
is assumed to be compact, it follows that the subsequence converges in norm
∗
in L2 (Ω) and so we have that |v|0,2∗ ,Ω = 1. Now |.|1,Ω is a norm on H01 (Ω)
equivalent to the usual norm (Poincaré’s inequality) and since the norm is
weakly lower semi-continuous, we have that
m ≤ |v|1,Ω ≤ lim inf |vnk |1,Ω = m.
k→∞

Thus v ∈ H01 (Ω) is a minimizer of J and extending it by zero outside Ω we

see that it achieves the best Sobolev constant (when p = 2) in RN . But we
saw in the last section that such minimizers never vanish and so we get a
∗
contradiction. Thus, the inclusion of H01 (Ω) in L2 (Ω) cannot be compact
∗
and so, a fortiori, the inclusion of H 1 (Ω) in L2 (Ω) cannot be compact either.
This is true for all p < N as well. 

28
8 The spaces W s,p(Ω)
In this section, we will try to define the spaces W s,p (Ω) when 1 < p < ∞ and
when s ∈ R is an arbitrary real number. We start with the case of negative
integers.
Proposition 8.1 Let Ω ⊂ RN be an open set and let 1 < p < ∞. Let F
belong to the dual space of W 1,p (Ω) (respectively, W01,p (Ω)). Then, there exist
0
f0 , f1 , · · · , fN ∈ Lp (Ω), where p0 is the conjugate exponent of p, such that,
for all v ∈ W 1,p (Ω) (respectively, v ∈ W01,p (Ω)), we have
Z N Z
X ∂v
F (v) = f0 v dx + fi dx.
Ω i=1 Ω
∂xi

If we define
N  
X  ∂v 
kvk1,p,Ω = |v|0,p,Ω +  
 ∂xi 
i=1 0,p,Ω

(which is equivalent to the norm defined in Section 2), then

kF k = max |fi |0,p0 ,Ω .
0≤i≤N

If Ω is bounded and if F is in the dual of W01,p (Ω), then we may take f0 = 0.

Proof: Let E = (Lp (Ω))N +1 and let T be the natural isometry from W 1,p (Ω)
(respectively, W01,p (Ω)) into E described in Section 2, i.e.
 
∂v ∂v
T (v) = v, ,···, .
∂x1 ∂xN
Let G ⊂ E denote the image of T and let S : G → W 1,p (Ω) (respectively,
W01,p (Ω)) be the inverse mapping of T . Consider the functional
h ∈ G 7→ F (S(h)).
By the Hahn-Banach theorem, we can extend this to a continuous linear
functional Φ on all of E, preserving the norm. Then, by the Riesz rep-
0
resentation theorem, we can find fi ∈ Lp (Ω), 0 ≤ i ≤ N such that, if
v = (v0 , v1 , · · · , vN ) ∈ E, then
N Z
X
Φ(v) = fi vi dx.
i=0 Ω
PN
Further, if we define kvkE = i=0 |vi |0,p,Ω , then
kΦk = max |fi |0,p0 ,Ω
0≤i≤N

and of course, kΦk = kF k by definition. Now, for any v ∈ W 1,p (Ω) (respec-
tively, W01,p (Ω)), we have v = S(T (v)) and so
Z N Z
X ∂v
F (v) = Φ(T (v)) = f0 v dx + fi dx.
Ω i=1 Ω
∂xi

29
If Ω is bounded then we can work (thanks to Poincaré’s inequality) with the
isometry T : W01,p (Ω) → (Lp (Ω))N given by
 
∂v ∂v
T (u) = ,···,
∂x1 ∂xN
and so we may take f0 = 0. This completes the proof. 
Now, if ϕ ∈ D(Ω), then, by definition of the distribution derivative,
Z N
X ∂f
F (ϕ) = f0 ϕ dx − < ,ϕ >
Ω i=1
∂x i

where the bracket < ., . > denotes the action of a distribution on an element
of D(Ω). Since D(Ω) is dense in W01,p (Ω), a continuous linear functional on
W01,p (Ω) is completely defined by its action on D(Ω) and so, in this case, we
can identify F with the distribution
N
X ∂fi
f0 − .
i=1
∂xi

This identification is not possible for functionals on W 1,p (Ω) since D(Ω) is not
dense there. Now, if m > 1 is any positive integer, any first order derivative
of an element in W m,p (Ω) falls in W m−1,p (Ω). For consistency, we would like
0
this to be true for m = 0 as well. Since derivatives of Lp (Ω) functions are in
0
the dual of W01,p (Ω), we therefore define this dual space as W −1,p (Ω).
Definition 8.1 Let 1 < p < ∞ and let m ≥ 1 be an integer. Then, the
0
space W −m,p (Ω) is the dual of the space W0m,p (Ω), where p0 is the conjugate
exponent of p. 
Let 1 < p < ∞ and let s > 0. Then the spaces W s,p (Ω) can be defined in a
variety of ways when s is not an integer. Usually interpolation theory in Lp
spaces is used. We will not go into these aspects here. The space W0s,p (Ω) will
0
be the closure of D(Ω) in W s,p (Ω) and its dual will be the space W −s,p (Ω)
where p0 is the conjugate exponent of p.
We will henceforth concentrate on the case p = 2. We saw that H m (RN )
can be defined via the Fourier transform. We can immediately generalize
this.
Definition 8.2 Let s > 0 be a real number. Then
s
b(ξ) ∈ L2 (RN )}.
H s (RN ) = {u ∈ L2 (RN ) | (1 + |ξ|2 ) 2 u
The norm in this space is given by
Z  21
kuks,RN = (1 + |ξ|2 )s |b
u(ξ)|2 dξ .
RN

If Ω ⊂ RN is a sufficiently smooth open subset, then H s (Ω) is the space of

restriction of functions in H s (RN ) to Ω. If Ω is a proper subset of RN , then
H0s (Ω) is the closure of D(Ω) in H s (Ω). The dual of H0s (Ω) (respectively,
H s (RN )) is the space H −s (Ω) (respectively, H −s (RN )). 

30
Let Ω ⊂ RN be a bounded open set with ‘sufficiently smooth’ boundary ∂Ω
and is such that Ω always lies to the same side of each connected component
of the boundary. Then, at each point of the boundary, there is a neighbour-
hood U and a bijective mapping T : Q → U , where Q is the unit cube in
RN centered at the origin, with the properties given in Definition 4.2. Since
∂Ω is compact, it can be covered by a finite number {Ui }m i=1 of such neigh-
∞
bourhoods. Let {Ti }i=1 be the corresponding maps. Let {ψi }m
m
i=1 be a C
m
partition of unity subordinate to the collection {Ui }i=1 .
Being a smooth (N − 1)-dimensional manifold, ∂Ω can be provided with
the (N − 1)- dimensional surface measure induced on it from RN . Thus, we
can easily define the spaces Lp (∂Ω), for 1 ≤ p ≤ ∞. Now, given u ∈ L2 (Ω),
we can write m
X
u = ψi u.
i=1

The function ψi u is supported inside the open set Ui . Consider the functions

vi (y 0 , 0) = (ψi u)(T (y 0 , 0)), (y 0 , 0) ∈ Q0 , 1 ≤ i ≤ m

where, as usual, y ∈ RN is written as (y 0 , yN ) with y 0 ∈ RN −1 . Since vi is

supported in Q0 , we can extend it to all of RN −1 by setting it to be zero
outside its support. It is easy to see that the maps u 7→ vi for 1 ≤ i ≤ m
all map L2 (∂Ω) into L2 (RN −1 ) and also maps smooth functions to smooth
functions. We now define, for s > 0,

H s (∂Ω) = {u ∈ L2 (∂Ω) | vi ∈ H s (RN −1 ), for all 1 ≤ i ≤ m}.

It can be checked that this definition is independent of the choice of the atlas
{Ui }m
i=1 on ∂Ω. We also define

H −s (∂Ω) = (H s (∂Ω))∗

(where the star denotes the dual space) for s > 0.

9 Trace theory
Sobolev spaces are the ideal functional analytic setting to study boundary
value problems. In that case given u ∈ W 1,p (Ω), where Ω ⊂ RN is a bounded
open set, we would like to assign meanings expressions like ‘u restricted to
∂Ω’ or ‘the outer normal derivative ∂u ∂ν
of u on ∂Ω’ and so on. But when
u ∈ W 1,p (Ω), it is a priori an element of Lp (Ω) and so is only defined almost
everywhere. Since ∂Ω has measure zero in RN , it is therefore not meaningful
to talk of the values u takes on the boundary.
However, since we have additional information on the derivatives of u,
when u is in some Sobolev space, we can indeed give such notions a meaning
consistent with our intuitive understanding of terms such as boundary value
and exterior normal derivative. We will now make this more precise.
While the theory outlined below can be done for all 1 < p < ∞, for
simplicity, we will restrict ourselves to the case p = 2.

31
 N −1
Theorem 9.1 There exists a continuous linear map γ0 : H 1 (RN 2
+ ) → L (R )
1 N N
such that, if v ∈ H (R+ ) is continuous on R+ , then γ0 (v) is the restriction
N −1
of v to ∂RN+ = R .
Proof: Let v ∈ D(RN ). Then
R∞
|v(x0 , 0)|2 = − 0 ∂x∂N (|v(x0 , xN )|2 ) dxN
R ∞ ∂v 0
= −2 0 ∂x N
(x , xN ).v(x0 , xN ) dxN 
R ∞  ∂v 0 2
≤ 0  ∂x (x , x )  + |v(x0 , xN )|2 dxN .

N
N

Integrating both sides with respect to x0 over RN −1 , we get

Z Z  2 !
∂v
|v(x0 , 0)|2 dx0 ≤
  2
 ∂xN  + |v| dx.
 
RN −1 N
R+

In other words, we have

 
v|
RN −1

0,RN −1
≤ kvk1,RN+ .

Since we know that (cf. Example 4.1 and Corollary 4.1) the restrictions of
functions in D(RN ) are dense in H 1 (RN
+ ), the above inequality implies that
the map
v 7→ v|RN −1
N −1
extends uniquely to a continuous linear map γ0 from H 1 (RN 2
+ ) into L (R ).
1 N N N
Now let v ∈ H (R+ ) be continuous on R+ . Extend v to all of R by
reflection on RN −1 (cf. Example 4.1). Choose a sequence εm ↓ 0 and let ρεm
be the corresponding mollifiers. Let ζ ∈ D(RN ) be such that 0 ≤ ζ ≤ 1, ζ ≡
1 on B(0; 1) and such that supp(ζ) ⊂ B(0; 2). Define ζk (x) = ζ(x/k). Then
(cf. Lemma 3.1), ρεm ∗ v(x) → v(x) for all x ∈ RN and also, since ζm ≡ 1 on
B(0; m), we have vm (x) → v(x) for all x ∈ RN , where
vm = ζm .(ρεm ∗ v).
We also know that (cf. Theorem 3.1) vm → v in H 1 (RN ) and hence in
N −1
H 1 (RN 2
+ ) as well. Thus γ0 (vm ) → γ0 (v) in L (R ). But, since vm ∈ D(RN ),
it follows that γ0 (vm ) is the restriction of vm to RN −1 and we saw that this
converges pointwise to the restriction of v to RN −1 . Thus, it follows that
γ0 (v) is the restriction of v to RN −1 . This completes the proof. 

It can be shown (cf. Kesavan [1]) that the range of γ0 is the space
1
H (RN −1 ) and that its kernel is H01 (RN
2
+ ). Thus we can interpret elements of
1 N 1 N
H0 (R+ ) as those of H (R+ ) which ‘vanish on the boundary’.
In the space H 2 (RN+ ), apart from γ0 , we can imitate the proof of the
N −1
preceding theorem to show the existence of a map γ1 : H 2 (RN 2
+ ) → L (R )
2 N 1 N
such that if v ∈ H (R+ ) ∩ C (R+ ), then

∂v 
γ1 (v) = − .
∂xN  N −1
R

32
 1 3
Its range would be H 2 (RN −1 ) while that of γ0 will be H 2 (∂Ω). The kernel
of the map
3 1
N −1
(γ0 , γ1 ) : H 2 (RN
+ ) → H (R
2 ) × H 2 (RN −1 )

is the space H02 (RN m N

+ ). More, generally, we can define maps γj : H (R+ ) →
L2 (RN −1 ) which, for a smooth function v is the restriction of (− ∂x
∂v j
N
) to
1
RN −1 and whose range is H m−j− 2 (RN −1 ) for 1 ≤ j ≤ m − 1. The kernel of
the map
1
(γ0 , γ1 , · · · , γm−1 ) : H m (RN m−1 m−j− 2
+ ) → Πj=0 H (RN −1 )

is H0m (RN
+ ).
Let us now turn to the case of a sufficently smooth bounded open set Ω
of RN . Let {Ui }m m
i=1 together with the associated maps {Ti }i=1 be an atlas
m ∞
for the boundary ∂Ω. Let {ψi }i=1 be an associated C partition of unity
subordinate to the collection {Ui }m 1
i=1 . If u ∈ H (Ω), then (after extension by
zero) we have that for each 1 ≤ i ≤ m, the functions (ψi u|Ui ∩Ω )◦Ti ∈ H 1 (RN
+)
1
N −1
and so we can define its trace γ0 as an element of H 2 (R ). Coming back
−1
by Ti , we can define the trace on Ui ∩ ∂Ω. Piecing these together, we get
the trace γ0 u ∈ L2 (∂Ω) and, by our definition of the spaces on ∂Ω, the range
1
will be precisely H 2 (∂Ω). Similarly, we can define higher order traces, which
generalize the notion of exterior normal derivatives of various orders.
Theorem 9.2 (Trace Theorem) Let Ω ⊂ RN be a bounded open set of class
C m+1 . Then there exist maps γ0 , γ1 , · · · , γm−1 from H m (Ω) into L2 (∂Ω) such
that
(i) if v ∈ H m (Ω) is sufficently smooth, then,

∂ m−1 v 
 
∂v 
γ0 (v) = v|∂Ω , γ1 (v) = , · · · , γm−1 (v) =
∂ν ∂Ω ∂ν m−1 ∂Ω

where ν is the unit outward normal on ∂Ω;

(ii) The range of the map (γ0 , γ1 , · · · , γm−1 ) is
1
Πm−1
j=0 H
m−j− 2
(∂Ω);

(iii) The kernel of the map (γ0 , γ1 , · · · , γm−1 ) is the space H0m (Ω). 

Theorem 9.3 (Green’s theorem) If Ω ⊂ RN is a bounded open set of class

C 1 and if u, v ∈ H 1 (Ω), then, for 1 ≤ i ≤ N ,
Z Z Z
∂v ∂u
u dx = − v dx + γ0 (u)γ0 (v)νi dσ
Ω ∂xi Ω ∂xi ∂Ω

where ν = (ν1 , · · · , νN ) is the unit outward normal on ∂Ω. In particular, if

one of the two functions is in H01 (Ω), we have
Z Z
∂v ∂u
u dx = − v dx.
Ω ∂xi Ω ∂xi

33
Proof: We know that C ∞ (Ω) is dense in H 1 (Ω). If {un } and {vn } are
sequences of smooth functions converging in H 1 (Ω) to u and v respectively,
we have, by the classical Green’s theorem,
Z Z Z
∂vm ∂um
um dx = − vm dx + um vm νi dσ
Ω ∂xi Ω ∂xi ∂Ω
for any 1 ≤ i ≤ N . The result now follows on passing to the limit as n → ∞.
If one of them is in H01 (Ω), then the integrand in the boundary integral van-
ishes. 

We conclude this section by showing that the kernel of the trace map γ0
is indeed H01 (Ω) in the one-dimensional case.
Theorem 9.4 Let 1 < p < ∞. Let I = (0, 1) ⊂ R. Then
W01,p (I) = {u ∈ W 1,p (I) | u(0) = u(1) = 0}.
Proof: Let u ∈ W01,p (I). Then there exists a sequence {un } in D(I) such
that un → u in W 1,p (I). But this implies that (cf. Example 6.2) that un → u
uniformly on I and so it follows that u(0) = u(1) = 0.

Conversely, assume that u ∈ W 1,p (I) (which is therefore continuous on

I) is such that u(0) = u(1) = 0. Then
Z x Z 1
0
u(x) = u (t) dt and u0 (t) dt = 0.
0 0
Since D(I) is denseRin L (I), let {vn } be a sequence in D(I) coverging to u0
p
1
in Lp (I). Set an = 0 vn (t) dt. Since I has finite measure, we also have that
vRn → u0 in L1 (I) and so it follows that an → 0. Now let ψ0 ∈ D(I) such that
1
0
ψ(t) dt = 1. Define
ψn (t) = vn (t) − an ψ0 (t).
Then ψn ∈ D(I) and its integral over I vanishes, by construction. Hence,
there exists un ∈ D(I) such that u0n = ψn (cf. Lemma 1.1).
Now,
Z x Z x
0 0
(un − u)(x) = (un − u )(t) dt = (ψn − u0 )(t) dt.
0 0
Thus
|(un − u)(x)| ≤ |ψn − u0 |0,p,I
which yields
|un − u|0,p,I ≤ |ψn − u0 |0,p,I
≤ |ψn − vn |0,p,I + |vn − u0 |0,p,I
= an |ψ0 |0,p,I + |vn − u0 |0,p,I
→ 0.
Also
|u0n − u0 |o,p,I = |ψn − u0 |0,p,I → 0
as already shown. Thus {un } is a sequence in D(I) which converges to u in
W 1,p (I) and so u ∈ W01,p (I) whih completes the proof. 

34
References
[1] Kesavan, S. Topics in Functional Analysis and Applications,
New Age International (formerly Wiley-Eastern), 1989.

35