Chapter

3
The variational
formulation of elliptic
PDEs

We now begin the theoretical study of elliptic partial differential equations and
boundary value problems. We will focus on one approach, which is called the
variational approach. There are other ways of solving elliptic problems. The varia-
tional approach is quite simple and well suited for a whole class of approximation
methods, as we will see later.

3.1 Model problems

Let us start with a few model problems. The simplest of all is a slight generalization
of the Poisson equation with a homogeneous Dirichlet boundary condition. Let
us be given a connected1 open Lipschitz subset Ω of Rd , a function c ∈ L∞ (Ω),
another function f ∈ L2 (Ω). We are looking for a function u : Ω̄ → R such that
�
−∆u + cu = f in Ω,
(3.1)
u = 0 on ∂ Ω.
We are going to transform the boundary value problem (3.1) into an entirely
different kind of problem that is amenable to an existence and uniqueness theory,
as well as the definition of approximation methods.
1 In the context of PDEs, all the open subsets considered will be connected, without further
mention. It is not that important, but PDEs set on open sets with several connected components are
basically several unrelated PDEs.
76 3. The variational formulation of elliptic PDEs

Proposition 3.1.1 Assume that u ∈ H 2 (Ω) solves the PDE in problem (3.1). Then,
for all v ∈ H01 (Ω)
� � �
∇u · ∇v dx + cuv dx = f v dx. (3.2)
Ω Ω Ω

Proof. We take an arbitrary v ∈ H01 (Ω), multiply the equation by v,

−(∆u)v + cuv = f v,

and integrate over Ω. Every term is integrable since u ∈ H 2 (Ω) hence ∆u ∈ L2 (Ω)
and v ∈ L2 (Ω) imply (∆u)v ∈ L1 (Ω), c ∈ L∞ (Ω), u ∈ L2 (Ω) and v ∈ L2 (Ω) imply
cuv ∈ L1 (Ω), and f ∈ L2 (Ω) implies f v ∈ L1 (Ω). We thus obtain
� � �
− (∆u)v dx + cuv dx = f v dx.
Ω Ω Ω

We now use Green’s formula (2.17), according to which

� � �
(∆u)v dx = − ∇u · ∇v dx + γ1 (u)γ0 (v) dΓ,
Ω Ω ∂Ω

and we conclude since v ∈ H01 (Ω) is equivalent to γ0 (v) = 0. �

Formulation (3.2) is called the variational formulation of problem (3.1). Ac-
tually, it is not entirely complete, since we have not yet decided in which space
to look for u. In fact, as we have seen in the previous section, the reasonable
way to impose the Dirichlet boundary condition is to require that u ∈ H01 (Ω). The
functions v are called test-functions.
Let us rewrite the variational formulation in a standard, abstract form. We let
V = H01 (Ω), it is a Hilbert space. Then we have a bilinear form on V ×V
�
a(u, v) = (∇u · ∇v + cuv) dx
Ω

and a linear form on V �

�(v) = f v dx.
Ω
The variational formulation then reads

∀ v ∈ V, a(u, v) = �(v), (3.3)

and we have shown that a solution of the boundary value problem with the addi-
tional regularity u ∈ H 2 (Ω) is a solution of the variational problem (3.3).
Now what about the reverse implication? Does a solution of the variational
problem solve the boundary value problem? The answer is basically yes, the two
problems are equivalent.
3.1. Model problems 77

Proposition 3.1.2 Assume that u ∈ V solves the variational problem (3.3). Then
we have
−∆u + cu = f in the sense of D � (Ω)
and ∆u ∈ L2 (Ω).
Proof. We have D(Ω) ⊂ H01 (Ω), therefore we can take v = ϕ ∈ D(Ω) as test-
function in (3.3). Let us examine each term separately.
First of all
� � d d d

Ω
∇u · ∇ϕ dx =
Ω i=1
∑ ∂iu∂iϕ dx = ∑ �∂iu, ∂iϕ� = ∑ −�∂iiu, ϕ� = −�∆u, ϕ�,
i=1 i=1
by definition of distributional derivatives. Similarly
� �
cuϕ dx = �cu, ϕ� and f ϕ dx = � f , ϕ�.
Ω Ω
Therefore, we have for all ϕ ∈ D(Ω)
�−∆u + cu − f , ϕ� = 0
or
−∆u + cu − f = 0 in the sense of D � (Ω)
and the PDE is satisfied in the sense of distributions The Dirichlet boundary
condition is also satisfied by the simple fact that u ∈ H01 (Ω), hence the boundary
value problem is solved.
To conclude, we note that ∆u = cu − f ∈ L2 (Ω). This also implies that the PDE
is satisfied almost everywhere. �

Remark 3.1.1 The two problems are thus equivalent, except for the fact that we
have assumed u ∈ H 2 (Ω) in one direction, and only recuperated ∆u ∈ L2 (Ω) in the
other.2 Actually, the assumption u ∈ H 2 (Ω) is somewhat artificial and made only
to make use of Green’s formula (2.17). It is possible to dispense with it with a little
more work, but that would take us too far.
It should be noted in any case, that if u ∈ H01 (Ω), ∆u ∈ L2 (Ω) and Ω is for
example of class C2 , then u ∈ H 2 (Ω). This is very profound result in elliptic
regularity theory, far beyond the scope of these notes. It is trivial in dimension one
though.
Of course, so far we have no indication that either problem has a solution. The
fact is that the variational formulation is significantly easier to treat, once the right
point of view is found. And the right point of view is an abstract point of view, as
is often the case. �
2 The
Laplacian is a specific linear combination of some of the second order derivatives. So it
being in L2 is a priori less than all individual second order derivatives, even those not appearing in
the Laplacian, being in L2 .
78 3. The variational formulation of elliptic PDEs

Before we start delving in the abstract, let us give a couple more model problems
of a different kind. First is a new boundary condition.
�
−∆u + cu = f in Ω,
∂u (3.4)
∂ n = g on ∂ Ω.

This is an example of a Neumann boundary condition. When g = 0, it is natu-

rally called a homogeneous Neumann boundary condition. In terms of modeling,
the Neumann condition is a flux condition. For instance, in the heat equilibrium
interpretation, the condition corresponds to an imposed heat flux through the bound-
ary, as opposed to the Dirichlet condition which imposes a temperature on the
boundary. The case g = 0 corresponds to perfect thermal insulation.
Let us derive the variational formulation informally. Assume first that u ∈ H 2 ,
take v ∈ H 1 (Ω), multiply, integrate and use Green’s formula to obtain
� � �
1
∀ v ∈ H (Ω), (∇u · ∇v + cuv) dx = f v dx + gγ0 (v) dΓ.
Ω Ω ∂Ω

Note the different test-function space and the additional boundary term in the
right-hand side.
The converse is more interesting. Let u ∈ H 2 (Ω) be a solution of the above
variational problem. Taking first v = ϕ ∈ D(Ω), we obtain

−∆u + cu = f in the sense of D � (Ω)

exactly as in the Dirichlet case. Of course, a test-function with compact support

does not see what happens on the boundary, and no information on the Neumann
condition is recovered. Thus, in a second step, we take v arbitrary in H 1 (Ω). By
Green’s formula again, we have
� � �
∇u · ∇v dx = − (∆u)v dx + γ1 (u)γ0 (v) dΓ.
Ω Ω ∂Ω

Recall that the normal trace γ1 (u) plays the role of the normal derivative. Since u
is a solution of the variational problem, it follows that
� � � �
(−∆u + cu)v dx + γ1 (u)γ0 (v) dΓ = f v dx + gγ0 (v) dΓ.
Ω ∂Ω Ω ∂Ω
� �
But we already know that Ω (−∆u + cu)v dx = Ω f v dx, hence we are left with
� �
γ1 (u)γ0 (v) dΓ = gγ0 (v) dΓ,
∂Ω ∂Ω

for all v ∈ H 1 (Ω). For simplicity, we assume here that g ∈ H 1/2 (∂ Ω), the image
of the trace γ0 and that Ω is smooth. Since u ∈ H 2 (Ω), it follows that γ1 (u) =
3.2. Abstract variational problems 79

∑di=1 γ0 (∂i u)ni ∈ H 1/2 (∂ Ω). Therefore, there exists v ∈ H 1 (Ω) such that γ0 (v) =
γ1 (u) − g. With this choice of v, we obtain
�
(γ1 (u) − g)2 dΓ = 0,
∂Ω
hence γ1 (u) = g which is the Neumann condition. �
The last hypotheses made are for brevity only. They are not at all necessary to
conclude. Another problem of interest is the non homogeneous Dirichlet problem.
�
−∆u + cu = f in Ω,
(3.5)
u = g on ∂ Ω,
with g ∈ H 1/2 (∂ Ω). This problem is reduced to the homogeneous problem by
taking a function G ∈ H 1 (Ω) such that γ0 (G) = g and setting U = u − G. Then
clearly U ∈ H01 (Ω) and −∆U + cU = −∆u + cu + ∆G − cG = f + ∆G − cG. Then
we just write the variational formulation of the homogeneous problem for U with
right-hand side F = f + ∆G − cG.
The Dirichlet and Neumann conditions can be mixed together, but not at the
same place on the boundary, yielding the so-called mixed problem. More precisely,
let Γ1 and Γ2 be two subsets of ∂ Ω such that Γ1 ∩ Γ2 = 0/ and Γ̄1 ∪ Γ̄2 = ∂ Ω. Then
the problem 
 −∆u + cu = f in Ω,
u = g1 on Γ1 , (3.6)
 ∂u
∂n = g 2 on Γ 2 .
The variational formulation for the mixed problem (in the case g1 = 0 for brevity,
if not follow the above route) is to let V = {v ∈ H 1 (Ω); γ0 (v) = 0 on Γ1 } and
� � �
∀ v ∈ V, (∇u · ∇v + cuv) dx = f v dx + g2 γ0 (v) dΓ,
Ω Ω Γ2
with u ∈ V .
Remark 3.1.2 An important rule of thumb to be remembered from the above
examples is that (homogeneous) Dirichlet conditions are taken into account in the
test-function space, whereas Neumann boundary conditions are taken into account
in the linear form via boundary integrals.

3.2 Abstract variational problems

We describe the general abstract framework for all variational problems. Let us
start with a quick review of Hilbert space theory. Let H be a real Hilbert space
with scalar product (·|·)H and norm � · �H . We first recall the Cauchy-Schwarz
inequality, which is really a hilbertian property.
80 3. The variational formulation of elliptic PDEs

Theorem 3.2.1 For all u, v ∈ H, we have

|(u, v)H | ≤ �u�H �v�H .

One of the most basic results in Hilbert space theory is the orthogonal projection
theorem, which we recall below.

Theorem 3.2.2 Let C be non empty, convex, closed subset of H. For all x ∈ H,
there exists a unique pC (x) ∈ C such that

�x − pC (x)�H = inf �x − y�H .

y∈C

The vector pC (x) is called the orthogonal projection of x on C. It is also character-

ized by the inequality

∀ y ∈ C, (x − pC (x)|y − pC (x))H ≤ 0.

In addition, if C is a closed vector subspace of H, then pC is a continuous linear

mapping from H to C which is also characterized by the equality

∀ y ∈ C, (x − pC (x)|y)H = 0.

pC (x)

y
x

Figure 1. The orthogonal projection on a closed convex C.

The orthogonal projection of x on C is thus the element of C closest to x and

the angle between x − pC (x) and y − pC (x) is larger than π2 . In particular, if x ∈ C,
then pC (x) = x. An important consequence of the last characterization in the case
of a closed vector subspace is that we can write H = C ⊕ C⊥ with continuous
orthogonal projections on each factor. Indeed, we have x = pC (x) + (x − pC (x)
with pC (x) ∈ C by construction and x − pC (x) ∈ C⊥ by the second characterization.
Hence H = C +C⊥ . To show that the sum is direct, it suffices to note that C ∩C⊥ =
{0} which is obvious since x ∈ C ∩C⊥ implies 0 = (x|x)H = �x�2H .
3.2. Abstract variational problems 81

E⊥ x

pE ⊥ (x)

pE (x)

Figure 2. The orthogonal projection on a closed vector subspace E.

Another important consequence is a characterization of dense subspaces.

Lemma 3.2.1 A vector subspace E of H is dense in H if and only of E ⊥ = {0}.

Proof. For any vector subspace, it is always true that E ⊥ = (Ē)⊥ . Let E be a dense
subspace, i.e., Ē = H. Then, of course E ⊥ = H ⊥ = {0}. Conversely, if E ⊥ = {0},
this implies that (Ē)⊥ = {0} and since H = (Ē)⊥ ⊕ Ē, it follows that Ē = H, and
E is dense in H. �
The Riesz theorem provides a canonical way of identifying a Hilbert space and
its dual.

Theorem 3.2.3 (Riesz) Let H be a Hilbert space and � an element of its dual H � .
There exists a unique u ∈ H such that

∀ v ∈ H, �(v) = (u|v)H .

Moreover
���H � = �u�H
and the linear mapping δ : H � → H, � �→ u, is an isometry.
82 3. The variational formulation of elliptic PDEs

Proof. If � = 0, then we set u = 0 to be the unique u in question. Let � �= 0. It is

thus a nonzero continuous linear form, hence its kernel ker � is a closed hyperplane
of H. Let us choose u0 ∈ (ker �)⊥ with �u0 �H = 1 (this is possible since ker �
is not dense). Since u0 ∈/ ker �, we have �(u0 ) �= 0 and for all v ∈ H, we can set
�(v)
w = v − �(u ) u0 so that
0

� �(v) � �(v)
�(w) = � v − u0 = �(v) − �(u0 ) = 0,
�(u0 ) �(u0 )
�(v)
and w ∈ ker �. Now writing v = �(u u0 + w and setting u = �(u0 )u0 ∈ (ker �)⊥ , we
0)
obtain � �(v) � �
�
(v|u)H = u0 �u + (w|u)H = �(v)(u0 |u0 )H = �(v),
�(u0 ) H
hence the existence of u.
For the uniqueness, assume u1 and u2 are two solutions, then for all v ∈ H,
(v|u1 − u2 )H = 0. This is true in particular for v = u1 − u2 , hence u1 = u2 .
The mapping δ is thus well defined and obviously linear. Finally, for the
isometry, we have on the one hand

���H � = sup |�(v)| = sup |(v|u)H | ≤ �v�H �u�H ≤ �u�H ,

�v�H ≤1 �v�H ≤1

by the Cauchy-Schwarz inequality. On the other hand, equality trivially holds

for � = 0, and for � �= 0, choosing v = �u�u H yields (v|u)H = �u�H with �v�H = 1,
hence the equality in this case too. �

Remark 3.2.1 The Riesz theorem shows that the dual of a Hilbert space is also a
Hilbert space for the scalar product (�1 |�2 )H � = (δ �1 |δ �2 )H , since it is not obvious
a priori that the dual norm is hilbertian. It is often used to identify H and H � via
the isometry δ or δ −1 . This identification is not systematic however. For example,
when we have two Hilbert spaces H and V such that V �→ H and V is dense in H,
the usual identification is to let

V �→ H = H � �→ V �

using the Riesz theorem for H, which is called the pivot space, but not for V . Such
is the case for H = L2 (Ω), V = H01 (Ω) in which case we have an identification
of V � as the space H −1 (Ω). Indeed, the scalar product used in the identification
of the pivot space and its dual is the duality bracket of an L2 function seen as a
distribution and a D test-function. This is not the case for any of the two scalar
products that H01 comes equipped with, and an identification using these scalar
products, which is also legitimate, does not yield a space of distributions. �
3.2. Abstract variational problems 83

We now come to abstract variational problems and how they are solved. This is
the Lax-Milgram theorem. This theorem is important, not because it is in any way
difficult, which it is not, but because it has a very wide range of applicability as we
will see later.

Theorem 3.2.4 (Lax-Milgram) Let V be a Hilbert space, a be a bilinear form

and � be a linear form. Assume that
i) The bilinear form a is continuous, i.e., there exists a constant M such that
|a(u, v)| ≤ M�u�V �v�V for all u, v ∈ V ,
ii) The bilinear form a is V -elliptic3 , i.e., there exists a constant α > 0 such
that a(v, v) ≥ α�v�V2 for all v ∈ V ,
iii) The linear form � is continuous, i.e., there exists a constant C such that
|�(v)| ≤ C�v�V for all v ∈ V .
There exists a unique u ∈ V that solves the abstract variational problem: Find
u ∈ V such that
∀ v ∈ V, a(u, v) = �(v). (3.7)

Proof. Let us start with the uniqueness. Let u1 and u2 be two solutions of prob-
lem (3.7). Since a is linear with respect to its first argument, it follows that
a(u1 − u2 , v) = 0 for all v ∈ V . In particular, for v = u1 − u2 , we obtain

0 = a(u1 − u2 , u1 − u2 ) ≥ α�u1 − u2 �V2 ,

so that �u1 − u2 �V = 0 since α > 0.

We next prove the existence of a solution. We first note that for all u ∈ V , the
mapping v �→ a(u, v) is linear (by bilinearity of a) and continuous (by i) continuity
of a). Therefore, there exists a unique element Au of V � such that a(u, v) =
�Au, v�V � ,V . Moreover, the bilinearity of a shows that the mapping A : V → V � thus
defined is linear. It is also continuous since for all v ∈ V , �v�V ≤ 1,

|�Au, v�V � ,V | = |a(u, v)| ≤ M�u�V �v�V ≤ M�u�V

so that
�Au�V � = sup |�Au, v�V � ,V | ≤ M�u�V .
�v�V ≤1

We rewrite the variational problem as: Find u ∈ V such that

∀ v ∈ V, �Au − �, v�V � ,V = 0

or
Au = �,
3 This condition is also sometimes called coerciveness.
84 3. The variational formulation of elliptic PDEs

and this is where the continuity of � is used

Thus, proving the existence is equivalent to showing that the mapping A is onto.
We do this in two independent steps: we show that im A is closed on the one hand
and that it is dense on the other hand.4 For the closedness of the image, we use
assuption ii) of V -ellipticity. Let �n be a sequence in im A such that �n → � in V � .
We want to show that � ∈ im A, which will imply that im A is closed. The sequence
�n is a Cauchy sequence in V � , and for all n, there exists un ∈ V such that Aun = �n .
By V -ellipticity,
1 1
�un − um �V2 ≤ a(un − um , un − um ) = �Aun − Aum , un − um �V � ,V
α α
1 1
= ��n − �m , un − um �V � ,V ≤ ��n − �m �V � �un − um �V ,
α α
by the definition of the dual norm. Therefore, if �un − um �V = 0 we are happy,
otherwise we divide by �un − um �V and in both cases
1
�un − um �V ≤ ��n − �m �V � ,
α
so that un is a Cauchy sequence in V . Since V is complete, there exists u ∈ V
such that un → u in V . Since A is continuous, it follows that �n = Aun → Au in V � .
Hence � = Au ∈ im A.
To show the density, we show that (im A)⊥ = {0}. We note that (Au|�)V � =
(δ Au|δ �)V = �Au, δ ��V � ,V = a(u, δ �). Let � ∈ (im A)⊥ . For all u ∈ V , we thus
have a(u, δ �) = 0. In particular, for u = δ �, we obtain 0 = a(δ �, δ �) ≥ α�δ ��V2 by
V -ellipticity. Since α > 0, it follows that � = 0. �

Remark 3.2.2 It should be noted that the Lax-Milgram theorem is not a particular
case of the Riesz theorem. It is actually more general, since it applies to bilinear
forms that are not necessarily symmetric, and it implies the Riesz theorem when
the bilinear form is just the scalar product.
Sometimes when the bilinear form a is symmetric, people think it advantageous
to apply Riesz’s theorem in place of the Lax-Milgram theorem. This is usually an
illusion: indeed, if a new scalar product defined by the bilinear form is introduced,
in order to apply Riesz’s theorem, it is necessary to show that the space equipped
with the new scalar product is still a Hilbert space, i.e., is complete. This is done
by V -ellipticity, hence nothing is gained (although this is the part that people who
think they are seeing a good deal usually forget). The continuity of the linear form
for the new norm must also be checked, which amounts to having the bilinear form
and linear form continuous for the original norm, again, no gain.
4 This is a pretty common strategy, to be kept in mind.
3.2. Abstract variational problems 85

The only case when Riesz’s theorem can be deemed advantageous over the
Lax-Milgram theorem, is when both above� facts to be checked are already known.
An example is the bilinear form a(u, v) = Ω ∇u · ∇v dx on V = H01 (Ω). �

Remark 3.2.3 In the case of complex Hilbert spaces and complex-valued varia-
tional problems, the Lax-Milgram theorem still holds true for a bilinear or sesquilin-
ear form. The V -ellipticity assumption can even be relaxed to only involve the real
part of a : ℜ(a(u, u)) ≥ α�u�2 (or the imaginary part), which is rather useful as
the imaginary part can then be pretty arbitrary (exercise). �

The linear form in the right-hand side of a variational problem should be

thought of as data. In this respect, the solution depends continuously on the data.
Proposition 3.2.1 The mapping V � → V , � �→ u defined by the Lax-Milgram theo-
rem is linear and continuous.
Proof. Let �1 , �2 be two linear forms and u1 , u2 the corresponding solutions to the
variational problem. For all λ ∈ R, we have for all v ∈ V ,

a(u1 + λ u2 , v) = a(u1 , v) + λ a(u2 , v) = �1 (v) + λ �2 (v) = (�1 + λ �2 )(v)

hence the linearity by uniqueness of the solution. For the continuity, we have

α�u�V2 ≤ a(u, u) = �(u) ≤ ���V � �u�V ,

hence
1
�u�V ≤ ���V � ,
α
and we have the continuity, with continuity constant α1 . �

Proposition 3.2.2 Let the hypotheses of the Lax-Milgram theorem be satisfied.

Assume in addition that the bilinear form a is symmetric. Then the solution u of the
variational problem (3.7) is also the unique solution of the minimization problem:
1
J(u) = inf J(v) with J(v) = a(v, v) − �(v). (3.8)
v∈V 2
Proof. Let u be the Lax-Milgram solution. For all v ∈ V , we let w = v − u and
1 1 1 1
J(v) = J(u + w) = a(u, u) + a(u, w) + a(w, u) + a(w, w) − �(u) − �(w)
2 2 2 2
1
= J(u) + a(u, w) − �(w) + a(w, w)
2
≥ J(u),
86 3. The variational formulation of elliptic PDEs

since a(w, w) ≥ 0. Make note of where the symmetry is used. Hence, u minimizes
J on V .
Conversely, assume that u minimizes J on V . Then, for all λ > 0 and all v ∈ V ,
we have J(u + λ v) ≥ J(u). Expanding the left-hand side, we get

1 λ2
a(u, u) + λ a(u, v) + a(v, v) − �(u) − λ �(v) ≥ J(u)
2 2
so that dividing by λ
λ
a(u, v) − l(v) + a(v, v) ≥ 0.
2
We then let λ → 0, hence
a(u, v) − l(v) ≥ 0,
and finally change v in −v to obtain

a(u, v) − l(v) = 0,

for all v ∈ V . �

Remark 3.2.4 Taking λ > 0, dividing by λ and then letting λ → 0 is quite clever,
and known as Minty’s trick. �

Remark 3.2.5 When the bilinear form a is not symmetric, we can still define the
functional J in the same fashion as before and try to minimize it. It is clear from the
above proof that the minimizing element u does not solve the variational problem
associated with a but the variational problem associated with the symmetric part
of a. Of course, when both variational problems are translated into PDEs, we get
entirely different equations. �

3.3 Application to the model problems, and more

We now apply the abstract results to concrete examples. We start with the first
model problem (3.1).

Proposition 3.3.1 Let Ω be a bounded open subset of Rn , f ∈ L2 (Ω), c ∈ L∞ (Ω).

Assume that c ≥ 0. Then the problem: Find u ∈ V = H01 (Ω) such that
� �
∀ v ∈ V, (∇u · ∇v + cuv) dx = f v dx,
Ω Ω

has one and only one solution.

3.3. Application to the model problems, and more 87

Proof. We already that V is a Hilbert space, for both scalar products that we defined
earlier. Of course �
a(u, v) = (∇u · ∇v + cuv) dx
Ω
clearly defines a bilinear form on V ×V and
�
�(v) = f v dx
Ω

a linear form on V . Hence, we have an abstract variational problem. Let us try

and apply the Lax-Milgram theorem. We need to check the hypotheses. For
definiteness, we choose to work with the full H 1 norm.
First of all, for all (u, v) ∈ V ×V ,
�� �
� �
|a(u, v)| = � (∇u · ∇v + cuv) dx�
�Ω
≤ |∇u · ∇v + cuv| dx
�Ω �
≤ |∇u · ∇v| dx + |cuv| dx
Ω Ω
≤ �∇u�L2 (Ω) �∇v�L2 (Ω) + �c�L∞ (Ω) �u�L2 (Ω) �v�L2 (Ω)
� �
≤ max 1, �c�L∞ (Ω) �u�H 1 (Ω) �v�H 1 (Ω) ,

by Cauchy-Schwarz to go from the third line to the fourth line, hence the continuity
of the bilinear form a.
Next is the V -ellipticity. For all v ∈ V , we have
� �
2 2
a(v, v) = (�∇v� + cv ) dx ≥ �∇v�2 dx ≥ α�v�2H 1 (Ω)
Ω Ω

with α = (C2 + 1)−1/2 > 0 by Corollary 2.6.3 where C is the Poincaré inequality
constant, and since c ≥ 0.
Finally, we check the continuity of the linear form. For all v ∈ V ,
�� �
� �
|�(v)| = � f v dx� ≤ � f �L2 (Ω) �v�L2 (Ω) ≤ � f �L2 (Ω) �v�H 1 (Ω)
Ω

by Cauchy-Schwarz again.
All the hypotheses of the Lax-Milgram theorem are satisfied, therefore there is
one and only one solution u ∈ V . �

Remark 3.3.1 Now is a time to celebrate since we have successfully solved our
first boundary value problem in arbitrary dimension. Indeed, we have already
88 3. The variational formulation of elliptic PDEs

seen that any solution of the variational problem is a solution of the PDE in the
distributional sense and in the L2 sense. The solution u depends continuously in
H 1 on f in L2 . Note that we have also solved the non homogeneous Dirichlet
problem at the same time. It is an instructive exercise to redo the proof using the
H 1 semi-norm in place of the full norm. The same ingredients are used, but not at
the same spots.
This is a case of a symmetric bilinear form, therefore the solution u also
minimizes the so-called energy functional
� �
1 2 2
J(v) = (�∇v� + cv ) dx − f v dx
2 Ω Ω

over V . �

Remark 3.3.2 It should be noted that the positivity condition c ≥ 0 is by no means

a necessary condition for existence and uniqueness via the Lax-Milgram theorem.
With a little more work, it is not too hard to allow the function c to take some
negative values. However, we have seen an example at the very beginning of these
notes with a negative function c for which existence and uniqueness fails.
One should also be aware that there is an existence and uniqueness theory that
goes beyond the Lax-Milgram theorem, which is only a sufficient condition for
existence and uniqueness. �

Let us now consider the non homogeneous Neumann problem (3.4). The
hypotheses are slightly different.

Proposition 3.3.2 Let Ω be a bounded open subset of Rn , f ∈ L2 (Ω), c ∈ L∞ (Ω),

g ∈ L2 (∂ Ω). Assume that there exists a constant η > 0 such that c ≥ η almost
everywhere. Then the problem: Find u ∈ V = H 1 (Ω) such that
� � �
∀ v ∈ V, (∇u · ∇v + cuv) dx = f v dx + gγ0 (v) dx,
Ω Ω ∂Ω

has one and only one solution.

Proof. We have a different Hilbert space (but known to be Hilbert, nothing to check
here), the same bilinear form and a different linear form
� �
�(v) = f v dx + gγ0 (v) dx.
Ω ∂Ω
3.3. Application to the model problems, and more 89

We have already shown that the bilinear form is continuous in the H 1 norm5 .
The V -ellipticity is clear since, for all v ∈ V ,
� � �
2 2 2
a(v, v) = (�∇v� + cv ) dx ≥ �∇v� dx + η v2 dx ≥ min(1, η)�v�2H 1 (Ω) ,
Ω Ω Ω

with min(1, η) > 0. The continuity of the linear form is also clear

|�(v)| ≤ � f �L2 (Ω) �v�L2 (Ω) + �g�L2 (∂ Ω) �γ0 (v)�L2 (∂ Ω)

≤ (� f �L2 (Ω) +C�g�L2 (∂ Ω) )�v�H 1 (Ω)

by Cauchy-Schwarz, where C is continuity constant of the trace mapping. �

The mixed problem (3.6) is practically entirely identical to the Neumann
problem.

Proposition 3.3.3 Same hypotheses as in Proposition 3.3.2 and let Γ1 and Γ2 be

two subsets of ∂ Ω such that Γ1 ∩ Γ2 = 0/ and Γ̄1 ∪ Γ̄2 = ∂ Ω. Then the problem:
Find u ∈ V = {v ∈ H 1 (Ω); γ0 (v) = 0 on Γ1 } such that
� � �
∀ v ∈ V, (∇u · ∇v + cuv) dx = f v dx + gγ0 (v) dΓ,
Ω Ω Γ2

has one and only one solution.

Proof. The only real difference with Proposition 3.3.2 lies with the space V , which
we do not know yet to be a Hilbert space. It suffices to show that V is a closed
subspace of H 1 (Ω). Let vn be a sequence in V such that vn → v in H 1 (Ω). By
continuity of the trace mapping, we have γ0 (vn ) → γ0 (v) in L2 (∂ Ω). Therefore,
there exists a subsequence γ0 (vn p ) that converges to γ0 (v) almost everywhere on
∂ Ω. Since γ0 (vn ) = 0 almost everywhere on Γ1 , it follows that γ0 (v) = 0 almost
everywhere on Γ1 , hence v ∈ V , which is thus closed. �
A natural question arises about the Neumann problem without a strictly positive
bound from below for the function c, in particular for c = 0. Now, this is an entirely
different problem from the previous ones. First we have to find the variational
formulation of the boundary value problem and show that it is equivalent to the
boundary value problem, then we have to apply the Lax-Milgram theorem.
Let us thus consider the Neumann problem
�
−∆u = f in Ω,
∂u (3.9)
∂ n = g on ∂ Ω,
5 If
we had worked with the semi-norm for the Dirichlet problem, we would have had to do the
continuity all over again here...
90 3. The variational formulation of elliptic PDEs

in a Lipschitz open set Ω in Rd . We see right away that things are going to be
different since we do not have uniqueness here. Indeed, if u is a solution, then u + s
is also a solution for any constant s. Furthermore, by Green’s formula (2.17) with
v = 1, it follows that if there is a solution, then, necessarily
� �
f dx + g dΓ = 0. (3.10)
Ω ∂Ω

If the data f , g does not satisfy the compatibility condition (3.10), there is thus no
solution. The two remarks, non uniqueness and non existence, are actually dual to
each other.
There are several ways of going around both problems, thus several variational
formulations6 . We choose to set,
� � �
1
V = v ∈ H (Ω); v dx = 0 .
Ω

This is well defined, since Ω is bounded and we thus have H 1 (Ω) ⊂ L2 (Ω) ⊂ L1 (Ω).
Note that V is the L2 -orthogonal to the one-dimensional space of constant functions,
which are the cause of non uniqueness.

Lemma 3.3.1 The space V is a Hilbert space for the scalar product of H 1 (Ω).

Proof. It suffices to show that V is closed. Let vn be a sequence in V such that

vn → v in H 1 (Ω). Of course, vn → v in L2 (Ω) and by Cauchy-Schwarz, vn → v in
L1 (Ω). Therefore � �
0= vn dx → v dx,
Ω Ω
and v ∈ V . �

Proposition 3.3.4 Assume that f ∈ L2 (Ω), g ∈ L�2 (∂ Ω) satisfy the compatibility

�
condition
�
(3.10). Then the triple V , a(u, v) = Ω ∇u · ∇v dx, �(v) = Ω f v dx +
gγ
∂Ω 0 (v) dΓ defines a variational formulation for the Neumann problem (3.9).

Proof. Multiplying the PDE by v ∈ V and using Green’s formula, we easily see
that if u solves problem (3.9), then we have for all v ∈ V , a(u, v) = �(v).
Conversely, let us be given a function u ∈ V such that for all v ∈ V , a(u, v) =
�(v). We would like to proceed as before and take v ∈ D(Ω) to deduce the PDE.
This does not work here because D(Ω) �⊂ V . For all ϕ ∈ D(Ω), we set
�
1
ψ =ϕ− ϕ(x) dx,
meas Ω Ω
6 We have always said the variational formulation, but there is no evidence that it is unique in
general.
3.3. Application to the model problems, and more 91

so that ψ ∈ V and �we can use ψ as a test-function. Now ϕ and ψ differ by a

1
constant k = meas Ω Ω ϕ(x) dx, therefore ∇ψ = ∇ϕ. We thus obtain,
� � �
∇u · ∇ϕ dx = f ψ dx + gψ dΓ
Ω �Ω ∂Ω �

= f (ϕ + k) dx + g(ϕ + k) dΓ
�Ω �� ∂Ω � � �
= f ϕ dx + k f dx + g dΓ = f ϕ dx,
Ω Ω ∂Ω Ω

since ϕ vanishes on ∂ Ω and f , g satisfy condition (3.10). So we can deduce right

away that −∆u = f in the sense of distributions, and since f ∈ L2 (Ω) in the sense
of L2 (Ω) as well.
We next pick an arbitrary v ∈ V and apply Green’s formula again. This yields
� � � �
f v dx + gγ0 (v) dΓ = − (∆u)v dx + γ1 (u)γ0 (v) dΓ.
Ω ∂Ω Ω ∂Ω

Hence, taking into account that −∆u = f , we obtain

�
(g − γ1 (u))γ0 (v) dΓ = 0
∂Ω

for all v ∈ V . Now� it is clear that γ0 (V ) = H 1/2 (∂ Ω). Indeed, �let is pick a
θ ∈ D(Ω) such that Ω θ dx = 1. Then, for all w ∈ H 1 (Ω), v = w − Ω w dxθ ∈ V
and γ0 (v) = γ0 (w). Therefore, there are enough test-functions in V to conclude that
γ1 (u) = g, since H 1/2 (∂ Ω) is dense in L2 (∂ Ω). �

Remark 3.3.3 We did not insist on the regularity needed to apply Green’s formula
or to define γ1 (u) as an element of H 1/2 , because it is possible to write down
slightly more complicated arguments that completely do away with such artificial
hypotheses. �

To apply the Lax-Milgram theorem, we need a new inequality.

Theorem 3.3.1 (Poincaré-Wirtinger inequality) Let Ω be a Lipschitz open sub-

set of Rd . There exists a constant C depending on Ω such that, for all v ∈ H 1 (Ω),
� � �
� 1 �
�v − v dx� ≤ C�∇v�L2 (Ω) . (3.11)
meas Ω Ω L2 (Ω)

Proof. We admit the Poincaré-Wirtinger inequality. �

92 3. The variational formulation of elliptic PDEs

Remark 3.3.4 Even though there is a certain similarity with the Poincaré inequal-
ity, there are major differences. In particular, the Poincaré-Wirtinger inequality
fails for open sets that are not regular enough (Lipschitz is sufficient for it to hold),
whereas no regularity is needed for Poincaré. Instead of proving it, let us just
note that the Poincaré-Wirtinger is at least reasonable, since both sides vanish for
constant functions v. �

Proposition 3.3.5 Let Ω be a Lipschitz open subset of Rd , f ∈ L2 (Ω) and g ∈

L2 (∂ Ω). Then the problem: Find u ∈ V such that
� � �
∀ v ∈ V, ∇u · ∇v dx = f v dx + gγ0 (v) dΓ,
Ω Ω ∂Ω

has one and only one solution.

Proof. We have already shown that V is a Hilbert space for the H 1 scalar product.
The continuity of both bilinear and linear forms have also already been proved.
Only the V -ellipticity remains.
�
For all v ∈ V , we have Ω v dx = 0, hence by the Poincaré-Wirtinger inequality
(3.11),
�v�2H 1 (Ω) = �v�2L2 (Ω) + �∇v�2L2 (Ω) ≤ (C2 + 1)�∇v�2L2 (Ω) .
Therefore,
a(v, v) = �∇v�2L2 (Ω) ≥ α�v�2H 1 (Ω)
1
with α = (C2 +1)
> 0. �

Remark 3.3.5 The compatibility condition (3.10) plays no role in the application
of the Lax-Milgram theorem. So exercise: What happens when it is not satisfied?
What exactly are we solving then?
Since the space V is a hyperplane of H 1 that is L2 orthogonal to the constants,
it follows that the general solution of the Neumann problem is of the form v + s,
where v ∈ V is the unique solution of the variational problem above and s ∈ R is
arbitrary. �

We now introduce a new kind of boundary condition, the Fourier condition

(also called the Robin condition or the third boundary condition). The boundary
value problem reads �
−∆u + cu = f in Ω,
(3.12)
bu + ∂∂ un = g on ∂ Ω,
where b and c are given functions. When b = 0, we recognize the Neumann
problem (and, in a sense, when b = +∞ the Dirichlet problem). This condition
3.3. Application to the model problems, and more 93

is called after Fourier who introduced it in the context of the heat equation. In
the heat interpretation, ∂∂ un represents the heat flux through the boundary. Let us
assume that we are modeling a situation in which the boundary is actually a very
thin wall that insulates Ω from the outside where the temperature is 0◦ . If g = 0,
the Fourier condition states that ∂∂ un = −bu, that is to say that the heat flux passing
through the wall is proportional to the temperature difference between the inside
and the outside. For this interpretation to be physically reasonable, it is clearly
necessary that b ≥ 0, i.e., the heat flows inwards when the outside is warmer than
the inside and conversely. It thus to be expected that the sign of b will play a role.
We follow the same pattern as before: First find a variational formulation for
the boundary value problem (3.12), second apply Lax-Milgram to prove existence
and uniqueness.
Proposition 3.3.6 Assume that f ∈ L2 (Ω), g ∈ L2 (∂ Ω), c ∈ L∞ (Ω) and b ∈ L∞ (∂ Ω).
Then the triple
V = H 1 (Ω),
� �
a(u, v) = (∇u · ∇v + cuv) dx + bγ0 (u)γ0 (v) dΓ,
�Ω � ∂Ω

�(v) = f v dx + gγ0 (v) dΓ,

Ω ∂Ω
defines a variational formulation for the Fourier problem (3.12).
Proof. As always, we multiply the PDE by v ∈ V and use Green’s formula,
� � �
(∇u · ∇v + cuv) dx = f v dx + γ1 (u)γ0 (v) dΓ
Ω �Ω �∂ Ω
= f v dx + (g − bγ0 (u))γ0 (v) dΓ,
Ω ∂Ω
hence
� � � �
(∇u · ∇v + cuv) dx + bγ0 (u)γ0 (v) dΓ = f v dx + gγ0 (v) dΓ, (3.13)
Ω ∂Ω Ω ∂Ω
for all v ∈ V .
Conversely, let us be given a solution u of the variational problem (3.13). Taking
first v = ϕ ∈ D(Ω), all the boundary integrals vanish and we obtain −∆u + cu = f
exactly as before. Taking then v arbitrary, using Green’s formula and the PDE just
obtained, we get
� � �
γ1 (u)γ0 (v) dΓ + bγ0 (u)γ0 (v) dΓ = gγ0 (v) dΓ,
∂Ω ∂Ω ∂Ω
so that �
(γ1 (u) + bγ0 (u) − g)γ0 (v) dΓ = 0,
∂Ω
for all v ∈ V = H 1 (Ω), hence the Fourier boundary condition. �
94 3. The variational formulation of elliptic PDEs

Remark 3.3.6 A natural question to ask is why not keep the term γ1 (u) in the
bilinear form? The answer is that, while it is true that γ1 (u) exists when u is a
solution of either the boundary value problem or the variational problem, it does
not exist for a general v ∈ H 1 (Ω), hence cannot appear in a bilinear form that is
defined on H 1 (Ω) × H 1 (Ω). Besides, how would b appear otherwise? �

Let us give a first existence and uniqueness result.

Proposition 3.3.7 Let Ω be a Lipschitz open subset of Rd , f ∈ L2 (Ω), g ∈ L2 (∂ Ω),

c ∈ L∞ (Ω) and b ∈ L∞ (∂ Ω). Assume that c ≥ η > 0 for some constant η and that
�b− �L∞ (∂ Ω) < min(1,η)
C2
, where Cγ0 is the continuity constant of the trace mapping.
γ0
Then the problem: Find u ∈ V such that
� � � �
∀ v ∈ V, (∇u · ∇v + cuv) dx + bγ0 (u)γ0 (v) dΓ = f v dx + gγ0 (v) dΓ,
Ω ∂Ω Ω ∂Ω

has one and only one solution.

Here b− = − min(0, b) denotes the negative part of b.

Proof. We check the hypotheses of the Lax-Milgram theorem. We already know
that V is a Hilbert space. The continuity of the bilinear form a has also already
been checked, except for the boundary integral terms
�� �
� �
� bγ0 (u)γ0 (v) dΓ� ≤ �b�L∞ (∂ Ω) �γ0 (u)�L2 (∂ Ω) �γ0 (v)�L2 (∂ Ω)
∂Ω
≤ Cγ20 �b�L∞ (∂ Ω) �u�H 1 (Ω) �v�H 1 (Ω)

for all u and v. The linear form is also known to be continuous. Let us check the
V -ellipticity. Obviously b ≥ −b− , thus
� �
2 2
(�∇v� + cv ) dx + bγ0 (v)2 dΓ
Ω ∂Ω
≥ min(1, η)�v�2H 1 (Ω) − �b− �L∞ (∂ Ω) �γ0 (v)�2L2 (∂ Ω)
� �
≥ min(1, η) −Cγ20 �b− �L∞ (∂ Ω) �v�2H 1 (Ω) ,

hence the V -ellipticity. �

Remark 3.3.7 Under the previous hypotheses, we have existence and uniqueness
via Lax-Milgram provided b is not too negative in some sense. �
3.3. Application to the model problems, and more 95

All these hypotheses only give sufficient conditions. Let us give another set of
such hypotheses.
Proposition 3.3.8 Same hypotheses except that we assume that c ≥ 0 and that
b ≥ µ > 0 for some constant µ. Then the Fourier problem has one and only one
solution.
Proof. The only point to be established is V -ellipticity. We use a compactness
argument by contradiction. For this we admit that Rellich’s theorem 2.7.3 is also
true in dimension d in a Lipschitz open set. We have
� � � �
2 2 2 2
(�∇v� + cv ) dx + bγ0 (v) dΓ ≥ �∇v� dx + µ γ0 (v)2 dΓ.
Ω ∂Ω Ω ∂Ω

Let us assume for contradiction that there is no constant α > 0 such that
� �
2
�∇v� dx + µ γ0 (v)2 dΓ ≥ α�v�2H 1 (Ω) .
Ω ∂Ω

This implies that for all n ∈ N∗ , there exists vn ∈ H 1 (Ω) such that
� �
1
�∇vn �2 dx + µ γ0 (vn )2 dΓ < �vn �2H 1 (Ω) .
Ω ∂Ω n
We can assume without loss of generality that
�vn �2H 1 (Ω) = 1, (3.14)

and we have � �
2
�∇vn � dx + µ γ0 (vn )2 dΓ → 0. (3.15)
Ω ∂Ω
Now vn is bounded in H 1 (Ω) by (3.14), thus relatively compact in L2 (Ω) by
Rellich’s theorem. We may extract a subsequence, still denoted vn , and v ∈ L2 (Ω)
such that vn → v in L2 (Ω). By (3.15), �∇vn �L2 (Ω) → 0, therefore, since ∇vn → ∇v
in D � (Ω), we have ∇v = 0 and v is constant on each connected component of Ω.
Moreover,
�vn − v�2H 1 (Ω) = �∇vn �2L2 (Ω) + �vn − v�2L2 (Ω) → 0 (3.16)
so that, by continuity of the trace mapping γ0 (vn ) → γ0 (v) in L2 (∂ Ω). By (3.15)
again, we also have �γ0 (vn )�L2 (∂ Ω) → 0 since µ > 0 and therefore γ0 (v) = 0.
It follows that v being a constant with zero trace vanishes in each connected
component, i.e., v = 0. We now realize that (3.14) and (3.16) contradict each other,
therefore our premise that there exists no V -ellipticity constant α is false. �

Remark 3.3.8 This is a typical compactness argument: we can prove that the
constant exists but we have no idea of its value. �
96 3. The variational formulation of elliptic PDEs

3.4 General second order elliptic problems

Up to now, the partial differential operator always was the Laplacian. Let us
rapidly consider more general second order elliptic operators in a Lipschitz open
subset Ω of Rd . We are given a d × d matrix-valued function A(x) = (ai j (x)) with
ai j ∈ C1 (Ω̄). Let u ∈ C2 (Ω) (we can lower this regularity considerably), then A∇u
is a vector field with components
d
(A∇u)i = ∑ ai j ∂ j u
j=1

whose divergence is given by

d
div (A∇u) = ∑ ∂i (A∇u)i
i=1
d d � d �
= ∑ a i j ∂i j u + ∑ ∑ i i j ∂ j u.
∂ a
i, j=1 j=1 i=1

The principal part of this operator ∑di, j=1 ai j ∂i j is of the second order. We will
consider the boundary value problem

 −div (A∇u) + cu = f in Ω,
u = h on Γ0 , (3.17)

bu + n · A∇u = g on Γ1 ,

where c, b, f , g and h are given functions and Γ0 , Γ1 a partition of ∂ Ω as in the

mixed problem. When A = I, we recognize −div (A∇u) = −∆u and n · A∇u = ∂∂ un ,
so that we are generalizing all the model problems seen up to now. First of all, we
reduce the study to the case h = 0 by subtracting a function with the appropriate
trace, as before.

Proposition 3.4.1 Assume that f ∈ L2 (Ω), g ∈ L2 (Γ1 ), c ∈ L∞ (Ω) and b ∈ L∞ (Γ1 ).

Then the triple

V = {v ∈ H 1 (Ω); γ0 (v) = 0 on Γ1 },
� �
a(u, v) = (A∇u · ∇v + cuv) dx + bγ0 (u)γ0 (v) dΓ,
Ω Γ1
� �
�(v) = f v dx + gγ0 (v) dΓ,
Ω Γ1

defines a variational formulation for problem (3.17).

3.4. General second order elliptic problems 97

Proof. The proof is routine, but we write it partially down for completeness.
Multiply the PDE by v ∈ V and integrate by parts. This yields first
� � d � � �
− ∑ i
∂
Ω i=1
(A∇u) i v dx + cuv dx = f v dx,
Ω Ω

then
� d � � d � � �

Ω i=1
∑ (A∇u)i∂iv dx − Γ1 i=1
∑ (A∇u)ini γ0(v) dΓ + cuv dx = f v dx,
Ω Ω

and finally
� � � �
(A∇u · ∇v + cuv) dx + bγ0 (u)γ0 (v) dΓ = f v dx + gγ0 (v) dΓ.
Ω Γ1 Ω Γ1

We leave the converse to the reader. �

Proposition 3.4.2 Let Ω be a Lipschitz open subset of Rd , f ∈ L2 (Ω), g ∈ L2 (Γ1 ),

c ∈ L∞ (Ω) and b ∈ L∞ (Γ1 ). We assume that the matrix A is uniformly elliptic, that
is to say that there exists a constant α > 0 such that
d
∑ ai j (x)ξi ξ j ≥ α�ξ �2
i, j=1

for all x ∈ Ω̄ and all ξ ∈ Rd . We assume in addition that c ≥ η > 0 for some
constant η and that b ≥ 0. Then the problem: Find u ∈ V = {v ∈ H 1 (Ω); γ0 (v) =
0 on Γ1 } such that
� � � �
∀ v ∈ V, (A∇u · ∇v + cuv) dx + bγ0 (u)γ0 (v) dΓ = f v dx + gγ0 (v) dΓ,
Ω Γ1 Ω Γ1

has one and only one solution.

Proof. That V is a Hilbert space and that � is continuous are already known facts.
We leave the proof of the continuity of the bilinear form, which is implied by
the boundedness of the matrix coefficients ai j (x). The V -ellipticity is also quite
obvious, since
� �
a(v, v) = (A∇v · ∇v + cv2 ) dx + bγ0 (v)2 dΓ
Ω Γ1
� �
≥α �∇v�2 dx + η v2 dx
Ω Ω
≥ min(α, η)�v�2H 1 (Ω) ,
hence the existence, uniqueness and continuous dependence of the solution on the
data by Lax-Milgram. �
98 3. The variational formulation of elliptic PDEs

Remark 3.4.1 When the matrix A is not symmetric, neither is the bilinear form a,
even though the principal part of the operator is symmetric since ∑di, j=1 ai j ∂i j =
a +a
∑di, j=1 2 ∂i j due to the fact that ∂i j = ∂ ji . When A is symmetric, then so is the
ij ji

bilinear form and we have an equivalent minimization problem with

� � � � �
1� 2 2
J(v) = (A∇v · ∇v + cv ) dx + bγ0 (v) dΓ − f v dx − gγ0 (v) dΓ,
2 Ω Γ1 Ω Γ1

minimized over V .
It is quite clear that we can reduce the regularity of A down to L∞ without
loosing the existence and uniqueness of the variational problem. The interpretation
in terms of PDEs stops at the divergence form −div (A∇u) + cu = f since we
cannot develop the divergence using Leibniz formula in this case. Such lack of
regularity is useful to model heterogeneous media. �

We now give an example of a non symmetric problem, the convection-diffusion

problem. Let us be given a vector field σ . The convection-diffusion problem reads
�
−∆u + σ · ∇u + cu = f in Ω,
(3.18)
u = 0 on ∂ Ω.

We have a diffusion term −∆u and a transport term σ · ∇u in the same equation
that compete with each other.

Proposition 3.4.3 Assume that f ∈ L2 (Ω), σ ∈ C1 (Ω̄; Rd ) and c ∈ L∞ (Ω). Then

the triple

V = H01 (Ω),
� � �
a(u, v) = ∇u · ∇v + (σ · ∇u + cu)v dx,
�Ω
�(v) = f v dx,
Ω

defines a variational formulation for problem (3.18).

Proof. This is really routine now. . . Note that the bilinear form is not symmetric.�

Proposition 3.4.4 Let Ω be a bounded open subset of Rd , f ∈ L2 (Ω), σ ∈ C1 (Ω̄; Rd )

and c ∈ L∞ (Ω). We assume that c − 12 div σ ≥ 0. Then the problem: Find u ∈ V
such that � � �
�
∀ v ∈ V, ∇u · ∇v + (σ · ∇u + cu)v dx = f v dx,
Ω Ω
has one and only one solution.
3.4. General second order elliptic problems 99

Proof. We just prove the V -ellipticity. We have

� � �
a(v, v) = �∇v�2 + cv2 + (σ · ∇v)v dx.
Ω

Let us integrate the last integral by parts

� � � d �
Ω
(σ · ∇v)v dx =
Ω i=1
∑ σi∂iv v dx
� � d � � � d � � � d �
=− ∑ i i
∂
Ω i=1
(σ v) v dx = − ∑ i i
∂ σ v2
Ω i=1
dx − ∑ i i v dx
σ ∂ v
Ω i=1
� �
=− div σ v2 dx − (σ · ∇v)v dx,
Ω Ω

since all boundary terms vanish, so that

� �
1
(σ · ∇v)v dx = − div σ v2 dx.
Ω 2 Ω

Therefore
� � � � �
1
a(v, v) = �∇v�2 + c − div σ v2 dx ≥ |v|2H 1 (Ω) ,
Ω 2

hence the result by the equivalence of the H 1 semi-norm and the H 1 norm on
H01 (Ω). �

Remark 3.4.2 We thus have existence and uniqueness if c = 0 and div σ = 0.

The case div σ = 0 is interesting because if σ represents the velocity field of
such a fluid as air or water, the divergence free condition is the expression of the
incompressibility of the fluid. Under usual experimental conditions, both fluids are
in fact considered to be incompressible. �

Let us now give a fourth order example, even though only second order prob-
lems were advertised in the section title. We consider a slight variant of the plate
problem with homogeneous Dirichlet boundary conditions
 2
 ∆ u + cu = f in Ω,
u = 0 on ∂ Ω, (3.19)
 ∂u
∂n = 0 on ∂ Ω,

with Ω bounded in Rd (d = 2 in the actual case of a plate).

100 3. The variational formulation of elliptic PDEs

The derivation of a variational formulation is again fairly routine, but since this
is our first (and only) fourth order problem, we give some detail. The variational
space for this Dirichlet problem is V = H02 (Ω) which incorporates the two boundary
conditions. Assume that u ∈ H 4 (Ω) ∩ H02 (Ω). Then ∆u ∈ H 2 (Ω) and we can use
Green’s formula
� �
2
(∆ u)v dx = (∆(∆u))v dx
Ω �Ω �
= ∆u∆v dx + (γ0 (v)γ1 (∆u) − γ1 (v)γ0 (∆u)) dΓ
Ω ∂Ω �
= ∆u∆v dx
Ω

since γ0 (v) = γ1 (v) = 0 for all v ∈ H02 (Ω). So we have our variational formulation
� �
∀ v ∈ V, (∆u∆v + cuv) dx = f v dx, (3.20)
Ω Ω

which is easily checked to give rise to a solution of the boundary value problem.
Let ∇2 v denote the collection of all d 2 second order partial derivatives of v. We
have

Lemma 3.4.1 The semi-norm �∇2 v�L2 (Ω) is a norm on H02 (Ω) that is equivalent
to the H 2 norm.

Proof. It is enough to establish a bound from below. Let v ∈ H02 (Ω). Then we have
∂i v ∈ H01 (Ω) for all i. Therefore �∇(∂i v)�2L2 (Ω) ≥ C2 �∂i v�2H 1 (Ω) , by Poincaré. Now
of course
�∂i v�2H 1 (Ω) = �∇(∂i v)�2L2 (Ω) + �∂i v�2L2 (Ω) ,

so summing in i, we get

d � �
�∇ 2
v�2L2 (Ω) =∑ �∇(∂i v)�2L2 (Ω) ≥ C �∇2 v�2L2 (Ω) + |v|2H 1 (Ω)
2
i=1
≥ C2 �∇2 v�2L2 (Ω) +C4 �v�2H 1 (Ω) ≥ C4 �v�2H 2 (Ω)

since v ∈ H01 (Ω) and C ≤ 1. �

Proposition 3.4.5 Let Ω be a bounded open subset of Rd , f ∈ L2 (Ω) and c ∈

L∞ (Ω). We assume that c ≥ 0. Then problem (3.20) has one and only one solution.
3.4. General second order elliptic problems 101

Proof. We just prove the V -ellipticity. We have

�
a(v, v) ≥ (∆v)2 dx.
Ω

We argue by density. Let ϕ ∈ D(Ω), since ∆ϕ = ∑di=1 ∂ii ϕ, we can write

� � � d �� d � d �

Ω
2
(∆ϕ) dx = ∑ ∂iiϕ ∑ ∂ j j ϕ
Ω i=1
dx = ∑ ∂ii ϕ∂ j j ϕ dx
j=1 i, j=1 Ω
d � d �
=− ∑ ∂i ϕ∂i j j ϕ dx = ∑ ∂i j ϕ∂i j ϕ dx
i, j=1 Ω i, j=1 Ω

with two successive integrations by parts, the first with respect to xi and the second
with respect to x j . Hence, for all ϕ ∈ D(Ω), we obtain
� d �

Ω
2
(∆ϕ) dx = ∑ (∂i j ϕ)2 dx = �∇2 ϕ�2L2 (Ω) .
i, j=1 Ω

Now, by definition, H02 (Ω) is the closure of D(Ω) in H 2 (Ω), thus for all v ∈ H02 (Ω),
there exists a sequence ϕn ∈ D(Ω) such that ϕn → v in H 2 (Ω). Passing to the limit
in the above equality, we thus get
�
(∆v)2 dx = �∇2 v�2L2 (Ω) ,
Ω

since ∂i j ϕn → ∂i j v in L2 (Ω), hence the result by Lemma 3.4.1. �

Remark 3.4.3 Notice the trick used in the above proof. To establish an equality for
H 2 functions, we need to use third derivatives, which make no sense as functions
in this context. However, the formula is valid for smooth functions, for which third
derivatives are not a problem, and since in the end the formula in question does not
involve any derivatives of order higher than two, it extends to H 2 by density.
The formula is actually surprising, since ∆v does not contain any derivative
∂i j v with i �= j, and only the sum of all ∂ii v derivatives. Its L2 norm squared is
nonetheless equal to the sum of the L2 norms squared of all individual second
derivatives. This is related to elliptic regularity, which was mentioned in passing
before. �

Remark 3.4.4 This is another symmetric problem, hence we have an equivalent

energy minimization formulation with
� �
1
J(v) = (∆v)2 dx − f v dx,
2 Ω Ω

to be minimized on H02 (Ω). �

102 3. The variational formulation of elliptic PDEs

To conclude this section, we discuss the general three point strategy for solving
elliptic problems that was repeatedly applied here. First we establish a variational
formulation: (homogeneous) Dirichlet boundary conditions are enforced by the test-
function space, which is included in H 1 for second order problems; we multiply the
PDE by a test-function—possibly assuming additional regularity on the solution—
and use integration by parts or Green’s formula to obtain the variational problem.
The bilinear form must be well-defined on the test-function space.
The second point is to check that the variational formulation actually gives
rise to a solution of the boundary value problem. This point is usually itself in
two steps: first obtain the PDE in the sense of distributions by using test-functions
in D, second retrieve Neumann or Fourier boundary conditions by using the full
test-function space. The first two points can appear somewhat formal because of
the assumed regularity on the solution that is not always easily obtained in the end.
This is not a real problem, since it is possible to write rigorous arguments, at the
expense of more theory than we need here.
The final third point is to try and apply the Lax-Milgram theorem, by making
precise regularity and possibly sign assumptions on the various functions that act
as data. Here we prove existence and uniqueness of the solution to the variational
problem.
A question that can be asked is what is the relevance of such solutions to a
boundary value problem, in which the partial derivatives are taken in a rather weak
sense. This is where elliptic regularity theory comes into play. Using elliptic
regularity, it is possible to show that the variational solution is indeed the classical
solution, provided the data (coefficients, right-hand side, boundary of Ω) is smooth
enough. �