[Trilinos-Users] null space removal

[Christopher K. Newman]

Hi Mike-

We have a nonlinear ocean model with many coupled pdes, which we are 
trying to solve for the steady state using jfnk with gmres with a finite 
difference jacobian as preconditioner.  Gmres does not seem to converge 
and upon examination of the finite difference jacobian with matlab a 
clear checkerboard nullspace mode appears in one of the components.  I 
believe it is manifested in a divergence term.  I am wondering if this 
null space accounts for the poor convergence and if it can be remedied 
by a projection orthogonal to the nullspace similar to what is done for 
the pure Neumann problem with a constant nullspace.

Chris

On 9/8/10 12:31 PM, Heroux, Michael A wrote:
> Hi Chris,
>
> This is something I have wanted to add for a long time, but as far as I know it is not generally available.  ML provides this capability itself, if you are using it.  It would be fairly straight-forward to add this.  Can you describe your specific situation?
>
> Thanks.
>
> Mike
>
>
> On 9/8/10 1:07 PM, "Newman, Christopher K. (LANL)"<cnewman at lanl.gov>  wrote:
>
> Is there a way to automatically remove a known null space during an iterative solve?  Dare I say, similar to KSPSetNullSpace in Petsc.  Or is the preferred method user implementation of a projection operator?
>
> Thanks,
>
> Chris
>    

-- 
Christopher K. Newman
Los Alamos National Laboratory, T-3     Email: cnewman at lanl.gov
PO Box 1663 MS B216                           Phone: 505 606 0647
Los Alamos, NM 87545                          Fax: 505 665 5926