MATHEMATICS OF COMPUTATION

Volume 73, Number 246, Pages 659689

S 0025-5718(03)01550-3
Article electronically published on October 17, 2003
SUBSTRUCTURING PRECONDITIONERS
FOR THE THREE FIELDS
DOMAIN DECOMPOSITION METHOD
SILVIA BERTOLUZZA
Abstract. We study a class of preconditioners based on substructuring, for
the discrete Steklov-Poincare operator arising in the three elds formulation of
domain decomposition in two dimensions. Under extremely general assump-
tions on the discretization spaces involved, an upper bound is provided on the
condition number of the preconditioned system, which is shown to grow at most
as log(H/h)
2
(H and h denoting, respectively, the diameter and the discretiza-
tion mesh-size of the subdomains). Extensive numerical testsperformed on
both a plain and a stabilized version of the methodconrm the optimality
of such bound.
1. Introduction
The use of nonconforming domain decomposition methods is becoming increas-
ingly popular. This is due to several factors: the possibility of easily coupling dif-
ferent discretizations in dierent subdomains without the need for imposing strong
matching conditions allows us to employ without any adaptation the technologies
developed for treating the problem with a single-domain approach. As an example,
in an adaptive strategy, renement can be carried out in each subdomain inde-
pendently. Moreover, it is possible to use a dierent type of discretization (nite
elements, spectral methods, wavelets) in the dierent subdomains, and therefore
to employ, in each subdomain, the best-suited method (spectral methods where
the solution is expected to be very smooth, nite elements where a complicated
geometry requires it, wavelets where isolated singularities in a regular background
are expected).
Among the dierent nonconforming domain decomposition method, we consider
here the three elds formulation, as proposed by F. Brezzi and L.D. Marini in [14].
Dierently from other nonconforming formulations [3, 18] of domain decomposition,
weak continuity is not imposed by requiring that the jump across the interface be
orthogonal to some multiplier space, but by introducing a space
h
, approximating
the traces of functions in H
1
() on the interface . The elements of
h
are to be
used as a Dirichlet boundary condition imposed by means of Lagrange multipli-
ers for the functions in the subdomains. As with other nonconforming methods,
great freedom is left in the choice of the discretization in the subdomains which,
Received by the editor November 6, 2000 and, in revised form, February 22, 2002.
2000 Mathematics Subject Classication. Primary 65N55, 65N22.
Key words and phrases. Three elds formulation, domain decomposition, stabilized methods,
preconditioning.
c 2003 American Mathematical Society
659
660 SILVIA BERTOLUZZA
after applying a suitable stabilization method if needed [2, 7], can be chosen inde-
pendently from each other. The disadvantage of introducing a third unknown eld
is compensated for by the fact that all subdomains are treated exactly in the same
way, which results in an easier implementation and possibly, when considering the
parallelization of the method, in an easier balancing of the load between processors.
We deal, in this paper, with the problem of the ecient solution of the linear
system arising when applying such a method. By applying a Schur complement
technique, the solution of the resulting discrete problem can be reduced to the
solution of an equation on the trace unknown, involving a Steklov-Poincare type
operator which can be shown to be a pseudo-dierential operator of order one.
The resulting linear system will then be solved by an iterative procedure, at each
step of which there will be the need of computing the action of such operator on
a given function. This reduces to solving independent Dirichlet problems on the
subdomains, a task which can be carried out in parallel, subdomain by subdomain
in a completely independent way.
In order to have an ecient scheme for solving the equation for the interface
unknown, we will then need two things:
ecient solvers for the local Dirichlet problems on the subdomains;
a preconditioner for the discrete Steklov-Poincare operator so that the num-
ber of iterations is kept as low as possible.
We will concentrate here on the second issue, by analyzing and testing a class
of preconditioners for the discrete Steklov-Poincare operator. The approach that
we will follow is the substructuring one, proposed by J.H. Bramble, J.E. Pasciak
and A.H. Schatz [9] in the framework of conforming domain decomposition and
already applied in the framework of nonconforming methods to the mortar method
in [1]. This consists in decomposing the trace space
h
in a direct sum of a coarse
subspace (essentially linked to the geometrical decomposition of in subdomains)
and of suitable local subspaces and then considering the related block-Jacobi type
preconditioners.
From the theoretical point of view, we study the dependence of the condition
number of the resulting preconditioned matrix on the mesh size of the discretiza-
tions, aiming at giving bounds which, under suitable assumptions, are independent
of the number and size of the subdomains. In particular we provide an upper bound
which applies to a wide variety of approximation spaces, including nite elements
and wavelets. Though only suboptimal in terms of the mesh size on the skele-
ton (the condition number of the preconditioned matrix grows like (log(H/h))
2
,
H and h being, respectively, the diameter of the subdomains and the mesh size
of the corresponding discretization), the estimate obtained is uniform with respect
to the number and size of subdomains. The resulting method is then potentially
almost perfectly scalable. Moreover this approach has the advantage over other
asymptotically better preconditioners for the three elds formulation ([7]) of be-
ing itself naturally computable in parallel (each edge in the skeleton being treated
independently of the others).
We conclude the paper by reporting a series of extensive numerical tests, per-
formed on both the plain and on a stabilized formulation as proposed in [7]. The
results of the tests are in agreement with the theory and conrm the polylogarithmic
growth (with respect to the mesh size of the discretization) of the preconditioned
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 661
system, as well as the independence (asymptotically) with respect to the number
and size of the subdomains.
2. Domain decomposition method: the three fields formulation
Let R
2
be a convex polygonal domain. We will consider the following simple
model problem: given f L
2
(), nd u satisfying
(1)
2

i,j=1

x
j
_
a
i,j
(x)
u
x
i
_
+a
0
(x)u = f in , u = 0 on .
We assume that for almost all x we have 0 a
0
(x) R and that the matrix
(a
ij
(x))
i,j=1,2
is symmetric positive denite, with smallest eigenvalue > 0 and
largest eigenvalue

, ,

independent of x.
In a domain decomposition framework, we will consider the three elds formula-
tion ([14]) of such a problem. Let =

K
k=1

k
,
k
polygons, which, for simplicity,
we will assume to be quadrangles. We set
k
=
k
, =

k

k
being the skele-
ton of the decomposition. Remark that dierently from [14] here we dene so
that it includes the external boundary of the domain , so that there is no need
of distinguishing between interior subdomains and subdomains which are adjacent
to the boundary . This results in a simplication both on the notational and
on the implementation point of view, since all subdomains are treated in the same
way. We will make the following regularity assumptions on the subdomains
k
:
(A1) The subdomains are regular in shape and the geometrical decomposition is
graded, that is,
(a) there exists a positive constant c
0
such that, for all k,
k
contains a
ball of diameter c
0
H
k
, it is contained in a ball of diameter H
k
, and
the length of each side is bounded from below by c
0
H
k
; moreover
any interior angle satises 0 < c
1
< < c
2
< (c
0
, c
1
, and c
2
independent of k);
(b) there exists a positive constant c
3
such that, if , k are such that
[
k

[ > 0, then it holds that

H
k
/H

c
3
.
For technical reasons, we also need to make the following assumption (needed
for the proof of (25) in the following), which in many cases can be shown to be
actually a consequence of the previous one:
(A2) There exists a constant C
0
such that for all y, denoting by
y
the line of
equation x
2
= y and setting K
y
= k : length(
y

k
) > 0, it holds that

kKy
H
k
C
0
.
We are interested here in explicitly studying the dependence of the estimates
that we are going to prove on the number and size of the subdomains. To this end,
in the following we will employ the notation A B (resp. A B) to say that
the quantity A is bounded from above (resp. from below) by cB, with a constant
c independent of k and of the H
k
s, as well as of any mesh size parameter. The
expression A B will stand for A B A.
For each k we let the norm and seminorm of H
1
(
k
) be dened as usual by
|u|
2
H
1
(
k
)
=
_

k
[u[
2
dx +
_

k
[u[
2
dx and [u[
2
H
1
(
k
)
=
_

k
[u[
2
dx. On the
662 SILVIA BERTOLUZZA
boundary
k
we let the norm of H
1/2
(
k
) be dened by
||
H
1/2
(
k
)
= inf
uH
1
(
k
)
u= on
k
|u|
H
1
(
k
)
.
For all s in ]0, 1[, we dene the H
s
(
k
) seminorm by
(2) [[
2
H
s
(
k
)
=
_

k
_

k
((x) (y))
2
[x y[
2s+1
ds(x) ds(y).
The space H
1/2
(
k
) is dened as the dual of H
1/2
(
k
).
For e
k
segment of extrema a and b, we dene the H
s
(e) seminorm as usual
by the integral expression (2) where the integrals are taken over e. On e we will
also consider the spaces H
s
0
(e) (s ,= 1/2) and H
1/2
00
(e) of functions whose extension
by zero is in H
s
(
k
) (s ,= 1/2) and H
1/2
(
k
), respectively, which we will equip
with the norm
||
2
H
1/2
00
(e)
= [[
2
H
1/2
(e)
+
_
e
[(x)[
2
[x a[
2s
ds(x) +
_
e
[(x)[
2
[x b[
2s
ds(x).
We recall that for s < 1/2 the two spaces H
s
(e) and H
s
0
(e) coincide, and the
two corresponding norms are equivalent. However, the constant in the equivalence
depends a priori on H
k
and on s. In particular it explodes as s converges to 1/2.
The behaviour of such a constant as s approaches the limit value 1/2 (see (44)) will
play a key role in the forthcoming analysis.
Remark also that, for H
1/2
00
(e), letting H
1/2
(
k
) denote the function
coinciding with on e and identically vanishing on
k
e, it holds that
(3) ||
H
1/2
00
(e)
[[
H
1/2
(
k
)
,
the constant in the equivalence being uniformly bounded, provided that the ratio
between the length of
k
and the length of e is uniformly bounded.
The functional setting for the three elds domain decomposition method is given
by the following spaces:
(4) V =
K

k=1
H
1
(
k
), =
K

k=1
H
1/2
(
k
),
and
(5) = L
2
() : there exists u H
1
0
(), u = on = H
1
0
()[

,
respectively equipped with the norms:
(6) |u|
V
=
_

k
|u
k
|
2
H
1
(
k
)
_
1/2
, ||

=
_

k
|
k
|
2
H
1/2
(
k
)
_
1/2
,
and (see [4])
(7) ||

= inf
uH
1
0
()
u= on
|u|
H
1
()

_

k
[[
2
H
1/2
(
k
)
_
1/2
.
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 663
Let a
k
: H
1
(
k
) H
1
(
k
) R denote the bilinear form corresponding to the
linear operator considered:
a
k
(w, v) =
_

k
_
_
2

i,j=1
a
ij
(x)
w
x
i
v
x
j
+a
0
(x)wv
_
_
dx.
Under the assumptions made on the matrix (a
ij
(x))
i,j=1,2
and on a
0
, the bilinear
forms a
k
are uniformly H
1
(
k
)-continuous, and semidenite. More precisely, there
exist positive constants and L independent of k and H
k
such that for all w, v in
H
1
(
k
)
[w[
2
H
1
(
k
)
a
k
(w, w), [a
k
(w, v)[ L|w|
H
1
(
k
)
|v|
H
1
(
k
)
.
The three elds formulation of equation (1) is the following: nd (u, , ) V
such that
(8)
_

_
k, v
k
H
1
(
k
),
k
H
1/2(
k
)
:
a
k
(u
k
, v
k
)
_

k
v
k

k
ds =
_

k
fv
k
dx,
_

k
u
k

k
ds
_

k
ds = 0,
and :

k
_

k
ds = 0.
It is known that this problem admits a unique solution (u, , ) where u is indeed
the solution of (1) and such that
(9)
k
=
u
k

k
a
on
k
, = u on ,
where
k
a
denotes the outer conormal derivative to the subdomain
k
. Remark that
in such a formulation the homogeneous Dirichlet boundary condition is embedded
in the denition of the space .
Equation (8) can be discretized by a Galerkin scheme: after choosing discretiza-
tion spaces V
h
=

K
k=1
V
k
h


K
k=1
H
1
(
k
),
h
=

K
k=1

k
h


K
k=1
H
1/2
(
k
)
and
h
, we consider the problem: nd (u
h
,
h
,
h
) V
h

h
, such that
664 SILVIA BERTOLUZZA
one has
(10)
_

_
k, v
k
h
V
k
h
,
k
h

k
h
:
a
k
(u
k
h
, v
k
h
)
_

k
v
k
h

k
h
ds =
_

k
fv
k
h
dx,
_

k
u
k
h

k
h
ds
_

k
h

h
ds = 0,
and
h

h
:

k
_

k
h

h
ds = 0.
Existence, uniqueness and stability of the solution of (10) rely on the validity of
suitable inf-sup conditions. More precisely we can make the following assumption:
(A3) The spaces V
k
h
,
k
h
and
h
are such that
(a)
k
h
contains the constant functions;
(b) the following compatibility condition between V
k
h
and
k
h
holds uni-
formly in k:
(11) inf

k
h

k
h
sup
u
k
h
V
k
h
_

k
h
u
k
h
ds
|u
k
h
|
H
1
(
k
)
|
k
h
|
H
1/2
(
k
)
> 0;
(c) the following compatibility condition between
h
and
h
holds:
(12) inf

h
sup

k
_

k
h

h
ds
|
h
|

|
h
|

> 0.
Remark that (A3)(a) implies that for all v
k
h
ker B
k
h
= v
k
h
V
k
h
:
_

k
v
k
h

k
h
=
0
k
h

k
h
it holds that
_

k
v
k
h
ds = 0 and then it is not dicult to see that a
is elliptic on ker B
k
h
. Then, by a well-known argument ([11]), there exists a unique
solution to problem (10), which converges with an optimal rate to the solution of
(8).
The linear system stemming from (10) takes the form
(13)
_
_
A B
T
0
B 0 C
T
0 C 0
_
_

_
_
u
h

h
_
_
=
_
_
f
h
0
0
_
_
,
(u
h
,
h
, and
h
being the vectors of the coecients of u
h
,
h
and
h
in the bases
chosen for V
h
,
h
and
h
, respectively). By a Schur complement argument the
solution of (13) can be reduced to a system in the unknown
h
of the form
(14) CA
1
C
T

h
= CA
1
_
f
h
0
_
, C = [ 0 C ], A =
_
A B
T
B 0
_
.
The matrix S = CA
1
C
T
does not need to be assembled. The system (14) can
rather be solved by an iterative technique (like for instance a conjugate gradient
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 665
method) and therefore only the action of S on a given vector needs to be imple-
mented. Multiplying by S implies the need for solving a linear system with matrix
A. This reduces, by a proper reordering of the unknowns, to independently solv-
ing K discrete Dirichlet problems with Lagrange multipliers in the K subdomains.
Moreover under the above assumptions it is possible to prove that the bilinear form
s :
h

h
R,
(15) s(
h
,
h
) =
T
h
S
h
, S = CA
1
C
T
,
corresponding to the Schur complement matrix S, satises the following continuity
and coercivity assumptions:
Lemma 2.1. If (A3) holds, then the bilinear form s is continuous and coercive
with respect to the norm:
s(
h
,
h
) |
h
|

|
h
|

, (16)
s(
h
,
h
) |
h
|
2

. (17)
Assumption (A3) is, in practice, quite restrictive. In particular the requirement
that both (A3)(b) and (A3)(c) hold simultaneously poses some serious limitation
on the choice of the discretization spaces for the three unknowns u, and . In
order to be more free in such a choice, and more specically in order to be allowed
to choose the discretization space for the interface unknown independently of
the choice of the other two discretization spaces, it is also possible to consider a
stabilized version. Dierent ways have been proposed for stabilizing the three elds
formulation ([2, 12, 13]). We consider here the approach of [7, 4]. For all k let
a bilinear form [, ] 1
2
,k
: H
1/2
(
k
) H
1/2
(
k
) R be given, satisfying the
following bounds: for all , H
1/2
(
k
)
(18) [, ] 1
2
,k
0, [, ] 1
2
,k
B[[
1/2,
k
[[
1/2,
k
,
for B positive constant independent of k and H
k
.
We consider then the following discrete stabilized formulation: nd u
h
V
h
,

h

h
,
h

h
such that it holds that
(19)
_

_
k, v
k
h
V
k
h
,
k
h

k
h
:
a
k
(u
k
h
, v
k
h
) +[u
k
h
, v
k
h
] 1
2
,k

_

k
v
k
h

k
h
ds [
h
, v
k
h
] 1
2
,k
=
_

k
fv
k
h
dx,
_

k
u
k
h

k
h
ds
_

k
h

h
ds = 0,
and
h

h
:

k
[u
k
h
,
h
] 1
2
,k
+

k
_

k
h

h
ds +

k
[
h
,
h
] 1
2
,k
= 0,
where > 0 is a parameter independent of the choice of the discretization spaces.
Such formulation is consistent with the original continuous problem, i.e., by sub-
stituting the solution (u, , ) of (8) at the place of (u
h
,
h
,
h
) in (19), we obtain
an identity.
The assumptions needed to guarantee existence and uniqueness of the solution
of the discrete problem, as well as an error estimate, can now be made on the one
666 SILVIA BERTOLUZZA
hand on the spaces V
h
and
h
and on the other hand in a completely independent
way, on the action of the bilinear forms [, ] 1
2
,k
on
h
[

k
. For example we can
replace assumption (A3)(c) with
(A3)(c

) There exists a positive constant b such that for all

h

h
it holds for
all k that
[
h
,
h
] 1
2
,k
b[
h
[
2
H
1/2
(
k
)
.
It is then possible to show that the discrete problem (19) admits a unique solution
satisfying an optimal error estimate ([4]). It is beyond the goals of this paper to
thoroughly discuss the construction of suitable bilinear forms [, ] 1
2
,k
. We recall that
this can be done with the aid of multiscale techniques and we refer to [7, 5, 10, 6]
for a detailed analysis of possible solutions.
The linear system stemming from such a problem takes the following form this
time:
(20)
_
_

A B
T
D
T
B 0 C
T
D C E
_
_

_
_
u
h

h
_
_
=
_
_
f
h
0
0
_
_
,
with

A = A+F (the matrices D, E and F deriving from the stabilizing terms).
Again, the solution of (20) can be reduced to a system in the unknown
h
, this
time taking the form
(21) S
h
:=
_
D

A
1
D
T
+E
_

h
= DA
1
_
f
h
0
_
with
(22)

A =
_

A B
T
B 0
_
, D = [ D C ].
Once again we let s :
h

h
R be the bilinear form corresponding to the
Schur complement matrix S
(23) s(
h
,
h
) =
T
h
S
h
,
and also for the stabilized formulation the following holds ([4]):
Lemma 2.2. Under assumptions (A3)(a), (b), (c

) the bilinear form s is continuous

and coercive with respect to the norm:
s(
h
,
h
) |
h
|

|
h
|

, (24)
s(
h
,
h
) |
h
|
2

. (25)
3. Preconditioning by substructuring
For both original and stabilized three elds formulation, we end up with an
equation on the interface space
h
of the form
S
h
= f
h
corresponding, through a relation of the form (23), to a bilinear form s :
h

h

R verifying for all
h
,
h

h
that
(26) [s(
h
,
h
)[ |
h
|

|
h
|

, s(
h
,
h
) |
h
|
2

.
Our aim is to provide a preconditioner for the matrix S. The approach that we will
follow is the one proposed, in the framework of conforming domain decomposition,
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 667
by J.H. Bramble, J.E. Pasciak and A.H. Schatz in [9] and already studied in [1] for
the mortar method case.
Let us consider a splitting of the interior part of the skeleton as the disjoint
union of M segments e
i
of extrema a
i
and b
i
=
M
_
i=1
e
i
,
(corresponding, for instance, to the splitting of as the union of segments of the
form
k

). For each k let E

k
individuate those segments that, together with

k
, build up the boundary of the subdomain
k
:

k
=
_
iE
k
e
i
, with E
k
= i : [e
i

k
[ > 0.
We make the following assumption:
(A4) There exists a constant c
4
such that it holds that
sup
k
sup
iE
k
[
k
[
[e
i
[
c
4
.
Note that (A4) implies that #(E
k
) c
4
for all k. This will allow us several times
to write inequalities of the form |

iE
k
q
i
|
2

iE
k
|q
i
|
2
, where | | stands for
various norms and seminorms (the constant in the inequality depending on c
4
).
Following [9], the main idea for constructing a preconditioner for the matrix S
is to decompose
h
as the direct sum of a coarse space L
H
and some local spaces

0
h,i
(one for each edge e
i
of the decomposition of ) and then considering the
corresponding block-Jacobi type preconditioners. More precisely, let L
H
,
L
H
= C
0
() : i = 1, . . . , M, [
ei
P
1
(e
i
), = 0 on ,
denote the subset of those functions of whose restriction to each segment e
i
is
linear. In order to construct the preconditioner we make the following assumptions
on the approximation space
h
:
(A5) The following two conditions hold:
(a) L
H

h
;
(b) for all
h

h
such that at the extrema a
i
and b
i
of some edge e
i
it
holds that
h
(a
i
) =
h
(b
i
) = 0, the function
0
h,i
dened as

0
h,i
=
h
on e
i
,
0
h,i
= 0 on e
i
veries

0
h,i

h
.
(A6)
h
H
1
() and for all
h

h
the following two inverse inequalities hold
for all k, k = 1, . . . , K, and for all r, t, 0 r < t 1:
[
h
[
H
t
(
k
)
h
rt
k
[
h
[
H
r
(
k
)
, (27)
[
h
[
H
t
(ei)
h
rt
k
[
h
[
H
r
(ei)
, i E
k
. (28)
Remark 3.1. Assumption (A5) is not at all restrictive. In a nite element framework
it reduces to asking that the extrema a
i
and b
i
of the segments e
i
into which the
skeleton is split be nodes of the nite element grid for
h
.
Remark 3.2. The parameter h
k
appearing in the inverse inequality (27) has here
the meaning of the smallest mesh size of the restriction to
k
of
h
.
668 SILVIA BERTOLUZZA
Under assumption (A5) it is easy to see that
h
can be split as

h
= L
H

0
h
,
h
=
H
+
0
h
,
with
0
h
dened as

0
h
=
h

h
,
h
= 0 at the extrema a
i
and b
i
of each edge e
i
of .
The subspace
0
h
veries

0
h
=
M

i=1

h

0
h
:
h
= 0 on e
i

M

i=1

0
h
[
ei
,
where the symbol means that the two spaces are isomorphic. Note that all
h

0
h
verify
h
[
ei
H
1/2
00
(e
i
) for all i, and in the following we will write |
h
|
H
1/2
00
(ei)
for |
h
[
ei
|
H
1/2
00
(ei)
(and analogously for other norms dened on portions of ).
For each i = 1, , M, we assume we are given a bilinear form s
i
:
0
h
[
ei

0
h
[
ei
R, satisfying the following continuity and positivity bounds uniformly in
h: for all ,
0
h
[
ei
:
(B1) s
i
(, ) ||
H
1/2
00
(ei)
||
H
1/2
00
(ei)
, s
i
(, ) ||
2
H
1/2
00
(ei)
.
Moreover we assume that we are given a bilinear form s
H
: L
H
L
H
R such
that for all
H
,
H
L
H
it holds that
(B2) s
H
(
H
,
H
) |
H
|

|
H
|

, s
H
(
H
,
H
) |
H
|
2

.
It is beyond the goals of this paper to fully discuss the dierent ways in which
the bilinear forms s
i
and s
H
can be constructed. In the following section we will
briey recall a possible solution and we refer to [9, 10, 16] for further discussions
on this topic.
Following [9], the preconditioner will be nally constructed with the aid of a
bilinear form s :
h

h
R assembled by blocks: for
h
,
h

h
, with
h
=

H
+
0
h
,
h
=
H
+
0
h
,
H
,
H
L
H
and
0
h
,
0
h

0
h
we set
(29) s(
h
,
h
) = s
H
(
H
,
H
) +
M

i=1
s
i
(
0
h
[
ei
,
0
h
[
ei
).
The main result of this paper is the following theorem.
Theorem 3.1. For all
h

h
it holds that
(30) s(
h
,
h
) s(
h
,
h
) max
k
_
1 + log
H
k
h
k
_
2
s(
h
,
h
).
By a well-known argument, Theorem 3.1 implies that we can derive the following
corollary, where we denote by S and

S, respectively, the matrices corresponding to
the Galerkin discretization of the bilinear forms s and s.
Corollary 3.1. It holds that
(31) cond(

S
1
S) max
k
_
1 + log
H
k
h
k
_
2
.
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 669
We recall that by the denition of the bilinear form s, the matrix S is indeed
the Schur complement matrix dened by either (14) or (21).
In order to prove Theorem 3.1, we will basically follow the guidelines of the proof
of the analogous theorems of [9, 1]. The main dierence is that here we consider a
general discretization space, not necessarily consisting in piecewise linear functions.
Moreover we will be working with functions and operators dened directly on the
interface.
In view of the denition of s
i
and s
H
on the one hand and of property (26) of
s on the other hand, proving (30) is equivalent to proving that
h

h
with

h
=
H
+
0
h
(
H
L
H
and
0
h
in
0
h
), it holds that
|
h
|
2

|
H
|
2

+
M

i=1
|
0
h
|
2
H
1/2
00
(ei)
max
k
_
1 + log
H
k
h
k
_
2
|
h
|
2

.
We start by observing that it holds that
|
h
|
2

|
H
|
2

+|
0
h
|
2

|
H
|
2

+
M

i=1
|
0
h
|
2
H
1/2
00
(ei)
,
where the last bound descends by splitting
0
h

0
h
as
0
h
=

M
i=1

0
h,i
which
implies (since #(E
k
) c
4
and since each i belongs to E
k
for at most two ks)
|
0
h
|
2

k
[
0
h
[
2
H
1/2
(
k
)

iE
k
[
0
h,i
[
2
H
1/2
(
k
)

M

i=1
|
0
h
|
2
H
1/2
00
(ei)
,
the last inequality deriving from the H
1/2
00
(e
i
) to H
1/2
(
k
) boundedness of the
trivial extension-by-zero operator.
Moreover, by direct computation, using the linearity of
H
, one can see that it
holds that
(32) |
H
|
2

k=1
[
H
[
2
H
1/2
(
k
)

M

i=1
(
H
(a
i
)
H
(b
i
))
2
,
a
i
and b
i
being the extrema of e
i
. Then, proving Theorem 3.1 reduces to proving
the following result:
Theorem 3.2. For all
h

h
with
h
=
H
+
0
h
(
H
L
H
and
0
h

0
h
), it
holds that
(33)
M

i=1
(
H
(a
i
)
H
(b
i
))
2
+
M

i=1
|
0
h
|
2
H
1/2
00
(ei)
max
k
_
1 + log
H
k
h
k
_
2
|
h
|
2

.
In order to prove Theorem 3.2, we will need several bounds. We start by proving
the following Lemma.
Lemma 3.1. Assume that
h
satises assumptions (A5) and (A6). Then the
following bounds hold:
(i) for all
h

h
such that
h
(p) = 0 for some p e
i
with i E
k
, we have
(34) |
h
|
2
L

(ei)

_
1 + log
H
k
h
k
_
[
h
[
2
H
1/2
(ei)
;
670 SILVIA BERTOLUZZA
(ii) for all
h

h
it holds that
(35) (
h
(a
i
)
h
(b
i
))
2

_
1 + log
H
k
h
k
_
[
h
[
2
H
1/2
(ei)
,
a
i
and b
i
being the extrema of e
i
, i E
k
;
(iii) for all
0
h

0
h
it holds that
(36)

iE
k
|
0
h
|
2
H
1/2
00
(ei)

_
1 + log
H
k
h
k
_
2
[
0
h
[
2
H
1/2
(
k
)
.
Proof. Let i E
k
. By assumptions (A1)(a) and (A4) we have that [e
i
[ H
k
. We
start by observing that for any H
1/2+
(e
i
) it holds that
(37) ||
2
L

(ei)
H
1
k
||
2
L
2
(ei)
+
H
2
k

[[
2
H
1/2+
(ei)
.
This is easily seen by observing that on the reference segment e = ]0, 1[ the con-
tinuity bound of the injection H
1/2+
( e) L

( e) depends on as follows (see

Appendix A, Lemma A.1): for all H
1/2+
( e)
||
2
L

( e)
||
2
L
2
( e)
+
1

[[
2
H
1/2+
( e)
.
A change of variable then yields (37). Let now
h

h
and R. By assumption
(A6) it holds that
(38)
H
2
k

[
h
[
2
H
1/2+
(ei)
=
H
2
k

[
h
[
2
H
1/2+
(ei)

H
2
k
h
2
k

[
h
[
2
H
1/2
(ei)
.
After choosing = 1/ log(H
k
/h
k
), inequality (37) for =
h
combined with
(38) yields:
(39) |
h
|
2
L

(ei)
H
1
k
|
h
|
2
L
2
(ei)
+ log
H
k
h
k
[
h
[
2
H
1/2
(ei)
.
In view of (39), (34) can then be proven by applying exactly the arguments of
[9], which we repeat here for completeness: let us choose to be the average of
h
on e
i
. Applying Poincare inequality gives
(40) H
1/2
k
|
h
|
L
2
(ei)
[
h
[
H
1/2
(ei)
which yields
(41) |
h
|
2
L

(ei)

_
1 + log
H
k
h
k
_
[
h
[
2
H
1/2
(ei)
.
Now we remark that
h
(p) = 0 for some p e
i
implies
(42) [[ |
h
|
L

(ei)
and we get (34) by triangular inequality.
(ii) then follows by applying (i) to the function (x) (a
i
) (which vanishes in
p = a
i
).
As far as (iii) is concerned, we start by remarking that for
0
h

0
h
and for
i E
k
, denoting by
0
h,i
the function coinciding with
0
h
on e
i
and identically
vanishing on
k
e
i
and using (3) and the inverse inequality (27), we obtain the
following bounds
(43) |
0
h
|
H
1/2
00
(ei)
[
0
h,i
[
H
1/2
(
k
)
h

k
[
0
h,i
[
H
1/2
(
k
)
[
0
h
[
H
1/2
0
(ei)
,
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 671
where the second inequality descends from (27) by applying a standard functional
spaces interpolation argument, since the spaces H
r
0
(e
i
), 0 < r < 1/2, can be ob-
tained as the interpolants of order 2r between L
2
(e
i
) and H
1/2
00
(e
i
). Again we bound
the H
1/2
0
(e
i
) seminorm of
0
h
by a change of variable, using the corresponding
bound on the reference segment e = ]0, 1[ . We recall once more that H
1/2
0
( e)
and H
1/2
( e) coincide, with equivalent norms, the constant in the equivalence de-
pending on . In particular it is possible to prove (see Appendix A, Corollary A.1)
that the following bound holds for all H
1/2
( e) and for all R:
(44) [[
H
1/2
0
( e)

1

| |
H
1/2
( e)
+
1

[[,
which rescales as
[[
H
1/2
0
(ei)

H

_
H
1/2
k
| |
L
2
(ei)
+[ [
H
1/2
(ei)
_
+
H

[[.
Again, taking =
0
h
and as its average on e
i
, we can apply on the one hand
Poincare inequality (40) and on the other hand the bound (42), which holds since

0
h
(a
i
) = 0, and we get
|
0
h
|
H
1/2
00
(ei)

H

[
0
h
[
H
1/2
(ei)
+
H

|
0
h
|
L

(ei)
.
We then take once again = 1/ log(H
k
/h
k
), and, thanks to bound (41), we obtain
|
0
h
|
H
1/2
00
(ei)

_
1 + log
H
k
h
k
_
[
0
h
[
H
1/2
(ei)
.
Since

iE
k
[ [
2
H
1/2
(ei)
[ [
2
H
1/2
(
k
)
, by squaring and then taking the sum over
i E
k
, we obtain (36).
Remark 3.3. Since for s < 1/2 the two spaces H
s
(e
i
) and H
s
0
(e
i
) coincide, the
seminorm [ [
H
s
0
(ei)
in equation (43) could be replaced by the standard H
s
(e
i
)
seminorm. In that case, however, the implicit constant in the bound would depend
on .
Analogously to what happens in [9], the main ingredient of the proof of Theorem
3.2 is then the following lemma.
Lemma 3.2. Let
0
h

0
h
and
H
L
H
. Then for all k it holds that
(45)

iE
k
|
0
h
|
2
H
1/2
00
(ei)

_
1 + log
H
k
h
k
_
2
[
0
h
+
H
[
2
H
1/2
(
k
)
.
Provided such lemma holds, (33) is obtained, thanks to (ii) of Lemma 3.1, by
applying (45) for
H
=
H
. In fact, for
h
=
H
+
0
h
,
H
L
H
,
0
h

0
h
, we can
write
M

i=1
|
0
h
|
2
H
1/2
00
(ei)
=
1
2
K

k=1

iE
k
|
0
h
|
2
H
1/2
00
(ei)

k
_
1 + log
H
k
h
k
_
2
[
h
[
2
H
1/2
(
k
)
max
k
_
1 + log
H
k
h
k
_
2
|
h
|
2

,
672 SILVIA BERTOLUZZA
which, combined with (35), yields (33). The only thing that we need to do then is
to prove Lemma 3.2.
Proof of Lemma 3.2. Let k, 1 k K. We start by introducing a function
0

k,0
h
,
(46)
k,0
h
=
h

h
[

k
:
h
= 0 at the cross points of
k
=
0
h
[

k
,

0
depending on
H
, dened as the unique element of
k,0
h
satisfying
(
0
, )
H
1/2
(
k
)
= (
H
, )
H
1/2
(
k
)
,
k,0
h
,
where (, )
H
1/2
(
k
)
is the bilinear form
(, )
H
1/2
(
k
)
=
_

k
_

k
((x) (y))((x) (y))
[x y[
2
dxdy
inducing the H
1/2
(
k
) seminorm. It is easy to check that [ [
H
1/2
(
k
)
is a norm
on
k,0
h
and then we can conclude by standard arguments that
(47) [
0
[
H
1/2
(
k
)
[
H
[
H
1/2
(
k
)
.
Now, adding and subtracting
0
at the left-hand side of (45), we can write:
(48)

iE
k
|
0
h
|
2
H
1/2
00
(ei)


iE
k
|
0
h
+
0
|
2
H
1/2
00
(ei)
+

iE
k
|
0
|
2
H
1/2
00
(ei)
.
Since
0
h
+
0

k,0
h
, by the denition of
0
we have that
(49) (
H

0
,
0
h
+
0
)
H
1/2
(
k
)
= 0,
which yields
[
0
h
+
H
[
2
H
1/2
(
k
)
[
0
h
+
0
[
2
H
1/2
(
k
)
.
The rst sum on the right-hand side of (48) can then be bound thanks to Lemma
3.1(iii); we have

iE
k
|
0
h
+
0
|
2
H
1/2
00
(ei)

_
1 + log
H
k
h
k
_
2
[
0
h
+
0
[
2
H
1/2
(
k
)
(50)

_
1 + log
H
k
h
k
_
2
[
0
h
+
H
[
2
H
1/2
(
k
).
(51)
Let us bound the second sum on the right-hand side of (48). We recall that we
have
(52) |
0
|
2
H
1/2
00
(ei)
= [
0
[
2
H
1/2
(ei)
+I
1
(
0
) +I
2
(
0
),
with
(53) I
1
(
0
) =
_
bi
ai
[
0
(x)[
2
[x a
i
[
dx, I
2
(
0
) =
_
bi
ai
[
0
(x)[
2
[x b
i
[
dx,
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 673
(we recall that a
i
and b
i
denote the two vertices of the edge e
i
). Now, using
equations (47), (32) and (35), since
0
h
vanishes at a
i
and b
i
, we can write

iE
k
[
0
[
2
H
1/2
(ei)
[
0
[
2
H
1/2
(
k
)
[
H
[
2
H
1/2
(
k
)
(54)

iE
k
(
H
(a
i
)
H
(b
i
))
2
(55)
=

iE
k
([
0
h
+
H
](a
i
) [
0
h
+
H
](b
i
))
2
(56)

_
1 + log
H
k
h
k
_
[
0
h
+
H
[
2
H
1/2
(
k
)
. (57)
Let us now bound I
1
(
0
). For notational simplicity let us identify e
i
= ]0, T[
with a
i
= 0 and b
i
= T. We have
I
1
(
0
) =
_
T
0
[
0
(x)[
2
[x[
dx (58)
=
_
T
0
[
0
(x) +
H
(x)
H
(x) +
H
(0)
H
(0)[
2
[x[
dx (59)

_
T
0
[
0
(x)
H
(x) +
H
(0)[
2
[x[
dx (60)
+
_
T
0
[
H
(x)
H
(0)[
2
[x[
dx. (61)
Let us set

=
0

H
. Remarking that

(0) =
H
(0), we have
_
T
0
[
0
(x)
H
(x) +
H
(0)[
2
[x[
dx =
_
T
0
[

(x)

(0)[
2
[x[
dx
=
_
h
k
0
[

(x)

(0)[
2
[x[
dx
+
_
T
h
k
[

(x)

(0)[
2
[x[
dx.
The rst term is bound, using Cauchy-Schwarzs inequality and inequality (28), by
_
h
k
0
[

(x)

(0)[
2
[x[
dx =
_
h
k
0
1
[x[

_
x
0

x
() d

2
dx

_
h
k
0
_
x
0
[

x
()[
2
d dx
h
k
[

[
2
H
1
(ei)
[

[
2
H
1/2
(ei)
,
while, using (34), we bound the second by
_
T
h
k
[

(x)

(0)[
2
[x[
dx |

(0)|
2
L

(ei)
_
T
h
k
1
x
dx

_
1 + log
H
k
h
k
_
2
[

[
2
H
1/2
(ei)
,
674 SILVIA BERTOLUZZA
where we used T H
k
. The second integral in (60) can be bound directly using
the linearity of
H
.
_
T
0
[
H
(x)
H
(0)[
2
[x[
dx (
H
(b
i
)
H
(a
i
))
2

_
1 + log
H
k
h
k
_
[
H
+
0
h
[
2
H
1/2
(ei)
where again we used (35) and the fact that
0
h
vanishes at a
i
and b
i
. Then, adding
up the dierent contributions, we conclude that
(62) I
1
(
0
)
_
1 + log
H
k
h
k
_
2
[
0

H
[
2
H
1/2
(ei)
+
_
1 + log
H
k
h
k
_
[
H
+
0
h
[
2
H
1/2
(ei)
.
By using the same arguments we get a similar bound for I
2
:
(63) I
2
(
0
)
_
1 + log
H
k
h
k
_
2
[
0

H
[
2
H
1/2
(ei)
+
_
1 + log
H
k
h
k
_
[
H
+
0
h
[
2
H
1/2
(ei)
.
We can now insert bounds (54), (62) and (63) back into (52) and sum over i
E
k
. By observing that for any w H
1/2
(
k
) it holds that

iE
k
[w[
2
H
1/2
(ei)

[w[
2
H
1/2
(
k
)
, we obtain the bound

iE
k
|
0
|
2
H
1/2
00
(ei)

_
1 + log
H
k
h
k
_
[
0
h
+
H
[
2
H
1/2
(
k
)
(64)
+
_
1 + log
H
k
h
k
_
2
[
0

H
[
2
H
1/2
(
k
)
. (65)
We now observe once again that equation (49) implies that [
H

0
[
2
H
1/2
(
k
)

[
H
+
0
h
[
2
H
1/2
(
k
)
and we get the thesis.
4. Numerical Tests
We will test the preconditioner we propose on the following problem: nd u such
that
(66)
_
u = f, in = ]0, 1[ ]0, 1[ ,
u = 0, on = .
We consider a uniform, geometrically conforming, decomposition of = ]0, 1[
]0, 1[ in K = N N equal square subdomains of size H H, H = 1/N.
In each subdomain
k
we take a uniform mesh T
k
composed of n
k
n
k
equal
square elements of size
k

k
,
k
= H/n
k
= 1/(Nn
k
). We then dene V
k
h
to be
the space of Q
1
nite elements on the mesh T
k
,
V
k
h
= u
h
C
0
() : u
h
[

Q
1
(), T
k
,
where Q
1
() denotes the space of polynomials of degree less than or equal to one
in each variable. The multiplier space
k
h
is then dened as the trace on
k
of
V
k
h
. With such a choice it is possible to prove that (A3) holds (see [4]).
The skeleton is split as =

2(N1)N
i=1
e
i
with e
i
=
k

for some
k = k(i) and = (i). We remark that [e
i
[ = H for all i. We consider a grid ( on
obtained by uniformly splitting each e
i
into L
i
elements of size h
i
and we dene

h
to be the space of P
1
nite elements on such a grid:

h
=
h
C
0
() :
h
[

P
1
(), (,
h
[

= 0.
With this choice, assumption (A6) holds for h
k
= min
iE
k
h
i
.
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 675
Letting
0
h,i
(resp.
0
h,i
) denote the vector of the values of
0
h
(resp
0
h
)
0
h
at
the interior nodes of the grid ([
ei
, we dene the bilinear form s
i
:
0
h,i
[
ei

0
h,i
[
ei
as
s
i
(
0
h
[
ei
,
0
h
[
ei
) = (
0
h,i
)
T
(R
i
)
1/2

0
h,i
,
where R
i
is the stiness matrix associated to the discretization of the operator
d
2
/dx
2
by P
1
nite element on e
i
with homogeneous Dirichlet boundary conditions
at the extrema a
i
and b
i
. With R
i
being positive denite, its square root is well
dened and it is computed and stored at the start.
As far as the coarse preconditioner s
H
is concerned, we considered two choices:
(BPS) Following [9], we set
(67) s
H
(
H
,
H
) := 2
M

i=1
(
H
(a
i
)
H
(b
i
))(
H
(a
i
)
H
(b
i
)).
(Laplace) Denoting by
H
and

H
the H
1
0
() functions obtained by lifting

H
and
H
harmonically in each subdomain, we set
(68) s
H
(
H
,
H
) :=
_

H
dx;
note that the set
H
,
H
L
H
coincides with the set of Q
1
nite elements
on the coarse grid corresponding to the decomposition =

k
, and
then applying the coarse preconditioner reduces to numerically solving the
Laplace equation on such space.
For all tests we set f = 1. We solved the linear system by a preconditioned
conjugate gradient method, using
h
= 0 as initial guess. We report the number
of iterations needed to reduce the residual of a factor 10
5
.
4.1. Plain formulation.
4.1.1. Conforming decomposition. We start by considering the case of a conforming
domain decomposition, that is, in the framework described above we set n
k
= n
for all k and L
i
= n for all i. In this case
k
= h
i
= h = H/n. With such a choice,
it is not dicult to check that the inf-sup condition (12) is satised uniformly in h
and then we are in the range of Lemma 2.1. It is easy to realize that the discrete
solution veries u
k
h
=
h
on
k
, and then the function u
h
, dened by u
h
[

k
= u
k
h
,
veries u
h
H
1
0
(). In such a case we obtain the same solution that we would
get using a conforming nite element method on a regular grid of NnNn square
elements of dimension h h. The use of a nonconforming method is therefore, in
this case, only a device to implement a solver for the discrete problem in an easily
parallelizable way.
In order to study both the dependence on H (size of the subdomains) and on h
we tested the preconditioner for dierent values of n in the range [5, 40] and N in
the range [4, 32]. We tested both coarse preconditioners for all combinations of N
and n. The results are summarized in Tables 1 and 2. As one can see, they are in
close agreement with the theory (note that in this case max
k
h
k
/H
k
= h/H = n).
676 SILVIA BERTOLUZZA
Table 1. Number of conjugate gradient iterations needed for re-
ducing the residual of a factor 10
5
, with coarse preconditioner
given by (68) for dierent combinations of the number K = N
2
of
subdomains and n of elements per edge (n
2
elements per subdo-
main).
n = 5 n = 10 n = 20 n = 40
K = N
2
# Iter. # Iter. # Iter. # Iter.
16 10 12 14 15
64 14 17 20 23
144 15 18 22 26
256 15 19 23 27
400 16 20 23 27
576 16 20 24 27
784 16 20 24 27
1024 16 20 24 27
Table 2. Number of conjugate gradient iterations needed for re-
ducing the residual of a factor 10
5
, with coarse preconditioner
given by (67) for dierent combinations of the number K = N
2
of
subdomains and n of elements per edge (n
2
elements per subdo-
main).
n = 5 n = 10 n = 20 n = 40
K = N
2
# Iter. # Iter. # Iter. # Iter.
16 8 9 11 13
64 10 13 17 22
144 11 15 20 26
256 13 16 22 30
400 13 17 23 32
576 13 17 24 33
784 14 18 25 33
1024 14 18 25 33
4.1.2. Nonconforming decomposition. We consider then the case of a truly noncon-
forming decomposition, without stabilization (formulation (10)). In our test for
e
i
=
k

, we set L
i
= minn
k
, n

. This discretization does not in general

satisfy the assumptions required for the uniform stability of the discrete problem.
Nevertheless, using a standard technique, it is not dicult to prove that (12) holds
provided that for all , k such that [

k
[ > 0 it either holds that n

= n
k
or
maxn

, n
k
/ minn

, n
k
C for some constant C > 1.
We considered two test congurations, both on a decomposition of into 5 5
subdomains. In the rst case (see Figure 1, top), for all subdomains except two,
we set n
k
= 10, while on the remaining two subdomains (subdomains 13 and 14,
if we number all subdomains row-wise) we set n
k
= n, where n varies in the set
[10, . . . , 52]. With the above criterion used to x L
i
, we have that L
i
= 10 for all
edges e
i
, except the common edge
13

14
(the two subdomains are adjacent),
where L
i
= maxn
13
, n
14
= n varies.
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 677
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(1)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(2)
Figure 1. The two congurations on which the preconditioner is
tested for the plain formulation (8) in the nonconforming case.
In the second conguration (see Figure 1, bottom) for all subdomains we again
set n
k
= 10, this time except that on a three-by-three block of subdomains where
n
k
= n, with n varying again between 10 and 52. Here, for 12 out of 40 edges of ,
we have that L
i
= n. For both congurations we tested both coarse preconditioners.
The results are provided in Table 3.
678 SILVIA BERTOLUZZA
Table 3. Number of conjugate gradient iterations needed for re-
ducing the residual of a factor 10
5
for Congurations 1 and 2
(corresponding, respectively, to the top and to the bottom decom-
positions illustrated in Figure 1). Both coarse preconditioners (67)
(BPS) and (68) (Laplace) are tested for dierent values of the
number n of elements per edge in the rened part of the skeleton
(one edge for Conguration 1 and twelve edges for Conguration
2). The number of elements per edge in the remaining part of the
skeleton is set to 10.
Laplace BPS
# It. # It.
n = 10 12 8
n = 17 18 11
n = 24 18 13
n = 31 18 13
n = 38 18 14
n = 45 19 14
n = 52 19 14
(a) Conguration 1.
Laplace BPS
# It. # It.
n = 10 12 8
n = 17 20 14
n = 24 23 16
n = 31 24 18
n = 38 25 19
n = 45 26 20
n = 52 27 20
(b) Conguration 2.
As one can easily see, though the theoretical estimate on the condition number
only depends on max
k
H
k
/h
k
, which takes the same value for both congurations, in
the rst case in which the overall discretization of the skeleton is sensibly coarser
the mesh being ne only in a small portion (in our case one) of the edges composing
the inuence of rening the mesh is weaker than in the second case, in which a
relevant portion of the skeleton is rened.
4.2. Stabilized formulation. We now consider the stabilized formulation (19).
Since the H
1/2
(
k
) seminorm is invariant under changes of scale, we can describe
the stabilizing form for subdomains scaled in such a way that [
k
[ = 1 (that is,
H = 1/4). We can then identify
k
with the circle T of unitary length. Following
the proposal of [7], the bilinear forms [, ] 1
2
,k
are designed by means of a wavelet
decomposition. More precisely, we restrict the number L
i
of elements of the i-th
edge e
i
to be a power of two,
L
i
= 2
ji
for some j
i
1,
so that for all k,
h
[

k
V
j
k
+2
where for j > 0, V
j
denotes the space of P
1
nite
elements on the uniform grid (
j
of T with mesh size 1/2
j
.
The sequence V
j

j0
forms a so-called multiresolution analysis of L
2
(T) and
it is well known (see for example [15]) that there exist several wavelet bases asso-
ciated with such a multiresolution analysis. More precisely there exist several P
1
compactly supported functions C
0
(R) dened on the uniform grid of mesh size
1 and integer nodes, such that, setting

m
=
+

n=
2
m/2
(2
m
(x n) ),
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 679
all functions
j
V
j
can be written as

j
=
0
+
j1

m=0
2
m

=1

m
,
0
constant,
and such that
[
j
[
2
H
1/2
(
k
)

j1

m=0
2
m

=1
2
m
[
m
[
2
.
Given the values of
j
at the nodes of the uniform grid, the coecients (
m
)
m,l
can be retrieved by a fast (O(2
j
)) wavelet transform (FWT).
If we denote by P
k
: L
2
(T) V
j
k
+2
the L
2
(T) orthogonal projection, for ,
L
2
(T) we can dene corresponding wavelet coecients

m
and

m
such that
P
k
() =
0
+
j
k

m=0
2
m

=1

m
, P
k
() =
0
+
j
k

m=0
2
m

=1

m
.
Then we dene the bilinear form [, ] 1
2
,k
as
[, ] 1
2
,k
=
j
k

m=0
2
m

=1
2
m

m
.
It is possible to prove ([7, 14]) that the bilinear forms thus dened satisfy assump-
tions (A4) and (18).
Remark 4.1. Note that the coecients

m
and

m
are not the classical wavelet
coecients of the functions and . These would in fact be dened as the scalar
product of, respectively, and with suitable dual functions

m
L
2
(T). The
coecients

m
and

m
are rather the scalar products of the dual functions

m
with the projections of and onto V
j
k
+2
, respectively. Naturally, such coecients
coincide with the classical ones, provided f V
j
k
+2
.
With this denition, the computation of [u
k
h

h
, v
k
h

h
] 1
2
,k
essentially reduces
to rst computing the nodal values of P
k
(u
k
h
), P
k
(v
k
h
), P
k
(
h
) and P
k
(
h
), respec-
tively, and then applying a FWT. As far as
h
and
h
are concerned, the choice
of j
k
implies that
h
,
h
V
j
k
+2
and then their nodal values are simply computed
by interpolation. As far as u
k
h
and v
k
h
are concerned, we need to compute their
L
2
projection on V
j
k
+2
. Observing that the mass matrix in V
j
k
+2
has a circulant
structure with bandwidth equal to three, this can also be done in O(2
j
k
) operations
by a modication of the Croute reduction algorithm for solving linear systems with
tridiagonal matrices. For further details on the implementation of the stabilized
formulation we refer to [8].
We tested the preconditioner (with both form (67) and (68) of the coarse bilinear
form) on four dierent splittings of into, respectively, 4 4, 8 8, 12 12 and
16 16 subdomains. In each case we randomly assigned the values of n
k
in such
a way that for 1/3 of the subdomains n
k
= 5, for about another third n
k
= 10,
and for the remaining subdomains n
k
= 15. The four congurations considered are
shown in Figure 2. For approximating the interface function , we used a uniform
discretization of the interface (L
i
= n for all i). For each conguration, we tested
the preconditioner for dierent values of n and of the stabilization parameter .
The function appearing in the denition of the stabilizing bilinear form is chosen
680 SILVIA BERTOLUZZA
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(a)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(b)
Figure 2. The four congurations on which the preconditioner is
tested for the stabilized formulation (10). (Continues)
to be the piecewise linear wavelet corresponding to the 2.2 B-spline biorthogonal
setting ([15]). Note (see Table 4) that for = 0 (no stabilization!) if n is big, the
CG algorithm does not converge within the maximum number of iterations (which
was set equals to 100), conrming the need for using some kind of stabilization.
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 681
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(c)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(d)
Figure 2. (continued)
As one can see from the results, the value of the parameter has an inuence on
the conditioning of the matrix

S
1
S. If is too small (see Table 5), the action of the
stabilization term is not sucient to compensate for the lack of validity of the inf-
sup condition, and therefore the system behaves almost like the one obtained using
the plain formulation. By increasing (see Tables 6 and 7), the situation improves
682 SILVIA BERTOLUZZA
dramatically and one can see that, for instance, for = .05 in the case of 16 16
subdomains with n = 32 (which corresponds to 15360 degrees of freedom for
h
)
the residual is reduced of a factor 10
5
in about 35 iterations, which is only slightly
worse than the result obtained for the analogous values in the conforming case.
However, when is further increased (see Table 8), the performances deteriorate.
This (as well as the dierence with respect to the conforming case) is probably
due to the fact that the constants in the estimates (26)which have an inuence
on the bound on the condition numberdepend on the ratio B/b (B and b being
the constants in (18) and in assumption (A3)(c

), respectively) amplied by the

factor . If such a ratio (which is always greater than or equal to one) in not close
enough to 1, the choice of the right balance between the original bilinear form and
the stabilizing term can become important from the point of view of the spectral
properties of the operator, which are essential for preconditioning.
Table 4. Number of conjugate gradient iterations needed to re-
duce the residual of a factor 10
5
on the four congurations de-
picted in Figure 2 with the plain formulation (10). Both coarse pre-
conditioners are tested for dierent values of the number n of ele-
ments per edge. As one can see for n 16, the CG method does not
converge within the maximum number of iteration (= 100), con-
rming the instability of the plain formulation for the discretization
considered.
(a) (b) (c) (d)
n # It. # It. # It. # It.
4 14 17 18 19
8 47 57 60 69
16
32
Laplace
(a) (b) (c) (d)
n # It. # It. # It. # It.
4 11 13 14 15
8 40 52 54 64
16
32
BPS
Table 5. Number of conjugate gradient iterations needed to re-
duce the residual of a factor 10
5
on the four congurations de-
picted in Figure 2 with the stabilized formulation (19) for =
.00125. Both coarse preconditioners are tested for dierent values
of the number n of elements per edge. One can clearly see that
the method converges very badly or does not converge. This is due
to the fact that, since the stabilization parameter is very small,
the action of the stabilization term is not sucient to compensate
for the lack of validity of the inf-sup condition, and therefore the
system behaves like the one obtained using the plain formulation.
(a) (b) (c) (d)
n # It. # It. # It. # It.
4 14 17 18 19
8 52 60 64 83
16 86
32 96
Laplace
(a) (b) (c) (d)
n # It. # It. # It. # It.
4 11 13 14 15
8 39 51 57 77
16 69 92
32 85
BPS
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 683
Table 6. Number of conjugate gradient iterations needed to re-
duce the residual of a factor 10
5
on the four congurations de-
picted in Figure 2 with the stabilized formulation (19) for = 1.
Both coarse preconditioners are tested for dierent values of the
number n of elements per edge.
(a) (b) (c) (d)
n # It. # It. # It. # It.
4 14 17 18 19
8 22 26 27 29
16 28 34 36 38
32 34 42 44 45
Laplace
(a) (b) (c) (d)
n # It. # It. # It. # It.
4 11 13 14 15
8 18 21 24 27
16 23 31 35 38
32 30 41 46 51
BPS
Table 7. Number of conjugate gradient iterations needed to re-
duce the residual of a factor 10
5
on the four congurations de-
picted in Figure 2 with the stabilized formulation (19) for = 5.
Both coarse preconditioners are tested for dierent values of the
number n of elements per edge.
(a) (b) (c) (d)
n # It. # It. # It. # It.
4 14 17 18 19
8 19 22 23 25
16 22 28 29 31
32 24 31 32 34
Laplace
(a) (b) (c) (d)
n # It. # It. # It. # It.
4 11 13 14 15
8 15 17 19 22
16 16 23 26 29
32 19 27 30 35
BPS
Table 8. Number of conjugate gradient iterations needed to re-
duce the residual of a factor 10
5
on the four congurations de-
picted in Figure 2 with the stabilized formulation (19) for = 20.
Both coarse preconditioners are tested for dierent values of the
number n of elements per edge.
(a) (b) (c) (d)
n # It. # It. # It. # It.
4 14 17 18 19
8 25 26 27 30
16 31 37 38 41
32 35 41 43 48
Laplace
(a) (b) (c) (d)
n # It. # It. # It. # It.
4 12 13 14 15
8 23 24 24 27
16 28 35 35 37
32 34 40 39 41
BPS
Appendix A: Some Sobolev space injection bounds
The aim of this section is to recall some known bounds for which we provide a
new proof based on the use of equivalent norms expressed through wavelet coe-
cients (see for instance [15]). Such proof will easily reveal the dependence of the
684 SILVIA BERTOLUZZA
constants in the bounds on the smoothness of the spaces considered and the way
these constants explode as the smoothness index approaches the limit of validity of
the bounds.
The key ingredient of the following arguments is the possibility of constructing
two orthonormal bases
B =
k
, k = 1, . . . , 2
j0

j,k
, j j
0
, k = 1, . . . , 2
j
,
B
0
=
0
k
, k = 1, . . . , 2
j0

0
j,k
, j j
0
, k = 1, . . . , 2
j
,
for L
2
(0, 1), tuned up to provide norm equivalences for Sobolev spaces, respectively,
without and with homogeneous boundary conditions. In particular, such bases can
be constructed in such a way that all u L
2
(0, 1) can be written both ways as
u =
2
j
0

k=1
u,
k
)
k
+

j=j0
2
j

k=1
u,
j,k
)
j,k
,
u =
2
j
0

k=1
u,
0
k
)
0
k
+

j=j0
2
j

k=1
u,
0
j,k
)
0
j,k
and that the following norm equivalences hold for all s with 0 s 1:
|u|
H
s
(0,1)

_
_
2
j
0

k=1
[u,
k
)[
2
+

j=j0
2
j

k=1
2
2js
[u,
j,k
)[
2
_
_
1/2
, (69)
|u|
H
s
0
(0,1)

_
_
2
j
0

k=1
[u,
0
k
)[
2
+

j=j0
2
j

k=1
2
2js
[u,
0
j,k
)[
2
_
_
1/2
, s ,= 1/2, (70)
|u|
H
1/2
00
(0,1)

_
_
2
j
0

k=1
[u,
0
k
)[
2
+

j=j0
2
j

k=1
2
j
[u,
0
j,k
)[
2
_
_
1/2
, (71)
with constants independent of s. Such bases are constructed in such a way that the
basis functions satisfy the following properties:
W.1. For all j j
0
and k = 1, . . . , 2
j
the functions
j,k
have a certain number
of vanishing moments; in particular
j,k
has zero mean value.
W.2. The basis functions are well localized and the supports of the basis functions
are properly scaled with respect to j: for some N xed, satisfying N
2
j02
, we have
supp
j,k
[(k N)2
j
, (k +N)2
j
] ]0, 1[ ,
supp
0
j,k
[(k N)2
j
, (k +N)2
j
] ]0, 1[ .
W.3. The interior functions of the two bases coincide:
j,k
=
0
j,k
, k =
N, . . . , 2
j
N; in other words the two bases dier only for those elements
that take into account boundary conditions.
W.4. The basis functions verify
(72) [
k
(x)[ C, [
0
k
(x)[ C, [
j,k
(x)[ 2
j/2
, [
0
j,k
(x)[ 2
j/2
.
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 685
Remark A.1. Note that the constants N and j
0
(and therefore 2
j0
) are xed once
for all and depend only on the choice of the wavelet basis. We will therefore consider
them as an O(1) as far as the estimates that we are going to prove are concerned.
The normalization estimate (72), together with property W.2 on the size of the
support, trivially yields the following bound on the L
1
norm of the basis functions
(73) |
j,k
|
L
1
(0,1)
2
j/2
, |
0
j,k
|
L
1
(0,1)
2
j/2
.
In the following we will make use of a certain number of algebraic inequalities:
we start by recalling the following bound on the sum of a geometric series
(74)

j1
2
aj

1
a
,
as well as the well-known boundedness property of the discrete convolution operator
(75)

j
a
j

j
b
j
[
2

_
_

j
[a
j
[
_
_
2
_
_

j
[b
j
[
2
_
_
.
Moreover, since N is a xed constant, we have
(76)
_
N

k=1
a
k
_
2

k=1
[a
k
[
2
.
The rst bound that we consider regards the injection H
s
(0, 1) L

(0, 1),
which is known to hold for all s > 1/2 [17]. As far as the dependence on s of the
constant in the bound is concerned, we have the following lemma.
Lemma A.1. The following bound holds for all u H
s
(0, 1) with s ]1/2, 1]:
|u|
L

(0,1)
|u|
L
2
(0,1)
+
1
_
s 1/2
[u[
H
s
(0,1)
.
Proof. Let u =

2
j
0
k=1
u,
k
)
k
+

j=j0

2
j
k=1
u,
j,k
)
j,k
. Using (72) and (69), as
well as properties W.1 and W.2 (which guarantees that for x and j xed,
j,k
(x) ,= 0
for at most 2N values of k), we have (with u =
_
1
0
u ds)
sup
x
[u(x)[ sup
x
[
2
j
0

k=1
u,
k
)
k
(x) +

j=j0
2
j

k=1
u,
j,k
)
j,k
(x)[
max
k
[u,
k
)[ +

jj0
max
k
[u u,
j,k
)[2
sj
2
(1/2s)j
|u|
L
2
(0,1)
+
_
_

jj0
4
(1/2s)j
_
_
1/2
_
_

jj0
2
j

k=1
[u u,
j,k
)[
2
2
2sj
_
_
1/2
|u|
L
2
(0,1)
+
1
_
s 1/2
|u u|
H
s
(0,1)
.
We conclude by applying Poincare inequality which yields |u u|
H
s
(0,1)

[u u[
H
s
(0,1)
= [u[
H
s
(0,1)
.
686 SILVIA BERTOLUZZA
The second issue that we consider is the equivalence between the spaces H
s
(0, 1)
and H
s
0
(0, 1). This is known to hold for all s [0, 1/2[ [17]. The dependence on s
of the constant in the bound is the object of the following lemma.
Lemma A.2. The following bound holds for all H
s
, 0 s < 1/2:
(77) ||
H
s
0
(0,1)

1
1/2 s
||
H
s
(0,1)
.
Proof. Let H
s
(0, 1) be expanded in terms of the basis B:
=
2
j
0

k=1

k
+

jj0
2
j

k=1

j,k

j,k
,
so that
||
2
H
s
(0,1)

2
j
0

k=1
[
k
[
2
L
2 +

jj0
2
j
2
j

k=1
[
j,k
[
2
.
We start by decomposing as the sum of four contributions:
=
0
+
left
+
center
+
right
with

0
=
2
j
0

k=1

k
,
center
=

jj0
2
j
N

k=N

j,k

j,k
=

jj0
2
j
N

k=N

j,k

0
j,k
, (78)

left
=

jj0
N1

k=1

j,k

j,k
,
right
=

jj0
2
j

k=2
j
N+1

j,k

j,k
(79)
and by triangular inequality we have
||
H
s
0
(0,1)
|
0
|
H
s
0
(0,1)
+|
left
|
H
s
0
(0,1)
+|
center
|
H
s
0
(0,1)
+|
right
|
H
s
0
(0,1)
.
In order to bound the H
s
0
(0, 1) norm of the four contributions, we will use the
equivalent norm (70). We start by observing that for all (j, n) with n = N, . . . , 2
j

N we have
k
,
0
j,n
) =
k
,
j,n
) = 0, while for the remaining couples (j, n) we
have
[
k
,
0
j,n
)[ |
k
|
L

(0,1)
|
0
j,n
|
L
1
(0,1)
2
j0/2
2
j/2
;
using (70) and observing that for all js the number of indexes n for which
k
,
0
j,n
)
,= 0 is bounded independently of j, we easily see that
|
k
|
2
H
s
0
(0,1)
2
j0
_
_
|
k
|
2
L
2 +

jj0
2
2js
2
j
_
_

1
1 2s
.
Using the fact that on R
2
j
0
the
2
and the
1
norms are equivalent (the constant
in the equivalence depending on j
0
, which, we recall, is a xed constant), we then
easily see that
|
0
|
2
H
s
0

1
1 2s
_
_
2
j
0

k=1
[
k
[
_
_
2

1
1 2s
2
j
0

k=1
[
k
[
2
.
We now consider
center
. This term contains the contribution of all functions
which are interior; i.e., they do not see the boundary conditions. More precisely,
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 687
we observe that for all j

, k

such that
j

,k
=
0
j

,k
it holds that
center
,
0
j

,k
) =

,k
, while for all other j

, k

,
center
,
0
j

,k
) = 0 and this yields
|
center
|
2
H
s
0
(0,1)


jj0
2
2sj
2
j
N

k=N
[
j,k
[
2
.
Let us now bound the H
s
0
(0, 1) norm of
left
, which is the contribution of those
basis functions which see the left boundary. Considering that for all N k

2
j
N, k N 1, it holds that
j,k
,
0
j

,k
) =
j,k
,
j

,k
) = 0 and that for
k N 1 and k

> 2
j
N we have supp
j,k
supp
0
j

,k
= , using (70), we can
write
|
left
|
2
H
s
0
(0,1)

2
j
0

=1
[
left
,
0
k
)[
2
+

=j0
2
2sj
N1

=1
[
left
,
0
j

,k
)[
2
.
It is not dicult to realize that the rst sum on the right-hand side can be bound
as follows:
2
j
0

=1
[
left
,
0
k
)[
2
|
left
|
2
L
2
(0,1)
||
H
s
(0,1)
.
In order to bound the second sum, let us now bound [
left
,
0
j

,k
)[. We observe
that, thanks to (72) and (73), for all j, k, j

, k

we can write
[
j,k
,
0
j

,k
)[ |
j,k
|
L

(0,1)
|
0
j

,k
|
L
1
(0,1)
2
(jj

)/2
as well as
[
j,k
,
0
j

,k
)[ |
j,k
|
L
1
(0,1)
|
0
j

,k
|
L

(0,1)
2
(j

j)/2
,
which yield
[
j,k
,
0
j

,k
)[ 2
|jj

|/2
.
Then we have
[
left
,
0
j

,k
)[

j=j0
N1

k=1
[
j,k
[ [
j,k
,
0
j

,k
)[

j=j0
2
|jj

|/2
N1

k=1
[
j,k
[.
Since the right-hand side of the above estimate does not depend on k

, using in-
equalities (74), (75) and (76), we can write

=j0
2
2j

s
N1

=1
[
left
,
0
j

,k
)[
2
(N 1)

=j0
2
2j

s
_
_

j=j0
2
|jj

|/2
N1

k=1
[
j,k
[
_
_
2

=j0
_
_

j=j0
2
(1/2s)|jj

|
N1

k=1
2
js
[
j,k
[
_
_
2

_
_

j=j0
2
(1/2s)j
_
_
2
_
_

j=j0
2
2js
N1

k=1
[
j,k
[
2
_
_
,
688 SILVIA BERTOLUZZA
which yields
|
left
|
2
H
s
0
(0,1)

1
(1/2 s)
2
_
_

j=0
2
2js
N

k=1
[
j,k
[
2
_
_
.
Analogously, we can prove
|
right
|
2
H
s
0
(0,1)

1
(1/2 s)
2
_
_

j=0
2
2js
2
j

k=2
j
N+1
[
j,k
[
2
_
_
.
Adding the bounds for |
0
|
H
s
0
(0,1)
, |
left
|
H
s
0
(0,1)
, |
center
|
H
s
0
(0,1)
and
|
right
|
H
s
0
(0,1)
, since for s in [0, 1/2[ it holds that 1 1/(1/2 s), in view of
the norm equivalence (69) we obtain the thesis.
Corollary A.1. For all H
1/2
(0, 1) and for all R it holds, for all s ]0, 1/2[,
that
||
H
s
0
(0,1)

1
1/2 s
| |
H
1/2
(0,1)
+
[[
_
1/2 s
.
Proof. Trivially it holds that
||
H
s
0
(0,1)
| |
H
s
0
(0,1)
+||
H
s
0
(0,1)
.
In view of bound (77) the only thing that needs to be proven is a bound on
||
H
s
0
(0,1)
, which can be done by direct computation, yielding the thesis.
References
[1] Y. Achdou, Y. Maday, and O. Widlund. Substructuring preconditioners for the mortar
method in dimension two. SIAM J. Numer. Anal., 36:551580, 1999. MR 99m:65233
[2] C. Baiocchi, F. Brezzi, and D. Marini. Stabilization of Galerkin methods and application
to domain decomposition. In Future Tendencies in Computer Science, Control and Applied
Mathematics. 1992. MR 94g:65119
[3] C. Bernardi, Y. Maday, and A.T. Patera. A new nonconforming approach to domain decom-
position: The mortar element method. In H. Brezis & J.-L. Lions, editors, Nonlinear Partial
Dierential Equations and their Applications, Coll`ege de France Seminar, volume XI of
Notes Math. Ser. 299, pages 1351. 1994. MR 95a:65201
[4] S. Bertoluzza. Analysis of a stabilized three elds domain decomposition method. Technical
Report 1175, I.A.N.-C.N.R., 2000.
[5] S. Bertoluzza. Wavelet stabilization of the Lagrange multiplier method. Numer. Math., 86:1
28, 2000. MR 2001h:65142
[6] S. Bertoluzza, C. Canuto, and A. Tabacco. Stable discretization of convection-diusion
problems via computable negative order inner products. SINUM, 38:10341055, 2000. MR
2001m:65163
[7] S. Bertoluzza and A. Kunoth. Wavelet stabilization and preconditioning for domain decom-
position. I.M.A. Jour. Numer. Anal., 20:533559, 2000. MR 2001h:65156
[8] S. Bertoluzza and G. Manzini. Wavelet stabilization of the three elds domain decomposition
method: Implementation and numerical tests. In preparation.
[9] J. H. Bramble, J. E. Pasciak, and A. H. Schatz. The construction of preconditioners for elliptic
problems by substructuring. Math. of Comp., 47(175):103134, 1986. MR 87m:65174
[10] J.H. Bramble, J.E. Pasciak, and P.S. Vassilevski. Computational scales of Sobolev norms with
application to preconditioning. Math. Comp., 69:463480, 2000. MR 2000k:65088
[11] F. Brezzi and M. Fortin. Mixed and Hybrid Finite Element Methods. Springer, 1991. MR
92d:65187
[12] F. Brezzi, L. Franca, D. Marini, and A. Russo. Stabilization techniques for domain decompo-
sition methods with nonmatching grids. In Proc. IX Domain Decomposition Methods Con-
ference.
THE THREE FIELDS DOMAIN DECOMPOSITION METHOD 689
[13] F. Brezzi and D. Marini. Error estimates for the three-eld formulation with bubble stabi-
lization. Math. of Comp., 70:911934, 2001. MR 2002b:65159
[14] F. Brezzi and D. Marini. A three-eld domain decomposition method. In A. Quarteroni,
J. Periaux, Y.A. Kuznetsov, and O.B. Widlund, editors, Domain Decomposition Methods
in Science and Engineering, volume 157 of American Mathematical Society, Contemporary
Mathematics, pages 2734, 1994. MR 95a:65202
[15] A. Cohen. Numerical analysis of wavelet methods. In P.G. Ciarlet and J.L. Lions, editors,
Handbook in Numerical Analysis, volume VII. Elsevier Science Publishers, North Holland,
2000. MR 2002c:65252
[16] W. Dahmen and A. Kunoth. Multilevel preconditioning. Numer. Math., 63:315344, 1992.
MR 93j:65065
[17] J.L. Lions and E. Magenes. Non-homogeneous Boundary Value Problems and Applications.
Springer, 1972. MR 50:2670, MR 50:2671
[18] P. Le Tallec and T. Sassi. Domain decomposition with nonmatching grids: Augmented La-
grangian approach. Math. Comp., 64:13671396, 1995. MR 95m:65212
Istituto di Matematica Applicata e Tecnologie Informatiche del Consiglio Nazionale
delle Ricerche, v. Ferrata 1, 27100 Pavia, Italy
E-mail address: [email protected]
URL: http://www.imati.cnr.it/~aivlis