meeting 2024 08 12 n57 - JacobPilawa/TriaxSchwarzschild_wiki_5 GitHub Wiki
Context
-
I've done a bit more work trying to figure out what is going on with our models. Given the strange behavior in high M/L we were seeing last time, I decided to run an additional scaling at 1.02 to explore the high M/L territory a bit more.
-
Additionally, I've done a bit of diagnosing of the individual results from each scale on its own to see why things are being driven to one side or the other.
-
Summary:
-
The additional scale = 1.02 models don't seem to appreciably alter anything in our results.
-
I can't seem to get the "best-scales", fiducial case to behavior normally and it's really frustrating and perplexing. The individual 1d panel landscapes look fine (aside from them genearlly missing the minimum for large and small scale values), so it really does not make sense to me why things are behaving as they are.
-
One thing I tried which moved the needle a bit is changing the assumed error on the chi2, but I don't think we have a good reason to do that currently, certainly not to the extent we need to altar our results.
-
One worry I have is that we're interpreting our results largely in the context of what is currently in the paper, which as a reminder uses the incorrect kinematic bins, so I'm really not sure how much we want to read into the differences between the old grid and the new grid. However, this doesn't explain some of the open issues we are seeing in this new grid.
- And to add to this, the indiviual results for the scale = 0.99, 1.0 already look to be fine, but are just discrepant with the past results which we know to be incorrect.
-
-
Next steps:
- I think it might be worth running a single 0.98 scale (to balance the high models?), and maybe one additional round with one more small scaling? The only thing I can think at this pointis that the rho0 landscape looks quite tilted, but this isn't reflected in the cornerplots so I doubt this is really driving things.
- It'd be interesting to see if Emily is getting the same results as I am (I think she will) just as a final sanity check that everything is fine on my end.
- This still leaves open the question about NNLS vs. kinem chi2, and what we want to end up quoting in our results. Again, if we ignore the old models using the old grid and the NNLS chi2/pretend we never saw those, I don't think we'd be as concerned with our current results. I think it's just simply hard to compare the previous models (again, which had the incorrect kinematics) with the new results coming from the corrected kinematics.
- As Emily checks things on her end, I'll return to our open questions on this front.
Plots
- First, here are some 1d panels summarizing the different scales/combined set of models/set of best-scaled models:
| Scale=0.97 | Scale=0.99 | Scale=1.00 | Scale=1.01 | Scale=1.02 | Scale=1.03 | Best Scales | All Scales |
|---|---|---|---|---|---|---|---|
| [images/240812/1d_panels_scale_0.97.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/1d_panels_scale_0.99.png) | [images/240812/1d_panels_scale_1.0.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/1d_panels_scale_1.01.png) | [images/240812/1d_panels_scale_1.02.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/1d_panels_scale_1.03.png) | [images/240812/1d_panels_best_scales.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/1d_panels.png) |
- And the resulting cornerplots:
| K | Scale=0.97 | Scale=0.99 | Scale=1.00 | Scale=1.01 | Scale=1.02 | Scale=1.03 | Best Scales | All Scales |
|---|---|---|---|---|---|---|---|---|
| 40 | [images/240812/scale_0.97_grid_alpha_K40_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_0.99_grid_alpha_K40_nu1.5-1.png) | [images/240812/scale_1.0_grid_alpha_K40_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_1.01_grid_alpha_K40_nu1.5-1.png) | [images/240812/scale_1.02_grid_alpha_K40_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_1.03_grid_alpha_K40_nu1.5-1.png) | images/240812/best_scales_grid_alpha_K40_nu1.5-1.png | ||||
| 50 | [images/240812/scale_0.97_grid_alpha_K50_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_0.99_grid_alpha_K50_nu1.5-1.png) | [images/240812/scale_1.0_grid_alpha_K50_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_1.01_grid_alpha_K50_nu1.5-1.png) | [images/240812/scale_1.02_grid_alpha_K50_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_1.03_grid_alpha_K50_nu1.5-1.png) | images/240812/best_scales_grid_alpha_K50_nu1.5-1.png | ||||
| 60 | [images/240812/scale_0.97_grid_alpha_K60_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_0.99_grid_alpha_K60_nu1.5-1.png) | [images/240812/scale_1.0_grid_alpha_K60_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_1.01_grid_alpha_K60_nu1.5-1.png) | [images/240812/scale_1.02_grid_alpha_K60_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_1.03_grid_alpha_K60_nu1.5-1.png) | images/240812/best_scales_grid_alpha_K60_nu1.5-1.png | ||||
| 80 | [images/240812/scale_0.97_grid_alpha_K80_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_0.99_grid_alpha_K80_nu1.5-1.png) | [images/240812/scale_1.0_grid_alpha_K80_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_1.01_grid_alpha_K80_nu1.5-1.png) | [images/240812/scale_1.02_grid_alpha_K80_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_1.03_grid_alpha_K80_nu1.5-1.png) | images/240812/best_scales_grid_alpha_K80_nu1.5-1.png | ||||
| 100 | [images/240812/scale_0.97_grid_alpha_K100_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_0.99_grid_alpha_K100_nu1.5-1.png) | [images/240812/scale_1.0_grid_alpha_K100_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_1.01_grid_alpha_K100_nu1.5-1.png) | [images/240812/scale_1.02_grid_alpha_K100_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240812/scale_1.03_grid_alpha_K100_nu1.5-1.png) | images/240812/best_scales_grid_alpha_K100_nu1.5-1.png |
- I also tried to mess around a bit with our GPR + dynesty routine, and found the most success in changing the error that we are assuming on the chi2. In the past, I've assumed an error of 0.5 on the chi2, and found little difference changing this around to reasonable values. I re-ran the best-scaled version of the models from above with an assumed error of 0.1 and was able to get at least marginally better-looking results:
