search

igmhub / baofit

Fits cosmological data to measure baryon acoustic oscillations.
9 stars 5 forks source link

Added MCMC-config option #4

Closed slosar closed 12 years ago

slosar commented 12 years ago

This time done in a much simpler way that modified only one file and really really should work, but it doesn't.... David, do you understand why?

dkirkby commented 12 years ago

This is harder than it sounds, at least in the general case. For example, if your mcmc-config floats a parameter that was previously fixed, how do we fill in the missing covariances that the MCMC algorithm expects? I think you were specifically interested in changing/removing priors, right? Why does the fit need different priors than MCMC?

slosar commented 12 years ago

I need for the following reasons:

You want boxpriors on anistropic params, but fitting with box priors often results in minuit failing. So the rigth thing to do is to fit with a strongish gaussian priors and then do an unbiased MCMC with boxpriors. In principle I think I should be able to just run MCMC with default errors as a diagonal proposal matrix...

anže

On Sun, 9 Sep 2012, dkirkby wrote:

This is harder than it sounds, at least in the general case. For example, if your mcmc-config floats a parameter that was previously fixed, how do we fill in the missing covariances that the MCMC algorithm expects? I think you were specifically interested in changing/removing priors, right? Why does the fit need different priors than MCMC?

Reply to this email directly or view it on GitHub: https://github.com/deepzot/baofit/pull/4#issuecomment-8407559

dkirkby commented 12 years ago

It sounds like the underlying problem is that minuit doesn't like box priors. I suspect this is because I made the edges too sharp and it doesn't like the discontinuity. I am going to add an option for softer edges and see if this helps.

dkirkby commented 12 years ago

I am closing this, assuming that the new soft box priors will eliminate most cases of box priors not working with MINUIT. Also, the new mcmc-reset option allows you to ignore the MINUIT covariance for MCMC (but does not let you change the priors).