Estimated burn in steps is zero

Use of Healpix, camb, CLASS, cosmomc, compilers, etc.
Shu-Rong Chen
Posts: 7
Joined: August 12 2015
Affiliation: National Taiwan University

Estimated burn in steps is zero

Hello,
I am running Cosmomc as a generic sampler using both single and multiple chains.
When I use getdist to check the data convergence, I get "RL: Estimated burn in step:0 (0 rows)" for both single and multiple chains (see below).

RL: Thin for Markov: 80
RL: Thin for indep samples: 81
RL: Estimated burn in steps: 0 (0 rows)
mean input multiplicity = 3.89792
using 11384 rows, processing 3 parameters
Approx indep samples: 548

I'm not sure whether it's reasonable to have zero estimated burn in.
Does anyone know how to fix it? Thanks so much in advance!

Antony Lewis
Posts: 1646
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: Estimated burn in steps is zero

Sounds odd, is this the python getdist? What are the corresponding RL stats in your .converge file?

Shu-Rong Chen
Posts: 7
Joined: August 12 2015
Affiliation: National Taiwan University

Estimated burn in steps is zero

I use Fortran getdist. The problem happens in both Feb 2015 and Oct 2012 versions.
The RL stats in .converge file are

Code: Select all

chain   markov_thin  indep_thin      nburn
1          80           81           0
2          50           51           0
3          42           43           0
4          52           53           0 

Antony Lewis
Posts: 1646
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: Estimated burn in steps is zero

OK thanks - can you try the python getdist for comparison? (which is replacing the Fortran one). If it gives the same, please email the chain files and I can have a closer look.

Shu-Rong Chen
Posts: 7
Joined: August 12 2015
Affiliation: National Taiwan University

Estimated burn in steps is zero

I find that the issue of zero burn-in depends on whether using the .paramnames file or not in the Fortran GetDist (Feb. 2015 version).
My previous post (which gives zero burn-in) dose not use the .paramnames file, but instead uses the following settings inside the distgeneric.ini:

Code: Select all

nparams=3
auto_label = T
#columnnum = 0
parameter_names =

When I use the .paramnames file, I get non-zero nburn and different indep_thin (see below).

Code: Select all

 chain  markov_thin  indep_thin    nburn
1          80          81         560
2          50          51         350
3          42         164         336
4          53          60        1248
I also try the Python Getdist (Jul 2015 version) and it also shows full Raftery & Lewis statistics as follows, but inde_thin and nburn are again different from the previous Fortran Getdist results.

Code: Select all

chain  markov_thin  indep_thin    nburn
0          80          80        560
1          50          51        350
2          42          59        294
3          52          60        728
Are these results reasonable? or I use something wrong?

In addition, I found that the marginal statistics results are the same in all cases, regardless of the difference in indep_thin and whether burn-in is zero or not. Is that normal?
Thanks a lot

Antony Lewis
Posts: 1646
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: Estimated burn in steps is zero

The convergence tests are independent of almost anything else. If things look OK with the latest python version all is probably fine.