[petsc-users] estimation of max and min eigenvalues in SLEPc

[Jose E. Roman]

El 03/06/2015, a las 17:47, Xujun Zhao escribió:

> Hi Jose,
> 
> Thank you for your reply. How about the computational cost compared to one EPSSolve() with all eigenvalues? what methods does SLEPc use for each solve? Because it may be cheaper for largest eigenvalue if the power method is used, but I don't if it is still so for smallest eigenvalue?
> 
> Xujun
> 

Don't compute all eigenvalues.

For the largest eigenvalue, don't use the power iteration. The default solver (Krylov-Schur) will be very fast for that. For the smallest eigenvalue, convergence may be slow if eigenvalues are small and poorly separated - it may be necessary to do shift-and-invert, in which case the cost may blow up.

Jose