Giuseppe Carleo

yes, good idea! just copying here the relevant work for energy minimization with Newton https://arxiv.org/pdf/cond-mat/0412634.pdf

this needs first a change in ```LocalOperator```...

I don't know if this is relevant, but if the interaction on the bond is not symmetric (i.e. O_{i,j} \neq O_{j,i}) then `GraphOperator` will pick an arbitrary ordering

Indeed this is the issue, since ```python >>> n_sites=3 >>> graph = nk.graph.Chain(n_sites, pbc=True) >>> print(graph.edges()) [(0, 1), (0, 2), (1, 2)] ``` whereas the loop by hand picks as...

that works with re-sampling, if I remember correctly, so you would also need to reweight the gradients etc with the correct factor... it could be implemented but I don't see...

I am also worried about duplicating a lot of code by merging this, isn't there a way to make this the default logic and specialize it when it comes to...

This is a broader topic for discussion I guess, but more generally, I am a bit worried about the combinatorial proliferation of functions that do almost the same things, maybe...

> I believe most of the pain and criticism here is that you are not familiar with multiple dispatch. This could be solved by writing documentation about it and tutorials...

sorry why is the current implementation inefficient? if you convert local operators to sparse matrices and then multiply by a vector, that is essentially optimal already I would argue

What do you mean that we first pass to dense? That is not possible otherwise you wouldn't be able to 20 spins, and we do them... On Thu, Oct 21,...