meeting 2024 01 29 n315 - JacobPilawa/TriaxSchwarzschild_wiki_5 GitHub Wiki

Context

  • I've run a few more tests following our exchange on email.
  • Specifically, I've run
    • 2 sets of scalings (+/- 2% on the mass parameters, which gets squared when modifiying the masses)
      • We discussed running an additional two sets of scalings with +/-5%, but I wanted to make sure I was handling these scalings properly before going wild with minimizations. I can still do this, and I think it would still be useful at this stage.
    • 1 additional test in which I mask the outer 4 Mitchell bins from the fit, which are at +/- 60" from the center.

Adding scalings

  • Like I said above, I ran forward with two sets of minimizations with scale factors of 0.98 and 1.02.
  • From the tests below, I haven't done any additional processing (e.g., selecting only the "best" version of each scaled model)
  • Takeaways:
    • Things are still looking good! The models are not jumping all over the place when I change the scalings, and you can pretty much match up the "base" models with their scaled versions.
    • The resulting parameters are all coming out roughly consistent with one another, again with the caveat that we still have to more thoroughly explore the space. I think this is a good start, though, all things considered.
    • Note that one of the 1.02 scaled models has become the new "best" model, but not very significantly (only changes the chi2 by ~40, down to 3487 from 3582)
  • The breakdown of the resulting models looks like:
Cutoff K Num. Models < min(chi2) + K
10 1
50 3
100 5
250 78
500 465
750 1027
1000 1579
1500 2342
2000 2584
5000 2598
all models 2598
Here are the resulting 1d chi2 vs. parameter panels, with the scalings colored differently from one another.
Plot
I've run a set of gpr+dynesty runs on these model points, and here are the resulting posteriors.
Cutoff K nu=0.5 nu=1.5
250
500
750
750

Masking out 4 mitchell bins

  • Removing the outer 4 Mitchell bins takes something like ~200 off the total chi2, which suggests that these 4 Mitchell bins were contributing ~50 per bin to the total.

  • While this helps in removing ~15% of the total chi2, the landscape shapes are nearly identical given the 866 points we have so far.

  • First, here's a breakdown of the number of models as a function of K for the 866 models with the outer 4 bins masked away:

Cutoff K Num. Models < min(chi2) + K
10 1
50 4
100 9
250 61
500 290
750 520
1000 680
1500 844
2000 866
5000 866
Here are the resulting 1d chi2 vs. parameter panels for the 866 models with and without the outer Mitchell bins masked, compared to the d1 and d3 results for the same sets of models.
d1 and d3 models (with all bins) d3 models (with outer 4 Mitchell bins masked)
And here's a comparison of the dynesty posteriors run on the 866 models with and without these Mitchell bins masked. I'm only presenting the nu = 1.5 case here for brevity.
Cutoff K all bins included outer 4 bins masked
500
1000
Lastly, here's a side-by-side comparison of the previous best-fitting model with all bins, compared to the best-fitting model without these outer Mitchell bins
Label All Bins Mitchell Masked
kinematics
beta
⚠️ **GitHub.com Fallback** ⚠️