[petsc-users] Krylov-Schur Tolerance

[Jose E. Roman]

For computing eigenvalues with smallest real part of generalized problems Ax=lambda Bx, it may be better to use a target value (instead of -eps_smallest_real). For instance, if you know that all eigenvalues are positive, use -eps_target 0 -eps_target_magnitude

What linear solvers are you using? In the default setting, the coefficient matrix for linear solves will be B, but with target=sigma the coefficient matrix will be A-sigma*B; this may make a difference. Also, in any case, if experiencing convergence problems I would suggest using MUMPS (see section 3.4.1 of SLEPc's users manual).

Jose



> El 17 feb 2017, a las 10:25, Christopher Pierce <cmpierce at WPI.EDU> escribió:
> 
> Hello All,
> 
> I'm trying to use the SLEPc Krylov-Schur implementation to solve a
> general eigenvalue problem.  I have a monitor on my solver and the
> solutions appear to converge correctly when using the approximation for
> the residual norm in the algorithm.  However, when the solutions are
> displayed and I retrieve the actual residual norm it is very large and
> increases with the size of the matrices I am working with.  In some
> cases it may be 10^17 times as large as the approximate norm.  I also
> don't get the eigenvalues I would expect for the system I am studying.
> 
> When I turn on the option "true residual" the solver fails to converge. 
> The residual norm shrinks to some limit (~10^-3) and then sits there for
> the remaining iterations.  As a note, I am solving for the eigenvalues
> with the smallest real part.  I have also tried the RQCG solver on the
> same problems and appear to get the correct results using it, but I'm
> looking to use the better scaling of the Krylov-Schur solver.
> 
> Does anyone know what could be causing this behavior?
> 
> Thanks,
> 
> Chris Pierce
> WPI Center for Computational Nanoscience
> 
>