meeting 2024 02 16 n315 - JacobPilawa/TriaxSchwarzschild_wiki

Context

I've re-run Grid C (rejection sampled models) minimizations on the set of kinematics with the outer 4 Mitchell bins masked for all our scalings. It looks like the shape of the resulting landscapes are still reasonable and not too dissimilar what we have been seeing with the old kinematics.
Main takeaway: I think the main thing to gain from these plots below is that the new, masked kinematics seem to want a slightly larger halo and thus a slightly smaller M/L ratio, but not by a super extreme amount. The halo bumps up from roughly 3e9 to 4e9, and the M/L drops by about 10-15% or so it seems.
- Additionally, the Mitchell h6 of our new current best model looks much better! Things are still a bit weird but we're definitely getting closer to a great fit the Mitchell h6 data specifically.

First, here are the 1d panels in both T space and angle space for the Grid C models with and without the masking procedure. To me it looks like the two sets of landscapes are consistent with one another at the level of sampling we currently have, with the caveat that we seem to want a slightly larger halo and slightly smaller mass to light ratio.

	No Mask	Masked
T Space
Angle Space

	No Mask	Masked
T space

I've also plotted here the comparsion of the chi2 before and after masking the outer 4 bins for each of the scalings here. I was a bit surprised to see the scatter in these and am still working out why this might be the case:

Plot

Depending on what we choose to do next, we might need to run regression + dynesty on the set of points we have which includes this out mask. I've run GPR + dynesty on the set of best scales and all scales for a few K and a few nu. These aren't to be interpreted that rigorously, but I wanted to get a sense of where things currently are at.
If we proceed with rejection sampling again, I think I'd be comfortable using the K=60, all scales posterior below. I know we generally don't like using all scalings to generate the surface for regression, but I think we'd be limitng ourselves too much using the posteriors from the set of best scalings. To my eye, the posteriors from all scalings match the 1d landscapes for best scales while being a bit more well-behaved where we already think the minimum is.