MCMC posterior distributions problem (NM fit perfect)

[gokhannY]

I have a weird problem. My NM fit is perfect for finding appropriate parameters, but MCMC posteriors "hit a wall" as described in other discussion

This is the NM fit results
nmfit.pdf

This is the MCMC fit results
lcrv.pdf

Somehow mass ratio and asini posteriors go to wrong direction and stuck in there, hence hitting a wall.

Here is the failed posteriors
failed.pdf

and lnprobability graph

lnprob.pdf

What can I do to fix this?

Here is the NM fit results

    requiv@primary@star@component: 4.443249539070683 solRad
   teffratio@binary@orbit@comp...: 0.6957623787199289
      incl@binary@orbit@component: 70.22895828050298 deg
       pblum@primary@V@lc@dataset: 6.8733465255469035 W
         q@binary@orbit@component: 0.4671437731594412
     asini@binary@orbit@component: 15.992735820875236 solRad

Lastly this is the MCMC graph
abc123.pdf

Forgot to mention that this is a semi-detached system with secondary filled Roche.

[amiszuda]

In general, whether using PHOEBE or another code, the key to understanding the dataset-dependent values of q and asini lies in the data itself. While these parameters primarily manifest in radial velocities (RVs), they do influence light curves—albeit to a lesser extent. Given the significantly larger quantity of photometric data compared to RVs, as well as the respective weighting of each, it’s natural for the photometric solution to have a stronger influence on these values. I’m not certain if PHOEBE automatically handles such cases by assigning different weights to balance the impact of each dataset on the respective parameters (@kecnry @aprsa), but in my experience, fixing q and asini to their RV-determined values and adding small priors during the MCMC process does the trick.

[gokhannY]

I agree with you. That's why my Gaussian around 0.1 and 1 for q and asini, respectively. I might try 0.02 for q, but the MCM calculation takes almost five days for this system.

I also attached the trace graph. The starting points are all over the place. The posteriors just stuck for some values for q and asini.

I might try to solve each data, rv and lc, independently in MCMC but that wouldn't be scientific.

trace.pdf

[kecnry]

I agree with @amiszuda - you have a few options within phoebe to address this:

• find what other parameter(s) might be able to account for the LC residuals that are "incorrectly" driving asini and q. This may or may not be possible depending on the situation.
• account for potential underestimation of the uncertainties on the photometry points, using sigmas_lnf (I guess we don't have a great tutorial on this still, but it is explained in the 2.3 release paper). I'm not sure it will completely avoid the problem you're seeing, though.
• do an MCMC run on the RVs alone and then hold q and asini fixed for the MCMC run on LCs+RVs. You can flip the constraint so that the second run can still vary the inclination while conserving asini.
• better yet (in my opinion), do an MCMC run on the RVs alone, and use those posteriors as the priors on the following MCMC run on LCs alone. This effectively treats the information from all the RVs equally as the information from all the LCs.
• use a custom cost function to manually override the weights of each point (I'm not a fan of this one, but it is technically possible).

Whatever you end up doing, since this is somewhat "non-standard", I think its important to be very clear in any publication of exactly what you did and why so that the results are reproducible and taken in the correct context.

[kecnry]

(I also converted this from an issue to a discussion so that people could find it easier - closed issues can be easily lost)

[gokhannY]

Thanks Kyle. I used the advice of @amiszuda and started the MCMC run for asini and q with smaller distributions as 0.5 and 0.02, respectively.

I believe your 4th suggestion is the best one among them.

https://phoebe-project.org/docs/2.4/tutorials/emcee_resampling

I believe this tutorial would satisfy my needs, if I'm not mistaken.

[gokhannY]

Firstly, sorry for delayed reply, but doing MCMC individually for RVs and LC did not work. As can be seen from the figures I attached.

This is for RVs alone and it's good as expected.
RV.pdf

This is for LC included, as you can see q and a goes to different way.
RV_LC.pdf

And this is the trace figure for each parameter.
RV_LC_trace.pdf

The weird thing, when I do the same analyses in Phoebe 1.0, everything is perfect. But I do not want to use Phoebe 1.0 just because it gives the expected result.
Phoebe1.pdf

I'm also attaching the code I used for secondary MCMC analyses that I included LC.

import phoebe
import numpy as np

b = phoebe.load('result_RV.bundle', import_from_newer=True)

b.enable_dataset('V')

b.set_value('lnprob_cutoff', solution='round_1', value=145) 

b.adopt_solution('round_1', adopt_values=False, adopt_distributions=True, distribution='resample')

b.set_value('init_from', solver='mcmc_last', value='resample')

b.add_distribution({'teffratio@binary': phoebe.gaussian_around(0.1),
    'requiv@primary': phoebe.gaussian_around(0.5),
    'incl@binary': phoebe.gaussian_around(2.0),
    'pblum@primary': phoebe.gaussian_around(2.)
    }, phoebe.gaussian_around(0.5), distribution='init_from_new')
    

b.set_value(qualifier='init_from', solver='mcmc_last', value=['resample', 'init_from_new'])

b.run_solver('mcmc_last', solution='round_2')

b.save('result_RV_LCs.bundle')

I'm guessing I do something wrong when I prepare the file.

[kecnry]

Sorry for the confusion - for this approach you'll want to apply the posteriors from the RVs as priors (not init_from - although using them for both would be reasonable) during the LC run.

[gokhannY]

Finally fixed, thanks for help guys. All I had to do was change this

b.set_value('init_from', solver='mcmc_last', value='resample')

to this

b.set_value('priors', solver='mcmc_last', value='resample')

Last.pdf

Hope, this would help to someone else.