NUMERICAL LINEAR ALGEBRA WITH APPLICATIONS

Numer. Linear Algebra Appl. 2006; 00:115 Prepared using nlaauth.cls [Version: 2002/09/18 v1.02]
Preconditioners for the discretized time-harmonic Maxwell
equations in mixed form
Chen Greif
1
and Dominik Schotzau
2
1
Department of Computer Science, University of British Columbia, Vancouver, B.C., V6T 1Z4, Canada,
[email protected]
2
Mathematics Department, University of British Columbia, Vancouver, B.C., V6T 1Z2, Canada,
[email protected]
SUMMARY
We introduce a new preconditioning technique for iteratively solving linear systems arising from
nite element discretization of the mixed formulation of the time-harmonic Maxwell equations. The
preconditioners are motivated by spectral equivalence properties of the discrete operators, but are
augmentation-free and Schur complement-free. We provide a complete spectral analysis, and show
that the eigenvalues of the preconditioned saddle point matrix are strongly clustered. The analytical
observations are accompanied by numerical results that demonstrate the scalability of the proposed
approach. Copyright c 2006 John Wiley & Sons, Ltd.
key words: time-harmonic Maxwell equations, nite element methods, saddle point linear systems,
preconditioners, Krylov subspace solvers
1. INTRODUCTION
We introduce new preconditioners for linear systems arising from nite element discretization of
the mixed formulation of the time-harmonic Maxwell equations in lossless media with perfectly
conducting boundaries [4, 5, 14, 19]. The following model problem with constant coecients
is considered: nd the vector eld u and the multiplier p such that
u k
2
u +p = f in ,
u = 0 in ,
u n = 0 on ,
p = 0 on .
(1)
Here R
3
is a simply connected polyhedron domain with a connected boundary , and n
denotes the outward unit normal on . The datum f is a given generic source (not necessarily
Contract/grant sponsor: Natural Sciences and Engineering Research Council of Canada.
Copyright c 2006 John Wiley & Sons, Ltd.
2 C. GREIF AND D. SCH

OTZAU
divergence-free), and the wave number satises k
2
=
2
, where 0 is the frequency, and
and are positive permittivity and permeability parameters. We assume that k
2
is not a
Maxwell eigenvalue and that
k
2
<1.
The introduction of the scalar variable p guarantees the stability and well-posedness of the
equations as k tends to 0, including the limit case k = 0; see the discussion in [5, Section 3].
Finite element discretization using Nedelec elements of the rst kind [18] for the
approximation of the vector eld and standard nodal elements for the multiplier yields a
saddle point linear system of the form

Ak
2
M B
T
B 0

u
p

g
0

, (2)
where now u R
n
and p R
m
are nite arrays representing the nite element approximations,
and g R
n
is the load vector associated with f. The matrix A R
nn
is symmetric positive
semidenite with nullity m, and corresponds to the discrete curl-curl operator; B R
mn
is
a discrete divergence operator with full row rank, and M R
nn
is the vector mass matrix.
It is possible to decouple (2) into two separate problems, using the discrete Helmholtz
decomposition [17, Section 7.2.1]. For p we obtain a standard Poisson equation, for which
many ecient solution methods exist. Then, once p is available, the high nullity of the discrete
curl-curl operator in the resulting equation for u can be dealt with by applying a procedure of
augmentation: the matrix A is replaced by A
W
= A+B
T
W
1
B, where W R
mm
is a weight
matrix, chosen so that A
W
is symmetric positive denite. This does not change the solution,
due to the divergence-free condition Bu = 0. Popular choices for W that have been considered
in the literature are scaled identity matrices or lumped mass matrices; see [13, pp. 319320]
and references therein. A similar approach in the context of nite volume methods has been
proposed in [11].
We note that if k = 0, a direct approach based on solving (A k
2
M)u = g automatically
enforces Bu = 0, provided the right hand side is divergence-free. A multigrid technique for
this case has been proposed in [8]. A regularization technique is introduced in [20] to deal with
the case k = 0, whereby A is replaced by A + M, where is a regularization parameter.
Algebraic multigrid is shown to converge even for small . The solution is divergence-free for
divergence-free data, but it changes with the parameter.
Leaving the saddle point system intact is a viable approach that works naturally for the
limiting case k = 0, which is our main interest in this paper. The saddle point matrix does not
have to be modied or regularized even if its (1,1) block is singular, and its structure lends itself
to eective block preconditioners. Indeed, there are several robust solution methods available
for solving saddle point systems [2].
An Uzawa-type algorithm for the saddle point system, coupled with a domain decomposition
approach, has been proposed in [16]. The original system is transformed into a new system by
augmentation with the scalar Laplacian as a weight matrix, and it is shown that the condition
number of the resulting preconditioned system grows logarithmically with respect to the ratio
between the subdomain diameter and the mesh size. The method incorporates augmentation
and is parameter-dependent. Its convergence properties rely on extreme eigenvalues of the
augmented Schur complement, which may be dicult to evaluate.
In this paper we introduce a new block diagonal preconditioning technique for the
iterative solution of the saddle point linear system. While it is motivated by spectral
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
PRECONDITIONERS FOR THE TIME-HARMONIC MAXWELL EQUATIONS 3
equivalence properties similar to those in [16] and by augmentation considerations, the
actual preconditioners are augmentation-free and parameter-free. Furthermore, convergence
of iterative solvers does not depend on a Schur complement. We show several equivalence
properties of the matrices, and present spectral bounds based on the stability constants of the
dierential operators.
Each iteration of our scheme requires solving for A + M, where > 0 is given. For
such systems solution techniques with linear complexity are available; see [1, 12, 15, 20] and
references therein. We show that the spectral distribution of the preconditioned matrices is
favorable for Krylov subspace solvers in terms of clustering of eigenvalues. We also derive
explicit expressions for the eigenvectors in terms of the null vectors of the discrete operators
A and B, making the convergence analysis complete.
Our numerical results indicate that the proposed technique scales extremely well with the
mesh size, both on uniformly and locally rened meshes. In this paper we only focus on the
performance of the outer solver, and do not consider computational issues related to how to
solve the inner iterations associated with (implicit) inversion of the preconditioner.
The remainder of the paper is structured as follows. In Section 2 we present the mixed nite
element formulation, make some necessary denitions, and discuss the algebraic properties of
the discrete operators. In Sections 3 and 4 we discuss spectral equivalence and augmentation.
In Section 5 we introduce and analyze the proposed preconditioning technique. In Section 6
we provide numerical examples that conrm the analysis and demonstrate the scalability of
our approach. Finally, in Section 7 we draw some conclusions.
2. MIXED FINITE ELEMENT FORMULATION
In this section we provide details on the nite element formulation leading to the saddle point
system (2).
2.1. Discretization
To discretize (1), we consider conforming and shape-regular partitions T
h
of into
tetrahedra K. We denote the diameter of the tetrahedron K by h
K
for all K T
h
and
dene h = max
KT
h
h
K
. Let {

(K) be the space of polynomials of degree on K and let

^

(K) be the space of Nedelec vector polynomials of the rst kind [17, 18]. The index is
chosen so that {
1
(K)
3
^

(K) {

(K)
3
. For 1, the nite element spaces for the
approximation of the electric eld and the multiplier are taken as
V
h
= v
h
H
0
(curl) [ v
h|K
^

(K), K T
h
,
Q
h
= q
h
H
1
0
() [ q
h|K
{

(K), K T
h
.
Here we use the Sobolev space
H
0
(curl) =

v L
2
()
3
: v L
2
()
3
, v n = 0 on

.
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
4 C. GREIF AND D. SCH

OTZAU
We consider the following nite element formulation: nd (u
h
, p
h
) V
h
Q
h
such that

(u
h
) (v
h
) dx k
2

u
h
v
h
dx +

v
h
p
h
dx =

f v
h
dx,

u
h
q
h
dx = 0
(3)
for all (v
h
, q
h
) V
h
Q
h
.
To transform (3) into matrix form, let '
j
`
n
j=1
and '
i
`
m
i=1
be standard nite element bases
for the spaces V
h
and Q
h
respectively:
V
h
= span'
j
`
n
j=1
, Q
h
= span'
i
`
m
i=1
. (4)
Dene
A
i,j
=

(
j
) (
i
) dx, 1 i, j n,
M
i,j
=

j

i
dx, 1 i, j n,
B
i,j
=

j

i
dx, 1 i m, 1 j n,
and let A R
nn
, M R
nn
, and B R
mn
be the corresponding matrices. Let us also
dene the scalar Laplace matrix on Q
h
as L = (L
i,j
)
m
i,j=1
R
mm
, where
L
i,j
=

j

i
dx. (5)
We further introduce the load vector g R
n
by setting
g
i
=

f
i
dx, 1 i n,
where f is the source term in (1). We identify nite element functions u
h
V
h
or p
h
Q
h
with
their coecient vectors u = (u
1
, . . . , u
n
)
T
R
n
and p = (p
1
, . . . , p
m
)
T
R
m
, with respect to
the bases (4). The nite element solution of (3) can now be computed by solving the saddle
point linear system (2).
2.2. Properties of the discrete operators
Let us now present a few key properties of the operators, using the well-known discrete
Helmholtz decomposition for Nedelec elements. To that end, note that Q
h
V
h
, and let
us introduce the matrix C R
nm
by setting

j
=
n

i=1
C
i,j

i
, j = 1, . . . , m.
For a function q
h
Q
h
given by q
h
=

m
j=1
q
j

j
, we then have
q
h
=
n

i=1
m

j=1
C
i,j
q
j

i
,
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
PRECONDITIONERS FOR THE TIME-HARMONIC MAXWELL EQUATIONS 5
so that for q = (q
1
, . . . , q
m
)
T
, we have that u = Cq is the coecient vector of u
h
= q
h
in the
basis '
i
`
n
i=1
.
We shall denote by ', ` the standard Euclidean inner product in R
n
or R
m
, and by null()
the null space of a matrix. For a given positive (semi)denite matrix W and a vector x, we
dene the (semi)norm
[x[
W
=

'Wx, x`.
Proposition 2.1. The following relations hold:
(i) R
n
= null(A) null(B).
(ii) For any u null(A) there is a unique q R
m
such that u = Cq.
(iii) 'Mu, Cq` = 'Bu, q` for u R
n
and q R
m
.
(iv) 'MCp, Cq` = 'Lp, q` for p, q R
m
.
(v) Let u null(A) with u = Cp. Then [u[
M
= [p[
L
.
Proof. The rst two relations readily follow from the discrete Helmholtz decomposition, see
for example [17, Section 7.2.1]. If u
h
and q
h
are the nite element functions associated with
the vectors u and q, then we have
'Mu, Cq` =

u
h
q
h
dx = 'Bu, q`,
which shows (iii). Relation (iv) follows similarly, and (v) follows from (iv). 2
Let us further show a few connections of C to the other matrices.
Proposition 2.2. The following relations hold:
(i) AC = 0.
(ii) BC = L.
(iii) MC = B
T
.
(iv) If the datum f is divergence-free, then C
T
g = 0.
Proof. The rst assertion is obvious since the null space of A is equal to the range of C, by
Proposition 2.1. The dening properties of B, C and L yield, for 1 i, j m,
(BC)
i,j
=
n

k=1
B
i,k
C
k,j
=

k=1
C
k,j

i
dx =

j

i
dx = L
i,j
.
This shows identity (ii). The third one follows similarly. Finally, to see (iv), note that for
1 j m, using integration by parts and the divergence-free condition, we obtain
(C
T
g)
j
=
n

i=1
C
i,j
g
i
=

f
j
dx =

( f)
j
dx = 0.
This completes the proof. 2
An orthogonality property with respect to the inner product 'M, ` is obtained as follows.
Let u
A
null(A) and u
B
null(B). Setting u
A
= Cq, we have
'Mu
A
, u
B
` = 'Mu
B
, Cq` = 'Bu
B
, q` = 0, (6)
by relation (iii) in Proposition 2.1. Consequently, we also have the following result.
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
6 C. GREIF AND D. SCH

OTZAU
Proposition 2.3. Let u = u
A
+ u
B
with u
A
null(A) and u
B
null(B). Then we have
[u[
2
M
= [u
A
[
2
M
+[u
B
[
2
M
.
Let us now present stability properties of the matrices A and B. First, by the Cauchy-
Schwarz inequality, we obviously have
['Au, v`[ [u[
A
[v[
A
, u, v R
n
.
A similar continuity property holds for B:
['Bv, q`[ [v[
M
[q[
L
, v R
n
, q R
m
. (7)
Secondly, the matrix A is positive denite on null(B) and
'Au, u`

[u[
2
A
+[u[
2
M

, u null(B), (8)
with a stability constant which is independent of the mesh size and only depends on the
shape regularity of the mesh and the approximation order [13, Theorem 4.7]. Note that, since
'Au, u` = [u[
2
A
, we must have 0 < < 1 and then also
[u[
2
A
[u[
2
M
, u null(B), (9)
with
=

1
. (10)
Finally, the matrix B satises the discrete inf-sup condition (see [17, p. 179] or [13, p. 319])
inf
0=qR
m
sup
0=vnull(A)
'Bv, q`
[v[
M
[q[
L
1. (11)
The above stated properties and the theory of mixed nite element methods [3] ensure that
(3) is well-posed and the saddle point system is uniquely solvable (provided that the mesh size
is suciently small). Moreover, it can been shown that asymptotically the method is optimally
convergent in the mesh size; see [17, Chapter 7].
3. SPECTRAL EQUIVALENCE PROPERTIES
Consider the augmented matrix
A
L
= A+B
T
L
1
B, (12)
where L is the scalar Laplacian dened in (5). The spectral equivalence properties derived
below motivate the preconditioners presented in Section 5.
Applying the discrete Helmholtz decomposition in Proposition 2.1, we have the following
result.
Lemma 3.1. Let u = u
A
+u
B
with u
A
null(A) and u
B
null(B). Then
[Bu[
L
1 = [u
A
[
M
.
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
PRECONDITIONERS FOR THE TIME-HARMONIC MAXWELL EQUATIONS 7
Proof. From Proposition 2.1, we have u
A
= Cp for a vector p R
m
. Using the identity BC = L
in Proposition 2.2, we obtain
[Bu[
2
L
1 = 'L
1
Bu, Bu` = 'L
1
Bu
A
, Bu
A
` = 'L
1
BCp, BCp` = 'Lp, p` = [p[
2
L
.
Since [p[
L
= [u
A
[
M
, the result follows. 2
As an immediate consequence of Lemma 3.1, we conclude that B
T
L
1
B and M are spectrally
equivalent on the null space of A.
Corollary 3.2. For any u in the null space of A the following relation holds:
'B
T
L
1
Bu, u` = 'Mu, u`.
Theorem 3.3. The matrices A
L
and A+M are spectrally equivalent:

'A
L
u, u`
'(A+M)u, u`
1,
for any u R
n
, where 0 < < 1 is the coercivity constant in (8).
By noticing that
'(A+M)u, u` = [u[
2
A
+[u[
2
M
,
the proof of Theorem 3.3 is readily obtained from the bounds in the subsequent lemma. We
note that a similar result can be found in [16, Theorem 3.1].
Lemma 3.4. The following relations hold:
(i) ['A
L
u, v`[

[u[
2
A
+[u[
2
M

1/2

[v[
2
A
+[v[
2
M

1/2
for u, v R
n
.
(ii) 'A
L
u, u`

[u[
2
A
+[u[
2
M

for u R
n
.
In (ii) 0 < < 1 is the coercivity constant given in (8).
Proof. By Proposition 2.1, we may decompose u and v into u = u
A
+ u
B
and v = v
A
+ v
B
with u
A
, v
A
null(A) and u
B
, v
B
null(B). Furthermore, there are vectors p and q in R
m
such that u
A
= Cp and v
A
= Cq.
Let us show the rst assertion. By the Cauchy-Schwarz inequality,
['Au, v`[ [u[
A
[v[
A
.
Similarly, the Cauchy-Schwarz inequality, Lemma 3.1 and the orthogonality in Proposition 2.3
yield
['B
T
L
1
Bu, v`[ = ['L
1
Bu, Bv`[ [Bu[
L
1[Bv[
L
1 = [u
A
[
M
[v
A
[
M
[u[
M
[v[
M
.
The rst assertion readily follows from summing the last two inequalities and applying again
the Cauchy-Schwarz inequality. To show the result in (ii), note that the stability property (8)
of the matrix A yields:
'Au, u` = 'Au
B
, u
B
`

[u
B
[
2
A
+[u
B
[
2
M

.
From Lemma 3.1,
'L
1
Bu, Bu` = [Bu[
2
L
1 = [u
A
[
2
M
,
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
8 C. GREIF AND D. SCH

OTZAU
and since 0 < < 1 we have
'A
L
u, u`

[u
A
[
2
M
+[u
B
[
2
A
+[u
B
[
2
M

.
By the orthogonality relation in Proposition 2.3 we have [u[
2
M
= [u
A
[
2
M
+ [u
B
[
2
M
, from which
relation (ii) follows. 2
We end this section by pointing out a connection between L and the Schur complement
associated with A
L
, S = BA
1
L
B
T
. The matrices S and L are spectrally equivalent; we have

'Sp, p`
'Lp, p`

1
,
for any p R
m
. Here, is the coercivity constant from (8). The proof is a consequence
of Lemma 3.4, the inf-sup condition in (11), and standard arguments for mixed nite
element methods [3]. We also refer the reader to [16, Theorem 3.3]. As a consequence,
the preconditioners we propose in the sequel are closely related to block preconditioners
that rely on forming approximations of the Schur complement. Such techniques have been
successfully used in a variety of applications, notably for the discretized Stokes and Navier-
Stokes equations [6, 7].
4. AUGMENTATION WITH THE SCALAR LAPLACIAN
We now turn our attention to the linear system and consider augmentation with the Laplacian
as a starting point. We will assume that Ak
2
M is nonsingular; this can always be achieved
by choosing the mesh size suciently small [17, Corollary 7.3].
Consider the matrix of (2):
/ =

Ak
2
M B
T
B 0

, (13)
and dene the symmetric positive denite block diagonal matrix
/
L
=

A
L
k
2
M 0
0 L

. (14)
We stress that /
L
will not be the preconditioner that we eventually use; it is only introduced
to lay the theoretical basis and motivation for the preconditioning approach that we propose
in Section 5. Note that A
L
k
2
M is symmetric positive denite for k suciently small.
Theorem 4.1. The matrix /
1
L
/ has two distinct eigenvalues, given by

+
= 1 ;

=
1
1 k
2
,
with algebraic multiplicities n and m respectively.
Proof. The matrix /
1
L
/ has a complete set of linearly independent eigenvectors that span
R
n+m
. The corresponding eigenvalue problem is

Ak
2
M B
T
B 0

v
q

Ak
2
M +B
T
L
1
B 0
0 L

v
q

.
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
PRECONDITIONERS FOR THE TIME-HARMONIC MAXWELL EQUATIONS 9
From the nonsingularity of /
1
L
/ it follows that = 0. Substituting q =
1
L
1
Bv, we obtain
for the rst block row
(Ak
2
M)v +B
T
L
1
Bv =
2
(Ak
2
M +B
T
L
1
B)v. (15)
By inspection it is straightforward to see that any vector v R
n
satises (15) with = 1,
and thus the latter is an eigenvalue of /
1
L
/, with eigenvectors of the form (v, L
1
Bv), where
v = 0. We claim that the eigenvalue = 1 has algebraic multiplicity n. (That is, there is no
other eigenvector associated with = 1 in addition to the above set.) Indeed, if the vectors
(v
i
, L
1
Bv
i
)
n+r
i=1
, with r 0, are linearly independent then necessarily v
i

n+r
i=1
are also
linearly independent, and the latter is impossible unless r = 0.
Let us now point out a specic set of eigenvectors for = 1, and derive expressions for the
remaining eigenpairs. According to Proposition 2.1 we can decompose v = v
A
+ v
B
, where
v
A
null(A) and v
B
null(B). We now show that if an eigenvector (v,
1
L
1
Bv) =
(v
A
+ v
B
,
1
L
1
Bv
A
) has a nonzero v
B
component, then its associated eigenvalue must
necessarily be = 1. Noting that by (6)
'M(v
A
+v
B
), v
B
` = [v
B
[
2
M
,
after taking inner products of (15) with v
B
and dividing by we get
( 1)

[v
B
[
2
A
k
2
[v
B
[
2
M

= 0.
Since the symmetric matrix Ak
2
M is nonsingular, it follows that for v
B
= 0 we must have
[v
B
[
2
A
k
2
[v
B
[
2
M
= '(Ak
2
M)v
B
, v
B
` = 0, and hence = 1.
Next, we argue that at least 2m of the vectors v must have a nonzero v
A
component.
Let us prove this by showing that assuming otherwise leads to a contradiction. Suppose the
eigenvectors are given by (v,
1
L
1
Bv) for a set of n + m choices of v. If the assumption
does not hold, then more than n m eigenvectors satisfy v = v
B
, and must be of the form
(v
B
, 0). But since the null space of B is of rank nm, there cannot be more than this number
of linearly independent vectors (v
B
, 0).
Since at least 2m of the eigenvectors satisfy v
A
= 0, and since the multiplicity of = 1
is n, it follows that at least m of the eigenvectors associated with = 1 satisfy v
A
= 0. Thus,
consider m such vectors, v = v
A
+v
B
with v
A
= 0. Then (15) reads

Av
B
k
2
M(v
A
+v
B
)

+B
T
L
1
Bv
A
=
2

Av
B
k
2
M(v
A
+v
B
) +B
T
L
1
Bv
A

.
Taking inner products with the vectors v
A
and noting that by (6)
'M(v
A
+v
B
), v
A
` = [v
A
[
2
M
,
and by Corollary 3.2 we have
'B
T
L
1
Bv
A
, v
A
` = 'Mv
A
, v
A
`,
it follows that
(
2
)k
2
[v
A
[
2
M
+ (
2
1)[v
A
[
2
M
= 0.
Hence we have
(1 k
2
)
2
+k
2
1 = 0, (16)
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
10 C. GREIF AND D. SCH

OTZAU
from which it follows that
+
= 1 and

=
1
1k
2
. We have thus shown that

is the only
possible eigenvalue that is not equal to 1, and its algebraic multiplicity must be equal to m.
This completes the proof. 2
The proof of Theorem 4.1 in fact shows that the eigenspace of /
1
L
/ can be expressed in
terms of the null vectors of A and B, as follows.
Corollary 4.2. Let v
i

m
i=1
be a basis for the null space of A and z
i

nm
i=1
a basis for
the null space of B. Then (v
i
, L
1
Bv
i
)
m
i=1
and (z
i
, 0)
nm
i=1
are n linearly independent
eigenvectors associated with the eigenvalue
+
. The vectors (v
i
, (1 k
2
)L
1
Bv
i
)
m
i=1
are m
linearly independent eigenvectors associated with the eigenvalue

. Grouped together, those

eigenvectors form a complete eigenspace that spans R
n+m
.
From Theorem 4.1 it follows that if MINRES were to be used for solving (2), with /
L
as
a preconditioner, then convergence would require merely two iterations, if roundo errors are
ignored. However, forming A
L
may be too computationally costly. We mention also that the
results in Theorem 5.2 can be extended to general algebraic settings, i.e. not necessarily to the
Maxwell operator, as we show in [10].
5. AUGMENTATION-FREE PRECONDITIONERS
Dene
{
M
= A+M, (17)
where = 1 k
2
. For the saddle point system (2) we consider preconditioning with
{
M,L
=

{
M
0
0 L

. (18)
Throughout, we will assume that preconditioned MINRES for the saddle point system is
used. A crucial factor in the speed of convergence of this method is the distribution of the
eigenvalues; strong clustering yields fast convergence [9, Section 3.1]. The choice of {
M
and
{
M,L
is motivated by the spectral equivalence results given in Theorem 3.3 and the eigenvalue
distribution observed in Theorem 4.1, which allow us to observe that {
M
A
L
k
2
M and
{
M,L
/
L
. Thus, the overall computational cost of the solution procedure will depend on
the ability to eciently solve linear systems whose associated matrices are A+M and L (or
approximations thereof). For solving the former we refer the reader to [1, 12, 15, 20].
Theorem 5.1. The matrix
{
1
M
(A
L
k
2
M)
has an eigenvalue = 1 of algebraic multiplicity m. The rest of the eigenvalues are bounded
as follows:
k
2
+ 1 k
2
< < 1, (19)
with dened in (10).
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
PRECONDITIONERS FOR THE TIME-HARMONIC MAXWELL EQUATIONS 11
Proof. The corresponding eigenvalue problem is
(Ak
2
M +B
T
L
1
B)v =

A+ (1 k
2
)M

v.
Suppose v = v
A
+v
B
, where v
A
null(A) and v
B
null(B). We then have
Av
B
k
2
M(v
A
+v
B
) +B
T
L
1
Bv
A
=

Av
B
+ (1 k
2
)M(v
A
+v
B
)

.
By linear independence considerations, there are at least m vectors v that satisfy v
A
= 0. For
m such vectors, taking inner products with v
A
and noting that by Corollary 3.2
'B
T
L
1
Bv
A
, v
A
` = [v
A
[
2
M
and that by (6) we have
'M(v
A
+v
B
), v
A
` = 'Mv
A
, v
A
` = [v
A
[
2
M
,
we get
(1 k
2
)[v
A
[
2
M
= (1 k
2
)[v
A
[
2
M
.
It follows that = 1 is an eigenvalue of multiplicity m.
For the rest of the eigenvectors we must have v
B
= 0, and now taking inner products with
v
B
and noting that
'B
T
L
1
Bv
A
, v
B
` = 'L
1
Bv
A
, Bv
B
` = 0,
and that by (6) we have
'M(v
A
+v
B
), v
B
` = 'Mv
B
, v
B
` = [v
B
[
2
M
,
it follows that
(1 )[v
B
[
2
A
=

(1 k
2
) +k
2

[v
B
[
2
M
. (20)
It is impossible to have = 1, since in this case (20) collapses into [v
B
[
M
= 0, which cannot
hold for v
B
= 0. We cannot have > 1 either, since that would imply that in (20) the left
hand side is negative but the right hand side is positive. (Recall that we assume k < 1.) We
conclude that we must have < 1.
From (9) we recall that for any u null(B), [u[
2
A
[u[
2
M
with =

1
> 0. Applying
this to (20) we conclude (1 k
2
) +k
2
(1 ), and since 1 k
2
+ > 0 we obtain (19).
Since can be either equal to 1 or satisfy (19), but not simultaneously both, the algebraic
multiplicities follow. 2
Theorem 5.2. Let / be the saddle point matrix (13). Then
+
= 1 and

=
1
1k
2
are
eigenvalues of the preconditioned matrix {
1
M,L
/, each with algebraic multiplicity m. The rest
of the eigenvalues satisfy the bound (19).
Proof. The eigenvalue problem for {
1
M,L
/ is

Ak
2
M B
T
B 0

v
q

A+ (1 k
2
)M 0
0 L

v
q

.
Setting q =
1
L
1
Bv and multiplying the resulting equation for v by , we have

(
2
)A+ ((1 k
2
)
2
+k
2
)M

v = B
T
L
1
Bv.
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
12 C. GREIF AND D. SCH

OTZAU
The rest of the proof follows by taking the same steps taken in the proof of Theorem 5.1.
We get m equations of the form
(1 k
2
)
2
+k
2
1 = 0,
from which
+
and

are obtained. Note that this quadratic equation is identical to equation

(16) for the eigenvalues of /
1
L
/, cf. Theorem 4.1. This reinforces that {
M,L
is an eective
sparse approximation of /
L
. Obtaining the bound (19) is done in a way identical to the last
part of the proof of Theorem 5.1. 2
For k = 0 the result of Theorem 5.2 simplies as follows.
Corollary 5.3. For the preconditioned matrix {
1
M,L
/ with k = 0, there are eigenvalues

= 1, with algebraic multiplicity m each. The rest of the eigenvalues satisfy < < 1.
6. NUMERICAL EXPERIMENTS
Our numerical experiments were performed using Matlab; for generating the meshes we used
the PDE toolbox. We implemented the two-dimensional version of the time-harmonic Maxwell
equations. The lowest order elements were used, i.e. = 1. The solutions of the preconditioned
systems in each iteration were computed exactly.
6.1. A smooth domain with a quasi-uniform grid
In this example the domain is the unit square. Uniformly rened meshes were constructed.
The rened meshes are obtained from the original one by subsequently dividing each triangle
into four congruent ones. The number of elements and matrix sizes are given in Table I.
Table I. Number of elements (Nel) and the size of the linear systems (n +m) for seven grids used in
Example 6.1.
Grid Nel n +m
G1 64 113
G2 256 481
G3 1024 1985
G4 4096 8065
G5 16384 32513
G6 65536 130561
G7 262144 523265
First, we set the right hand side function so that the exact solution is given by
u(x, y) =

u
1
(x, y)
u
2
(x, y)

1 y
2
1 x
2

and p 0. The datum f in this case is divergence-free. We ran MINRES with the
preconditioner {
M,L
. The counts of the outer iterations are given in Table II. The inner
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
PRECONDITIONERS FOR THE TIME-HARMONIC MAXWELL EQUATIONS 13
iterations were solved by the conjugate gradient method, preconditioned with incomplete
Cholesky factorization, using a tight convergence tolerance. As expected, the outer solver
scales extremely well with hardly any sensitivity to the mesh size and the wave number.
Table II. Iteration counts for Example 6.1 with a divergence-free right hand side, for various meshes
and values of k, using MINRES for solving the saddle point system with the preconditioner P
M,L
.
The outer iteration was stopped once the initial relative residual was reduced by a factor of 10
10
.
Grid k = 0 k =
1
8
k =
1
4
k =
1
2
G1 5 5 5 5
G2 5 5 5 5
G3 5 5 5 5
G4 6 6 5 6
G5 6 6 6 6
G6 6 6 6 6
G7 6 6 6 6
Table III. Iteration counts for Example 6.1 with a right hand side that is not divergence-free,
for various meshes and values of k, using MINRES for solving the saddle point system with the
preconditioner P
M,L
. The outer iteration was stopped once the initial relative residual was reduced
by a factor of 10
10
.
Grid k = 0 k =
1
8
k =
1
4
k =
1
2
G1 5 5 5 5
G2 6 6 6 6
G3 6 6 6 6
G4 6 6 6 7
G5 7 7 7 7
G6 7 7 7 7
G7 7 7 7 7
We also ran the saddle point solver on an example with a right hand side function that was
not divergence-free. We took the same u as above, and p = (1 x
2
)(1 y
2
). The iteration
counts are given in Table III. As before, the solver scales very well. Figure 1 depicts the
eigenvalues of the preconditioned matrix {
1
M,L
/ for grid G2 with k =
1
4
. This linear system
has 481 degrees of freedom, with n = 368 and m = 113. As is expected from Theorem 5.2,
the m negative eigenvalues of the matrix are equal to
1
1k
2
=
16
15
= 1.0666 . . . , and for
the positive ones, m of them are equal to 1 and the remaining n m eigenvalues are bounded
away from 0 and below 1. In our computations we observed strong clustering beyond what can
be concluded from Theorem 5.2. Three of the positive eigenvalues are between 0.7 and 0.9,
with the smallest equal to 0.706. . . , and four additional ones are between 0.9 and 0.95. The
remaining 361 eigenvalues are all between 0.95 and 1, with 113 of them identically equal to 1,
again as is known by the same theorem. This clustering eect explains the fast convergence of
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
14 C. GREIF AND D. SCH

OTZAU
the preconditioned iterative solver.
Figure 1. Plot of the eigenvalues of the preconditioned matrix P
1
M,L
K, for k =
1
4
, for grid G2 in
Example 6.1.
0 50 100 150 200 250 300 350 400 450 500
1.5
1
0.5
0
0.5
1
1.5
6.2. An L-shaped domain with locally rened grids
In this example we consider an L-shaped domain, as depicted in Figure 6.2. The meshes were
locally rened at the nonconvex corner at the origin; the number of elements and sizes are
given in Table IV. Four of the ve grids that were used are depicted in Figure 2. We set up the
problem so that the right hand side function is equal to 1 throughout the domain. As in the
previous example, we applied MINRES, preconditioned by {
M,L
, to the saddle point system.
Table V demonstrates the scalability of the solvers: the outer iteration counts do not seem to
be sensitive to changes in the mesh size.
Table IV. Number of elements (Nel) and the size of the linear systems (n + m) for ve grids used in
Example 6.2.
Grid Nel n +m
L1 258 451
L2 458 813
L3 1403 2608
L4 5164 9927
L5 19339 37882
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
PRECONDITIONERS FOR THE TIME-HARMONIC MAXWELL EQUATIONS 15
Figure 2. Grids L1 through L4 for Example 6.2.
1 0.5 0 0.5 1
1
0.5
0
0.5
1
Lshaped grid L1
1 0.5 0 0.5 1
1
0.5
0
0.5
1
Lshaped grid L2
1 0.5 0 0.5 1
1
0.5
0
0.5
1
Lshaped grid L3
1 0.5 0 0.5 1
1
0.5
0
0.5
1
Lshaped grid L4
Table V. Iteration counts for Example 6.2 with various meshes and values of k, using MINRES for
solving the saddle point system with the preconditioner P
M,L
. The outer iteration was stopped once
the initial relative residual was reduced by a factor of 10
10
.
Grid k = 0 k =
1
8
k =
1
4
k =
1
2
L1 5 5 5 5
L2 5 5 5 5
L3 5 5 5 5
L4 5 5 5 5
L5 4 4 4 4
7. CONCLUSIONS
We have introduced a new augmentation-free and Schur complement-free block diagonal
preconditioning approach for solving the discretized mixed formulation of the time-harmonic
Maxwell equations. We have presented a complete spectral analysis, and have shown that the
outer iteration counts are hardly sensitive to changes in the mesh size or in small values of the
wave number.
We have limited the discussion in this paper to the convergence of the outer iterations, relying
on the assumption that robust solution techniques exist for solving A + M. Future research
will focus on further computational aspects of our solution technique, and we will explore
using ecient inner solvers. Finally, we will investigate whether similar preconditioners can be
applied to problems in three dimensions and problems with variable coecients.
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls
16 C. GREIF AND D. SCH

OTZAU
REFERENCES
1. D.N. Arnold, R.S. Falk, and R. Winther. Multigrid in H(div) and H(curl). Numer. Math., 85:197217,
2000.
2. M. Benzi, G.H. Golub, and J. Liesen. Numerical solution of saddle point problems. Acta Numerica,
14:1137, 2005.
3. F. Brezzi and M. Fortin. Mixed and hybrid nite element methods. In Springer Series in Computational
Mathematics, volume 15. SpringerVerlag, New York, 1991.
4. Z. Chen, Q. Du, and J. Zou. Finite element methods with matching and nonmatching meshes for Maxwell
equations with discontinuous coecients. SIAM J. Numer. Anal., 37:15421570, 2000.
5. L. Demkowicz and L. Vardapetyan. Modeling of electromagnetic absorption/scattering problems using
hpadaptive nite elements. Comput. Methods Appl. Mech. Engrg., 152:103124, 1998.
6. H.C. Elman, D.J. Silvester, and A.J. Wathen. Iterative methods for problems in computational uid
dynamics. In R.H. Chan, T.F. Chan, and G.H. Golub, editors, Iterative Methods in Scientic Computing,
pages 271327. Springer-Verlag, Singapore, 1997.
7. H.C. Elman, D.J. Silvester, and A.J. Wathen. Finite Elements and Fast Iterative Solvers. Oxford
University Press, 2005.
8. J. Gopalakrishnan, J.E. Pasciak, and L.F. Demkowicz. Analysis of a multigrid algorithm for time harmonic
Maxwell equations. SIAM J. Numer. Anal., 42(1):90108, 2004.
9. A. Greenbaum. Iterative Methods for Solving Linear Systems. SIAM, 1997.
10. C. Greif and D. Schotzau. Preconditioners for saddle point linear systems with highly singular (1,1)
blocks. ETNA, Special Volume on Saddle Point Problems, 22:114121, 2006.
11. E. Haber, U.M. Ascher, D. Aruliah, and D. Oldenburg. Fast simulation of 3D electromagnetic problems
using potentials. J. Comput. Phys., 163:150171, 2000.
12. R. Hiptmair. Multigrid method for Maxwells equations. SIAM J. Numer. Anal., 36:204225, 1998.
13. R. Hiptmair. Finite elements in computational electromagnetism. Acta Numerica, 11:237339, 2002.
14. P. Houston, I. Perugia, A. Schneebeli, and D. Schotzau. Mixed discontinuous Galerkin approximation of
the Maxwell operator: the indenite case. Model. Math. Anal. Numer., 39:727754, 2005.
15. J.J. Hu, R.S. Tuminaro, P.B. Bochev, C.J. Garasi, and A.C. Robinson. Toward an h-independent algebraic
multigrid method for Maxwells equations. SIAM J. Sci. Comput., 27(5):16691688, 2006.
16. Q. Hu and J. Zou. Substructuring preconditioners for saddle-point problems arising from Maxwells
equations in three dimensions. Math. Comp., 73(245):3561, 2004.
17. P. Monk. Finite element methods for Maxwells equations. Oxford University Press, New York, 2003.
18. J.C. Nedelec. Mixed nite elements in R
3
. Numer. Math., 35:315341, 1980.
19. I. Perugia, D. Schotzau, and P. Monk. Stabilized interior penalty methods for the time-harmonic Maxwell
equations. Comput. Methods Appl. Mech. Engrg., 191:46754697, 2002.
20. S. Reitzinger and J. Schoberl. An algebraic multigrid method for nite element discretizations with edge
elements. Numer. Linear Algebra Appl., 9:223238, 2002.
Copyright c 2006 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2006; 00:115
Prepared using nlaauth.cls