Fitter convergence issues

[bencebecsy]

Many people reported the following error occasionally popping up:

StepProblem: Unable to improve chi2 even with very small steps.

The consensus seems to be that this can be fixed by specifying to use the GLSFitter instead of the auto option which usually selects the DownhillGLSFitter from pint. This can be done in pta_replicator like this (Thanks to Emiko's PR #6):

fakepsr.fit(fitter='gls')

The question remains why this error is happening at the first place. So this might be good to investigate.

Practically the question is whether we should switch the default to be the GLSFitter when using pta_replicator? Arguably the quality of the fit is not as essential in simulated datasets as in real data, so maybe this is okay, and would prevent a lot of people from running into this annoying issue.

[AaronDJohnson]

Maybe we should bring in some of the PINT folks to say something about this issue?

[bencebecsy]

That's a good idea. I mentioned this in a related issue in the pint repo, but I'll ping @abhisrkckl and @dlakaplan here in case they have some thoughts on this.

[bencebecsy]

I wonder what tempo2 does in terms of fitting? In libstempo I've never seen any errors about fits, but maybe that has more to do with errors not being too informative there - I guess it's possible that the plethora of recent segfaults when trying to simulate things in libstempo are indeed underlying fit convergence things that don't give a meaningful error?

[abhisrkckl]

See nanograv/PINT#1946