meeting 2024 01 26 n315 - JacobPilawa/TriaxSchwarzschild_wiki

Context

I've run the 1000 models with higher rho0, and have some diagnostics below on the grid from yesterday, this new grid, and the combined grids.
- It looks like the minimum in rho0 is near the middle of our new grid now that we've extended the range of allowed values!
- The shape landscapes are still rather flat, but it's not obvious to me that anything specific is broken.
- One interesting note that I'm looking into is that there seems to be one model that performs exceptionally well, with a chi2 that is ~150 lower than anything else in our grids. I'm not entirely sure why that's the case, other than maybe we got a bit lucky with the specific input parameters? Not sure.
I've also compiled a list of all the input choices I had to make so that we can ensure there are no bugs in what I'm inputting/in the models.

Input Choices

Here are all the input files used in the modeling and the various choices I've made:

aperture_gmos/aperture_mitchell.dat

The only real number that I control in these files are the kin*gridtheta values, which represent the difference between the MGE +x axis and the +x axis of the kinematics.
These numbers are computed as PA_MGE - PA_kinematics:
- The MGE PA comes from Jonelle's paper, which uses 44.31 E of N
- The kinematic PAs come from the files on BOX.
  - For GMOS, these values are listed in /MASSIVE/GMOS_kinematics/N0315/N0315_PA_note.txt, which lists the +x axis of the orbit modeling files as 42 deg E of N.
  - For Mitchell, the numbers are located in /MASSIVE/Mitchell_reduced_data/NGC0315/kinematics_paperversion/unfolded/more_files/NGC0315-s2-unfolded-bininfo.txt, which lists the gal pa as 42.4 deg E of N.
  - This means that the input files should have kin1gridtheta and kin2gridtheta as 2.31 and 1.91, respectively.

bins_gmos/bins_mitchell.dat

I don't make any choices for these files. These are generated when preparing the orbit model files and associated the apertures with specific kinematics.

kindata.dat

This is our normal kindata.dat file, which comes from the Gh8Mh6 minimization without an additive polynomial (adeg=-1) and with a multiplicative polynomial of degree 3 (mdeg=3).
This file contains 12 moments total, 8 of which are measured for the GMOS data and are from my own spectral fits, and 6 moments are fit for Mitchell (which I obtained from the files on BOX).
- I've masked the singular overlapping Mitchell bin by increasing the error on each moment to 10000.
- I've added dummy moments for GMOS h9 through h12 and Mitchell h7 through h12 by assigning all the values to 0.0, and uncertainties identical to the h7/h8 errors for GMOS and h5/h6 errors for Mitchell.

mge.dat

This is taken directly from Jonelle's paper, and is the least masked/processed form of the MGE. Specifically, this is MGE A
Note that the last two components of the MGE are Spitzer IRAC1 MGE components, scaled to match the WFC3/IR data and included as fixed components when fitting MGEs to the HST images.

orbitstart_loop/box.dat

These are just auxillary files which point to other files in our models. No changes or choices.

orblib_loop/box.in

The only specific changes I had to make here relate to the PSFs we assumed for each dataset:
- I obtained the PSF for GMOS from the file /MASSIVE/GMOS_kinematics/N0315/N315_GMOS_PSF.txt, which says that the average weighted PSF = 0.38". The code accepts the equivalent "sigma" for this PSF, which I obtain as sigma = 0.38/(2sqrt(2ln(2))) = 0.161.
- I am using the same PSF for Mitchell as we have used for the other galaxies, which has a sigma = 0.5, which is equivalent to a 1.177" PSF.
Other notable parameters that I kept constant from past orbit modeling:
- innermost mass bin starts at 0.1"
- outermost mass bin starts at 100" and extends to 316"
- 200 dynamical time integration for the box library; 2000 dynamical times for the loop library

parameters_loop/box.in

Many of these parameters are changed on the fly (mass components and angles).
I'm using:
- nEner = 40 for both
- nI2 = 18, 9 for the loop and box library respectively
- nI3 = 9, 9 for both
- logrmin = 10^-2 arcsec (minimum IC radius in arcsec)
- logrmin = 10^2.5 arcsec (maximum IC radius in arcsec)
- interplogmin = 10^-8
- interplogmax = 10^2.6
- gnfw halo with gamma = 0

triaxmass.in

Auxillary file which points to other files.

triaxmassbin.in

Auxillary file which points to other files.
Also contains the same PSF information present in orblib_loop/box.dat.

triaxnnls.in

Mostly points to other files.
I've ensured that this file correctly assumes we have 12 total moments.
I've also checked to make sure that this file reads in both the loop orbit library and box orbit library.

Modeling Results

Notes on the new grid model points

I ran a second set of 1000 dither 1 models, keeping the same ranges in the other parameters but updating the rho0 range to [4e9,7e9] in code units.
- Note that 60 models got caught waiting for open nodes on Expanse, so I went forward without them for now. I can go back and re-scrape the data if we need, but 60/2000 should be a negligible amount of models missing.

Here are the actual model points I submitted for this higher rho0 grid

T Space	Angle Space

Combined Results/Analysis

I've reprocessed our results with the two cubes combined, as well as each cube individually. Here are some of those diagnositcs:
- Note that the dynesty runs for the low rho0 grid in yesterdays page are a bit different from the ones presented here. I still had the jacknife routine I have been running on N57 enabled when I ran dynesty yesterday, so the plots posted on the previous page were splitting the available data into thirds before running dynesty. The results posted here fixed this issue, but the answers or conclusions don't appreciably change because the grid is already so coarse/the minima were a bit missed in yesterday's results.

Cutoff K	Num. Models < min(chi2)+K
10	1
50	1
100	1
250	30
500	175
1000	572
1500	866
2000	1107
5000	1659

1d chi2 vs. parameters

low rho0 grid	high rho0 grid	combined grid

Dynesty Results (nu=0.5)

cutoff K	low rho0 grid	high rho0 grid	combined grid
500
1000
1500
2000

Dynesty Results (nu=1.5)

cutoff K	low rho0 grid	high rho0 grid	combined grid
500
1000
1500
2000

Single model that performs abnormally well...

As you can see in the plots above, there is a single model that had a substantially smaller kinematic chi2 than the other models, by about ~125.
I was curious if the results above substantially change if I remove that single model from the landscape, so here are those results:
First, here's the updated N(less than min(chi2)+K)) without that lowest chi2 model:

Cutoff K	Num. Models < min(chi2)+K
10	3
50	12
100	28
250	87
500	283
1000	658
1500	940
2000	1154
5000	1677

Dynesty Results on the Combined Grid after removing the single best fitting model: note that the nu=1.5, high K cutoff cases are currently getting stuck in dynesty, but I wanted to get this update out quickly so I scrapped them for now. They're still running and I'll let you know if things appreciably change when they're done, but for now I think the K = 500 results give the general picture.