meeting 2025 04 03 n57 - JacobPilawa/TriaxSchwarzschild_wiki_5 GitHub Wiki

Context

  • I reminimized the 2200 N57 models with the newest set of kinematics and have some diagnositcs below. Nothing is obviosuly broken and the fits are at the very least sensible! A great start to the reminimization.
  • Takeaways:

    • Generally speaking, the 1d chi2 panels are in large agreement in shape with the old minimization with the exception of Tmaj. You might remember that Tmaj actually caused us a bunch of trouble in old N57 minimizations. In the old tests, the landscape was extremely flat and we had a "double minimum" situation which we had fixed by modifying the prior on Tmaj.

      • This issues appears to have disappeared, and the new Tmaj profile is MUCH more reasonbly shaped, with the minimum nearer Tmaj~0.9 rather than ~0.55 which we had been "assuming" in the past.

    • It seems like the new M/L might be slightly smaller than we had initially found too, so I think there's a general "tilt" in the mass panels that could potentially be solved with scalings.

      • In the past, we used s=(0.97,0.99,0.995,1.0,1.005,1.01,1.02,1.03), which I think we should also add here and then run our "select best scalings" routine.

    • The chi2 values for the new minimization are ~500 higher than the old minimization which is a bit strange, but this could be in part due to the shifting of Tmaj/the M/L or BH "miss" that we're currently seeing. With that said, our reduced chi2 is still less than one, with the current best model having 1690/(2158 + 416) = 0.859 or so, larger than the other galaxies which were closer to 0.7-ish.

Diagnostics

  • First, here's the 1d panels for the updated minimizations, with the last column being the scale=1.0 case for the old minimization.
Colored by: BH ML Rho0 T Tmaj Tmin OLD VERSION
[images/250403/color_by_bh.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250403/color_by_ml.png) [images/250403/color_by_rho0.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250403/color_by_T.png) [images/250403/color_by_Tmaj.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250403/color_by_Tmin.png) images/250403/old_panels.png
  • And here are some diagnostics on the current best fitting model. I've included the radial kinematics as well as heatmaps showing the chi2 per bin and moment, and the chi2 summed over the moments:
Radial Moments Full Heatmap Chi2 by Moment
New Best Fit [images/250403/current_best_fit.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250403/heatmap_full.png) images/250403/heatmap_sum.png
Old Best Fit [images/250403/old_best_fit.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250403/heatmap_full_old.png) images/250403/heatmap_sum_old.png
  • And here's a quick comparison of the DIFFERENCE in the heatmaps above (old minus new). It looks like there's almost a 200 difference in the chi2 of the sigma profiles. It could potentially be because of sub-optimal parameter choices for the current grid? It's not entirely clear right now but I'll keep looking into it.
Full Heatmap Difference (Old Minus New) Difference in Moments (Old Minus New)
[images/250403/heatmap_full_diff.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250403/heatmap_sum_diff.png)
  • And here's a quick comparison of the NNLS and kinem chi2 for the new minimization vs. the old minimization:
    • Note that these both exclude dummies and include all the bins in the chi2 calculation. We had tested masking the outer 4 Mitchell bins in the old minimizations, but it's unclear we still need to do that for the new minimization so I included everything for now.
NNLS Comparsion Kinem Comparsion
[images/250403/NNLS_comparison.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250403/kinem_comparison.png)

Chi2 Follow Up Questions

  • As seen above, the chi2's for the new kinematics are ~400-500 higher than the chi2's using the old kinematics:

    • I think I figured out the reason for this: it seems this discrepency is largely due to us masking the additional two features in the spectra and the chip gap. The uncertainties in the moments when masking these additional regions is systematically smaller than when they were not masked, and this propagates to our chi2 values.

  • First, here's a plot comparing the "published" errors to the errors from this most recent minimization (with the two additional masks + chip gap region masked).

    • Note that the V error agrees very well between the two cases, but all other moments have smaller errors with the new kinematics compared to the old kinematics.
Old vs. New Error Comparison
images/250403/error_comparison.png
  • One concern I had looking at this is that maybe the nMC=100 case hadn't converged on the error size yet, so I ran a quick case using nMC=300 to get a sense of how much the errors changes. The answer is that they don't change by much:
nMC=100 vs. nMC=300
images/250403/nmc300_test.png
  • So since the nMC didn't seem to be the issue, I compared the case where I have masked the additional two "spikes" we saw as well as the chip gap compared to the kinematics without these additional masks. This seems to show that, for many bins, the errors are in very nice agreement (likely these are cases where the additional masks or chip gap are not poorly fit). However there are definitely lots of points with much larger uncertainties when the regions are not masked.
BOX Mask vs. BOX Mask + Additional Regions
images/250403/chip_gap_test.png
  • I did a quick test where I took the latest set of minimizations and simply replaced the GMOS errors with the errors from the BOX Mask alone case, and the resulting chi2s are MUCH better. Here are those 1d panels:
Replaced Errors 1d Panels
images/250403/replaced_errors.png

Additional Diagnostics

  • Following up on the email exchanges before the meeting. Here's a quick histogram showing the uncerainties for the three cases. While the medians are sort of close to each other, I think the key here is that the case with the smallest mask have bins that are much higher than the others which won't appreciably change the median, but will change the chi2's quite subtantially.
Histograms of the Errors
images/250403/mask_error_comparison.png
  • And here are a comparison of the spectra, MC fits, and RMS for the 10 most discrepant (sigma) errors. They are ordered from most discrepant between the cases to least discrepant (but still quite discrepant).
BOX Mask BOX Mask + Edits BOX Mask + Edits + Chip Gap
[images/250403/CaseZ_discrep.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250403/CaseZ_mask_edits_discrep.png) images/250403/CaseZ_mask_edits_and_chip_gap_discrep.png
  • I also included a random sample of 10 spectra (a true random sample) to get a sense of the "typical" cases. These are not ordered in any particular way.
BOX Mask BOX Mask + Edits BOX Mask + Edits + Chip Gap
[images/250403/CaseZ_random.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250403/CaseZ_mask_edits_random.png) images/250403/CaseZ_mask_edits_and_chip_gap_random.png

N315 Error Comparison

  • I quickly made one other plot showing the error histograms above but with the most recent N315 errors as well:
with N315
images/250403/mask_error_comparison_with_N315.png