Re: Adding a non-Hermitian Eigensolver

Question

Following from 1, I'm interested in adding a non-Hermitian eigenvalue solver to ITensor. Looking at iterativesolvers.h, ITensor implements a version of Davidson which uses ITensor's various structs. I was wondering if it's possible to use a third-party eigensolver routine (such as Arnoldi from ARPACK) in ITensor. This would involve extracting the raw matrix data from the matrix object, which could then be passed to ARPACK. Is this feasible, or would I be better off trying to implement an algorithm within ITensor's framework (similarly to how Davidson is implemented)?

PS Thank you for your great work!

Answer 1

Hi, yes we do need to and plan to add an Arnoldi code similar to Davidson. We actually had one in an older version of ITensor, and just need to put it back in.

So it is on our to-do list, and Matt Fishman has this as a high priority (he was the one who recently also added the gmres algorithm for iteratively solving linear equations).

I don't think we will use ARPACK, since it would become a dependency for ITensor and I've heard there are some issues with ARPACK and that it is not maintained much anymore.

Best regards,
Miles

Comments

Thank you Miles, that's quite exciting and will be of immediate benefit to my research.