Cell-centered discretization methods

for 2nd order elliptic problems

Jean E. Roberts, Jerome Jare
INRIA-Rocquencourt
Ali Saada
ENIT
1
Outline
Introduction
A menagerie of discretization methods
The model problem
Some Sobolev spaces
Primal formulation
The case of discontinuous coecients
Mixed nite elements
The space H(div, )
Mixed formulation
The lowest order Raviart-Thomas mixed nite element
Mixed nite element formulation
The resulting linear system
Abstract theory of approximation
Functional framework
Babuskas theory
2
Brezzis theory
Application to the homogeneous Dirichlet problem
Application to the discrete problem
Error estimates
Abstract error estimates
Interpolation errors for the RT nite element
Application to the homogeneous Dirichlet problem
Finite volume methods
The basic idea of the nite volume method
Finite volume formulation
Continuous and discontinuous coecients
Reducing the mixed method to the nite volume method for rectangles
Reducing the mixed method to the nite volume method for triangles
Error analysis of this reduction in the case of triangles
Other equivalent formulations
The mixed-hybrid nite elements formulation
Nonconforming nite elements
3
Rening with nonmatching grid
4
INTRODUCTION
5
Vertex-centered approximation methods
The degrees of freedom are located at the vertices of the mesh
Finite dierences
Finite elements Q
1
P
1
Vertex-centered nite
volumes
control volume
6
Cell-centered approximation methods
Mixed nite elements
on rectangles on triangles
Finite volumes
Unknowns : average value in each cell Control volume = cell
Nodal methods in neutronics
7
The model problem
div(K

grad p) = f in
p = p on if Dirichlet
K
p
n
= g on if Neumann
n

For ow in porous media :

p, pressure K, permeability u = K

grad p, Darcy velocity
K =
_
_
k
1
k
12
k
12
k
2
_
_
, 0 < (Kx, x) , x R
2
.
8
The Sobolev space H
1
().
H
0
() = L
2
() q
2
0,
=
_

q
2
(x)dx
H
1
() = {q L
2
();

grad q (L
2
())
2
}
q
2
1,
= q
2
0,
+|q|
2
1,
|q|
2
1,
=
_

|

grad q|
2
(x)dx
The trace q|

of q H
1
() is in H
1/2
().
The trace
q
n
|

of q H
1
() is in H
1/2
(), the dual space of H
1/2
().
9
Weak primal formulation
Assume K L

(), f L
2
().
Neumann boundary conditions: g H
1/2
().
Find p H
1
() such that
_

K

grad p

grad q =
_

fq < g, q >, q H
1
().
Dirichlet boundary conditions: p H
1/2
().
Find p V
p
= {q H
1
(), q = p on } such that
_

K

grad p

grad q =
_

fq, q V
0
.
10
An example with a discontinuous K
In one dimension, =]0, 2[, f = 0, p(0) = 0, p(2) = 1.
0 1 2
K = 10 K = 100
~ ~ ~
u
x
= 0 = u constant thus very smooth.
p continuous, piecewise linear,
p
x
discontinuous at x = 1 = p is not smooth.
u has a physical meaning and is a good mathematical and numerical unknown.
11
APPROXIMATION WITH MIXED FINITE ELEMENTS
12
Mixed formulation
Write the elliptic problem as a system of rst order equations:
div u = f, u = K

grad p, in
p = p on
D
, u n = g on
N
,
N

D
= = .
Assume that f L
2
() so divu L
2
().
Therefore we take u H(div, ) = {v (L
2
())
2
; div v L
2
()}.
Multiply the second equation by K
1
, then by v, integrate over and by parts.
We obtain
_

(K
1
u) v
_

pdivv = < p, v n >

Recall Greens formula:
_

grad q v +
_

q divv =
_

qv n.
It is sucient to take p L
2
(), u H(div, ).
13
Properties of H(div, )
H(div, ) = {v (L
2
())
2
; div v L
2
()} is an Hilbert space with norm
v
H(div,)
= v
L
2
()
+divv
L
2
()
.
Traces (v n)|

of functions v of H(div, ) are in H

1/2
(),
so boundary data must be such that p H
1/2
(
D
), g H
1/2
(
N
).
14
Notations
The data are f L
2
(), p H
1/2
(
D
), g H
1/2
(
N
).
The spaces are M= L
2
(), W = H(div, ), W
g
= {v W; vn = g on
N
}.
Introduce the forms
a : (L
2
())
2
(L
2
())
2
R, a(u, v) =
_

(K
1
u) v,
b : W M R, b(v, q) =
_

q divv,
l
W
: W R, l
W
(v) =
_

D
p v n,
l
M
: M R, l
M
(v) =
_

fq.
15
Mixed formulation
The problem is
P
m
_

_
Find u W
g
and p M such that
a(u, v) b(v, p) = l
W
(v), v W
0
,
b(u, q) = l
M
(q), q M.
16
Discretization of the domain
1. Let T
h
be a discretization of . Card(T
h
) = ne = number of cells.
A
h
be the set of edges. Card(A
h
) = na = number of edges
h the largest diameter of the cells.
a rectangular mesh a triangular mesh
t
t
t
t
t
t
t
t
t
t
t
tt
t
t
t
t
t
t
t
t
t
t
t
tt
t
t
t
t
t
t
t
t
t
t
t
tt
t
t
t
t
t
tt
2. Dene
an approximation p
h
of p in M
h
, a nite dimensional subset of M
an approximation u
h
of u in W
h
, a nite dimensional subset of W.
17
Approximation space for the scalar unknown
M
h
= the space of functions q
h
in M which are
constant over each triangle

$
$
$
$
$
$
$
$
$
$
$
$$
d
d
d
d
d
d
d

constant over each rectangle

dim M
h
= ne = number of elements
The degrees of freedom are
p
T
an approximation of the average value of p over the cell T, T T
h
.
A basis is {
T
}
TT
h
such that
T
(x) =
_
_
_
1 if x T
0 otherwise
Then p
h
=

TT
h
p
T

T
18
Approximation of the vector unknown
W
h
= {v
h
W; v
h
|T W
T
, T T
h
}.
on each triangle
W
T
= {v
h
; v
h|T
= a
_
_
x
1
x
2
_
_
+
_
_
b
c
_
_
}
on each rectangle
W
T
= {v
h
; v
h|T
=
_
_
ax
1
+b
cx
2
+d
_
_
}
Functions v W
T
are uniquely dened by
_
E
v n
T
, E T.
19
On remarque que divW
h
= M
h
.
dimW
h
= na = number of edges.
The degrees of freedom for W
h
are
u
E
an approximation of the ow rate of u across E,
_
E
u n
E
, E A
h
,
n
E
a chosen unit normal to E.
A basis of W
h
is {v
E
}
EA
h
such that
_
F
v
E
n
F
=
E,F
, F A
h
.
Then, u
h
=

EA
h
u
E
v
E
.
20
Basis functions of W
h
For rectangles v
E
is:
E
For triangles v
E
is:
E
21
The approximation problem
Assume the data p, g are piecewise constant on the edges E .
Introduce W
hg
= {v W
h
; v = g on
N
}
The approximation problem is
P
mh
_

_
Find u
h
W
hg
and p
h
M
h
such that
a(u
h
, v
h
) b(v
h
, p
h
) = l
W
(v
h
), v
h
W
h0
,
b(u
h
, q) = l
M
(q
h
), q
h
M
h
.
22
Discrete equations
The unknowns are the degrees of freedom:
Find {p
T
}
TT
h
, {u
E
}
EA
h
such that
_

K
1

FA
h
u
F
v
F
v
E

_

TT
h
p
T

T
divv
E
=
_

D
pv
E
n, E A
h
, E
N
_

div

EA
h
u
E
v
E

T
=
_

f
T
, T T
h
u
E
= g|E|, E
N
(assuming n
E
= n)
Find {p
T
}
TT
h
, {u
E
}
EA
h
such that

FA
h
u
F
_

K
1
v
F
v
E

TT
h
p
T
_

T
divv
E
=
_

D
pv
E
n, E A
h
, E
N

EA
h
u
E
_

divv
E

T
=
_

f
T
, T T
h
u
E
= g|E|, E
N
.
23
Algebraic system
This leads to the linear system
_
_
A
t
D
D 0
_
_
_
_
U
P
_
_
=
_
_
F
v
F
q
_
_
with P = {p
T
}
TT
h
, U = {u
E
}
EA
h
,E
N
.
This linear system is not positive-denite.
For triangles A has 5 nonzero entries per row
For quadrilaterals A has 7 nonzero entries per row
24
On a rectangular mesh
A rectangular
mesh for
i, j + 1
ij i + 1, j
j + 1/2
j 1/2
j
i + 1/2 i 1/2 i E
n
1
T
n
2
h
1i
h
2j
E '
T
c
We take n
E
= n
1
if E is vertical, n
E
= n
2
if E is horizontal.
Note that

EA
h
u
E
_

divv
E

T
=

ET
u
E
_
T
divv
E
.
Thus the second discrete equation gives
u
i+1/2,j
u
i,j1/2
+u
i,j+1/2
u
i1/2,j
=
_
T
ij
f
25
Consider now the rst discrete equation.
Denote N(E) the set of the 2 cells adjacent to E if E
1 cell adjacent to E if E
If E = E
i+1/2,j
, E
D
:

TT
h
p
T
_

T
divv
E
=

TN(E)
p
T
_
T
divv
E
= p
ij
p
i+1,j
.
If E = E
i+1/2,j
, E
D
:

TT
h
p
T
_

T
divv
E
= p
ij
assuming E lies on the right of the domain,

TT
h
p
T
_

T
divv
E
= p
i+1,j
assuming E lies on the left of the domain.
26
If E = E
i+1/2,j
, E :

FA
h
u
F
_

K
1
v
F
v
E
=

TN(E)

FT
u
F
_

K
1
v
F
v
E
=
u
i+1/2,j
_
T
ij
K
1
v
i+1/2,j
v
i+1/2,j
+u
i1/2,j
_
T
ij
K
1
v
i1/2,j
v
i+1/2,j
+
u
i,j+1/2
_
T
ij
K
1
v
i,j+1/2
v
i+1/2,j
+u
i,j1/2
_
T
ij
K
1
v
i,j1/2
v
i+1/2,j
+
u
i+1/2,j
_
T
i+1,j
K
1
v
i+1/2,j
v
i+1/2,j
+u
i+3/2,j
_
T
i+1,j
K
1
v
i+3/2,j
v
i+1/2,j
+
u
i+1,j+1/2
_
T
i+1,j
K
1
v
i+1,j+1/2
v
i+1/2,j
+u
i+1,j1/2
_
T
i+1,j
K
1
v
i+1,j1/2
v
i+1/2,j
If E , say for instance i = 0 (left boundary):

FA
h
u
F
_

K
1
v
F
v
E
=

TN(E)

FT
u
F
_

K
1
v
F
v
E
=
u
1/2,j
_
T
1,j
K
1
v
1/2,j
v
1/2,j
+u
3/2,j
_
T
1,j
K
1
v
3/2,j
v
1/2,j
+
u
1,j+1/2
_
T
1,j
K
1
v
1,j+1/2
v
1/2,j
+u
1,j1/2
_
T
1,j
K
1
v
1,j1/2
v
1/2,j
27
Denote K
1
=
_
_

1

12

12

2
_
_
,
with
1
= k
2
/,
2
= k
1
/,
12
= k
12
/, = k
1
k
1
(k
12
)
2
.
_
T
ij
K
1
v
i+1/2,j
v
i+1/2,j
=

1
ij
3
h
1i
h
2j
_
T
ij
K
1
v
i1/2,j
v
i+1/2,j
=

1
ij
6
h
1i
h
2j
_
T
ij
K
1
v
i,j+1/2
v
i+1/2,j
=

12
ij
4
28
When K is diagonal
Products of basis functions for vertical edges by basis functions for horizontal
edges vanish.
If E = E
i+1/2,j
, E :

FA
h
u
F
_

K
1
v
F
v
E
=

TN(E)

FT
u
F
_

K
1
v
F
v
E
=
u
i+1/2,j
_
_
T
ij
1
k
1
ij
v
i+1/2,j
v
i+1/2,j
+
_
T
i+1,j
1
k
1
i,j+1
v
i+1/2,j
v
i+1/2,j
_
+
u
i1/2,j
_
T
ij
1
k
1
ij
v
i1/2,j
v
i+1/2,j
+u
i+3/2,j
_
T
i+1,j
1
k
1
i,j+1
v
i+3/2,j
v
i+1/2,j
If E , say for instance i = 0 (left boundary):

FA
h
u
F
_

K
1
v
F
v
E
=

TN(E)

FT
u
F
_

K
1
v
F
v
E
=
u
1/2,j
_
T
1,j
1
k
1
1j
v
1/2,j
v
1/2,j
+u
3/2,j
_
T
1,j
1
k
1
1,j+1
v
3/2,j
v
1/2,j
29
Using V, H as indices for the vertical and the horizontal edges we can write the
linear system as
_

_
A
V
0
t
D
V
0 A
H

t
D
H
D
V
D
H
0
_

_
_

_
U
V
U
H
P
_

_
=
_

_
F
vV
F
vH
F
q
_

_
Matrices A
V
et A
H
are tridiagonal.
30
FINITE VOLUMES
31
The PDE as a conservation law.
divu = f is a conservation law.
This equation is obtained in two steps:
1. Write mass balance in a small volume T
_
T
u n
. .
echange
=
_
T
f
..
exterior
contribution
_

_
> 0 source
< 0 well
= 0 what goes in goes out
2. Make T innitely small
32
Finite volumes as a method which models the physics
In a nite volume method:
1. skip step 2
2. T is a triangle or a rectangle
3. p is constant over T
4. give a rule to calculate u n on the edges
33
Classical formulation of cell-centered nite volumes
divu = f, u = K

grad p dans
p = p sur
A rectangular
mesh for
i, j + 1
ij i + 1, j
j + 1/2
j 1/2
j
i + 1/2 i 1/2 i E
n
1
T
n
2
h
1i
h
2j
E '
T
c
The unknowns :
p
T
an approximation of the average value of p over the cell T.
u
E
an approximation of
_
E
u n
E
, n
E
= n
1
if E vertical
n
E
= n
2
if E horizontal
34
Rectangular nite volumes
Integrate over the cell T
ij
:
_
T
ij
divu =
_
T
ij
u n =
_
T
ij
f
T
ij
Tn

%
E
i+1/2,j
_
E
i,j1/2
u n +
_
E
i+1/2,j
u n +
_
E
i,j+1/2
u n +
_
E
i1/2,j
u n =
_
T
ij
f
This equation is approximated by
u
i+1/2,j
u
i,j1/2
+u
i,j+1/2
u
i1/2,j
=
_
T
ij
f
A nite volume consists in specifying
how to calculate the u
i+1/2,j
s in terms of the p
ij
s.
35
Assume K is diagonal: K =
_
_
k
1
0
0 k
2
_
_
.
It is natural to approximate u = K

grad p
by u
i+
1
2
,j
= k
1
i+
1
2
,j
p
i+1,j
p
i,j
h
1i
+h
1,i+1
2
h
2j
T
i,j
T
i+1,j
E
i+
1
2
,j
How to choose k
1
i+
1
2
,j
?
k
1
i+
1
2
,j
=
_

_
1
2
(k
1
i,j
+k
1
i+1,j
) the arithmetic average
1
1
2
(
1
k
1
i,j
+
1
k
1
i+1,j
)
the harmonic average
36
Arithmetic average:
1
2
(s +r)
For a xed s:
_
_
_
r

r0

s
2
0 5 10 15 20 25 30 35 40 45 50
0
5
10
15
20
25
30
Harmonic average:
1
1
2
(
1
s
+
1
r
)
For a xed s:
_
_
_
r
2s
r0
0
0 5 10 15 20 25 30 35 40 45 50
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
The harmonic average is better
since it allows no ow to enter a cell with 0 permeability.
37
Combining
u
i+
1
2
,j
u
i
1
2
,j
+u
i,j+
1
2
u
i,j
1
2
=
_
T
ij
f
with
u
i+
1
2
,j
= k
1
i+
1
2
,j
p
i+1,j
p
i,j
h
1,i+1
+h
1i
2
h
2j
we obtain
k
1
i+
1
2
,j
p
i+1,j
p
i,j
h
1,i+1
+h
1i
2
h
2j
+k
1
i
1
2
,j
p
ij
p
i1,j
h
1i
+h
1,i1
2
h
2j
k
2
i,j+
1
2
p
i,j+1
p
i,j
h
2,j+1
+h
2j
2
h
1i
+k
2
i,j
1
2
p
ij
p
i1,j
h
2j
+h
2,j1
2
h
1i
=
_
T
ij
f
38
We divide by h
1i
h
2j
and that gives the ve point scheme:
p
i+1,j
k
1
i+
1
2
,j
h
1i
h
1,i+1
+h
1i
2
p
i1,j
k
1
i
1
2
,j
h
1i
h
1,i
+h
1i1
2
p
i,j+1
k
2
i,j+
1
2
h
2j
h
2,j+1
+h
2j
2
p
i,j1
k
2
i,j
1
2
h
2j
h
2,j
+h
2,j1
2
+p
ij
_
_
k
1
i+
1
2
,j
h
1i
h
1,i+1
+h
1i
2
+
k
1
i
1
2
,j
h
1i
h
1i
+h
1,i1
2
+
k
2
i,j+
1
2
h
2j
h
2,j+1
+h
2j
2
+
k
2
i,j
1
2
h
2j
h
2j
+h
2,j1
2
_
_
=
1
h
1i
h
2j
_
T
ij
f
When h = h
1
= h
2
and K are constant, with k = k
1
= k
2
, this equation results
in the standard equation:
k
h
2
[p
i+1,j
p
i1,j
p
i,j+1
p
i,j1
+ 4p
i,j
] =
1
h
1i
h
2j
_
T
ij
f
39
On the Dirichlet boundary
D
: p = p.
Say for instance for the left boundary:
u1
2
,j
= k
1
1
2
,j
p
1,j
p
j
1
2
h
1,1
h
2j
with k
1
1
2
,j
= k
1
1,j
j
1
The corresponding equation is now
p
2,j
k
1
3
2
,j
h
1,1
h
1,2
+h
1,1
2
p
1,j+1
k
2
1,j+
1
2
h
2,j
h
2,j+1
+h
2,j
2
p
1,j1
k
2
1,j
1
2
h
2j
h
2,j
+h
2,j1
2
+p
1,1
_
_
k
1
3
2
,j
h
1,1
h
1,2
+h
1,1
2
+
k
1
1,j
h
2
1,1
2
+
k
2
1,j+
1
2
h
2,j
h
2,j+1
+h
2,j
2
+
k
2
1,j
1
2
h
2,j
h
2,j
+h
2,j1
2
_
_
=
1
h
1,1
h
2,j
_
T
ij
f +p
j
k
1
1,j
h
2
1,1
2
40
On the Neumann boundary
N
: u n = g.
Say for instance for the left boundary:
u1
2
,j
= gh
2j
j
1
Then we obtain
p
2,j
k
1
3
2
,j
h
1,1
h
1,2
+h
1,1
2
p
1,j+1
k
2
1,j+
1
2
h
2j
h
2,j+1
+h
2j
2
p
1,j1
k
2
1,j
1
2
h
2j
h
2,j
+h
2,j1
2
+p
1,1
_
_
k
1
3
2
,j
h
1,1
h
1,2
+h
1,1
2
+
k
2
1,j+
1
2
h
2j
h
2,j+1
+h
2j
2
+
k
2
1,j
1
2
h
2j
h
2j
+h
2,j1
2
_
_
=
1
h
1,1
h
2j
_
T
ij
f
g
h
1,1
41
Mixed nite elements vs nite volumes
Assume still that K is diagonal.
Use trapezoidal rule to calculate the coecients of A. Then
_
_
T
ij
1
k
1
ij
v
i+1/2,j
v
i+1/2,j
+
_
T
i+1,j
1
k
1
i,j+1
v
i+1/2,j
v
i+1/2,j
_
=
1
2h
2j
_
h
1i
k
1
ij
+
h
1,i+1
k
1
i+1
_
_
T
ij
1
k
1
ij
v
i1/2,j
v
i+1/2,j
= 0
_
T
i+1,j
1
k
1
i,j+1
v
i+3/2,j
v
i+1/2,j
= 0
42
Therefore the matrix A
V
becomes diagonal. Its rows correspond to the
equations:
u
i+
1
2
,j
=
_
k
1
h
1
_
i+
1
2
,j
(p
i+1,j
p
i,j
)h
2j
with
_
k
1
h
1
_
i+
1
2
,j
=
1
1
2
_
h
1i
k
1
ij
+
h
1,i+1
k
1
i,j+1
_
= the harmonic average of
k
1
h
1
.
This formula for u
i+
1
2
,j
is slightly dierent from that given before for a standard
nite volume method.
It is natural since one can realize that the coecient in front of (p
i+1,j
p
i,j
)h
2j
is actually like k
1
/h
1
(and not just k
1
).
It gives better results in cases where there is also a sharp change in h
1
.
43
What did we actually do to obtain nite volumes from mixed
nite elements ?
We approximated a by a
h
such that
a(u
h
, v
h
) =
_

K
1
u
h
v
h
=

TT
h
_
T
K
1
u
h
v
h
a
h
(u
h
, v
h
) =

TT
h
_
T
K
1
u
h
v
h
where
_
T
is an approximate integral over T calculated with the trapezoidal rule in
x
1
for the vertical edges and in x
2
for horizontal edges.
44
The bilinear form a
h
can be rewritten as
a
h
(u
h
, v
h
) =

TT
h

FT

T,F
_
F
u
h
n
F
_
F
v
h
n
F
.
with
T(E),E
=
1
2|E|
h
1
T
(E)
k
1
T
(E)
for a vertical edge E.
This gives a matrix A
h
corresponding to a
h
which is diagonal.
a
h
(v
E
, v
E
) =
T
1
(E),E
+
T
2
(E),E
, T
1
(E), T
2
(E) N(E), if E ,
a
h
(v
E
, v
E
) =
T(E),E
if E ,
a
h
(v
E
, v
F
) = 0 if E = F.
45
The new approximate formulation is
P

mh
_

_
Find u

h
W
hg
and p

h
M
h
such that
a
h
(u

h
, v
h
) b(v
h
, p

h
) = l
W
(v
h
), v
h
W
h0
,
b(u

h
, q
h
) = l
M
(q
h
), q
h
M
h
.
which is equivalent to the cell-centered nite volume formulation on rectangles.
46
The algebraic system is now
_
_
A
h

t
D
D 0
_
_
_
_
U

_
_
=
_
_
F
v
F
q
_
_
.
The row equation associated with v
h
= v
E
reads now
(
T
1
(E),E
+
T
2
(E),E
)u

E
p

T
1
(E)
+p

T
2
(E)
= 0, E ,

T,E
u

E
p

T
1
(E)
+p
E
= 0, E .
One can know eliminate U

to obtain the linear system for P

(DA
1
h
t
D) P = F
q
DA
1
h
F
v
where DA
1
h
t
D is still a sparse matrix (5 diagonals).
Did we lose accuracy by replacing a by a
h
?
47
Le cas des triangles
Base de

W
T
:
v
T,E
i
=
1
2|T|
_
_
x
1
S
i
1
x
2
S
i
2
_
_
=
1
2|T|

S
i
M,
i = 1, 2, 3
Les v
T,E
i
verient
_
E
j
v
T,E
i
n
T
j
=
ij
.
e
e
e
e
e
e
e
e
e
e
e
e
e
e

B
r
r
r
c
S
1
S
3
S
2
E
2
E
3
E
1
n
T
2
n
T
3
n
T
1
M
y
S
i
_
S
i
1
S
i
2
_
M
_
x
1
x
2
_
48
Base de W
h
:
v
E
(x) =
_

_
v
T
1
(E),E
(x) si x T
1
(E)
v
T
2
(E),E
(x) si x T
2
(E)
0 ailleurs
E
49
Volumes nis triangulaires
Comme pour les rectangles on approche a par a
h
de sorte que
a(u
h
, v
h
) =

TT
h
_
T
K
1
u
h
v
h
. .
a
T
(u
h
,v
h
)
a
h
(u
h
, v
h
) =

TT
h
3

i=1

T,E
i
_
E
i
u
h
n
E
i
_
E
i
v
h
n
E
i
.
. .
a
T
h
(u
h
,v
h
)
La matrice de a
h
est diagonale.
50
Trouver u

h
W
h
et p

h
M
h
tels que
a
h
(u

h
, v
h
) b(p

h
, v
h
) = g(v
h
), v
h
W
h
,
b(u

h
, q
h
) = f(q
h
), q
h


M
h
.
Le syst`eme algebrique secrit maintenant
_
_
A
h

t
B
B 0
_
_
_
_
U

_
_
=
_
_
F
F
_
_
.
La ligne du syst`eme correspondant `a q
E
secrit maintenant
(
T
1
(E),E
+
T
2
(E),E
)u

E
p

T
1
(E)
+p

T
2
(E)
= 0,
E ,

T,E
u

E
p

T
1
(E)
+p
E
= 0, E .
51
On peut donc maintenant eliminer U

en maintenant la structure creuse de la

matrice en P

.
Reste `a choisir les coecients
T,E
i
de sorte que la precision ne soit pas aectee.
52
Estimations derreur
a et b sont comme avant, veriant les hypoth`eses de continuite, de V-ellipticite,
et la condition inf-sup .
Theor`eme : Hypoth`eses sur a
h
: il existe A

independantes de h telles que

(H1) a
h
(u
h
, v
h
) A

u
h

W
v
h

W
, u
h
, v
h
W
h
(H2) a
h
(v
h
, v
h
)

v
h

2
W
, v
h
V
h
= {v
h
W
h
| b(v
h
, q
h
) = 0, q
h
M
h
}.
Alors il existe C telle que
u u

W
+p p

M
C
_
inf
v
h
W
h
_
u v
h

W
+ sup

h
W
h
a(v
h
,
h
) a
h
(v
h
,
h
)

W
_
+ inf
q
h
M
h
(p q
h

M
)
_
.
53
On connait dej`a les erreurs dinterpolation :
inf
q
h
M
h
pq
h

M
Chp
H
1
()
, inf
v
h
W
h
(uv
h

W
) Ch(u
H
1
()
+divu
H
1
()
).
Il reste `a verier les hypoth`eses (H1) et (H2) pour appliquer le theor`eme, et `a
evaluer lerreur a a
h
.
54
Verication de lhypoth`ese (H2)
Les coecients
T,E
1
=
1
4|T|
(K
1
T

S
1
S
3
)

S
2
S
1
,

T,E
2
=
1
4|T|
(K
1
T

S
2
S
1
)

S
3
S
2
,
T,E
3
=
1
4|T|
(K
1
T

S
3
S
2
)

S
1
S
3
.
sont choisis de sorte que
Lemme :
a
T
(u
h
, v
h
) = a
T
h
(u
h
, v
h
), u
h
, v
h
V
T
= {v W
T
; divv = 0} = (P
0
(T))
2
. (1)
Corollaire : lhypoth`ese (H2) est veriee.
En eet, pour v
h
V
h
,
a
h
(v
h
, v
h
) =

TT
h
a
T
h
(v
h
, v
h
) =

TT
h
a
T
(v
h
, v
h
) = a(v
h
, v
h
) v
h

2
W
.
55
Verication de lhypoth`ese (H1)
On va montrer que
|a
T
h
(u
h
, v
h
)| Cu
h

0,T
v
h

0,T
, u
h
, v
h
W
h
(2)
ce qui implique (H1).
Passage `a lelement de reference : etc...
56
Estimation de lerreur a a
h
Lemme :
inf
v
0
(P
0
(T))
2
v
h
v
0

0,T
Chdivv
h

0,T
, v
h
W
h
. (3)
En eet, soit v
0
V
T
tel que v
h
= v
0
+(v
T,1
+v
T,2
+v
T,3
) avec =
|T|divv
h
3
.
v
h
v
0

2
0,T
=
2
1
4|T|
2
_
T
(

S
1
M +

S
2
M +

S
3
M)
2
=
2
1
12|T|

i=1,2,3
(

S
1
M
1
+

S
2
M
2
+

S
3
M
3
)
2
o` u M
1
, M
2
, M
3
sont les milieux des aretes opposees aux sommets S
1
, S
2
, S
3
.
57
etc...
Donc
v
h
v
0

2
0,T
=
1
144
(
3

i=1
|E
i
|
2
)divv
h

2
0,T
ce qui implique le lemme.
58
On peut maintenant estimer a a
h
.
Proposition :
inf
v
h
W
h
sup

h
W
h
a(v
h
,
h
) a
h
(v
h
,
h
)

W
Chu
W
. (4)
Preuve : Pour tout v
h
,
h
W
T
, il existe v
0
,
0
V
T
veriant
v
h
v
0

0,T
Chdivv
h

0,T
,
h

0,T
Chdiv
h

0,T
. (5)
Alors
a
T
(v
h
,
h
) a
T
h
(v
h
,
h
) = a
T
(v
h
,
h

0
) +a
T
(v
h
v
0
,
0
) +a
T
(v
0
,
0
)
[a
T
h
(v
h
,
h

0
) +a
T
h
(v
h
v
0
,
0
) +a
T
h
(v
0
,
0
)]
59
Legalite (1) et les inegalites (2),(5) impliquent
|a
T
(v
h
,
h
) a
T
h
(v
h
,
h
)| C(v
h

0,T

0,T
+v
h
v
0

0,T

0,T
)
Ch
T
(v
h

W
T

W
T
).
Donc, pour tout v
h
,
h
W
h
on a
a(v
h
,
h
) a
h
(v
h
,
h
)| Ch(v
h

W
),
sup

h
W
h
a(v
h
,
h
) a
h
(v
h
,
h
)

W
Chv
h

W
.
Si v
h
=
h
u alors
sup

h
W
h
a(
h
u,
h
) a
h
(
h
u,
h
)

W
Ch
h
u
W
Chu
W
et linegalite (4) est vraie.
60
Fin des estimations derreur
Grace aux estimations dinterpolation et `a lestimation (4), le theor`eme 1 nous
permet de conclure :
u u

W
+p p

M
Ch(p
1,
+u
1,
+divu
1,
).
Remarque : Pour que lanalyse ci-dessous fonctionne il faut que les coecients

T,E
i
, i = 1, 2, 3 soient strictement positifs
= les angles des triangles de T
h
doivent etre tous aigus.
61
Qualities and diculties of the nite volume method
On rectangles it is simple, leads to a positive-denite matrix, and satises
the maximum principle
Dicult to formulate for diormed rectangles and triangles
For triangles restriction on the meshes
Complicated formulation when K is not diagonal
62
OTHER CELL-CENTERED DISCRETIZATION METHODS
63
Mixed-hybrid nite elements
Instead of calculating u
h
W
h
, we now calculate
u

h
W

h
= {v
h
(L
2
())
2
; v
h
|T W
T
, T T
h
}
Functions of W

h
are not required to have their ux continuous across the edges.
Continuity of the ux will now be written explicitely.
We need also
N
h
= {
h

EA
h

E
,
E
R}.
64
The mixed-hybrid formulation is
Find u

h
W

hg
, p

h
M
h
,
h
N
h
such that
_
T
K
1
u

h
v
h

_
T
p

h
divv
h
+

ET
_
E

h
v
h
n
T
=
_

D
pv
h
n
T
,
v
h
W

h
, T T
h
_
T
div u
h
q
h
=
_
T
f q
h
, q
h
M
h
, T T
h

TT
h
,TE
u

h
n
T

h
= 0, E A
h
, E ,
h
N
h
u

h
n|
E
= g|E|, E
N
,

h
|
E
= p, E
D
.

h
represents a trace of the pressure on the edges E A
h
.
We check easily that p

h
= p
h
, u

h
|
T
= u
h
|
T
, T T
h
.
65
The linear system
_

_
A

t
D
t
B
D 0 0
B 0 I
D
_

_
_

_
U

_
=
_

_
F
v
F
q
F

_
A

is block diagonal; we can eliminate U

= A
(1)
(F
v
+
t
DP +
t
B) to get
_
_
DA
(1) t
D DA
(1) t
B
BA
(1) t
D BA
(1) t
B +I
D
_
_
_
_
P

_
_
=
_
_
F
q
DA
(1)
F
v
F

BA
(1)
F
v
_
_
The matrix DA
(1) t
D is diagonal, so we can eliminate P:
P = (DA
(1) t
D)
1
[F
q
DA
(1)
F
v
(DA
(1) t
B)] to obtain
H = G (H sparse)
66
where H = BA
(1) t
B +I
D
(BA
(1) t
D)(DA
(1) t
D)
1
(DA
(1) t
B)
G = F

BA
(1)
(BA
(1) t
D)(DA
(1) t
D)
1
(F
q
DA
(1)
F
v
)
67
Properties of the matrix H
H is sparse
The number of nonzeros in the line E is equal to the number of neighbouring
edges + 1 (for E itself) (7 for a rectangular mesh).
H is positive denite
To prove it, assuming (H, ) = 0 we have to show that this implies = 0.
Then
((BA
(1) t
B+I
D
), ) ((BA
(1) t
D)(DA
(1) t
D)
1
(DA
(1) t
B), ) = 0
Introduce P = (DA
(1) t
D)
1
((DA
(1) t
B)). We obtain
(A
(1) t
B,
t
B) + (I
D
, ) (A
(1) t
DP, B) = 0
68
But equation for P implies that
((DA
(1) t
D)P, P) + ((DA
(1) t
B), P) = 0.
Adding to the previous equation gives
(A
(1)
(
t
DP +
t
B),
t
DP +
t
B) + (I
D
, ) = 0
Since A
(1)
is positive denite and I
D
is positive semi-denite, this implies that
t
DP +
t
B = 0 and
E
= 0, E
D
.
Equation
t
DP +
t
B = 0 says actually that
P
T

E
= 0, E T, T T
h
which means that the pressure is constant over .
But from
E
= 0, E
D
it follows that P = 0 and = 0.
69
A PROJECT
70
The problem
div u = f, u = K

grad p, in = (0.1) (0, 1)
p = 100 on
1
, u n = 0 on
2
, u n = 10 on
3
.
3/4

3
2
2
1
1/4
1/4
3/4
1. Solve for K =
_
_
10 0
0 10
_
_
, K =
_
_
10 0
0 100
_
_
, K =
_
_
10 9
9 10
_
_
2. Solve for K =
_
_
1 0
0 1
_
_
in the shaded area, K =
_
_
10 0
0 10
_
_
outside the
shaded area.
3. Calculate the errors p p
h

0
, u u
h

0
.
71
A remark on checking the program
One check of the program follows from choosing u W
h
.
Since W
h
W, M
h
M problem P
m
implies
_
_
_
a(u, v
h
) b(v
h
, p) = l
W
(v
h
), v W
h0
,
b(u, q
h
) = l
M
(q
h
), q
h
M
h
.
Denoting
h
p the L
2
-projection of p, since div is a mapping from W
h
onto M
h
,
we have
_

divv
h
(p
h
p) = 0 for all v
h
W
h
.
Therefore u W
h
,
h
p M
h
is the unique solution to the equations :
_
_
_
a(u, v
h
) b(v
h
,
h
p) = l
W
(v
h
), v W
h0
,
b(u, q
h
) = l
M
(q
h
), q
h
M
h
.
72
Thus, to check the program,
1. choose u W
h
2. calculate directly
_
E
u n
E
, E A
h
as well as
1
|T|
_
T
p, T T
h
.
These numbers should be equal to u
E,EA
h
, p
T,TT
h
calculated by the program
up to the machine accuracy.
73