CosmoMC: Importance sampling

Use of Cobaya. camb, CLASS, cosmomc, compilers, etc.
Post Reply
gongbo zhao
Posts: 73
Joined: January 04 2005
Affiliation: NAOC
Contact:

CosmoMC: Importance sampling

Post by gongbo zhao » August 29 2005

Dear all:
I have set action = 0 and indep_sample = 3 (is it reasonable?)in the params.ini and generated the .data and .txt chain files using WMAP and SN data. Now I want to add SDSS data to tighten the constraints. What I did is set action =1 and
redo_likelihoods = T
redo_theory = T
redo_cls = T
redo_pk = T
redo_skip = 0
redo_outroot = blah
redo_thin = 1
#If large difference in log likelihoods may need to offset to give sensible weights
#for exp(difference in likelihoods)
redo_likeoffset = 0

I am puzzled when I check the output .txt file. The first column (weights) is of order 10^-7 or smaller . It is buggy when I getdist . Did I make any mistakes ? Thanks in advance.

ps. I have attempted to set redo_likeoffset = 10 but the resulted "weight" in .txt is still weird.

Cheers,
Gongbo Zhao

Antony Lewis
Posts: 1936
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: CosmoMC: Importance sampling

Post by Antony Lewis » August 31 2005

Unfortunately cosmomc doesn't at the moment automatically compute a rescaling so the weights are sensible. You have to set redo_likeoffset so that the weights come out of order unity for sensible models. Since the weight is exp(-\chi^2/2), the redo_offset should be around the number of independent new data points times 2.

(you expect some very small weights - those models ruled out by the new data - but for numerical reasons you want good fit models to have weights of order unity).

indep_sample=3 is OK, though importance sampling will be faster if you increase it. Small numbers also increase the size of the .data files that need to be stored. Generally you want the number to be somewhat less that the correlation length of the chains.

gongbo zhao
Posts: 73
Joined: January 04 2005
Affiliation: NAOC
Contact:

CosmoMC: Importance sampling

Post by gongbo zhao » September 01 2005

Dear Antony:
Thanks for your help. Now I am adding SDSS data . is it right to set redo_offset = 40 ? I have tested it, the weight is huge (of order 10^13 ) ,but I can getdist the chain files and get correct results . Would you please give me some further comments ? Thanks again.

Cheers,
Gongbo Zhao

Post Reply