Some small amount of numerical variations are to be expected from parallel load distribution, but the sparse-gnr integration test requires unreasonably large floating point tolerances to have any chance of succeeding. I believe this is because eigsh uses a random starting vector, which has a butterfly effect after relaxation along the negative modes.
Some small amount of numerical variations are to be expected from parallel load distribution, but the
sparse-gnr
integration test requires unreasonably large floating point tolerances to have any chance of succeeding. I believe this is because eigsh uses a random starting vector, which has a butterfly effect after relaxation along the negative modes.