[petsc-users] SLEPc EPS Tolerance

[Eda Oktay]

Hello,

I am currently using EPS to find 2 smallest eigenvalues and
corresponding eigenvectors of a matrix in order to be used in spectral
graph partitioning algorithms. I know that default tolerance is E-8
and I also tried the solver with tolerances E-2, E-4 and E-6 expecting
that in every matrix I used, I get the most accurate solution (After
partitioning the matrix, I look at edge cut for accuracy) when
tolerance is E-8.

I even tried E-10 and the result was still the same, sometimes E-4 or
E-2 gives better results.

I was wondering if this is normal, if normal then how can we explain
this? Because in theory, one can expect that if tolerance decreases,
accuracy increases.

Thanks!

Eda