meeting 2024 10 10 n315 - JacobPilawa/TriaxSchwarzschild_wiki_5 GitHub Wiki
Context
Following up on the work from last time, largely centering around re-minimizing and expanding Grid B:
I've reminimized the original Grid B data points but with the corrected kinematics + new binaries.
Takeaways/Moving Forward: These results are looking great! Grid A and Grid B appear largely consistent with one another (and pretty much consistent to the old results). Currently I had only run the s=1.0 Grid B models, but I can go ahead and add the same scalings we used for Grid A (+/- 0.5%, 1%, 2%, 3%) and continue our processing.
I've also run an additional 3000 dither 1 models covering the low Tminor portion of the space. I'm calling these points Grid Bprime.
Takeaways/Moving Forward: These models are looking really great, and I think Grid Bprime will be a valuable addition to our final set of models. The minima for BH, ML, Rho0, T, Tmaj seem consistent with what we see in Grid A/B, and the added Tminor portion of the space appears to be quite flat which suggests we are actaully finding better fitting models.
For a bit of a comparison, this bullet compares the original Grid B dither 1 vs dither 3 models, which suggests that the dither 1 models are ~436 chi2 better than the dither 3 models. If we take the results posted below for Grid Bprime's dither 1 models and add a similar offset, the chi2 landscapes seems quite comparable to the Grid A/B landscapes we currently have (which have minimum nera ~3100-3200 compared to the ~2700 minimum of the current d1 Grid B prime).
Additionally, I've got the Number of Models vs. K below, which seems to recover what we expected from our meeting -- about ~1000 models are within a delta chi2 of ~350 from the minimum, which I believe was the more conservative cutoff we suggested in the meeting.
Grid B Reminimization
First, here's a quick comparison of both Grid A and Grid B NNLS vs. kinem diagnostics. Note that these only show the base models which have been re-minimized with the new binaries/corrected kinematics.
The trend that we have seen in the past is again present, with the kinem and NNLS tracking each other quite well in Grid B. Nothing really new to takeaway from these plots other than things are tracking each other quite well.
Grid A
Grid B
And here's the joined 1d panels for the base models for Grid A and Grid B (total of ~1500 models). Note that I am plotting only the base models here for simplicity, but my plan is to add additional scalings for Grid B and Grid B prime, but I wanted to make sure the s=1.0 results looked alright first.
It seems like the results are really not too different from what we saw in the original Grid B.
Scale = 1.0 Grid A + B Models
OLD Version of Grid A + B
And for completeness, here are the different chi2 panel (FOR GRID B ONLY) we have been considering (NNLS vs. Kinem, with and without dummies). Nothing new here but just keeping track of our different quantites!
NNLS (No Dummies)
NNLS (with Dummies)
Kinem (No Dummies)
Kinem (with Dummies)
with Scalings (10/15 Update)
I ended up running the scalings for Grid B at our usual values of +/-0.5%, +/-1%, +/-2%, +/-3%, and here are some of those diagnostics put together with the from Grid A + scalings.
As usual, things are coming together nicely and the grids are consistent with one another, generally showing the same trend that we've seen.
Grid A + B All Scales (Ts)
Grid A + B All Scales (UPQ)
Here's the "best-scaled" versions of Grids A + B:
Best Scaled Models
OLD Best Scaled Models
And here are the distribtuions of which models are getting picked as the best/worst scales:
Grid A Only
Grid B Only
Total
For completeness, here are the 1d panels for the Grid B scalings individually (these are the NNLS chi2 without dummies)
scale=0.97
scale=0.98
scale=0.99
scale=0.995
scale=1.0
scale=1.005
scale=1.01
scale=1.02
scale=1.03
And here's a quick comparison showing Grid A and Grid B, along with the best scalings, so that we can get an idea of how much we improve by adding our scalings. Roughly speaking, it looks like we add ~50% of the coverage within a Delta Chi2 of ~300 or so.
Plot
Zoomed
Grid Bprime (Dither 1 Models)
I've run a set of models which should complement the existing Grid B points, but adds the portion of the space at low Tmin that we had initially trimmed from our parameter space. We had initially trimmed out the low Tminor portion of the space because it didn't seem the chi2 favored this area, but then our original Grid C pulled us back to that corner.
Here's the points I submitted for this new grid. I chose 3000 models (from our estimates in the meeting) to run with dither 1, with the hopes that we rerun a subset of the well-fitting models here. The key part of the parameter space that we're covering here is Tmin = [0.001, 0.4] which had been excluded from our original Grid B.
Plot
And here are the main results, showing the 1d panels for this newest d1 grid and the number of models as a function of K:
1d Panels (No Zoom)
1d Panels (Zoomed)
Models vs. K
One last plot: Here's our current Grid A + B 1d panels from above, but I've added a set of green points here. These green points are the dither 1 results after applying the fitting function from this bulletpoint. The result after applying the formula is remarkably consistent with the Grid A and B dither 3 grids.
