meeting 2023 06 06 - JacobPilawa/TriaxSchwarzschild_wiki_5 GitHub Wiki

This is a quick update showing a quick set of results we discussed in the last meeting – specifically, this bullet tests selecting only "good fitting" models to be included in the chi2+Norb GPR + dynesty calculation.

In the past, we had been determining which models get selected based on the sum of (chi2+Norb), and then taking all models within some number $K$ from the minimum. I am now making this cut on the chi2 values alone (as if we were running a standard recovery test), and use those models with the correpsondign Norb to make these cuts.

I tried a few different quick tests, including changing the cutoff value used (for our normal recoveries, I use a K=40). For these tests, I tried both a K=30 and K=50, and it doesn't seem to make a difference. I also tried tuning the assumed error on the quantity chi2 + Norb up to error = 1.0, beyond which the errors are too large to make meaningful constraints.

In summary:

• I have a few other tests I could still do for these results (look a bit more at the Norb surfaces near the minimum, or trying 1d axisymmetric tests) but simply including only well-fitting models doesn't seem to improve our GPR and dynesty routines. In these results, we still see that our GPR's are running away to the corners, and increasing the K further will start to include some even more problematic model points causing even more issues.