meeting 2025 03 03 n315 - JacobPilawa/TriaxSchwarzschild_wiki_5 GitHub Wiki
Context
- Here's some diagnostics relating to pPXF/my fits vs. Emily's fits/the updated pPXF code, and where we go from here.
Diagnostics
- I started by obtaining the most up to date version of the pPXF fitter, and I re-ran our code with the settings we discussed at the end of last week.
- There's a few things I noticed in doing the re-reduction:
- My first attempts with the new code were a bit strange and the results were not in perfect agreement. Specifically when I was re-running things on my end with the new code, about ~1/3 of my spectra were....very strange. They had very, very large sigma values which caused h4/h6 to be all wonky, too. But strangely this only impacted a subset of all the spectra.
- It turns out that the default RunPPXF() call has the bias parameter set to bias=0.2, whereas I've always had bias=0.0. When I update this value to be bias=0.2, I can reproduce Emily's preliminary fits exactly:
- What's very strange about this is that my old code with bias=0.0 did not have these issues when fitting the spectra/didn't result in such high sigmas for some of the bins. Basically, I'm not sure why changing bias to 0.0 is resulting in such a large swing in the kinematics, and why this is so different from the previous results I obtained with that bias value.
| Bias = 0.0 Reproduction of Emily's Fits (Prelim Fits) | Bias = 0.2 Reproduction of Emily's Fits (Prelim Fits) |
|---|---|
| [images/250303/bias_0p0_repro.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250303/bias_0p2_repro.png) |
- I was now interested in how the kinematics change when changing from the Barth stars to the stars that I used when also keeping track of the bias parameter. I mostly want to ensure that the new code can reproduce my old code + compare how change away from the Barth stars to my stars impacts our kinematics.
- These plots compare preliminary fits with the new code to the preliminary fits currently in the paper, and thus using the old code. It's encouraging that with the new code I can reproduce my old preliminary fits very well regardless of the bias parameter I am assuming.
| Bias = 0.0, Off-by-one library (Reproducing Myself) | Bias = 0.2, Off-by-one Library (Reproducing Myself) |
|---|---|
| [images/250303/mystars_0p0.png]]](/JacobPilawa/TriaxSchwarzschild_wiki_5/wiki/[[images/250303/mystars_0p2.png) |
Comparing A Few Cases
-
Now that I seem to be able to reproduce both Emily's results and my results with the preliminary spectral fits, I wanted to put up a few comparisons of the MC kinematics rather than just the preliminary fits.
-
First as a sanity check, here's a one-to-one comparing my fits to Emily's fits from the end of last week with the exact same settings/stars:
- Thankfully, the results are virtually identical to one another.
| Reproducing Emily's Fits |
|---|
| images/250303/mc_barth_0p2_bias.png |
- And here's a comparison of my fiducial kinematics from the paper and their reproduction with the most up to date code. The only difference (aside from code versions) is that the new code uses nMC=100. I've already checked above that the fiducial values agree well, so I think any noise here is real in the sense that the scatter doesn't appear much bigger than the version above comparing fiducial fits.
| Reproducing My Original Kinematics |
|---|
| images/250303/mc_barth_0p0_bias_original_vs_repro.png |
So where does this leave us?
- I'll put this plot here for quick reference, which compares on a one-to-one plot the "correct" kinematics from Emily's fits (which use bias = 0.2 and the Barth Table 2 library) compared to my reproduced fits (which use bias = 0.2 and the off-by-one library).
| Emily's Final Version (barth stars, bias = 0.2) vs. My Final (off-by-one stars, bias = 0.2) |
|---|
| images/250303/final_final.png |
-
So what's the long story short of the results above?
-
I've got the latest RunPPXF() scripts from Github, and am able to reproduce Emily's results from the end of last week.
- Notably, I've used a bias of 0.0 for all my fits up to now, and it seems like Emily used 0.2 (the default in the RunPPXF() script at the end of last week) in her reductions. Strangely, when I rerun the spectra with bias = 0.0 vs. bias = 0.2, a good deal of my spectral fits with the new code favor extremely high sigmas/push some of the higher order moments to the boundaries, whereas others are essentially in perfect agreement. I am quite puzzled while this is the case and am still trying to track that down.
-
I'm also able to reproduce my fiducial fits with the new code, and the results with the INCORRECT stars are less sensitive to the choice of bias parameter. Even though I am able to reproduce these, they are slightly offset from Emily's reproduction due to the different library being used (hers using Barth stars, mine using the off-by-one indexed stars).
-
-
The net result is that we really have two sets of kinematics which I've plotted against each other just above:
- My original kinematics, which I can reproduce with the newest RunPPXF() script with either bias=0.0 or bias=0.2 (the results are not sensitive to this value) and ensuring that I use the off-by-one set of stars instead of the Barth stars.
- Updated kinematics which are produced with the Barth library and the bias parameter set to 0.2. This is essentially the kinematics that Emily had produced at the end of last week. There are a few kinematic bins in this reduction though that are against the boundary/favor very large and unphyiscal sigma which might end up throwing things off in the fits, but I can always mask these bins if we think they're problematic.
-
So moving forward, I think I should reminimize two sets of kinematics (the two plotted against each other above) like we've been discussing recently and hope things don't appreciably change around...:
- In one set of the kinematics, I'll keep everything the same as our current results, but I'll replace our current uncertainties with the uncertainties from the nMc=100 case (new code, off-by-one library, bias=0.2 for consistency across cases). This test would be simply to see how much of an impact our uncertainties are having on our parameter constraints.
- The second set of kinematics would be the same set that Emily produced/that I was able to reproduce with bias = 0.2 and the Barth library. I'm really not sure what to expect from this test, other than a slim hope that things "move around" in such a way that our current results are not appreciably changed. Because there's a slight tilt between these kinematics + my fiducial kinematics, my gut says that this might not be so clean.