Convergence of chains when a parameter has a one tail distribution

Use of Healpix, camb, CLASS, cosmomc, compilers, etc.
Post Reply
Daniel Francisco Boriero
Posts: 16
Joined: March 09 2009
Affiliation: Bielefeld

Convergence of chains when a parameter has a one tail distri

Post by Daniel Francisco Boriero » December 08 2016

Dear All,

I am testing a model with exponential parametrization of the reionization and I am facing the problem to run chains on cosmomc with a parameter (the exponent) which has a one tail distribution.

Do you have any advice to help the convergence of the Monte-Carlo? I would like to test a few decades of this parameter, but the distribution becomes flat quickly and it is taking forever to reach convergence.


Thanks a lot,
Daniel

Antony Lewis
Posts: 1369
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: Convergence of chains when a parameter has a one tail di

Post by Antony Lewis » December 08 2016

Re-parameterize, e.g. using the log of that variable as the MCMC parameter? (though note that this also changes the implicit flat prior).

Daniel Francisco Boriero
Posts: 16
Joined: March 09 2009
Affiliation: Bielefeld

Convergence of chains when a parameter has a one tail distri

Post by Daniel Francisco Boriero » December 08 2016

Indeed, the log would work, but since the parametrization chosen has some physical meaning, I would have to fix the prior to transform for a flat prior on the original parameter.


Thanks Antony,
Daniel

Daniel Francisco Boriero
Posts: 16
Joined: March 09 2009
Affiliation: Bielefeld

Convergence of chains when a parameter has a one tail distri

Post by Daniel Francisco Boriero » December 13 2016

Dear Antony,

To fix the prior, would you sugest to use the function AdjustPrior(), on GetDist?

Moreover, is it still enough to process the chains only using these lines below?

! coldata(1,i) = coldata(1,i)*exp(-chisq/2)
! coldata(2,i) = coldata(2,i) + chisq/2


Thanks,
Daniel

Antony Lewis
Posts: 1369
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: Convergence of chains when a parameter has a one tail di

Post by Antony Lewis » December 13 2016

I think that should work. (you may also want to adjust the chi2_xxx derived parameters, if you use them for anything)

Daniel Francisco Boriero
Posts: 16
Joined: March 09 2009
Affiliation: Bielefeld

Convergence of chains when a parameter has a one tail distri

Post by Daniel Francisco Boriero » December 13 2016

Ok, I am only interested in the total chi2 anyway. I was concerned that these new chi2_xxx derived parameters had changed the meaning of the first two columns, making the AdjustPrior() deprecated.

Thanks for explaining.


Cheers,

Post Reply