Short Period High Eccentricity After MCMC

[ella918]

Hi I'm a student working with @stephtdouglas measuring periods for binaries. For stars of short periods, the original rejection sampling (without MCMC) is resulting in short eccentricities which is what we are expecting. However, when MCMC is done on stars with 1 rejection sample and with short periods, the orbits are being pushed towards higher eccentricities. This is occurring on 7 stars in our sample.

Below is an example of this, a star with a Gaia ID of 2128124963389008384.

This is the period vs eccentricity plot after rejection sampling

This is the period vs eccentricity after MCMC

Here is the script I am using
https://github.com/ella918/Summer-Research/blob/main/RunTheJokerOnePrior.py

[adrn]

👋 Could you share a plot of the radial velocity data vs. time? If that's proprietary, you can email me instead of posting here!

[ella918]

Here is a plot of the RVs vs time, thanks!

[adrn]

Hm ok, nothing looks immediately strange about the RV data! One quick question: are the RV uncertainties very small? What are the typical values?

Have you checked the convergence statistics for the MCMC run? Does it look like that ran successfully? For example, if you print az.summary(mcmc_samples) in your script, what does that table look like? (and maybe try running with chain=4 for better statistics). Another thing to check is to increase the number of prior samples you use in TheJoker to more like 10 million.

[ella918]

Hi sorry for the delay! I was able to create a MWE for this issue and attached it to this comment along with the output files when running it with 50 million samples and 8 chains. Please let me know if you need anything else. Thank you!

https://github.com/ella918/Summer-Research/blob/main/MWE_1.py
https://github.com/ella918/Summer-Research/blob/main/MWE_RVData.csv (data that I used)

Period vs Eccentricity after rejection sampling

Period vs Eccentricity after MCMC

Trace plot

[nogroundNJU]

Hi there! I have encountered the same problem, eccentricity will become larger after MCMC, but I think the MCMC results didn't converge very well

the eccentricity in the returned rejection sample was 0.05, and I used 50M priors and 20 chains when doing the MCMC

[ella918]

We believe that the problem might stem from the version of arviz, so we decided to test that ourselves. I tried to create a new environment using the more up-to-date version of arviz (0.20.0), but was not able to get all of the dependencies for thejoker to install properly. I had to pip install thejoker without dependencies and still got an error when trying to download pymc. The error comes from pkg_resouces in setuptools (version 75.8.0) and part of the error message is pasted below.

AttributeError: module 'pkgutil' has no attribute 'ImpImporter'. Did you mean: 'zipimporter'?

This environment has numpy version 2.2.3, astropy version 6.1.4, and h5py version 3.12.1. All downloaded using pip version 25.0.1.

@stephtdouglas thinks you might have to update the dependencies because we could not figure out the right combination. Thanks!

[stephtdouglas]

Updating to note that at least a partial fix can be achieved by adding init='adapt_full' to the pm.sample call. (Some of our stars are still having this issue, but it's likely data-related. )