Chapter 2

Sobolev spaces

2.1 Weakly differentiable functions

In this first paragraph we introduce the notion of weak derivatives and define the Sobolev
spaces of weakly differentiable functions. This generalizes the notion of differentiability to a
class of functions which are not differentiable in the classical sense.
We will sometimes refer to distributions and the notion of derivatives in the sense of
distributions, which are assumed to be known. However, we will recall all the required
definitions and results to make this chapter self-contained.

2.1.1 One derivative in Lp in dimension one

We begin with the one dimensional case. Let I be an open and non-empty interval of R. The
key observation behind the definition of the weak derivative is the integration by parts. For
u P C 1 pIq and φ P C08 pIq we have
ż ż
u1 φ dx “ ´ uφ1 dx. (2.1)
I I

The right-hand side makes sense even when u is not differentiable. This is how we define the
function u1 which appears in the left-hand side.

Definition 2.1. Let u P L1loc pIq. We say that u has a weak derivative in L1loc pIq if there
exists v P L1loc pIq such that
ż ż
@φ P C08 pIq, ´ uφ1 dx “ vφ dx. (2.2)
I I

In this case we denote by u1 this function v.

Definition 2.2. We denote by W 1,p pIq the set of functions u P Lp pIq with a weak derivative
u1 P Lp pIq. We also write H 1 pIq for W 1,2 pIq.

Of course, if u is differentiable in the usual sense, then the derivatives in the usual and
in the weak senses coincide. However u and u1 are not necessarily in Lp pIq, so u is not nec-
essarily in W 1,p pIq (if I is a compact interval, then continuous functions are integrable and
hence, in this particular case, continuously differentiable functions on I belong to W 1,p pIq).

The weak derivative is just the derivative in the sense of distributions. A function u P
L1loc pIq defines a distribution Tu on I. This distribution has a derivative Tu1 P D1 pIq. Saying
that the derivative of u belongs to L1loc pIq means that Tu1 is the distribution defined by a
function in L1loc pIq. In other words, for some v P L1loc pIq we have Tu1 “ Tv in D1 pIq. Then
a function u P Lp pIq belongs to W 1,p pIq if and only if u1 P Lp pIq, where u1 is understood in
the sense of distributions.
Notice also that a function v satisfying (2.2) is necessarily unique (up to equality almost
everywhere), so there is no ambiguity in the definition of u1 .

3
M2RI - Elliptic PDE and evolution problems

Example 2.3. We consider on s0, 1r the function u : x ÞÑ x´ 4 . Then u belongs to L2 ps0, 1rq
1

but its derivative u1 : x ÞÑ ´ 14 x´ 4 is not in L2 ps0, 1rq, so u is not in H 1 ps0, 1rq. We similarly
5

consider u : x ÞÑ x´ 4 on s1, `8r. Then u1 P L2 ps1, `8rq but u R L2 ps1, `8rq, so u R

1

H ps1, `8rq.
1

On the other hand, the function u ÞÑ x 4 belongs to H 1 ps0, 1rq and x ÞÑ x´ 4 belongs to
3 3

H ps1, `8rq.
1

Exercise 1. Let p P r1, `8s and α P R. Does the function x ÞÑ xα belongs to W 1,p ps0, 1rq
? W 1,p ps1, `8rq ? W 1,p ps0, `8rq ?
These first examples concern functions differentiable in the usual sense. But W 1,p pIq
contains functions which are only differentiable in the weak sense.
Example 2.4. We consider on s ´ 1, 1r the map u : x ÞÑ |x|. Then its derivative in the sense
of distributions is given by #
1 ´1 if x ă 0,
u : x ÞÑ
1 if x ą 0.
Then u belongs to W 1,p ps ´ 1, 1rq for any p P r1, `8s.
A function in Lp pIq can have a derivative in the sense of distributions which is not a
function.
Example 2.5. The Heaviside function
#
1 if x ą 0,
H : x ÞÑ
0 if x ă 0,

belongs to Lp ps ´ 1, 1rq. However, there is no function v P Lp ps ´ 1, 1rq such that

ż1 ż1
8
@φ P C0 ps ´ 1, 1rq, vφ dx “ ´ Hφ1 dx “ φp0q,
´1 ´1

so H has no weak derivative in Lp ps ´ 1, 1rq, and hence it does not belong to W 1,p ps ´ 1, 1rq.
With the vocabulary of distributions the derivative of H in the sense of distributions is the
Dirac distribution H 1 “ δ, and δ is not associated to any function in Lp ps ´ 1, 1rq. Similarly, a
piecewise C 1 function on I which is not continuous cannot be in W 1,p pIq (with the vocabulary
of distributions, by the jump formula the derivative in the sense of distributions of such a
function involves Dirac distributions and cannot belong to Lp pIq).
More generally, we can prove that a function in W 1,p pIq is necessarily continuous on I.
We recall that a function in Lp pIq or in W 1,p pIq is in fact a equivalence class of functions
which are pairwise almost everywhere equal. When we say that u P W 1,p pIq is continuous,
this means that one of the representatives of u is continuous.
Proposition 2.6. Let p P r1, `8s and u P W 1,p pIq. Then u has a representative ũ P Lp pIq
such that, for x, y P I, ż y
ũpyq ´ ũpxq “ u1 psq ds.
x

In particular ũ is continuous. If p ą 1 then ũ is even p´1p -Hölder continuous on I (when

p “ `8 this means that ũ is Lipschitz continuous). Moreover, if I is not bounded and if
p Ps1, `8r then ũ goes to 0 at infinity.
Finally, for all p P r1, `8s, ũ is bounded and hence u P L8 pIq.
For the proof we recall the following results (prove them as an exercice if not already
known).
Lemma 2.7. Let u P L1loc pIq be such that
ż
@φ P C08 pIq, uφ1 dx “ 0.
I

There exists a constant α such that u “ α almost everywhere.

4 J. Royer - Université Toulouse 3

 Sobolev spaces

Lemma 2.8. Let w P L1loc pIq and x0 P I. Then the map

żx
v ÞÑ wpsq ds
x0

is well defined, it is continuous on I, and

ż ż
@φ P C08 pIq, vφ1 dx “ ´ wφ dx.
I I

Now we can prove Proposition 2.6:

Proof. We fix x0 P I. For x P I we set
żx
vpxq “ u1 psq ds.
x0

This makes sense since u1 P Lp pIq Ă L1loc pIq. Then, by Lemma 2.8, v is continuous and its
derivative in the sense of distributions is u1 . By Lemma 2.7, there exists a constant α such
that u ´ v “ α almost everywhere. We set ũ “ v ` α.
For x, y P I we have
ży
ũpyq ´ ũpxq “ vpyq ´ vpxq “ u1 psq ds.
x

If p “ 1 then for some x0 P I we have |ũpyq| ď |ũpx0 q| ` }u1 }L1 pIq .

If p “ `8 then |ũpyq ´ ũpxq| ď |y ´ x| }u1 }L8 pIq , so ũ is }u1 }L8 pIq -Lipschitz continuous.
If p Ps1, `8r, we have by the Hölder inequality
ˇż y ˇ ˆż ˙ p1
ˇ ˇ ˇ ˇ p´1 ˇ 1 ˇp
|ũpyq ´ ũpxq| ď ˇˇ ˇu1 psqˇ dsˇˇ ď |y ´ x| p ˇu psqˇ ds .
x I

p´1
This proves that ũ is p -Hölder continuous, and in particular uniformly continuous. All the
statements of the proposition follow.
Exercise 2. Let α P R. For x P R we set
#
xα e´x if x ě 0,
uα pxq “
´ |x| e´|x|
α
if x ă 0.

1. Prove that uα P C 1 pRq if α ą 1.

2. Prove that uα P H 1 pRq if α ą 12 .
Exercise 3. Let u P Lp pIq. Prove that u P W 1,p pIq if and only if there exists v P Lp pIq
such that ż ż
1 1
@φ P C0 pIq, uφ “ ´ vφ.
I I

Exercise 4. 1. Let u` P C01 pr0, `8rq.For x P R we set

#
u` pxq if x ě 0,
upxq “
0 if x ă 0.

Does u belong to H 1 pRq ? to C 1 pRq ?

2. Same questions with #
u` pxq if x ě 0,
upxq “
u` p´xq if x ă 0,
3. Same questions with
#
u` pxq if x ě 0,
upxq “
´3u` p´xq ` 4u` p´x{2q if x ă 0.

2020-2021 5
M2RI - Elliptic PDE and evolution problems

2.1.2 General definitions

The above definitions can be extended in any dimension d P N˚ and we can consider any
order k P N of derivatives.
Definition 2.9. Let Ω be an open subset of Rd . For p P r1, `8s and k P N we set
� (
W k,p pΩq “ u P Lp pΩq : B α u P Lp pΩq for all α P Nd with |α| ď k ,

where B α u is the derivative of u in the sense of distributions. In other words, a function

u P Lp pΩq belongs to W k,p pΩq if for all α P Nd such that |α| ď k there exists vα P Lp pΩq such
that ż ż
@φ P C08 pΩq, u B α φ dx “ p´1q|α| vα φ dx. (2.3)
Ω Ω

In this case vα is unique (up to equality almost everywhere) and we set B α u “ vα . We also
set H k pΩq “ W k,2 pΩq.
Remark 2.10. By the Riesz Theorem and by density of C08 pΩq in L2 pΩq, a function u P
L1loc pΩq belongs to H k pΩq is and only if for all α P Nd with |α| ď k there exists Cα ą 0 such
that ˇż ˇ
8
ˇ ˇ
@φ P C0 pΩq, ˇ uB φ dxˇˇ ď Cα }φ}L2 pΩq .
ˇ α
Ω
´α
Example 2.11. Let α ą 0. For x P Bp0, 1qz t0u we set upxq “ |x| . Then u P Lp pBp1qq if and
´α´2
only if αp ă d. On the other hand u is of class C 1 on Bp1qz t0u and ∇upxq “ ´α |x| x
for all x P Bp0, 1qz t0u. Thus ∇u P Lp pBp1qq if and only if pα ` 1qp ă d. This proves that if
α ě dp ´ 1 then u is not in W 1,p pBp0, 1qq. Now assume that α ă dp ´ 1.
Let φ P C08 pBp0, 1qq. Since u P L1 pBp0, 1qq we have by the dominated convergence
theorem ż ż
´α ´α
´ |x| ∇φ dx “ ´ lim |x| ∇φ dx.
Bp0,1q εÑ0 Bp0,1qzBp0,εq

For ε Ps0, 1r we have by the Green formula

ż ż ż
´α ´α ´α´2
´ |x| ∇φ dx “ ´ |x| φν dσpxq ´ α |x| xφ dx,
Bp0,1qzBp0,εq Sp0,εq Bp0,1qzBp0,εq

where Sp0, εq is the sphere of radius ε. On the one hand we have

ˇż ˇ
ˇ ˇ
ˇ ´α ˇ
ˇ |x| φν dσpxqˇ ď }φ}8 |Sp0, 1q| εd´1´α ÝÝÝÑ 0,
ˇ Sp0,εq ˇ εÑ0

and on the other hand,

ż ż
´α´2 ´α´2
|x| xφ dx ÝÝÝÑ |x| xφ dx.
Bp0,1qzBp0,εq εÑ0 Bp0,1q

´α´2
This proves that the map x ÞÑ ´α |x| x is the gradient of u on Bp0, 1q in the weak sense.
And hence ∇u P Lp pBp0, 1qq. Finally we have proved that u P W 1,p pBp0, 1qq if and only if
pα ` 1qp ă d.
Exercise 5. Let d ě 2. We denote by B the unit ball in Rd . Let u P C 1 pBz t0uq such that
∇u (well defined on Bz t0u) is in L1loc pBq.
1. For ε Ps0, 1s we denote by Bpεq the ball of radius ε. Prove that
ż
εd´1
d´1
∇upxq dx ÝÝÝÑ 0.
BzBpεq |x| εÑ0

2. For ε Ps0, 1s we denote by Spεq the sphere of radius ε. We set S “ Sp1q. Prove that
ż ż ż
d´1 εd´1
|u| ď ε |u| ` d´1
|∇u| dx.
Spεq S BzBpεq |x|

6 J. Royer - Université Toulouse 3

 Sobolev spaces

3. Prove that u P L1loc pBq.

4. Prove that for j P �1, d� and φ P C08 pBq we have
ż ż
@φ P C08 pBq, ´ u Bj φ dx “ Bj u φ dx
B B

Exercise 6. Does the map x ÞÑ lnp|lnp|x|q|q belong to W 1,d pBp0, 1qq (we recall that Bp0, 1q
is the unit ball of Rd ) ?
Example 2.11 and Exercise 6 show that the results of Proposition 2.6 are only valid in
dimension 1. In higher dimensions, a function in W 1,p pΩq is not necessarily continuous.

In the following proposition we give some basic properties for the set W k,p pΩq (the proof
is left as an exercice). We define C08 pΩq as the restrictions to Ω of functions in C08 pRd q.
Proposition 2.12. Let p P r1, `8s, k P N˚ and α “ pα1 , . . . , αd q P Nd with |α| ď k. Let
u P W k,p pΩq.
(i) We have B α u P W k´|α|,p pΩq and for β P Nd with |β| ď k´|α| we have B β pB α uq “ B α`β u.
(ii) Let ω be an open subset of Ω. Then the restriction u |ω of u on ω belongs to W k,p pωq
and B α puω q “ pB α uq|ω .

(iii) Let χ P C08 pΩq. Then χu P W k,p pΩq and

ÿ ˆα˙
α
B pχuq “ B β χB α´β u,
β
βďα

where we have set ˆ ˙

α α!
“ , α! “ α1 ! . . . αd !.
β β!pα ´ βq!

When Ω “ Rd and p “ 2 we can use the Fourier transform to give a simple characterisation
of H k pRd q. Notice that in Definition 2.9 we can see the derivatives of u in the sense of
tempered distributions. This means that we can replace C08 pRd q by SpRd q in (2.3).
Proposition 2.13. Let k P N˚ and u P L2 pRd q. Then B α u P L2 pΩq if and only if the map
ξ ÞÑ piξqα ûpξq belongs to L2 pRd q (and, in this case, the latter is the Fourier transform of the
former). Then u P H k pRd q if and only if
ż
` 2 ˘k
1 ` |ξ| |ûpξq|2 dξ ă `8. (2.4)
Rd

Proof. Let u P L2 pRd q. For φ P SpRd q we have

ż ż ż
piyqα ûφ dy “ {
upiyqα φ dy “ p´1q|α| uB α φ̂ dy. (2.5)
Rd Rd Rd

Assume that B α u P L2 pRd q for all α P Nd with |α| ď k. Then for such an α (2.5) gives
ż ż ż
piyqα ûφ dy “ B α uφ̂ dy “ By
α uφ dy,
Rd Rd Rd

so the map y ÞÑ piyqα ûpyq belongs to L2 pRd q, and it is the Fourier transform of B α u.
Conversely, let u P L2 pRd q be such that (2.4) holds. For α P Nd with |α| ď k (2.5) applied
with φ̌ gives ˇż ˇ ˇż ˇ
ˇ ˇ ˇ ˇ }y α û}L2 pRd q
ˇ uB α
φ dy ˇ “ ˇ y α
û φ̌ dy ˇď }φ}L2 pRd q .
ˇ d ˇ ˇ d ˇ d
R R p2πq 2
This proves that B α u P L2 pRd q and the proof is complete.
Remark 2.14. If u P L2 pRd q is such that Δu belongs to L2 pRd q, then u belongs to H 2 pRd q.
This remark does not hold on a general domain (see Remark 3.17 below).

2020-2021 7
M2RI - Elliptic PDE and evolution problems

In (2.4), k is the number of derivatives in L2 pRd q. In particular it is an integer. But it

makes sense to write the same condition with any real exponent. This is a way to define
derivatives of real order, which will turn out to be useful. By Proposition 2.13, the following
definition coincides with the previous one when s P N.
Definition 2.15. Let s ě 0. We define H s pRd q as the set of functions u P L2 pRd q such that
ż
` 2 ˘s
1 ` |ξ| |ûpξq|2 dξ ă `8.
Rd

Exercise 7. Let p P r1, `8s and u P W 1,p pRd q. Let ρ P C08 pRd q. We recall that pρ ˚ uq P
C 8 pRd q. Prove that for j P �1, d� we have
Bj pρ ˚ uq “ ρ ˚ pBj uq.
Deduce that pρ ˚ uq P W 1,p pRd q.
Exercise ş8. Let u P W 1,8 pRd q. Let B be a compact subset of Rd . Let ρ P C08 pRd , R` q
such that Rd ρ dx “ 1. For ε ą 0 and x P Rd we set ρε pxq “ ρpx{εq.
1. Prove that there exists a sequence pεn qnPN going to 0 such that if we set un “ ρεn ˚ u for
all n P N then un pxq goes to upxq for almost all x P B.
2. Prove that for all n P N we have }∇un }L8 pRd q ď }∇u}L8 pRd q .
3. Prove that for almost all x, y P B we have |upxq ´ upyq| ď }∇u}L8 pRd q |x ´ y|.
4. Prove that u has a representative which is }∇u}L8 pRd q -Lipschitz (and in particular contin-
uous).

2.2 Topology on the Sobolev spaces

In this section we define the norms on the Sobolev spaces we have just defined, and we give
the properties of these new functional spaces. In the particular case of Sobolev spaces on
the Euclidean space, we prove that smooth functions are dense in the Sobolev space, and
we show on some examples how this important result is used to generalize some properties
known for regular functions. The density of smooth functions in the general case will be
discussed in the following section.

2.2.1 Banach spaces

Let Ω be an open subset of Rd . Let p P r1, `8s and k P N. For u P W k,p pΩq we set
¨ ˛ p1
ÿ p
}u}W k,p pΩq “ ˝ }B α u}Lp pΩq ‚ . (2.6)
|α|ďk

This defines a norm on W k,p pΩq. We could also consider the quantity
ÿ
}B α u}Lp pΩq , (2.7)
|α|ďk

which defines an equivalent norm on W k,p pΩq.

On H k pΩq we define an inner product by setting, for u, v P H k pΩq,
ÿ
�u, v�H k pΩq “ �B α u, B α v�L2 pΩq . (2.8)
|α|ďk

The corresponding norm is exactly (2.6) with p “ 2.

Remark 2.16. With the notation of Proposition 2.12, we observe that
}B α u}W k´|α|,p pΩq ď }u}W k,p pΩq ,
for ω Ă Ω we have
}u}W k,p pωq ď }u}W k,p pΩq ,
and for χ P C08 pΩq there exists Cχ ą 0 independant of u such that
}χu}W k,p pΩq ď Cχ }u}W k,p pΩq .

8 J. Royer - Université Toulouse 3

 Sobolev spaces

Theorem 2.17. Let k P N and p P r1, `8s. The Sobolev space W k,p pΩq, endowed with the
norm (2.7) or (2.6), is a Banach space. In particular, H k pΩq with the inner product (2.8) is
a Hilbert space.

Proof. Let pun qnPN be a Cauchy sequence in W k,p pΩq. The sequences pB α un qnPN for |α| ď k
are Cauchy sequences in Lp pΩq. Since Lp pΩq is complete by the Riesz-Fisher theorem, there
exist vα P Lp pΩq for |α| ď k such that B α un goes to vα . For |α| ď k and φ P C08 pΩq we have
ż ż ż ż
p´1q|α| v0 B α φ dx “ p´1q|α| lim un B α φ dx “ lim B α un φ dx “ vα φ dx.
Ω nÑ`8 Ω nÑ`8 Ω Ω

This proves that in the sense of distributions we have B α v0 “ vα P Lp pΩq. Then v0 P W k,p pΩq
and ÿ
p 2
}un ´ v0 }W k,p pΩq “ }B α un ´ vα }Lp pΩq ÝÝÝÝÝÑ 0.
nÑ`8
|α|ďk

Thus the sequence pun qnPN has a limit in W k,p pΩq. This proves that W k,p pΩq is complete.

The proofs of the following two results are omitted (see [Brézis]).

Theorem 2.18. If p Ps1, `8r then W k,p pΩq is reflexive.

Theorem 2.19. If p P r1, `8r then W k,p pΩq is separable.

Proposition 2.20. Let s ě 0. Then the map

ˆż ˙ 12
` 2 ˘s
pu, vq ÞÑ 1 ` |ξ| ûpξqv̂pξq dξ
Rd

defines a scalar product on H s pRd q. When s is an integer, this norm is equivalent to the
norm defined by (2.6) with p “ 2.

2.2.2 Approximation by smooth functions

We know that for p P r1, `8r the set C08 pΩq of smooth and compactly supported functions
on the open set Ω is dense in Lp pΩq. In this paragraph we will see in what sense we can
approach functions in W k,p pΩq by smooth functions.
More precisely, we prove the density of smooth functions in the Sobolev spaces when
Ω “ Rd . This will not be the case in general domains. Since the closure of C08 pΩq in
W k,p pΩq will play an important role in applications, we introduce the following notation.

Definition 2.21. For k P N and p P r1, `8r we denote by W0k,p pΩq the closure of C08 pΩq in
W k,p pΩq. We also set H0k pΩq “ W01,2 pΩq.

Exercise 9. For x Ps ´ 1, 1r we set upxq “ 1. Prove that for p P r1, `8s there is no sequence
pun qnPN in C08 ps ´ 1, 1rq which goes to u in W 1,p ps ´ 1, 1rq.

As in Lp pRd q, the proofs will rely on regularization by convolution with

ş a sequence of
mollifiers. Let ρ P C08 pRd , r0, 1sq be supported in Bp0, 1q and such that Rd ρ dx “ 1. For
n P N˚ and x P Rd we set ρn pxq “ nd ρpnxq.

Lemma 2.22.` 1Let

˘ Ω be an open subset of R . 8
d
Let n P N˚ and let ω be an open subset of Ω
such that B x, n Ă Ω for all x P ω. Let ρn P C0 pRd q be as above and let . Let u P W k,p pΩq.
Then ρn ˚ u P C 8 pRd q X W k,p pωq and for |α| ď k we have in the weak sense on ω

B α pρn ˚ uq “ ρn ˚ pB α uq.

Proof. We prove the case k “ 1, and the general case follows by induction. Let j P �1, d� and
φ P C08 pωq. We have
ż ż ż
´ pρn ˚ uqpxqBj φpxq dx “ ´ ρn pyq upx ´ yqBj φpxq dx dy.
ω 1
Bp0, n q ω

2020-2021 9
M2RI - Elliptic PDE and evolution problems

` ˘
For y P B 0, n1 the map x ÞÑ upx ´ yq belongs to W 1,p pωq, so
ż ż ż ż
´ pρn ˚ uqpxqBj φpxq dx “ ρn pyq Bj upx ´ yqφpxq dx dy “ pρn ˚ Bj uqpxqφpxq dx.
ω 1
Bp0, n q ω ω

The conclusion follows.

Notice that the lemma applies in particular with ω “ Ω “ Rd .
Given v P W k,p pRd q we set for n P N˚

vn “ ρn ˚ v.

Then vn P C 8 pRd q X W k,p pRd q and for α P Nd with |α| ď k we have

B α vn “ ρn ˚ pB α vq. (2.9)

In particular,
}vn ´ v}W k,p pRd q ÝÝÝÝÝÑ 0.
nÑ`8

Moreover, if v is compactly supported then vn P C08 pRd q for all n P N˚ .

Statement (2.9) can be seen as a particular case on Rd of the more general following result.
For the following two proofs we also consider χ P C08 pRd q supported in the ball Bp0, 2q of
radius 2 and equal to 1 on Bp0, 1q. Then for m P N˚ and x P Rd we set χm pxq “ χp m x
q.
Theorem 2.23. Let p P r1, `8r. Then C08 pRd q is dense in W k,p pRd q. In other words we
have W0k,p pRd q “ W k,p pRd q.
Proof. Let u P W k,p pRd q and ε ą 0. Let α P Nd with |α| ď k. By Proposition 2.12 we have
χm u P W k,p pRd q for all m P N˚ and
ÿ ˆα˙ › ›
α α
}B pχm uq ´ B u}Lp pRd q ď ›B α´β pχm ´ 1qB β u› p d .
β L pR q
0ďβďα

By the dominated convergence theorem we have for β ď α

ż
› α´β ›p › › ˇ β ˇ
›B χm B β u›Lp pRd q ď ›B α´β pχm ´ 1q›L8 pRd q ˇB upxqˇp dx ÝÝÝÝÝÑ 0,
|x|ěm mÑ`8

so there exists m P N˚ such that

ε
}u ´ χm u}W k,p pRd q ď .
2
We set v “ χm u, and for n P N˚ we set vn “ ρn ˚ v. Then vn P C08 pRd q and for all α P Nd
with |α| ď k we have by Lemma 2.22

}B α vn ´ B α v}Lp pRd q “ }ρn ˚ pB α vq ´ B α v}Lp pRd q ÝÝÝÝÝÑ 0.

nÑ`8

Then, if for n P N˚ large enough we set uε “ vn we have uε P C08 pRd q and

ε
}uε ´ v}W k,p pRd q ď ,
2
so finally
}uε ´ u}W k,p pRd q ď ε.
Remark 2.24. For any ε ą 0 the function uε constructed in the previous proof is such that
}uε }L8 pRd q ď }u}L8 pRd q .
The conclusion of Theorem 2.23 does not hold in a general domain Ω. In other words,
W0k,p pΩq ‰ W k,p pΩq (see Exercise 9 and Proposition 2.42 below). However we have the
following weaker result of approximation by regular functions on any compact subset of Ω.
For a result of approximation on the whole domain Ω we refer to Proposition 2.32 below.

10 J. Royer - Université Toulouse 3

 Sobolev spaces

Theorem 2.25. Let p P r1, `8r and k P N. Let Ω be an open subset of Rd . Let u P W k,p pΩq.
There exists a sequence pun qnPN in C08 pRd q such that un|Ω goes to u in Lp pΩq and for any
open bounded subset ω such that ω Ă Ω we have
› ›
›un|ω ´ u|ω › k,p ÝÝÝÝÝÑ 0.
W pωq nÑ`8

Proof. For x P Rd we set

#
upxq if x P Ω,
vpxq “
0 if x P Rd zΩ.

Then we set vn “ ρn ˚ v P C 8 pRd q and un “ χn vn . We have

}un ´ u}Lp pΩq ď }un ´ v}Lp pRd q ď }χn pρn ˚ vq ´ χn v}Lp pΩq ` }χn v ´ v}Lp pΩq ÝÝÝÝÝÑ 0.
nÑ`8

` ˘ `
Let N P N be so large that B x, N1 Ă Ω and χN “ 1 on B x, N1 q for all x P ω. Then for
|α| ď k we have by Lemma 2.22

}B α pun ´ uq}Lp pωq “ }B α pvn ´ vq}Lp pωq “ }ρn ˚ pB α vq ´ B α v}Lp pωq ÝÝÝÝÝÑ 0.
nÑ`8

This proves that χvn goes to u in W k,p pωq.

Exercise 10. Let Ω be a bounded subset of Rd and p P r1, `8r. Prove that W01,p pΩq is the
closure of C0k pΩq in W 1,p pΩq.

2.2.3 Examples of properties proved by density

It is not always convenient to prove results about differentiation in the weak sense, and
most of the properties of Sobolev spaces are proved by density. We first prove the result for
regular functions (smooth, or of class C k for a property in W k,p ), and then the general case
is deduced by density.
Here we give some examples of results which are already known for regular functions and
which can be extended in the suitable Sobolev spaces by density.

We begin with the integration by parts on Rd .

Proposition 2.26 (Green Formula on Rd ). Let Ω be an open subset of Rd and u, v P H01 pΩq.
For j P �1, d� we have
ż ż
pBj uqv dx “ ´ upBj vq dx.
Ω Ω

Proof. Let pun qnPN and pvn qnPN be sequences in C08 pRd q which go to u and v in H 1 pRd q.
The Green formula for smooth and compactly supported functions gives, for all n P N,
ż ż
pBj un qvn dx “ ´ un pBj vn q dx.
Ω Ω

Taking the limit n Ñ `8 gives the result.

We continue with the product of differentiable functions. If u and v are continuously

differentiable, then so is the product uv. The same result holds for weak derivatives. Notice
that in this result and the following we do not take functions in W01,p pΩq. The approximation
by regular functions is given by Theorem 2.25.

Proposition 2.27 (Differentiation of a product). Let Ω be an open subset of R d . Let p P

r1, `8s and u, v P W 1,p pΩq X L8 pΩq. Then uv P W 1,p pΩq and, for j P �1, d�,

Bj puvq “ pBj uqv ` upBj vq. (2.10)

2020-2021 11
M2RI - Elliptic PDE and evolution problems

Proof. Assume that p ă `8. Let pun qnPN be a sequence in C08 pRd q as given by Theorem
2.25. After extraction of a subsequence if necessary, we can assume that un pxq tends to upxq
for almost all x P Rd . By Remark 2.24, we can also assume that }un }L8 pRd q ď }u}L8 pRd q for
all n P N. By Proposition 2.12, we have un v P W 1,p for all n P N and, for j P �1, d� and
φ P C08 pRd q, ż ż
` ˘
´ un vBj φ dx “ pBj un qv ` un pBj vq φ dx.
Rd Rd

The limit n Ñ `8 yields (2.10). In particular Bj puvq P Lp pRd q, and the proof is complete if
p ă `8.
Now assume that p “ `8. Then uv and pBj uqv ` upBj vq are in L8 pRd q. Let φ P C08 pRd q.
Let χ P C08 pRd q be equal to 1 on a neighborhood of supppφq. Then χu and χv are in
W 1,p pRd q for any p P r1, `8r so
ż ż ż
` ˘
´ uvBj φ dx “ ´ χu χvBj φ dx “ Bj pχuq χv ` χu Bj pχvq φ dx
Rd
ż R Rd
d

` ˘
“ pBj uqv ` upBj vq φ dx.
Rd

This proves (2.10) and concludes the proof.

Then we discuss the chain rule, which will be important in particular for changes of
variables.
Proposition 2.28 (Chain rule). Let Ω1 and Ω2 be two open subsets in Rd , and let Φ “
pΦ1 , . . . , Φd q : Ω1 Ñ Ω2 be a diffeomorphism of class C 1 . We assume that JacpΦq and
JacpΦ´1 q are bounded on Ω1 and Ω2 , respectively. Let p P r1, `8s. Then for u P W 1,p pΩ2 q
we have u ˝ Φ P W 1,p pΩ1 q and for j P �1, d�,

ÿ
d
` ˘
Bj pu ˝ Φq “ pBk uq ˝ Φ Bj Φk .
k“1

In particular there exists CΦ ą 0 such that

}u ˝ Φ}W 1,p pΩ2 q ď CΦ }u}W 1,p pΩ1 q .

Proof. Assume that p ă `8. Let pun qnPN be a sequence in C08 pRd q which goes to u in the
sense of Theorem 2.25. Let ψ P C08 pΩ1 q, K1 “ supppψq and K2 “ ΦpK1 q. Then K2 is a
compact of Ω2 . For n P N and ψ P C08 pRd q we have
ż d ż
ÿ
´ pun ˝ ΦqBj ψ dx “ pBk un ˝ ΦqBj Φk ψ dx. (2.11)
Rd k“1 R
d

By the change of variables y “ Φpxq we have

ż
p p
}un ˝ Φ ´ u ˝ Φ}Lp pK1 q “ |pun ´ uqpΦpxqq| dx
żK1
pˇ ˇ
“ |pun ´ uqpyq| ˇJΦ´1 pyqˇ dy
K
› 2 ›
ď ›JΦ´1 ›L8 pK2 q }un ´ u}Lp pK2 q
p

ÝÝÝÝÝÑ 0.
nÑ`8

We similarly have, for all j, k P �1, d�,

}pBj un ˝ ΦqBk Φj ´ pBj u ˝ ΦqBk Φj }Lp pK1 q

ď }Bk Φj }L8 pK1 q }pBj un ˝ Φq ´ pBj u ˝ Φq}Lp pK1 q ÝÝÝÝÝÑ 0.
nÑ`8

We take the limit n Ñ `8 in (2.11) and conclude when p ă `8. The case p “ `8
follows as in the proof of Proposition 2.27.

12 J. Royer - Université Toulouse 3

 Sobolev spaces

We finish this paragraph with the characterisation of H 1 pRd q by the difference quotients.
We recall that the classical notion of differentiability is defined by looking at the limit at
each point of the difference quotient. The following result gives a link between this point of
view and the weak derivative.
For u P L2 pRd q and h P Rd z t0u we define the diffence quotient Dh u P L2 pRd q by
upx ` hq ´ upxq
Dh upxq “ . (2.12)
|h|
Notice that for u, v P L2 pRd q we have
ż ż
pDh uqv dx “ upD´h vq dx. (2.13)
Rd Rd

Proposition 2.29. Let u P L2 pΩq.

(i) Assume that u P H 1 pΩq. Then for h P Rd z t0u we have
}Dh u}L2 pRd q ď }∇u}L2 pRd q .

(ii) Assume that there exists C ą 0 such that for all h P Rd z t0u we have
}Dh u}L2 pΩq ď C.

Then u P H 1 pRd q and for all j P �1, d� we have

› ›
› Bu ›
› ›
› Bxj › 2 d ď C.
L pR q

Proof. ‚ Assume that u P C08 pRd q. Then for x P Rd we have by the Cauchy-Schwarz
inequality
ˆż 1 ˙2 ż1
2 2 2
|upx ` hq ´ upxq| ď |∇upx ` thq| |h| dt ď |h| |∇upx ` thq| dt.
0 0

Then, by the Fubini Theorem and the change of variables y “ x ` th,

ż ż1 ż
2 2 2 2 2
|upx ` hq ´ upxq| dx ď |h| |∇upx ` thq| dx dt ď |h| }∇u}L2 pRd q .
Rd t“0 xPRd

This gives the first property.

‚ We denote by pe1 , . . . , ed q the canonical basis of Rd . Let j P �1, d�. Let φ P C08 pRd q. For
t P R˚ we have
ˇ� � ˇ ˇ� � ˇ
ˇ ˇ ˇ ˇ
ˇ u, D´tej φ L2 pRd q ˇ “ ˇ Dtej u, φ L2 pRd q ˇ ď C }φ}L2 pRd q .

On the other hand, u P L1loc pRd q so by the dominated convergence theorem we have
ˇ ˇ ˇ � � ˇ
ˇ ˇ ˇ ˇ
ˇ´ �u, Bj φ�L2 pRd q ˇ “ ˇlim u, D´tej φ L2 pRd q ˇ ď C }φ}L2 pRd q .
tÑ0

By the Riesz Theorem there exists vj P L2 pRd q such that

@φ P C08 pRd q, ´ �u, Bj φ�L2 pRd q “ �vj , φ�L2 pRd q .

This proves that u P H 1 pRd q with, for all j P �1, d�,

› ›
› Bu ›
› ›
› Bxj › 2 d “ }vj }L2 pRd q ď C.
L pR q

The proof is complete.

Exercise 11. 1. Prove that the first statement of Proposition 2.29 holds in W 1,p pRd q for
any p P r1, `8r.
2. Prove that the second statement of Proposition 2.29 holds in W 1,p pRd q for any p Ps1, `8r.

2020-2021 13
M2RI - Elliptic PDE and evolution problems

2.3 Sobolev spaces on domains with boundary

In the previous section we have given some properties of the Sobolev spaces on Rd , or local
properties in general domains. In this section we look more carefully at the behavior of
functions in Sobolev spaces at the boundary of the domain.

The model case will be the half space

� (
Rd` “ x “ px1 , . . . , xd q P Rd : x1 ą 0 .

This is the simplest case because the boundary BRd` “ t0u ˆ Rd´1 is flat. Then, if the open
subset Ω of Rd is sufficiently regular, the boundary BΩ can be locally straightened out and,
with a partition of unity and a change of variables for each part, the problem on Ω is reduced
to a problem far from the boundary (where we can apply the results on Rd ) and a finite
number of problems on Rd` .

It is the purpose of this section to make these ideas clearer and to deduce some results
for the Sobolev spaces on bounded subsets.

2.3.1 Regular domains

Let k P N˚ Y t8u. We recall that an open subset Ω of Rd is said to be of class C k if for any
w P BΩ there exist an orthonormal basis β “ pβ1 , . . . , βd q of Rd , an open subset O of Rd´1 ,
a, b P R with a ă b and an application ϕ : O Ñsa, br of class C k such that U defined by
# +
ÿd
U“ xj βj , px2 , . . . , xd q P O, x1 Psa, br
j“1

is a neighborhood of w in Rd and
# +
ÿd
‰ “
ΩXU “ xj βj , x1 “ px2 , . . . , xd q P O, x1 P ϕpx1 q, b .
j“1

In particular, in U the boundary BΩ is the graph of ϕ in the basis β. We can always construct
the basis β with the vectors of the canonical basis pe1 , . . . , ed q, possibly in a different order.
For x1 “ px2 , . . . , xd q P O we set

ÿ
d
ϕ̃px1 q “ ϕpx1 qβ1 ` xj βj .
j“2

Then BΩ X U is also the image of O by ϕ̃.

Given w P BΩ X V and x1 “ px2 , . . . , xd q P O such that w “ φ̃px1 q, the outward normal
derivative to Ω at point w is defined by
řd
´β1 ` j“2 Bj ϕpx1 qβj
νpwq “ b . (2.14)
2
1 ` |∇ϕpx1 q|

The is the only vector such that νpwqKTw pBΩq, |νpwq| “ 1 and, for some t0 ą 0,
#
P Ω, @t Ps ´ t0 , 0s,
w ` tνpwq
R Ω, @t Ps0, t0 r.

We define on BΩ the topology and the corresponding Borel σ-algebra inherited from the
usual structure on Rd . We define the Lebesgue measure of a Borel set B P BΩ X U as follows:
ż b
1B pϕ̃pxqq 1 ` |∇ϕpx1 q| dx1 .
2
σpBq “
O

14 J. Royer - Université Toulouse 3

 Sobolev spaces

Thus, if f is an integrable function on BΩ X U we have

ż ż b
f pϕ̃pxqq 1 ` |∇ϕpx1 q| dx1 .
2
f dσ “
BΩXU O

Then we can define Lebesgue spaces on BΩ as on any measure space.

řd
For x “ j“1 xj βj P U we set

` ˘ ÿ
d
Φpxq “ x1 ´ ϕpx2 , . . . , xd q e1 ` xj ej .
j“2

Then Φ is of classe C k and it is injective. So it defines a bijection on its image denoted

by W. Then W is open in Rd and the inverse Φ´1 of Φ is of class C k on W (Φ defines a
diffeomorphism of class C k from U to W). Moreover we have

ΦpU X Ωq “ W X Rd` .

Notice also that for x1 P O we have Φpϕ̃px1 qq “ p0, x1 q, and then Wj X BRd` “ t0u ˆ O.
The interest of this change of variables is to transform a function supported in Ω X U to
a function on Rd` , where the properties of Sobolev spaces are easier.
Notice that if Ω is bounded then its boundary BΩ is compact. This is not necessary but
it will simplify the discussion (an unbounded open subset can also have a compact boundary,
but we will not consider this situation here).

Now let Ω be a bounded open subset of Rd of class C k for some k ě 1. There exist
N P N˚ , open subsets U1 , . . . , UN , W1 , . . . , WN of Rd and diffeomorphisms Φj : Uj Ñ Wj of
ŤN
class C k such that BΩ Ă j“1 Uj and for all j P �1, N � we have ΦpΩ X Uj q “ Rd` X Wj .
ŤN
If we set Ω “ U0 then j“0 Uj is an open cover of Ω. We consider a corresponding
partition of unity pχj q0ďjďN (χj P C08 pRd , r0, 1sq is supported in Uj for all j P �0, N � and
řN
j“0 χj “ 1 on Ω).
řN
For u P W 1,p pΩq we set uj “ χj u for all j P �0, N �. Then u “ j“0 uj , uj P W 1,p pΩq
for all j P �0, N �, u0 is supported in a compact subset of Ω, and uj is supported in a com-
pact subset of Ω X Uj for all j P �1, N �. In particular, the extension of u0 by 0 on Rd is in
W 1,p pRd q, and puj ˝ Φ´1 q belongs to W 1,p pRd` X Wj q (and can be extended by 0 to a function
in W 1,p pRd` q) for all j P �1, N �.

We will use this setting to prove results for Sobolev spaces on Ω.

2.3.2 Extension
We begin with a result of extension. In order to deduce results in W 1,p pΩq from results on
W 1,p pRd q it is natural to extend functions in W 1,p pΩq to functions in W 1,p pRd q (notice that
in the proof of Theorem 2.25 we were able to prove results on W k,p pωq for ω ĂĂ Ω precisely
because we had a function with a nice behavior on a bigger domain).
It is clear, at least in dimension 1, than extending functions by 0 outside Ω does not
always give a function in W 1,p pRd q. However, we have seen in Exercise 4 that in dimension
1 we can indeed extend a function in H 1 pR˚` q to a function in H 1 pRq. We generalize this
observation to the case of a function in W 1,p pRd` q and then, by the argument described above,
to the case of a function in W 1,p pΩq for a regular bounded open subset Ω.

Proposition 2.30. Let p P r1, `8s. For u P Lp pRd` q and x “ px1 , . . . , xd q “ px1 , x1 q P Rd
we set #
upx1 , x1 q if x1 ą 0,
pP uqpxq “
up´x1 , x1 q if x1 ă 0.

2020-2021 15
M2RI - Elliptic PDE and evolution problems

1
Then P u P Lp pRd q and }P u}Lp pRd q “ 2 p }u}Lp pRd q . For u P W 1,p pRd` q we have P u P
`

W 1,p pRd q with

B1 pP uq “ P̃ pB1 uq and Bj pP uq “ P pBj uq, 2 ď j ď d,
where #
vpx1 , x1 q if x1 ą 0,
pP̃ vqpxq “
´vp´x1 , x1 q if x1 ă 0.
In particular, P defines a continuous extension from W 1,p pRd` q to W 1,p pRd q.
p
Proof. ‚ We set Rd´ “ Rd zRd` . It is easy to see that }P u|Rd´ }pLp pRd q “ }u}Lp pRd q if p ă `8,
´ `
p p
so P u P Lp pRd q with }P u}Lp pRd q “ 2 }u}Lp pRd q . If p “ `8 we have }P u}L8 pRd q “ }u}L8 pRd q .
` `

‚ For x “ px1 , . . . , xd q P Rd we set σpxq “ p´x1 , x2 , . . . , xd q. Let j P �2, d�. Let φ P C08 pRd q.
If φ is supported in Rd` we have
ż ż ż ż
´ P u Bj φ dx “ ´ u Bj φ dx “ Bj u φ dx “ P pBj uq φ dx.
Rd Rd
` Rd
` Rd

If φ is supported in Rd´ then, similarly,

ż ż ż
´ P u Bj φ dx “ ´ pu ˝ σq Bj φ dx “ ´ u pBj φ ˝ σq dx
Rd Rd
´ Rd
`
ż ż ż
“´ u Bj pφ ˝ σq dx “ Bj u pφ ˝ σq dx “ ppBj u ˝ σqφ dx
Rd
` Rd
` Rd
´
ż
“ P pBj uq φ dx.
Rd

We consider the general case. Let χ P C08 pR, r0, 1sq be even, equal to 1 on [-1,1] and supported
in ]-2,2[. For n P N and x P Rd we set χn pxq “ χpnx1 q. Since p1 ´ χn qφ is supported outside
BRd` we have
ż ż ż
` ˘
´ P up1 ´ χn qBj φ dx “ ´ P uBj p1 ´ χn qφ dx “ P pBj uqp1 ´ χn qφ dx.
Rd Rd Rd
`

By the dominated convergence theorem this yields

ż ż
´ P u Bj φ dx “ P pBj uqφ dx.
Rd Rd

This proves that in the weak sense we have Bj pP uq “ P pBj uq. In particular Bj pP uq P Lp pRd q
1
with }Bj pP uq}Lp pRd q “ 2 p }Bj u}Lp pRd q .
`

‚ We proceed similarly for the first partial derivative. We observe that for φ P C 08 pRd q we
now have B1 pφ ˝ σq “ ´pB1 φq ˝ σ, so if φ is supported outside BRd` we now have
ż ż
´ P u B1 φ dx “ P̃ pB1 uqφ dx.
Rd Rd

On the other hand p1 ´ χn q does not commute with the partial derivative B1 . But the
additional term is estimated as follows. Let R ą 0 be such that φ is supported in R ˆ
Bd´1 p0, Rq (Bd´1 p0, Rq is the ball of radius R in Rd´1 ). Since χ is even we have
ˇż ˇ ˇˇż ˇ
ˇ
ˇ ˇ ˇ ˇ
ˇ P u B χ φ dx ˇ“ˇ u B χ pφ ´ φ ˝ σq dx ˇ
ˇ d 1 n ˇ ˇ d 1 n
ˇ
R R`
ż n2 ż
› › ˇ ˇˇ ˇ
ď n ›χ1 ›8 ˇupx1 , x1 qˇ ˇφpx1 , x1 q ´ φp´x1 , x1 qˇ dx1 dx1
x1 “0 x1 PBd´1 p0,Rq
ż 2 ż
› › n ˇ ˇ
ď 4 ›χ1 ›8 }B1 φ}8 ˇupx1 , x1 qˇ dx1 dx1
x1 “0 x1 PBd´1 p0,Rq

ÝÝÝÝÝ
Ñ 0.
nÑ`8

The conclusion follows as above.

16 J. Royer - Université Toulouse 3

 Sobolev spaces

Theorem 2.31. Let Ω be an open bounded subset of class C 1 in Rd . Let p P r1, `8s. Let
O be an open subset of Rd such that Ω Ă O. Then there exists a bounded linear operator
P : W 1,p pΩq Ñ W 1,p pRd q (which is also bounded for the norm of Lp pΩq) such that P u is
supported in O and pP uq|Ω “ u for all u P W 1,p pΩq.
Proof. Let u P W 1,p pΩq. We use the notation introduced in Paragraph 2.3.1. Without loss
of generality we can assume that Uj Ă O and Wj is symmetric with respect to BRd` for all
j P �1, N � (for instance Wj is a ball centered on BRd` ). We denote by v0 the extension of
u0 by 0 on Rd . We have }v0 }W 1,p pRd q “ }u0 }W 1,p pΩq . Let j P �1, N �. We denote by ṽj the
extension of uj ˝ Φ´1 j on Wj given by Proposition 2.30. It is supported in a compact subset
of Wj , and ṽj ˝ Φj is compactly supported in Uj . Then we denote by vj the extension by
0 of ṽj ˝ Φj on Rd . By Propositions 2.28 and 2.30 and Remark 2.16 there exist constants
CΦ , CΦ´1 , CP , Cχj ą 0 independant of u such that
› ›
}vj }W 1,p pRd q “ }ṽj ˝ Φj }W 1,p pUj q ď CΦ }ṽj }W 1,p pWj q ď CP CΦ ›uj ˝ Φ´1 ›W 1,p pW XRd q
` j

ď C P CΦ C }uj }W 1,p pUj XΩq ď CP CΦ C

Φ´1 Cχj }u}W 1,p pΩq .
Φ´1
řN
Finally we set P u “ j“0 vj , and P u P W 1,p pRd q satisfies all the required properties.

We recall that given an open subset Ω of Rd , we denote by C08 pΩq the set of restrictions
to Ω of functions in C08 pRd q. When k “ 1, the following result follows from Theorem 2.31
(applied with O “ Rd ) and Theorem 2.23. For the general case we can extend Theorem 2.31
to the case k ě 2, we can also give a direct proof.
Proposition 2.32. Let p P r1, `8r and k P N. Let Ω be equal to Rd` or be a bounded open
subset of Rd . Let u P W k,p pRd` q. There exists a sequence pun qnPN of functions in C08 pΩq
such that
}un ´ u}W k,p pΩq ÝÝÝÝÝÑ 0.
nÑ`8

Proof. We prove the case Ω bounded. The case Ω “ Rd` is more direct and is left as an
exercice. Let w P BΩ. We use the notation of Paragraph 2.3.1. Let u P W 1,p pΩq be
supported in Ω X U . We denote by ũ the extension of u by 0 on Rd . For τ ą 0 we set
# +
ÿ
d
Uτ “ xj βj , px2 , . . . , xd q P O, x1 Psa, b ´ τ r .
j“1

There exists τ0 ą 0 such that supppuq Ă Uτ . For τ Ps0, τ0 s and x P Uτ we set

uτ pxq “ upx ` τ β1 q.
We extend uτ by 0 on UzUτ . The restriction of uτ to U X Ω is in W k,p pU X Ωq and the
derivatives of uτ up to order k are the translations of the corresponding derivatives of u:
B α puτ q “ pB α uqτ .
By continuity in Lp pRd q of the translation we have
p
ÿ p
ÿ p
}uτ ´ u}W k,p pU XΩq “ }B α uτ ´ B α u}Lp pU XΩq “ }vτα ´ v α }Lp pRd q ÝÝÝÑ 0,
τ Ñ0
|α|ďk |α|ďk

where we have denoted by vτα and v α the extensions by 0 of B α uτ and B α u (then vτα pxq “
v α px ` τ β1 q almost everywhere on Rd ).
Now let τ Ps0, τ0 s be fixed. There exists η0 ą 0 if we set
ď
V“ Bpx, η0 q,
xPsupppuτ qXΩ

then for all y Pş V we have y ` τ β1 P U X Ω. Let ρ P C08 pRd , r0, 1sq be supported in Bp0, 1q
and such that Rd ρ “ 1. For η Ps0, η0 s and x P Rd we set ρη pxq “ η ´d ρpx{ηq. For η Ps0, η0 s
we set uητ “ ρη ˚ uτ . Its restriction to U X Ω belongs to C08 pΩq. Since uτ P W k,p pVq we can
prove as in the proof of Theorem 2.23 that
}uητ ´ uτ }W k,p pU XΩq ÝÝÝÑ 0.
ηÑ0

It remains to chose τ ą 0 small enough and then η ą 0 small enough to conclude.

2020-2021 17
M2RI - Elliptic PDE and evolution problems

2.4 Sobolev inequalities

In this section we prove some inclusions between Sobolev spaces. The inclusions between
Lebesgue spaces are already known. In particular we know that Lp pRd q is never included in
Lq pRd q if p ‰ q. The purpose here is to prove that if we add information about the derivatives
then we get better results. In particular we will prove (continuous) inclusions of the form
W 1,p pRd q Ă Lq pRd q for suitable pairs pp, qq.
As for Lebesgue spaces, we get stronger results on a bounded domain Ω. In this case
we will prove compact inclusions. For instance, H 1 pΩq is compactly embedded in L2 pΩq.
This means that if a sequence of functions in H 1 pΩq is bounded, then it has a convergent
subsequence for the L2 pΩq norm. This result will be of great importance for the analysis of
PDEs. We will already use this fact in the following section (see the proof of Theorem 2.49).

2.4.1 Morrey’s inequality

We have already seen in Proposition 2.6 that if I is an interval of R then for any p P r1, `8s
we have W 1,p pIq Ă L8 pIq (and that a function in W 1,p pIq is also Hölder-continuous
‰ if p ą 1).
This result only holds in dimension 1. Indeed if 1 ď p ă d then for α P ´ dp ` 1, 0r the
function x ÞÑ |x| belongs to W 1,p pBp0, 1qq but not to L8 pBp0, 1qq. The purpose of the
α

following theorem is to prove that we recover a result analogous to Proposition 2.6 if p ą d.

Exercise 12. Find u P W 1,d pRd q such that u R L8 pRd q.

Theorem 2.33 (Morrey’s inequality). Let p Psd, `8r. We have

d
W 1,p pRd q Ă L8 pRd q X C 0,1´ p pRd q

and there exists C ą 0 such that, for u P W 1,p pRd q,

}u}L8 pRd q ď C }u}W 1,p pRd q ,

1´ d
@x1 , x2 P Rd , |upx1 q ´ upx2 q| ď C |x1 ´ x2 | p
}∇u}Lp pRd q .

In dimension 1, we have used the fundamental theorem of calculus to compare upxq to

upx0 q for some fixed x0 . It gave a one-dimensional integral which was controled by the
norm of u1 . In higher dimension we can still write the fundamental theorem of calculus for
regular functions but the corresponding one-dimensional integral is not controled by the d-
dimensional integral which defines the norm of ∇u. The trick in the following proof is to
compare upxq to the mean value of u on an open subset of Rd . This will give a d-dimensional
integral controled as stated in the theorem.

Proof. ‚ We consider u P C08 pRd q. The general case will follow by density. Let x P Rd and
let O be an open subset of Rd . We set

δpx, Oq “ sup |y ´ x| .
yPO

For y P O and h “ ph1 , . . . , hd q “ y ´ x we have

ż1ˇ ˇ
ˇd ˇ
|upyq ´ upxq| ď ˇ upx ` thqˇ dt
ˇ dt ˇ
0
ż1 ÿ d
ď |hj | |Bj upx ` thq| dt
0 j“1

d ż1
ÿ
ď δpx, Oq |Bj upx ` thq| dt.
j“1 0

For t Ps0, 1s we set

tpO ´ xq “ ttpy ´ xq, y P Ou .

18 J. Royer - Université Toulouse 3

 Sobolev spaces

If we set ż
1
uO “ upyq dy,
|O| O
then we have
ż
1
|upxq ´ uO | ď |upxq ´ upyq| dy
|O| yPO
ż ż1 ÿ
d
δpx, Oq
ď |Bj upx ` thq| dt dh
|O| hPpO´xq 0 j“1
ż1 d ż
δpx, Oq 1 ÿ
ď |Bj upx ` ηq| dη dt.
|O| 0 td j“1 ηPtpO´xq

By the Hölder inequality we have for t P r0, 1s

˜ż ¸ p1
d ż
ÿ ÿ
d
p´1
p
|Bj upx ` ηq| dη ď |Bj upx1 ` ηq| dη |tpO ´ xq| p

j“1 tpO´xq j“1 tpO´xq

dpp´1q p´1
ďt p |O| p
}∇u}Lp pRd q ,

so
ż1 δpx, Oq |O|
´p
1
}∇u}Lp pRd q
´p
1
´d
|upxq ´ uO | ď δpx, Oq |O| }∇u}Lp pRd q t p dt “ d
. (2.15)
0 1´ p

‚ Now let x1 , x2 P Rd and let O be the open ball with diameter rx1 , x2 s. We have δpx1 , Oq “
d
δpx2 , Oq “ |x1 ´ x2 | and |O| “ 2cdd |x1 ´ x2 | where cd is the size of the unit ball in Rd . Thus
d ´p
1
21` p cd 1´ d
|upx1 q ´ upx2 q| ď |upx1 q ´ uO | ` |upx2 q ´ uO | ď |x1 ´ x2 | p
}∇u}Lp pRd q .
1 ´ dp

This gives the second statement. Now for x P Rd we apply (2.15) with O “ Bpx, 1q, the ball
of center x and radius 1. The Hölder inequality gives
´p
1
|uO | ď cd }u}Lp pRd q ,

so ˜ ¸
´1 1
|upxq| ď cd p }u}Lp pRd q ` d
}∇u}Lp pRd q .
1´ p

This completes the proof.

2.4.2 Gagliardo-Nirenberg Inequality

In this paragraph we consider the case p ď d. This is particularly interesting for the common
case p “ 2.
We want to prove that if we control ∇u in some Lebesgue space, then we can control u
in another Lebesgue space. Assume that there exists q P r1, `8r and C ą 0 such that

@v P C08 pRd q, }v}Lq pRd q ď C }∇v}Lp pRd q . (2.16)

Let u P W 1,p pRd qz t0u. For λ ą 0 and x P Rd we set uλ pxq “ upλxq. Then for all λ ą 0 we
have
λ´ q }u}Lq pRd q “ }uλ }Lq pRd q ď C }∇uλ }Lp pRd q “ Cλ1´ p }u}Lp pRd q .
d d

Letting λ go to 0 or to `8 we see that we necessarily have

d d
´ “1´ . (2.17)
q p

2020-2021 19
M2RI - Elliptic PDE and evolution problems

In the following theorem we prove that if (2.17) holds then we indeed have (2.16). For
p P r1, dr we define p˚ P r1, `8r by
pd 1 1 1
p˚ “ , “ ´ . (2.18)
d´p p˚ p d
Notice that we have p˚ ą p and p˚ Ñ `8 if p Ñ d.
Theorem 2.34 (Gagliardo-Nirenberg-Sobolev inequality). Let p P r1, dr and let p ˚ be defined
by (2.18). There exists C ą 0 such that for all u P C01 pRd q we have

}u}Lp˚ pRd q ď C }∇u}Lp pRd q .

Proof. ‚ Let u P C01 pRd q. For x “ px1 , . . . , xd q P Rd and j P �1, d� we have

ˇż xj ˇ
ˇ ˇ
|upxq| “ ˇˇ Bj upx1 , . . . , xj´1 , t, xj`1 , . . . , xd q dtˇˇ ď vj px̃j qd´1
´8

where x̃j “ px1 , . . . , xj´1 , xj`1 , . . . , xd q and

ˆż ˙ d´1
1

vj px̃j q “ |∇upx1 , . . . , xj´1 , t, xj`1 , . . . , xd q| dt .

R

This gives
d ź
d
|upxq| d´1 ď vj px̃j q.
j“1

Now we prove by induction on d ě 2 that if we set

ź
d
v : x P Rd ÞÑ vj px̃j q,
j“1

then we have
ź
d
}v}L1 pRd q ď }ṽj }Ld´1 pRd´1 q . (2.19)
j“1

The case d “ 2 is easy. Assume that (2.19) is true up to the dimension d ´ 1 for some d ě 3.
We fix x1 P R and see v as a function of x1 “ px2 , . . . , xd q. By the Hölder inequality we have

ż ˜ż ¸ d´2
ź
d
d´1
d´1

vpx1 , x1 q dx1 ď }ṽ1 }Ld´1 pRd´1 q ṽj px1 , x̃1j q d´2 dx2 . . . dxd ,
Rd´1 Rd´1 j“2

where for j P �2, d� we have set x̃1j “ px2 , . . . , xj´1 , xj`1 , . . . , xd q. The induction assumption
gives
ż źd ź d ˆż ˙ d´2
1
d´1
1 d´2 1 1 d´1 1
ṽj px1 , x̃j q dx ď ṽj px1 , x̃j q dx̃j
Rd´1 j“2 j“2 Rd´2

and hence
ż d ˆż
ź ˙ d´1
1

1 1
vpx1 , x q dx ď }ṽ1 }Ld´1 pRd´1 q ṽj px1 , x̃1j qd´1 dx̃1j .
Rd´1 j“2 Rd´2

After integration over x1 P R we get, by the Hölder inequality,

ż d ˆż
ź ˙ d´1
1

d´1
vpxq dx ď }ṽ1 }Ld´1 pRd´1 q ṽj px̃j q dx̃j .
Rd j“2 Rd´1

This is (2.19). We deduce

ż ˆż ˙ d´1
d
d
|upxq| d´1
dx ď |∇upxq| dx ,
Rd Rd

20 J. Royer - Université Toulouse 3

 Sobolev spaces

which gives the result for u P C01 pRd q when p “ 1.

γ´1 γ´1
‚ Let γ ą 1. The case p “ 1 applied to |u| u (still in C01 pRd q, with gradient γ |u| ∇u)
gives
ˆż ˙ d´1 ż ˆż ˙ p´1 ˆż ˙ p1
γd d pγ´1qd p
γ´1 p
|u| d´1 ďγ |u| |∇u| dx ď γ |u| d´1
dx |∇u| dx .
Rd Rd Rd Rd
(2.20)
If we choose
ppd ´ 1q
γ“ ą1
d´p
we have
γd pγ ´ 1qp dp
“ “ “ p˚ ,
d´1 p´1 d´p
and the conclusion follows for u P C01 pRd q. The general case u P W 1,p pRd q follows by
density.
In Theorem 2.34 we have only used the fact that ∇u P Lp pRd q. If u is also in Lp pRd q
˚
we have better conclusions. We know that Lp pRd q X Lp pRd q Ă Lq pRd q (with continuous
˚
inclusion) for any q P rp, p s. This is the first statement of the following theorem. The second
statement is about the limit case p “ d. Notice that Theorem 2.34 does not hold with p “ d
and p˚ “ `8 (see Exercise 12), but for u P W 1,d pRd q we have a result similar to the case
p ă d.
Theorem 2.35. (i) Let p P r1, dr. Then for all q P rp, p˚ s we have W 1,p pRd q Ă Lq pRd q
with continuous injection.
(ii) For all q P rd, `8r we have W 1,d pRd q Ă Lq pRd q with continuous injection.

“ γd ‰We prove the second statement. We prove 1by induction

Proof. on γ ě d ´ 1 that for q P
d, d´1 there exists Cq ą 0 such that, for all u P Cc pRd q,

}u}Lq pRd q ď Cq }u}W 1,d pRd q . (2.21)

The result will follow by density. (2.21) is clear when γ “ d ´ 1. We assume that it is proved
up to γ ´ 1 for some γ ě d. Let u P Cc1 pRd q. We use estimate (2.20) from the previous proof
with p “ d. With the induction assumption this gives
γ γ´1 γ
}u} γd ď γ }u} pγ´1qd }∇u}Ld pRd q ď γC γ´1
pγ´1qd }u}W 1,d pRd q .
L d´1 pRd q L d´1 pRd q d´1

γd
“ γd ‰
This gives (2.21) for q “ d´1 . The case q P d, d´1 follows since u belongs to Ld pRd q.

2.4.3 Sobolev embeddings on a bounded domain

So far we have only proved inclusions between Sobolev spaces on Rd . Our purpose in this
paragraph is to prove analogous results for Sobolev spaces on a bounded open subset Ω. For
this, we will use the extension operator of Theorem 2.31 to deduce inequalities on Ω from
their analogs on Rd .
However, as said in introduction, we will get better results on Ω. For instance we recall
that Lp pΩq Ă Lq pΩq if p ą q. This will automatically improve the result of Theorem 2.34 (in
particular the discussion before Theorem 2.34 is not valid on a bounded domain).
Another very important difference between the case of Rd and the case of a bounded
domain is that some inclusions will be not only continuous but also compact.
Definition 2.36. Let X and Y be Banach spaces. A bounded linear operator T : X Ñ Y is
said to be compact if for any bounded sequence pun qnPN P X N , the sequence pT un qnPN has a
convergent subsequence in Y . Equivalently, T is compact if T pBX q is compact in Y , where
BX is the unit ball in X.
Compactness for a set of functions is usually given by the Ascoli-Arzelá Theorem, which
we recall now.

2020-2021 21
M2RI - Elliptic PDE and evolution problems

Theorem 2.37 (Ascoli-Arzelá Theorem). Let K be a compact metric space and let F be a
bounded subset of CpK, Rq. We assume that F is equicontinuous:

@ε ą 0, Dδ ą 0, @f P F, @x, y P K, dpx, yq ď δ ùñ |f pxq ´ f pyq| ď ε.

Then the closure F of F in CpKq is compact.

The results of Theorem 2.35 and 2.33 are extended to bounded domains as follows.

Theorem 2.38. Let Ω be a bounded open subset of class C 1 in Rd . Let p P r1, `8s. Then
we have the following compact inclusions.

(i) If p ă d then for all q P r1, p˚ r we have W 1,p pΩq ĂĂ Lq pΩq.

(ii) For all q P rd, `8r we have W 1,d pΩq ĂĂ Lq pΩq.

(iii) If p ą d then we have W 1,p pΩq ĂĂ C 0 pΩq.

In particular we always have W 1,p pΩq ĂĂ Lp pΩq.

Proof of Theorem 2.38. ‚ We begin with the last case. By the extension Theorem 2.31, we
can see functions in W 1,p pΩq as functions in W 1,p pRd q supported in some fixed compact of
Rd . If p ă `8, the conclusion follows from the Morrey inequality (Theorem 2.33) and the
Ascoli-Arzelá Theorem 2.37. Since W 1,`8 pΩq is continuously embedded in W 1,p pΩq for any
p Psd, `8r, it is also compactly embedded in C 0 pΩq.
‚ Assume that (i) is proved and let q P rd, `8r. Then there exists p P r1, dr such that
q ă p˚ . Then we have
W 1,d pΩq Ă W 1,p pΩq ĂĂ Lq pΩq,
where the first inclusion is continuous (since Ω is bounded) and the second is compact by (i).
Thus it only remains to prove (i).
‚ Let q P r1, p˚ r. We consider a sequence pun qnPN bounded in W 1,p pΩq. As above, we
identify this sequence with a sequence (still denoted by pun qnPN ) bounded in W 1,p pRd q such
that the functions un are supported in the same şbounded open subset U . Let ρ P C08 pRd , R` q
be supported
` x ˘ in the unitε ball and such8 that Rd
ρ “ 1. For ε ą 0 and x P Rd we set
ρε “ εd ρ ε , and then un “ ρε ˚ un P C pR q. Let ε ą 0. For n P N and x P Rd we have
1 d

|uεn pxq| ď }ρε }L8 pRd q }un }L1 pU q

and
|∇uεn pxq| ď }∇ρε }L8 pRd q }un }L1 pU q ,

so the sequence puεn qnPN is bounded in C 0 pRd q and uniformly equicontinuous. Moreover
the functions uεn are supported in a common bounded set V of Rd , so by the Ascoli-Arzelá
Theorem 2.37 there exists a subsequence puεnk qkPN which converges uniformly in V and hence
in U . This gives › ›
› ›
lim sup ›uεnj ´ uεnk › “ 0.
j,kÑ`8 Lq pU q

‚ We already know that uεn goes to un as ε Ñ 0 in Lq pU q for all n P N. We prove that this
convergence is uniform with respect to n. Let v P C01 pRd q be supported in U . For ε ą 0 we
set vε “ ρε ˚ v. Then for x P Rd we have
ż ż ż1
` ˘
vε pxq ´ vpxq “ ρpyq vpx ´ εyq ´ vpxq dy “ ´ε ρpyq ∇vpx ´ εtyq ¨ y dt dy,
Bp0,1q Bp0,1q 0

and hence
ż ż ż1ż
}vε ´ v}L1 pU q “ |vε pxq ´ vpxq| dx ď ε ρpyq |∇vpx ´ εtyq| dx dt dy
U Bp0,1q 0 U (2.22)
ď ε }∇v}L1 pU q .

22 J. Royer - Université Toulouse 3

 Sobolev spaces

By density, the same estimate holds for any v P W 1,p pRd q supported in U (note that if
vm P C 1 pRd q goes to v in W 1,p pRd q then ρε ˚ vn goes to ρε ˚ v in L1 pRd q). Let θ Ps0, 1r be
such that
1 1´θ
“θ` ˚ .
q p
By (2.22) applied with v “ un and the Gagliardo-Nirenberg inquality (Theorem 2.34 there
exists C ą 0 independant on u, n or ε such that
θ 1´θ
}uεn ´ un }Lq pU q ď }uεn ´ un }L1 pU q }uεn ´ un }Lp˚ pRd q ď Cεθ }∇un }Lp pU q .

This proves that uεn goes to un in Lq pU q as ε Ñ 0 uniformly with respect to n P N. Then for
any η ą 0 we get
› ›
lim sup ›unj ´ unk ›Lq pU q ď η.
j,kÑ`8

Using a standard diagonal argument, we obtain a subsequence which goes to 0 in L q pU q and

hence in Lq pΩq.

Exercise 13. Let p P r1, dr. Prove that we have the continuous inclusion W 1,p pBp0, 1qq Ă
˚
Lp pBp0, 1qq, but that this inclusion is not compact.

2.5 Traces
We recall that functions in the Sobolev spaces are not really functions, but equivalence classes
of functions pairwise almost everywhere equal. In particular, for u in some Sobolev space
W k,p pΩq, it does not make sense to consider the value of u at some point x0 P Ω.

We have seen in Proposition 2.6 that, in dimension 1, an element u of W 1,p pIq has a
continuous reprentative ũ. It is reasonnable to consider ũpx0 q as the value of u at x0 . Indeed,
if ṽ is another representative of u then ṽpx0 q can be far from ũpx0 q, but for almost all x P I
“close to x0 ” then ṽpxq is equal to ũpxq and hence “close to ũpx0 q”.
However, this possible definition only works in dimension 1, since in higher dimension an
element of W 1,p pΩq does not necessarily have a continuous representative.

In applications, it is not crucial to give the value of a function at a point, but we are
interested in what happens at the boundary of the domain. This will be important for in-
stance for integration by parts (Green formula in higher dimension), where the value of the
function at the boundary appears. For regular domains, the boundary is a submanifold of
dimension pd ´ 1q. This is still of dimension 0 for the Lebesgue measure on Ω, but if d ě 2
this is in some sense “bigger” than a point.

Our purpose in this section is the following. Given a regular open subset Ω of Rd and
u P W 1,p pΩq, we want to give a natural sense to the restriction of u on the boundary BΩ, in
such a way that if u belongs to C 0 pΩq then the new definition coincides with the usual one.

2.5.1 Trace
As explained in the previous section, we begin our analysis with the model case Ω “ Rd` and
then, using a partition of unity and changes of variables, we will give a more general result.

Proposition 2.39. Let p P r1, `8r. There exists C ą 0 such that for u P C08 pRd` q we have
p p
}up0, ¨q}Lp pRd´1 q ď C }u}W 1,p pRd q .
`

For the proof we only have to integrate over Rd´1 the one-dimensional case which is very
close to Proposition 2.6:

2020-2021 23
M2RI - Elliptic PDE and evolution problems

Proof. For x1 P Rd´1 we have

ż `8
ˇ ˇ ˇ ˇˇ ˇ
ˇup0, x1 qˇp ď p ˇBx1 ups, x1 qˇ ˇups, x1 qˇp´1 ds
0

so, by the Hölder and Young inequalities,

ˆż `8 ˙ p1 ˆż `8 ˙ p´1
ˇ ˇ ˇ ˇ ˇ ˇ p
ˇup0, x1 qˇp ď p ˇBx1 ups, x1 qˇp ds ˇups, x1 qˇp ds
0 0
ż `8 ż `8
ˇ ˇ ˇ ˇ
ď ˇBx1 ups, x1 qˇp ds ` pp ´ 1q ˇups, x1 qˇp ds.
0 0
1
After integration over x P R we get
p p p
}up0, ¨q}Lp pRd´1 q ď pp ´ 1q }u}Lp pRd q ` }Bx1 u}Lp pRd q ,
`

and the conclusion follows.

Theorem 2.40. Let Ω be an open subset of Rd of class C 1 . Let p P r1, `8r. There is a
unique bounded linear operator

γ0 : W 1,p pΩq Ñ Lp pBΩq

such that
@u P W 1,p pΩq X C 0 pΩq, γ0 puq “ u|BΩ .
Proof. Let u P C08 pΩq. We use the notation of Paragraph 2.3.1. Let j P �1, N �. We have
ż ż
ˇ ˇ b
p
|u| dσ “ ˇupϕ̃px1 qqˇp 1 ` |∇ϕpx1 q|2 dx1
BΩXUj Oj
ż ż
ˇ ˇ ˇ ˇ
ď Cϕ ˇupϕ̃px1 qqˇp dx1 “ Cϕ ˇpu ˝ Φ´1 qˇp dx1 ,
Oj BRd
`

b
where Cϕ “ supx1 PO 1 ` |∇ϕpx1 q| and pu ˝ Φ´1 q has been extended by 0 on Rd` . By
2

Propositions 2.39 and 2.28 there exists Cj ą 0 independant of u such that

ż
› ›p
|u| dσ ď Cϕ C ›u ˝ Φ´1 ›W XRd ď Cj }uj }W 1,p pΩq .
p p
j `
BΩXUj

Then,
N ›
ÿ › ÿ
N
› › › ›
›u|BΩ › p ď ›u j › ď Cj }uj }W 1,p pΩq .
L pΩq |BΩ
Lp pΩq
j“1 j“1

Finally, there exists C ą 0 such that for all u P C08 pΩq we have
› ›
›u|BΩ › p ď C }u}W 1,p pΩq .
L pBΩq

Since C08 pΩq is dense in W 1,p pΩq, the map u P C08 pΩq ÞÑ u|BΩ P Lp pBΩq extends to a unique
continuous map on W 1,p pΩq. Moreover, if u P W 1,p pΩq X C 0 pΩq then the sequence pun qnPN
given by the proof of Proposition 2.32 goes uniformly to u and hence the restriction of un
goes to the restriction of u uniformly on BΩ, and hence in Lp pBΩq.
The following notation is motivated by Theorem 2.47 below:
Definition 2.41. When p “ 2 we denote by H 1{2 pBΩq the range of γ0 : H 1 pΩq Ñ L2 pΩq.
We do not discuss the properties of H 1{2 pBΩq here. However we will use in the following
chapter that even if γ0 is not surjective, H 1{2 pBΩq is dense in L2 pBΩq.
Proposition 2.42. Let Ω be an open subset of Rd of class C 1 . Let p P r1, `8r and u P
W 1,p pΩq. Then we have
γ0 puq “ 0 ðñ u P W01,p pΩq.

24 J. Royer - Université Toulouse 3

 Sobolev spaces

Proof. ‚ Assume that u P W01,p pRd q. Then there is a sequence pun qnPN in C08 pΩq going to
u in W 1,p pΩq. Since γ0 pun q “ 0 for all n P N and γ0 is continuous, we have γ0 puq “ 0.
‚ For the converse, we consider the case Ω “ Rd` and u supported in a bounded domain.
Then, with a partition of unity and changes of variables as above, we get the general case.
So let u P W 1,p pRd` q such that γ0 puq “ 0. Let pun qnPN be a sequence in C08 pRd` q which goes
to u in W 1,p pRd` q (see Proposition 2.32). Let n P N and x1 ą 0. For x1 P Rd´1 we have by
the Hölder inequality
ˆ ż x1 ˙p
ˇ ˇ ˇ ˇ ˇ ˇ
ˇun px1 , x1 qˇp ď ˇun p0, x1 qˇ ` ˇ∇un ps, x1 qˇ ds
0
ˆż x1 ˙p
ˇ
p´1 ˇ 1 ˇ
ˇp ˇ ˇ
1 ˇ
ď2 un p0, x q ` 2 p´1 ˇ∇un ps, x q ds
0
ż x1
ˇ ˇp ˇ ˇ
ď 2p´1 ˇun p0, x1 qˇ ` 2p´1 xp´1 ˇ∇un ps, x1 qˇp ds,
1
0

so for ε ą 0
żεż ż żε
ˇ ˇ ˇ ˇ
ˇun px1 , x1 qˇp dx1 dx1 ď 2p´1 ε }γ0 pun q}p p d ` 2p´1 εp ˇ∇un ps, x1 qˇp ds dx1 .
L pR q `
0 Rd´1 Rd´1 0

Taking the limit n Ñ 0 yields, by continuity of the trace,

p p
}u}Lp ps0,εrˆRd´1 q ď 2p´1 εp }∇u}Lp ps0,εrˆRd´1 q . (2.23)

Let χ P C 8 pR` , r0, 1sq, equal to 1 on r0, 1s and equal to 0 on r2, `8r. Then for n P N˚
and x “ px1 , . . . , xd q P Rd` we set χn pxq “ χpnx1 q. For n P N˚ we set un “ p1 ´ χn qu, so
that un P C08 pRd` q. By the dominated convergence theorem, we have

}un ´ u}Lp pRd q “ }χn u}Lp pRd q ÝÝÝÝÝÑ 0.

` ` nÑ`8

For n P N˚ we have
∇pun ´ uq “ p1 ´ χn q∇u ´ uB1 χn .
The first term goes to 0 in Lp pRd` q. For the second term we use (2.23) to write
ż 2 ż
n ˇ 1 ˇ ˇ ˇ
p
}uB1 χn }Lp “n p ˇχ pnx1 qˇp ˇupx1 , x1 qˇp dx1 dx1
x1 “ n
1
x1 PRp´1
› 1›
ď2 2p´1 ›χ › }∇u}p p 2
8 L ps0, rˆRd´1 q
n

ÝÝÝÝÝ
Ñ 0.
nÑ`8

This proves that

}un ´ u}W 1,p pRd q ÝÝÝÝÝÑ 0,
` nÑ`8

and hence u P W01,p pRd` q.

Exercise 14. Find an open domain Ω and u P W 1,8 pΩq such that u|BΩ “ 0 but u is not in
the closure of C08 pΩq in W 1,`8 pΩq.

2.5.2 Normal derivative

Let Ω be a bounded open subset of class C 1 in Rd . For the rest of this section we only
consider the case p “ 2.

Let u P H 2 pΩq. For j P �1, d� the derivative Bj u belongs to H 1 pΩq and hence has a trace
on BΩ. Then we set
ÿd
γ1 puq “ Bν u “ γ0 pBj uqνj P L2 pBΩq,
j“1

2020-2021 25
M2RI - Elliptic PDE and evolution problems

where ν “ pν1 , . . . , νd q is the outward normal derivative (see (2.14)). Notice that if u belongs
to C 1 pΩq then on BΩ we have
Bν u “ ∇u ¨ ν.
This defines a continuous function γ1 from H 2 pΩq to L2 pΩq. We can prove (see Theorem
2.47 below for the case Ω “ Rd` ) that
� (
Bν u, u P H 2 pΩq “ H 1{2 pΩq.

2.5.3 Green Formula

As said above, one of the motivations for the definition of the traces is the generalization
of the Green Formula to functions which are not regular in the usual sense. The following
results are deduced from the regular analogsş by density of regular
ş functions andş continuity
of the traces.
ş For u P W 1,1
pΩq we can write BΩ
u dσ instead of γ puq dσ and BΩ Bν u v dσ
BΩ 0
instead of BΩ γ1 puqγ0 pvq dσ.
Theorem 2.43. Let u, v P H 1 pΩq. Then for j P �1, d� we have
ż ż ż
u Bj v dx “ uv dσ ´ Bj u v dx
Ω BΩ Ω

Theorem 2.44. Let u P H pΩq and v P H pΩq. Then we have

2 1

ż ż ż
´ Δu v dx “ ´ Bν u v dσ ` ∇u ¨ ∇v dx.
Ω BΩ Ω

2.5.4 Appendix
In this additional paragraph we continue the discussion about traces and the Green formula.
In particular we define, via the Green Formula, a normal derivative for functions which are
not in H 2 pΩq.

We have denoted by H 1{2 pΩq Ă L2 pΩq the range of the trace γ0 defined on H 1 pΩq. This
is a vector space, which can be endowed with the following norm.

}g}H 1{2 pBΩq “ inf }w}H 1 pΩq .

wPH 1
γ0 pwq“g
� (
We notice that Hg1 pΩq “ w P H 1 pΩq : γ0 pwq “ g is a nonempty (by definition of H 1{2 pBΩq)
and closed (since γ0 is continuous) affine subspace (since γ0 is linear) of the Hilbert space
H 1 pΩq, so by the Hilbert projection theorem there exists a unique Rpgq P Hg1 pΩq such that

}g}H 1{2 pBΩq “ }Rpgq}H 1 pΩq .

Moreover Rpgq is the only solution in Hg1 pΩq of

@v P H01 pΩq, �Rpgq, v�H 1 pΩq “ 0.

From this we can deduce that the application which maps g P H 1{2 pΩq to Rpgq P H 1 pΩq is
linear, and then that H 1{2 pBΩq is a Banach space:
Proposition 2.45. H 1{2 pBΩq is a Banach space.
Proof. Let pgn qnPN be a Cauchy sequence in H 1{2 pΩq. Then pRpgn qqnPN is a Cauchy sequence
in H 1 pΩq. Since H 1 pΩq is complete, Rpgn q tends to some w in H 1 pΩq. We set g “ γ0 pwq P
H 1{2 pBΩq. Then we have

}gn ´ g}H 1{2 pBΩq “ }Rpg ´ gn q}H 1 pΩq “ }Rpgq ´ Rpgn q}H 1 pΩq ÝÝÝÝÝÑ 0.
nÑ`8

This proves that the sequence pgn qnPN has a limit in H 1{2 pBΩq, and hence that H 1{2 pBΩq is
complete.

26 J. Royer - Université Toulouse 3

 Sobolev spaces

We denote by H ´1{2 pBΩq the dual of H 1{2 pBΩq.

Proposition 2.46. Let u P H 1 pΩq such that Δu P L2 pΩq. Then the map
ż
` ˘
g P H 1{2 pΩq ÞÑ Δu vg ` ∇u ¨ ∇vg dx, (2.24)
Ω

where vg P H 1 pΩq satisfies γ0 pvg q “ g is well defined (the definition does not depend on the
choice of vg ) and defines a continuous linear map on H 1{2 pBΩq which we denote by Bν u.

We recall that in a general domain Ω the assumptions that u P H 1 pΩq and Δu P L2 pΩq
do not imply that u P H 2 pΩq.

Proof. We first observe that if v1 and v2 in H 1 pΩq are such that γ0 pw1 q “ γ0 pv2 q “ g then
v1 ´ v2 belongs to H01 pΩq, so there exists a sequence pφn qnPN in C08 pΩq which goes to v1 ´ v2
in H 1 pΩq. For all n P N we have
ż
` ˘
Δu φn ` ∇u ¨ ∇φn dx “ �Δu, φn �D1 pΩq,DpΩq ` �∇u, ∇φn �D1 pΩq,DpΩq “ 0,
Ω

so, taking the limit n Ñ `8,

ż ż
` ˘ ` ˘
Δu v1 ` ∇u ¨ ∇v1 dx “ Δu v2 ` ∇u ¨ ∇v2 dx.
Ω Ω

This proves that the definition in (2.24) does not depend on the choice of vg , and the map
Bν u is well-defined on H 1{2 pBΩq.
For g P H 1{2 pBΩq we have
ˇż ˇ
ˇ ` ˘ ˇ ´ ¯
ˇ Δu vg ` ∇u ¨ ∇vg dxˇˇ ď }Δu}L2 pΩq ` }∇u}L2 pΩq }vg }H 1 pΩq ,
ˇ
Ω

and hence
ˇż ˇ
ˇ ` ˘ ˇ ´ ¯
ˇ ` ¨ ˇ ď }Δu} 2
ˇ Δu v g ∇u ∇v g dx ˇ L pΩq ` }∇u}L2 pΩq }g}H 1{2 pBΩq .
Ω

This proves that the map Bν u is continuous on H 1{2 pBΩq. Since it is also linear, this defines
an element of H ´1{2 pBΩq.

By definition, we have the following Green formula for u, v P H 1 pΩq such that Δu P L2 :
ż ż
´ Δu v dx “ ´ �Bν u, v�H ´1{2 pBΩq,H 1{2 pBΩq ` ∇u ¨ ∇v. (2.25)
Ω Ω

We finish this section about traces by giving a general result on Ω “ Rd` by means of
the Fourier transform. This will in particular ensure that the two definitions of H 1{2 on
Rd´1 » BRd` are equivalent, and that the trace on H 1 pΩq and the normal trace on H 2 pΩq
have the same range.

Theorem 2.47. Let k P N and s ą k ` 12 . Then the map

"
SpRd q Ñ SpRd´1 q
u ÞÑ B1k up0, ¨q
1
has a unique continuous extansion γk : H s pRd q Ñ H s´k´ 2 pRd´1 q. Moreover, γk is surjective
1
and there exists a continuous linear map Rk : H s´k´ 2 pRd´1 q Ñ H s pRd q such that

γk ˝ Rk “ Id 1 .
H s´k´ 2 pRd´1 q

2020-2021 27
M2RI - Elliptic PDE and evolution problems

Proof. ‚ We first observe that for m P N, η ą 0 and σ ą m`1 we have, with the change of
? 2
variable t “ ηθ
ż ż
` ˘´σ m`1
tm η ` t2 dt “ η 2 ´σ Cm,σ , where Cm,σ “ θm p1 ` θ2 q´σ dθ. (2.26)
R R

‚ Let φ P SpRd q. For x1 P Rd´1 we have by the inversion formula

ż ˆż ˙
1 ix1 ¨ξ 1
B1k φp0, x1 q “ e piξ 1 qk
φ̂pξ 1 , ξ 1
q dξ 1 dξ 1 ,
p2πqd Rd´1 R

so the Fourier transform (in Rd´1 ) of B1k φp0, ¨q is given by

ż
1 1
g : ξ ÞÑ piξ1 qk φ̂pξ1 , ξ 1 q dξ1 . (2.27)
2π R

By the Cauchy-Schwarz inequality and (2.26) applied with η “ 1 ` |ξ 1 |2 we have, for all
ξ 1 P Rd´1 ,
ˆż ˙ ˆż ˙
2ˇ
ˇ 1 ˇ2 1 2
` ˘
1 2 s
` ˘
1 2 ´s
4π gpξ q ďˇ 2
|φ̂pξ1 , ξ q| 1 ` ξ1 ` |ξ | dξ1 2k 2
ξ1 1 ` ξ1 ` |ξ | dξ1
R
ż
` ˘´ps´k´ 12 q ` ˘s
ď C2k,s 1 ` |ξ 1 |2 |φ̂pξ1 , ξ 1 q|2 1 ` ξ12 ` |ξ 1 |2 dξ1 .

` ˘s´k´ 12
Multiplying by 1 ` |ξ 1 |2 and integrating over ξ 1 P Rd´1 gives
› k ›
›pB1 φqp0, ¨q›2 s´k´ 1 d´1 ď C2k,s }φ}2 s d .
2 pR q H pR q
H 4π 2
This proves the first statement of the theorem.
‚ Now we prove that γk is surjective with a continuous right inverse. We begin with
v P SpRd´1 q. Let g P SpRd´1 q be the Fourier transform of v on Rd´1 . The expression (2.27)
suggests to find f such that
ż
1 1
gpξ q “ piξ1 qk f pξ1 , ξ 1 q dξ1 . (2.28)
2π R
` ˘
Let N ą 12 s ´ k ´ 12 . For ξ “ pξ1 , ξ 1 q P Rd we set
` ˘N
2π p´iqk 1 ` |ξ 1 |2 1
f pξq “ ` 1 gpξ q.
2 ˘N ` 2 ` 2
k
Ck,N ` 12 1 ` |ξ|

In particular, for all ξ 1 P Rd´1 the map ξ1 ÞÑ p´iξ1 qk f pξ1 , ξ 1 q is integrable on R and (2.28)
holds by (2.26). Moreover, by (2.26) again we have
ż
2
p1 ` |ξ|2 qs |f pξq| dξ
Rd
ż ˆż ˙
4π 2 ˇ 1 ˇ2
1 2 2N ˇ 2 ´p2N `k`1´sq
“ 2 p1 ` |ξ | q gpξ q ˇ p1 ` |ξ| q dξ1 dξ
Ck,N ` 1 Rd´1 R
2
ż
4π 2 C0,2N `k`1´s ˇ ˇ2 1 ˇ ˇ2
“ 2 p1 ` ˇξ 1 ˇ qs´k´ 2 ˇgpξ 1 qˇ dξ 1
Ck,N ` 1 Rd´1
2

Then if we denote by u the inverse Fourier transform of f we obtain that u P H s pRd q and

2 4π 2 C0,2N `k`1´s 2
}u}H s pRd q ď 2 }v} s´k´ 21 d´1 . (2.29)
Ck,N H pR q
` 1
2

Moreover (2.28) ensures that γk puq “ v. Thus we have defined a map Rk : SpRd´1 q Ñ H s pRd q
1
such that γk ˝ Rk “ Id. By (2.29), Rk extends to a continuous map from H s´k´ 2 pRd´1 q to
H pR q, and the proof is complete.
s d

28 J. Royer - Université Toulouse 3

 Sobolev spaces

2.6 Poincaré Inequality

In Theorem 2.34 we have given an estimate with the norm }∇u}Lp pRd q and not the full norm
}u}W 1,p pRd q . In application, and in particular for the analysis of second order PDEs, we will
often be in the situation where we only control the norm of the gradient of the function and
not the function itself.
It turns out that in some particular situations, the norm of the function is in fact controled
by the norm of the gradient:

}u}Lp pΩq ď C }∇u}Lp pΩq . (2.30)

In this case, }∇u}Lp pΩq defines a norm on W 1,p pΩq, equivalent to }u}W 1,p pΩq . An inequality
like (2.30) is called a Poincaré inequality. This is the subject of this paragraph.

Before giving precise statements, we notice that a Poincaré inquality cannot hold in a
space which contains constant functions. In an unbounded domain, troubles can come from
slowly varying functions. For instance on R we consider for n P N˚ the function un defined
by #
1 ´ |x| if |x| ď n,
un pxq “ n
0 if |x| ą n.
2
and }u1 }L2 pRq “
2
Then we have }u}L2 pRq “ 2n
3 n.
2
A Poincaré inequality cannot hold in H 1 pRq.

In fact, we have discussed all the problems to prove a Poincaré inequality. Roughly speak-
ing, on a bounded domain, and if we remove constant functions, a Poincaré inquality holds.
The first way to remove constant functions is to consider only functions vanishing at the
boundary.

We first recall that Lemma 2.7 also holds in higher dimension.

Proposition 2.48. Let Ω be an open connected subset of Rd . Let u P L1loc pΩq be such that
∇u “ 0 (in the sense of distributions). Then there exists a constant α such that u “ α almost
everywhere.
Proof. We proceed by induction on the dimension. The case d “ 1 is already known. We
assume that d ě 2 and that the result is knownśd up to the dimension d ´ 1.
It is enough to consider the case Ω “ j“1 saj , bj r. Let χ P C08 psa1 , b1 rq be such that
ş b1 śd
a1
χpx1 q dx1 “ 1. For x1 P Ω1 “ j“2 saj , bj r we set
ż b1
vpx1 q “ upx1 , x1 qχpx1 q dx1 .
a1

This defines a function v P L1loc pΩ1 q. For ψ P C08 pΩ1 q and j P �2, d� we have
ż ż
´ vpx1 qBj ψpx1 q dx1 “ ´ upx1 , x1 qχpx1 qBj ψpx1 q dx1 dx1
Ω1
żΩ
` ˘
“ ´ upx1 , x1 qBj χpx1 qψpx1 q dx1 dx1
Ω
“ 0.

This proves that, in the sense of distributions, we have ∇v “ 0 on Ω1 By the induction

assumption there exists α such that v “ α almost everywhere on Ω1 .
Now let φ P C08 pΩq. For x “ px1 , x1 q P Ω we set
ż b1
φ̃px1 q “ φpx1 , x1 q dx1
a1

and ż x1
` ˘
ζpxq “ φpt, x1 q ´ χptqφ̃px1 q dt.
a1

2020-2021 29
M2RI - Elliptic PDE and evolution problems

Then ζ P C08 pΩq and φ “ Bx1 ζ ` χ b φ̃, so

ż ż b1 ż ż ż ż
uφ dx “ upx1 , x1 qχpx1 qφ̃px1 q dx1 dx1 “ v φ̃ dx1 “ α φ̃ dx1 “ α φ dx.
Ω a1 Ω1 Ω1 Ω1 Ω

This proves that u “ α almost everywhere on Ω.

Now we can prove the Poincaré inequality.
Theorem 2.49. Let Ω be an open bounded subset of Rd . Let p P r1, `8r. Then there exists
C ą 0 such that
@u P W01,p pΩq, }u}Lp pΩq ď C }∇u}Lp pΩq .
Proof. Assume by contradiction that the statement is not true. Then for all n P N there
exists un P W01,p pΩq such that

}un }Lp pΩq ą n }∇un }Lp pΩq .

Since this inequality can be divided by }un }Lp pΩq (which cannot be 0), we can assume without
loss of generality that }un }Lp pΩq “ 1 for all n P N. Then

}∇un }Lp pΩq ÝÝÝÝÝÑ 0, (2.31)

nÑ`8

and the sequence pun qnPN is bounded in W 1,p pΩq. Since W 1,p pΩq is compactly embedded in
Lp pΩq (see Theorem 2.38), there exists an increasing sequence pnk qkPN P NN and v P Lp pΩq
such that
}unk ´ v}Lp pΩq ÝÝÝÝÑ 0.
kÑ`8

With (2.31), this implies that the sequence punk qkPN is a Cauchy sequence in W 1,p pΩq. Since
W 1,p pΩq is complete (see Theorem 2.17), the sequence punk qkPN has a limit in W 1,p pΩq. This
limit is necessarily v. In particular v belongs to W 1,p pΩq, and by (2.31) we have ∇v “ 0.
By Proposition 2.48, v is constant on each connected component of Ω. Since unk belongs to
W01,p pΩq for all k P N, we also have v P W01,p pΩq, so v “ 0, which gives a contradiction with
the fact that }unk }L2 pΩq “ 1 for all k P N.
Notice that the proof of Theorem 2.49 does not give any clue about the constant C of
the inequality. We now give a similar result, with a more constructive proof. Moreover the
open set Ω is only required to be bounded in one direction. This means that Ω is included
in a strip of the form � (
Ω Ă x P Rd , x ¨ e Psa, br ,
for some e P Rd , |e| “ 1 and a, b P R.
Theorem 2.50 (Poincaré inequality). Let Ω be an open subset of R d , bounded in one direc-
tion. Let p P r1, `8r. Then there exists CΩ ą 0 such that, for all u P W01,p pΩq,

}u}Lp pΩq ď CΩ }∇u}Lp pΩq .

For instance, we can take CΩ “ pb ´ aqp.

Proof. ‚ It is enough to prove the estimate for u P C08 pΩq. Then the result will follow
by density of C08 pΩq in W01,p pΩq. We can extend u by 0, this gives a function in C08 pRd q
supported in Ω.
‚ We first consider the one-dimensional case. Then there exists a, b P R such that Ω Ăsa, br.
Then we can extend u as a function in C08 psa, brq which vanishes outside Ω. Then for all
x Psa, br we have pupxqp q1 “ pu1 pxqupxqp´1 so, by the Hölder inequality
żb
ˇ ˇ p´1 › ›
|upxq| ď p ˇu1 psqˇ |upsq| ds ď p }u}Lp pΩq ›u1 ›Lp pΩq .
p p´1

After integration over sa, br this gives

p´1 › ›
}u}Lp pΩq ď pb ´ aqp }u}Lp pΩq ›u1 ›Lp pΩq ,
p

30 J. Royer - Université Toulouse 3

 Sobolev spaces

p´1
and the conclusion follows after simplification by }u}Lp pΩq .
‚ Now we consider the general case. Let pf1 , . . . , fd q be an orthonormal basis of Rd such
that � (
supppuq Ă y1 f1 ` y 1 f 1 : y1 Psa, br, y 1 P Rd´1 ,
řd
for some a, b P R, where for y 1 “ py2 , . . . , yd q P Rd´1 we have set y 1 f 1 “ j“2 yj fj . By a
change of variables and using the one-dimensional case we can write
ż ż żb
ˇ ˇ
p
|upxq| dx “ ˇupy1 f1 ` y 1 f 1 qˇp dy1 dy 1
Ω y 1 PRd´1 y1 “a
ż żb ˇ ˇp
ˇ B ˇ
1 1 ˇ 1
ď ppb ´ aqpq p ˇ
ˇ By1 upy1 f1 ` y f qˇ dy1 dy
y 1 PRd´1 y1 “a
ż żb
ˇ ˇ
ď ppb ´ aqpqp ˇ∇upy1 f1 ` y 1 f 1 qˇp dy1 dy 1
y 1 PRd´1 y1 “a
ż
p
ď ppb ´ aqpqp |∇upxq| dx.
Ω

The conclusion follows with C “ pb ´ aqp.

It can be important in application to have an explicit constant for the Poincaré inequality.
Computing the optimal constant for particular sets Ω requires more work, and we do not
discuss this issue here, but we already have an upper bound.
After Theorem 2.50, the interest of the proof given for Theorem 2.49 is not clear. For
the proof of Theorem 2.50 we have really used the fact that the function u vanishes at the
boundary. While for the proof of Theorem 2.49 we have in fact only used the property that
the only constant function is 0. The interest of the proof of Theorem 2.49 is that it can
be used in any such situation. For instance, we give the following version of the Poincaré
inequality.

For a bounded open subset Ω we define

" ż *
Ă 1,p 1,p
W pΩq “ u P W pΩq : u dx “ 0 . (2.32)
Ω

Ă 1,p pΩq such that ∇u “ 0 is u “ 0.

Notice that if Ω is connected then the only function u P W

Theorem 2.51 (Poincaré-Wirtinger inequality). Let Ω be an open, connected and bounded

Ă 1,p pΩq,
subset of Rd . Let p P r1, `8s Then there exists C ą 0 such that, for all u P W

Ă 1,p pΩq,
@u P W }u}Lp pΩq ď C }∇u}Lp pΩq .

Exercise 15. Let Ω be an open, bounded and connected subset of Rd . Let p P r1, `8s.
1. Prove Theorem 2.51.
2. For u P W 1,p pΩq we set ˇż ˇ
ˇ ˇ
N puq “ }∇u}Lp pΩq ` ˇˇ u dxˇˇ .
Ω

Prove that N is a norm on W 1,p

pΩq, equivalent to the usual one.

2.7 The dual of H01 pΩq

Definition 2.52. We denote by H ´1 pΩq the dual space of H01 pΩq.

We recall that the dual space of H01 pΩq is the set of continuous linear forms on H01 pΩq. It
is endowed with the norm defined by

|ϕpuq|
}ϕ}H ´1 pΩq “ sup .
uPH01 pΩqzt0u }u}H 1 pΩq
0

2020-2021 31
M2RI - Elliptic PDE and evolution problems

We usually write �ϕ, u� (or �ϕ, u�H ´1 pΩq,H 1 pΩq ) instead of ϕpuq. Notice that if H 1 pΩq ‰ H01 pΩq
0
then H ´1 pΩq is not the dual space of H 1 pΩq.

We recall that by the Riesz Theorem, we can identify a Hilbert space with its dual.
However, in this kind of context we usually already identify L2 pΩq with its dual. With this
identification we have
H01 pΩq Ă L2 pΩq Ă H ´1 pΩq,
with continuous injections. The first inclusion is clear by definition of the Sovolev space
H01 pΩq. Now a function u P L2 pΩq is identified with the continuous linear form on L2 pΩq
defined by
v ÞÑ �u, v�L2 pΩq . (2.33)
By restriction, this also defines a continuous linear form on H01 pΩq. In this sense, we can say
that u belongs to H ´1 pΩq. However, all the elements of H ´1 pΩq cannot be identified with a
function in L2 pΩq. For instance, on R, the Dirac distribution

δ : v ÞÑ vp0q (2.34)

defines a continuous linear for on H 1 pRq “ H01 pRq, and it is not of the form (2.33) (notice
that this example is specific to the dimension 1, a Dirac distribution is not in H ´1 pΩq in
dimension d ě 2, however with the trace Theorem we can generalize this example in higher
dimension, see Exercise 16).

Let f P L2 pΩq and F P L2 pΩ, Rd q. Then ϕ “ f ´ div F , where the derivatives are
understood in the sense of distributions, also defines a continuous linear form on H01 pΩq
(which is not necessarily in L2 pΩq). For v P H01 pΩq it is given by

ÿ
d
ϕpvq “ �f, v� ` �Fj , Bj u� .
j“1

In particular we have
ÿ
d
}ϕ}H ´1 pΩq ď }f }L2 pΩq ` }Fj }L2 pΩq . (2.35)
j“1

In fact, using the Riesz Theorem in H01 pΩq we see that any ϕ P H ´1 pΩq can be written in
this form with u P H01 pΩq and F “ ∇u. Moreover, in this case we have an equality in (2.35).
See Theorem 5.9.1 in [Evans] (see Exercise 17 for the particular case of the Dirac distribution
(2.34)).

Exercise 16. Let f P L2 pRq. Prove that the map

ż
8 2
v P C0 pR q ÞÑ f pxqvpx, 0q dx
R

extends to a continuous linear form on H 1 pR2 q.

Exercise 17. Find u P H 1 pRq such that
ż ż
1
@v P H pRq, vp0q “ uv ` u1 v 1 .
R R

32 J. Royer - Université Toulouse 3

 Sobolev spaces

2.8 Exercises
Exercise 18. Show that there is no continuous linear map γ : L2 pR˚` q Ñ R such that
γpuq “ 0 for all u P C 0 pr0, `8rq X L2 pR˚` q.
Exercise 19. For which values of k P N, p P r1, `8s and q P r1, `8s do we have W k,p pRd q Ă
Lq pRd q ?
Exercise 20. 1. Let u P H01 pR2` q. For px1 , x2 q P R2 we set
#
upx1 , x2 q if x1 ą 0,
ũpx1 , x2 q “
0 if x1 ď 0.

Prove that ũ P H 1 pR2 q and give an expression for the derivatives of ũ. In particular, what
can we say about }ũ}H 1 pR2 q ?
2. Let Ω be an open subset of R2 . Let u P H01 pΩq. Prove that the extension of u by 0 on Rd
belongs to H 1 pR2 q.
Exercise 21. We recall that for s P r0, 1s and u P L2 pRd q we have set
ż
2 ` 2 ˘s 2
}u}H s pRd q “ 1 ` |ξ| |ûpξq| dξ.
Rd

Then H s pRd q is the set of u P L2 pRd q such that }u}H s pRd q ă `8.
1. Check that H 0 pRd q “ L2 pRd q and that H 1 pRd q coincides with the space already defined,
with equivalent norm.
2. Let s Ps0, 1r. Prove that there exists C ą 0 such that for all u P SpRd q we have
ż ż 2 ż
|upxq ´ upyq| 2s 2
d`2s
dx dy “ C |ξ| |ûpξq| dξ.
yPRd xPRd |x ´ y| ξPRd

3. Deduce that the quantity

˜ ż ż ¸ 12
2
2 |upxq ´ upyq|
}u}L2 pRd q ` d`2s
dx dy
yPRd xPRd |x ´ y|

defines a norm on H s pRd q, equivalent to the norm defined above.

Exercise 22. In this exercise we prove that for u P H 1 pRd q (real valued) we have |u| P
H 1 pRd q, ∇u “ 0 almost everywhere on u´1 pt0uq and ∇ |u| “ signpuq∇u on u´1 pRd z t0uq.
1. Let G : R Ñ R be of class C 1 , globally Lipschitz and such that Gp0q “ 0.
a. Show that G1 is bounded on R.
b. Prove that G ˝ u P H 1 pRd q with ∇pG ˝ uq “ pG1 ˝ uq∇u.
2. For t P R we set
# #
1 if x ą 0, 1 if x ě 0,
H´ ptq “ and H` ptq “
0 if x ď 0, 0 if x ă 0.

For n P N˚ we set $
&1 if t ě n ,
’ 1

Hn ptq “ nt if 0 ď t ď 1
n,
’
%
0 if t ď 0.
şt
Then we set Vn ptq “ ´8 Hn psq ds.
a. Prove that pVn ˝ uq P H 1 pRd q with ∇pVn ˝ uq “ pHn ˝ uq∇u.
b. For t P R we set #
t if t ą 0,
gptq “
0 if t ă 0.
Prove that pg ˝ uq P H 1 pRd q with ∇pg ˝ uq “ pH´ ˝ uq∇u.
c. Prove that ∇pg ˝ uq “ pH` ˝ uq∇u.
d. Deduce that ∇u “ 0 almost everywhere on u´1 pt0uq.
3. Conclude.

2020-2021 33