Eigensolver not converged

[mfherbst]

Since #964 is in I quite frequently get something like

┌ Warning: Eigensolver not converged
│   n_iter =
│    42-element Vector{Int64}:
│     16
│      7
│      4
│     14
│      ⋮
│     18
│      2
│     13
└ @ DFTK ~/.julia/packages/DFTK/Yozrw/src/scf/self_consistent_field.jl:76

for the first few iterations. I think this is because due to the ordering (involving a Reighleigh-Ritz) on an eigenpair the residual simply increases above the threshold again.

@antoine-levitt Any idea what to do ? It probably does not really matter in practice, so should we ignore it ? Or somehow modify how we check for converegence ?

[antoine-levitt]

Hm this is annoying. I think what happens is that the LOBPCG detects the Nth eigenvalue as converged and the N+1th as not converged, but then switches them. It means that we return a non-converged eigenvector, which is breaking the promise. That sort business has always been a hack, we should just get rid of it, I'll try to see what I can do