Significance of Centre, Starting width and Propose width in CosmoMC 2016

Use of Healpix, camb, CLASS, cosmomc, compilers, etc.
Post Reply
Shouvik Roychoudhury
Posts: 31
Joined: August 14 2016
Affiliation: IIT Bombay

Significance of Centre, Starting width and Propose width in

Post by Shouvik Roychoudhury » October 09 2016

I would like to understand the effect of different Centre, Starting width and Propose width on the outputs, for chains, for the same Min and Max values in CosmoMC 2016.
1.If I take the centre near Min once and the centre near the Max later, how different can I expect my results would be?
2. Also, for some data set combinations large enough starting widths are resulting in failure to start the MCMC chains whereas the same starting width works for some other data set combinations (center, min, max and propose width being also same in two cases). Is there some specific reason for that?
3. If I take different starting widths or propose widths (considering all the different values are wide enough), but same Min, Max and Centre, would output results be different?

Antony Lewis
Posts: 1659
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: Significance of Centre, Starting width and Propose width

Post by Antony Lewis » October 10 2016

After chains converge, these settings should make no difference. They just affect the time it takes to converge. If you know where the high likelihood region is, save time by centering roughly on it (but you want some dispersion of different chains, so you can check convergence and potentially find other nearby minima).

With very wide widths you may hit other prior constraints, causing all the initially proposed models to be ruled out.

Shouvik Roychoudhury
Posts: 31
Joined: August 14 2016
Affiliation: IIT Bombay

Significance of Centre, Starting width and Propose width in

Post by Shouvik Roychoudhury » October 12 2016

Is there is any format in CosmoMC in which I can use only the Min & Max values, without providing any centre and the widths?

Antony Lewis
Posts: 1659
Joined: September 23 2004
Affiliation: University of Sussex
Contact:

Re: Significance of Centre, Starting width and Propose width

Post by Antony Lewis » October 12 2016

No. but you could the initial width to max-min and center anywhere in the range.

Shouvik Roychoudhury
Posts: 31
Joined: August 14 2016
Affiliation: IIT Bombay

Significance of Centre, Starting width and Propose width in

Post by Shouvik Roychoudhury » October 13 2016

Yes, I actually tried something similar already but if the starting width is quite large then the MCMC chains are not starting for some combination of data sets, as I mentioned earlier. But since you say the centre and the widths are not important as long as the chains converge, I guess I shouldn't worry much about setting narrower ranges for the widths. Thanks!

Post Reply