MCMC chain running for very long.

Use of Healpix, camb, CLASS, cosmomc, compilers, etc.
Post Reply
Shouvik Roychoudhury
Posts: 21
Joined: August 14 2016
Affiliation: Harish-Chandra Research Institute

MCMC chain running for very long.

Post by Shouvik Roychoudhury » April 06 2017

Hi,

I am using the Nov 2016 version of CosmoMC. I am facing a weird problem I have not faced before. The problem is my MCMC chain is running for more than 1 month (usually, earlier, it got finished within 5−6 days). R−1 (quoted in the .converge_stat file in the chains folder) goes below 0.01 (takes more than 1 week to do that) but fails to reach the MPI_Limit_Converge_Err for at least one of the parameters (I am using the default convergence settings for batch2) and then starts increasing and goes as high as around 0.2 and starts decreasing again but very slowly and has not converged yet. (R−1 > 0.01 still). This has happened with various data set combinations but same parameter set.

This doesn't happen when I am using base LCDM model. Happens only when I am adding a parameter. So my guess would be that the cause is somehow linked to the parameter. Can it be due to the fact that there is no related covmat file available in the planck_covmats folder for the new parameter and a bad starting covmat (I am using base_TT_lowTEB_plik.covmat)? What can be a remedy?

Any help is appreciated.

Thanks,
Shouvik

Antony Lewis
Posts: 1354
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: MCMC chain running for very long.

Post by Antony Lewis » April 06 2017

It could just have a weak secondary mode, which is very hard for MCMC to explore. If so you may be able to see it in your current posteriors.
Certainly no point running it for that long! If it's not converging it probably won't.

Shouvik Roychoudhury
Posts: 21
Joined: August 14 2016
Affiliation: Harish-Chandra Research Institute

MCMC chain running for very long.

Post by Shouvik Roychoudhury » April 06 2017

By a secondary weak mode, do you mean a small secondary peak in the 1-D posterior?

Antony Lewis
Posts: 1354
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: MCMC chain running for very long.

Post by Antony Lewis » April 07 2017

in the N-D posterior. Could also be wide low-amplitude tail, or something more complicated.

Post Reply