You also need to be careful about assuming the outputs of eigh are sorted, I think I found they weren’t necessarily here. You can also see some comments about eigenvalue problems in that thread. They’re non-trivial for gradient-based samplers to work with.