meeting 2024 10 22 n315 - JacobPilawa/TriaxSchwarzschild_wiki_5 GitHub Wiki

Plots

  • First, here are the 1d panels for the base models from Grid A, B, Bprime, C, and D. For Grids A and B, we currently have [0.97, 0.98, 0.99, 0.995, 1, 1.005, 1.01, 1.02, and 1.03] for the scalings. We currently only have the s = 1.0 models for Grid Bprime (we can certainly add this). And for Grids C and D, we should nominally have scales for [0.99,0.995,0.998,1.0,1.002,1.005,1.01], but I messed up submitting the 1.002 and 0.998 models and am adding them currently (they're running on Expanse right now).

  • It seems to suggest what we saw in the original grid of models -- there does appear to be a low Tmin, Tmaj minimum that our models are finding, and this is confirmed with the scalings and the addition of the Bprime models. It's not obvious to me that our coverage is bad at this point.

Case T Space Angle Space UPQ Space
All Models
Best Scales
  • And here's a quick comparison of the chi2 for each of the grids (best scaled), which do indeed show that we were closing in on the minimum quite well:
Plot
For completeness, here are the 1d panels for all scalings for Grids C and D alone as we currently have them. I'll update this when s = 0.998 and s = 1.002 are completed.
Plot

d1 vs. d3 Grid B Prime

  • Dither 1 vs. Dither 3 Bprime Results:
    • Note -- I hadn't noticed this when I scraped the data, but it looks like only ~2/3 of the models ended up minimizing last night, with the other ~1/3 getting cancelled due to Expanse filesystem issues (but looks like they never started so we didn't lose hours or anything). Those are now minimizing and will be added shortly!
    • With that said, I think that this looks reasonable for well-fitting models. There is a bit of scatter in the resulting trend line (roughly a width of ~100 on either side), but still roughly consistent with what we've been seeing. I also think this is somewhat expected since we're only covering a range of ~200 in chi2 with ~200 model points, so I expected the trend to be a bit noisier.
    • I can certainly come back to this grid, run the remaining models, and add scalings, but I'm not sure if this is really needed given where Grids C and D land, but we can dsicuss in the meeting.
NNLS Kinem

GPR + Dynesty on Combined Grid

  • First, here's a plot showing the number of models vs K for the best-scaled models:
Plot
  • I've also started to run our results through dynesty + GPR, since I think we're at a point to start scrutinizing the results and comparing them to our old models. I've run GPR + dynesty for on the "best-scaled" models with the results shown below. In short, the results are quite consistent with one another across Ks and across nu's.
    • Note that I'll run these with a higher number of samples -- currently these were quickly made using nIter = 1, (they still took ~20 mins each), so I just have to have a bit more time to post the polished results. I don't expect much to change from here, however.
K=60 K=80 K=100
nu=0.5
nu=1.5
  • And a few more cornerplots, same as above, but showing angles/shapes/rho0 together. Note that I will change the limits here but needed plots quickly:
    • Note that the results here (in particular for the K = 100, nu = 1.5 cases) look great, and are consistent with our old results (see below):
K=60 K=80 K=100
nu=0.5
nu=1.5
  • Here's the most recent cornerplot using the old data (from the 10/02 page) for comparison:
Plot (Angles) Plot (Ts)

(from 10/18 Page) Other Plots

  • I have a few other plots here just for completeness. The first is simply comparing the NNLS and kinem chi2 for the different grids we have so far -- again, nothing too surprising from these. The two plots are identical, but the Zoomed in one only plots the rejection-sampled grids to better look at the well-fitting models.
Kinem vs. NNLS Kinem vs. NNLS (Zoomed)
And here, side-by-side, are the 1d panels comparing the kinem chi2 and the NNLS chi2 (same as the panels above). Both of these do not include the dummy moments.
Case T Space Angle Space UPQ Space
NNLS
Kinem
And lastly, a side-by-side comparison of the histograms for each of the grids in NNLS vs. Kinem space. Again, the NNLS panel is the same as above, and both panels exclude dummy moments.
NNLS Kinem
⚠️ **GitHub.com Fallback** ⚠️