[petsc-users] Does the ciss solver ignore deflation space?

[Jose E. Roman]

> El 12 jul 2017, a las 20:09, Giacomo Mulas <gmulas at oa-cagliari.inaf.it> escribió:
> 
> Hello, I am trying to experiment with the ciss solver to see if it can be
> more efficient for my calculation. I need to compute the eigenpairs of a
> sparse real symmetric matrix. I don't know in advance how many I will need
> to compute, the convergence condition is on the subspace of converged
> eigenvectors. I don't know in advance the range(s) in which "good"
> eigensolutions (for my problem) will be.
> So my approach with other solvers has been to compute some
> eigensolutions for each call to epssolve, each time setting a deflation
> space with previously found eigenvectors. Not superefficient but it works,
> and it scales.
> However, when I attempt this with the ciss solver I keep getting the same
> eigenpairs over and over, regardless of the deflation space. Any clue as to
> why this happens and/or how to cure this?
> 
> Thanks in advance, best regards
> Giacomo

The current implementation of CISS ignores the deflation space. I don't know if it makes sense to set constraints in the contour integral solver. The idea of CISS is that it computes all eigenvalues contained inside the region. So in your case, a possible strategy could be to use small, disjoint but adjacent regions, and repeat until all wanted eigenvalues are discovered. But CISS is generally quite costly. With default parameters, it will require many factorizations per each region.

Jose