On Sun, Jul 22, 2012 at 1:11 PM, Umut Tabak <u.tabak at tudelft.nl> wrote:
> Well, basically, I am not interested in time domain response. What I would > like to do is to find the eigenvalues/vectors of the system so it is in the > frequency domain. What I was doing it generally is the fact that I first > factorize the operator matrix with the normal factorization operation and > use it to do multiple solves in my Block Lanczos eigenvalue solver. Then in > my performance evaluations I saw that this is the point that I should make > faster, then I realized that I could solve this particular system, that is > pinned in your words, faster with iterative methods almost %20 percent > faster. And this is the reason why I am trying to dig under. > How many grid points per wavelength? > > >> basically the operator is singular however for my problem I can delete >> one of the rows of the matrix, for this case, I and get a non-singular >> operator that I can continue my operations, basically, I am getting a >> matrix with size n-1, where original problem size is n. > > > This is often bad for iterative solvers. See the User's Manual section > on solving singular systems. What is the condition number of the original > operator minus the zero eigenvalue (instead of "pinning" on point)? > > This is not clear to me... You mean something like projecting the original > operator on the on the zero eigenvector, some kind of a deflation. > See the User's Manual section. As long as the preconditioner is stable, convergence is as good as for the nonsingular problem by removing the null space on each iteration. -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120722/d85090e8/attachment.html>
