Numerical Methods For Partial Differential Equations: Joachim SCH Oberl April 6, 2009

Numerical Methods for
Partial Dierential Equations

Joachim Sch oberl
April 6, 2009
2
Contents
1 Introduction 5
1.1 Classication of PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Weak formulation of the Poisson Equation . . . . . . . . . . . . . . . . . . 6
1.3 The Finite Element Method . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 The abstract theory 11
2.1 Basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Projection onto subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Riesz Representation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Symmetric variational problems . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Coercive variational problems . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5.1 Approximation of coercive variational problems . . . . . . . . . . . 20
2.6 Inf-sup stable variational problems . . . . . . . . . . . . . . . . . . . . . . 21
2.6.1 Approximation of inf-sup stable variational problems . . . . . . . . 23
3 Sobolev Spaces 25
3.1 Generalized derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Trace theorems and their applications . . . . . . . . . . . . . . . . . . . . . 28
3.3.1 The trace space H
1/2
. . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 Equivalent norms on H
1
and on sub-spaces . . . . . . . . . . . . . . . . . . 37
4 The weak formulation of the Poisson equation 41
4.1 Shift theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5 Finite Element Method 45
5.1 Finite element system assembling . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Finite element error analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.3 A posteriori error estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.4 Non-conforming Finite Element Methods . . . . . . . . . . . . . . . . . . . 63
6 Linear Equation Solvers 69
6.1 Direct linear equation solvers . . . . . . . . . . . . . . . . . . . . . . . . . 70
3
4 CONTENTS
6.2 Iterative equation solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.3 Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
7 Mixed Methods 95
7.1 Weak formulation of Dirichlet boundary conditions . . . . . . . . . . . . . 95
7.2 A Mixed method for the ux . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.3 Abstract theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.4 Analysis of the model problems . . . . . . . . . . . . . . . . . . . . . . . . 100
7.5 Approximation of mixed systems . . . . . . . . . . . . . . . . . . . . . . . 107
8 Applications 111
8.1 The Navier Stokes equation . . . . . . . . . . . . . . . . . . . . . . . . . . 111
8.2 Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
8.3 Maxwell equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
9 Parabolic partial dierential equations 129
9.1 Semi-discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
9.2 Time integration methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
10 Hyperbolic partial dierential equations 135
Chapter 1
Introduction
Dierential equations are equations for an unknown function involving dierential opera-
tors. An ordinary dierential equation (ODE) requires dierentiation with respect to one
variable. A partial dierential equation (PDE) involves partial dierentiation with respect
to two or more variables.
1.1 Classication of PDEs
The general form of a linear PDE of second order is: nd u : R
d
R such that
d
i,j=1

x
i
a
i,j
(x)
u(x)
x
j
+
d
i=1
b
i
(x)
u(x)
x
i
+ c(x)u(x) = f(x). (1.1)
The coecients a
i,j
(x), b
i
(x), c(x) and the right hand side f(x) are given functions. In
addition, certain type of boundary conditions are required. The behavior of the PDE
depends on the type of the dierential operator
L :=
d
i,j=1
x
i
a
i,j
x
j
+
d
i=1
b
i
x
i
+ c.
Replace

x
i
by s
i
. Then
d
i,j=1
s
i
a
i,j
s
j
+
d
i=1
b
i
s
i
+ c = 0
describes a quartic shape in R
d
. We treat the following cases:
1. In the case of a (positive or negative) denite matrix a = (a
i,j
) this is an ellipse,
and the corresponding PDE is called elliptic. A simple example is a = I, b = 0, and
c = 0, i..e.
2
u
x
2
i
= f.
5
6 CHAPTER 1. INTRODUCTION
Elliptic PDEs require boundary conditions.
2. If the matrix a is semi-denite, has the one-dimensional kernel spanv, and b v = 0,
then the shape is a parabola. Thus, the PDE is called parabolic. A simple example
is
d1
i=1
2
u
x
2
i
+
u
x
d
= f.
Often, the distinguished direction corresponds to time. This type of equation requires
boundary conditiosn on the d1-dimensional boundary, and initial conditions in the
dierent direction.
3. If the matrix a has d 1 positive, and one negative (or vise versa) eigenvalues, then
the shape is a hyperbola. The PDE is called hyperbolic. The simplest one is
d1
i=1
2
u
x
2
i
+

2
u
x
2
d
= f.
Again, the distinguished direction often corresponds to time. Now, two initial con-
ditions are needed.
4. If the matrix a is zero, then the PDE degenerates to the rst order PDE
b
i
u
x
i
+ cu = f.
Boundary conditions are needed at a part of the boundary.
These cases behave very dierently. We will establish theories for the individual cases.
A more general classiccation, for more positive or negative eigenvalues, and systems of
PDEs is possible. The type of the PDE may also change for dierent points x.
1.2 Weak formulation of the Poisson Equation
The most elementary and thus most popular PDE is the Poisson equation
u = f in , (1.2)
with the boundary conditions
u = u
D
on
D
,
u
n
= g on
N
,
u
n
+ u = g on
R
.
(1.3)
The domain is an open and bounded subset of R
d
, where the problem dimension d is
usually 1, 2 or 3. For d = 1, the equation is not a PDE, but an ODE. The boundary
1.2. WEAK FORMULATION OF THE POISSON EQUATION 7
:= consists of the three non-overlapping parts
D
,
N
, and
R
. The outer unit
normal vector is called n. The Laplace dierential operator is :=
d
i=1
2
x
2
i
, the normal
derivative at the boundary is

n
:=
d
i=1
n
i
x
i
. Given are the functions f, u
D
and g in
proper function spaces (e.g., f L
2
()). We search for the unknown function u, again, in
a proper function space dened later.
The boundary conditions are called
Dirichlet boundary condition on
D
. The function value is prescribed,
Neumann boundary condition on
N
. The normal derivative is prescribed,
Robin boundary condition on
R
. An ane linear relation between the function
value and the normal derivative is prescribed.
Exactly one boundary condition must be specied on each part of the boundary.
We transform equation (1.2) together with the boundary conditions (1.3) into its weak
form. For this, we multiply (1.2) by smooth functions (called test functions) and integrate
over the domain:
uv dx =
fv dx (1.4)
We do so for suciently many test functions v in a proper function space. Next, we apply
Gauss theorem
div p dx =
p nds to the function p := u v to obtain
div(u v) dx =
u nv ds
From the product rule there follows div(uv) = uv +u v. Together we obtain
u v dx
u
n
v ds =
fv dx.
Up to now, we only used the dierential equation in the domain. Next, we incorporate
the boundary conditions. The Neumann and Robin b.c. are very natural (and thus are
called natural boundary conditions). We simply replace
u
n
by g and u + g on
N
and
R
, respectively. Putting unknown terms to the left, and known terms to the right hand
side, we obtain
u v dx +
R
uv ds
D
u
n
v ds =
fv dx +
N
+
R
gv ds.
Finally, we use the information of the Dirichlet boundary condition. We work brute force
and simple keep the Dirichlet condition in strong sense. At the same time, we only allow
test functions v fullling v = 0 on
D
. We obtain the
Weak form of the Poisson equation:
Find u such that u = u
D
on
D
and
u v dx +
R
uv ds =
fv dx +
N
+
R
gv ds (1.5)
v such that v = 0 on
D
.
We still did not dene the function space in which we search for the solution u. A proper
choice is
V := v L
2
() : u [L
2
()]
d
and u[
L
2
().
It is a complete space, and, together with the inner product
(u, v)
V
:= (u, v)
L
2
()
+ (u, v)
L
2
()
+ (u, v)
L
2
()
it is a Hilbert space. Now, we see that f L
2
() and g L
2
() is useful. The Dirichlet
b.c. u
D
must be chosen such that there exists an u V with u = u
D
on
D
. By denition
of the space, all terms are well dened. We will see later, that the problem indeed has a
unique solution in V .
1.3 The Finite Element Method
Now, we are developing a numerical method for approximating the weak form (1.5). For
this, we decompose the domain into triangles T. We call the set T = T triangulation.
The set ^ = x
j
is the set of nodes. By means of this triangulation, we dene the nite
element space, V
h
:
V
h
:= v C() : v[
T
is ane linear T T
This is a sub-space of V . The derivatives (in weak sense, see below) are piecewise constant,
and thus, belong to [L
2
()]
2
. The function v
h
V
h
is uniquely dened by its values v(x
j
)
in the nodes x
j
^. We decompose the set of nodes as
^ = ^
D
^
f
,
where ^
T
are the nodes on the Dirichlet boundary, and ^
f
are all others (f as free). The
nite element approximation is dened as
Find u
h
such that u
h
(x) = u
D
(x) x ^
D
and
u
h
v
h
dx +
R
u
h
v
h
ds =
fv
h
dx +
N
+
R
gv
h
ds (1.6)
v
h
V
h
such that v
h
(x) = 0 x ^
D
1.3. THE FINITE ELEMENT METHOD 9
Now it is time to choose a basis for V
h
. The most convenient one is the nodal basis
i
characterized as
i
(x
j
) =
i,j
. (1.7)
The Kronecker- is dened to be 1 for i = j, and 0 else. These are the popular hat
functions. We represent the nite element solution with respect to this basis:
u
h
(x) =
u
i
i
(x) (1.8)
By the nodal-basis property (1.7) there holds u
h
(x
j
) =
i
u
i
i
(x
j
) = u
j
. We have to
determine the coecients u
i
R
N
, with N = [^[. The N
D
:= [^
D
[ values according to
nodes on
D
are given explicitely:
u
j
= u
h
(x
j
) = u
D
(x
j
) x
j

D
The others have to be determined from the variational equation (1.6). It is equivalent to
fulll (1.6) for the whole space v
h
V
h
: v
h
(x) = 0 x
j
^
D
, or just for its basis
i
: x
i
^
f
associated to the free nodes:
i

j
dx+
j
ds
u
i
=
f
j
dx+
N
+
R
g
j
ds (1.9)

j
such that x
j
^
f
We have inserted the basis expansion (1.8). We dene the matrix A = (A
ji
) R
NN
and
the vector f = (f
j
) R
N
as
A
ji
:=
i

j
dx +
j
ds,
f
j
:=
f
j
dx +
N
+
R
g
j
ds.
According to Dirichlet- and free nodes they are splitted as
A =
A
DD
A
DD
A
fD
A
ff
and f =
f
D
f
f
.
Now, we obtain the system of linear equations for u = (u
i
) R
N
, u = (u
D
, u
f
):
I 0
A
fD
A
ff
u
D
u
f
u
D
f
f
. (1.10)
At all, we have N coecients u
i
. N
D
are given explicitely from the Dirichlet values. These
are N
f
equations to determine the remaining ones. Using the known u
D
, we can reformulate
it as symmetric system of equations for u
f
R
N
f
:
A
ff
u
f
= f
f
A
fD
u
D
Chapter 2
The abstract theory
In this chapter we develop the abstract framework for variational problems.
2.1 Basic properties
Denition 1. A vector space V is a set with the operations + : V V V and
: R V V such that for all u, v V and , R there holds
u + v = v + u
(u + v) + w = u + (v + w)
(u + v) = u + v, ( + ) u = u + u
Examples are R
n
, the continuous functions C
0
, or the Lebesgue space L
2
.
Denition 2. A normed vector space (V, | |) is a vector space with the operation |.| :
V R being a norm, i.e., for u, v V and R there holds
|u + v| |u| +|v|
|u| = [[ |u|
|u| = 0 u = 0
Examples are (C
0
, | |
sup
), or (C
0
, | |
L
2
).
Denition 3. In a complete normed vector space, Cauchy sequences (u
n
) V
N
converge
to an u V . A complete normed vector space is called Banach space.
Examples of Banach spaces are (L
2
, | |
L
2
), (C
0
, | |
sup
), but not (C
0
, | |
L
2
).
Denition 4. The closure of a normed vector-space (W, | |
V
), denoted as W
||
V
is the
smallest complete space containing W.
11
12 CHAPTER 2. THE ABSTRACT THEORY
Example: C
||
L
2
= L
2
.
Denition 5. A functional or a linear form l() on V is a linear mapping l() : V R.
The canonical norm for linear forms is the dual norm
|l|
V
:= sup
0,=vV
l(v)
|v|
.
A linear form l is called bounded if the norm is nite. The vector space of all bounded
linear forms on V is called the dual space V
.
An example for a bounded linear form is l() : L
2
R : v
v dx.
Denition 6. A bilinear form A(, ) on V is a mapping A : V V R which is linear
in u and in v. It is called symmetric if A(u, v) = A(v, u) for all u, v V .
Examples are the bilinear form A(u, v) =
uv dx on L
2
, or A(u, v) := u
T
Av on R
n
, where
A is a (symmetric) matrix.
Denition 7. A symmetric bilinear form A(, ) is called an inner product if it satises
(v, v) 0 v V
(v, v) = 0 v = 0
Often, is is denoted as (, )
A
, (, )
V
, or simply (, ).
An examples on R is u
T
Av, where A is a symmetric and positive denite matrix.
Denition 8. An inner product space is a vector space V together with an inner product
(, )
V
.
Lemma 9. Cauchy Schwarz inequality. If A(, ) is a symmetric bilinear form such
that A(v, v) 0 for all v V , then there holds
A(u, v) A(u, u)
1/2
A(v, v)
1/2
Proof: For t R there holds
0 A(u tv, u tv) = A(u, u) 2tA(u, v) + t
2
A(v, v)
If A(v, v) = 0, then A(u, u) 2tA(u, v) 0 for all t R, which forces A(u, v) = 0, and the
inequality holds trivially. Else, if A(v, v) = 0, set t = A(u, v)/A(v, v), and obtain
0 A(u, u) A(u, v)
2
/A(v, v),
which is equivalent to the statement. 2
Lemma 10. |v|
V
:= (v, v)
1/2
V
denes a norm on the inner product space (V, (, )
V
).
2.1. BASIC PROPERTIES 13
Denition 11. An inner product space (V, (, )
V
) which is complete with respect to | |
V
is called a Hilbert space.
Denition 12. A closed subspace S of an Hilbert space V is a subset which is a vector
space, and which is complete with respect to | |
V
.
A nite dimensional subspace is always a closed subspace.
Lemma 13. Let T be a continuous linear operator from the Hilbert space V to the Hilbert
space W. The kernel of T, ker T := v V : Tv = 0 is a closed subspace of V .
Proof: First we observe that ker T is a vector space. Now, let (u
n
) ker T
N
converge
to u V . Since T is continuous, Tu
n
Tu, and thus Tu = 0 and u ker T. 2
Lemma 14. Let S be a subspace (not necessarily closed) of V . Then
S
:= v V : (v, w) = 0 w S
is a closed subspace.
The proof is similar to Lemma 13.
Denition 15. Let V and W be vector spaces. A linear operator T : V W is a linear
mapping from V to W. The operator is called bounded if its operator-norm
|T|
V W
:= sup
0,=vV
|Tv|
W
|v|
V
is nite.
An example is the dierential operator on the according space
d
dx
: (C
1
(0, 1), | |
sup
+
|
d
dx
|
sup
) (C(0, 1), | |
sup
).
Lemma 16. A bounded linear operator is continuous.
Proof. Let v
n
v, i.e. |v
n
v|
V
0. Then |Tv
n
Tv| |T|
V W
|v
n
v|
V
converges
to 0, i.e. Tv
n
Tv. Thus T is continuous.
Denition 17. A dense subspace S of V is such that every element of V can be ap-
proximated by elements of S, i.e.
> 0 u V v S such that |u v|
V
.
Lemma 18 (extension principle). Let S be a dense subspace of the normed space V , and
let W be a complete space. Let T : S W be a bounded linear operator with respect to the
norm |T|
V W
. Then, the operator can be uniquely extended onto V .
Proof. Let u V , and let v
n
be a sequence such that v
n
u. Thus, v
n
is Cauchy. Tv
n
is a well dened sequence in W. Since T is continuous, Tv
n
is also Cauchy. Since W
is complete, there exists a limit w such that Tv
n
w. The limit is independent of the
sequence, and thus Tu can be dened as the limit w.
Denition 19. A bounded linear operator T : V W is called compact if for every
bounded sequence (u
n
) V
N
, the sequence (Tu
n
) contains a convergent sub-sequence.
Lemma 20. Let V, W be Hilbert spaces. A operator is compact if and only if there exists
a complete orthogonal system (u
n
) and values
n
0 such that
(u
n
, u
m
)
V
=
n,m
(Tu
n
, Tu
m
)
W
=
n
n,m
This is the eigensystem of the operator K : V V
: u (Tu, T)
W
.
Proof. (sketch) There exists an maximizing element of
(Tv,Tv)
W
(v,v)
V
. Scale it to |v|
V
= 1
and call it u
1
, and
1
=
(Tu
1
,Tu
1
)
W
(u
1
,u
1
)
V
. Repeat the procedure on the V -complement of u
1
to
generate u
2
, and so on.
2.2 Projection onto subspaces
In the Euklidean space R
2
one can project orthogonally onto a line through the origin, i.e.,
onto a sub-space. The same geometric operation can be dened for closed subspaces of
Hilbert spaces.
Theorem 21. Let S be a closed subspace of the Hilbert space V . Let u V . Then there
exists a unique closest point u
0
S:
|u u
0
| |u v| v V
There holds
u u
0
S
Proof: Let d := inf
vS
|uv|, and let (v
n
) be a minimizing sequence such that |uv
n
| d.
We rst check that there holds
|v
n
v
m
|
2
= 2 |v
n
u|
2
+ 2 |v
m
u|
2
4 |1/2(v
n
+ v
m
) u|
2
.
Since 1/2(v
n
+v
m
) S, there holds |1/2(v
n
+v
m
)u| d. We proof that (v
n
) is a Cauchy
sequence: Fix > 0, choose N N such that for n > N there holds |u v
n
|
2
d
2
+
2
.
Thus for all n, m > N there holds
|v
n
v
m
|
2
2(d
2
+
2
) + 2(d
2
+
2
) 4d
2
= 4
2
.
Thus, v
n
converge to some u
0
V . Since S is closed, u
0
S. By continuity of the norm,
|u u
0
| = d.
2.3. RIESZ REPRESENTATION THEOREM 15
Fix some arbitrary w S, and dene (t) := |uu
0
tw
. .. .
S
|
2
. () is a convex function,
it takes its unique minimum d at t = 0. Thus
0 =
d(t)
dt
[
t=0
= 2(u u
0
, w) 2t(w, w)[
t=0
= 2(u u
0
, w)
We obtained u u
0
S. If there were two minimizers u
0
= u
1
, then u
0
u
1
= (u
0
u)
(u
1
u)S and u
0
u
1
S, which implies u
0
u
1
= 0, a contradiction. 2
Theorem 21 says that given an u V , we can uniquely decompose it as
u = u
0
+ u
1
, u
0
S u
1
S
This allows to dene the operators P

S
: V S and P
S
: V S
as
P
S
u := u
0
P
S
u := (I P
S
)u = u
1
Theorem 22. P
S
and P
S
are linear operators.
Denition 23. A linear operator P is called a projection if P
2
= P. A projector is
called orthogonal, if (Pu, v) = (u, Pv).
Lemma 24. The operators P
S
and P
S
are both orthogonal projectors.
Proof: For u S there holds Pu = u. Since Pu S, there holds P
2
u = Pu. It is
orthogonal since
(Pu, v) = (Pu, v Pv + Pv) = ( Pu
....
S
, v Pv
. .. .
S
) + (Pu, Pv) = (Pu, Pv).

With the same argument there holds (u, Pv) = (Pu, Pv). The co-projector P
S
= I P
S
is a projector since
(I P
S
)
2
= I 2P
S
+ P
2
S
= I P
S
.
It is orthogonal since ((I P
S
)u, v) = (u, v) (P
s
u, v) = (u, v) (u, P
S
v) = (u, (I P
S
)v)
2
2.3 Riesz Representation Theorem
Let u V . Then, we can dene the related continuous linear functional l
u
() V
by
l
u
(v) := (u, v)
V
v V.
The opposite is also true:
Theorem 25. Riesz Representation Theorem. Any continuous linear functional l on
a Hilbert space V can be represented uniquely as
l(v) = (u
l
, v) (2.1)
for some u
l
V . Furthermore, we have
|l|
V
= |u
l
|
V
.
Proof: First, we show uniqueness. Assume that u
1
= u
2
both fulll (2.1). This leads to
the contradiction
0 = l(u
1
u
2
) l(u
1
u
2
)
= (u
1
, u
1
u2) (u
2
, u
1
u
2
) = |u
1
u
2
|
2
.
Next, we construct the u
l
. For this, dene S := ker l. This is a closed subspace.
Case 1: S
= 0. Then, S = V , i.e., l = 0. So take u

l
= 0.
Case 2: S
= 0. Pick some 0 = z S
. There holds l(z) = 0 (otherwise, z S S
=
0). Now dene
u
l
:=
l(z)
|z|
2
z S
Then
(u
l
, v) = ( u
l
....
S
, v l(v)/l(z)z
. .. .
S
) + (u
l
, l(v)/l(z)z)
= l(z)/|z|
2
(z, l(v)/l(z)z)
= l(v)
Finally, we prove |l|
V
= |u
l
|
V
:
|l|
V
= sup
0,=vV
l(v)
|v|
= sup
v
(u
l
, v)
V
|v|
V
|u
l
|
V
and
|u| =
l(z)
|z|
2
|z| =
l(z)
|z|
|l|
V
.
2.4 Symmetric variational problems
Take the function space C
1
(), and dene the bilinear form
A(u, v) :=
uv +
uv ds
2.5. COERCIVE VARIATIONAL PROBLEMS 17
and the linear form
f(v) :=
fv dx
The bilinear form is non-negative, and A(u, u) = 0 implies u = 0. Thus A(, ) is an inner
product, and provides the norm |v|
A
:= A(v, v)
1/2
. The normed vector space (C
1
, |.|
A
) is
not complete. Dene
V := C
1
()
|.|
A
,
which is a Hilbert space per denition. If we can show that there exists a constant c such
that
f(v) =
fv dx c|v|
A
v V
then f(.) is a continuous linear functional on V . We will prove this later. In this case, the
Riesz representation theorem tells that there exists an unique u V such that
A(u, v) = f(v).
This shows that the weak form has a unique solution in V .
Next, take the nite dimensional (closed) nite element subspace V
h
V . The nite
element solution u
h
V
h
was dened by
A(u
h
, v
h
) = f(v
h
) v
h
V
h
,
This means
A(u u
h
, v
h
) = A(u, v
h
) A(u
h
, v
h
) = f(v
h
) f(v
h
) = 0
u
h
is the projection of u onto V
h
, i.e.,
|u u
h
|
A
|u v
h
|
A
v
h
V
h
The error u u
h
is orthogonal to V
h
.
2.5 Coercive variational problems
In this chapter we discuss variational problems posed in Hilbert spaces. Let V be a Hilbert
space, and let A(, ) : V V R be a bilinear form which is
coercive (also known as elliptic)
A(u, u)
1
|u|
2
V
u V, (2.2)
and continuous
A(u, v)
2
|u|
V
|v|
V
u, v V, (2.3)
with bounds
1
and
2
in R
+
. It is not necessarily symmetric. Let f(.) : V R be a
continuous linear form on V , i.e.,
f(v) |f|
V
|v|
V
.
We are posing the variational problem: nd u V such that
A(u, v) = f(v) v V.
Example 26. Diusion-reaction equation:
Consider the PDE
div(a(x)u) + c(x)u = f in ,
with Neumann boundary conditions. Let V be the Hilbert space generated by the inner
product (u, v)
V
:= (u, v)
L
2
+(u, v)
L
2
. The variational formulation of the PDE involves
the bilinear form
A(u, v) =
(a(x)u) v dx +
c(x)uv dx.
Assume that the coecients a(x) and c(x) fulll a(x) R
dd
, a(x) symmetric and
1

min
(a(x))
max
(a(x))
2
, and c(x) such that
1
c(x)
2
almost everywhere. Then
A(, ) is coercive with constant
1
= min
1
,
1
and
2
= max
2
,
2
.
Example 27. Diusion-convection-reaction equation:
The partial dierential equation
u + b u + u = f in
with Dirichlet boundary conditions u = 0 on leads to the bilinear form
A(u, v) =
uv dx +
b u v dx +
uv dx.
If div b 0, what is an important case arising from incompressible ow elds (div b = 0),
then A(, ) is coercive and continuous w.r.t. the same norm as above.
Instead of the linear form f(), we will often write f V
. The evaluation is written

as the duality product
'f, v`
V
V
= f(v).
Lemma 28. A continuous bilinear form A(, ) : V V R induces a continuous linear
operator A : V V
via
'Au, v` = A(u, v) u, v V.
The operator norm |A|
V V
is bounded by the continuity bound
2
of A(, ).
2.5. COERCIVE VARIATIONAL PROBLEMS 19
Proof: For every u V , A(u, ) is a bounded linear form on V with norm
|A(u, )|
V
= sup
vV
A(u, v)
|v|
V
sup
vV
2
|u|
V
|v|
V
|v|
V
=
2
|u|
V
Thus, we can dene the operator A : u V A(u, ) V
. It is linear, and its operator

norm is bounded by
|A|
V V
= sup
uV
|Au|
V
|u|
V
= sup
uV
sup
vV
'Au, v`
V
V
|u|
V
|v|
V
= sup
uV
sup
vV
A(u, v)
|u|
V
|v|
V
sup
uV
sup
vV
2
|u|
V
|v|
V
|u|
V
|v|
V
=
2
.
2
Using this notation, we can write the variational problem as operator equation: nd
u V such that
Au = f (in V
).
Theorem 29 (Banachs contraction mapping theorem). Given a Banach space V and a
mapping T : V V , satisfying the Lipschitz condition
|T(v
1
) T(v
2
)| L|v
1
v
2
| v
1
, v
2
V
for a xed L [0, 1). Then there exists a unique u V such that
u = T(u),
i.e. the mapping T has a unique xed point u. The iteration u
1
V given, compute
u
k+1
:= T(u
k
)
converges to u with convergence rate L:
|u u
k+1
| L|u u
k
|
Theorem 30 (Lax Milgram). Given a Hilbert space V , a coercive and continuous bilinear
form A(, ), and a continuous linear form f(.). Then there exists a unique u V solving
A(u, v) = f(v) v V.
There holds
|u|
V

1
1
|f|
V
(2.4)
Proof: Start from the operator equation Au = f. Let J
V
: V
V be the Riesz
isomorphism dened by
(J
V
g, v)
V
= g(v) v V, g V
.
Then the operator equation is equivalent to
J
V
Au = J
V
f (in V ),
and to the xed point equation (with some 0 = R chosen below)
u = u J
V
(Au f). (2.5)
We will verify that
T(v) := v J
V
(Av f)
is a contraction mapping, i.e., |T(v
1
)T(v
2
)|
V
L|v
1
v
2
|
V
with some Lipschitz constant
L [0, 1). Let v
1
, v
2
V , and set v = v
1
v
2
. Then
|T(v
1
) T(v
2
)|
2
V
= |v
1
J
V
(Av
1
f) v
2
J
V
(Av
2
f)|
2
V
= |v J
V
Av|
2
V
= |v|
2
V
2(J
V
Av, v)
V
+
2
|J
V
Av|
2
V
= |v|
2
V
2 'Av, v` +
2
|Av|
2
V
= |v|
2
V
2A(v, v) +
2
|Av|
2
V
|v|
2
V
2
1
|v|
2
V
+
2
2
2
|v|
2
V
= (1 2
1
+
2
2
2
)|v
1
v
2
|
2
V
Now, we choose =
1
/
2
2
, and obtain a Lipschitz constant
L
2
= 1
2
1
/
2
2
[0, 1).
Banachs contraction mapping theorem state that (2.5) has a unique xed point. Finally,
we obtain the bound (2.4) from
|u|
2
V

1
1
A(u, u) =
1
1
f(u)
1
1
|f|
V
|u|
V
,
and dividing by one factor |u|. 2
2.5.1 Approximation of coercive variational problems
Now, let V
h
be a closed subspace of V . We compute the approximation u
h
V
h
by the
Galerkin method
A(u
h
, v
h
) = f(v
h
) v
h
V
h
. (2.6)
This variational problem is uniquely solvable by Lax-Milgram, since, (V
h
, |.|
V
) is a Hilbert
space, and continuity and coercivity on V
h
are inherited from the original problem on V .
The next theorem says, that the solution dened by the Galerkin method is, up to a
constant factor, as good as the best possible approximation in the nite dimensional space.
Theorem 31 (Cea). The approximation error of the Galerkin method is quasi optimal
|u u
h
|
V

2
/
1
inf
vV
h
|u v
h
|
V
2.6. INF-SUP STABLE VARIATIONAL PROBLEMS 21
Proof: A fundamental property is the Galerkin orthogonality
A(u u
h
, w
h
) = A(u, w
h
) A(u
h
, w
h
) = f(w
h
) f(w
h
) = 0 w
h
V
h
.
Now, pick an arbitrary v
h
V
h
, and bound
|u u
h
|
2
V

1
1
A(u u
h
, u u
h
)
=
1
1
A(u u
h
, u v
h
) +
1
1
A(u u
h
, v
h
u
h
. .. .
V
h
)

2
/
1
|u u
h
|
V
|u v
h
|
V
.
Divide one factor |u u
h
|. Since v
h
V
h
was arbitrary, the estimation holds true also for
the inmum in V
h
. 2
If A(, ) is additionally symmetric, then it is an inner product. In this case, the coer-
civity and continuity properties are equivalent to to
1
|u|
2
V
A(u, u)
2
|u|
2
V
u V.
The generated norm |.|
A
is an equivalent norm to |.|
V
. In the symmetric case, we can
use the orthogonal projection with respect to (., .)
A
to improve the bounds to
|u u
h
|
2
V

1
1
|u u
h
|
2
A

1
1
inf
v
h
V
h
|u v
h
|
2
A

2
/
1
|u v
h
|
2
V
.
The factor in the quasi-optimality estimate is now the square root of the general, non-
symmetric case.
2.6 Inf-sup stable variational problems
The coercivity condition is by no means a necessary condition for a stable solvable system.
A simple, stable problem with non-coercive bilinear form is to choose V = R
2
, and the
bilinear form B(u, v) = u
1
v
1
u
2
v
2
. The solution of B(u, v) = f
T
v is u
1
= f
1
and u
2
= f
2
.
We will follow the convention to call coercive bilinear forms A(, ), and the more general
ones B(, ).
Let V and W be Hilbert spaces, and B(, ) : V W R be a continuous bilinear form
with bound
B(u, v)
2
|u|
V
|v|
W
u V, v W. (2.7)
The general condition is the inf-sup condition
inf
uV
u=0
sup
vW
v=0
B(u, v)
|u|
V
|v|
W

1
. (2.8)
Dene the linear operator B : V W
by 'Bu, v`
W
W
= B(u, v). The inf-sup
condition can be reformulated as
sup
vW
'Bu, v`
|v|
W

1
|u|
V
, u V
and, using the denition of the dual norm,
|Bu|
W

1
|u|
V
. (2.9)
We immediately obtain that B is one to one, since
Bu = 0 u = 0
Lemma 32. Assume that the continuous bilinear form B(, ) fullls the inf-sup condition
(2.8). Then the according operator B has closed range.
Proof: Let Bu
n
be a Cauchy sequence in W
. From (2.9) we conclude that also u

n
is
Cauchy in V . Since V is complete, u
n
converges to some u V . By continuity of B, the
sequence Bu
n
converges to Bu W
. 2
The inf-sup condition (2.8) does not imply that B is onto W
. To insure that, we can

pose an inf-sup condition the other way around:
inf
vW
v=0
sup
uV
u=0
B(u, v)
|u|
V
|v|
W

1
. (2.10)
It will be sucient to state the weaker condition
sup
uV
u=0
B(u, v)
|u|
V
|v|
W
> 0 v W. (2.11)
Theorem 33. Assume that the continuous bilinear form B(, ) fullls the inf-sup condition
(2.8) and condition (2.11). Then, the variational problem: nd u V such that
B(u, v) = f(v) v W (2.12)
has a unique solution. The solution depends continuously on the right hand side:
|u|
V

1
1
|f|
W
Proof: We have to show that the range R(B) = W
. The Hilbert space W
can be split
into the orthogonal, closed subspaces
W
= R(B) R(B)
.
Assume that there exists some 0 = g R(B)
. This means that

(Bu, g)
W
= 0 u V.
2.6. INF-SUP STABLE VARIATIONAL PROBLEMS 23
Let v
g
W be the Riesz representation of g, i.e., (v
g
, w)
W
= g(w) for all w W. This v
g
is in contradiction to the assumption (2.11)
sup
uV
B(u, v
g
)
|u|
V
= sup
uV
(Bu, g)
W
|u|
V
= 0.
Thus, R(B)
= 0 and R(B) = W
. 2
Example 34. A coercive bilinear form is inf-sup stable.
Example 35. A complex symmetric variational problem:
Consider the complex valued PDE
u + iu = f,
with Dirichlet boundary conditions, f L
2
, and i =
1. The weak form for the real

system u = (u
r
, u
i
) V
2
is
(u
r
, v
r
)
L
2
+ (u
i
, v
r
)
L
2
= (f, v
r
) v
r
V
(u
r
, v
i
)
L
2
(u
i
, v
i
)
L
2
= (f, v
i
) v
i
V
(2.13)
We can add up both lines, and dene the large bilinear form B(, ) : V
2
V
2
R by
B((u
r
, u
i
), (v
r
, v
i
)) = (u
r
, v
r
) + (u
i
, v
r
) + (u
r
, v
i
) (u
i
, v
i
)
With respect to the norm |v|
V
= (|v|
2
L
2
+|v|
2
L
2
)
1/2
, the bilinear form is continuous, and
fullls the inf-sup conditions (exercises !) Thus, the variational formulation: nd u V
2
such that
B(u, v) = (f, v
r
) (f, v
i
) v V
2
is stable solvable.
2.6.1 Approximation of inf-sup stable variational problems
Again, to approximate (2.12), we pick nite dimensional subspaces V
h
V and W
h
W,
and pose the nite dimensional variational problem: nd u
h
V
h
such that
B(u
h
, v
h
) = f(v
h
) v
h
W
h
.
But now, in contrast to the coercive case, the solvability of the nite dimensional equation
does not follow from the solvability conditions of the original problem on V W. E.g.,
take the example in R
2
above, and choose the subspaces V
h
= W
h
= span(1, 1).
We have to pose an extra inf-sup condition for the discrete problem:
inf
u
h
V
h
u
h
=0
sup
v
h
W
h
v
h
=0
B(u
h
, v
h
)
|u
h
|
V
|v
h
|
W

1h
. (2.14)
On a nite dimensional space, one to one is equivalent to onto, and we can skip the second
condition.
Theorem 36. Assume that B(, ) is continuous with bound
2
, and B(, ) fullls the
discrete inf-sup condition with bound
1
. Then there holds the quasi-optimal error estimate
|u u
h
| (1 +
2
/
1
) inf
v
h
V
h
|u v
h
| (2.15)
Proof: Again, there holds the Galerkin orthogonality B(u, w
h
) = B(u
h
, w
h
) for all w
h
V
h
.
Again, choose an arbitrary v
h
V
h
:
|u u
h
|
V
|u v
h
|
V
+|v
h
u
h
|
V
|u v
h
|
V
+
1
1h
sup
w
h
W
h
B(v
h
u
h
, w
h
)
|w
h
|
V
= |u v
h
|
V
+
1
1h
sup
w
h
W
h
B(v
h
u, w
h
)
|w
h
|
V
|u v
h
|
V
+
1
1h
sup
w
h
W
h
2
|v
h
u|
V
|w
h
|
W
|w
h
|
W
= (1 +
2
/
1h
)|u v
h
|
V
.
Chapter 3
Sobolev Spaces
In this section, we introduce the concept of generalized derivatives, we dene families of
normed function spaces, and prove inequalities between them. Let be an open subset of
R
d
, either bounded or unbounded.
3.1 Generalized derivatives
Let = (
1
, . . . ,
d
) N
d
0
be a multi-index, [[ =
i
, and dene the classical dierential
operator for functions in C
()
D

x
1

x
n
d
.
For a function u C
(
), the support is dened as
suppu := x : u(x) = 0.
This is a compact set if and only if it is bounded. We say u has compact support in , if
supp u . If is a bounded domain, then u has compact support in if and only if u
vanishes in a neighbourhood of .
The space of smooth functions with compact support is denoted as
T() := C
0
() := u C
() : u has compact support in . (3.1)

For a smooth function u C
[[
(), there holds the formula of integration by parts
udx = (1)
[[
uD
dx T(). (3.2)
The L
2
inner product with a function u in C() denes the linear functional on T
u() := 'u, `
T
T
:=
udx.
25
26 CHAPTER 3. SOBOLEV SPACES
We call these functionals in T
t
distributions. When u is a function, we identify it with
the generated distribution. The formula (3.2) is valid for functions u C
. The strong
regularity is needed only on the left hand side. Thus, we use the less demanding right hand
side to extend the denition of dierentiation for distributions:
Denition 37. For u T
t
, we dene g T
t
to be the generalized derivative D
g
u of u by
'g, `
T
T
= (1)
[[
'u, D
`
T
T
T
If u C
, then D
g
coincides with D
.
The function space of locally integrable functions on is called
L
loc
1
() = u : u
K
L
1
(K) compact K .
It contains functions which can behave very badly near . E.g., e
e
1/x
is in L
1
loc
(0, 1). If
is unbounded, then the constant function 1 is in L
loc
1
, but not in L
1
.
Denition 38. For u L
loc
1
, we call g the weak derivative D
w
u, if g L
loc
1
satises
g(x)(x) dx = (1)
[[
u(x)D
(x) dx T.
The weak derivative is more general than the classical derivative, but more restrictive
than the generalized derivative.
Example 39. Let = (1, 1) and
u(x) =
1 + x x 0
1 x x > 0
Then,
g(x) =
1 x 0
1 x > 0
is the rst generalized derivative D

1
g
of u, which is also a weak derivative. The second
generalized derivative h is
'h, ` = 2(0) T
It is not a weak derivative.
In the following, we will focus on weak derivatives. Unless it is essential we will skip
the sub-scripts w and g.
3.2. SOBOLEV SPACES 27
3.2 Sobolev spaces
For k N
0
and 1 p < , we dene the Sobolev norms
|u|
W
k
p
()
:=
[[k
|D
u|
p
Lp
1/p
,
for k N
0
we set
|u|
W
k
()
:= max
[[k
|D
u|
L
.
In both cases, we dene the Sobolev spaces via
W
k
p
() = u L
loc
1
: |u|
W
k
p
<
In the previous chapter we have seen the importance of complete spaces. This is the
case for Sobolev spaces:
Theorem 40. The Sobolev space W
k
p
() is a Banach space.
Proof: Let v
j
be a Cauchy sequence with respect to | |
W
k
p
. This implies that D
v
j
is a
Cauchy sequence in L
p
, and thus converges to some v
in |.|
Lp
.
We verify that D
v
j
v
implies
v
j
dx
dx for all T. Let K be

the compact support of . There holds
(D
v
j
v
)dx =
K
(D
v
j
v
)dx
|D
v
j
v
|
L
1
(K)
||
L
|D
v
j
v
|
Lp(K)
||
L
0
Finally, we have to check that v
is the weak derivative of v:
dx = lim
j
v
j
dx
= lim
j
(1)
[[
v
j
D
dx =
= (1)
vD
dx.
2
An alternative denition of Sobolev spaces were to take the closure of smooth functions
in the domain, i.e.,
W
k
p
:= C
() : |.|
W
k
p

|.|
W
k
p
.
A third one is to take continuously dierentiable functions up to the boundary
W
k
p
:= C
()
|.|
W
k
p
.
Under moderate restrictions, these denitions lead to the same spaces:
Theorem 41. Let 1 p < . Then

W
k
p
= W
k
p
.
Denition 42. The domain has a Lipschitz boundary, , if there exists a collection
of open sets O
i
, a positive parameter , an integer N and a nite number L, such that for
all x the ball of radius centered at x is contained in some O
i
, no more than N of
the sets O
i
intersect non-trivially, and each part of the boundary O
i
is a graph of a
Lipschitz function
i
: R
d1
R with Lipschitz norm bounded by L.
Theorem 43. Assume that has a Lipschitz boundary, and let 1 p < . Then
W
k
p
= W
k
p
.
The case W
k
2
is special, it is a Hilbert space. We denote it by
H
k
() := W
k
2
().
The inner product is
(u, v)
H
k :=
[[k
(D
u, D
v)
L
2
In the following, we will prove most theorems for the Hilbert spaces H
k
, and state the
general results for W
k
p
.
3.3 Trace theorems and their applications
We consider boundary values of functions in Sobolev spaces. Clearly, this is not well dened
for H
0
= L
2
. But, as we will see, in H
1
and higher order Sobolev spaces, it makes sense
to talk about u[
. The denition of traces is essential to formulate boundary conditions

of PDEs in weak form.
We start in one dimension. Let u C
1
([0, h]) with some h > 0. Then, we can bound
u(0) =
1
x
h
u(x)[
x=0
=
h
0
1
x
h
u(x)
t
dx
=
h
0
1
h
u(x) +
1
x
h
u
t
(x) dx
1
h
L
2
|u|
L
2
+
1
x
h
L
2
|u
t
|
L
2
h
1/2
|u|
L
2
(0,h)
+ h
1/2
|u
t
|
L
2
(0,h)
.
This estimate includes the scaling with the interval length h. If we are not interested in the
scaling, we apply Cauchy-Schwarz in R
2
, and combine the L
2
norm and the H
1
semi-norm
|u
t
|
L
2
to the full H
1
norm and obtain
[u(0)[
h
1/2
+ h
1/2
|u|
2
L
2
+|u
t
|
2
L
2
= c |u|
H
1.
Next, we extend the trace operator to the whole Sobolev space H
1
:
3.3. TRACE THEOREMS AND THEIR APPLICATIONS 29
Theorem 44. There is a well dened and continuous trace operator
tr : H
1
((0, h)) R
whose restriction to C
1
([0, h]) coincides with
u u(0).
Proof: Use that C
1
([0, h]) is dense in H
1
(0, h). Take a sequence u
j
in C
1
([0, h]) con-
verging to u in H
1
-norm. The values u
j
(0) are Cauchy, and thus converge to an u
0
. The
limit is independent of the choice of the sequence u
j
. This allows to dene tr u := u
0
. 2
Now, we extend this 1D result to domains in more dimensions. Let be bounded,
be Lipschitz, and consists of M pieces
i
of smoothness C
1
.
We can construct the following covering of a neighbourhood of in : Let Q = (0, 1)
2
.
For 1 i M, let s
i
C
1
(Q, ) be invertible and such that |s
t
i
|
L
c, |(s
t
i
)
1
|
L
c,
and det s
t
i
> 0. The domains S
i
:= s
i
(Q) are such that s
i
((0, 1) 0) =
i
, and the
parameterizations match on s
i
(0, 1 (0, 1)).
Theorem 45. There exists a well dened and continuous operator
tr : H
1
() L
2
()
which coincides with u[
for u C
1
().
Proof: Again, we prove that
tr : C
1
() L
2
() : u u[
is a bounded operator w.r.t. the norms |.|

H
1 and L
2
, and conclude by density. We use
the partitioning of into the pieces
i
, and transform to the simple square domain
Q = (0, 1)
2
. Dene the functions u
i
on Q = (0, 1)
2
as
u
i
( x) = u(s
i
( x))
We transfer the L
2
norm to the simple domain:
| tr u|
2
L
2
()
=
M
i=1
i
u(x)
2
dx
=
M
i=1
1
0
u(s
i
(, 0))
2
s
i
(, 0)
d
c
M
i=1
1
0
u
i
(, 0)
2
d
To transform the H
1
-norm, we dierentiate with respect to x by applying the chain rule
d u
i
d x
( x) =
du
dx
(s
i
( x))
ds
i
d x
( x).
Solving for
du
dx
is
du
dx
(s
i
( x)) =
d u
i
d x
( x)
ds
d x
1
( x)
The bounds onto s
t
and (s
t
)
1
imply that
c
1
[
x
u[ [
x
u[ c [
x
u[
We start from the right hand side of the stated estimate:
|u|
2
H
1
()

M
i=1
S
i
[
x
u[
2
dx
=
M
i=1
Q
[
x
u(s
i
( x))[
2
det(s
t
) d x
c
M
i=1
Q
[
x
u( x)[
2
d x
We got a lower bound for det(s
t
) = (det(s
t
)
1
)
1
from the upper bound for (s
t
)
1
.
It remains to prove the trace estimate on Q. Here, we apply the previous one dimen-
sional result
[u(, 0)[
2
c
1
0
u(, )
2
+
u(, )
d (0, 1)
The result follows from integrating over
1
0
[u(, 0)[
2
d c
1
0
1
0
u(, )
2
+
u(, )
d d
c |u|
2
H
1
(Q)
.
2
Considering the trace operator from H
1
() to L
2
() is not sharp with respect to the
norms. We will improve the embedding later.
By means of the trace operator we can dene the sub-space
H
1
0
() = u H
1
() : tr u = 0
It is a true sub-space, since u = 1 does belong to H
1
, but not to H
1
0
. It is a closed sub-space,
since it is the kernel of a continuous operator.
By means of the trace inequality, one veries that the linear functional
g(v) :=
N
g tr v dx
is bounded on H
1
.
Integration by parts
The denition of the trace allows us to perform integration by parts in H
1
:
udx =
udiv dx +
tr u ndx [C
1
()]
2
The denition of the weak derivative (e.g. the weak gradient) looks similar. It allows only
test functions with compact support in , i.e., having zero boundary values. Only by
choosing a normed space, for which the trace operator is well dened, we can state and
prove integration by parts. Again, the short proof is based on the density of C
1
() in H
1
.
Sobolev spaces over sub-domains
Let consist of M Lipschitz-continuous sub-domains
i
such that
=
M
i=1
i

i

j
= if i = j
The interfaces are
ij
=
i

j
. The outer normal vector of
i
is n
i
.
Theorem 46. Let u L
2
() such that
u
i
:= u[
i
is in H
1
(
i
), and g
i
= u
i
is its weak gradient
the traces on common interfaces coincide:
tr
ij
u
i
= tr
ij
u
j
Then u belongs to H
1
(). Its weak gradient g = u fullls g[
i
= g
i
.
Proof: We have to verify that g L
2
()
d
, dened by g[
i
= g
i
, is the weak gradient of
u, i.e.,
g dx =
u div dx [C
0
()]
d
We are using Greens formula on the sub-domains
g dx =
M
i=1
i
g
i
dx =
M
i=1
i
u
i
dx
=
M
i=1
i
u
i
div dx +
i
tr u
i
n
i
ds
=
udiv dx +
ij
ij
tr
ij
u
i
n
i
+ tr
ij
u
j
n
j
ds
=
u div dx
We have used that = 0 on , and n
i
= n
j
on
ij
. 2
Applications of this theorem are (conforming nodal) nite element spaces. The parti-
tioning
i
is the mesh. On each sub-domain, i.e., on each element T, the functions are
polynomials and thus in H
1
(T). The nite element functions are constructed to be contin-
uous, i.e., the traces match on the interfaces. Thus, the nite element space is a sub-space
of H
1
.
Extension operators
Some estimates are elementary to verify on simple domains such as squares Q. One tech-
nique to transfer these results to general domains is to extend a function u H
1
() onto
a larger square Q, apply the result for the square, and restrict the result onto the general
domain . This is now the motivation to study extension operators.
We construct a non-overlapping covering S
i
of a neighbourhood of on both sides.
Let =
i
consist of smooth parts. Let s : (0, 1) (1, 1) S
i
: (, ) x be an
invertible function such that
s
i
((0, 1) (0, 1)) = S
i

s
i
((0, 1) 0) =
i
s
i
((0, 1) (1, 0)) = S
i
`
Assume that |
ds
i
dx
|
L
and |
ds
i
dx
1
|
L
are bounded.
This denes an invertible mapping x x(x) from the inside to the outside by
x(x) = s
i
((x), (x)).
The mapping preserve the boundary
i
. The transformations s
i
should be such that x x
is consistent at the interfaces between S
i
and S
j
.
With the ipping operator f : (, ) (, ), the mapping is the composite x(x) =
s
i
(f(s
1
i
)). From that, we obtain the bound
d x
dx
ds
dx
ds
dx
.
Dene the domain

= S
1
. . . S
M
.
We dene the extension operator by
(Eu)( x) = u(x) x S
i
(Eu)(x) = u(x) x
(3.3)
Theorem 47. The extension operator E : H
1
() H
1
(
) is well dened and bounded

with respect to the norms
|Eu|
L
2
(
e
)
c |u|
L
2
()
and
|Eu|
L
2
(
e
)
c |u|
L
2
()
Proof: Let u C
1
(). First, we prove the estimates for the individual pieces S
i
:
S
i
\
Eu( x)
2
d x =
S
i
u(x)
2
det
d x
dx
dx c|u|
2
L
2
(S
i
)
For the derivatives we use
dEu( x)
d x
=
du(x( x))
d x
=
du
dx
dx
d x
.
Since
dx
d x
and (
dx
d x
)
1
=
d x
dx
are bounded, one obtains
[
x
Eu( x)[ [
x
u(x)[,
and
S
i
\
[
x
Eu[
2
dx c
S
i
[u[
2
dx
These estimates prove that E is a bounded operator into H
1
on the sub-domains S
i
` .
The construction was such that for u C
1
(), the extension Eu is continuous across ,
and also across the individual S
i
. By Theorem 46, Eu belongs to H
1
(
), and
|Eu|
2
L
2
()
= |u|
2
+
M
i=1
|u|
2
S
i
\
c|u|
2
L
2
()
,
By density, we get the result for H
1
(). Let u
j
C
1
() u, than u
j
is Cauchy, Eu
j
is
Cauchy in H
1
(
), and thus converges to u H

1
(
).
The extension of functions from H
1
0
() onto larger domains is trivial: Extension by 0
is a bounded operator. One can extend functions from H
1
() into H
1
0
(
), and further, to
an arbitrary domain by extension by 0.
For x = s
i
(, ), , (0, 1)
2
, dene the extension
E
0
u( x) = (1 ) u(x)
This extension vanishes at

Theorem 48. The extension E
0
is an extension from H
1
() to H
1
0
(
). It is bounded
w.r.t.
|E
0
u|
H
1
(
e
)
c|u|
H
1
()
Proof: Exercises
In this case, it is not possible to bound the gradient term only by gradients. To see this,
take the constant function on . The gradient vanishes, but the extension is not constant.
3.3.1 The trace space H
1/2
The trace operator is continuous from H
1
() into L
2
(). But, not every g L
2
() is
a trace of some u H
1
(). We will motivate why the trace space is the fractional order
Sobolev space H
1/2
().
We introduce a stronger space, such that the trace operator is still continuous, and
onto. Let V = H
1
(), and dene the trace space as the range of the trace operator
W = tr u : u H
1
()
with the norm
| tr u|
W
= inf
vV
tr u=tr v
|v|
V
. (3.4)
This is indeed a norm on W. The trace operator is continuous from V W with norm 1.
Lemma 49. The space (W, |.|
W
) is a Banach space. For all g W there exists an u V
such that tr u = g and |u|
V
= |g|
W
Proof: The kernel space V
0
:= v : tr v = 0 is a closed sub-space of V . If tr u = tr v,
then z := u v V
0
. We can rewrite
| tr u|
W
= inf
zV
0
|u z|
V
= |u P
V
0
u|
V
u V
Now, let g
n
= tr u
n
W be a Cauchy sequence. This does not imply that u
n
is Cauchy,
but P
V
0
u
n
is Cauchy in V :
|P
V
0
(u
n
u
m
)|
V
= | tr (u
n
u
m
)|
W
.
The P
V
0
u
n
converge to some u V
0
, and g
n
converge to g := tr u. 2
The minimizer in (3.4) fullls
tr u = g and (u, v)
V
= 0 v V
0
.
This means that u is the solution of the weak form of the Dirichlet problem
u + u = 0 in
u = g on .
To give an explicit characterization of the norm |.|
W
, we introduce Hilbert space
interpolation:
Let V
1
V
0
be two Hilbert spaces, such that V
1
is dense in V
0
, and the embedding
operator id : V
1
V
0
is compact. We can pose the eigen-value problem: Find z V
1
,
R such that
(z, v)
V
1
= (z, v)
V
0
v V.
There exists a sequence of eigen-pairs (z
k
,
k
) such that
k
. The z
k
form an or-
thonormal basis in V
0
, and an orthogonal basis in V
1
.
The converse is also true. If z
k
is a basis for V
0
, and the eigenvalues
k
, then the
embedding V
1
V
0
is compact.
Given u V
0
, it can be expanded in the orthonormal eigen-vector basis:
u =
k=0
u
k
z
k
with u
k
= (u, z
k
)
V
0
The |.|
V
0
- norm of u is
|u|
2
V
0
= (
k
u
k
z
k
,
l
u
l
z
l
)
V
0
=
k,l
u
k
u
l
(z
k
, z
l
)
V
0
=
k
u
2
k
.
If u V
1
, then
|u|
2
V
1
= (
k
u
k
z
k
,
l
u
l
z
l
)
V
0
=
k,l
u
k
u
l
(z
k
, z
l
)
V
1
=
k,l
u
k
u
l
k
(z
k
, z
l
)
V
0
=
k
u
2
k
k
The sub-space space V
1
consists of all u =
u
k
z
k
such that
k

k
u
2
k
is nite. This suggests
the denition of the interpolation norm
|u|
2
Vs
=
k
(u, z
k
)
2
V
0
s
k
,
and the interpolation space V
s
= [V
0
, V
1
]
s
as
V
s
= u V
0
: |u|
Vs
< .
We have been fast with using innite sums. To make everything precise, one rst works
with nite dimensional sub-spaces u : n N and u =
n
k=1
u
k
z
k
, and takes the closure.
In our case, we apply Hilbert space interpolation to H
1
(0, 1) L
2
(0, 1). The eigen-value
problem is to nd z
k
H
1
and
k
R such that
(z
k
, v)
L
2
+ (z
t
k
, v
t
)
L
2
=
k
(z
k
, v)
L
2
v H
1
By denition of the weak derivative, there holds (z
t
k
)
t
= (1
k
)z
k
, i.e., z
k
H
2
. Since
H
2
C
0
, there holds also z C
2
, and a weak solution is also a solution of the strong form
z
k
z
tt
k
=
k
z
k
on (0, 1)
z
t
k
(0) = z
t
k
(1) = 0
(3.5)
All solutions, normalized to |z
k
|
L
2
= 1, are
z
0
= 1
0
= 1
and, for k N,
z
k
(x) =
2 cos(kx)
k
= 1 + k
2
2
.
Indeed, expanding u L
2
in the cos-basis u = u
0
+
k=1
u
k
2 cos(kx), one has

|u|
2
L
2
=
k=0
(u, z
k
)
2
L
2
and
|u|
2
H
1 =
k=0
(1 + k
2
2
)(u, z
k
)
2
L
2
Dierentiation adds a factor k. Hilbert space interpolation allows to dene the fractional
order Sobolev norm (s (0, 1))
|u|
2
H
s
(0,1)
=
k=0
(1 + k
2
2
)
s
(u, z
k
)
2
L
2
We consider the trace tr [
E
of H
1
((0, 1)
2
) onto one edge E = (0, 1) 0. For g
W
E
:= tr H
1
((0, 1)
2
), the norm |g|
W
is dened by
|g|
W
= |u
g
|
H
1.
Here, u
g
solves the Dirichlet problem u
g
[
E
= g, and (u
g
, v)
H
1 = 0 v H
1
such that
tr
E
v = 0.
Since W L
2
(E), we can expand g in the L
2
-orthonormal cosine basis z
k
g(x) =
g
n
z
k
(x)
The Dirichlet problems for the z
k
,
u
k
+ u
k
= 0 in
u
k
= z
k
on E
u
k
n
= 0 on ` E,
3.4. EQUIVALENT NORMS ON H
1
AND ON SUB-SPACES 37
have the explicit solution
u
0
(x, y) = 1
and
u
k
(x, y) =
2 cos(kx)
e
k(1y)
+ e
k(1y)
e
k
+ e
k
.
The asymptotic is
|u
k
|
2
L
2
(k + 1)
1
and
|u
k
|
2
L
2
k
Furthermore, the u
k
are orthogonal in (., .)
H
1. Thus u
g
=
n
g
n
u
k
has the norm
|u
g
|
2
H
1 =
g
2
n
|u
k
|
2
H
1
g
2
n
(1 + k).
This norm is equivalent to H
1/2
(E).
We have proven that the trace space onto one edge is the interpolation space H
1/2
(E).
This is also true for general domains (Lipschitz, with piecewise smooth boundary).
3.4 Equivalent norms on H
1
and on sub-spaces
The intention is to formulate 2
nd
order variational problems in the Hilbert space H
1
. We
want to apply the Lax-Milgram theory for continuous and coercive bilinear forms A(., .).
We present techniques to prove coercivity.
The idea is the following. In the norm
|v|
2
H
1 = |v|
2
L
2
+|v|
2
L
2
,
the | |
L
2
-semi-norm is the dominating part up to the constant functions. The L
2
norm
is necessary to obtain a norm. We want to replace the L
2
norm by some dierent term
(e.g., the L
2
-norm on a part of , or the L
2
-norm on ), and want to obtain an equivalent
norm.
We formulate an abstract theorem relating a norm |.|
V
to a semi-norm |.|
A
. An
equivalent theorem was proven by Tartar.
Theorem 50 (Tartar). Let (V, (., .)
V
) and (W, (., .)
W
) be Hilbert spaces, such that the
embedding id : V W is compact. Let A(., .) be a non-negative, symmetric and V -
continuous bilinear form with kernel V
0
= v : A(v, v) = 0. Assume that
|v|
2
V
|v|
2
W
+|v|
2
A
v V (3.6)
Then there holds
1. The kernel V
0
is nite dimensional. On the factor space V/V
0
, A(., .) is an equivalent
norm to the quotient norm
|u|
A
inf
vV
0
|u v|
V
u V (3.7)
2. Let B(., .) be a continuous, non-negative, symmetric bilinear form on V such that
A(., .) + B(., .) is an inner product. Then there holds
|v|
2
V
|v|
2
A
+|v|
2
B
v V
3. Let V
1
V be a closed sub-space such that V
0
V
1
= 0. Then there holds
|v|
V
|v|
A
v V
1
Proof: 1. Assume that V
0
is not nite dimensional. Then there exists an (., .)
V
-orthonormal
sequence u
k
V
0
. Since the embedding id : V W is compact, it has a sub-sequence
converging in |.|
W
. But, since
2 = |u
k
u
l
|
2
V
|u
k
u
l
|
2
W
+|u
k
u
l
|
A
= |u
k
u
l
|
2
W
for k = l, u
k
is not Cauchy in W. This is a contradiction to an innite dimensional kernel
space V
0
. We prove the equivalence (3.7). To bound the left hand side by the right hand
side, we use that V
0
= ker A, and norm equivalence (3.6):
|u|
A
= inf
vV
0
|u v|
A
inf
vV
0
|u v|
V
The quotient norm is equal to |P
V
0
u|. We have to prove that |P
V
0
u|
V
|P
V
0
u|
A
for
all u V . This follows after proving |u|
V
|u|
A
for all u V
0
. Assume that this is not
true. I.e., there exists a V -orthogonal sequence (u
k
) such that |u
k
|
A
k
1
|u
k
|
V
. Extract
a sub-sequence converging in |.|
W
, and call it u
k
again. From the norm equivalence (3.6)
there follows
2 = |u
k
u
l
|
2
V
_ |u
k
u
l
|
W
+|u
k
u
l
|
A
0
2. On V
0
, |.|
B
is a norm. Since V
0
is nite dimensional, it is equivalent to |.|
V
, say with
bounds
c
1
|v|
2
V
|v|
2
B
c
2
|v|
2
V
v V
0
From 1. we know that
c
3
|v|
2
V
|v|
2
A
c
4
|v|
2
V
v V
0
.
3.4. EQUIVALENT NORMS ON H
1
AND ON SUB-SPACES 39
Now, we bound
|u|
2
V
= |P
V
0
u|
2
V
+|P
V
0
u|
2
V
1
c
1
| P
V
0
u
....
uP
V
0
u
|
2
B
+|P
V
0
|
2
V
2
c
1
|u|
2
B
+ c
2
|P
V
0
u|
2
V
+|P
V
0
u|
2
V
=
2
c
1
|u|
2
B
+
1
c
2
1 +
2c
2
c
1
|P
V
0
u|
2
A
_ |u|
2
B
+|u|
2
A
3. Dene B(u, v) = (P
V
1
u, P
V
1
u
)
V
. Then A(., .) +B(., .) is an inner product: A(u, u) +
B(u, u) = 0 implies that u V
0
and u V
1
, thus u = 0. From 2. there follows that
A(., .) +B(., .) is equivalent to (., .)
V
. The result follows from reducing the equivalence to
V
1
.
2
We want to apply Tartars theorem to the case V = H
1
, W = L
2
, and |v|
A
= |v|
L
2
.
The theorem requires that the embedding id : H
1
L
2
is compact. This is indeed true
for bounded domains :
Theorem 51. The embedding of H
k
H
l
for k > l is compact.
We sketch a proof for the embedding H
1
L
2
. First, prove the compact embedding
H
1
0
(Q) L
2
(Q) for a square Q, w.l.o.g. set Q = (0, 1)
2
. The eigen-value problem: Find
z H
1
0
(Q) and such that
(z, v)
L
2
+ (z, v)
L
2
= (u, v)
L
2
v H
1
0
(Q)
has eigen-vectors z
k,l
= sin(kx)sin(ly), and eigen-values 1 + k
2
2
+ l
2
2
. The
eigen-vectors are dense in L
2
. Thus, the embedding is compact.
On a general domain Q, we can extend H
1
() into H
1
0
(Q), embed H
1
0
(Q) into
L
2
(Q), and restrict L
2
(Q) onto L
2
(). This is the composite of two continuous and a
compact mapping, and thus is compact. 2
The kernel V
0
of the semi-norm |v| is the constant function.
Theorem 52 (Friedrichs inequality). Let
D
be of positive measure [
D
[. Let
V
D
= v H
1
() : tr
D
v = 0. Then
|v|
L
2
_ |v|
L
2
v V
D
Proof: The intersection V
0
V
D
is trivial 0. Thus, Theorem 50, 3. implies the
equivalence
|v|
2
V
= |v|
2
L
2
+|v|
2
L
2
|v|
L
2
.
2
Theorem 53 (Poincare inequality). There holds
|v|
2
H
1
()
|v|
2
L
2
+ (
v dx)
2
Proof: B(u, v) := (
u dx)(
v dx) is a continuous bilinear form on H

1
, and (u, v)+
B(u, v) is an inner product. Thus, Theorem 50, 2. implies the stated equivalence. 2
Let have positive measure [[ in R
d
. Then
|u|
2
H
1
()
|v|
2
L
2
()
+|v|
L
2
()
,
Let have positive measure [[ in R
d1
. Then
|u|
2
H
1
()
|v|
2
L
2
()
+|v|
L
2
()
,
Theorem 54 (Bramble Hilbert lemma). Let U be some Hilbert space, and L : H
k
U be
a continuous linear operator such that Lq = 0 for polynomials q P
k1
. Then there holds
|Lv|
U
[v[
H
k.
Proof: The embedding H
k
H
k1
is compact. The V -continuous, symmetric and
non-negative bilinear form A(u, v) =
:[[=k
(
u,
v) has the kernel P

p1
. Decompose
|u|
2
H
k
= |u|
2
H
k1
+ A(u, u). By Theorem 50, 1, there holds
|u|
A
inf
vV
0
|u v|
H
k
The same holds for the bilinear-form
A
2
(u, v) := (Lu, Lv)
U
+ A(u, v)
Thus
|u|
A
2
inf
vV
0
|u v|
H
k u V
Equalizing both implies that
(Lu, Lu)
U
|u|
2
A
2
|u|
2
A
u V,
i.e., the claim.
We will need point evaluation of functions in Sobolev spaces H
s
. This is possible, we
u H
s
implies that u is continuous.
Theorem 55 (Sobolevs embedding theorem). Let R
d
with Lipschitz boundary. If
u H
s
with s > d/2, then u L
with
|u|
L
_ |u|
H
s
There is a function in C
0
within the L
equivalence class.
Chapter 4
The weak formulation of the Poisson
equation
We are now able to give a precise denition of the weak formulation of the Poisson problem
as introduced in Section 1.2, and analyze the existence and uniqueness of a weak solution.
Let be a bounded domain. Its boundary is decomposed as =
D

N

R
according to Dirichlet, Neumann and Robin boundary conditions.
Let
u
D
H
1/2
(
D
),
f L
2
(),
g L
2
(
N

R
),
L
(
D
), 0.
Assume that there holds
(a) The Dirichlet part has positive measure [
D
[ > 0,
(b) or the Robin term has positive contribution
R
dx > 0.
Dene the Hilbert space
V := H
1
(),
the closed sub-space
V
0
= v : tr
D
v = 0,
and the linear manifold
V
D
= u V : tr
D
u = u
D
.
Dene the bilinear form A(., .) : V V R
A(u, v) =
uv dx +
R
uv ds
41
42 CHAPTER 4. THE WEAK FORMULATION OF THE POISSON EQUATION
and the linear form
f(v) =
fv dx +
R
gv dx.
Theorem 56. The weak formulation of the Poisson problem
Find u V
D
such that
A(u, v) = f(v) v V
0
(4.1)
has a unique solution u.
Proof: The bilinear-form A(., .) and the linear-form f(.) are continuous on V . Tartars
theorem of equivalent norms proves that A(., .) is coercive on V
0
.
Since u
D
is in the closed range of tr
D
, there exists an u
D
V
D
such that
tr u
D
= u
D
and | u
D
|
V
_ |u
D
|
H
1/2
(
D
)
Now, pose the problem: Find z V
0
such that
A(z, v) = f(v) A( u
D
, v) v V
0
.
The right hand side is the evaluation of the continuous linear form f(.) A( u
D
, .) on
V
0
. Due to Lax-Milgram, there exists a unique solution z. Then, u := u
D
+z solves (4.1).
The choice of u
D
is not unique, but, the constructed u is unique. 2
4.1 Shift theorems
Let us restrict to Dirichlet boundary conditions u
D
= 0 on the whole boundary. The
variational problem: Find u V
0
such that
A(u, v) = f(v) v V
0
is well dened for all f V
0
, and, due to Lax-Milgram there holds
|u|
V
0
c|f|
V
0
.
Vice versa, the bilinear-form denes the linear functional A(u, .) with norm
|A(u, .)|
V
0
c|u|
V
0
This dual space is called H
1
:
H
1
:= [H
1
0
()]
4.1. SHIFT THEOREMS 43

Since H
1
0
L
2
, there is L
2
H
1
(). All negative spaces are dened as H
s
() :=
[H
s
0
]
(), for s R
+
. There holds
. . . H
2
0
H
1
0
L
2
H
1
H
2
. . .
The solution operator of the weak formulation is smoothing twice. The statements of
shift theorem are that for s > 0, the solution operator maps also
f H
1+s
u H
1+s
with norm bounds
|u|
H
1+s _ |f|
H
1+s.
In this case, we call the problem H
1+s
- regular.
Theorem 57 (Shift theorem).
(a) Assume that is convex. Then, the Dirichlet problem is H
2
regular.
(b) Let s 2. Assume that C
s
. Then, the Dirichlet problem is H
s
-regular.
We give a proof of (a) for the square (0, )
2
by Fourier series. Let
V
N
= spansin(kx) sin(ly) : 1 k, l N
For an u =
N
k,l=1
u
kl
sin(kx) sin(ly) V
N
, there holds
|u|
2
H
2 = |u|
2
L
2
+|
x
u|
2
L
2
+|
y
u|
2
L
2
+|
2
x
u|
2
L
2
+|
x
y
u|
2
L
2
+|
2
y
u|
2
k,l=1
(1 + k
2
+ l
2
+ k
4
+ k
2
l
2
+ l
4
)u
2
kl
k,l=1
(k
4
+ l
4
)u
2
kl
,
and, for f = u,
| u|
2
L
2
=
N
k,l=1
(k
2
+ l
2
)
2
u
2
kl

N
k,l=1
(k
4
+ l
4
)u
2
kl
.
Thus we have |u|
H
2 |u|
L
2
= |f|
L
2
for u V
N
. The rest requires a closure argument:
There is v : v V
N
= V
N
, and V
N
is dense in L
2
. 2
Indeed, on non-smooth non-convex domains, the H
2
-regularity is not true. Take the
sector of the unit-disc
= (r cos , r sin ) : 0 < r < 1, 0 < <
44 CHAPTER 4. THE WEAK FORMULATION OF THE POISSON EQUATION
with (, 2). Set = / < 1. The function
u = (1 r
2
)r
sin()
is in H
1
0
, and fullls u = (4 + 4)r
sin() L
2
. Thus u is the solution of a Dirichlet
problem. But u H
2
.
On non-convex domains one can specify the regularity in terms of weighted Sobolev
spaces. Let be a polygonal domain containing M vertices V
i
. Let
i
be the interior angle
at V
i
. If the vertex belongs to a non-convex corner (
i
> ), then choose some
i
(1

, 1)
Dene
w(x) =
non-convex
Vertices V
i
[x V
i
[
i
Theorem 58. If f is such that wf L
2
. Then f H
1
, and the solution u of the
Dirichlet problem fullls
|wD
2
u|
L
2
_ |wf|
L
2
.
Chapter 5
Finite Element Method
Ciarlets denition of a nite element is:
Denition 59 (Finite element). A nite element is a triple (T, V
T
,
T
), where
1. T is a bounded set
2. V
T
is function space on T of nite dimension N
T
3.
T
=
1
T
, . . . ,
N
T
T
is a set of linearly independent functionals on V
T
.
The nodal basis
1
T
. . .
N
T
T
for V
T
is the basis dual to
T
, i.e.,
i
T
(
j
T
) =
ij
Barycentric coordinates are useful to express the nodal basis functions.
Finite elements with point evaluation functionals are called Lagrange nite elements,
elements using also derivatives are called Hermite nite elements.
Usual function spaces on T R
2
are
P
p
:= spanx
i
y
j
: 0 i, 0 j, i + j p
Q
p
:= spanx
i
y
j
: 0 i p, 0 j p
Examples for nite elements are
A linear line segment
A quadratic line segment
A Hermite line segment
A constant triangle
A linear triangle
A non-conforming triangle
45
46 CHAPTER 5. FINITE ELEMENT METHOD
A Morley triangle
A Raviart-Thomas triangle
The local nodal interpolation operator dened for functions v C
m
(T) is
I
T
v :=
N
T
=1
T
(v)
T
It is a projection.
Two nite elements (T, V
T
,
T
) and (
T, V
b
T
,
b
T
) are called equivalent if there exists an
invertible function F such that
T = F(
T)
V
T
= v F
1
: v V
b
T

T
=
T
i
: V
T
R : v
T
i
(v F)
Two elements are called ane equivalent, if F is an ane-linear function.
Lagrangian nite elements dened above are equivalent. The Hermite elements are not
equivalent.
Two nite elements are called interpolation equivalent if there holds
I
T
(v) F = I
b
T
(v F)
Lemma 60. Equivalent elements are interpolation equivalent
The Hermite elements dene above are also interpolation equivalent.
A regular triangulation T = T
1
, . . . , T
M
of a domain is the subdivision of a domain
into closed triangles T
i
such that = T
i
and T
i
T
j
is
either empty
or an common edge of T
i
and T
j
or T
i
= T
j
in the case i = j.
In a wider sense, a triangulation may consist of dierent element shapes such as segments,
triangles, quadrilaterals, tetrahedra, hexhedra, prisms, pyramids.
A nite element complex (T, V
T
,
T
) is a set of nite elements dened on the geo-
metric elements of the triangulation T .
It is convenient to construct nite element complexes such that all its nite elements
are ane equivalent to one reference nite element (
T,

V
T
,

T
). The transformation F
T
is
such that T = F
T
(
T).
Examples: linear reference line segment on (0, 1).
5.1. FINITE ELEMENT SYSTEM ASSEMBLING 47
The nite element complex allows the denition of the global interpolation operator for
C
m
-smooth functions by
I
T
v
[T
= I
T
v
T
T T
The nite element space is
V
T
:= v = I
T
w : w C
m
()
We say that V
T
has regularity r if V
T
C
r
. If V
T
= C
0
, the regularity is dened as 1.
Examples:
The P
1
- triangle with vertex nodes leads to regularity 0.
The P
1
- triangle with edge midpoint nodes leads to regularity 1.
The P
0
- triangle leads to regularity 1.
For smooth functions, functionals
T,
and
e
T,
sitting in the same location are equiv-
alent. The set of global functionals =
1
, . . . ,
N
is the linearly independent set of
functionals containing all (equivalence classes of) local functionals.
The connectivity matrix C
T
R
NN
T
is dened such that the local functionals are
derived from the global ones by
T
(u) = C
t
T
(u)
Examples in 1D and 2D
The nodal basis for the global nite element space is the basis in V
T
dual to the global
functionals
j
, i.e.,
j
(
i
) =
ij
There holds
i
[
T
= I
T
i
=
N
T
=1
T
(
i
)
T
=
N
T
=1
(C
t
T
(
i
))
T
=
N
T
=1
(C
t
T
e
i
)
T
=
N
T
=1
C
T,i
T
5.1 Finite element system assembling
As a rst step, we assume there are no Dirichlet boundary conditions. The nite element
problem is
Find u
h
V
T
such that : A(u
h
, v
h
) = f(v
h
) v
h
V
T
(5.1)
The nodal basis and the dual functionals provides the one to one relation between R
N
and
V
T
:
R
N
u u
h
V
T
with u
h
=
N
i=1
i
u
i
and u
i
=
i
(u
h
).
Using the nodal basis expansion of u
h
in (5.1), and testing only with the set of basis
functions, one has
A(
N
i=1
u
i
i
,
j
) = f(
j
) j = 1 . . . N
With
A
ji
= A(
i
,
j
) and f
j
= f(
j
),
one obtains the linear system of equations
Au = f
The preferred way to compute the matrix A and vector f is a sum over element contribu-
tions. The restrictions of the bilinear and linear form to the elements are
A
T
(u, v) =
T
u v dx +
T
uv ds
and
f
T
(v) =
T
fv dx +
T
gv ds
Then
A(u, v) =
TT
A
T
(u, v) f(v) =
TT
f
T
(v)
On each element, one denes the N
T
N
T
element matrix and element vector in terms
of the local basis on T:
A
T,
= A
T
(
T
,
T
) f
T,
= f
T
(
T
)
Then, the global matrix and the global vector are
A =
TT
C
T
A
T
C
t
T
and
f =
TT
C
T
f
T
Namely,
f
i
= f(
i
) =
TT
f
T
(
i
[
T
) =
TT
f
T
(
C
T,i
T
)
=
TT
C
T,i
f
T
(
T
) =
TT
C
T,i
f
5.2. FINITE ELEMENT ERROR ANALYSIS 49

and
A
ij
=
TT
A(
i
[
T
,
j
[
T
) =
TT
A(
C
T,i
T
,
C
T,j
T
)
=
TT
C
T,i
A
T,
C
T,j
On the elements T, the integrands are smooth functions. Thus, numerical integration
rules can be applied.
In the case of Dirichlet boundary conditions, let
D
1, . . . , N correspond to the
vertices x
i
at the Dirichlet boundary, and
f
= 1, . . . N `
D
.
We have the equations
i
D
A
ji
u
i
+
i
f
A
ji
u
i
= f
j
j
f
Inserting u
i
= u
D
(x
i
) for i
i
results in the reduced system
i
f
A
ji
u
i
= f
j
i
D
A
ji
u
D
(x
i
)
An alternative approach is to approximate Dirichlet boundary conditions by Robin b.c.,
u
n
+ u = u
D
, with large parameter .
5.2 Finite element error analysis
Let u be the solution of the variational problem, and u
h
its Galerkin approximation in the
nite element sub-space V
h
. Ceas Lemma bounds the nite element error u u
h
by the
best approximation error
|u u
h
|
V
C inf
vV
h
|u v
h
|
V
.
The constant factor C is the ratio of the continuity bound and the coercivity bound of the
bilinear form A(., .).
Provided that the solution u is suciently smooth, we can take the nite element
interpolant to bound the best approximation error:
inf
vV
h
|u v
h
|
V
|u I
T
u|
V
In the following, we will bound the interpolation error.
Lemma 61. Let

T and T be d-dimensional domains related by the invertible ane linear
transformation F
T
:

T T
F
T
(x) = a + Bx,
where a R
d
and B is a regular matrix in R
dd
. Then there holds:
|u F
T
|
L
2
(
b
T)
= (det B)
1/2
|u|
L
2
(T)
(5.2)
x
im
. . .

x
i
1
(u F
T
) =
d
jm=1
. . .
d
j
1
=1

x
jm
. . .

x
j
1
u
F
T
B
jm,im
. . . B
j
1
,i
1
(5.3)
[u F[
H
m
(
b
T)
_ (det B)
1/2
|B|
m
[u[
H
m
(T)
(5.4)
Proof: Transformation of integrals, chain rule. 2
We dene the diameter of the element T
h
T
= diamT
A triangulation is called shape regular, if all its elements fulll
[T[ _ h
2
T
with a good constant 1. If one studies convergence, one considers families of triangu-
lations with decreasing element sizes h
T
. In that case, the family of triangulations is called
shape regular, if there is a common constant C such that all elements of all triangulations
fulll [T[ Ch
2
T
.
Lemma 62. Let F
T
= a + Bx be the mapping from the reference triangle to the triangle
T. Let [T[ _ h
2
T
. Then there holds
|B
T
| h
T
|B
1
T
| h
1
T
The following lemma is the basis for the error estimate. This lemma is the main
application for the Bramble Hilbert lemma. Sometimes, it is called the Bramble Hilbert
lemma itself:
Lemma 63. Let (T, V
T
,
T
) be a nite element such that the element space V
T
contains
polynomials up to order P
k
. Then there holds
|v I
T
v|
H
1 C[v[
H
m v H
m
(T)
for all m > d/2, m 1, and m k + 1.
Proof: First, we prove that id I
T
is a bounded operator from H
m
to H
1
:
|v I
T
v|
H
1 |v|
H
1 +|I
T
v|
H
1 = |v|
H
1 +|
(v)
|
H
1
|v|
H
1 +
|
H
1[
(v)[
|v|
H
m
The last step used that for H
m
, with m > d/2, point evaluation is continuous. Now, let
v P
k
(T). Since P
k
V
T
, and I
T
is a projection on V
T
, there holds v I
T
v = 0. The
Bramble Hilbert Lemma applied for U = H
1
and L = id I
T
proves the result. 2
To bound the nite element interpolation error, we will transform functions from the
elements T to the reference element

T.
Theorem 64. Let T be a shape regular triangulation of . Let V
T
be a C
0
-regular nite
element space such that all local spaces contain P
1
. Then there holds
|v I
T
v|
L
2
()
_
TT
h
4
T
[v[
H
2
(T)
2
1/2
v H
2
()
[v I
T
v[
2
H
1
()
C
TT
h
2
T
[v[
H
2
(T)
2
v H
2
()
Proof: We prove the H
1
estimate, the L
2
one follows the same lines. The interpolation
error on each element is transformed to the interpolation error on one reference element:
[v I
T
v[
2
H
1
()
=
TT
[(id I
T
)v
T
[
2
H
1
(T)
_
TT
(det B
T
) |B
1
T
|
2
[(id I
T
)v
T
F
T
[
2
H
1
(
b
T)
=
TT
(det B
T
)|B
1
T
|
2
|(id I
b
T
)(v
T
F
T
)|
2
H
1
(
b
T)
On the reference element

T we apply the Bramble-Hilbert lemma. Then, we transform
back to the individual elements:
[v I
T
v[
2
H
1
()
_
TT
(det B
T
)|B
1
T
|
2
[v
T
F
T
[
2
H
m
(
b
T)
_
TT
(det B
T
) |B
1
T
|
2
(det B
1
T
) |B
T
|
4
[v
T
[
2
H
2
(T)
TT
h
2
T
|v
T
|
2
H
2
(T)
.
2
A triangulation is called quasi uniform, is all elements are essentially of the same
size, i.e., there exists one global h such that
h h
T
T T .
On a quasi-uniform mesh, there hold the interpolation error estimates
|u I
T
u|
L
2
()
_ h
2
[u[
H
2
[u I
T
u[
H
1
()
_ h[u[
H
2
We are interested in the rate of the error in terms of the mesh-size h.
Theorem 65 (Finite element error estimate). Assume that
the solution u of the weak bvp is in H
2
,
the triangulation T is quasi-uniform of mesh-size h,
the element spaces contain P
1
.
Then, the nite element error is bounded by
|u u
h
|
H
1 _ h[u[
H
2
Error estimates in L
2
-norm
The above theorem bounds the error in the L
2
-norm of the function, and the L
2
-norm of
the derivatives with the same rate in terms of h. This is obtained by the natural norm of
the variational formulation.
The interpolation error suggests a faster convergence in the weaker norm L
2
. Under
certain circumstances, the nite element error measured in L
2
also decays faster. The
considered variational problem is
Find u V : A(u, v) = f(v) v V.
We dene the dual problem as
Find w V : A(v, w) = f(v) v V.
In the case of a symmetric bilinear form, the primal and the dual problem coincide.
Theorem 66 (Aubin-Nitsche). Assume that
the dual weak bvp is H
2
regular
the triangulation T is quasi-uniform of mesh-size h,
the element spaces contain P
1
.
Then, there holds the L
2
-error estimate
|u u
h
|
L
2
_ h
2
[u[
H
2
Proof: Solve the dual problem with the error u u
h
as right hand side:
Find w V : A(v, w) = (u u
h
, v)
L
2
v V.
Since the dual problem is H
2
regular, there holds w H
2
, and |w|
H
2 _ |u u
h
|
L
2
.
Choose the test function v := u u
h
to obtain the squared norm
A(u u
h
, w) = (u u
h
, u u
h
)
L
2
.
Using the Galerkin orthogonality A(u u
h
, v
h
) = 0 for all v
h
V
h
, we can insert I
T
w:
|u u
h
|
2
L
2
= A(u u
h
, w I
T
w).
Next we use continuity of A(., .) and the interpolation error estimates:
|u u
h
|
2
L
2
_ |u u
h
|
H
1 |w I
T
w|
H
1 _ |u u
h
|
H
1 h[w[
H
2.
From H
2
regularity:
|u u
h
|
2
L
2
_ h|u u
h
|
H
1 |u u
h
|
L
2
,
and, after dividing one factor
|u u
h
|
L
2
_ h|u u
h
|
H
1 _ h
2
|u|
H
2.
2
Approximation of Dirichlet boundary conditions
Till now, we have neglected Dirichlet boundary conditions. In this case, the continuous
problem is
Find u V
D
: A(u, v) = f(v) v V
0
,
where
V
D
= v H
1
: tr
D
v = u
D
and V
0
= v H
1
: tr
D
v = 0.
The nite element problem is
Find u
h
V
hD
: A(u
h
, v
h
) = f(v
h
) v
h
V
h0
,
where
V
hD
= I
T
v : v V
D
and V
h0
= I
T
v : v V
0
.
The denition of V
hD
coincides with v
h
V
h
: v
h
(x
i
) = u
D
(x
i
) vertices x
i
on
D
.
There holds V
h0
V
0
, but, in general, there does not hold V
hD
V
D
.
Theorem 67 (Error estimate for Dirichlet boundary conditions). Assume that
A(., .) is coercive on V
h0
:
A(v
h
, v
h
)
1
|v
h
|
2
V
v
h
V
h0
A(., .) is continuous on V :
A(u, v)
2
|u|
V
|v|
V
u, v V
Then there holds the nite element error estimate
|u u
h
|
H
1 _ h[u[
H
2
Proof: To make use of the coercivity of A(., .), we need an element in V
h0
. There holds
Galerkin orthogonality A(u u
h
, v
h
) = 0 v
h
V
h0
:
|u u
h
|
2
V
= |u I
h
u + I
h
u u
h
|
2
V
2 |u I
h
u|
2
V
+ 2 |I
h
u u
h
|
2
V
2 |u I
h
u|
2
V
+
2
1
A(I
h
u u
h
, I
h
u u
h
)
2|u I
h
u|
2
V
+
2
1
A(I
h
u u, I
h
u u
h
) +
2
1
A(u u
h
, I
h
u u
h
)
2 |u I
h
u|
2
V
+
2
2
1
|I
h
u u||I
h
u u
h
| + 0
2 |u I
h
u|
2
V
+
2
2
1
|I
h
u u|(|I
h
u u| +|u u
h
|)
= (2 +
2
2
1
) |u I
h
u|
2
V
+
2
2
1
|u I
h
u|
V
|u u
h
|
V
Next, we apply ab
1
2
a
2
+
1
2
b
2
for a =
2
2
1
|u I
h
u|
V
and b = |u u
h
|
V
:
|u u
h
|
2
V
(2 +
2
2
1
)|u I
h
u|
2
V
+ 2
2
2
2
1
|u I
h
u|
2
V
+
1
2
|u u
h
|
2
V
Moving the term
1
2
|u u
h
| to the left, we obtain
|u u
h
|
2
V
_ |u I
h
u|
2
V
_ h|u|
H
2
2
High order elements
One can obtain faster convergence, if the solution is smooth, and elements of higher order
are used:
Theorem 68. Assume that
the solution is smooth: u H
m
for m 2
all element spaces V
T
contain polynomials P
p
for p 1
the mesh is quasi-uniform
Then there holds
h
1
|u I
h
u|
L
2
+|u I
h
u|
H
1 _ h
min|m1,k
|u|
H
m
The proof is analogous to the case m = 2 and k = 1. The constants in the estimates
depend on the Sobolev index m and on the polynomial order p. Nodal interpolation is
instable (i.e., the constant grow with p) for increasing order p. There exist better choices
to bound the best approximation error.
5.3. A POSTERIORI ERROR ESTIMATES 55
Graded meshes around vertex singularities
On non-convex meshes domains, the solution is in general not in H
2
, but in some weighted
Sobolev space. The information of the weight can be used to construct proper locally
rened meshes.
On a sector domain with a non-convex corner of angle > , the solution is bounded
in the weighted Sobolev norm
|r
D
2
u|
L
2
C,
with =

. One may choose a mesh such that

h
T
hr
T
, T T
where r
T
is the distance of the center of the element to the singular corner, and h R
+
is
a global mesh size parameter.
We bound the interpolation error:
|u I
T
u|
2
H
1 _
TT
h
2
T
[u[
H
2
(T)

TT
h
2
[r
D
2
u[
L
2
(T)
h
2
|r
D
2
u|
2
L
2
()
_ C h
2
The number of elements in the domain can be roughly estimated by the integral over
the density of elements. The density is number of elements per unit volume, i.e., the inverse
of the area of the element:
N
el

[T[
1
dx =
h
2
r
2
dx = h
2
r
2
dx Ch
2
In two dimensions, and (0, 1), the integral is nite.
Combining the two estimates, one obtains a relation between the error and the number
of elements:
|u I
T
u|
2
V
_ N
1
el
This is the same order of convergence as in the H
2
regular case !
5.3 A posteriori error estimates
We will derive methods to estimate the error of the computed nite element approximation.
Such a posteriori error estimates may use the nite element solution u
h
, and input data
such as the source term f.
(u
h
, f)
An error estimator is called reliable, if it is an upper bound for the error, i.e., there
exists a constant C
1
such that
|u u
h
|
V
C
1
(u
h
, f) (5.5)
An error estimator is ecient, if it is a lower bound for the error, i.e., there exists a
constant C
2
such that
|u u
h
|
V
C
2
(u
h
, f). (5.6)
The constants may depend on the domain, and the shape of the triangles, but may not
depend on the source term f, or the (unknown) solution u.
One use of the a posteriori error estimator is to know the accuracy of the nite element
approximation. A second one is to guide the construction of a new mesh to improve the
accuracy of a new nite element approximation.
The usual error estimators are dened as sum over element contributions:
2
(u
h
, f) =
TT
2
T
(u
h
, f)
The local contributions should correspond to the local error. For the common error
estimators there hold the local eciency estimates
|u u
h
|
H
1
(
T
)
C
2
T
(u
h
, f).
The patch
T
contains T and all its neighbor elements.
In the following, we consider the Poisson equation u = f with homogenous Dirichlet
boundary conditions u = 0 on . We choose piecewise linear nite elements on triangles.
The Zienkiewicz Zhu error estimator
The simplest a posteriori error estimator is the one by Zienkiewicz and Zhu, the so called
ZZ error estimator.
The error is measured in the H
1
-semi norm:
|u u
h
|
L
2
Dene the gradient p = u and the discrete gradient p
h
= u
h
. The discrete gradient
p
h
is a constant on each element. Let p
h
be the p.w. linear and continuous nite element
function obtained by averaging the element values of p
h
in the vertices:
p
h
(x
i
) =
1
[T : x
i
T[
T:x
i
T
p
h[T
for all vertices x
i
The hope is that the averaged gradient is a much better approximation to the true gradient,
i.e.,
|p p
h
|
L
2
|p p
h
|
L
2
(5.7)
holds with a small constant < 1. This property is known as super-convergence.It is
indeed true on (locally) uniform meshes, and smoothness assumptions onto the source
term f.
The ZZ error estimator replaces the true gradient in the error p p
h
by the good
approximation p
h
:
(u
h
) = | p
h
p
h
|
L
2
()
If the super-convergence property (5.7) is fullled, than the ZZ error estimator is reli-
able:
|u u
h
|
L
2
= |p p
h
|
L
2
|p
h
p
h
|
L
2
+|p p
h
|
L
2
|p
h
p
h
|
L
2
+ |p p
h
|
L
2
,
and
|u u
h
|
L
2

1
1
|p
h
p
h
|
L
2
.
It is also ecient, a similar short application of the triangle inequality.
There is a rigorous analysis of the ZZ error estimator, e.g., by showing equivalence to
the following residual error estimator.
The residual error estimator
The idea is to compute the residual of the Poisson equation
f + u
h
,
in the natural norm H
1
. The classical -operator cannot be applied to u
h
, since the rst
derivatives, u
h
, are non-continuous across element boundaries. One can compute the
residuals on the elements
f
[T
+ u
h[T
T T ,
and one can also compute the violation of the continuity of the gradients on the edge
E = T
1
T
2
. We dene the normal-jump term
u
h
n
:=
u
h
n
1
[
T
1
+
u
h
n
2
[
T
2
.
The residual error estimator is
res
(u
h
, f)
2
:=
res
T
(u
h
, f)
2
with the element contributions
res
T
(u
h
, f)
2
:= h
2
T
|f + u
h
|
2
L
2
(T)
+
E:ET
E
h
E
u
h
n
2
L
2
(E)
.
The scaling with h
T
corresponds to the natural H
1
norm of the residual.
To show the reliability of the residual error estimator, we need a new quasi-interpolation
operator, the Clement- operator
h
. In contrast to the interpolation operator, this operator
is well dened for functions in L
2
.
We dene the vertex patch of all elements connected with the vertex x
x
=
T:xT
T,
the edge patch consisting of all elemenets connected with the edge E
E
=
T:ET,=
T,
and the element patch consisting of the element T and all its neighbors
T
=
:TT
,=
T
t
.
The nodal interpolation operator I
h
was dened as
I
h
v =
x
i
1
v(x
i
)
i
,
where
i
are the nodal basis functions. Now, we replace the nodal value v(x
i
) be a local
mean value.
Denition 69 (Clement quasi-interpolation operator). For each vertex x,let v
x
be the
mean value of v on the patch
x
, i.e.,
v
x
=
1
[
x
[
x
v dx.
The Clement operator is
h
v :=
x
i
1
v
x
i
i
.
In the case of homogeneous Dirichlet boundary values, the sum contains only inner vertices.
Theorem 70. The Clement operator satises the following continuity and approximation
estimates:
|
h
v|
L
2
(T)
_ |v|
L
2
(
T
)
|v
h
v|
L
2
(T)
_ h
T
|v|
L
2
(
T
)
|v
h
v|
L
2
(E)
_ h
1/2
E
|v|
L
2
(
E
)
Proof: First, choose a reference patch
T
of dimension 1. The quasi-interpolation
operator is bounded on H
1
(
T
):
|v
h
v|
L
2
(
b
T)
+|(v
h
v)|
L
2
(
b
T)
_ |v|
H
1
(b
T
)
(5.8)
If v is constant on
T
, then the mean values in the vertices take the same values, and
also (
h
v)
[T
is the same constant. The constant function (on
T
) is in the kernel of
|v
h
v|
H
1
(T)
. Due to the Bramble-Hilbert lemma, we can replace the norm on the right
hand side of (5.8) by the semi-norm:
|v
h
v|
L
2
(
b
T)
+|(v
h
v)|
L
2
(
b
T)
_ |v|
L
2
(b
T
)
(5.9)
The rest follows from scaling. Let F : x hx scale the reference patch
T
to the actual
patch
T
. Then
|v
h
v|
L
2
(T)
+ h|(v
h
v)|
L
2
(T)
_ h|v|
L
2
(
T
)
The estimate for the edge term is similar. One needs the scaling of integrals form the
reference edge

E to E:
|v|
L
2
(E)
= h
1/2
E
|v F|
L
2
(

E)
Theorem 71. The residual error estimator is reliable:
|u u
h
| _
res
(u
h
, f)
Proof: From the coercivity of A(., .) we get
|u u
h
|
H
1 _
A(u u
h
, u u
h
)
|u u
h
|
H
1
sup
0,=vH
1
A(u u
h
, v)
|v|
H
1
.
The Galerkin orthogonality A(u u
h
, v
h
) = 0 for all v
h
V
h
allows to insert the Clement
interpolant in the numerator. It is well dened for v H
1
:
|u u
h
|
H
1 sup
0,=vH
1
A(u u
h
, v
h
v)
|v|
H
1
.
We use that the true solution u fullls A(u, v) = f(v), and insert the denitions of A(., .)
and f(.):
A(u u
h
, v
h
v) = f(v
h
v) A(u
h
, v
h
v)
=
fv dx
u
h
(v
h
v) dx
=
TT
T
fv dx
TT
T
u
h
(v
h
v) dx
On each T, the nite element function u
h
is a polynomial. This allows integration by parts
on each element:
A(u u
h
, v
h
v) =
TT
T
fv dx
TT
T
u
h
(v
h
v) dx +
T
u
h
n
(v
h
v) ds
All inner edges E have contributions from normal derivatives from their two adjacent
triangles T
E,1
and T
E,2
. On boundary edges, v
h
v vanishes.
A(u u
h
, v
h
v)
=
T
(f + u
h
)(v
h
v) dx +
u
h
n
[
T
E,1
+
u
h
n
[
T
E,2
(v
h
v) ds
=
T
(f + u
h
)(v
h
v) dx +
u
h
n
(v
h
v) ds
Applying Cauchy-Schwarz rst on L
2
(T) and L
2
(E), and then in R
n
:
A(u u
h
, v
h
v)
T
|f + u
h
|
L
2
(T)
|v
h
v|
L
2
(T)
+
u
h
n
L
2
(E)
|v
h
v|
L
2
(E)
=
T
h
T
|f + u
h
|
L
2
(T)
h
1
T
|v
h
v|
L
2
(T)
+
E
h
1/2
E
u
h
n
L
2
(E)
h
1/2
E
|v
h
v|
L
2
(E)
T
h
2
T
|f + u
h
|
2
L
2
(T)
1/2
T
h
2
T
|v
h
v|
2
L
2
(T)
1/2
+
+
E
h
E
u
h
n
2
L
2
(E)
1/2
E
h
1
E
|v
h
v|
2
L
2
(E)
1/2
We apply the approximation estimates of the Clement operator, and use that only a
bounded number of patches are overlapping:
T
h
2
T
|v
h
v|
2
L
2
(T)
_
T
|v|
2
L
2
(
T
)
_ |v|
2
L
2
()
,
and similar for the edges
E
h
1
E
|v
h
v|
2
L
2
(E)
|v|
2
L
2
()
.
Combining the steps above we observe
|u u
h
|
V
_ sup
vH
1
A(u u
h
, v
h
v)
|v|
1
H
_ sup
V H
1
T
h
2
T
|f + u
h
|
2
L
2
(T)
+
E
h
E
u
h
n
2
L
2
(E)
1/2
|v|
L
2
()
|v|
H
1
T
h
2
T
|f + u
h
|
2
L
2
(T)
+
E
h
E
u
h
n
2
L
2
(E)
1/2
,
what is the reliability of the error estimator
res
(u
h
, f)
Theorem 72. If the source term f is piecewise polynomial on the mesh, then the error
estimator
res
is ecient:
|u u
h
|
V
_
res
(u
h
, f)
Goal driven error estimates
The above error estimators estimate the error in the energy norm V . Some applications
require to compute certain values (such as point values, average values, line integrals, uxes
through surfaces, ...). These values are descibed by linear functionals b : V R. We want
to design a method such that the error in this goal, i.e.,
b(u) b(u
h
)
is small. The technique is to solve additionally the dual problem, where the right hand
side is the goal functional:
Find w V : A(v, w) = b(v) v V.
Usually, one cannot solve the dual problem either, and one applies a Galerkin method also
for the dual problem:
Find w
h
V
h
: A(v
h
, w
h
) = b(v
h
) v
h
V
h
.
In the case of point values, the solution of the dual problem is the Green function (which
is not in H
1
). The error in the goal is
b(u u
h
) = A(u u
h
, w) = A(u u
h
, w w
h
).
A rigorous upper bound for the error in the goal is obtained by using continuity of the
bilinear-form, and energy error estimates
1
and
2
for the primal and dual problem,
respectively:
[b(u u
h
)[ _ |u u
h
|
V
|w w
h
|
V
_
1
(u
h
, f)
2
(w
h
, b).
A good heuristic is the following (unfortunately, not correct) estimate
b(uu
h
) = A(uu
h
, ww
h
) _
TT
|uu
h
|
H
1
(T)
|ww
h
|
H
1
(T)
_
1
T
(u
h
, f)
2
T
(w
h
, b)
(5.10)
The last step would require a local reliability estimate. But, this is not true.
We can interpret (5.10) that way: The local estimators
2
T
(w
h
) provide a way for
weighting the primal local estimators according to the desired goal.
Mesh renement algorithms
A posteriori error estimates are used to control recursive mesh renement:
Start with initial mesh T
Loop
compute fe solution u
h
on T
compute error estimator
T
(u
h
, f)
if tolerance then stop
rene elements with large
T
to obtain a new mesh
The mesh renement algorithm has to take care of
generating a sequence of regular meshes
generating a sequence of shape regular meshes
Red-Green Renement:
A marked element is split into four equivalent elements (called red renement):
But, the obtained mesh is not regular. To avoid such irregular nodes, also neighboring
elements must be split (called green closure):
If one continues to rene that way, the shape of the elements may get worse and worse:
5.4. NON-CONFORMING FINITE ELEMENT METHODS 63
A solution is that elements of the green closure will not be further rened. Instead, remove
the green closure, and replace it by red renement.
Marked edge bisection:
Each triangle has one marked edge. The triangle is only rened by cutting from the middle
of the marked edge to the opposite vertex. The marked edges of the new triangles are the
edges of the old triangle.
If there occurs an irregular node, then also the neighbor triangle must be rened.
To ensure nite termination, one has to avoid cycles in the initial mesh. This can be
obtained by rst sorting the edges (e.g., by length), end then, always choose the largest
edges as marked edge.
Both of these renement algorithms are also possible in 3D.
5.4 Non-conforming Finite Element Methods
In a conforming nite element method, one chooses a sub-space V
h
V , and denes the
nite element approximation as
Find u
h
V
h
: A(u
h
, v
h
) = f(v
h
) v
h
V
h
For reasons of simpler implementation, or even of higher accuracy, the conforming frame-
work is often violated. Examples are:
The nite element space V
h
is not a sub-space of V = H
m
. Examples are the non-
conforming P
1
triangle, and the Morley element for approximation of H
2
.
The Dirichlet boundary conditions are interpolated in the boundary vertices.
The curved domain is approximated by straight sided elements
The bilinear-form and the linear-form are approximated by inexact numerical inte-
gration
The lemmas by Strang are the extension of Ceas lemma to the non-conforming setting.
The First Lemma of Strang
In the rst step, let V
h
V , but the bilinear-form and the linear-form are replaced by
mesh-dependent forms
A
h
(., .) : V
h
V
h
R
and
f
h
(.) : V
h
R.
We do not assume that A
h
and f
h
are dened on V . We assume that the bilinear-forms
A
h
are uniformly coercive, i.e., there exists an
1
independent of the mesh-size such that
A
h
(v
h
, v
h
)
1
|v
h
|
2
V
v
h
V
h
The nite element problem is dened as
Find u
h
V
h
: A
h
(u
h
, v
h
) = f
h
(v
h
) v
h
V
h
Lemma 73 (First Lemma of Strang). Assume that
A(., .) is continuous on V
A
h
(., .) is uniformly coercive
Then there holds
|u u
h
| _ inf
v
h
V
h
|u v
h
| + sup
w
h
V
h
[A(v
h
, w
h
) A
h
(v
h
, w
h
)[
|w
h
|
+ sup
w
h
V
h
f(w
h
) f
h
(w
h
)
|w
h
|
Proof: Choose an arbitrary v
h
V
h
, and set w
h
:= u
h
v
h
. We use the uniform
coercivity, and the denitions of u and u
h
:
1
|u
h
v
h
|
2
V
A
h
(u
h
v
h
, u
h
v
h
) = A
h
(u
h
v
h
, w
h
)
= A(u v
h
, w
h
) + [A(v
h
, w
h
) A
h
(v
h
, w
h
)] + [A
h
(u
h
, w
h
) A(u, w
h
)]
= A(u v
h
, w
h
) + [A(v
h
, w
h
) A
h
(v
h
, w
h
)] + [f
h
(w
h
) f(w
h
)]
Divide by |u
h
v
h
| = |w
h
|, and use the continuity of A(., .):
|u
h
v
h
| _ |u v
h
| +
[A(v
h
, w
h
) A
h
(v
h
, w
h
)[
|w
h
|
+
[f(w
h
) f
h
(w
h
)[
|w
h
|
(5.11)
Using the triangle inequality, the error |u u
h
| is bounded by
|u u
h
| inf
v
h
V
h
|u v
h
| +|v
h
u
h
|
The combination with (5.11) proves the result. 2
Example: Lumping of the L
2
bilinear-form:
Dene the H
1
- bilinear-form
A(u, v) =
u v +
uv dx,
and perform Galerkin discretization with P
1
triangles. The second term leads to a non-
diagonal matrix. The vertex integration rule
T
v dx
[T[
3
3
=1
v(x
T,
)
is exact for v P
1
. We apply this integration rule for the term
uv dx:
A
h
(u, v) =
u v +
TT
[T[
3
3
=1
u(x
T,
)v(x
T,
)
The bilinear form is now dened only for u, v V
h
. The integration is not exact, since
uv P
2
on each triangle.
Inserting the nodal basis
i
, we obtain a diagonal matrix for the second term:
i
(x
T,
)
j
(x
T,
) =
1 for x
i
= x
j
= x
T,
0 else
To apply the rst lemma of Strang, we have to verify the uniform coercivity
T
[T[
3
3
=1
[v
h
(x
T,
)[
2

1
T
[v
h
[
2
dx v
h
V
h
, (5.12)
which is done by transformation to the reference element. The consistency error can be
estimated by
[
T
u
h
v
h
dx
[T[
3
3
=1
u
h
(x
)v
h
(x
)[ _ h
2
T
|u
h
|
L
2
(T)
|v
h
|
L
2
(T)
(5.13)
Summation over the elements give
A(u
h
, v
h
) A
h
(u
h
, v
h
) _ h
2
|u
h
|
H
1
()
|v
h
|
H
1
()
The rst lemma of Strang proves that this modication of the bilinear-form preserves the
order of the discretization error:
|u u
h
|
H
1 _ inf
v
h
V
h
|u v
h
|
H
1 + sup
w
h
V
h
[A(v
h
, w
h
) A
h
(v
h
, w
h
)[
|w
h
|
H
1
_ |u I
h
u|
H
1 + sup
w
h
V
h
[A(I
h
u, w
h
) A
h
(I
h
u, w
h
)[
|w
h
|
H
1
_ h|u|
H
2 + sup
w
h
V
h
h
2
|I
h
u|
H
1|w
h
|
H
1
|w
h
|
H
1
_ h|u|
H
2
A diagonal L
2
matrix has some advantages:
It avoids oscillations in boundary layers (exercises!)
In explicit time integration methods for parabolic or hyperbolic problems, one has to
solve linear equations with the L
2
-matrix. This becomes cheap for diagonal matrices.
The Second Lemma of Strang
In the following, we will also skip the requirement V
h
V . Thus, the norm |.|
V
cannot
be used on V
h
, and it will be replaced by mesh-dependent norms |.|
h
. These norms must
be dened for V +V
h
. As well, the mesh-dependent forms A
h
(., .) and f
h
(.) are dened on
V + V
h
. We assume
uniform coercivity:
A
h
(v
h
, v
h
)
1
|v
h
|
2
h
v
h
V
h
continuity:
A
h
(u, v
h
)
2
|u|
h
|v
h
|
h
u V + V
h
, v
h
V
h
The error can now be measured only in the discrete norm |u u
h
|
V
h
.
Lemma 74. Under the above assumptions there holds
|u u
h
|
h
_ inf
v
h
V
h
|u v
h
|
h
+ sup
w
h
V
h
[A
h
(u, w
h
) f
h
(w
h
)[
|w
h
|
h
(5.14)
Remark: The rst term in (5.14) is the approximation error, the second one is called
consistency error.
Proof: Let v
h
V
h
. Again, set w
h
= u
h
v
h
, and use the V
h
-coercivity:
1
|u
h
v
h
|
2
h
A
h
(u
h
v
h
, u
h
v
h
) = A
h
(u
h
v
h
, w
h
)
= A
h
(u v
h
, w
h
) + [f
h
(w
h
) A
h
(u, w
h
)]
Again, divide by |u
h
v
h
|, and use continuity of A
h
(., .):
|u
h
v
h
|
h
_ |u v
h
|
h
+
A
h
(u, w
h
) f
h
(w
h
)
|w
h
|
h
The rest follows from the triangle inequality. 2
The non-conforming P
1
triangle
The non-conforming P
1
triangle is also called the Crouzeix-Raviart element.
The nite element space generated by the non-conforming P
1
element is
V
nc
h
:= v L
2
: v
[T
P
1
(T), and v is continuous in edge mid-points
The functions in V
nc
h
are not continuous across edges, and thus, V
nc
h
is not a sub-space
of H
1
. We have to extend the bilinear-form and the norm in the following way:
A
h
(u, v) =
TT
T
uv dx u, v V + V
nc
h
and
|v|
2
h
:=
TT
|v|
2
L
2
(T)
v V + V
nc
h
We consider the Dirichlet-problem with u = 0 on
D
.
We will apply the second lemma of Strang.
The continuous P
1
nite element space V
c
h
is a sub-space of V
nc
h
. Let I
h
: H
2
V
c
h
be
the nodal interpolation operator.
To bound the approximation term in (5.14), we use the inclusion V
c
h
V
nc
h
:
inf
v
h
V
nc
h
|u v
h
|
h
|u I
h
u|
H
1 _ h|u|
H
2
We have to bound the consistency term
r(w
h
) = A
h
(u, w
h
) f(w
h
)
=
T
uw
h
T
fw
h
dx
=
T
u
n
w
h
ds
T
(u + f) w
h
ds
=
T
u
n
w
h
ds
Let E be an edge of the triangle T. Dene the mean value w
h
E
. If E is an inner edge,
then the mean value on the corresponding edge of the neighbor element is the same. The
normal derivative
u
n
on the neighbor element is (up to the sign) the same. If E is an edge
on the Dirichlet boundary, then the mean value is 0. This allows to subtract edge mean
values:
r(w
h
) =
ET
E
u
n
(w
h
w
h
E
) ds
Since
E
w
h
w
h
E
ds = 0, we may insert the constant function
I
h
u
n
on each edge:
r(w
h
) =
ET
u
n

I
h
u
n
(w
h
w
h
E
) ds
Apply Cauchy-Schwarz on L
2
(E):
r(w
h
) =
ET
|(u I
h
u)|
L
2
(E)
|w
h
w
h
E
|
L
2
(E)
To estimate these terms, we transform to the reference element

T, where we apply the
Bramble Hilbert lemma. Let T = F
T
(
T), and set

u = u F
T
w
h
= w
h
F
T
There hold the scaling estimates
[w
h
[
H
1
(T)
[ w
h
[
H
1
(
b
T)
|w
h
w
h
E
|
L
2
(E)
h
1/2
E
| w
h
w
h
b
E
|
L
2
(
b
E)
[u[
H
2
(T)
h
1
T
[ u[
H
2
(
b
T)
|(u I
h
u)|
L
2
(E)
h
1/2
E
|( u

I
h
u)|
L
2
(E)
On the reference element, we apply the Bramble Hilbert lemma, once for w
h
, and once for
u. The linear operator
L : H
1
(
T) L
2
(
E) : w
h
w
h
w
h
b
E
is bounded on H
1
(
T) (trace theorem), and Lw = 0 for w P

0
, thus
| w
h
w
h
b
E
|
L
2
(
b
E)
_ [ w
h
[
H
1
(
b
T)
Similar for the term in u: There is |(uI
h
u)|
L
2
(E)
_ |u|
H
2
(T)
, and uI
h
u vanishes for
u P
1
.
Rescaling to the element T leads to
|w
h
w
h
E
|
L
2
(E)
_ h
1/2
[w
h
[
H
1
(T)
|(u I
h
u)|
L
2
(E)
_ h
1/2
[u[
H
2
(T)
This bounds the consistency term
r(w
h
) _
T
h[u[
H
2
(T)
[w
h
[
H
1
(T)
_ h|u|
H
2
()
|w
h
|
h
.
The second lemma of Strang gives the error estimate
|u u
h
| _ h|u|
H
2
There are several applications where the non-conforming P
1
triangle is of advantage:
The L
2
matrix is diagonal (exercises)
It can be used for the approximation of problems in uid dynamics described by the
Navier Stokes equations (see later).
The nite element matrix has exactly 5 non-zero entries in each row associated with
inner edges. That allows simplications in the matrix generation code.
Chapter 6
Linear Equation Solvers
The nite element method, or other discretization schemes, lead to linear systems of equa-
tions
Au = f.
The matrices are typically
of large dimension N (10
4
10
8
unknowns)
and sparse, i.e., there are only a few non-zero elements per row.
A matrix entry A
ij
is non-zero, if there exists a nite element connected with both
degrees of freedom i and j.
A 1D model problem: Dirichlet problem on the interval. A uniform grid with n ele-
ments. The matrix is
A =
2 1
1 2 1
.
.
.
.
.
.
.
.
.
1 2 1
1 2
(n1)(n1)
,
A 2D model problem: Dirichlet problem on a unit-square. A uniform grid with 2n
2
triangles. The unknowns are enumerated lexicographically:
h
1
n
2 n1
(n1)
2
69
70 CHAPTER 6. LINEAR EQUATION SOLVERS
The FEM - matrix of dimension N = (n 1)
2
is
A =
D I
I D I
.
.
.
.
.
.
.
.
.
I D I
I D
with D =
4 1
1 4 1
.
.
.
.
.
.
.
.
.
1 4 1
1 4
(n1)(n1)
,
and the (n 1) (n 1) identity matrix I.
6.1 Direct linear equation solvers
Direct solvers are factorization methods such as LU-decomposition, or Cholesky factoriza-
tion. They require in general O(N
3
) = O(n
6
) operations, and O(N
2
) = O(n
4
) memory. A
fast machine can perform about 10
9
operations per second. This corresponds to
n N time memory
10 10
2
1 ms 80 kB
100 10
4
16 min 800 MB
1000 10
6
30 years 8 TB
A band-matrix of (one-sided) band-width b is a matrix with
A
ij
= 0 for [i j[ > b
The LU-factorization maintains the band-width. L and U are triangular factors of band-
width b. A banded factorization method costs O(Nb
2
) operations, and O(Nb) memory.
For the 1D example, the band-with is 1. Time and memory are O(n). For the 2D example,
the band width is O(n). The time complexity is O(n
4
), the memory complexity is O(n
3
).
This corresponds to
n time memory
10 10 s 8 kB
100 0.1 s 8 MB
1000 16 min 8 GB
Block-elimination methods
By splitting the unknowns into two groups, we rewrite the equation Au = f as a block
system
A
11
A
12
A
21
A
22
u
1
u
2
f
1
f
2
.
6.1. DIRECT LINEAR EQUATION SOLVERS 71
First, expressing u
1
from the rst row gives
u
1
= A
1
11
(f
1
A
12
u
2
),
and the Schur-complement equation to determine u
2
(A
22
A
21
A
1
11
A
12
. .. .
=:S
)u
2
= f
2
A
21
A
1
11
f
1
.
This block-factorization is used in sub-structuring algorithms: Decompose the domain into
mm sub-domains, each one containing 2
n
m
n
m
triangles. Split the unknowns into interior
(I), and coupling (C) unknowns.
m = 2
The interior ones corresponding to dierent sub-domains have no connection in the matrix.
The block matrix is
A
I
A
IC
A
CI
A
C
A
I,1
A
IC,1
.
.
.
.
.
.
A
I,m
2 A
IC,m
2
A
CI,1
A
CI,m
2 A
C
Factorizing the block-diagonal interior block A

I
splits into m
2
independent factorization
problems. If one uses a banded factorization, the costs are
m
2
n
m
4
=
n
4
m
2
Computing the Schur complement
S = A
C
A
CI
A
1
I
A
IC
= A
C

m
2
i=1
A
CI,i
A
1
I,i
A
IC,i
is of the same cost. The Schur complement is of size mn, and has band-width n. Thus,
the factorization costs O(mn
3
). The total costs are of order
n
4
m
2
+ mn
3
Equilibrating both terms lead to the optimal number of sub-domains m = n
1/3
, and to the
asymptotic costs
n
3.33
If a parallel computer is used, the factorization of A
I
and the computation of Schur com-
plements can be performed in parallel.
The hierarchical sub-structuring algorithm, known as nested dissection, eliminates in-
terior unknowns hierarchically:
Let n = 2
L
. On level l, with 1 l L, one has 4
l
sub-domains. Each sub-domain has
O(2
Ll
) unknowns. The factorization of the inner blocks on level l costs
4
l
(2
Ll
)
3
= 2
3Ll
Forming the Schur-complement is of the same cost. The sum over all levels is
L
l=1
2
3Ll
= 2
3L
1
2
+
1
4
+ . . .
2
3L
The factorization costs are O(n
3
). Storing the matrices on each level costs
4
l
(2
Ll
)
2
= 2
2L
.
The total memory is O(L 2
2L
) = O(n
2
log n).
This corresponds to
n time memory
10 1 s 3 kB
100 1 ms 500 kB
1000 1 s 150 MB
A corresponding sparse factorization algorithm for matrices arising from unstructured
meshes is based on minimum degree ordering. Successively, the unknowns with the least
connections in the matrix graph are eliminated.
In 2D, a direct method with optimal ordering is very ecient. In 3D, the situation is
worse for the direct solver. There holds N = n
3
, time complexity = O(N
2
), and memory
= O(N
1.33
).
6.2. ITERATIVE EQUATION SOLVERS 73
6.2 Iterative equation solvers
Iterative equation solvers improve the accuracy of approximative solution by an successive
process. This requires in general much less memory, and, depending on the problem and
on the method, may be (much) faster.
The Richardson iteration
A simple iterative method is the preconditioned Richardson iteration (also known as simple
iteration, or Picard iteration):
start with arbitrary u
0
for k = 0, 1, . . . convergence
d
k
= f Au
k
w
k
= C
1
d
k
u
k+1
= u
k
+ w
k
Here, is a damping parameter which may be necessary to ensure convergence. The
matrix C is called a preconditioner. It should fulll
1. C is a good approximation to A
2. the matrix-vector multiplication w = C
1
d should be cheap
A simple choice is C = diag A, the Jacobi preconditioner. The application of C
1
is cheap.
The quality of the approximation C A will be estimated below. The optimal choice for
the rst criterion would be C = A. But, of course, w = C
1
d is in general not cheap.
Combining the steps, the iteration can be written as
u
k+1
= u
k
+ C
1
(f Au
k
)
Let u be the solution of the equation Au = f. We are interested in the behavior of the
error u
k
u:
u
k+1
u = u
k
u + C
1
(f Au
k
)
= u
k
u + C
1
(Au Au
k
)
= (I C
1
A)(u
k
u)
We call the matrix
M = I C
1
A
the iteration matrix. The error transition can be estimated by
|u
k+1
u| |M| |u
k
u|.
The matrix norm is the associated matrix norm to some vector norm. If := |M| < 1,
then the error is reduced. The error after k steps is
|u
k
u|
k
|u
0
u|
To reduce the error by a factor (e.g., = 10
8
), one needs
N
its
=
log
log
iterations.
We will focus on the symmetric (A = A
T
) and positive denite (u
T
Au > 0 for u =
0) case (short: SPD). Then it makes sense to choose symmetric and positive denite
preconditioners C = C
T
. Eigenvalue decomposition allows a sharp analysis. Pose the
generalized eigenvalue problem
Az = Cz.
Let (
i
, z
i
) be the set of eigen-pairs. The spectrum is C
1
A =
i
. The eigen-vectors
z
i
are normalized to
|z
i
|
C
= 1
The eigenvalues can are bounded from below and from above by the Rayleigh quotient:
min
v
v
T
Av
v
T
Cv

i
max
v
v
T
Av
v
T
Cv
The ratio of largest to smallest eigen-value is the relative spectral condition number
=

N
1
We will establish the spectral bounds
1
v
T
Cv v
T
Av
2
v
T
Cv v R
N
,
which allow to bound the eigenvalues
i
[
1
,
2
],
and the condition number

2
1
.
A vector v can be expressed in terms of the eigen-vector basis z
i
as v =
v
i
e
i
. There
holds
|v|
2
C
=
v
2
i
|v|
2
A
=
i
v
2
i
Lemma 75. The iteration matrix M can be bounded in A-norm and in C-norm:
|M|
A
sup
[1,
2
]
[1 [
|M|
C
sup
[1,
2
]
[1 [
Proof: Express v =
v
i
z
i
. Then
Mv = (I C
1
A)v =
v
i
(I C
1
A)z
i
=
v
i
(1
i
)z
i
The norm is
|Mv|
2
A
=
i
v
2
i
(1
i
)
2
sup
i
(1
i
)
2
i
v
2
i
sup
[
1
,
2
]
(1 )
2
|v|
2
A
and thus
|M|
A
= sup
vR
n
|Mv|
A
|v|
A
sup
[
1
,
2
]
[1 [.
The proof is equivalent for |M|
C
. 2
1
1
1
2
1
2
1
1
The optimal choice of the relaxation parameter is such that
1
1
= (1
2
),
i.e.,
=
2
1
+
2
The convergence factor is
1
1
=

2
2
+
1
.
Assume we knew sharp spectral bounds
1
=
1
and
2
=
N
. Then the convergence factor
is
|M| =
1
+ 1
1
2

The number of iterations to reduce the error by a factor is
N
its
=
log
log

log
2/
= log
1

2
Take the 1D stiness matrix of dimension N N:
A =
2 1
1 2 1
.
.
.
.
.
.
.
.
.
1 2 1
1 2
,
and the trivial preconditioner C = I. The eigen-vectors z
i
and eigen-values
i
are
z
i
=
sin
ij
N + 1
j=1,...N
i
= 2 2 cos(
i
N + 1
)
The extremal eigenvalues are
1
= 2 2 cos(

N + 1
)

2
N
2
N
= 2 2 cos(
N
N + 1
4

2
N
2
.
The optimal damping is
=
2
1
+
N
=
1
2
,
and the convergence factor is
|M| 1
2
1
N
1
2
2
N
2
The number of iterations is
N
its
log
1
N
2
For the 2D model problem with N = (n 1)
2
, the condition number behaves like
n
2
.
The costs to achieve a relative accuracy are
N
its
Costs-per-iteration log
1
n
2
N log
1
n
4
The costs per digit are comparable to the band-factorization. The memory requirement is
optimal O(N).
The gradient method
It is not always feasible to nd the optimal relaxation parameter a priori. The gradient
method is a modication to the Richardson method to nd automatically the optimal
relaxation parameter :
The rst steps are identic:
d
k
= f Au
k
w
k
= C
1
d
k
Now, perform the update
u
k+1
= u
k
+ w
k
such that the error is minimal in energy norm:
Find such that |u u
k+1
|
A
= min!
Although the error cannot be computed, this minimization is possible:
|u u
k+1
|
2
A
= |u u
k
w
k
|
2
A
= (u u
k
)
T
A(u u
k
) 2(u u
k
)
T
Aw
k
+
2
(w
k
)
T
Aw
k
This is a convex function in . It takes its minimum at
0 = 2(u u
k
)
T
Aw
k
+ 2
opt
(w
k
)
T
Aw
k
,
i.e.,
opt
=
w
k
A(u u
k
)
(w
k
)
T
Aw
k
=
w
k
d
k
(w
k
)
T
Aw
k
Since the gradient method gives optimal error reduction in energy norm, its convergence
rate can be estimated by the Richardson iteration with optimal choice of the relaxation
parameter:
|u u
k+1
|
A

1
+ 1
|u u
k
|
A
The Chebyshev method
We have found the optimal choice of the relaxation parameter for one step of the iteration.
If we perform m iterations, the overall rate of convergence can be improved by choosing
variable relaxation parameters
1
, . . .
m
.
The m-step iteration matrix is
M = M
m
. . . M
2
M
1
= (I
m
C
1
A) . . . (I
1
C
1
A).
By diagonalization, the A-norm and C-norm are bounded by
|M| max
[
1
,
N
]
[(1
1
) . . . (1
m
)[
The goal is to optimize
1
, . . .
m
:
min
1
,...m
max
[
1
,
N
]
[(1
1
) . . . (1
m
)[
This is a polynomial in , of order m, and p(0) = 1:
min
pP
m
p(0)=1
max
[
1
,
N
]
[p()[. (6.1)
This optimization problem can be solved explicitely by means of Chebyshev polynomials.
These are the polynomials dened by
T
m
(x) =
cos(m arccos(x)) [x[ 1

cosh(m arccosh(x)) [x[ > 1
The T
m
fulll the recurrence relation
T
0
(x) = 1
T
1
(x) = x
T
m+1
(x) = 2xT
m
(x) T
m1
(x)
The T
m
fulll also
T
m
(x) =
1
2
(x +
x
2
1)
m
+ (x +
x
2
1)
m
The optimum of (6.1) is

p(x) =
T
m
2x
1
T
m
1
= C
m
T
m
2x
1
The numerator is bounded by 1 for the range

1
x
2
. The factor C
m
can be
computed as
C
m
=
2c
m
1 + c
2m
with c =
2
+
1
Using the condition number we have
c 1
2
,
and
C
m
(1
2
)
m
Now, an error reduction by a factor of can be achieved in
N
its
log
1
steps. The original method by choosing m dierent relaxation parameters

k
is not a good
choice, since
it is not numerically stable
one has to know a priori the number of iterations
The recurrence relation for the Chebyshev polynomials leads to a practicable iterative
method called Chebyshev iteration.
The conjugate gradient method
The conjugate gradient algorithm automatically nds the optimal relaxation parameters
for the best k-step approximation.
Let p
0
, p
1
, . . . be a nite sequence of A-orthogonal vectos, and set
V
k
= spanp
0
, . . . p
k1
We want to approximate the solution u in the linear manifold u

0
+ V
k
:
min
vu
0
+V
k
|u v|
A
We represent u
k
as
u
k
= u
0
+
k1
l=0
l
p
l
The optimality criteria are
0 = (u u
k
, p
j
)
A
= (u u
0
k1
l=0
l
p
l
, p
j
)
A
0 j < k.
The coecients
l
follow from the A-orthogonality:
l
=
(u u
0
)
T
Ap
l
p
T
l
Ap
l
=
(f Au
0
)
T
p
l
p
T
l
Ap
l
The
l
are computable, since the A-inner product was chosen. The best approximations
can be computed recursively:
u
k+1
= u
k
+
k
p
k
Since u
k
u
0
V
k
, and p
k
A
V
k
, there holds
k
=
(f Au
k
)
T
p
k
p
T
k
Ap
k
.
Any k-step simple iteration approximates the solution u
k
in the manifold
u
0
+/
k
(d
0
)
with the Krylov space
/
k
(d
0
) = C
1
d
0
, C
1
AC
1
d
0
, . . . , C
1
(AC
1
)
k1
d
0
.
Here, d
0
= f Au
0
is the initial residual. The conjugate gradient method computes an
A-orthogonal basis of the Krylov-space. The term conjugate is equivalent to A-orthogonal.
Conjugate Gradient Algorithm:
Choose u
0
, compute d
0
= f Au
0
, set p
0
= C
1
d
0
.
for k = 0, 1, 2, . . . compute
k
=
d
T
k
p
k
p
T
k
Ap
k
u
k+1
= u
k
+
k
p
k
d
k+1
= d
k
k
Ap
k
k
=
d
T
k+1
C
1
Ap
k
p
T
k
Ap
k
p
k+1
= C
1
d
k+1
+
k
p
k
Remark 76. In exact arithmetic, the conjugate gradient algorithm terminates at a nite
number of steps k N.
Theorem 77. The conjugate gradient algorithm fullls for k k
1. The sequence p
k
is A-orthogonal. It spans the Krylov-space /
k
(d
0
)
2. The u
k
minimizes
min
vu
0
+|
k
(d
0
)
|u v|
A
3. There holds the orthogonality
d
T
k
p
l
= 0 l < k
Proof: Per induction in k. We assume
p
T
k
Ap
l
= 0 l < k
d
T
k
p
l
= 0 l < k
This is obvious for k = 0. We prove the property for k + 1: For l < k there holds
d
T
k+1
p
l
= (d
k
k
Ap
k
)
T
p
l
= d
T
k
p
l
k
p
T
k
Ap
l
= 0
per induction. For l = k there is
d
T
k+1
p
k
= (d
k
k
Ap
k
)
T
p
k
= d
T
k
p
k
d
T
k
p
k
p
T
k
Ap
k
p
T
k
Ap
k
= 0.
Next, prove the A-orthogonality of the p
k
. For l < k we have
(p
k+1
, p
l
)
A
= (C
1
d
k+1
+
k
p
k
, p
l
)
A
= d
T
k+1
C
1
Ap
l
There is
C
1
Ap
l
spanp
0
, . . . p
k
,
and d
T
k+1
p
j
= 0 for j k. For l = k there is
(p
k+1
, p
k
)
A
= (C
1
d
k+1
+
k
p
k
, p
k
)
A
= (C
1
d
k+1
, p
k
)
A
d
T
k+1
C
1
Ap
k
p
T
k
Ap
k
p
T
k
Ap
k
= 0
2
The coecients
k
and
k
should be computed by the equivalent, and numerically more
stable expressions
k
=
d
T
k
C
1
d
k
p
T
k
Ap
k
k
=
d
T
k+1
C
1
d
k+1
d
T
k
C
1
d
k
.
Theorem 78. The conjugate gradient iteration converges with the rate
|u u
k
|
A

+ 1
k
Proof: The conjugate gradient gives the best approximation in the Krylov space. Thus,
it can be bounded by the Chebyshev method leading to that rate.
The conjugate gradient iteration is stopped as soon as a convergence criterion is fullled.
Ideally, on wants to reduce the error in the energy norm by a factor :
|u u
k
|
A
|u u
0
|
A
But, the energy error cannot be computed. We rewrite
|u u
k
|
2
A
= |A
1
(f Au
k
)|
2
A
= |A
1
d
k
|
2
A
= d
T
k
A
1
d
k
If C is a good approxiamtion to A, then also C
1
is one to A
1
. The error can be
approximated by
d
T
k
C
1
d
k
.
This scalar is needed in the conjugate gradient iteration, nevertheless.
For solving the 2D model problem with C = I, the time complexity is
log
1
N
= log
1
n
3
The costs for one digit are comparable to the recursive sub-structuring algorithm. In 3D,
the conjugate gradient method has better time complexity.
6.3 Preconditioning
In the following, let the symmetric and positive denite matrix A arise from the nite
element discretization of the H
1
-elliptic and continuous bilinear-form A(., .). We construct
preconditioners C such that the preconditioning action
w = C
1
d
is eciently computable, and estimate the spectral bounds
1
u
T
Cu u
T
Au
2
u
T
Cu u R
N
The analysis of the preconditioner is performed in the nite element framework. For
this, dene the Galerkin isomorphism
G : R
N
V
h
: u u =
u
i
i
,
where
i
are the fe besis functions. Its dual is
G
: V
h
R
N
: d() (d(
i
))
i=1,...,N
.
To distinguish vectors and the corresponding nite element functions, we write vectors
u R
N
with underlines (when necessary).
The evaluation of the quadratic form is
u
T
Au = A(Gu, Gu) |Gu|
2
H
1
The Jacobi Preconditioner
The Jacobi preconditioner C is
C = diag A.
The preconditioning action is written as
C
1
d =
N
i=1
e
i
(e
T
i
Ae
i
)
1
e
T
i
d
Here, e
i
is the i
th
unit-vector. Thus, e
T
i
Ae
i
gives the i
th
diagonal element A
ii
of the matrix,
which is
A
ii
= A(
i
,
i
) |
i
|
2
H
1.
The quadratic form generated by the preconditioner is
u
T
Cu =
N
i=1
u
2
i
|
i
|
2
A

N
i=1
u
2
i
|
i
|
2
H
1
6.3. PRECONDITIONING 83
Theorem 79. Let h be the minimal mesh-size of a shape-regular triangulation. Then there
holds
h
2
u
T
Cu _ u
T
Au _ u
T
Cu (6.2)
Proof: We start to prove the right inequality
u
T
Au = |
i
u
i
i
|
2
A
_ u
T
Cu =
i
u
2
i
|
i
|
2
A
.
We dene the interaction matrix O with entries
O
ij
=
1 A(
i
,
j
) = 0
0 else
On a shape regular mesh, only a (small) nite number of basis functions have overlapping
support. Thus, O has a small number of entries 1 per row. There holds
|
i
u
i
i
|
2
A
=
j
u
i
u
j
A(
i
,
j
)
=
j
u
i
u
j
O
ij
A(
i
,
j
)
j
(u
i
|
i
|
A
)O
ij
(u
j
|
j
|
A
)
(O)
i
(u
i
|
i
|
A
)
2
= (O)u
T
Cu.
The spectral radius (O) = max
xR
N
x
T
Ox
|x|
2
is bounded by the (small) nite row-sum norm
of O.
The other estimate is proven element by element. Note that
u
T
Au |u|
2
H
1
()
=
T
|
i
u
i
i
|
2
H
1
(T)
and
u
T
Cu
i
|u
i
i
|
2
H
1
()
=
i
|u
i
i
|
2
H
1
(T)
.
We prove the inequality for each individual element. The triangle T has diameter h
T
. On
T, we expand u in terms of the element shape functions
, namely u[
T
=
3
=1
u
.
We transform to the reference element

T:
|
|
2
H
1
(T)
= |
|
2
L
2
(T)
+|
|
2
L
2
(T)
h
2
T
|
|
2
L
2
(
b
T)
+|
|
2
L
2
(
b
T)
h
2
T
|
|
2
L
2
(
b
T)
and
|u
|
2
H
1
(T)
=
|u
|
2
L
2
(T)
+
|u
|
2
L
2
(T)
h
2
T
|u
|
2
L
2
(
b
T)
+
|u
|
2
L
2
(
b
T)
_
|u
|
2
L
2
(
b
T)
Both, (u)
|
L
2
(
b
T)
and u
|u
|
2
L
2
(
b
T)
1/2
are norms on R
3
. Since all
norms in R
3
are equivalent, we have
|u
|
2
H
1
(T)
_ h
2
T
|
|
2
H
1
(T)
. (6.3)
By summing over all elements and choosing h = min h
T
, we have proven the left inequality
of (6.2). 2
Remark: Inequality (6.3) is sharp. To prove this, choose u
= 1.
Block-Jacobi preconditioners
Instead of choosing the diagonal, one can choose a block-diagonal of A, e.g.,
In the case of systems of PDEs, choose blocks consisting of all degrees of freedom
sitting in one vertex. E.g., mechanical deformations (u
x
, u
y
, u
z
).
For high order elements, choose blocks consisting of all degrees of freedom associated
to the edges (faces, inner) of the elements.
On anisotropic tensor product meshes, choose blocks consisting of unknowns in the
short direction
Domain decomposition methods: Choose blocks consisting of the unknowns in a
sub-domain
Decompose the unknowns into M blocks, the block i has dimension N
i
. Dene the
rectangular embedding matrices
E
i
R
NN
i
i = 1, . . . , M.
E
i
consists of N
i
unit vectors corresponding to the unknowns in the block i. Each u R
N
can be uniquely written as
u =
M
i=1
E
i
u
i
with u
i
R
N
i
The diagonal blocks are
A
i
= E
T
i
AE
i
i = 1, . . . , M.
The block Jacobi preconditioner is
C
1
d =
M
i=1
E
i
A
1
i
E
T
i
d
The quadratic form induced by C can be written as
u
T
Cu =
i
u
T
i
A
i
u
i
=
i
|GE
i
u
i
|
2
A
where u =
E
i
u
i
.
Example: Discretize the unit interval I = (0, 1) into n elements of approximate size
h 1/n. Split the unknowns into two blocks, the left half and the right half, and dene
the corresponding block-Jacobi preconditioner.
Set
I = I
1
T
n/2
I
2
,
with I
1
= (0, x
n/2
), T
n/2
= [x
n/2
, x
n/2+1
], and I
2
= (x
n/2+1
, 1). Decompose
u = E
1
u
1
+ E
2
u
2
.
The corresponding nite element functions are u
i
= GE
i
u
i
. There holds
u
1
(x) =
Gu(x) x I
1
linear x T
0 x I
2
,
and u
2
vice versa. The quadratic form is
u
T
Cu =
i
u
i
A
i
u
i
=
i
|GE
i
u
i
|
2
A
Evaluation gives
|u
1
|
2
A
= |u
1
|
2
H
1
(I
1
)
+|u
1
|
2
H
1
(T)
|u
1
|
2
H
1
(I
1
)
+ h
1
[u(x
n/2
)[
2
_ |u|
2
H
1
(I)
+ h
1
|u|
2
H
1
(I)
(trace theorem)
h
1
|u|
2
A
,
and thus
u
T
Cu =
i
|u
i
|
2
A
_ h
1
|u|
2
A
h
1
u
T
Au.
The situation is the same in R
d
.
Exercise: Sub-divide the interval I into M sub-domains of approximative size H 1/M.
What are the sprectral bounds of the block-Jacobi preconditioner ?
Additive Schwarz preconditioners
The next generalization is an overlapping block Jacobi preconditioner. For i = 1, . . . , M
let E
i
R
NN
i
be rectangular matrices such that each u R
N
can be (not necessarily
uniquely) written as
u =
M
i=1
E
i
u
i
with u
i
R
N
i
Again, the overlapping block-Jacobi preconditioning action is
C
1
d =
M
i=1
E
i
A
1
i
E
T
i
d
Example: Choose the unit-interval problem from above. The block 1 contains all nodes in
(0, 3/4), and the block 2 contains nodes in (1/4, 1). The blocks overlap, the decomposition
is not unique.
The columns of the matrices E
i
are not necessarily unit-vectors, but are linearly in-
dependent. In this general setting, the preconditioner is called Additive Schwarz precon-
ditioner. The following lemma gives a useful representation of the quadratic form. It
was proven in similar forms by many authors (Nepomnyaschikh, Lions, Dryja+Widlund,
Zhang, Xu, Oswald, Griebel, ...) and is called also Lemma of many fathers, or Lions
Lemma:
Lemma 80 (Additive Schwarz lemma). There holds
u
T
Cu = inf
u
i
R
N
i
u=
P
E
i
u
i
M
i=1
u
T
i
A
i
u
i
Proof: The right hand side is a constrained minimization problem of a convex function.
The feasible set is non-empty, the CMP has a unique solution. It is solved by means of
Lagrange multipliers. Dene the Lagrange-function (with Lagrange multipliers R
N
):
L((u
i
), ) =
u
T
i
Au
i
+
T
(u
E
i
u
i
).
Its stationary point (a saddle point) is the solution of the CMP:
0 =
u
i
L((u
i
), ) = 2A
i
u
i
+ E
T
i

0 =
L((u
i
), ) = u
E
i
u
i
The rst line gives
u
i
=
1
2
A
1
i
E
T
i
.
Use it in the second line to obtain
0 = u
1
2
E
i
A
1
i
E
i
= u
1
2
C
1
,
i.e., = 2Cu, and
u
i
= A
1
i
E
T
i
Cu.
The minimal value is
u
T
i
A
i
u
i
=
u
T
CE
i
A
1
i
A
i
A
1
i
E
T
i
Cu
=
u
T
CE
i
A
1
i
E
T
i
Cu
= u
T
CC
1
Cu = u
T
Cu
2
Next, we rewrite the additive Schwarz iteration matrix
I C
1
A = I
M
i=1
E
i
A
1
i
E
T
i
A
in the fe framework. Let
V
i
= GE
i
R
N
i
V
h
be the sub-space corresponding to the range of E
i
, and dene the A-orthogonal projection
P
i
: V
h
V
i
: A(P
i
u, v
i
) = A(u, v
i
) v
i
V
i
Lemma 81. Set u = Gu, the application of the iteration matrix is u = (I C
1
A)u, and
set u = G u. Then there holds
u =
I
M
i=1
P
i
u.
Proof: Let w
i
= A
1
i
E
T
i
Au. Then
u = u GE
i
w
i
.
There holds w
i
:= GE
i
w
i
V
i
, and
A(GE
i
w
i
, GE
i
v
i
) = v
T
i
E
T
i
AE
i
w
i
= v
T
i
A
i
w
i
= v
T
i
E
T
i
Au
= A(Gu, GE
i
v
i
) v
i
R
N
i
,
i.e., w
i
= P
i
u. 2
The additive Schwarz preconditioner is dened by the space splitting
V =
M
i=1
V
i
If the spaces V
i
are A-orthogonal, then
i
P
i
= I, and (with = 1), and the iteration
matrix is M = 0.
The reformulation of the additive Schwarz lemma 80 in the nite element framework is
Lemma 82 (Additive Schwarz lemma). Let u = Gu. There holds
u
T
Cu = inf
u
i
V
i
u=
P
u
i
M
i=1
|u
i
|
2
A
Example: Let
A(u, v) =
1
0
u
t
v
t
+
uv dx
with 0 < 1. The bilinear-form is H
1
-elliptic and continuous, but the bounds depend
on the parameter . Let C
J
be the Jacobi preconditioner. The proof of Theorem 79 shows
that
h
2
u
T
C
J
u _ u
T
Au _ u
T
C
J
u.
The non-robust lower bound is sharp: Take u = (1, . . . , 1)
T
.
The solution is to add the additional sub-space
V
0
= span1 = GE
0
R
1
to the AS preconditioner (with E
0
R
N1
consisting of 1-entries). The preconditioning
action is
C
1
d = diagA
1
d + E
0
(E
T
0
AE
0
)
1
E
T
0
d.
The spectral bounds are robust in :
h
2
u
T
Cu _ u
T
Au _ u
T
Cu,
namely
u
T
Cu = inf
u
i
V
i
u=
P
M
0
u
i
M
i=0
|u
i
|
2
A
= inf
u
0
V
0
|u
0
|
2
A
+ inf
u
i
V
i
uu
0
=
P
M
1
u
i
M
i=1
|u
i
|
2
A
_ inf
u
0
V
0
|u
0
|
2
A
+ h
2
|u u
0
|
2
H
1
The last step was the result of the Jacobi preconditioner applied to (u, v)
H
1. Finally, we
choose u
0
=
1
0
u dx to obtain
u
T
Cu _ |u
0
|
2
A
+ h
2
|u u
0
|
2
H
1
_ |u
0
|
2
L
2
+ h
2
|(u u
0
)|
2
L
2
_ |u|
2
L
2
+ h
2
|u|
2
L
2
= h
2
|u|
2
A
Overlapping domain decomposition preconditioning
Let =
M
i=1
i
be a decomposition of into M sub-domains of diameter H. Let

i
be
such that
i
dist
i
` ,
i
_ H,
and only a small number of

i
are overlapping. Choose a nite element mesh of mesh
size h H, and the nite element space is V
h
. The overlapping domain decomposition
preconditioner is the additive Schwarz preconditioner dened by the sub-space splitting
V
h
=
V
i
with V
i
= V
h
H
1
0
(
i
).
The bilinear-form A(., .) is H
1
-elliptic and continuous. The implementation takes the sub-
matrices of A with nodes inside the enlarged sub-domains

i
.
Lemma 83. The overlapping domain decomposition preconditioner fullls the spectral es-
timates
H
2
u
T
Cu _ uAu _ u
T
Cu.
Proof: The upper bound is generic. For the lower bound, we construct an explicit
decomposition u =
u
i
.
There exists a partition of unity
i
such that
0
i
1, supp
i

i
,
M
i=1
i
= 1
and
|
i
|
L
_ H
1
.
Let
h
: L
2
V
h
be a Clement-type quasi-interpolation operator such that
h
is a
projection on V
h
, and
|
h
v|
L
2
_ |v|
L
2
, and |
h
v|
L
2
_ |v|
L
2
.
For given u V
h
, we choose the decomposition
u
i
=
h
(
i
u).
Indeed u
i
V
i
is a decomposition of u V
h
:
u
i
=
h
(
i
u) =
h
i
)u
=
h
u = u
The lower bound follows from
u
T
Cu = inf
u=
P
v
i
i
|v
i
|
2
A
i
|u
i
|
2
A
_
i
|u
i
|
2
H
1
=
i
|
h
(
i
u)|
2
H
1
_
i
|
i
u|
2
H
1
=
|
i
u|
2
L
2
(
e
i
)
+|(
i
u)|
2
L
2
(
e
i
)
|
i
u|
2
L
2
(
e
i
)
+|(
i
)u|
2
L
2
(
e
i
)
+|
i
u|
2
L
2
(
e
i
)
|u|
2
L
2
(
e
i
)
+ H
2
|u|
2
L
2
(
e
i
)
+|u|
2
L
2
(
e
i
)
_ |u|
2
L
2
()
+ H
2
|u|
2
L
2
()
+|u|
2
L
2
()
_ H
2
|u|
2
A
.
2
Overlapping DD preconditioning with coarse grid correction
The local DD preconditioner above gets worse, if the number of sub-domains increases. In
the limit, if H h, the DD preconditioner is comparable to the Jacobi preconditioner.
To overcome this degeneration, we add one more subspace. Let T
H
be a coarse mesh of
mesh-size H, and T
h
is the ne mesh generated by sub-division of T
H
. Let V
H
be the nite
element space on T
H
. The sub-domains of the domain decomposition are of the same size
as the coarse grid.
The sub-space decomposition is
V
h
= V
H
+
M
i=1
V
i
.
Let G
H
: R
N
H
V
H
be the Galerkin isomorphism on the coarse grid, i.e.,
G
H
u
H
=
N
H
i=1
u
H,i
H
i
The coarse space fullls V
H
V
h
. Thus, every coarse grid basis
H
i
can be written as
linear combination of ne grid basis functions
h
j
:
H
i
=
N
j=1
E
H,ji
h
j
.
Example:
H h
V V
The rst basis function
H
1
is
H
1
=
h
1
+
1
2
h
2
The whole matrix is
E
H
=
1
1/2 1/2
1
1/2 1/2
1
.
There holds
G
H
u
H
= G
h
E
H
u
H
.
Proof:
G
H
u
H
=
N
H
i=1
u
H,i
H
i
=
N
H
i=1
N
h
j=1
u
H,i
E
H,ji
h
j
=
N
h
j=1
h
j
(E
H
u
H
)
j
= GEu
H
The matrix E
H
transforms the coecients u
H
w.r.t. the coarse grid basis to the coecients
u
h
= E
H
u
H
w.r.t. the ne grid basis. It is called prolongation matrix.
The DD preconditioner with coarse grid correction is
C
1
d =
i
E
i
A
1
i
E
T
i
d + E
H
(E
T
H
AE
H
)
1
E
T
H
d
The rst part is the local DD preconditioner from above. The second part is the coarse
grid correction step. The matrix E
T
H
(called restriction matrix) transfers the defect d from
the ne grid to a defect vector on the coarse grid. Then, the coarse grid problem with
matrix E
T
H
AE
H
is solved. Finally, the result is prolongated to the ne grid.
The matrix A
H
:= E
T
H
AE
H
is the Galerkin matrix w.r.t. the coarse grid basis:
A
H,ij
= e
T
j
E
T
H
AE
H
e
i
= A(G
h
E
H
e
i
, G
h
E
H
e
j
)
= A(G
H
e
i
, G
H
e
j
) = A(
H
i
,
H
j
).
Theorem 84. The overlapping domain decomposition preconditioner with coarse grid sys-
tem fullls the optimal spectral estimates
u
T
Cu _ u
T
Au _ u
T
Cu.
Proof: The quadratic form generated by the preconditioner is
u
T
Cu = inf
u
H
V
H
,u
i
V
i
u=u
H
+
P
u
i
|u
H
|
2
A
+
M
i=1
|u
i
|
2
A
.
Again, the upper bound u
T
Au _ u
T
Cu follows from the nite overlap of the spaces
V
H
, V
1
, . . . V
M
. To prove the lower bound, we come up with an explicit decomposition.
We split the minimization into two parts:
u
T
Cu = inf
u
H
V
H
inf
u
i
V
i
uu
H
=
P
u
i
|u
H
|
2
A
+
M
i=1
|u
i
|
2
A
(6.4)
In the analysis of the DD precondition without coarse grid system we have observed that
inf
u
i
V
i
uu
H
=
P
u
i
M
i=1
|u
i
|
2
A
_ H
2
|u u
H
|
2
L
2
+|(u u
H
)|
2
L
2
Using this in (6.4) gives
u
T
Cu _ inf
u
H
V
H
|u
H
|
2
A
+ H
2
|u u
H
|
2
L
2
+|(u u
H
)|
2
L
2
_ inf
u
H
V
H
|u
H
|
2
L
2
+ H
2
|u u
H
|
2
L
2
+|u|
2
L
2
To continue, we introduce a Clement operator

H
: H
1
V
H
being continuous in the
H
1
-semi-norm, and approximating in L
2
-norm:
|
H
u|
2
L
2
+ H
2
|u
H
u|
2
L
2
_ |u|
2
L
2
Choosing now u
H
:=
H
u in the minimization problem we obtain the result:
u
T
Cu _ |
H
u|
2
A
+ H
2
|u
H
u|
2
L
2
+|u|
2
L
2
_ |u|
2
|u|
2
A
2
The inverse factor H
2
we have to pay for the local decomposition could be compensated
by the approximation on the coarse grid.
The costs for the setup depend on the underlying direct solver for the coarse grid
problem and the local problems. Let the factorization step have time complexity N
. Let
N be the number of unknowns at the ne grid, and M the number of sub-domains. Then
the costs to factor the coarse grid problem and the M local problems are of order
M
+ M
N
M
Equilibrating both terms gives the optimal choice of number of sub-domains

M = N

21
,
and the asymptotic costs
N

2
21
.
Example: A Cholesky factorization using bandwidth optimization for 2D problems has
time complexity N
2
. The optimal choice is M = N
2/3
, leading to the costs of
N
4/3
.
Multi-level preconditioners
The preconditioner above uses two grids, the ne one where the equations are solved,
and an articial coarse grid. Instead of two grids, one can use a whole hierarchy of grids
T
0
, T
1
, . . . , T
/
= T . The according nite element spaces are
V
0
V
1
. . . V
L
= V
h
.
Let E
l
be the prolongation matrix from level l to the nest level L. Dene
A
l
= E
T
l
AE
l
and D
l
= diagA
l
.
Then, the multi-level preconditioner is
C
1
= E
0
A
1
0
E
T
0
+
L
l=1
E
l
D
1
l
E
T
l
The setup, and the application of the preconditioner takes O(N) operations. One can
show that the multi-level preconditioner fullls optimal spectral bounds
u
T
Cu _ u
T
Au _ u
T
Cu.
An iterative method with multi-level preconditioning solves the matrix equation Au = f
of size N with O(N) operations !
Chapter 7
Mixed Methods
A mixed method is a variational formulation involving two function spaces, and a bilinear-
form of a special saddle point structure. Usually, it is obtained from variational problems
with constraints.
7.1 Weak formulation of Dirichlet boundary condi-
tions
We start with the Poisson problem
u = f in , (7.1)
and boundary conditions
u = u
D
on
D
,
u
n
= 0 on
N
.
In contrast to the earlier method, we multiply equation (7.1) with test functions v H
1
(without imposing Dirichlet constraints), and integrate by parts. Using the Neumann
boundary conditions, we obtain
uv dx
D
u
n
v ds =
fv dx
The normal derivative
u
n
is not known on
D
. We simply call it :
:=
u
n
To pose the Dirichlet boundary condition, we multiply u = u
D
by suciently many test
functions, and integrate over
D
:
D
uds =
D
u
D
ds ?
95
96 CHAPTER 7. MIXED METHODS
Combining both equations, we get the system of equations: Find u V = H
1
() and
Q =? such that
u v dx +
D
vds =
fv dx v V,
D
uds =
D
u
D
ds Q.
(7.2)
A similar formulation can be obtained for interface conditions.
7.2 A Mixed method for the ux
We start from the second order pde
div(au) = f in ,
u = u
D
on
D
a
u
n
= g on
N
Next, we introduce the ux variable := au to rewrite the equations as: Find u and
such that
a
1
u = 0, (7.3)
div = f, (7.4)
u = u
D
on
D
n = g on
N
.
We want to derive a variational formulation for the system of equations. For that, we
multiply the rst equations by vector-valued test functions , the second equation by test
functions v, and integrate:
(a
1
) dx
u dx = 0
div v dx =
fv dx v
We would like to have the second term of the rst equation of the same structure as the
rst term in the second equation. This can be obtained by integration by parts applied to
either one of them. The interesting case is to integrate by parts in the rst line to obtain:
a dx +
div u dx
n
u ds
n
u ds = 0.
7.3. ABSTRACT THEORY 97
Here, we make use of the boundary conditions. On the Dirichlet boundary, we know
u = u
D
, and use that in the equation. The Neumann boundary condition n = g must
be put into the approximation space, it becomes an essential boundary condition. Thus,
it is enough to choose test functions of the sub-space fullling n = 0. The problem is
now the following. The space V will be xed later. Find V,
n
= g on
N
, and u Q
such that
(a
1
) dx +
div u dx =
D
u
D
n
ds ,
n
= 0 on
N
div v dx =
fv dx v
The derivatives are put onto the ux unknown (and its test function ). We dont have
to derive the primal unknown u. This will give us better approximation for the uxes than
for the scalar. That is one of the reasons to use this mixed method.
7.3 Abstract theory
A mixed variational formulation involves two Hilbert spaces V and Q, bilinear-forms
a(u, v) : V V R,
b(u, q) : V Q R,
and continuous linear-forms
f(v) : V R,
g(q) : Q R.
The problem is to nd u V and p Q such that
a(u, v) + b(v, p) = f(v) v V,
b(u, q) = g(q) q Q.
(7.5)
The two examples from above are of this form.
Instead of considering this as a system of equations, one can look at the mixed method
as one variational problem on the product spaces V Q. For this, simply add both lines,
and search for (u, p) V Q such that
a(u, v) + b(u, q) + b(v, p) = f(v) + g(q) (v, q) V Q.
Dene the big bilinear-form B(., .) : (V Q) (V Q) R as
B((u, p), (v, q)) = a(u, v) + b(u, q) + b(v, p),
to write the whole system as single variational problem
Find (u, p) V Q : B((u, p), (v, q)) = f(v) + g(q) (v, q) V Q
By the Riesz-representation theorem, we can dene operators:
A : V V : u Au : (Au, v)
V
= a(u, v) v V
B : V Q : u Bu : (Bu, q)
Q
= b(u, q) q Q
B
: Q V : p B
p : (B
p, v)
V
= b(v, p) v V.
By means of these operators, we can write the mixed variational problem as operator
equation
Au + B
p = J
V
f,
Bu = J
Q
g.
(7.6)
Here, we used the Riesz-isomorphisms J
V
: V
V and J
Q
: Q
Q.
In the interesting examples, the operator B has a large kernel:
V
0
:= v : Bv = 0
Lemma 85. Assume that B
Q is closed in V . Then there holds the V -orthogonal decom-

position
V = V
0
+ B
Q
Proof: There holds
V
0
= v : Bv = 0
= v : (Bv, q)
Q
= 0 q Q
= v : (v, B
q)
V
= 0 q Q.
This means, V
0
is the V -orthogonal complement to B
Q. 2
Now, we will give conditions to ensure a unique solution of a mixed problem:
Theorem 86 (Brezzis theorem). Assume that a(., .) and b(., .) are continuous bilinear-
forms
a(u, v)
2
|u|
V
|v|
V
u, v V, (7.7)
b(u, q)
2
|u|
V
|q|
Q
u V, q Q. (7.8)
Assume there holds coercivity of a(., .) on the kernel,i.e.,
a(u, u)
1
|u|
2
V
u V
0
, (7.9)
and there holds the LBB (Ladyshenskaja-Babuska-Brezzi) condition
sup
uV
b(u, q)
|u|
V

1
|q|
Q
q Q. (7.10)
Then, the mixed problem is uniquely solvable. The solution fullls the stability estimate
|u|
V
+|p|
Q
c|f|
V
+|g|
Q
,
with the constant c depending on
1
,
2
,
1
,
2
.
7.3. ABSTRACT THEORY 99
Proof: The big bilinear-form B(., .) is continuous
B((u, p), (v, q)) _ (|u| +|p|) (|v| +|q|).
We prove that it fullls the inf sup condition
inf
v,q
sup
u,p
B((u, p), (v, q))
(|v|
V
+|q|
Q
)(|u|
V
+|p|
Q
)
.
Then, we use Theorem 33 (by Babuska-Aziz) to conclude continuous solvability.
To prove the inf sup-condition, we choose arbitrary v V and q Q. We will
construct u V and p Q such that
|u|
V
+|p|
Q
_ |v|
V
+|q|
Q
and
B((u, p), (v, q)) = |v|
2
V
+|q|
2
Q
.
First, we use (7.10) to choose u
1
V such that
b(u
1
, q) = |q|
2
Q
and |u
1
|
V
2
1
1
|q|
Q
.
Next, we solve a problem on the kernel:
Find u
0
V
0
: a(u
0
, w
0
) = (v, w
0
)
V
a(u
1
, w
0
) w
0
V
0
Due to assumption (7.9), the left hand side is a coercive bilinear-form on V
0
. The right
hand side is a continuous linear-form. By Lax-Milgram, the problem has a unique solution
fullling
|u
0
|
V
_ |v|
V
+|u
1
|
V
We set
u = u
0
+ u
1
.
By the Riesz-isomorphism, we dene a z V such that
(z, w)
V
= (v, w)
V
a(u, w) w V
By construction, it fullls z
V
V
0
. The LBB condition implies
|p|
Q

1
1
sup
v
b(v, p)
|v|
V
=
1
1
sup
v
(v, B
p)
V
|v|
V
=
1
1
|B
p|
V
,
and thus B
Q is closed, and z B
Q. Take the p Q such that

z = B
p.
It fullls
|p|
Q

1
1
|z|
V
_ |v|
V
+|q|
Q
Concluding, we have constructed u and p such that
|u|
V
+|p|
Q
_ |v|
V
+|q|
Q
,
and
B((u, p), (v, q)) = a(u, v) + b(v, p) + b(u, q)
= a(u, v) + (z, v)
V
+ b(u, q)
= a(u, v) + (v, v)
V
a(u, v) + b(u, q)
= |v|
2
V
+ b(u
1
, q)
= |v|
2
V
+|q|
2
Q
.
2
7.4 Analysis of the model problems
Now, we apply the abstract framework to the two model problems.
Weak formulation of Dirichlet boundary conditions
The problem is well posed for the spaces
V = H
1
() and Q = H
1/2
(
D
)
Remember, H
1/2
(
D
) is the dual to H
1/2
(
D
). The later one is the trace space of H
1
(),
the norm fullls
|u
D
|
H
1/2
(
D
)
inf
wH
1
tr w=u
D
|w|
H
1
()
.
The bilinear-forms are
a(u, v) =
uv dx
b(u, ) = ', tr u`
H
1/2
H
1/2
To be precise, the integral
D
u dx is extended to the duality product ', u`. For regular
functions ( L
2
(
D
)), we can write the L
2
-inner product.
Theorem 87. The mixed problem (7.2) has a unique solution u H
1
() and
H
1/2
(
D
).
Proof: The spaces V and Q, and the bilinear-forms a(., .) and b(., .) fulll the assump-
tions of Theorem 86. The kernel space V
0
is
V
0
= u :
D
udx = 0 L
2
(
D
) = u : tr
D
u = 0
7.4. ANALYSIS OF THE MODEL PROBLEMS 101
The continuity of a(., .) on V is clear. It is not coercive on V , but, due to Friedrichs
inequality, it is coercive on V
0
.
The bilinear-form b(., .) is continuous on V Q:
b(u, ) = ', tr u`
H
1/2
H
1/2
||
H
1/2| tr u|
H
1/2
(
D
)
_ ||
Q
|u|
H
1 = ||
Q
|u|
V
The LBB - condition of b(., .) follows more or less from the denition of norms:
|q|
Q
= sup
uH
1/2
'q, u`
|u|
H
1/2
sup
uH
1/2
'q, u`
inf
wH
1
()
tr w=u
|w|
H
1
()
= sup
uH
1/2
sup
wH
1
()
tr w=u
'q, u`
|w|
H
1
= sup
wH
1
'q, tr w`
|w|
H
1
= sup
wV
b(w, q)
|w|
V
Mixed method for the uxes
This mixed method requires the function space H(div, ):
Denition 88. A measurable function g is called the weak divergence of on R
d
if
there holds

g dx =
dx C
0
()
The function space H(div) is dened as
H(div, ) := [L
2
()]
d
: div L
2
,
its norms is
||
H(div)
=
||
2
L
2
+| div |
2
L
2
1/2
The mixed method is formulated on the spaces
V = H(div) Q = L
2
The bilinear-forms are
a(, ) =
a
1
dx , V
b(, v) =
div v dx V, v Q
We assume that the symmetric matrix a R
dd
and its inverse a
1
are bounded.
Theorem 89. The mixed problem for the uxes is well posed.
Proof: We check the conditions of the theorem of Brezzi: The bilinear-forms are
bounded, namely
a(, ) =
a
1
dx |a
1
|
L
||
L
2
||
L
2
_ ||
V
||
V
and
b(, v) =
div v dx | div |
L
2
|v|
L
2
||
V
|v|
Q
.
The kernel space V
0
= : b(, v) = 0 v Q is
V
0
= H(div) : div = 0
There holds the kernel-ellipticity of a(., .). Let V
0
. Then
a(, ) =

T
a
1
dx inf
x
min
(a
1
)
[[
2
dx _ ||
2
L
2
= ||
2
H(div)
We are left to verify the LBB condition
sup
H(div)
div v dx
||
H(div)
_ |v|
L
2
v L
2
. (7.11)
For given v L
2
, we will construct a ux satisfying the inequality. For this, we solve
the articial Poisson problem = v with Dirichlet boundary conditions = 0 on .
The solution satises ||
L
2
_ |v|
L
2
. Set = . There holds div = v. Its norm is
||
2
H(div)
= ||
2
L
2
+| div |
2
L
2
= ||
2
L
2
+|v|
2
L
2
_ |v|
2
L
2
.
Using it in (7.11), we get the result
div v dx
||
H(div)
=
v
2
dx
||
H(div)
_ |v|
L
2
.
The function space H(div)
The mixed formulation has motivated the denition of the function space H(div). Now,
we will study some properties of this space. We will also construct nite elements for
the approximation of functions in H(div). In Section 3.3.1, we have investigated traces of
functions in H
1
. Now, we apply similar techniques to the space H(div). Again, the proofs
are based on the density of smooth functions.
For a function in H
1
, the boundary values are well dened by the trace operator. For
a vector-valued function in H(div), only the normal-component is well dened on the
boundary:
Theorem 90. There exists a normal-trace operator
tr
n
: H(div) H
1/2
()
such that for H(div) [C()]
d
it coincides with its normal component
tr
n
= n on .
Proof: For smooth functions, the trace operator gives the normal component on the
boundary. We have to verify that this operator is bounded as operator from H(div) to
H
1/2
(). Then, by density, we can extend the trace operator to H(div). Let
H(div) [C
1
()]
d
:
| tr
n
|
H
1/2 = sup
H
1/2
()
nds
||
H
1/2
sup
H
1
()
n tr ds
||
H
1
= sup
H
1
()
( tr ) nds
||
H
1
= sup
H
1
()
div() dx
||
H
1
= sup
H
1
()
(div )dx +
dx
||
H
1
sup
H
1
()
| div |
L
2
||
L
2
+||
L
2
||
L
2
||
H
1
||
2
L
2
+| div |
2
L
2
1/2
= ||
H(div)
2
Lemma 91. There holds integration by parts
dx +
(div )dx = 'tr

n
, tr `
H
1/2
H
1/2
for all H(div) and H
1
().
Proof: By density of smooth functions, and continuity of the trace operators.
Now, let
1
, . . .
M
be a non-overlapping partitioning of . In Section 3.3.1, we have
proven that functions which are in H
1
(
i
), and which are continuous across the boundaries
ij
=
i

j
, are in H
1
(). A similar property holds for functions in H(div).
Theorem 92. Let [L
2
()]
d
such that
i
H(div,
i
)
tr
n,i
[
i
= tr
n,j
[
j
on
ij
.
Then H(div, ), and
(div )[
i
= div ([
i
).
The proof follows the lines of Theorem 46.
We want to compute with functions in H(div). For this, we need nite elements for
this space. The characterization by sub-domains allows the denition of nite element
sub-spaces of H(div). Let T = T be a triangulation of . One family of elements are
the BDM (Brezzi-Douglas-Marini) elements. The space is
V
h
= [L
2
]
2
: [
T
[P
k
]
d
, n continuous across edges.
This nite element space is larger than the piece-wise polynomial H
1
-nite element space
of the same order. The nite element functions can have non-continuous tangential com-
ponents across edges.
The cheapest element for H(div) is the lowest order Raviart-Thomas element RT0.
The nite element (T, V
T
,
i
) is dened by the space of shape functions V
T
, and linear
functionals
i
. The element space is
V
T
=
a
b
+ c
x
y
: a, b, c R
,
the linear functionals are the integrals of the normal components on the three edges of the
triangle
i
() =
e
i
nds i = 1, 2, 3
The three functionals are linearly independent on V
T
. This means, for each choice of
1
,
2
,
3
, there exists three unique numbers a, b, c R such that
=
a
b
+ c
x
y
.
satises
i
() =
i
.
Exercise: Compute the shape functions for the RT0 - reference triangle.
The global nite element functions are dened as follows. Given one value
i
for each
edge e
i
of the triangulation. The corresponding RT0 nite element function is dened
by
[
T
V
T
and
e
i
[
T
n
e
i
ds =
i
for all edges e
i
T and all triangles T T .
We have to verify that this construction gives a function in H(div). For each ele-
ment, [
T
is a linear polynomial, and thus in H(div, T). The normal components must be
continuous. By construction, there holds
e
[
T,i
nds =
e
[
T,j
nds
for the edge e = T
i
T
j
. The normal component is continuous since n
e
is constant on
an edge: Points (x, y) on the edge e fulll xn
x
+ yn
y
is constant. There holds
n
e
=
a
b
+ c
x
y
n
x
n
y
= an
x
+ bn
y
+ c(xn
x
+ yn
y
) = constant
The global RT0-basis functions
RT
i
are associated to the edges, and satisfy
e
i
RT
j
n
e
ds =
ij
i, j = 1, . . . N
edges
By this basis, we can dene the RT - interpolation operator
I
RT
h
=
edges e
i
e
i
n
e
ds
RT
i
It is a projection on V
h
. The interpolation operator preserves the divergence in mean:
Lemma 93. The RT0 - interpolation operator satises
T
div I
h
dx =
T
div dx
for all triangles T T .
Let P
h
be the L
2
projection onto piece-wise constant nite element functions. This
is: Let Q
h
= q L
2
: q[
T
= const T T . Then P
h
p is dened by P
h
p Q
h
and
P
h
p q
h
dx =
Pp q
h
dx q
h
Q
h
. This is equivalent to P
h
p satises P
h
p Q
h
and
T
P
h
p dx =
T
p dx T T .
The Raviart-Thomas nite elements are piecewise linear. Thus, the divergence is piece-
wise constant. From div I
h
Q
h
and Lemma 93 there follows
div I
h
= P
h
div .
This relation is known as commuting diagram property:
H(div)
div
L
2
I
h
P
h
V
RT
h
div
Q
h
(7.12)
The analysis of the approximation error is based on the transformation to the reference
element. For H
1
nite elements, interpolation on the element T is equivalent to interpo-
lation on the reference element

T, i.e., (I
h
v) F
T
=

I
h
(v F
T
). This is not true for the
H(div) elements: The transformation F changes the direction of the normal vector. Thus
e
nds =
e
nds.
The Piola transformation is the remedy:
Denition 94 (Piola Transformation). Let F :

T T be the mapping from the reference
element

T to the element T. Let L
2
(
T). Then, the Piola transformation

= {
is dened by
(F( x)) = (det F
t
)
1
F
t
(x).
The Piola transformation satises:
Lemma 95. Let H(div,

T), and = { . Then there holds
(div )(F( x)) = (det F
t
)
1
div
Let e be an edge of the reference element, and e = F( e). Then
e
nds =
e
nds
Proof: Let C
0
(
T), and (F( x)) = (x). Then there holds
T
div dx =
T
dx
=
b
T
(det F
t
)
1
F
t
(F
t
)
T

(det F
t
) d x
=
b
T
d x =
b
T
div d x
=
T
(det F
t
)
1
(div )dx.
Since C
0
is dense in L
2
(T), there follows the rst claim. To prove the second one, we show
that
e
( n)ds =
e
( n) dx
holds for all C
(T), = 0 on T ` e. Then, let 1 on the edge e:
e
( n)ds =
T
div() dx =
b
T
div( ) d x =
e
( n) d s.
2
Lemma 96. The Raviart-Thomas triangle T and the Raviart-Thomas reference triangle
are interpolation equivalent:
I
RT
h
{ = {
I
RT
h

Proof: The element spaces are equivalent, i.e., V
T
= {V
b
T
, and the functionals
i
() =
e
nds are preserved by the Piola transformation.
7.5. APPROXIMATION OF MIXED SYSTEMS 107
Theorem 97. The Raviart-Thomas interpolation operator satises the approximation
properties
| I
RT
h
|
L
2
()
_ h||
L
2
()
| div div I
RT
h
|
L
2
()
_ h|div |
L
2
()
Proof: Transformation to the reference element, using that the interpolation preserves
constant polynomials, and the Bramble Hilbert lemma. The estimate for the divergence
uses the commuting diagram property
| div(I I
RT
h
)|
L
2
= |(I P
h
) div |
L
2
_ h| div |
L
2
2
7.5 Approximation of mixed systems
We apply a Galerkin-approximation for the mixed system. For this, we choose (nite
element) sub-spaces V
h
V and Q
h
Q, and dene the Galerkin approximation (u
h
, p
h
)
V
h
Q
h
by
B((u
h
, p
h
), (v
h
, q
h
)) = f(v
h
) + g(q
h
) v
h
V
h
q
h
Q
h
.
Theorem 98. Assume that the nite element spaces fulll the discrete stability condition
inf
vV
h
,qQ
h
sup
uV
h
,pQ
h
B((u, p), (v, q))
(|v|
V
+|q|)(|u|
V
+|p|
Q
)
. (7.13)
Then the discretization error is bounded by the best-approximation error
|u u
h
|
V
+|p p
h
|
Q
_ inf
v
h
V
h
,q
h
Q
h
|u v
h
|
V
+|p q
h
|
Q
Proof: Theorem 36 applied to the big system B((u, p), (v, q)). 2
The stability on the continuous level V Q does not imply the discrete stability !
Usually, one checks the conditions of Brezzi on the discrete level to prove stability of
B(., .) on the discrete levels. The continuity of a(., .) and b(., .) are inherited from the
continuous levels. The stability conditions have to be checked separately. The discretet
kernel ellipticity
a(v
h
, v
h
) _ |v
h
|
2
V
v
h
V
0h
= v
h
V
h
: b(v
h
, q
h
) = 0 q
h
Q
h
, (7.14)
and the discrete LBB condition
sup
u
h
V
h
b(u
h
, q
h
)
|u
h
|
V
_ |q
h
|
Q
q
h
Q
h
. (7.15)
The discrete LBB condition is posed for less dual variables q
h
in Q
h
Q, but the space in
the supremum is also smaller. It does not follow from the LBB condition on the continuous
levels.
There is a canonical technique to derive the discrete LBB condition from the continuous
one:
Lemma 99. Assume there exists a quasi-interpolation operator
h
: V V
h
which is continuous
|
h
v|
V
_ |v|
V
v V,
and which satises
b(
h
v, q
h
) = b(v, q
h
) q
h
Q
h
.
Then, the continuous LBB condition implies the discrete one.
Proof: For all p
h
Q
h
there holds
sup
v
h
V
h
b(v
h
, p
h
)
|v
h
|
V
sup
vV
b(
h
v, p
h
)
|
h
v|
V
_ sup
vV
b(v, p
h
)
|v|
V
_ |p
h
|
Q
2
Approximation of the mixed method for the ux
Choose the pair of nite element spaces, the Raviart Thomas spaces
V
h
= v H(div) : v[
T
V
RT
T
V = H(div)
and the space of piece-wise constants
Q
h
= q L
2
: q[
T
P
0
Q = L
2
.
Pose the discrete mixed problem: Find (
h
, u
h
) V
h
Q
h
such that
(a
1
h
)
h
dx +
div
h
u
h
dx =
D
u
D
n
ds
h
V
h
div
h
v
h
dx =
fv
h
dx v
h
Q
h
.
(7.16)
Lemma 100 (Discrete Stability). The discrete mixed variational problem (7.16) is well
posed.
Proof: By Brezzis theorem. Continuity of the bilinear-form and the linear-form follow
from the continuous level. We prove the kernel ellipticity: Since
div V
h
Q
h
,
7.5. APPROXIMATION OF MIXED SYSTEMS 109
there holds
div
h
q
h
dx = 0 q
h
Q
h
div
h
= 0,
and thus V
0h
V
0
. In this special case, the discrete kernel ellipticity is simple the restriction
of the continuous one to V
0h
. We are left with the discrete LBB condition. We would like
to apply Lemma 99. The quasi-interpolation operator is the Raviar-Thomas interpolation
operator I
RT
h
. The abstract condition
b(I
RT
h
, v
h
) = b(, v
h
) v
h
Q
h
reads as
T
div I
RT
h
dx =
T
div dx,
which was proven in Lemma 93. But, the interpolation operator is not continuous on
H(div). The edge-integrals are not well dened on H(div). We have to include the sub-
space [H
1
]
d
H(div). There holds
|I
RT
h
|
H(div)
_ ||
H
1 [H
1
]
d
,
and the stronger LBB condition (see Section on Stokes below)
sup
[H
1
]
d
(div , v)
L
2
||
H
1
|v|
L
2
v L
2
.
We follow the proof of Lemma 99: For all v
h
Q
h
there holds
sup
h
V
h
b(
h
, v
h
)
|
h
|
V
sup
[H
1
]
d
(div I
RT
h
, v
h
)
|I
RT
h
|
V
_ sup
[H
1
]
d
(div , v
h
)
||
H
1
_ |v
h
|
L
2
.
Brezzis theorem now proves that the discrete problem is well posed, i.e., it fullls the
discrete inf-sup condition.
2
Theorem 101 (A priori estimate). The mixed nite element methods for the uxes satises
the error estimates
|
h
|
L
2
+| div(
h
)|
L
2
+|u u
h
|
L
2
_ h (||
H
1 +|u|
H
1 +|f|
H
1) (7.17)
Proof: By discrete stability, one can bound the discretization error by the best approx-
imation error
|
h
|
H(div)
+|u u
h
|
L
2
_ inf
h
V
h
v
h
Q
h
|
h
|
H(div)
+|u v
h
|
L
2
.
The best approximation error is bounded by the interpolation error. The rst term is
(using the commuting diagram property and div = f)
inf
h
V
h
|
h
|
L
2
+| div(
h
)|
L
2
|I
RT
h
|
L
2
+|(IP
0
) div |
L
2
_ h (||
H
1 +|f|
H
1) .
The second term is
inf
v
h
Q
h
|u v
h
|
L
2
|u P
0
u|
L
2
_ h|u|
H
1.
2
The smoothness requirements onto the solution of (7.17) are fullled for problems on
convex domains, and smooth (constant) coecients a. There holds |u|
H
2 _ |f|
L
2
. Since
= au, there follows ||
H
1 _ |f|
L
2
. The mixed method requires more smoothness
onto the right hand side data, f H
1
. It can be reduced to H
1
on sub-domains, what is
a realistic assumption. On non-convex domains, u is in general not in H
2
(and not in
H
1
). Again, weighted Sobolev spaces can be used to prove similar estimates on properly
rened meshes.
Approximation of the mixed method for Dirichlet boundary conditions
A possibility is to choose continuous and piece-wise linear nite element spaces on the
domain and on the boundary
V
h
= v C() : v[
T
P
1
T,
Q
h
= C() : [
E
P
1
E .
Theorem 102. The discrete mixed method is well posed.
Proof: Exercises.
Chapter 8
Applications
We investigate numerical methods for equations describing real life problems.
8.1 The Navier Stokes equation
The Navier Stokes equation describe the ow of a uid (such as water or air). The in-
compressible Navier Stokes equation models incompressible uids (such as water). The
stationary N.-St. equation models a solution in steady state (no change in time).
The eld variables are the uid velocity u = (u
x
, u
y
, u
z
), and the pressure p. Conserva-
tion of momentum is
u + (u )u p = f
The rst term describes friction of the uid ( is called viscosity). The second one arises
from conservation of momentum of moving particles. It is called the convective term ( is
the density). The source term f models forces, mainly gravity. The incompressibility of
the uid is described by
div u = 0.
Dierent types of boundary conditions onto u and p are possible.
The Navier Stokes equation is nonlinear. In general, no unique solution is guaranteed.
The common approach to nd a solution is the so called Oseen iteration: Given u
k
, nd
the next iterate (u
k+1
, p
k+1
) by solving
u
k+1
+ (u
k
)u
k+1
p
k+1
= f
div u
k+1
= 0.
Under reasonable conditions, this Oseen equation is uniquely solvable. Since u
k
is the
solution of the old step, it satises div u
k
= 0. Furthermore, we assume that the velocity
u
k
is bounded in L
-norm.
111
112 CHAPTER 8. APPLICATIONS
From now on, we continue to investigate the Oseen equation. Given a vector-eld
w = (w
x
, w
y
, w
z
) [L
]
3
such that div w = 0. Find u and p such that
u + (w )u p = f
div u = 0.
We have removed the viscosity by rescaling the equation. The factor / is incorporated
into the vector-eld w.
As usual, we go over to the weak formulation: Find u V = [H
1
]
3
and p Q = L
2
such that
uv + (w )u v dx +
div v p dx =
fv dx v V
div u q dx = 0 q Q.
(8.1)
This variational problem is a mixed formulation. It satises the conditions of Brezzi:
The bilinear forms are
a(u, v) =
uv + (w )uv dx,
b(u, q) =
div u q dx.
Both forms are continuous. The form a(., .) is non-symmetric. In a(., .), the x, y, and
z components of u and v are independent. To investigate a(., .), it is enough to consider
scalar bilinear-forms. We dene the inow and outow boundaries
i
= x : w n < 0,
o
= x : w n 0.
If we pose Dirichlet boundary conditions on
i
, then a(., .) is coercive (see example 27, and
exercises). The ratio of the continuity bound and the coercivity bound depends on the
norm of the convection w. With increasing w, the problem is getting worse.
The form b(., .) satises the LBB condition:
sup
u[H
1
0,D
]
3
div u q dx
|u|
H
1
_ |q|
L
2
q L
2
.
In the case of (partial) Dirichlet boundary conditions (H
1
0,D
= u : u = 0 on
D
),
this condition is very nontrivial to prove. If there are only Dirichlet b.c., one has to use
Q = L
0
2
= q :
q dx = 0.
Under these conditions, Brezzis theorem proves a unique solution of the Oseen equa-
tion.
8.2. ELASTICITY 113
Finite elements for Navier-Stokes equation
We want to approximate the Oseen equation by a Galerkin method: Find u
h
V
h
and
p
h
Q
h
such that
u
h
v
h
+ (w )u
h
v
h
dx +
div v
h
p
h
dx =
fv
h
dx v
h
V
h
div u
h
q
h
dx = 0 q
h
Q.
(8.2)
To obtain convergence u
h
u and p
h
p, it is important to choose proper approxi-
mation spaces V
h
and Q
h
. Using the simplest elements, namely continuous and piece-wise
linear elements for V
h
[H
1
]
3
, and piece-wise constants for Q
h
L
2
does not work. The
discrete LBB condition is not fullled: In 2D, there are asymptotically twice as many
triangles than vertices, i.e., dim V
h
dim Q
h
, and
div u
h
q
h
dx = 0 q
h
Q
h
implies
u
h
0.
The simplest spaces which lead to convergence are the non-conforming P
1
element for
the velocities, and piece-wise constant elements for the pressure. The arguments are
There are unknowns on the edges to construct a Fortin operator satisfying
e
u nds =
e
(I
h
u) nds,
and thus proving the discrete LBB condition.
The error due to the non-conforming space V
h
V is of the same order as the
approximation error (see Section 5.4).
8.2 Elasticity
We start with a one-dimensional model. Take a beam which is loaded by a force density f
in longitudinal (x) direction. We are interested in the displacement u(x) in x direction.
The variables are
The strain : It describes the elongation. Take two points x and y on the beam.
After deformation, their distance is y +u(y) (x +u(x)). The relative elongation of
the beam is
y + u(y) (x + u(x)) (y x)
y x
=
u(y) u(x)
y x
.
In the limit y x, this is u
t
. We dene the strain as
= u
t
.
The stress : It describes internal forces. If we cut the piece (x, y) out of the beam,
we have to apply forces at x and y to keep that piece in equilibrium. This force is
called stress . Equilibrium is
(y) (x) +
y
x
f(s) ds = 0,
or
t
= f
Hooks law postulates a linear relation between the strain and the stress:
= E.
Combining the three equations
= u
t
= E
t
= f
leads to the second order equation for the displacement u:
(Eu
t
)
t
= f.
Boundary conditions are
Dirichlet b.c.: Prescribe the displacement at the boundary
Neumann b.c: Prescibe the stress at the boundary
Elasticity in more dimensions
We want to compute the deformation of the body R
d
.
The body is loaded with a volume force density f : R
d
.
The displacement is described by a the vector-valued function
u : R
d
.
The strain becomes a symmetric tensor in R
dd
. The elongation in the direction of
the unit-vector n is
n
T
n.
The (linearized!) relation between the displacement u and the strain is now
ij
=
1
2
u
i
x
j
+
u
j
x
i
,
or, in compact form
= (u) =
1
2
u + (u)
T
.
8.2. ELASTICITY 115
If the displacement is a pure translation (u = const), then the strain vanishes. Also,
if the displacement is a linearized (!) rotation, (in two dimensions u = (u
x
, u
y
) =
(y, x), the strain vanishes. We call these deformations the rigid body motions:
R
2D
=
a
1
a
2
+ b
y
x
: a
1
, a
2
, b R
R
3D
=
a + b x : a, b R
3
The stress becomes a tensor R

dd
. Consider the part V . To keep V in
equilibrium, on has to apply the surface force density n at V :
V
nds +
V
f dx = 0.
Apply Gauss theorem to obtain the dierential form
div = f.
The div-operator is applied for each row of . A further hypothesis, equilibrium of
angular momentum, implies that is symmetric.
Hooks law is now a relation between two second order tensors:
ij
=
kl
D
ijkl
kl
,
in short
= D,
where D is a fourth order tensor. For an isotropic material (same properties in all
directions), the matrial law has the special structure
= 2 + tr I.
The two parameters and are called Lames parameters. The trace tr is dened
as tr =
d
i=1
ii
.
Collecting the equations
= (u) = D div = f
leads to
div D(u) = f.
Multiplication with test-functions v : R
d
, and integrating by parts leads to
D(u) : v dx =
f v dx v
The operator : is the inner product for matrices, A : B =
ij
A
ij
B
ij
. Next, we use that
= D(u) is symmetric. Thus, : v = : (v)
T
= :
1
2
v + (v)
T
.
The equations of elasticity in weak form read as: Find u V = [H
1
0,D
()]
d
such that
D(u) : (v) dx =
f v dx v V.
Displacement (Dirichlet) boundary conditions (u = u
D
at
D
) are essential b.c., and
are put into the space V . Neumann boundary conditions (natural b.c.) model surface
forces sigman = g, and lead to the additional term
N
g v ds on the right hand side.
The bilinear-form in the case of an isotropic material reads as
2(u) : (v) + div u div v dx.

We assume a positive denite material law
D : _ : symmetric R
dd
Theorem 103. Assume that the Dirichlet boundary
D
has positive measure. Then the
equations of elasticity are well posed in [H
1
]
d
.
Proof: Continuity of the bilinear-form and the linear-form are clear. Ellipticity of the
bilinear-form follows from the positive denite matrial law, and the (non-trivial) Korn
inequality
(u) : (v) dx _ |u|

2
H
1
()
u [H
1
0,D
]
d
The Lax-Milgram theorem proves a unique solution u. 2
The discretization of the elasticity problem is straight forward. Take a nite dimensional
sub-space V
h
V , and perform Galerkin projection. One may use the standard nodal
nite elements for each component.
Structural mechanics
Many engineering applications involve thin structures (walls of a building, body of a car,
...). On thin structures, the standard approach has a problem: One observed that the
simulation results get worse as the thickness decreases. The explanation is that the constant
in Korns inequality gets small for thin structures. To understand and overcome this
problem, we go over to beam, plate and shell models.
8.2. ELASTICITY 117
We consider a thin (t <1) two-dimensional body
= I (t/2, t/2) with I = (0, 1)
The goal is to derive a system of one-dimensional equations to describe the two-dimensional
deformation. This we obtain by a semi-discretization. Dene
V
M
=
u
x
(x, y)
u
y
(x, y)
V : u
x
(x, y) =
Mx
i=0
u
i
x
(x)y
i
, u
y
(x, y) =
My
i=0
u
i
y
(x)y
i
.
This function space on R
2
is isomorph to a one-dimensional function space with values
in R
Mx+My+2
. We perform semi-discretization by searching for u

V
M
such that
A( u, v) = f( v) v

V
M
.
As M
x
, M
y
,

V
M
V , and we obtain convergence u u.
The lowest order (qualitative) good approximating semi-discrete space is to set M
x
= 1
and M
y
= 0. This is
V =
U(x) (x)y
w(x)
Evaluating the bilinear-form (of an isotropic material) leads to

A
U y
w
U y
= (2 + )t
1
0
U
t

U
t
dx +
(2 + )
t
3
12
1
0
t
+ 2
t
2
1
0
(w
t
)( w
t

) dx
The meaning of the three functions is as follows. The function U(x) is the average
(over the cross section) longitudinal displacement, w(x) is the vertical displacement. The
function is the linearized rotation of the normal vector.
We assume that the load f(x, y) does not depend on y. Then, the linear form is
f
U y
= t
1
0
f
x

U dx + t
1
0
f
y
wdx
The semi-discretization in this space leads to two decoupled problems. The rst one
describes the longitudinal displacement: Find U H
1
(I) such that
(2 + )t
1
0
U
t

U
t
dx = t
1
0
f
x

U
t
dx U
t
H
1
(I).
The small thickness parameter t cancels out. It is a simple second order problem for the
longitudinal displacement.
The second problems involves the 1D functions w and : Find (w, ) V =? such that
(2 + )
t
3
12
1
0
t
dx + t
1
0
(w
t
)( w
t

) dx = t
1
0
f
y
wdx ( w,

) V
The rst term models bending. The derivative of the rotation is (approximative) the
curvature of the deformed beam. The second one is called the shear term: For thin beams,
the angle tan is approximatively w
t
. This term measures the dierence w
t
. This
second problem is called the Timoshenko beam model.
For simplication, we skip the parameters and , and the constants. We rescale the
equation by dividing by t
3
: Find (w, ) such that
t
dx +
1
t
2
(w
t
)( w
t

) dx =
t
2
f wdx. (8.3)
This scaling in t is natural. With t 0, and a force density f t
2
, the deformation
converges to a limit. We dene the scaled force density
f = t
2
f
In principle, this is a well posed problem in [H
1
]
2
:
Lemma 104. Assume boundary conditions w(0) = (0) = 0. The bilinear-form
A((w, ), ( w,

)) of (8.3) is continuous
A((w, ), ( w,

)) _ t
2
(|w|
H
1 +||
H
1)(| w|
H
1 +|
|
H
1)
and coercive
A((w, ), (w, )) |w|
2
H
1 +||
2
H
1
Proof: ...
As the thickness t becomes small, the ratio of the continuity and coercivity bounds
becomes large ! This ratio occurs in the error estimates, and indicates problems. Really,
numerical computations show bad convergence for small thickness t.
The large coecient in front of the term
(w
t
)( w
t

) forces the dierence w
t
to be small. If we use piece-wise linear nite elements for w and , then w

t
h
is a piece-wise
constant function, and
h
is continuous. If w
t
h

h
0, then
h
must be a constant
function !
The idea is to weaken the term with the large coecient. We plug in the projection P
0
into piece-wise constant functions: Find (w
h
,
h
) such that

t
h
t
h
dx +
1
t
2
P
0
(w
t
h
h
) P
0
( w
t
h
h
) dx =

f w
h
dx. (8.4)
Now, there are nite element functions w
h
and
h
fullling P
0
(w
t
h
h
) 0.
8.2. ELASTICITY 119
In the engineering community there are many such tricks to modify the bilinear-form.
Our goal is to understand and analyze the obtained method.
Again, the key is a mixed method. Start from equation (8.3) and introduce a new
variable
p = t
2
(w
t
). (8.5)
Using the new variable in (8.3), and formulating the denition (8.5) of p in weak form leads
to the bigger system: Find (w, ) V and p Q such that
t
dx +
( w
t
)p dx =

f wdx (w, ) V
(w
t
) p dx t
2
p p dx = 0 p Q.
(8.6)
This is a mixed formulation of the abstract structure: Find u V and p Q such that
a(u, v) + b(v, p) = f(v) v V,
b(u, q) t
2
c(p, q) = 0 q Q.
(8.7)
The big advantage now is that the parameter t does not occur in the denominator, and
the limit t 0 can be performed.
This is a family of well posed problems.
Theorem 105 (extended Brezzi). Assume that the assumptions of Theorem 86 are true.
Furthermore, assume that
a(u, u) 0,
and c(p, q) is a symmetric, continuous and non-negative bilinear-form. Then, the big form
B((u, p), (v, q)) = a(u, v) + b(u, q) + b(v, p) t
2
c(p, q)
is continuous and stable uniformly in t [0, 1].
We check Brezzis condition for the beam model. The spaces are V = [H
1
]
2
and Q = L
2
.
Continuity of the bilinear-forms a(., .), b(., .), and c(., .) is clear. The LBB condition is
sup
w,
(w
t
)q dx
|w|
H
1 +||
H
1
_ |q|
L
2
We construct a candidate for the supremum:
w(x) =
x
0
q(s) ds and = 0
Then

(w
t
)q dx
|w|
H
1 +||
H
1
_
q
2
dx
|w
t
|
= |q|
L
2
Finally, we have to check kernel ellipticity. The kernel is
V
0
= (w, ) : = w
t
.
On V
0
there holds
|w|
1
H
1 +||
2
H
1 _ |w
t
|
2
+||
2
H
1 = ||
2
L
2
+||
2
H
1
_ |
t
|
L
2
= a((w, ), (w, ))
The lowest order nite element discretization of the mixed system is to choose contin-
uous and piece-wise linear elements for w
h
and
h
, and piecewise constants for p
h
. The
discrete problem reads as: Find (w
h
,
h
) V
h
and p
h
Q
h
such that

t
h
t
h
dx +
( w
t
h
h
)p
h
dx =

f w
h
dx (w
h
,
h
) V
h
(w
t
h
h
) p
h
dx t
2
p
h
p
h
dx = 0 p
h
Q
h
.
(8.8)
This is a inf-sup stable system on the discrete spaces V
h
and Q
h
. This means, we obtain
the uniform a priori error estimate
|(w w
h
,
h
)|
H
1
+|p p
h
|
L
2
_ inf
w
h
,
h
, p
h
|(w w
h
,

h
)|
H
1
+|p p
h
|
L
2
_ h|w|
H
2 +||
H
2 +|p|
H
1
The required regularity is realistic.
The second equation of the discrete mixed system (8.8) states that
p
h
= t
2
P
0
(w
t
h
h
)
If we insert this observation into the rst row, we obtain exactly the discretization method
(8.4) ! Here, the mixed formulation is a tool for analyzing a non-standard (primal) dis-
cretization method. Both formulations are equivalent. They produce exactly the same
nite element functions. The mixed formulation is the key for the error estimates.
The two pictures below show simulations of a Timoshenko beam. It is xed at the left
end, the load density is constant one. We compute the vertical deformation w(1) at the
right boundary. We vary the thickness t between 10
1
and 10
3
. The left pictures shows
the result of a standard conforming method, the right picture shows the results of the
method using the projection. As the thickness decreases, the standard method becomes
worse. Unless h is less than t, the results are completely wrong ! The improved method
converges uniformly well with respect to t:
0
0.02
0.04
0.06
0.08
0.1
0.12
1 10 100 1000 10000
w
(
1
)
Elements
t=1e-1
t=1e-2
t=1e-3
0
0.02
0.04
0.06
0.08
0.1
0.12
1 10 100 1000
w
(
1
)
Elements
t=1e-1
t=1e-2
t=1e-3
8.3. MAXWELL EQUATIONS 121
8.3 Maxwell equations
Maxwell equations describe electro-magnetic elds. We consider the special case of sta-
tionary magnetic elds. Maxwell equations are three-dimensional.
A magnetic eld is caused by an electric current. We suppose that a current density
j [L
2
()]
3
is given. (Stationary) currents do not have sources, i.e., div j = 0.
The involved (unknown) elds are
The magnetic ux B (in German: Induktion). The ux is free of sources, i.e.,
div B = 0.
The magnetic eld intensity H (in German: magnetische Feldst arke). The eld is
related to the current density by Henrys law:
S
j nds =
S
H ds Surfaces S
By Stokes Theorem, one can derive Henrys law in dierential form:
curl H = j
The dierential operator is curl = rot = . Both elds are related by a material law.
The coecient is called permeability:
B = H
The coecient is 10
3
to 10
4
times larger in iron (and other ferro-magnetic metals) as
in most other media (air). In a larger range, the function B(H) is also highly non-linear.
Collecting the equations we have
div B = 0 B = H curl H = j (8.9)
In principle, Maxwell equations are valid in the whole R
3
. For simulation, we have to
truncate the domain and have to introduce articial boundary conditions.
The picture below shows the magnetic eld caused by a tangential current density in a
coil:
Compare these equations to the diusion equation div au = f. Here, we could
introduce new unknowns g = u and = ag. On simply connected domains, g is a
gradient eld if and only if curl g = 0. We could reformulate the equations as: Find vector
elds g and such that
curl g = 0 = ag div = f.
The system of magnetostatic equations looks similar. Only, the right hand side data is
applied to the curl-equation, instead of the div-equation. In a similar way as curl g = 0
allows to introduce a scalar eld u such that g = u, div B = 0 allows to introduce a
vector potential A such that
B = curl A.
Inserting the vector-potential into the equations (8.9), one obtains the second order equa-
tion
curl
1
curl A = j. (8.10)
The two original elds B and H can be obtained from the vector potential A.
The vector-potential A is not uniquely dened by (8.10). One may add a gradient eld
to A, and the equation is still true. To obtain a unique solution, the so called Coloumb-
Gauging can be applied:
div A = 0. (8.11)
As usual, we go over to the weak form. Equations (8.10) and (8.11) together become:
Find A such that
1
curl A curl v dx =
j v dx v ?
and
A dx = 0.
We want to choose the same space for A and the according test functions v. But, then
we have more equations than unknowns. The system is still solvable, since we have made
the assumption div j = 0, and thus j is in the range of the curl- operator. To obtain a
symmetric system, we add a new scalar variable . The problem is now: Find A V =?
and Q = H
1
/R such that

1
curl A curl v dx +
v dx =
j v dx v V
A dx = 0 Q
(8.12)
The proper space V is the H(curl):
H(curl) = v [L
2
()]
3
: curl v [L
2
()]
3
Again, the dierential operator curl is understood in the weak sense. The canonical norm
is
|v|
H(curl)
=
|v|
2
L
2
+| curl v|
2
L
2
1/2
.
Similar to H
1
and H(div), there exists a trace operator for H(curl). Now, only the
tangential components of the boundary values are well dened:
Theorem 106 (Trace theorem). There exists a tangential trace operator tr
v : H(curl)
W() such that
tr
v = (v[
for smooth functions v [C()]

3
.
Theorem 107. Let =
i
. Assume that u[
i
H(curl,
i
), and the tangential traces
are continuous across the interfaces
ij
. Then u H(curl, ).
The theorems are according to the ones we have proven for H(div). But, the proofs (in
R
3
) are more involved.
The gradient operator relates the space H
1
and H(curl):
: H
1
H(curl)
Furthermore, the kernel space
H
0
(curl) = v H(curl) : curl v = 0
is exactly the range of the gradient:
H
0
(curl) = H
1
Theorem 108. The mixed system (8.12) is a well posed problem on H(curl) H
1
/R.
Proof: The bilinear-forms
a(A, v) =

1
curl A curl v dx
and
b(v, ) =
v dx
are continuous w.r.t. the norms of V = H(curl) and Q = H
1
/R.
The LBB-condition in this case is trivial. Choose v = :
sup
vH(curl)
vdx
|v|
H(curl)
dx
||
H(curl)
=
||
2
L
2
||
L
2
= ||
L
2
||
Q
The dicult part is the kernel coercivity of a(., .). The norm involves also the L
2
-norm,
while the bilinear-form only involves the semi-norm | curl v|
L
2
. Coercivity cannot hold on
the whole V : Take a gradient function . On the kernel, the L
2
-norm is bounded by the
semi-norm:
|v|
L
2
_ | curl v| v V
0
,
where
V
0
= v H(curl) :
vdx = 0 H
1
This is a Friedrichs-like inequality.

Finite elements in H(curl)
We construct nite elements in three dimensions. The trace theorem implies that functions
in H(curl) have continuous tangential components across element boundaries (=faces).
We design tetrahedral nite elements. The pragmatic approach is to choose the element
space as V
T
= P
1
, and choose the degrees of freedom as the tangential component along
the edges in the end-points of the edges. The dimension of the space is 3 dimP
1
=
3 4 = 12, the degrees of freedom are 2 per edge, i.e., 2 6 = 12. They are also linearly
independent. In each face, the tangential component has 2 components, and is linear.
Thus, the tangential component has dimension 6. These 6 values are dened by the 6
degrees of freedom of the 3 edges in the face. Neighboring elements share this 6 degrees of
freedom in the face, and thus have the same tangential component.
There is a cheaper element, called Nedelec, or edge-element. It has the same accuracy
for the curl-part (the B-eld) as the P
1
-element. It is similar to the Raviart-Thomas
element. It contains all constants, and some linear polynomials. All 3 components are
dened in common. The element space is
V
T
= a + b x : a, b R
3
.
These are 6 coecients. For each of the 6 edges of a tetrahedron, one chooses the integral
of the tangential component along the edge
E
i
(u) =
E
i
u
E
i
ds.
Lemma 109. The basis function
E
i
associated with the edge E
i
is
E
i
=
E
1
i
E
2
i
E
2
i
E
1
i
,
where E
1
i
and E
2
i
are the two vertex numbers of the edge, and
1
, . . .
4
are the vertex shape
functions.
Proof:
These functions are in V
T
If i = j, then
E
j
(
E
i
) = 0.

E
i
(
E
i
) = 1
Thus, edge elements belong to H(curl). Next, we will see that they have also very
interesting properties.
The deRham complex
The spaces H
1
, H(curl), H(div), and L
2
form a sequence:
H
1

H(curl)
curl
H(div)
div
L
2
Since H
1
[L
2
]
3
, and curl = 0, the gradients of H
1
functions belong to H(curl).
Similar, since curl H(curl) [L
2
]
3
, and div curl = 0, the curls of H(curl) functions belong
to H(div).
The sequence is a complete sequence. This means that the kernel of the right dierential
operator is exactly the range of the left one (on simply connected domains). We have used
this property already in the analysis of the mixed system.
The same property holds on the discrete level: Let
W
h
be the nodal nite element sub-space of H
1
V
h
be the Nedelec (edge) nite element sub-space of H(curl)
Q
h
be the Raviart-Thomas (face) nite element sub-space of H(div)
S
h
be the piece-wise constant nite element sub-space of L
2
Theorem 110. The nite element spaces form a complete sequence
W
h
V
h
curl
Q
h
div
S
h
Now, we discretize the mixed formulation (8.12) by choosing edge-nite elements for
H(curl), and nodal nite elements for H
1
: Find A
h
V
h
and
h
W
h
such that

1
curl A
h
curl v
h
dx +

h
v
h
dx =
j v
h
dx v
h
V
h
A
h

h
dx = 0
h
W
h
(8.13)
The stability follows (roughly) from the discrete sequence property. The verication of the
LBB condition is the same as on the continuous level. The kernel of the a(., .)- form are
the discrete gradients, the kernel of the b(., .)-form is orthogonal to the gradients. This
implies solvability. The discrete kernel-coercivity (with h-independent constants) is true
(nontrivial).
The complete sequences on the continuous level and on the discrete level are con-
nected in the deRham complex: Choose the canonical interpolation operators (vertex-
interpolation I
W
, edge-interpolation I
V
, face-interpolation I
Q
, L
2
-projection I
S
). This
relates the continuous level to the discrete level:
H
1

H(curl)
curl
H(div)
div
L
2
I
W
I
V
I
Q
I
S
W
h
V
h
curl
Q
h
div
S
h
.
(8.14)
Theorem 111. The diagram (8.14) commutes:
I
V
= I
W
I
Q
curl = curl I
V
I
S
div = div I
Q
Proof: We prove the rst part. Note that the ranges of both, I
W
and I
V
, are in
V
h
. Two functions in V
h
coincide if and only if all functionals coincide. It remains to prove
that
E
(I
W
w) ds =
E
(I
V
w) ds
Per denition of the interpolation operator I
V
there holds
E
(I
V
w) ds =
E
w ds
Integrating the tangential derivative gives the dierence
E
w ds =
E
w
ds = w(E
2
) w(E
1
)
Starting with the left term, and using the property of the nodal interpolation operator, we
obtain
E
(I
W
w) ds = (I
W
w)(E
2
) (I
W
w)(E
1
) = w(E
2
) w(E
1
).
We have already proven the commutativity of the H(div) L
2
part of the diagram. The
middle one involves Stokes theorem. 2
This is the key for interpolation error estimates. E.g., in H(curl) there holds
|u I
V
u|
2
H(curl)
= |u I
V
u|
2
L
2
+| curl(I I
V
)u|
2
L
2
= |u I
V
u|
2
L
2
+|(I I
Q
) curl u|
2
L
2
_ h
2
|u|
2
H
1 + h
2
| curl u|
2
H
1
Since the estimates for the L
2
-term and the curl-term are separate, one can also scale each
of them by an arbitrary coecient.
The sequence is also compatible with transformations. Let F :

T T be an (element)
transformation. Choose
w(F(x)) = w(x)
v(F(x)) = (F
t
)
T
v(x) (covariant transformation)
q(F(x)) = (det F
t
)
1
(F
t
)q(x) (Piola-transformation)
s(F(x)) = (det F
t
)
1
s(x)
Then
v = w v = w
q = curl v q = curl v
s = div q s = div q
Using these transformation rules, the implementation of matrix assembling for H(curl)-
equations is very similar to the assembling for H
1
problems (mapping to reference element).
Chapter 9
Parabolic partial dierential
equations
PDEs involving rst order derivatives in time, and an ellitpic dierential operator in space,
are called parabolic PDEs. For example, time dependent heat ow is described by a
parabolic PDE.
Let R
d
, and Q = (0, T). Consider the initial-boundary value problem
u(x, t)
t
div(a(x)
x
u(x, t)) = f(x, t) (x, t) Q,
with boundary conditions
u(x, t) = u
D
(x, t) (x, t)
D
(0, T),
a(x)
u
n
= g(x, t) (x, t)
N
(0, T),
and initial conditions
u(x, 0) = u
0
(x) x .
Weak formulation in space: Find u : [0, T] H
1
0,D
() such that
t
u(x, t)v(x) dx +
au(x, t) v(x, t) dx =
f(x, t)v(x, t) dx +
N
g(x, t)v(x, t) dx
v H
1
0,D
, t (0, T]
In abstact form: Find u : [0, T] V s.t.
(u
t
(t), v)
L
2
+ a(u(t), v) = 'f(t), v` v V, t (0, T]
In operator form (with 'Au, v` = a(u, v)):
u
t
(t) + Au(t) = f(t) V
129
130 CHAPTER 9. PARABOLIC PARTIAL DIFFERENTIAL EQUATIONS
Function spaces:
X = L
2
((0, T), V ) X
= L
2
((0, T), V
)
with norms
|v|
X
=
T
0
|v(t)|
2
V
dt
1/2
|v|
X
=
T
0
|v(t)|
2
V
dt
1/2
Denition 112. Let u L
2
((0, T), V ). It has a weak derivative w L
2
((0, T), V
) if
T
0
(t) 'w, v`
V
V
dt =
T
0
t
(t)(u, v)
L
2
dt v V, C
0
(0, T)
Denition 113.
H
1
((0, T), V ; L
2
) = v L
2
((0, T), V ) : v
t
L
2
((0, T), V
)
with norm
|v|
2
H
1 = |v|
2
X
+|v
t
|
2
X
.
This space is a one-dimensional Sobolev space with range in a Hilbert space.
Theorem 114 (Trace theorem). Point evaluation is continuous:
max
t[0,T]
|v(t)|
L
2
_ |v|
H
1
This allows the formulation of the initial value u(0) = u
0
.
Theorem 115. Assume that a(., .) is coercive
a(u, u)
1
|u|
2
V
u V
and continuous
a(u, v)
2
|u|
V
|v|
V
u, v V.
Then, the parabolic problem has a unique solution depending continuously on the right hand
side and the initial conditions:
|u|
H
1
((0,T),V ;L
2
)
_ |u
0
|
L
2
+|f|
L
2
((0,T),V
.
We only prove stability: Choose test functions v = u(t):
(u
t
(t), u(t))
L
2
+ a(u(t), u(t)) = 'f(t), u(t)`
Use that
d
dt
|u(t)|
2
L
2
= 2(u
t
(t), u(t))
L
2
,
9.1. SEMI-DISCRETIZATION 131
and integrate the equation over (0, T):
1
2
|u(T)|
2
L
2
|u
0
|
2
L
2
T
0
'f(s), u(s)` a(u(s), u(s)) ds
T
0
|f(s)|
V
|u(s)|
V

1
|u(s)|
2
V
ds
|f|
X
|u|
X

1
|u|
2
X
Since |u(T)| 0, one has
1
|u|
2
X
|f|
X
|u|
X

1
2
|u
0
|
L
2
Solving the quadatic inequality, one obtains the bound
|u|
X

1
2
1
|f|
X
+
|f|
2
X
+ 2
1
|u
0
|
2
L
2
The bound |u
t
|
L
2
((0,T),V
)
follows from u
t
(t) = f(t) Au(t).
9.1 Semi-discretization
We start with a discretization in space. Choose a (nite element) sub-space V
h
V . The
Galerkin discretiztaion is: Find u : [0, T] V
h
such that
(u
t
h
(t), v
h
)
L
2
+ a(u
h
(t), v
h
) = 'f(t), v
h
` v
h
V
h
, t (0, T],
(u
h
(0), v
h
)
L
2
= (u
0
, v
h
)
L
2
v
h
V
h
.
Choose a basis
1
, . . .
N
of V
h
. Expand the solution w.r.t. this basis:
u
h
(x, t) =
N
i=1
u
i
(t)
i
(x),
and choose test functions v =
j
. With the matrices
M = ((
j
,
i
)
L
2
)
i,j=1,...,N
A = (a(
j
,
i
))
i,j=1,...,N
,
and the t-dependent vector
f(t) = ('f(t),
j
`)
i=1,...,N
,
one obtains the system of ordinary dierential equations (ODEs)
Mu
t
(t) + Au(t) = f(t), u(0) = u
0
In general, the (mass) matrix M is non-diagonal. In the case of the (inexact) vertex
integration rules, or non-conforming P
1
-elements, M is a diagonal matrix. Then, this ODE
can be eciently reduced to explicit form
u
t
(t) + M
1
Au(t) = f(t)
Theorem 116. There holds the error estimate
|u u
h
|
H
1
((0,T),V ;L
2
)
_ |(I R
h
)u|
H
1
((0,T),V ;L
2
)
,
where R
h
is the Ritz projector
R
h
: V V
h
: a(R
h
u, v
h
) = a(u, v
h
) u V, v
h
V
h
.
Proof: The error is split into two parts:
u(t) u
h
(t) = u(t) R
h
u(t)
. .. .
(t)
+R
h
u(t) u
h
(t)
. .. .
h
The rst part, u(t) R
h
u(t) is the elliptic discretization error, which can be bounded by
Ceas lemma. To bound the second term, we use the properties for the continuous and the
discrete formulation:
'f, v
h
` = (u
t
, v
h
) + a(u, v
h
) = (u
t
, v
h
) + a(R
h
u, v
h
)
= (u
t
h
, v
h
) + a(u
h
, v
h
),
i.e.,
(u
t
u
t
h
, v
h
) + a(R
h
u u
h
, v
h
) = 0,
or
(R
h
u
t
u
t
h
, v
h
) + a(R
h
u u
h
, v
h
) = (R
h
u
t
u
t
, v
h
).
With the abbreviations from above we obtain the discrete parabolic equation for
h
:
(
t
h
, v
h
) + a(
h
, v
h
) = (
t
, v
h
)
h
(0) = (I R
h
)u(0).
The stability estimate, and the trace theorem bounds
|
h
|
H
1
((0,T),V ;L
2
)
_ |(I R
h
)u(0)|
L
2
()
+|
t
|
L
2
((0,T),V
)
_ |(I R
h
)u|
H
1
((0,T),V ;L
2
)
2
9.2. TIME INTEGRATION METHODS 133
9.2 Time integration methods
Next, we discuss methods for solving the system of ODEs:
Mu
t
(t) + Au(t) = f(t) (9.1)
u(0) = u
0
We focus on simple time integration rules and the specic properties arising from the
space-discretization of parabolic PDEs. Let
0 = t
0
< t
1
< t
m
= T,
a partitioning of the interval [0, T]. Dene
j
= t
j+1
t
j
. Integrating (9.1) over the
intervalls leads to
Mu(t
j+1
) u(t
j
) +
t
j+1
t
j
Au(s) ds =
t
j+1
t
j
f(s) ds.
Next, we replace the integrals by numerical integration rules. The left-sided rectangle
rule leads to
Mu(t
j+1
) u(t
j
) +
j
Au(t
j
) =
j
f(t
j
)
With the notation u
j
= u(t
j
), this leads to the sequence of linear equations
Mu
j+1
= Mu
j
+
j
(f
j
Au
j
)
In the case of a diagoal M-matrix, this is an explicit formulae for the new time step !
Using the right-sided rectangle rule leads to
Mu
j+1
u
j
+
j
Au
j+1
=
j
f
j+1
,
or
(M +
j
A)u
j+1
= Mu
j
+
j
f
j+1
.
In case of the right-side rule, a linear system must be solve in any case. Thus, this method
is called an implicit time integration method. These two special cases are called the explicit
Euler method, and the implicit Euler method. A third simple choice is the trapezoidal rule
leading to
(M +

j
2
A)u
j+1
= Mu
j
+

j
2
(f
j
+ f
j+1
Au
j
)
It is also an implcit method. Since the trapezoidal integration rule is more accurate, we
expect a more accurate method for approximating the ODE.
All single-step time integration methods can be written in the form
u
j+1
= G
j
(u
j
, f
j
),
where G
j
is linear in both arguments and shall be continuous with bounds
|G
j
(u
j
, f
j
)|
M
L|u
j
|
M
+
j
l |f
j
|
M
1,
with L 1.
Lemma 117. The time integration method fullls the stability estimate
|u
j
|
M
L
j
|u
0
|
M
+ lL
j
j1
i=0
i
|f
i
|
M
1 (9.2)
The explicit Euler method is written as
u
j+1
= (I M
1
A)u + M
1
f
j
,
and has bounds
L = max1,
max
(M
1
A) 1 max1,

h
2
l = 1
If > h
2
, the powers L
j
become very large. This means that the explicit Euler method
becomes instable. Thus, for the explicit Euler method, the time-step must not be greater
than ch
2
.
The implicit Euler method is written as
u
j+1
= (M + A)
1
Mu
j
+ (M + A)
1
f
j
,
and has the bounds
L = 1
l = 1
The method is stable for any time-step . Such a method is called A-stable.
Lemma 118. The time discretization error e
j
:= u(t
j
) u
j
of the implicit Euler method
satises the dierence equation
Me
j+1
e
j
+ Ae
j+1
= d
j
,
where the d
j
satisfy
d
j
=
t
j+1
t
j
f(s) Au(s) ds
j
f(t
j+1
) Au(t
j+1
).
Lemma 119. The error of the integration rule can be estimated by
|d
j
| _ |(f Au)
t
|
L
= |u
tt
|
L
Convergence of the time-discretization method follows from stability plus approxima-
tion:
Theorem 120. The error of the implicit Euler method satises
|u(t
j
) u
j
|
M

j
i=0
|d
j
| _ |u
tt
|
L(0,T)
The trapezoidal rule is A-stable, too. It is based on a more accurate integration rule,
and leads to second order convergence O(
2
). Convergence of higher order can be obtained
by Runge-Kutta methods.
Chapter 10
Hyperbolic partial dierential
equations
Wave phenomena are are described by PDEs involving second order time derivatives:
2
u(x, t)
t
2
div(a(x)u) = f
They require two initial conditions
u(x, 0) = u
0
(x),
u(x, t)
t
= v
0
(x).
Space discretization is accoring to parablic problems, and lead to the second order ODE
Mu
tt
(t) + Au(t) = f,
u(0) = u
0
u
t
(0) = v
0
.
By introducing a new function v = u
t
, the second order ODE can be reduced to the rst
order system
u
t
= v
Mv
t
= f Au,
u(0) = u
0
v(0) = v
0
.
Time integration methods for rst order systems can be applied.
135

Numerical Methods For Partial Differential Equations: Joachim SCH Oberl April 6, 2009

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Numerical Methods For Partial Differential Equations: Joachim SCH Oberl April 6, 2009

Uploaded by

Copyright:

Available Formats

Numerical Methods for

Partial Dierential Equations

p nds to the function p := u v to obtain

This allows to dene the operators P

) + (Pu, Pv) = (Pu, Pv).

= 0. Then, S = V , i.e., l = 0. So take u

. There holds l(z) = 0 (otherwise, z S S

. The evaluation is written

. It is linear, and its operator

. From (2.9) we conclude that also u

. To insure that, we can

Proof: We have to show that the range R(B) = W

. The Hilbert space W

. This means that

1. The weak form for the real

() : u has compact support in . (3.1)

is the rst generalized derivative D

dx for all T. Let K be

is the weak derivative of v:

. The denition of traces is essential to formulate boundary conditions

is a bounded operator w.r.t. the norms |.|

) is well dened and bounded

), and thus converges to u H

34 CHAPTER 3. SOBOLEV SPACES

2 cos(kx), one has

v dx) is a continuous bilinear form on H

v) has the kernel P

4.1. SHIFT THEOREMS 43

5.2. FINITE ELEMENT ERROR ANALYSIS 49

. One may choose a mesh such that

T), and set

T) (trace theorem), and Lw = 0 for w P

Factorizing the block-diagonal interior block A

76 CHAPTER 6. LINEAR EQUATION SOLVERS

cos(m arccos(x)) [x[ 1

The optimum of (6.1) is

The numerator is bounded by 1 for the range

steps. The original method by choosing m dierent relaxation parameters

We want to approximate the solution u in the linear manifold u

To continue, we introduce a Clement operator

Equilibrating both terms gives the optimal choice of number of sub-domains

Q is closed in V . Then there holds the V -orthogonal decom-

Q. Take the p Q such that

(div )dx = 'tr

T). Then, the Piola transformation

T), and (F( x)) = (x). Then there holds

(T), = 0 on T ` e. Then, let 1 on the edge e:

The stress becomes a tensor R

2(u) : (v) + div u div v dx.

(u) : (v) dx _ |u|

Evaluating the bilinear-form (of an isotropic material) leads to

to be small. If we use piece-wise linear nite elements for w and , then w

for smooth functions v [C()]

This is a Friedrichs-like inequality.

You might also like