HinLeung622
commented
4 years ago @grd349 New runs, exactly the same setup as above except for using beta(10,10) priors in place of the old beta(1.1,1.1) ones, and lowering the number of tuning steps from 7000 to 2000 since it takes much longer to run on colab.
-
100 M67 dwarfs that are estimated to have mass 0.8-1.35Msun:
3 chains agree with each other and 1 chain does not. However, the age, Y and MLT are estimated to be much lower than previous results in this run, which seems to suggest some problems happening within the sampling process. -
100 M67 dwarfs that are estimated to have mass 0.8-1.2Msun:
Much better sampling here, and the estimated values are slightly higher, closer to previous estimates.
Since the sampler in the higher mass sample did not technically converge, this test still might not be a strong enough test regarding the effect in the fundamentals' estimation uncertainties of samples with and without the 1.2-1.35Msun group. However, I think it is safe to say that the sampler has a much better time working on the lowest 0.8-1.2Msun mass stars.
(Also, maybe these results suggest dwarfs-only HBM runs result in lower age estimates than runs with subgiants (ie. back during our report stage)? Probably not enough evidence about that yet.)
grd349
Interestingly, the posterior has shifted to lower age and helium compared to the last successful run, despite having all four chains consistent with each other. Though the estimations are still within 2 sigmas of the previous estimation values, so I guess it still passes as simple variations in measurements.
Again, similar to the results above, much lower age and Y estimations than previous results. A reassuring finding is that the estimation uncertainties on all 4 fundamentals decreased thanks to the larger sample size. So perhaps it really is the case that only running on low mass dwarfs cause lower age and Y estimations?
For reference, this is our estimates previously using subgiants+main sequence:
And literature values:
First of all, great convergence, but the same situation seems to be happening, the age and Y estimations seems to be lower than previous estimations and literature values.
Note: for the last plot, only 100 random stars were picked for the sampling, so not all blue dots are sampled for, hence not having corresponding red dots.
So it seems the mixture model is working for binaries. But just as before, it seems the mean age is even lower and MLT higher than the previous runs without binaries.
Looks like it is doing rather well. The q value came out suspiciously close to the testval 0.7. The same situation occurs as in previous dwarfs-only runs, where the age and Y estimates comes out to be much lower than literature and our report sub-giants+dwarfs values. Here is CMD of the data and the HBM true_mG + true_Bp_Rp estimates:
I guess it looks reasonable? I expected the sequence of HBM true values to be at the bottom edge of the clusters of data, but it seems to be in the middle currently. Do you think this is expected behaviour?
Not sure if the results make any sense, since delta_mG came out to be much lower than the number we previously used, which seems to have caused a lower q fraction as well. Maybe the HBM is mistakenly treating some of the higher spreaded singular stars as multiples? Could this suggest I shouldn't be allowing delta_mG as a free variable?
The mean estimations for age and Y went back up to acceptable regions, and the mean feh went back down to 0.10 levels. The only problem is that q is of a super low value (~80% of the stars are doubles??) while delta_mG is basically 0. It seems this particular HBM run is unable to treat both the partial pool and mixture models, so the spreads in the partial pool "accounted for" the more luminous doubles, leaving delta_mG to be basically 0.
Basically the same thing is seen in this plot. Though curiously, the sequence of red dots (HBM mean guesses) and the MS of the data (blue) don't run parallel to each other. Is this possibly a sign of the stellar evolution code failing at lower masses?
So it was a success in terms of sampling, the three distributions are very distinct. The HBM guesses are also more aligned with the high magnitude edge of the data, which was one of the concerns you raised previously. However, the super low helium and mean age are still present.
As you can see, the helium and age are still very low with just the main singular dwarfs. So perhaps is a problem with the data/data reduction?
And for these fundamentals:
Age = 2.5 Gyr, M = 0.8-1.2 Msun, feh = 0.1, Y = 0.24, MLT = 2.1 (basically what our current max pool mixture models are predicting)
Both red regions overlap nicely with the blue data dots, meaning it is reasonable that both satisfies the data in the sampler's logic. The only major difference between the two is the first combination covers a more luminous region than the second, meaning the stars would be estimated to have lower masses than the second.
There isn't much of a binary band to speak of
@grd349 The last issue thread was getting very long so I opt for starting a new one, since we are (kind of) pass that ravel equals to zero error.
Last night I ran 2 HBM runs:
Max pool model with 100 M67 dwarfs taken just under the hook:
The red large dots represent a "max pool" generated cluster with masses from 0.8 to 1.2, which means this dataset includes stars with mass > 1.2 Msun. The prior on mass hence runs from 0.8 to 1.5
g mag obs error = x100, tuning steps = 7000, sampling steps = 2000
Sampler not converged, and oddly, the mean value results from this max pool model are somewhat different from the partial pool and the ones we got for the report, especially for Y.
Max pool model with 100 M67 dwarfs taken considerably lower luminosity under the hook:
Red dots are of the exact same generated cluster set up, so this sample should be with stars only with mass 0.8-1.2 Msun. The HBM model for this run is identical to the last except for shrinking the mass prior to 0.8 to 1.2
g mag obs error = x100, tuning steps = 7000, sampling steps = 2000
Somehow rather worse sampling results, and very strangely, the posteriors on mass goes up above 1.2. Maybe I made some mistake that I am yet to be able to locate, but I am not sure what happened there.
This test was meant to check if discarding the >1.2Msun dwarfs would result in a much worse estimate in the cluster-wide fundamentals, but the samplers are still having a hard time sampling, so I don't think the test answers the question yet.
I have started an additional run identical to no.2 except for using beta(10,10) on all priors, will report when that gives results.