Dear Bruno > > > 2) The largest eigenvalue is computed instead of the largest one, and, I > guess, the smallest eigenvalue is computed instead of the largest one. > If you use ARPACK through deal without patching it, then the only > method available is shift-inverse. The thing is that when you know the > largest eigenvalue of the inverse you can get the smallest eigenvalue > of the operator. So yes, the names are reversed. > Yup, using the patch solved this issue. At least the smallest eigenvalue is smaller than the largest.
> > > 3) I get complex eigenvalues in some cases and I am quite sure, that all > eigenvalues of the system are real. > I've not had this problem. Maybe it's because of numerical error? Eigenvalue: (1.03342,0.0069945) with arpack tolerance 1e-8. Is it possible? > > > I noticed that in some cases the "smallest" eigenvalue does not match > maximum eigenvalue obtained from power method. > Are you sure that the power method has converged? Also, try using > ARPACK's regular mode instead of shift-inverse. I think it might be > better to find the largest eigenvalue. > The deal.II power method converged with residual: 9.5052e-05 I switched to regular mode and obtained 1.04192 from arpack and 1.03981 from power method. The tolerance of arpack was set to 1e-8. The difference is growing with mesh size, so I assume the tolerance is set for eigenvector, resulting in much smaller tolerance for the eigenvalue. Also, I got the smallest eigenvalues equal to 0 in some cases, while I'm sure that the matrix is not singular. I can solve a linear system with the same matrix using Krylov solvers without a problem. Michał -- The deal.II project is located at http://www.dealii.org/ For mailing list/forum options, see https://groups.google.com/d/forum/dealii?hl=en --- You received this message because you are subscribed to the Google Groups "deal.II User Group" group. To unsubscribe from this group and stop receiving emails from it, send an email to dealii+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
