29.04.2014 Views

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

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

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

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

4 Ulrich Langer and Clemens Pechste<strong>in</strong><br />

dimensions the s<strong>in</strong>gle layer <strong>potential</strong> operator V i is always elliptic, whereas <strong>in</strong> two<br />

dimensions, due to the logarithm <strong>in</strong> the f<strong>un</strong>damental solution, it is only elliptic if<br />

diam Ω i < 1. This property can always be achieved by a suitable coord<strong>in</strong>ate scal<strong>in</strong>g.<br />

After discretiz<strong>in</strong>g and elim<strong>in</strong>at<strong>in</strong>g t i, one obta<strong>in</strong>s the symmetric and stable approximation<br />

Si, BEM<br />

h := D i, h + ` α i<br />

2<br />

Mi, ⊤ h + Ki, ⊤ −1` h<br />

´V<br />

αi<br />

i, h 2<br />

M i, h + K i, h´<br />

, (8)<br />

where V i, h , K i, h , D i, h are the bo<strong>un</strong>dary element matrices correspond<strong>in</strong>g to V i, K i,<br />

D i, respectively, and M i, h is a mass matrix.<br />

Note that the two approximations Si, FEM<br />

h and Si, BEM<br />

h are compatible and both<br />

spectrally equivalent to the Galerk<strong>in</strong> matrices of the exact Steklov-Po<strong>in</strong>caré operators<br />

S i. The application of Si, FEM<br />

h or Si, BEM<br />

h simply corresponds to the solution of<br />

local Dirichlet <strong>problems</strong>. For details we refer to [11, 12].<br />

2.2 Tear<strong>in</strong>g and Interconnect<strong>in</strong>g<br />

Introduc<strong>in</strong>g separate variables u i on the local subdoma<strong>in</strong>s, one can re-en<strong>for</strong>ce the<br />

cont<strong>in</strong>uity of the solution u across <strong>in</strong>terfaces Γ ij by the constra<strong>in</strong>ts<br />

X p<br />

Bi ui = 0 , (9)<br />

i=1<br />

where the B i are <strong>in</strong>cidence matrices.<br />

Problem (2) can be written as a constra<strong>in</strong>t m<strong>in</strong>imization problem, as well as a<br />

saddle po<strong>in</strong>t problem <strong>in</strong>volv<strong>in</strong>g Lagrange multipliers. Us<strong>in</strong>g the notion of the pseudo<strong>in</strong>verse<br />

( † ) and a special projection P address<strong>in</strong>g the kernels of the sub<strong>problems</strong>, it<br />

is possible to elim<strong>in</strong>ate the primal <strong>un</strong>knowns u i. F<strong>in</strong>ally, one obta<strong>in</strong>s the discrete<br />

dual <strong>FETI</strong>/<strong>BETI</strong> <strong>for</strong>mulation, to f<strong>in</strong>d the Lagrange multiplier λ such that<br />

where the <strong>FETI</strong>/<strong>BETI</strong> operator F is def<strong>in</strong>ed by<br />

F = B<br />

where B = ˆB i˜p<br />

, i=1 SFEM/BEM h<br />

:=<br />

P T F λ = d, (10)<br />

FEM/BEM˜† ˆSh<br />

B ⊤ = X p<br />

˜†<br />

Bi<br />

ˆSFEM/BEM<br />

i=1<br />

i,h<br />

Bi ⊤ , (11)<br />

FEM/BEM˜p<br />

ˆSi, h<br />

. The application of the pseudo-<br />

i=1<br />

<strong>in</strong>verses [S FEM/BEM<br />

i, h<br />

] † can be realized by the simple solution of regularized local Neumann<br />

<strong>problems</strong>. S<strong>in</strong>ce F is symmetric positive def<strong>in</strong>ite on range(P), one can solve<br />

the dual problem (10) by a preconditioned conjugate gradient subspace iteration.<br />

The preconditioner<br />

M −1<br />

S,α = (BD −1<br />

α B ⊤ ) −1 BD α S FEM/BEM<br />

h<br />

D αB ⊤ (BD −1<br />

α B ⊤ ) −1 , (12)<br />

first <strong>in</strong>troduced and fully analyzed by Klawonn and Widl<strong>un</strong>d [7], satisfies the quasioptimal<br />

condition number estimate<br />

κ(PM −1<br />

S,αP T P T FP) ≤ C(1 + log(H/h)) 2 , (13)<br />

<strong>in</strong>dependent of the values–and there<strong>for</strong>e possible jumps–of the coefficients α i. It is<br />

well known that an appropriate norm of the<br />

√iteration error of the conjugate gradient<br />

method will decrease at least by a factor 2` κ−1<br />

´n<br />

√ κ+1<br />

<strong>in</strong> n steps. The robustness with

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!