[CosmoMC] - convergence is not achieved after ~ week

Use of Healpix, camb, CLASS, cosmomc, compilers, etc.
Post Reply
Ira *Wolfson
Posts: 64
Joined: January 24 2013
Affiliation: BGU
Contact:

[CosmoMC] - convergence is not achieved after ~ week

Post by Ira *Wolfson » June 20 2017

Hi,

I am currently running 3 instances of CosmoMC with the following .ini configuration:

Code: Select allp>

#general settings
#Bicep-Keck-Planck, varying cosmological parameters
DEFAULT(batch2/BKPlanck.ini)

#general settings
DEFAULT(batch2/common.ini)

#e.g. to vary r in addition to standard 6:
#(for r>0 also need compute_tensors=T)
compute_tensors = T
param[r] = 0.03 0 2 0.04 0.04
param[tau] = 0.07

#high for new runs
MPI_Max_R_ProposeUpdate = 30

propose_matrix= planck_covmats/base_TT_lowTEB_plik.covmat

#Folder where files (chains, checkpoints, etc.) are stored
root_dir = chains/TEST/

#Root name for files produced
file_root=test1
#action= 0 runs chains, 1 importance samples, 2 minimizes
#use action=4 just to quickly test likelihoods
action = 0

#expected result for -(log like)
test_check_compare = 28.337

num_threads = 0

#if you want to get theory cl for test point
#test_output_root = output_cl_root
[b]
start_at_bestfit =F #T[/b]
feedback=2
use_fast_slow = T

checkpoint = T
stop_on_error=F
#sampling_method=7 is a new fast-slow scheme good for Planck
sampling_method = 7
dragging_steps  = 3
propose_scale = 2

#Set >0 to make data files for importance sampling
indep_sample=0

#these are just small speedups for testing
get_sigma8=T

#Uncomment this if you dont want one 0.06eV neutrino by default
#num_massive_neutrinos=
#to vary parameters set param[name]= center, min, max, start width, propose width
#param[mnu] = 0 0 0 0 0

The difference between the 3 runs are the following:
2 runs on CosmoMC-master - one that starts at best fit and one that doesn't.
and 1 run on CosmoMC-Nov2016 that doesn't start at best fit.


All of these runs, at this point are running for over a week. And some for more.

The 2 runs that don't start at best fit show convergence statistics, at the level of ~0.2

All previous tests (i.e. test.ini, test_planck.ini) were successful.

Any idea what might be going wrong?

It seems as maybe there is some noise that prevents the MC from settling into the minimum point of the underlying data's phase space.

Also I have hunch it might have to do with the optical depth?

– Ira

Antony Lewis
Posts: 1352
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: [CosmoMC] - convergence is not achieved after ~ week

Post by Antony Lewis » June 20 2017

It's probably because you have large numbers of unconstrained standard LCDM parameters ([tex]n_s[/tex], [tex]\Omega_b h^2[/tex], etc), since you are not including Planck high L.

For BKP only runs there's a supplied batch2/BKPlanckonly.ini (likewise for newer BK14).

Post Reply