Coupled FETI/BETI solvers for nonlinear potential problems in (un ...

Coupled FETI/BETI solvers for nonlinear
potential problems in (un)bounded domains
Ulrich Langer 1,2 and Clemens Pechstein 2
1 Institute of Computational Mathematics, Johannes Kepler University,
Altenberger Str. 69, 4040 Linz, Austria ulanger@numa.uni-linz.ac.at
2 Special Research Program SFB F013, Johannes Kepler University, Altenberger
Str. 69, 4040 Linz, Austria clemens.pechstein@numa.uni-linz.ac.at
Summary. In nonlinear electromagnetic field computations, one is not only faced
with large jumps of material coefficients across material interfaces but also with
high variation in these coefficients even inside homogeneous materials due to the
nonlinearity. The radiation condition can conveniently be taken into account by a
coupled boundary integral and domain integral variational formulation. The coupled
finite and boundary element discretization leads to large-scale nonlinear algebraic
systems. In this paper we propose special inexact Newton methods where the Jacobi
systems arising in every step of the Newton method are solved by a special preconditioned
finite and boundary element tearing and interconnecting solver. The numerical
experiments show that the preconditioner proposed in the paper can handle
large jumps in the coefficients across the material interfaces as well as high variation
in these coefficients on the subdomains. Furthermore, the convergence does not
deteriorate if many inner subdomains touch the unbounded exterior subdomain.
Key words: Nonlinear problems, electromagnetic field computations, domain decomposition,
FEM, BEM, Newton’s methods, unbounded computational domains
1 Introduction
Domain decomposition (DD) methods of iterative substructuring type, such as the
widely used classical finite element tearing and interconnecting (FETI) methods [3],
dual-primal FETI (FETI-DP) methods [4] and balanced domain decomposition by
constraints (BDDC) [5], are powerful methods for large-scale field computations on
parallel computers. In the classical FETI method, the finite element subspaces are
treated separately on each subdomain including its boundary. The global continuity
across the interfaces is enforced by Lagrange multipliers, which leads to a saddle
point problem that can be solved iteratively via its dual problem. From the Lagrange
multipliers, the solution can easily be computed. The iteration process is nothing
but a preconditioned conjugate gradient (PCG) subspace iteration.
To obtain a fast method, a careful choice of the preconditioner is essential. For
the case that the coefficients of the underlying elliptic partial differential equation

2 Ulrich Langer and Clemens Pechstein
(PDE) are constant in each subdomain, quasi-optimal preconditioners are available.
It is proved that the condition number grows proportionally to (1+log(H/h)) 2 , where
h is the mesh size and H the average diameter of the subdomains. Moreover, the
preconditioners are robust with respect to jumps in the coefficients across subdomain
interfaces [7, 13]. The PCG subspace iteration involves the use of standard Dirichlet
and Neumann solvers.
To summarize, the main success of FETI, FETI-DP and BDDC methods is
certainly due to their rather general structure, wide applicability, moderate complexity,
robustness, and finally their scalability with respect to parallel computing.
For a comprehensive introduction to domain decomposition methods, FETI and
FETI-DP methods, we refer to the monograph by Toselli and Widlund [18].
Recently, Langer and Steinbach have introduced the boundary element tearing
and interconnecting (BETI) methods [11] as a boundary element counterpart of
the FETI methods, as well as the coupled FETI/BETI methods [12]. The BETI
method uses boundary-element-based analogs of the FETI operators. Due to spectral
arguments, all the properties of FETI methods mentioned above remain valid for
BETI methods. Furthermore, inexact and data-sparse techniques are available, cf.
[9].
Coupling boundary element and finite element discretizations, one can benefit
from the advantages of both discretization techniques. For instance, in electromagnetics,
source terms and nonlinearities can be treated more efficiently by the finite
element method (FEM) than by the boundary element method (BEM), whereas unbounded
domains, moving parts and air regions can efficiently be handled by BEM.
We refer to [2] for the symmetric coupling of finite and boundary elements, and to
[6, 8] for using this coupling technique to construct domain decomposition solvers.
In this contribution, we use coupled FETI/BETI methods to solve nonlinear
potential problems as they appear in nonlinear magnetostatics in two dimensions,
div(ν(x, |∇u(x)|)∇u(x)) = f(x) for x ∈ Ω . (1)
Additionally we may have suitable transmission conditions, Dirichlet boundary conditions
and/or–in case Ω is unbounded–a radiation condition.
In Section 2 we give a review of the coupled FETI/BETI methods for the case
that the coefficient ν is constant in the subdomains Ω i, considering a bounded
domain with Dirichlet boundary conditions.
Solving the nonlinear problem (1) in a bounded domain Ω with FETI/BETI
methods has been investigated by the present authors in [10]. We give an outline of
the main issues in Section 3. Applying Newton’s method to the global formulation,
the spectrum of the Jacobi matrices in the nonlinear subdomains may show high
variation, especially if there are singularities in the solution. We propose a special
preconditioner to overcome these difficulties and show its good numerical behavior
in a typical magnetostatic model problem.
Section 4 is devoted to unbounded domains. We show how the BETI operators
change and give some analytic results. Up to now, we can only prove a suboptimal
condition number estimate whereas the performance of the numerical experiments
is far more promising.

Coupled FETI/BETI solvers – nonlinear, unbounded 3
2 Coupled FETI/BETI Methods
Let Ω ⊂ R d (where d = 2, 3) be a bounded, connected domain with a Lipschitz
boundary Γ and the outward unit normal vector n. We assume that Ω is decomposed
into p non-overlapping simply-connected Lipschitz domains Ω i, i. e. Ω = S p
i=1
Ωi. It
is assumed that the diameters H i = diam Ω i are all of comparable size and bounded
by the maximal diameter H. We define the local boundaries Γ i = ∂Ω i and the
interfaces Γ ij = Ω i ∩ Ω j, and denote the outward unit normal vector on Γ i by
n i. In the following, we consider the Poisson problem with homogeneous Dirichlet
boundary conditions and piecewise constant coefficients, to find u satisfying
−div(α i∇u) = f in Ω i ,
u = 0 on Γ ,
α i
∂u
∂n i
+ α j
∂u
∂n j
= 0 on Γ ij ,
(2)
with α i = const. The generalization to inhomogeneous and mixed boundary conditions
is straightforward.
2.1 Dirichlet-to-Neumann maps
The solution u i of the local subproblem
−div(α i∇u i) = 0 in Ω i , u i = g i on Γ i , (3)
∂u
defines the Steklov-Poincaré operator S i g i := α i i ∂n i
, mapping the Dirichlet trace g i
to the corresponding Neumann trace. The contribution of a right hand side f i to the
∂v
Neumann trace of the solution is described by the Newton potential N i f i = −α i i ∂n i
,
where v i solves
−div(α i∇v i) = f i in Ω i , v i = 0 on Γ i . (4)
These operators can be approximated by the FEM. Fixing a triangulation T i, h of
Ω i with a mesh size h, we denote by K i the FEM stiffness matrix and partition it
according to boundary unknowns (subscript Γ) and inner unknowns (subscript I).
The Schur complement
S FEM
i, h := K II, i − K ΓI, i K −1
II, i K IΓ, i (5)
is a symmetric and stable approximation of S i, and
N FEM
i, h f i
:= f Γ, i
− K ΓI, i K −1
II, i f I i
(6)
is a stable approximation of the Newton potential N i, cf. [17].
On the other hand, one can approximate the Steklov-Poincaré operator by means
of the BEM, cf. [6, 17]. The solution of (3) satisfies the Caldéron equation
„ « „ αi
« „ «
gi
= 2
I − K i V i gi
α
t i D i i 2
I + K i
′ (7)
t i
where t i is the co-normal derivative and V i, K i, K ′ i, D i are the usual boundary
integral operators, the single layer potential operator, the double layer potential
operator, its adjoint, and the hypersingular integral operator, respectively. In three

4 Ulrich Langer and Clemens Pechstein
dimensions the single layer potential operator V i is always elliptic, whereas in two
dimensions, due to the logarithm in the fundamental solution, it is only elliptic if
diam Ω i < 1. This property can always be achieved by a suitable coordinate scaling.
After discretizing and eliminating t i, one obtains the symmetric and stable approximation
Si, BEM
h := D i, h + ` α i
2
Mi, ⊤ h + Ki, ⊤ −1` h
´V
αi
i, h 2
M i, h + K i, h´
, (8)
where V i, h , K i, h , D i, h are the boundary element matrices corresponding to V i, K i,
D i, respectively, and M i, h is a mass matrix.
Note that the two approximations Si, FEM
h and Si, BEM
h are compatible and both
spectrally equivalent to the Galerkin matrices of the exact Steklov-Poincaré operators
S i. The application of Si, FEM
h or Si, BEM
h simply corresponds to the solution of
local Dirichlet problems. For details we refer to [11, 12].
2.2 Tearing and Interconnecting
Introducing separate variables u i on the local subdomains, one can re-enforce the
continuity of the solution u across interfaces Γ ij by the constraints
X p
Bi ui = 0 , (9)
i=1
where the B i are incidence matrices.
Problem (2) can be written as a constraint minimization problem, as well as a
saddle point problem involving Lagrange multipliers. Using the notion of the pseudoinverse
( † ) and a special projection P addressing the kernels of the subproblems, it
is possible to eliminate the primal unknowns u i. Finally, one obtains the discrete
dual FETI/BETI formulation, to find the Lagrange multiplier λ such that
where the FETI/BETI operator F is defined by
F = B
where B = ˆB i˜p
, i=1 SFEM/BEM h
:=
P T F λ = d, (10)
FEM/BEM˜† ˆSh
B ⊤ = X p
˜†
Bi
ˆSFEM/BEM
i=1
i,h
Bi ⊤ , (11)
FEM/BEM˜p
ˆSi, h
. The application of the pseudo-
i=1
inverses [S FEM/BEM
i, h
] † can be realized by the simple solution of regularized local Neumann
problems. Since F is symmetric positive definite on range(P), one can solve
the dual problem (10) by a preconditioned conjugate gradient subspace iteration.
The preconditioner
M −1
S,α = (BD −1
α B ⊤ ) −1 BD α S FEM/BEM
h
D αB ⊤ (BD −1
α B ⊤ ) −1 , (12)
first introduced and fully analyzed by Klawonn and Widlund [7], satisfies the quasioptimal
condition number estimate
κ(PM −1
S,αP T P T FP) ≤ C(1 + log(H/h)) 2 , (13)
independent of the values–and therefore possible jumps–of the coefficients α i. It is
well known that an appropriate norm of the
√iteration error of the conjugate gradient
method will decrease at least by a factor 2` κ−1
´n
√ κ+1
in n steps. The robustness with

Coupled FETI/BETI solvers – nonlinear, unbounded 5
respect to the jumps in the coefficients α i is due to the special diagonal scaling
matrix D α, involving weighted mean values of α i on cross points, interface lines and
interfaces between the subdomains. We further point out that one step of the PCG
subspace iteration is performed by the solution of one local Dirichlet and one local
Neumann problem on each of the subdomains, and the application of the projection
and some scaling matrices, which are both global operations but of a rather small
dimension. This is why these tearing and interconnecting methods are most suitable
for parallelization. For a more detailed description see e. g. [7, 10, 12]
2.3 Varying coefficients
In this subsection, we consider a varying matrix coefficient A i(x) instead of a constant
coefficient α i on some of the subdomains discretized by the FEM. We assume
A i(x) to be constant on the finite elements T ∈ T i, h . In order to determine the
amount of variance, we introduce the spectral variance measure
m SV (A i) := sup x∈Ω i
α i (x)
inf x∈Ωi α i (x) , (14)
where α i(x) and α i
(x) denote the maximal and minimal local eigenvalues of A i(x),
respectively. The application of a preconditioner with Steklov-Poincaré operators
corresponding to constant coefficients leads in the worst case to a condition number
proportional to max i m SV (A i), which is not at all acceptable in magnetostatic
applications. In [10] the present authors have proposed a new preconditioner M c−1
S,A
based on a varying scalar coefficient bα i(x) which can be easily computed from A i(x),
together with a suitable diagonal scaling matrix D bα . If the local anisotropy measure
m anis(A i) := sup x∈Ωi
α i (x)
α i (x)
(15)
is moderate, our new preconditioner works fine, cf. Table 1 and [10].
3 Nonlinear Problems
We now consider the following nonlinear magnetostatic model problem:
−div[ν i(|∇u|)∇u] = f in Ω i ,
u = 0 on Γ ,
ν i(|∇u|) ∂u
∂n i
+ ν j(|∇u|) ∂u
∂n j
= 0 on Γ ij .
Assuming that the functions t ↦→ ν i(t)t : [0, ∞) → [0, ∞) are strongly monotonically
increasing and piecewise C 2 , (16) is uniquely solvable in the weak sense and
the corresponding Newton iteration converges locally at a quadratic rate. For our
computations, we generated these material curves from noisy measurements using
the robust interproximation technique introduced by Pechstein and Jüttler in
[15]. In the linearized problems a varying matrix coefficient ζ i(∇u (k)
h
(x)) appears.
In our numerical experiments, it turns out that typically the anisotropy measure
m anis(ζ i(∇u (k)
h
)) is small, whereas the spectral variance measure m SV(ζ i(∇u (k)
h ))
becomes rather large. As one can observe in Table 1, with our new preconditioner
cM −1
S,A
such linearized Newton-problems can be solved satisfactorily. In order to get
a good initial guess for Newton’s iteration, it is convenient to set up a hierarchy
of nested grids and use coarse grid solutions as initial guesses on finer levels. For
details, see [10].
(16)

6 Ulrich Langer and Clemens Pechstein
d.o.f. Lagr. H/h m anis(ζ) m SV (ζ) Newt. PCG-steps ref
806 408 6.3 12.0 187.4 6 14.0 12
3539 777 12.6 12.0 469.9 4 17.8 14
14357 1515 25.3 12.0 852.4 4 21.3 17
57833 2991 20.6 12.0 1496.4 3 25.3 19
232145 5943 101.2 12.4 2670.9 4 27.8 21
Table 1. Average number of FETI PCG iterations to get a reduction of 10 −8 in
the residual of linear subproblems during Newton’s iteration, with 70 subdomains,
compared to a linear reference problem (ref.) of the same size.
4 Unbounded domains
In this section we allow that one of the subdomains, namely Ω 0, is the unbounded
exterior ext(Γ 0) of its boundary Γ 0, where the interior int(Γ 0) is bounded. Usually,
for a typical domain decomposition, H 0 := diam int(Γ 0) ≫ H i for i ≠ 0. In Ω 0, we
assume that the homogeneous Poisson equation is satisfied together with a suitable
radiation condition, e.g. for d = 3,
−α 0∆u = 0 in Ω 0 , |u(x)| = O(|x| −1 ) for |x| → ∞ , (17)
cf. [14, 16]. In magnetostatic field computations, α 0 could equal 1/µ 0, where µ 0 is
the permeability of vacuum. The Dirichlet-to-Neumann map on Ω 0 reads
S0 ext = D 0 + ` α 0
´ `
2
I − K 0
′ V−1 α0
0 2
I − K 0´
. (18)
This means the difference between the interior and exterior Steklov-Poincaré operator
is just the sign in front of the double layer potential operators K and K ′ . As
a consequence, S0 ext is always a one-to-one mapping, and Ω 0 must be treated as a
non-floating subdomain, cf. [7]. It can be shown that
〈S ext
0 v, v〉 ≃ |v| 2 H 1/2 (Γ 0 ) + 1
H 0
‖v‖ 2 L 2 (Γ 0 ) , (19)
where in two dimensions the coordinates have to be scaled such that H 0 ≃ 1. The
BEM-approximation of S0 ext works analogously to (8).
By means of Sobolev interpolation techniques and the auxiliary results stated in
[7, 18], one can show that the FETI/BETI preconditioner defined in (12) yields the
following condition number estimate in the presence of one exterior domain Ω 0,
“ “ ”” 2
κ(P M −1
S,α P ⊤ P ⊤ H
F P) ≤ C 1 + log max F
F ⊂Γ
h F
max
H 0
S F ⊂Γ
H 0 F
, (20)
where Γ S = S p
i=1
Γi is the skeleton of the domain decomposition, and F runs over
interfaces of subdomains, i. e. F = Γ ij for some i ≠ j, and H F := diam F. By h F
we denote the minimal mesh size on the interface F. As in the standard case, C is
independent of the diameters H i, H F, the mesh size h F and possible jumps in the
coefficients α i.
Note that ( H 0
H F
) d−1 is proportional to the number of subdomains touching Γ 0.
Our first numerical experiments show even better performance than expected from

Coupled FETI/BETI solvers – nonlinear, unbounded 7
p H 0/H F (H F/h F) 4 8 16 32
10 3 (PCG) 8 9 11 13
37 6 9 11 13 16
145 12 9 13 15 17
577 24 11 14 16 –
2305 48 12 – – –
Table 2. Iteration numbers of the PCG subspace iteration of the FETI/BETI
method in presence of a twodimensional exterior domain. p: Number of subdomains.
the estimate (20). In Table 2 one can observe a logarithmic instead of a linear growth
in the condition numbers, when increasing H 0
H F
.
Whenever computing on parallel machines with such exterior domains, one has
to keep the load balancing of the processors in mind. Our suggestion is either to use
a special decomposition with different mesh sizes to obtain balance, or to treat the
exterior subdomain on a group of processors and apply an inner parallelization of
the corresponding local problem, e.g. using the techniques described in [1] and the
references therein. In experiments, we have observed that for nonlinear problems in
unbounded domains, a coarse solution of a homogeneous Dirichlet problem (without
considering the exterior domain) may serve as a good initial guess for the Newton
iteration on the fine level including the exterior domain.
5 Conclusion
In this work we gave an overview on the standard FETI and BETI methods for
linear large-scale potential problems on bounded domains, and an outline on how
to extend and use these techniques in the presence of nonlinearities and unbounded
domains. Our numerical results are rather promising in both cases.
For a proof of the effectiveness of our new preconditioner that can be applied
to nonlinear problems, an analysis of FETI/BETI methods for varying coefficients
is needed, which is certainly a challenging subject of future research. For the unbounded
case, we could give gave some analysis leading to a suboptimal condition
number estimate, while our numerical experiments show an even better, quasioptimal
behavior.
Acknowledgments
The authors would like to thank Olaf Steinbach and Günther Of (Graz University of
Technology, Austria) for fruitful discussions, and gratefully acknowledge the financial
support of the Austrian Grid project and the FWF (Austrian Science Funds)
Special Research Program SFB F013. We remark that in our numerical experiments
we have used PARDISO (www.computational.unibas.ch/cs/scicomp) as a sparse
direct linear solver and Steinbach’s BEM-package OSTBEM.

8 Ulrich Langer and Clemens Pechstein
References
1. Ainsworth, M. and Guo, B.: Analysis of iterative sub-structuring techniques for
boundary element approximation of the hypersingular operator in three dimensions,
Appl. Anal. 81(2), 241–280 (2002)
2. Costabel, M.: Symmetric coupling of finite elements and boundary elements. In:
Brebbia, C.A., Kuhn, G., Wendland, W.L. (ed.) Boundary elements IX, 411–420,
Springer, Berlin (1987)
3. Farhat, C. and Roux, F.-X.: A method of finite element tearing and interconnecting
and its parallel solution algorithm, Int. J. Numer. Meth. Engrg., 32,
1205–1227 (1991)
4. Farhat, C., Lesoinne, M., and Pierson, K.: A scalable dual-primal domain decomposition
method, Numer. Linear Algebra Appl., 7, 687–714 (2000)
5. Dohrmann, C.R.: A preconditioner for substructuring based on constrained energy
minimization, SIAM J. Sci. Comput., 25(1), 246–258 (2003)
6. Hsiao, G.C., Steinbach, O., Wendland, W.L.: Domain decomposition methods
via boundary integral equations, J. Comput. Appl. Math., 125, 521–537 (2000)
7. Klawonn, A., Widlund, O.B.: FETI and Neumann-Neumann iterative substructuring
methods: Connections and new results, Comm. Pure Appl. Math., 54,
57–90 (2001)
8. Langer, U.: Parallel iterative solution of symmetric coupled FE/BE-equations via
domain decomposition, Contemp. Math., 157, 335–344 (1994)
9. Langer, U., Of, G., Steinbach, O., Zulehner, W.: Inexact data-sparse boundary
element tearing and interconnecting methods, SIAM J. Sci. Comp., 29(1), 290–
314 (2007)
10. Langer, U., Pechstein, C.: Coupled finite and boundary element tearing and
interconnecting solvers for nonlinear potential problems, ZAMM - J. Appl. Math.
Mech., 86(12), 915–931 (2006)
11. Langer, U., Steinbach, O.: Boundary element tearing and interconnecting methods,
Computing, 71(3), 205–228 (2003)
12. Langer, U., Steinbach, O.: Coupled boundary and finite element tearing and
interconnecting methods. In: Kornhuber, R., Hoppe, R., Periaux, J., Pironneau,
O., Widlund, O., Xu, J. (ed.) Proceedings of the 15th int. conference on domain
decomposition. LNCSE, 40, 83–97, Springer, Heidelberg (2004)
13. Mandel, J., Dohrmann, C.R., Tezaur, R.: An algebraic theory for primal and
dual substructuring methods by constraints, App. Num. Math., 54, 167–193
(2005)
14. McLean, W.: Strongly elliptic systems and boundary integral equations, Cambridge
University Press (2000)
15. Pechstein, C., Jüttler, B.: Monotonicity-preserving interproximation of B-Hcurves,
J. Comp. Appl. Math., 196, 45–57 (2006)
16. Steinbach, O.: Numerische Näherungsverfahren für elliptische Randwertprobleme.
Finite Elemente und Randelemente. B.G. Teubner, Stuttgart, Leipzig,
Wiesbaden (2003)
17. Steinbach, O.: Stability estimates for hybrid coupled domain decomposition
methods, Lecture Notes in Mathematics, 1809, Springer, Berlin Heidelberg
(2003)
18. Toselli, A., Widlund, O.: Domain decomposition methods – Algorithms and
theory, Springer Series in Computational Mathematics, 34, Springer, New York
(2005)