Fix issue 468 (Levenberg-Marquardt damping term)

[dwpaley]

The damping term doesn't blow up so rapidly when the refinement is fully converged (i.e. objective stops changing)

https://github.com/cctbx/cctbx_project/issues/468

@luc-j-bourhis @nksauter Any objection?

[nksauter]

No objections that I know of at this time.

[dwpaley]

I see the tst_commandline_anomrefine failure on Centos 8/python36, but I couldn't reproduce it in 300 trials in a VM. I know that was one of the builds affected by the PR from last night, so I'll rebase and see if the tests pass.

[luc-j-bourhis]

Interesting. I would say that if the damping terms blows up when the refinement is fully converged, that's because the termination criteria on gradient and step are too optimistic (methods had_too_small_a_step and has_gradient_converged_to_zero governed by step_threshold and gradient_threshold respectively). Could you try reducing those two values and see whether it solves your problem? The theory behind the test you want to change is that rho shall not be too small. So actually, some implementations impose rho > eps for some small positive eps instead.

Then, @pcxod L-M is used in Olex 2, isn't it? Any opinion on that?

[dwpaley]

Actually this only happens when step_threshold and gradient_threshold are set to 0. In that case it runs iterations until actual_decrease is identically 0; then it starts doing

      else:
        self.non_linear_ls.step_backward()
        self.mu *= nu
        nu *= 2

until (in ~30 more iterations) mu is in the 1e150 range and the damped step norm becomes identically 0 and the refinement stops in the middle of the iteration. Then the problem is only detectable because the esds are calculated based on the (extremely over-) damped normal matrix.

If you want to leave the rho test as it is, I suggested a different solution in the issue I wrote, where the iteration always goes to the end, finishing with build_up. Then esds will be calculated from the un-damped normal equations as Sheldrick recommends: http://shelx.uni-goettingen.de/shelxl_html.php#DAMP The only problem is, if the refinement was unstable (hence the need for L-M refinement) then it might crash with a Cholesky failure when attempting the final un-damped solution.

Otherwise some other more reliable way to detect convergence. I have more thoughts on this but it might exceed the scope of this PR.

Dan