meeting 2024 08 13 n57 - JacobPilawa/TriaxSchwarzschild_wiki_5 GitHub Wiki

Context

  • I've got a few updates following the meeting Emily and I had this morning -- we talked in depth about some of the issues related to both the GPR and dynesty routines, and the on-going questions regarding the NNLS vs. kinem chi2 values.

  • On the GPR/dynesty front:

    • We discussed how it's probably not that surprising that the individual scale results are noisy given that the uncertainties on M/L are already ~10%, and so the 3% shift due to the velocity scaling is quite-significant both in terms of shifting the median model location and the distribution of points about that model location.
    • One thing we can do to mitigate this is to switch one of the parameters in our GPR/dynesty routine from nu=1.5 to nu=0.5, which provides a more flexible Gaussian process/is more sensitive to the individual points in our landscapes, so I've re-run our tests from before with the updated nu=0.5 value.
    • The result is that the landscapes look qualitatively what I would expect given the 1d panels, and so I'm much more comfortable now with those results/the consistency we see across scales.

  • On the NNLS/Kinem front:

    • Emily and I discussed a few things to look at which I've started to do below. The various test we discussed include;

      • Looking more systematically at the NNLS vs. kinem chi2 to see trends with parameters.
      • Re-writing out scraping routine to include the aperture histories, allowing us to compute full LOSVDs for our bins.

    • In summary:

      • I think I'm narrowing in on the problem! There seems to be a correlation between discrepent NNLS vs kinem models and the rho0 associated with the model. The low rho0 models really do seem to track the 1-to-1 line quite well, but it seems like the most discrepent points when looking at the NNLS vs. kinem chi2 is really coming from the highest halo (rho0) models), and in turn, this maps really directly to the misfitting of the outermost sigma data.
      • I have some plots showing this below, and I want to extend this to comparing the Mitchell vs. GMOS chi2s which should make this a little clearer.
      • In addition to looking at the spatial variation, Emily and I also discussed how to scrape the LOSVD information. I'm going to write some code to scrape this information and will generate the LOSVD plots we discussed in our previous emails.

  • On the LOSVDS:

    • I've finally gotten all the data scraped and am working on translating our old plotting routine to the new txz/npz system. It's quite a hefty package of functions and tools so it's taking a bit of time, but I am making progress.
    • These changes have been pushed to TriOS_Tools and have been checked by Emily. Everything is now fully up to date.

Plots

  • Starting off with the same 1D panels as last time, but now with the updated cornerplots:
Scale=0.97 Scale=0.99 Scale=1.00 Scale=1.01 Scale=1.02 Scale=1.03 Best Scales All Scales
  • And the cornerplots with nu=0.5:
K Scale=0.97 Scale=0.99 Scale=1.00 Scale=1.01 Scale=1.02 Scale=1.03 Best Scales
40
50
60
80
100

Kinem vs. NNLS Diagnostics

  • Here are some new plots which compare the kinematic chi2 to the NNLS chi2 for the newest grid of models (with the corrected binning scheme).
  • For these plots, I've added a 1-to-1 line and colored by the different parameters of the model. To me, the rho0 panel stands out greatly. It seems to me that the largest halos are resulting in the largest discrepency in the kinem vs. NNLS chi2, which makes me inclined to believe that this is something in the othermost kinematics (perhaps the outer sigma)?
    • At least to me, the rho0 panel seems to show that the lowest rho0 models seem to roughly track the 1-to-1 line much better than the large rho0 models systematically. This doesn't seem to be as clear for the other parameters.
    • Perhaps the reason this isn't the case for H15 is due to the quality of the outermost data/how well the models are able to fit these outermost bits? Currently still unclear but I'm leaning that way and hope to explore it a bit more.
Colored by BH ML Rho0 T Tmaj Tmin
All Models
Base Models

What is driving this difference?

One obvious question is what is driving the discrepency between the NNLS and kinem chi2, so I started to investigate that a bit. Specfically, I've take a small slice of models from the plots above between 1310 and 1320 in NNLS chi2 and have begun to look at the resulting kinematics for the models in this small slice. One thing that stood out to me was the outermost sigma points as a function of total chi2 seem to be WAY more misfit in the discrepent models than the well-agreeing models. Here are some radial kinematics for the 40 models in this small slice, ordered from best fitting to worst fitting. Note that as you scroll down, the fits get worse but by eye, this is largely driven by the outermost sigma.
Total Chi2 Kinematics
1303.91
1304.57
1305.84
1308.05
1310.44
1312.36
1312.37
1312.64
1312.99
1314.02
1316.72
1316.75
1318.15
1318.85
1319.32
1319.65
1321.25
1321.74
1323.78
1323.88
1325.2
1330.46
1330.97
1331.5
1332.42
1334.42
1334.44
1336.32
1336.43
1339.65
1340.87
1341.03
1342.27
1343.08
1345.94
1346.02
1347.99
1352.47
1366.14
  • You can collapse the plots in the bullet above into a single plot split by moment. In the plot below, I've taken the same models in the thing slice, and plotted the kinem chi2 associated with each moment as a function of total kinematic chi2. Immediately, we see that, as the fit (total chi2) gets worse, the chi2 associated with sigma is essentially linear, and this is mostly coming from the outermost bins' sigma fit than any other moment which scatter randomly about their mean.
Total chi2 vs. Moment Chi2
  • And interestingly, this trend seems to only appear in the Mitchell data. Here is the same plot as above but now I'm computing the moment chi2 from the two datasets separately. The GMOS data do not appear to have the same trend in the sigma panel as we see in the Mitchell data, suggesting that the GMOS data are relatively insenitive to whatever issue is causing this:
GMOS Mitchell
  • I went a bit further and here are a few more plots focusing on the outer portions of the data. In these plots, we're looking at the final 4 bins, the 8th-4th outer bins, 12th-8th outer bins, and then 16th to 12th. You can see that the trend of increasing chi2 in sigma with increased total chi2 really seems to impact the outermost bins preferentially.
Bins Plot
Last 4
Last 8-4
Last 12-8
Last 16-12
⚠️ **GitHub.com Fallback** ⚠️