meeting 2024 08 08 n57 - JacobPilawa/TriaxSchwarzschild_wiki_5 GitHub Wiki
Context
- I've been trying to make sense of the NNLS and kinem chi2 questions we have, but got caught down some side issues that I am still trying to figure out. In this bullet, I'll post some of these diagnostics with what I've been able to make sense of:
- I've re-minimized the models we have with four different scale factors (0.97, 0.99, 1.0, 1.01, 1.03), with 1.0 being the base models we've been working with. This brings the total number of re-run models from ~2000 models to ~10000.
- It seems like the additional sampling is helping out our recovery and making things more consistent with our original results, but I'm still sorting out a strange quirk. For some reason that I'm still trying to figure out, the GPR + dynesty routine seems to be pushing the M/L minimum to the edge of the grid, despite there being a pretty obvious minimum in the 1d panels.
- I think that one thing I'd like to try is do one more >1.0 scale factor to improve the sampling at high M/L. To me, the 1d panel showing the best scaled version of the models seems to have a bit of an imbalance in the M/L distribution, and I think the lack of points at high M/L might be driving this strange behavior.
- It's really weird because the 1d base model panels look very, very similar to the best-scaled selection, but they favor totally different locations in the 6D parameter space. It's really mind-boggling that simply adding in the scaled models is causing GPR and dynesty to behave so strangely?
Plots
- First, here's are the 1d panels for all the scaled models (red, orange, gold, green, blue) maps to (0.97, 0.99, 1.0, 1.01, 1.03). The black, smaller points are the "best-scaled" versions of the models. I've included one other plot showing the base models (scale = 1.0) compared to the "best-scaled" version as well. I want to emphasize how similar the 1D panels are in the "best-scaled" and "base" case comparison, because these points are what go into our GPR + dynesty:
| All Scales | Base vs. Best |
|---|---|
| [images/240808/1d_panels.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240808/1d_panels_best_base.png) |
- Feeding these exact points into our GPR and dynesty routine, I obtain the following cornerplots. It makes no sense to me why the GPR and dynesty is favoring the high M/L portion of the space. The only thing I can think of that could have an impact is the relative paucity of points at high M/L, and thus maybe GPR and dynesty is extrapolating strangely? I'm still working this out, though.
| K | Base Models | Best Scaled Models |
|---|---|---|
| 40 | [images/240808/base_grid_alpha_K40_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240808/best_scales_grid_alpha_K40_nu1.5-1.png) | |
| 50 | [images/240808/base_grid_alpha_K50_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240808/best_scales_grid_alpha_K50_nu1.5-1.png) | |
| 60 | [images/240808/base_grid_alpha_K60_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240808/best_scales_grid_alpha_K60_nu1.5-1.png) | |
| 80 | [images/240808/base_grid_alpha_K80_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240808/best_scales_grid_alpha_K80_nu1.5-1.png) | |
| 100 | [images/240808/base_grid_alpha_K100_nu1.5-1.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/240808/best_scales_grid_alpha_K100_nu1.5-1.png) |
Additional Diagnostics
- More to come soon, but have been working through the dynesty/GPR from above.