Make Lanczos algorithm more robust about impuritymodel HOT 7 OPEN

I have done some investigation of a few methods.

1. Using blocks of vectors for the Lanczos method, instead of a single vector. This is useful if we want to calculate offdiagonal elements in the Greens function. Numerically, this increases stability a bit as well.
2. Partial reorthogonalization is probably the best compromise between computational cost and numerical accuracy, link to old paper
3. Full reorthogonalization is easy to implement, but veeeery computationally expensive.

In general, all methods that adds some orthogonalization-step to the method will require that we save all the Krylov vectors we generate. This increases the memory footprint of the calculation quite a bit. In my experience, the Greens function does not converge faster when we add orthogonalization. I have not actually looked at the numerical accuracy of the resulting Greens function though.

from impuritymodel.

Restarting the algorithm after a certain number of iterations seems to be a popular approach, e.g. Thick-Restart Lanczos method: https://zenodo.org/records/1236144

from impuritymodel.

Not sure how big this issue is.

from impuritymodel.

No, I am not sure that this will actually help with calculating the Greens function. The restarted methods work by focusing on certain eigenvalues/eigenvectors. If what you want is just certain eigenvalues/eigenvectors this is great, but for the Greens function to converge quickly, what eigenvalues should we focus on?

from impuritymodel.

I don't know. Perhaps it's not suitable for calculating Green's functions.

from impuritymodel.

The way we calculate Greens functions relies on the tridiagonalization of the Hamiltonian. With restarted methods I think we replace the tridiagonal structure with something different, but maybe there is a clever way to calculate the Greens function (we only need one element of the inverse of the Hamiltonian).

from impuritymodel.

When I quickly read about restarted methods I agree it does seem the matrix is not tridiagonal. So probably the three implementation ideas you outlined above are more promising.

from impuritymodel.