Fitting dispersion too larger, how to fix it

[tpos8]

Why I constrain the parameters gausses scale so small, but its MCMC fitting in-probability result always arise. Even I change the walk-step to 1000 to fit it, the final result always deviate it reasonable value. So please you can help me to resolve this question, thanks!

[kecnry]

In general I think this is often a sign that the chosen parameterization is too restrictive to allow the model to account for the remaining residuals. So I'd start by trying to understand the cause of the spread in the observations by inspecting in time-space - it could be a real physical effect that varies with time (spots etc), an imprecise orbital period or "eclipse" timing variations causing a horizontal shift, or something that can be treated as noise. Depending on the answer to that, you want to introduce the appropriate flexibility into the forward model. I also notice that you're not marginalizing over passband luminosity or third light, so the model doesn't have freedom to scale or move vertically (apart from the inclination) - I'd definitely add least pblum into your final sampling run to make sure any parameter correlations between inclination and luminosity are included in your posteriors.

It also looks like your data uncertainties may be underestimated (if the vertical spread in observations is noise, then that spread is much larger than the individual error bars that are visible). In that case, I'd suggest marginalizing over sigmas_lnf as well

Lastly, it looks like the sampler is still converging/burning in, so the spread may improve after it finds the solution and gets to fully explore the parameter space. I'd suggest using optimizers as much as possible before switching to MCMC and spend the time carefully constructing the initializing distributions so that sampling run is as efficient as possible.

Hope this helps!
Kyle

[tpos8]

Hi Kyle,

Thank you for your reply—your guidance will be very helpful, and I’ll take some time to implement it
. If I run into any issues, I’ll be sure to reach out again for further assistance.
Thanks！
YangYang Lin

[tpos8]

Hi Kyle,

Thank you again for your helpful guidance.
I tried to find information about the sigmas_lnf parameter on the PHOEBE documentation website, but I couldn't locate a detailed description.
Could you kindly point me to a page or resource where this parameter is explained?

Thanks a lot for your help!

Best regards,
YangYang Lin

[kecnry]

See section 3.5 (and included citations) in the PHOEBE 2.3 release paper and emcee's tutorial on fitting a line

[tpos8]

Hello, Kyle， how should I operate to incorporate the masses of the primary star and the companion star into the mcmc fitting

[kecnry]

The error message tells you that it cannot be included in the sampling because it is a constrained parameter, and suggest flipping the constraint (or removing it from sampling). What have you tried?