You are on page 1of 21

ECOLE POLYTECHNIQUE

CENTRE DE MATH

EMATIQUES APPLIQU

EES
UMR CNRS 7641
91128 PALAISEAU CEDEX (FRANCE). Tel: 01 69 33 41 50. Fax: 01 69 33 30 11
http://www.cmap.polytechnique.fr/
An optimized Schwarz algorithm for
the compressible Euler equations
Victorita Dolean and Frederic Nataf
R.I. N
0
556 October 2004
An optimized Schwarz algorithm for the compressible Euler
equations
V. Dolean

and F. Nataf

5th October 2004
Abstract
In this work we design new interface transmission conditions for a domain decomposition Schwarz
algorithm for the Euler equations in 2 dimensions. These new interface conditions are designed to
improve the convergence properties of the Schwarz algorithm. These conditions depend on a few
parameters and they generalize the classical ones. Numerical results illustrate the eectiveness of the
new interface conditions.
1 Introduction
When solving the compressible Euler equations by an implicit scheme the nonlinear system is usually
solved by Newtons method. At each step of this method we have to solve a linear system which is non-
symmetric and very ill conditioned. That is why a domain decomposition method is used. In a previous
paper [DLN04] we formulated a Schwarz algorithm (interface iteration which relies on the successive
solving of the local decomposed problems and the transmission of the result at the interface) involving
transmission conditions that are derived naturally from a weak formulation of the underlying boundary
value problem. We also studied the convergence of the proposed algorithm from a quantitative point
of view in the two and three dimension overlapping and non-overlapping cases by applying a Fourier
analysis. For the sake of the analysis we limited ourselves to the cases of two and three subdomain
decompositions and we provided analytical expressions of the convergence rate of the Schwarz algorithm
applied to the linearized equations.
Various works and studies exist when dealing with Schwarz algorithms applied to the scalar problems:
classical and optimized transmission conditions as well as preconditionning aspects were treated for ex-
ample in the case of the Poisson, advection-diusion equations [JN00, JNR01], Helmholtz [GMN02], or
other simple systems reducible to scalar problems, on conforming and non-conforming meshes. There are
also such methods for linear systems such as time harmonic Maxwell equations [CDJP97][DJR92] [AG04]
or on the linear elasticity. To our knowledge, little is known about complex systems. When dealing with
systems we can mention some classical works by Quarteroni and al. [Qua90] [QS96], Bjorhus [Bj95] or
Cai et al. [CFS98]. As far as the optimized interface conditions are concerned, we can mention [DLN02]
based on the Smith factorization . The most related work to our study belongs to Clerc [Cle98] and it
describes the principle of building very simple interface conditions for a general hyperbolic system which
we will apply and extend to Euler system.

Universite d

Evry, Dept. Maths., 91000



Evry Cedex and CMAP, CNRS UMR7641,

Ecole Polytechnique, 91128 Palaiseau


Cedex, France, dolean@cmapx.polytechnique.fr, http://www.cmap.polytechnique.fr/dolean

CMAP, CNRS UMR7641, Ecole Polytechnique, 91128 Palaiseau, France, nataf@cmapx.polytechnique.fr,


http://www.cmap.polytechnique.fr/nataf
1
In this work we formulate and analyze the convergence of the Schwarz algorithm with new interface
conditions inspired by [Cle98], depending on 2 parameters whose value is determined by minimizing the
norm of the convergence rate.
The paper is organized as follows. In the section 2 we rst formulate the Schwarz algorithm for a
general linear hyperbolic system of PDEs with general interface conditions built in order to have a well-
posed problem. The convergence rate is computed in the Fourier space as a function of some parameters.
In the section 3 we present the discretization method as well as the discrete counterpart of the con-
sidered problem. We will further estimate the convergence rate at the discrete level. We will nd the
optimal parameters of the interface conditions at the discrete level.
In the section 4, we use the new optimal interface conditions in Euler computations which illustrate the
improvement over the classical interface conditions (rst described in [QS96]). An appendix containing
the solution of the optimization problem in the non-overlapping case, where we can obtain some analytical
results, concludes this work.
2 A Schwarz algorithm with general interface conditions
2.1 A well-posed boundary value problem
In this section we briey review the main denitions and properties of hyperbolic systems of conservation
laws that are of interest to our study. Then we introduce a Schwarz algorithm which is based on general
transmission conditions at subdomain interfaces that take into account the hyperbolic nature of the
problem. In addition, we recall some existing results concerning the convergence of the algorithm. We
consider here a general non-linear system of conservation laws which has the form:
(1)
W
t
+
d

i=1
F
i
(W)
x
i
= 0, W R
q
where d denotes the space dimension and q the dimension of the system. The ux functions F
i
are
assumed dierentiable with respect to the state vector W = W(x, t). In the general case, the ux
functions are non-linear functions of W. Under the hypothesis that the solution is regular, we can also
write a non-conservative (or quasi-linear) equivalent form of equation (1) :
(2)
W
t
+
d

i=1
A
i
(W)
W
x
i
= 0
where the A
i
are the Jacobian matrices of the ux vectors. Assume that we rst proceed to an integration
in time of (1) using a backward Euler implicit scheme involving a linearization of the ux functions and
eventually we symmetrize it (we know that when the system admits an entropy it can be symmetrized
by multiplying it by the hessian matrix of this entropy). This operation results in the linearized system:
(3) L(W)
Id
t
W +
d

i=1
A
i
W
x
i
= f
where W W
n+1
W
n
and W
n+1
= W(x, (n + 1)t), and A
i
is a shorthand for A
i
(W
n
).
In the following we will dene the boundary conditions that have to be imposed when solving the
problem on a domain R
d
. We denote by A
n
=
d

i=1
A
i
n
i
, the linear combination of jacobian matrices
2
by the components of the outward normal vector at the boundary of the domain . This matrix is real,
symmetric and can be diagonalized
A
n
= T
n
T
1
,
n
= diag(
i
)
It can also be splitted in negative and positive part using this diagonalization
_
_
_
A
n
= A
+
n
+ A

n
A

n
= T

n
T
1

+
n
= diag(max(
i
, 0)),

n
= diag(min(
i
, 0))
This corresponds to a decomposition with local characteristic variables. A more general splitting in
negative(positive) denite parts, A
neg
n
and A
pos
n
of A
n
can be done such that these matrices satisfy the
following properties:
(4)
_
_
_
A
n
= A
neg
n
+ A
pos
n
rank(A
neg,pos
n
) = rank(A

n
)
A
pos
n
= A
neg
n
In the scalar case the only possible choice is A
neg
n
= A

n
. Using the previous formalism we can dene
the following boundary condition:
A
neg
n
W = A
neg
n
g, on
Remark 1 In the case of a classical decomposition in negative and positive part this boundary condition
has the physical meaning of the incoming ux in domain . By extension of the properties found in this
case we call the last equality of (4) conservation property because it insures that the out-ow quantity
(given by the positive part of the jacobian ux matrix with oposite direction of the normal) is retrieved
out of the in-ow quantity imposed by the boundary condition (given the negative part of the jacobian
ux matrix).
Within this framework we have the following result concerning the boundary value problem associated
to the system that can be found in [Cle98] :
Theorem 1 If f L
2
()
q
and g satises

A
neg
n
g g

< and there exists a constant C


0
independent
of x such that the following inequality be respected in the sense of symmetric positive denite matrices:
_
Id
t

xi
A
i
_
C
0
Id > 0
then there exists a unique, W L
2
()
q
with

A
i

xi
W L
2
()
q
solution of the boundary value problem:
(5)
_

_
L(W) =
Id
t
W +
d

i=1
A
i

xi
W = f in
A
neg
n
W = A
neg
n
g on
The unique solution W of (5) satises the estimate:
(6) C
0
|W|
2
L
2
()
+
_

A
pos
n
W W
_

A
neg
n
W W
As the boundary value problem (5) is well-posed, the decomposition (4) enables the design of a domain
decomposition method.
3
2.2 Schwarz algorithm with general interface conditions
We consider a decomposition of the domain into N overlapping or non-overlapping subdomains

=
N
_
i=1

i
. We denote by n
ij
the outward normal to the interface
ij
bewteen
i
and a neighboring subdomain

j
. Let W
(0)
i
denote the initial appoximation of the solution in subdomain
i
. A general formulation of
a Schwarz algorithm for computing (W
p+1
i
)
1iN
from (W
p
i
)
1iN
(where p denes the iteration of the
Schwarz algorithm) reads :
(7)
_

_
LW
p+1
i
= f in
i
A
neg
nij
W
p+1
i
= A
neg
nij
W
p
j
on
ij
=
i

j
A
neg
nij
W
p+1
i
= A
neg
nij
g on
i
where A
neg
nij
and A
pos
nij
satisfy (4). We have the following result concerning the convergence of the Schwarz
algorithm in the non-overlapping case, due to([Cle98]):
Theorem 2 If we denote by E
p
i
= W
p
i
W
i
the error vector associated to the restriction to the i-th
subdomain of the global solution of the problem. Then, the Schwarz algorithm converges in the following
sense :
_

_
lim
p
|E
p
i
|
L
2
(i)
q = 0
lim
p
|
d

j=1
A
j

j
E
p
i
|
L
2
(i)
q = 0
The convergence rate of the algorithm dened by (7) depends of the choice of the decomposition of
A
n
ij
into a negative and a positive part satisfying (4). In order to choose the decomposition (4) we need
to relate this choice to the convergence rate of (7).
2.3 Convergence rate of the algorithm with general interface conditions
We consider a two-subdomain non-overlapping or overlapping decomposition of the domain = R
d
,

1
=] , [R
d1
and
2
=], [R
d1
with and study the convergence of the Schwarz
algorithm in the subsonic case. A Fourier analysis applied to the linearized equations allows us to derive
the convergence rate of the -th Fourier component of the error. We will rst briey recall the technique
of Fourier transform which was already described in detail in [DLN04]. The vector of Fourier variables is
denoted by = (
j
, j = 2, . . . , d). Let (E
p
i
)(x) = (W
p
i
W
i
)(x) be the error vector in the ith subdomain
at the pth iteration of the Schwarz algorithm and:

E(x
1
,
2
, . . . ,
d
) = TE(x
1
,
2
, . . . ,
d
) =
_
R
d1
e
i2x2...i
d
x
d
E(x
1
, . . . , x
d
)dx
2
. . . dx
d
the Fourier symbol of the error vector. This transformation is useful only if the A
i
matrices are constant
which is the case here because we have considered the linearized form of the Euler equations around a
constant state

W. The Schwarz algorithm in the Fourier space ( R
d1
) can be written as follows:
(8)
_

_
d
dx
1

E
p+1
1
= M()

E
p+1
1
, x <
/
neg
(

E
p+1
1
) = /
neg
(

E
p
2
), on x =
_

_
d
dx
1

E
p+1
2
= M()

E
p+1
2
, x >
/
pos
(

E
p+1
2
) = /
pos
(

E
p
1
), on x =
4
where we denoted by /
neg
= A
neg
n
, /
pos
= A
pos
n
with n = (1, 0) the outward normal to the domain
1
and:
(9) M() = A
1
1
_
1
t
Id +
d

i=2
A
i

i1
_
We thus obtain local problems that for a given are very simple ODEs whose solutions can be expressed
as linear combinations of the eigenvectors of M() :
(10)

E
p
i
(x
1
, ) =
q

j=1

i,p
j
()e
j()x1
V
j
()
where
j
() are the eigenvalues of M(). Here we have assumed that the eigenvectors V
j
() of M()
are linearly independent. Furthermore, we require that these solutions are bounded at innity ( and
+ respectively). We deduce that in the decomposition of

E
1
(x
1
, ) (respectively

E
2
(x
1
, )) we must
keep only the eigenvectors corresponding to the negative (respectively the positive) real parts of the
eigenvalues. Taking into account these considerations we replace the expressions of the local solutions
(10) into the interface conditions (8) to obtain the interface iterations on the coecients:
_

_
(
1,p+1
j
)
j,(j )<0
() = T
1
_
(
2,p
j
)
j,(j )>0
()
_
(
2,p+1
j
)
j,(j )>0
() = T
2
_
(
1,p
j
)
j,(j )<0
()
_
Then, the convergence rate of the -th component of the error vector of the Schwarz algorithm can
be computed as the spectral radius of one of the iteration matrices T
1
T
2
() or T
2
T
1
():

2
2

2
Schwarz2
= (T
1
T
2
) = (T
2
T
1
)
2.4 The 2D Euler equations
After having dened in a general frame the well-possedness of the boundary value problem associated to
a general equation and the convergence of the Schwarz algorithm applied to this class of problems, we
will concentrate ourselves on the conservative Euler equations in two-dimensions:
(11)
W
t
+.F(W) = 0 , W = (, V , E)
T
, =
_

x
,

y
_
T
.
In the above expressions, is the density, V = (u, v)
T
is the velocity vector, E is the total energy
per unit of volume and p is the pressure. In equation (11), W = W(x, t) is the vector of conservative
variables, x and t respectively denote the space and time variables and F(W) = (F
1
(W) , F
2
(W))
T
is
the conservative ux vector whose components are given by
F
1
(W) =
_
_
_
_
u
u
2
+ p
uv
u(E + p)
_
_
_
_
, F
2
(W) =
_
_
_
_
v
uv
v
2
+ p
v(E + p)
_
_
_
_
.
The pressure is deduced from the other variables using the state equation for a perfect gas p =
(
s
1)(E
1
2
| V |
2
) where
s
is the ratio of the specic heats (
s
= 1.4 for the air).
5
2.5 A new type of interface conditions
We will apply now the method described previously to the computation of the convergence rate of the
Schwarz algorithm applied to the two-dimensional subsonic Euler equations. In the supersonic case there
is only one decomposition satisfying (4), that is: /
pos
= A
n
and /
neg
= 0 and the convergence follows
in 2 steps. Therefore the only case of interest is the subsonic one.
The starting point of our analysis is given by the linearized form of the Euler equations (11) which
are of the form (3) where we replace W by W and to whom we applied a change of variable

W = T
1
W
based on the eigenvector factorization of A
1
= T

A
1
T
1
. In the following we will abandon the symbol):
W
ct
+ A
1

x
W + A
2

y
W = 0
characterized by the following jacobian matrices:
(12) A
1
= diag(M
n
1 , M
n
+ 1 , M
n
, M
n
) A
2
=
_
_
_
_
_
_
_
_
_
_
M
t
0
1

2
0
0 M
t
1

2
0
1

2
1

2
M
t
0
0 0 0 M
t
_
_
_
_
_
_
_
_
_
_
where M
n
=
u
c
, M
t
=
v
c
denote respectively the normal and the tangential Mach number. Before
estimating the convergence rate we will derive the general transmission conditions at the interface by
splitting the matrix A
1
into a positive and negative part.
We have the following general result concerning this decomposition:
Lemma 1 Let
1
= M
n
1,
2
= M
n
+ 1,
3
=
4
= M
n
. Suppose we deal with a subsonic ow:
0 < u < c so that
1
< 0,
2,3,4
> 0. Any decomposition of A
1
= A
n
, n = (1, 0) which satises (4) has
to be of the form:
/
neg
=
1
a
1
u u
t
, u = (a
1
, a
2
, a
3
, a
4
)
t
/
pos
= A
n
/
neg
.
where (a
1
, a
2
, a
3
, a
4
) R
4
satises a
1

1
< 0 and
a
1

1
+
a
2
2
a
1

2
+
a
2
3
a
1

3
+
a
2
4
a
1

4
= 1.
Proof The fact that each 1-rank symmetric matrix is of the form v v
t
is straightforward. In order to
have a negative matrix we need to take: /
neg
= v v
t
with v = (c
1
, c
2
, c
3
, c
4
)
t
. Indeed, for each vector
x = (x
1
, x
2
, x
3
, x
4
)
t
we have:
x
t
/
neg
x = (x
t
v)
2
0
A necessary condition for /
pos
to be of rank 3 is det(A
n
/
neg
) = 0, that is:
(13) 0 =
c
2
1

1
+
c
2
2

2
+
c
2
3

3
+
c
2
4

3
+ 1 1 +
c
2
1

1
which implies that c
1
,= 0. Thus, without loss of generality and in order to simplify the writing of the
interface conditions, in the sequel we will take /
neg
=
1
a
1
u u
t
with u = (a
1
, a
2
, a
3
, a
4
)
t
. On the other
hand,
(14) /
pos
= A
n
/
neg
6
is of a maximum rank 3 i det(/
pos
) = 0, that is:
(15)
a
1

1
+
a
2
2
a
1

2
+
a
2
3
a
1

3
+
a
2
4
a
1

3
= 1.
In the same time from A
n
= /
pos
+/
neg
and the fact that rank(/
neg
) = 1 and rank(A
n
) = 4 we infer
that rank(/
pos
) 3, therefore rank(/
neg
) = 3. Relation (15) also implies that

1
a
1
=

1
a
1
_
a
2
2

2
+
a
2
3

3
+
a
2
4

3
_
0.
Under these hypothesis we can show that /
pos
is positive. First of all we can see that if a
1
=
1
this result is obvious as we are in the case of the classical transmission conditions. Suppose now that
a
1
] ,
1
[. Then, by using the above relation and the Cauchy-Schwarz inequality applied to the
vectors
_
a
2

2
,
a
3

3
,
a
4

4
_
and
_
x
2

2
, x
3

3
, x
4

4
_
:
a
1

1
(
1
a
1
)(
2
x
2
2
+
3
x
2
3
+
3
x
2
4
) =
_
a
2
2

2
+
a
2
3

3
+
a
2
4

3
_
(
2
x
2
2
+
3
x
2
3
+
3
x
2
4
) (a
2
x
2
+ a
3
x
3
+ a
4
x
4
)
2
From the previous inequality we have a minoration of
2
x
2
2
+
3
x
2
3
+
3
x
2
4
by a quadartic term in y =
a
2
x
2
+ a
3
x
3
+ a
4
x
4
and then by using it
x
t
/
pos
x = (
1
a
1
)x
2
1
2x
1
y
1
a
1
y
2
+
2
x
2
2
+
3
x
2
3
+
3
x
2
4
(
1
a
1
)x
2
1
2x
1
y +
y
2

1
a
1
=
_
_

1
a
1
x
1

1
a
1
_
2
0.
we get the desired result.
We will proceed now to the estimation of the convergence rate using some results from [DLN04]. The
matrix M corresponding to (9) writes as:
(16) M() =
_
_
_
_
_
_
_
_
_
_
_
_
a
M
n
1
0
i

2(M
n
1)
0
0
a
1 + Mn
i

2(1 + Mn)
0
i

2M
n
i

2M
n
a
M
n
0
0 0 0
a
u
_
_
_
_
_
_
_
_
_
_
_
_
with =
1
ct
, a = + iM
t
. We obtain the following expressions for the eigenvalues and the corre-
sponding eigenvectors of the matrix M() :

1
() =
aM
n
R()
1 M
2
n
, V
1
() =
_

(R() + a)(1 + M
n
)

2
,
(R() a)(1 M
n
)

2
, i(1 M
2
n
), 0
_
T

2
() =
aM
n
+ R()
1 M
2
n
, V
2
() =
_
(R() a)(1 + M
n
)

2
,
(R() + a)(1 M
n
)

2
, i(1 M
2
n
), 0
_
T

3,4
() =
a
M
n
, V
3
() =
_

iM
n

2
,
iM
n

2
, a, 0
_
T
, V
4
() = [0, 0, 0, 1]
T
7
where R() =
_
a
2
+
2
(1 M
2
n
). We recall that we made the assumption that the ow is subsonic i.e.
M < 1; this also means that M
n
< 1 and M
t
< 1 since M
2
= M
2
n
+M
2
t
. Finally we also assume that the
ow is such that u > 0, in other words we have that 0 < M
n
< 1. Under the assumption 0 < u < c we
have that '(
1
()) < 0 and '(
2,3,4
()) > 0.
Following the technique described in section 2.3 we estimate the convergence rate in the non-overlapping
case and we use the non-dimensioned wave-number

= ct and if we drop the bar symbol, we get for
the general interface conditions the following:
(17)
_

2
2,novr
() =

1
4M
n
(1 M
n
)(1 + M
n
)R()a
2
1
(a + M
n
R())
D
1
D
2

D
1
= R()[a
1
(1 + M
n
) a
2
(1 M
n
)] + a[a
1
(1 + M
n
) + a
2
(1 M
n
)] i

2a
3
(1 M
2
n
)
D
2
= M
n
a
1
[R()[a
1
(1 + M
n
) a
2
(1 M
n
)] + a[a
1
(1 + M
n
) + a
2
(1 M
n
)]]
+ a
3
(1 M
2
n
)[a
3
(R + a) iM
n
a
1

2]
Remark 2 The expression (17) gives the convergence rate in the classical case for a
1
= (1 M
n
) =

1
(0) and a
2
= a
3
= a
4
= 0, which corresponds to the classical transmission conditions. Moreover,
theorem 2 proves that this quantity is always strictly inferior to 1 as the algorithm is convergent.
In order to simplify our optimization problem we will take a
3
= 0, we can thus reduce the number
of parameters to 2, a
1
and a
2
, as we can see from (15) that a
4
can be expressed as a function of a
1
, a
2
and a
3
. We can also see that the convergence rate is a real quantity when the ow is normal to the
interface M
t
= 0. In the same time for the optimization purpose only we introduce the parameters:
b
1
= a
1
/(1 M
n
) and b
2
= a
2
/(1 + M
n
) which provide a simpler form of the convergence rate:
(18)
2
2,novr
() =

1
4b
1
(a + M
n
R())R()
(R()(b
1
+ b
2
) + a(b
1
b
2
))
2
(Mn + 1)

From (15) we get the intervals in which the 2 new parameters lie:
(19)
_

_
b
1
1
1
=]1, [, b
2
1
2
(b
1
)
b
2
2

1 M
n
1 + M
n
b
1
(b
1
1) 1
2
(b
1
) =
_

_
1 M
n
1 + M
n
b
1
(b
1
1),
_
1 M
n
1 + M
n
b
1
(b
1
1)
_
Before proceeding to the analysis of the general case we recall some results found in the classical case
obtained in [DLN04]. The asymptotic convergence rate in the non-overlapping case:
(20) lim
k+

2,novr
(k) =

_
1 3M
n
1 + M
n
_
2
+
8M
n
M
2
t
(1 + M
n
)
3
< 1
is always strictly inferior to 1. Moreover, in the particular case M

n
= 1/3 and M
t
= 0, this limit becomes
null. The inequality (20) has a numerical meaning. For a given discretization, let
max
denote the largest
frequency supported by the numerical grid. This largest frequency is of the order /h with h a typical
mesh size. The convergence rate in a numerical computation made on this grid can be estimated by

h
2
= max
||<max

2
(). From (20), we have that
h
2
max
||<max

2
() < 1. This means that for ner
and ner grids, the number of iterations may increase slightly but should not go to innity. Thus the
optimization problem with respect to the parameters b
1
and b
2
, makes sense:
(21) min
(b1,b2)I1I2(b1)
max
0
()
8
The solving of this problem is quite a tedious task even in the non-overlapping case, where we can
obtain analytical expression of the parameters only for some values of the Mach number (see the appendix
for details). In the same time, we have to analyze the convergence of the overlapping algorithm. Indeed,
standard discretizations of the interface conditions correspond to overlapping decompositions with an
overlap of size = h, h being the mesh size, as seen in [CFS98] and [DLN04]. By applying the procedure
described in section 2.3 to the overlapping case we have the following expression of the convergence rate:
(22)
_

2
2,ovr
=

Ae
(2(k)1(k))

+ (B + C)e
(3(k)1(k))

A =
a + M
n
R()
a M
n
R()

_
b
1
(R() a) + b
2
(R() + a)
b
1
(R() + a) + b
2
(R() a)
_
2
B =
2M
n
(b
1
(1 M
n
) + b
2
(1 + M
n
))R()(R() a)(R() + a)
(1 M
2
n
)(a M
n
R())(b
1
(R() + a) + b
2
(R() a))
2
C =
4((1 M
n
)(b
2
1
b
1
) b
2
2
(Mn + 1))(a + M
n
R())
(1 M
2
n
)(b
1
(R() + a) + b
2
(R() a))
2
where

=

ct
denotes the non-dimensioned overlap between subdomains.
Analytic optimization with respect to b
1
and b
2
seems out of reach. We will have to use numerical
procedures of optimization. In order to get closer to the numerical simulations we will estimate the conver-
gence rate for the discretized equations with general transmission conditions, both in the non-overlapping
and the overlapping case and then optimize numerically this quantity in order to get the best parameters
for the convergence.
3 Optimized interface conditions
In this section we study the convergence of the Schwarz algorithm with general interface conditions
applied to the discrete Euler equations. First, we consider a well-posed boundary value problem dened
on a half plane as described in [DLN04]. This BVP is discretized using a nite volume scheme where the
ux at the interface of the nite volume cells is computed using a Roe [Roe81] type solver. Afterwards,
we formulate a Schwarz algorithm whose convergence rate is estimated in a discrete context.
3.1 Discretization by a nite volume method
We consider rst the following BVP dened on the domain
1
=] , [R
(23)
_
_
_
W
ct
+ A
1
W
x
+ A
2
W
y
= f, for x <
/
neg
W = g for x = ,
In order to discretize the BVP (23) we consider a regular quadrilateral grid where a vertex v
ij
is charac-
terized by
v
ij
=
__
i
1
2
_
x ,
_
j
1
2
_
y
_
for i 0 and j Z.
We associate to each vertex a nite volume cell, C
ij
= [(i 1)x , ix] [(j 1)x , jx] which is a
rectangle having as a center the vertex v
ij
. A rst order vertex centered nite volume formulation for
the discretization of (23) simply writes (see for example [Cle98])
(24)
W
i,j
ct
+
1
[C
ij
[

eCij
[e[
e
= f,
9
where [C
ij
[ denotes the area of the cell C
ij
, [e[ the length of the edge e and W
i,j
the average value of the
unknown on the cell C
ij
W
i,j
=
1
[C
ij
[
_
Cij
W(x, y)dxdy.
Here, the elementary ux
e
ij
across edge [e[ is computed by a Roe type scheme

e
= A
+
n
W
i,j
+ A

n
W
k,l
,
where n = (n
x
, n
y
) is the outward normal to the the edge e, A
n
= n
x
A
1
+ n
y
A
2
and C
kl
is the
neighboring cell of C
ij
sharing the edge e with it. In the present case, we easily see that C
kl
is such that
(k, l) (i 1, j), (i + 1, j), (i, j 1), (i, j + 1) and the four edges of a cell have the following lengths
and outward normal vectors
[e
1
[ = x , n
1
= ( 0, 1), [e
2
[ = y , n
2
= ( 1, 0)
[e
3
[ = x , n
3
= ( 0, 1), [e
4
[ = y , n
4
= (1, 0)
which allows us to rewrite (24) as
(25)
W
ij
ct
+
[A
1
[W
i,j
+ A

1
W
i+1,j
A
+
1
W
i1,j
x
+
[A
2
[W
i,j
+ A

2
W
i,j+1
A
+
2
W
i,j1
y
= f, i l
2
..
where = l
2
x. We will further denote

x =
x
ct
and

y =
y
ct
, the non dimensionned counterpart
of the mesh size in x and y directions.
In the following we will detail the equations (25) in order to emphasize the use of the new boundary
conditions (here we denoted by w
l
i,j
the l-th component of the vector W
i,j
):
(26)
_

_
w
1
i,j

1 M
n
x
(w
1
i+1,j
w
1
i,j
) +
1
y
_
1 + M
t
2
w
1
i,j

1 M
t
4
w
1
i,j+1

1 + 3M
t
4
w
1
i,j1
_
+
1
y
_
1 M
t
2
w
2
i,j

1 M
t
4
w
2
i,j+1

1 M
t
4
w
2
i,j1
_
+
1
y
_
M
t

2
w
3
i,j
+
1 M
t
2

2
w
2
i,j+1

1 + M
t
2

2
w
3
i,j1
_
= f
1
i,j
w
2
i,j
+
1 + M
n
x
(w
2
i,j
w
2
i1,j
) +
1
y
_
1 + M
t
2
w
2
i,j

1 M
t
4
w
2
i,j+1

1 + 3M
t
4
w
2
i,j1
_
+
1
y
_
1 M
t
2
w
1
i,j

1 M
t
4
w
1
i,j+1

1 M
t
4
w
1
i,j1
_
+
1
y
_
M
t

2
w
3
i,j
+
1 M
t
2

2
w
2
i,j+1

1 + M
t
2

2
w
3
i,j1
_
= f
2
i,j
w
3
i,j
+
M
n
x
(w
3
i,j
w
3
i1,j
) +
1
y
_
w
3
i,j

1 M
t
2
w
3
i,j+1
+
1 + M
t
2
w
3
i,j1
_
+
1
y
_
M
t

2
w
1
i,j
+
1 M
t
2

2
w
1
i,j+1

1 + M
t
2

2
w
1
i,j1
_
+
1
y
_
M
t

2
w
2
i,j
+
1 M
t
2

2
w
2
i,j+1

1 + M
t
2

2
w
2
i,j1
_
= f
3
i,j
w
4
i,j
+
M
n
x
(w
4
i,j
w
4
i1,j
) +
M
t
y
[w
4
i,j
w
4
i,j1
] = f
4
i,j
.
10
As for the equations at x = (i.e. for i = l
2
), we use the last 3 equations of (26) but not the rst
one because of the unknown w
1
l2+1,j
which is not dened in the domain. We will provide the missing
information from the boundary condition:
a
1
w
1
l2,j
+ a
2
w
2
l2,j
+ a
4
w
4
l2,j
= g
j
, j Z.
obtaining a linear system where the number of unknowns and the number of equations are the same.
For the discretized BVP in the domain
2
=], [R
(27)
_
_
_
W
ct
+ A
1
W
x
+ A
2
W
y
= f, for x >
/
pos
W = g for x = ,
we obtain inside the domain the discrete equations (26). If we denote = l
1
x, on the points of the
boundary, that is at x = (i.e. for i = l
1
) we can only keep the rst equation of (26) and add three more
boundary conditions:
_

_
a
2
w
1
l1,j
+
_

a
2
2
a
1
_
w
2
l1,j

a
2
a
4
a
1
w
4
l1,j
= g
j,2
, j Z.

3
w
3
l1,j
= g
j,3
, j Z.
a
4
w
1
l1,j

a
2
a
4
a
1
w
2
l1,j
+
_

a
2
4
a
1
_
w
4
l1,j
= g
j,4
, j Z.
3.2 Optimization of the convergence rate for the discrete Schwarz algorithm
Because of the linearity of the problem studied we can consider directly the algorithm applied to the
homogeneous problem, in term of the error vector. We will further look for a solution under the following
form
(28) W
i,j
=

k
3

l=1

kl
e
(i
1
2
)
l
(k)x
e
Ijky
V
l
(k)
where I
2
= 1. By introducing this expression into the discrete equation (25) we get that for each k,

l
(k) and V
l
(k) have to be the solution of
_
Id +
[A
1
[ + A

1
e

l
(k)x
A
+
1
e

l
(k)x
x
+
[A
2
[ + A

2
e
Iky
A
+
2
e
Iky
y
_
V
l
(k) = 0.
If we denote by L
l
(k) =
e

l
(k)
1
x
and by e
y
(k) =
e
Iky
1
x
and
_

_
h
1
(k) =
1 M
t
4
e
y
(k) +
1 + 3M
t
4
e
y
(k)
e
y
(k)y + 1
h
2
(k) =
1 M
t
4
e
y
(k) +
1 M
t
4
e
y
(k)
e
y
(k)y + 1
h
3
(k) =
1 M
t
2

2
e
y
(k) +
1 + M
t
2

2
e
y
(k)
e
y
(k)y + 1
h
4
(k) =
1 M
t
2
e
y
(k) +
1 + M
t
2
e
y
(k)
e
y
(k)y + 1
11
and by doing the calculation we can see that the discrete eigenvalues L
l
(k) and the corresponding eigen-
vectors V
l
(k) = [V
l,1
(k), V
l,2
(k), V
l,3
(k), 0]
t
, l = 1, 2, 3 satisfy the following:
_

_
V
l,1
(k)[1 (1 M
n
)L
l
(k)] + h
1
(k)V
l,1
+ h
2
(k)V
l,2
(k) + h
3
(k)V
l,3
(k) = 0
V
l,2
(k)
_
1 +
1 + M
n
L
l
(k)x + 1
_
+ h
1
(k)V
l,2
+ h
2
(k)V
l,1
(k) + h
3
(k)V
l,3
(k) = 0
V
l,3
(k)
_
1 +
M
n
L
l
(k)x + 1
_
+ h
4
(k)V
l,3
+ h
3
(k)V
l,1
(k) + h
3
(k)V
l,2
(k) = 0
V
4
(k) = [0, 0, 0, 1]
t
.
We note that lim
x,y0
L
l
(k) =
l
(k) where
l
(k) are the eigenvalues of the matrix /(k) given by
(9) and V
l
(k) can be found up to a multiplicative constant. Therefore L
l
(k) and V
l
(k) can be seen as
the discrete counterpart of the eigenvalues and eigenvectors obtained in the continuous case. Moreover,
L
l
(k) cannot be expressed analytically in a simple form as it is the root of a third order polynomial
P(L
l
(k), x, y) whose expression is not detailed here but whose coecients tend, as x and y tend
to zero, to those of the characteristic polynomial of the matrix /(k). Thus, we can conclude using
continuity arguments, that for a small x and y, P(L
l
(k), x, y) possesses the same number of roots
with positive real part as lim
x,y0
P(L
l
(k), x, y).
The discrete counterpart of the Schwarz algorithm is:
(29)

1
:
_

_
W
p+1
i,j
ct
+
[A
1
[W
p+1
i,j
+ A

1
W
p+1
i+1,j
A
+
1
W
p+1
i1,j
x
+
[A
2
[W
p+1
i,j
+ A

2
W
p+1
i,j+1
A
+
2
W
p+1
i,j1
y
= f, i < l
2
.
/
neg
W
p+1
i,j
= /
neg
W
p
i,j
, i = l
2

2
:
_

_
W
p+1
i,j
ct
+
[A
1
[W
p+1
i,j
+ A

1
W
p+1
i+1,j
A
+
1
W
p+1
i1,j
x
+
[A
2
[W
p+1
i,j
+ A

2
W
p+1
i,j+1
A
+
2
W
p+1
i,j1
y
= f, i > l
1
.
/
pos
W
p+1
i,j
= /
pos
W
p
i,j
, i = l
1
where = l
2
x and = l
1
x.
If we assume that the ow is subsonic that is, if we adopt the same hypotheses as in section 2 then,
in each subdomain, the solution has the form
(30)
W
i,j
=

k1
e
(i
1
2
)1(k)x
e
Ijky
V
1
(k) for i l
2
.
W
i,j
=

k
_

k2
e
(i
1
2
)2(k)x
e
Ijky
V
2
(k)
+
k3
e
(i
1
2
)3(k)x
e
Ijky
V
3
(k) +
k4
e
(i
1
2
)3(k)x
e
Ijky
V
4
(k)
_
for i l
1
.
By introducing these expressions in the interface conditions of (29) we get the discrete convergence
rate
12
Table 1: Overlapping Schwarz algorithm
Numerical vs. theoretical parameters
M
n
b
th
1
b
th
2
b
num
1
b
num
2
0.1 1.6 -0.8 1.6 -0.9
0.2 1.3 -0.5 1.4 -0.6
0.3 1.25 -0.3 1.25 -0.45
0.4 1.08 -0.15 1.08 -0.28
0.5 1.03 -0.08 1.02 -0.23
0.6 1.0 0.0 1.0 0.0
0.7 1.02 0.06 1.01 0.04
0.8 1.03 0.08 1.02 0.06
0.9 1.06 0.08 1.04 0.06
(31)

2
2
(k, x, M
n
, M
t
) =

((v
3
31
v
1
31
)b
2
+ (v
1
31
v
3
21
v
3
31
v
1
21
)b
1
)(b
1
(M
n
1) + b
2
v
2
21
(M
n
+ 1))
((v
3
31
v
2
31
)b
2
+ (v
2
31
v
3
21
v
3
31
v
2
21
)b
1
)(b
1
(M
n
1) + b
2
v
1
21
(M
n
+ 1))
E
2

((v
2
31
v
1
31
)b
2
+ (v
1
31
v
2
21
v
2
31
v
1
21
)b
1
)(b
1
(M
n
1) + b
2
v
3
21
(M
n
+ 1))
((v
3
31
v
2
31
)b
2
+ (v
2
31
v
3
21
v
3
31
v
2
21
)b
1
)(b
1
(M
n
1) + b
2
v
1
21
(M
n
+ 1))
E
3
+
((b
2
1
b
1
)(1 M
n
) b
2
2
(M
n
+ 1))(v
1
31
(v
3
21
v
2
21
) + v
3
31
(v
2
21
v
1
21
) + v
2
31
(v
1
21
v
3
21
))
((v
3
31
v
2
31
)b
2
+ (v
2
31
v
3
21
v
3
31
v
2
21
)b
1
)(b
1
(M
n
1) + b
2
v
1
21
(M
n
+ 1))
E
3

,
where we denoted by v
i
j1
=
V
i,j
(k)
V
i,1
(k)
and E
j
= e
(j(k)1(k))(l2l1)x
, j = 2, 3.
Optimizing the convergence rate with respect to the 2 parameters is already a very dicult task at
the continuous level in the non-overlapping case, we could not carry on such a process and obtaining
analytical results at the discrete level in the overlapping case (which is our case of interest). Therefore,
we will get the theoretical optimized parameters at the discrete level by means of a numerical algorithm,
by calculating the following
(32)
(b
1
, b
2
) = max
kD
h

2
2
(k, x, M
n
, M
t
, b
1
, b
2
)
min
(b1,b2)I
h
(b
1
, b
2
)
where T
h
is a uniform partition of the interval [0, /x] and 1
h
1 a discretization by means of a uniform
grid of a subset of the domain of the admissible values of the parameters. This kind of calculations are
done once for all for a given pair (M
n
, M
t
) before the beginning of the Schwarz iterations. An example of
such a result is given in the gure 1 Mach number M
n
= 0.2. The computed parameters from the relation
(32) will be further refered to with a superscript th. The theoretical estimates are compared afterwards
with the numerical ones obtained by running the Schwarz algorithm with dierent pairs of parameters
which lie in a an interval such that the algorithm is convergent. We are thus able to estimate the optimal
values for b
1
and b
2
from these numerical computations. These values will be referred to by a superscript
num.
4 Implementation and numerical results
We present here a set of results of numerical experiments that are concerned with the evaluation of the
inuence of the interface conditions on the convergence of the non-overlapping Schwarz algorithm of the
13
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2
b2
b
1
Theoretical optimization: Mn=0.2, predicted precision after 20 iterations

5
.
2

5
.
2

5
.
2

4
.
8

4
.
8

4
.
8

4
.
8

4
.
8

4
.
8

4
.
6

4
.
6

4
.
6

4
.
6

4
.
6

4
.
6

4
.
4

4
.
4

4
.
4

4
.
4

4
.
4

4
.
4

4
.
2

4
.
2

4
.
2

4
.
2

4
.
2

4
.
2

4
3
.
8

3
.
8

3
.
8

3
.
8

3
.
8

3
.
8

3
.
6

3
.
6

3
.
6

3
.
6

3
.
6

3
.
6

3
.
4

3
.
4

3
.
4

3
.
4

3
.
4

3
.
4

3
.
2

3
.
2

3
.
2

3
.
2

3
.
2

2
.
8

2
.
8

2
.
8

2
.
8

2
.
8

2
.
6

2
.
6

2
.
4

2
.
4

2
.
2

2
.
2

1
.
8

1
.
8

1
.
6

1
.
6

1
.
4

1
.
4

1
.
2

1
.
2

0
.
8

0
.
8

0
.
6

0
.
6

0
.
4

0
.
4

0
.
2

0
.
2
0
0
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
Figure 1: Isovalues of the predicted reduction factor of the error after 20 iterations via formula (32)
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2
b2
b
1
Numerical optimization:Mn=0.2, precision attained in 20 iterations

5
.
6

5
.
6

5
.
6

5
.
6

5
.
6

5
.
2

5
.
2

5
.
2

5
.
2

5
.
2

4
.
8

4
.
8

4
.
8

4
.8

4
.
8

4
.
8

4
.
4

4
.
4

4
.
4
4.4

4
.
4

4
.
4

3
.
6

3
.
6

3
.
6

3
.
6

3
.6

3
.
6

3
.
6

3
.
2

3
.
2

3
.
2

3
.
2

3
.
2

3
.
2

2
.
8

2
.
8

2
.
8

2
.
8

2
.
8

2
.
4

2
.
4

2
.
4

1
.
6

1
.
6

1
.
2

1
.
2

0
.
8

0
.
8

0
.
4

0
.
4
0
0
6
5
4
3
2
1
0
Figure 2: Isovalues of the reduction factor of the error after 20 iterations for the nite volume code
14
Table 2: Overlapping Schwarz algorithm
Classical vs. optimized counts for dierent values of M
n
M
n
IT
num
0
IT
num
op
M
n
IT
num
0
IT
num
op
0.1 48 19 0.5 22 18
0.2 41 20 0.7 20 16
0.3 32 20 0.8 22 15
0.4 26 19 0.9 18 12
form. The computational domain is given by the rectangle [0 , 1] [0 , 1]. The numerical investigation
is limited to the resolution of the linear system resulting from the rst implicit time step using a Courant
number CFL=100.
In all these calculations we considered a model problem: a ow normal to the interface (that is when
M
t
= 0). In gures 1 and 2 we can see an example of a theoretical and numerical estimation of the
reduction factor of the error. We illustrate here the level curves which represent the log of the precision
after 20 iterations for dierent values of the parameters (b
1
, b
2
), the minimum being attained in this case
for b
th
1
= 1.3 and b
th
2
= 0.5, b
num
1
= 1.4 and b
num
2
= 0.6. We can see that we have good theoretical esti-
mates of these parameters we can therefore use them in the interface conditions of the Schwarz algorithm.
Table 2 summarizes the number of Schwarz iterations required to reduce the initial linear residual by
a factor 10
6
for dierent values of the reference Mach number with the optimal parameters b
num
1
and
b
num
1
. Here we denoted by IT
num
0
and IT
num
op
the observed (numerical) iteration number for classical and
optimized interface conditions in order to achieve a convergence with a threshlod = 10
6
. The same
results are presented in gure 4. In the gure 3 we compare the theoretical estimated iteration number
in the classical and optimized case. Comparing gures 3 and 4 we can see that the theoretical prediction
are very close to the numerical tests.
The conclusion of these numerical tests is, on one hand, that the theoretical prediction is very close
to the numerical results: we can get by a numerical optimization (32) a very good estimate of optimal
parameters (b
1
, b
2
)). On the other hand, the gain, in number of iterations, provided by the optimized
interface conditions, is very promising for low Mach numbers, where the classical algorithm doesnt give
optimal results. We can note that the optimized convergence rate is monotone with respect to the normal
Mach number while the classical one isnt. For bigger Mach numbers, for instance, those who are close
to 1, the classical algorithm already has a very good behaviour so the optimization is less useful. In the
same time we studied here the zero order and therefore very simple transmission conditions. The use of
higher order conditions (see [GMN02]) is a possible way that can be further studied to obtain even better
convergence results.
5 Appendix
In the following we place ourselves in the cas M
t
= 0, therefore the convergence rate has the expression:
(33) (R, b
1
, b
2
) = 1
4Rb
1
(1 + MR)
(R(b
1
+ b
2
) + b
1
b
2
)
2
(1 + M)
where R =
_
1 +
2
(1 M
2
) is a real quantity depending only on the wave number and the Mach
number M = M
n
. We have the following optimization result:
15
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
10
20
30
40
50
60
Mn
I
t
e
r
Theoretical iteration number, eps=1.0e6
Classical Ovr.
Optimized Ovr.
Figure 3: Theoretical iteration number: classical vs. optimized conditions
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
5
10
15
20
25
30
35
40
45
50
Mn
I
t
e
r
Numerical iteration number, eps=1.0e6
Classical Ovr.
Optimized Ovr.
Figure 4: Numerical iteration number: classical vs. optimized conditions
16
Theorem 3 The optimization problem
min
(b1,b2)I1I2
max
R1
[(R, b
1
, b
2
)[
possesses an analytical solution if M
_
0.125,
5 + 3

17
16
_
which is given by
_

_
b
1
= 1 +
(

M + 1

M)
2
8
_
M(M + 1)
b
2
=
_
M(M + 1) (1 M)
(

M + 1 +

M)

M + 1
b
1
Before proceeding to the proof we will rst formulate some properties of the convergence rate:
Lemma 2 We will denote by R
s
the local extremum (solution of the equation

R
(R, b
1
, b
2
) = 0) when
it exists. If the convergence rate has two distinct roots (there exists R
1,2
> 1 such that (R
1
, b
1
, b
2
) =
(R
2
, b
1
, b
2
) = 0) then the solution of the optimization problem satises the relations:
(34) max(1, b
1
, b
2
), (, b
1
, b
2
) = (R
s
, b
1
, b
2
)
and this equation gives an admissible solution ((b
1
, b
2
) D) only for M [0.125, M
0
] [0.125, 0.55].
Moreover the optimum veries (1, b
1
, b
2
) = (, b
1
, b
2
) for M [0.125,
5 + 3

17
16
].
Proof Before proceeding to the analysis we notice that for a xed value R > 1 the convergence rate
grows monotonically in the parameters b
1,2
:
(35)

b
1
=
4(MR + 1)R(Rb
1
Rb
2
+ b
1
+ b
2
)
((Rb
1
+ Rb
2
+ b
1
b
2
)
3
(1 + M))
> 0

b
2
=
8(MR + 1)Rb
1
(R 1)
((Rb
1
+ Rb
2
+ b
1
b
2
)
3
(1 + M)
> 0
When has two roots R
1,2
> 1 then the parameters satisfy the following
(36) b
2
1

2
(b
1
) =
_
_
4Mb
1
M + 1
b
1
,
Mb
1

_
b
1
(b
1
1)
M + 1
_
Moreover it admits one extremum R
s
which is the solution of its derivative:
(37) R
s
=
b
1
b
2
b
1
+ b
2
2M(b
1
b
2
)
By imposing that R
s
be bigger than 1 we get an additional condition on b
2
:
b
2

2M + 1
M + 1
b
1
which gives together with b
2
1

2
(b
1
) a new restriction on b
1
:
(38) b
1
1
1
=
_
1,
(2M + 1)
2
4M(M + 1)
_
17
The convergence rate for this value (R
s
) will have an opposite sign to
(1, b
1
, b
2
) =
b
1
1
b
1
> 0
and its modulus decreases monotonically in the parameters b
1,2
:
(39)
[(R
s
, b
1
, b
2
)[
b
1
=
(M 1)b
2
1
(M + 1)b
2
2
(b
1
b
2
)
2
(Mb
1
Mb
2
b
1
b
2
)
2
(1 + M)
< 0
[(R
s
, b
1
, b
2
)[
b
2
=
2b
1
(Mb
1
(M + 1)b
2
(b
1
b
2
)
2
(Mb
1
Mb
2
b
1
b
2
)
2
(1 + M)
< 0
From the previous remarks we deduce that the positive quantities (1, b
1
, b
2
) and (, b
1
, b
2
) are
increasing with respect to the parameters b
1
and b
2
and (R
s
, b
1
, b
2
) is decreasing. Therefore is
minimized if the positive maximum of and the negative maximum are equal in modulus, that is:
(40) max(1, b
1
, b
2
), (, b
1
, b
2
) = (R
s
, b
1
, b
2
)
The fact that (1, b
1
, b
2
) = (, b
1
, b
2
) comes out by supposing the contrary, that is for example:
(1, b
1
, b
2
) > (, b
1
, b
2
) > 0
The arbitrary small changes b
1
and b
2
of the parameters lead to arbitrary small change in (, b
1
, b
2
)
which is unimportant and this inequality will be preserved. The changes in (1, b
1
, b
2
) and (R
s
, b
1
, b
2
)
are given by:
(41)
(1, b
1
, b
2
) = b
1

b
1
(1, b
1
, b
2
) + b
2

b
2
(1, b
1
, b
2
)
(R
s
, b
1
, b
2
) = R
s

R
(R
s
, b
1
, b
2
) + b
1

b
1
(R
s
, b
1
, b
2
) + b
2

b
2
(R
s
, b
1
, b
2
)
= b
1

b
1
(R
s
, b
1
, b
2
) + b
2

b
2
(R
s
, b
1
, b
2
)
and the extremum will be now located in R
s
+R
s
. From the relations (35) and (41) we see that we can
decrease b
1
and increase b
2
such that (R
s
, b
1
, b
2
) increases while (1, b
1
, b
2
) decreases which contradicts
the fact that we had an optimum. Therefore the optimum is attained for those values of the parameters
where the equalities (34) hold. In the case when, for a given Mach number this cannot be satised (the
solution to these equations gives admissible values to the parameters only for certain values of the Mach
numbers as we will see later) we still have the weaker condition given by (40).
In the following we will determine which are the values of mach number for which one of the relations
(34) or (40) hold. In order to have an admissible solution we have to check rst that b
2
1
2
(b
1
) where
b
1
and b
2
are the solutions of (34):
(42)
_

_
b
1
= 1 +
(

M + 1

M)
2
8
_
M(M + 1)
b
2
=
_
M(M + 1) (1 M)
(

M + 1 +

M)

M + 1
b
1
By solving the inequality which characterizes this inclusion we get that it is possible only for the values
of the Mach number lying in the interval
_
1
8
,
5 + 3

17
16
_
= [0.125, 0.46]. When the Mach number doesnt
18
lie in this interval we still have the equation (40) veried. We can distinguish two possible cases:
Case 1a. (1, b
1
, b
2
) = (R
s
, b
1
, b
2
) > lim
R
(R, b
1
, b
2
).
In this case by solving the rst equation we get the value of the parameter b
2
in function of b
1
:
(43) b
2
(b
1
, M) =
(4Mb
1
2M 2
_
4b
2
1
6b
1
+ 2)b
1
2(2b
1
1)(1 + M)
and the inequation gives
b
1
> 1 +
(

M + 1

M)
2
8
_
M(M + 1)
Afterwards we have to check that for a given Mach number b
2
(b
1
, M) 1
2
(b
1
) 1

2
(b
1
). We nd
numerically that this is true only when M [0.125, 0.55] but we cannot get an analytical solution.
Case 1b. (R
s
, b
1
, b
2
) = lim
R
(R, b
1
, b
2
) > (1, b
1
, b
2
).
After some tedious calculations we get that the solution of the min-max problem is found in the
previous case.
Now we will proceed to the proof of the theorem:
Proof If M ]0.125, 0.55[, according to the results given by the lemma, the solution is given by a
pair (b
1
, b
2
) where b
1
lies in an interval given by
b
1
[b
2
(b
1
, M) 1
2
(b
1
) 1

2
(b
1
)
Moreover the value of the convergence rate is increasing with respect to b
1
:
(1, b
1
, b
2
)
b
1
=
1
b
2
1
> 0
so the value of b
1
which minimizes the convergence rate will be given by:
(44)
mininf b
1
[b
2
(b
1
, M) 1
2
(b
1
) 1

2
(b
1
) , inf b
1
[b
2
(b
1
, M) 1
2
(b
1
) 1

2
(b
1
) =
= inf b
1
[b
2
(b
1
, M) 1
2
(b
1
) 1

2
(b
1
)
Moreover if M [0.125,
5 + 3

17
16
[ the inmum of the relation (44) is given by the formula (42).
References
[AG04] Ana Alonso-Rodr`guez and Luca Gerardo-Giorda. New nonoverlapping domain decomposi-
tion methods for the harmonic maxwell system. Technical Report 529, CMAP - Ecole Poly-
technique, 2004.
[Bj95] M. Bjrhus. Semi-discrete subdomain iteration for hyperbolic systems. Technical Report 4,
NTNU, 1995.
[CDJP97] P. Collino, G. Delbue, P. Joly, and A. Piacentini. A new interface condition in the non-
overlapping domain decomposition for the Maxwell equations. Comput. Methods Appl. Mech.
Engrg., 148:195207, 1997.
[CFS98] X.-C. Cai, C. Farhat, and M. Sarkis. A minimum overlap restricted additive Schwarz precon-
ditioner and appication in 3D ow simulations. In C. Farhat J. Mandel and X.-C. Cai, editors,
Proceedings of the 10th Domain Decomposition Methods in Sciences and Engineering, volume
218 of Contemporary Mathematics, pages 479485. AMS, 1998.
19
[Cle98] S. Clerc. Non-overlapping Schwarz method for systems of rst order equations. Cont. Math,
218:408416, 1998.
[DJR92] Bruno Despres, Patrick Joly, and Jean E. Roberts. A domain decomposition method for the
harmonic Maxwell equations. In Iterative methods in linear algebra (Brussels, 1991), pages
475484, Amsterdam, 1992. North-Holland.
[DLN02] V. Dolean, S. Lanteri, and F. Nataf. Construction of interface conditions for solving compress-
ible euler equations by non-overlapping domain decomposition methods. Int. J. Numer. Meth.
Fluids, 40:14851492, 2002.
[DLN04] V. Dolean, S. Lanteri, and F. Nataf. Convergence analysis of a schwarz type domain de-
composition method for the solution of the euler equations. Appl. Num. Math., 49:153186,
2004.
[GMN02] M.-J. Gander, F. Magoul`es, and F. Nataf. Optimized Schwarz methods without overlap for
the Helmholtz equation. SIAM J. Sci. Comput., 24-1:3860, 2002.
[JN00] C. Japhet and F. Nataf. The best interface conditions for domain decomposition methods:
absorbing boundary conditions. In L. Tourrette, editor, Articial Boundary Conditions, with
Applications to Computational Fluid Dynamics Problems, pages 348373. Nova Science, 2000.
[JNR01] C. Japhet, F. Nataf, and F. Rogier. The Optimized Order 2 method. application to convection-
diusion problems. Future Generation Computer Systems, 18:1730, 2001.
[QS96] A. Quarteroni and L. Stolcis. Homogeneous and heterogeneous domain decomposition methods
for compressible ow at high reynolds numbers. Technical Report 33, CRS4, 1996.
[Qua90] A. Quarteroni. Domain decomposition methods for systems of conservation laws : spectral
collocation approximation. SIAM J. Sci. Stat. Comput., 11:10291052, 1990.
[Roe81] P.L. Roe. Approximate Riemann solvers, parameter vectors and dierence schemes. J. Comput.
Phys., 43:357372, 1981.
20

You might also like