meeting 2024 10 04 n315 - JacobPilawa/TriaxSchwarzschild_wiki_5 GitHub Wiki

Context

We're back to working on NGC 315 with many lessons learned from N57.
As a quick reminder, there were a few things we wanted to clear up as we revisit N315:
- The old kinematics had the outer bin flipping issue.
- The old kinematics/posteriors were computed using the kinematic chi2 (without including any dummies). We think we want to switch to NNLS without dummies moving forward.
- The old kinematics were run with the old set of triaxnnls binaries (this isn't necessarily an issue, but it's something to be aware of when looking at any of the old results).
So in this bullet:
- I've rescraped the old data's GH moments (allowing us to compute the NNLS chi2 per bin/per moment for the old data).
- I've also reprocessed the kinematics to be correct. This entailed:
  - Swapping the positions of the outer Mitchell bins prior to point-symmetrization (and then PAS'ing the data afterwards).
  - Re-minimizing the Grid A models with the new kinematics, and then scraping so that we have the set of kinematics and GH moments for all the data.
Main takeaways: I think that the results are looking great so far!
- The 1d panels (which are currently quite undersampled) seem to be showing roughly the same shape for the landscape, at roughly the same location.
- On NNLS vs. Kinem:
  - Both the old minimizations and the re-run minimizations show more or less the behavior we would expect: the dummies seem to provide a constant offset but without changing the 1-to-1 behavior, and the NNLS and kinem seem to track each other quite well with a bit of scatter for the less-well-fitting models.
- On the old minimizations vs. the new minimizations:
  - There are really two changes between the old data and the new data. In addition to fixing the outer bin flipping issue, the new data are minimized with the new binaries. We thus would expect there to be a slight difference in the chi2 between the new and old both in terms of the bin swapping and in terms of using the new binaries. It seems to me like the bin swapping issues was very negligible, since the difference between the new and old binary chi2 is ~5 or so, about the same magntiude as we saw for N57 and H15.
Moving forward:
- I think all of this points to us being able to more or less reminimize the libraries with the new data and continue on our way.
- I think I'd prefer to start by reminimizing the scalings we had for Grid A before moving straight to reminimizing Grid B, that way we have a greater sample of models to make these diagnostic plots. Happy to hear your thoughts!

Plots

1d Panels

First, here are the 1d panels for the ~866 models in Grid A. These are all scale = 1.0 cases, but we can certainly add more scales in the mix if we deem fit. We really are quite undersampled here, so I'm not bothering with GPR/dynesty or taking these too seriously since this was a large, uninformed grid of points.

Case	Kinem (No Dummies)	Kinem (With Dummies)	NNLS (No Dummies)	NNLS (With Dummies)
Old Data	[images/241004/rescraped_kinem_no_dummies.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/241004/rescraped_kinem_with_dummies.png)	[images/241004/rescraped_nnls_no_dummies.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/241004/rescraped_nnls_with_dummies.png)
Rerun Data	[images/241004/rerun_kinem_no_dummies.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/241004/rerun_kinem_with_dummies.png)	[images/241004/rerun_nnls_no_dummies.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/241004/rerun_nnls_with_dummies.png)

Comparisons of NNLS vs. Kinem for OLD Minimizations

After rescraping the old data, we now have access to the NNLS chi2 per bin and per moment for the old minimizations. Again, these minimizations are quite flawed (old binaries, outer bin issue), but comparisons between the different cases are qualitatively similar to what we were seeing with N57/the other galaxies. The NNLS and Kinem chi2 seems to track each other decently, and the dummies seem to only impact the chi2s by a fixed amount.

Plot
images/241004/old_kinem_nnls_comparison.png

Comparisons of NNLS vs. Kinem for New Minimizations

We find an extremely similar story for the new set of minimizations. Again, these results are obtained with the new binaries AND after correcting the Mitchell binning issue. Even with those changes, we find a similar story where the dummies seem to only add a constant to the landscapes, and the NNLS and kinematic chi2 seem to track each other quite well with a little bit of scatter for less-well-fitting points.

Plot
images/241004/new_kinem_nnls_comparison.png

Comparing chi2 between Old and new Minimizations

Perhaps the most interesting comparison is to look at the chi2 for the same quantity but before and after correcting the binning and switching binaries. Given these two changes, I am quite happy with the differences being comparable to what we saw for N57/H15 -- there is at most a constant shift between the old and new cases, and I think this is very encouraging for recovering our results obtained before. We again see a difference between the old and new cases of ~a few, and I am assuming for now this is likely the result of the new binaries being used (the new binaries want, on average, a lower chi2).

One-to-One	Histogram of Differences
[images/241004/line_comparisons.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/241004/hist_comparisons.png)