Weak Formulations

39

PART III
Weak Formulation of Elliptic Boundary Value
Problems

Weak Formulations

40

A M ODEL B OUNDARY VALUE P ROBLEM I

Assume that is an open bounded set in Rd and its boundary = is
Lipschitz continuous
Consider an example Boundary Value Problem (the Poisson equation)
(
u = f , in ,
u = 0,

on ,

Given f C(), a classical solution of the above problem is a function

T
u C2 () C() which satisfies the above equation and the boundary
conditions pointwise
Note that existence of such classical solutions for more general problems is
hard to show ....
Introduction of weak solutions allow one to remove some of the high
smoothness requirements

Weak Formulations

41

A M ODEL B OUNDARY VALUE P ROBLEM II

Multiply th equation by a smooth test function v C0 () and integrate
over the domain , then use integration by parts

uv dx =

u v dx =
{z
|
F

Note that the boundary term

f v dx
}

H u
n v d = 0, since v 0 on .

The equation (F) makes sense for much weaker assumptions:

C0 () is dense in H01 (), so as the test function we can take any
v H01 ()
For the RHS it is enough to assume that f H 1 () = (H01 ())0
Thus, the weak formulation of the boundary value problem becomes
u H01 (),

u v dx =

f v dx, v H01 ()

Weak Formulations

42

A M ODEL B OUNDARY VALUE P ROBLEM III

Relation between classical and weak solutions:
classical solutions are also weak solutions
the converse is not true, unless extra regularity is added
Set V = H01 () and define:
a bilinear form a(, ) : V V R such that
a(u, v) =

u v dx, u, v V

a linear functional l : V R such that

l(v) =

f v dx, v V

Then the weak formulation of the problem is to find u V such that

a(u, v) = l(v),

v V

Weak Formulations

43

A M ODEL B OUNDARY VALUE P ROBLEM IV

Define the following differential operator A associated with the boundary
value problem as
A : H01 () H 1 (),

hAu, vi = a(u, v), u, v H01 (),

where h, i denotes the duality pairing between H01 () and H 1 (), i.e.,
l(v) = hl, vi for l H 1 () and v H01 ()
Then the weak formulation of the boundary value problem can be rewritten
as a linear operator equation in a dual space
Au = l, in H 1 ()
Thus, the formalism of weak formulation allows one to convert a differential
equation to an equality of functionals
Note that the weak formulation does not explicitly state the boundary
conditions (they are incorporated into the definition of the function spaces)
Weak formulations directly lead to Galerkintype numerical methods

Weak Formulations

44

L AX M ILGRAM L EMMA I
Theorem There exists a onetoone correspondence between linear
continuous operators A : V V 0 and continuous bilinear forms
a : V V R given by the formula
hAu, vi = a(u, v), u, v V
Consequently, properties of elliptic boundary value problems defined with A
can be studied using the properties of the bilinear form a
Definition The operator A (resp. the bilinear form a) is said to be
V elliptic iff hAv, vi kvkV2 , v V (resp. a(v, v) kvkV2 , v V )
Theorem Assume that K is a nonempty closed subspace of the Hilbert
space V , a : V V R is bilinear, symmetric , bounded and V elliptic,
1
l V 0 . Let
E(v) = a(v, v) l(v), v V
2
Then there exists a unique u K such that E(u) = infvK E(v) which is also
the unique solution of the equation
u K, a(u, v) = l(v), v K

Weak Formulations

45

L AX M ILGRAM L EMMA II
Note that, in practice, the bilinear form a(, ) may not necessarily be
symmetric; The LaxMilgram Lemma is essential for proving existence and
uniqueness of certain operator equations
LaxMilgram Lemma - Assume V is a Hilbert space, a(, ) is a bounded,
V elliptic bilinear form on V and l V 0 . Then there is a unique solution of
the problem
u V, a(u, v) = l(v), v V
Illustration consider the case V = R and a simple linear equation with the
corresponding weak formulation
x R,
ax = l,
x R,

axy = ly,

y R

To ensure existence of solutions we need:

0 < a < , i.e., the bilinear form a(x, y) = axy must be continuous and
Relliptic,
|l| < , i.e., the linear functional l(y) = ly must be bounded

Weak Formulations

46

A M ODEL B OUNDARY VALUE P ROBLEM V

N ON H OMOGENEOUS D IRICHLET BC S
Consider the problem

u = f ,

in ,

u = g,

on ,

Assume that g H 1/2 (); since (by the Trace Theorem )

(H 1 ()) = H 1/2 (), we have the existence of a function G H 1 () such
that G = g
Thus, we can set u = w + G , where w solves the problem with homogeneous
Dirichlet boundary conditions
(
w = f + G, in ,
on ,

w = 0,

The corresponding weak formulation is

w H01 (),

w v dx =

( f v G v) dx, v H01 ()

Existence of a solution follows from the LaxMilgram Lemma.

Weak Formulations

47

A M ODEL B OUNDARY VALUE P ROBLEM VI

N ON H OMOGENEOUS N EUMANN BC S
Consider the Helmholtz equation (the corresponding Poisson equation has

nonunique solutions)
u + u = f , in ,
u
= g, on ,
n
T
Note that the classical solution, if exists, u C 2 () C1 ()

The weak formulation is

1

u H (),

(v v + uv) dx =

f v dx +

gvd, v H 1 ()

Assuming V = H 1 () and
Z
a(u, v) = (v v + uv) dx,

l(v) =

f v dx +

gvd

we can apply the LaxMilgram lemma which guarantees existence of the

weak solutions.

Weak Formulations

48

A M ODEL B OUNDARY VALUE P ROBLEM VII

G ENERAL E LLIPTIC P ROBLEMS
Consider a general elliptic boundary value problem

j (Ai j i u) + Bi i u +Cu = f , in ,

u = 0, on D ,

Ri j (i u)n j = g, on N ,
where the boundary = D

N with D

N = 0/

The following assumptions are made regarding the data:

Ai j , B j ,C L ()
> 0 : Ai j i j ||2 , Rd , a.e. in
f L2 (), g L2 (N )

Weak Formulations

49

A M ODEL B OUNDARY VALUE P ROBLEM VIII

G ENERAL E LLIPTIC P ROBLEMS
The weak formulation is obtained by setting V = H1D () and
a(u, v) =

l(v) =

(Ai j i u j v + Bi (i u)v +Cuv) dx,

f v dx +

gvd

Existence of weak solutions again follows from the application of the

LaxMilgram Lemma

Weak Formulations

50

T HE G ALERKIN M ETHOD I
The Galerkin Method provides a natural framework for finding
finitedimensional approximation of weak solutions of elliptic boundary
value problems
Assume that a(, ) is
bounded (i.e., |a(u, v) MkukV kvkV , u, v V ), and
V elliptic (i.e., a(v, v) c0 kvkV2 , v V )
Given an Ndimensional subspace VN V , consider the problem
uN VN ,

a(uN , v) = l(v),

v VN

With the above assumptions, LaxMilgram Lemma guarantees existence of

an unique solution uN

Weak Formulations

51

T HE G ALERKIN M ETHOD II
How to find uN in practice?
N
Consider a set of basis functions {}N
i=1 such that VN = span{}i=1 and set
N

uN =

j j

j=1

In terms of v take the basis functions i , i = 1, . . . , N

The weak formulation becomes equivalent to a linear algebraic system
A = b,
where:
= ( j ) RN is the vector of unknown coefficients,


A = a(i , j ) RNN is the stiffness matrix
b = (l(i )) RN is the load vector

Approximate solutions with increasing accuracy can be calculated by

considering a sequence of nested spaces VN1 VN2 V

Weak Formulations

52

T HE G ALERKIN M ETHOD III

The RitzGalerkin Method can be used when the bilinear form a(, ) is
symmetric, i.e., a(u, v) = a(v, u), u, v V
Evidently, the original problem is equivalent to the minimization problem
u V,

E(u) = inf E(v),

vV

where the energy functional is E(v) = 12 a(v, v) l(v).

Note that by considering the directional (Gateaux) differential of E(v) we
obtain
E 0 (u; u0 ) = a(u, u0 ) l(u0 ) = 0
as the necessary condition for optimality
In the finitedimensional setting VN V we have
uN VN ,

E(uN ) = inf E(v),

vVN

which can be solved using standard minimization techniques.

When a(, ) is symmetric, the Galerkin and RitzGalerkin methods are
equivalent.

Weak Formulations

53

T HE G ALERKIN M ETHOD IV
As regards error estimation in the Galerkin method, the key result is Ceas
inequality
Lemma Assume V is a Hilbert space, VN V is a subspace, a(, ) is a
bounded, V elliptic bilinear on V , and l V 0 . Let u V be the solution of
the problem
u V, a(u, v) = l(v), v V
and uN VN be the Galerkin approximation defined in
uN VN ,

a(uN , v) = l(v),

v VN

Then there exists a constant C such that

ku uN kV C inf ku vkV
vVN

Weak Formulations

54

T HE G ALERKIN M ETHOD V
Proof of Ceas Lemma
Subtracting the equation for uN from that for u and taking v VN we get
a(u uN , v) = 0, v VN
Using this relation together with V ellipticity and boundedness of a(, )
we get
C0 ku uN kV2 a(u uN , u uN )
= a(u uN , u v) Mku uN kV ku vkV
Thus ku uN kV Cku vkV for any arbitrary v VN

Therefore, to estimate the error of the Galerkin solution, it is sufficient to

estimate the approximation error infvVN ku vkV
When a(, ) is symmetric, it defines an inner product on V whose associated
p
norm kvka = a(v, v) is equivalent to kvkV . With respect to this new inner
product the error of the Galerkin solution u uN is orthogonal to VN

Weak Formulations

55

T HE G ALERKIN M ETHOD VI
Consequences of Ceas Lemma for convergence analysis
Proposition Make the assumptions stated in Ceas Lemma. Assume
additionally that VN1 VN2 is a sequence of subspaces of V with the
property
[

VNi = V

i1

The the Galerkin method converges, i.e.,

ku uNi kV 0, as i
Proof By the above (density) assumption we can find a sequence u i VNi ,
i 1 such that
ku ui kV 0, as i
Applying Ceas inequality we have
ku uNi kV Cku ui kV

Weak Formulations

56

T HE P ETROVG ALERKIN M ETHOD I

Sometimes the solution u and the test function v may belong to different
Hilbert spaces, respectively, U and V
Given a : U V R and l V 0 , the boundary value problem may have the
following weak form
u U

a(u, v) = l(v), v V

Existence of solutions of this problem is addressed by the generalized

LaxMilgram Lemma
Such problem can be solved approximately using the PetrovGalerkin
method given by
uN UN

a(uN , vN ) = l(vN ), vN VN ,

where UN U, VN V , and dim(UN ) = dim(VN ) = N.

The PetrovGalerkin method can be proven to be convergent, if certain
compatibility conditions for the spaces UN and VN are satisfied (the
BabuslaBrezzi conditions)