[Trilinos-Users] ML on 2x2 block structures?

[Nico Schlömer]

Hi all,

I have this Hermitian, positive definite linear problem here, and I'd 
like to speed up CG performance by preconditioning with ML.
There are N complex-valued unknowns. In one implementation, the matrix 
is constructed as a real-valued 2Nx2N Epetra_FECrsMatrix/_Operator by 
splitting up the real and imaginary parts of the unknowns. The matrix 
can be viewed as consisting of 2x2 blocks.
Using ML for this problem does not perform well, and I think one reason 
for this is that ML has no notion of this block structure, and thus may 
perform interpolation between the real and imaginary parts, or similar 
undesirable things.

Is there a canonical way to treat problems like this in ML?

Cheers,
Nico