Page 1 of 1

CosmoMC: Fisher Matrix

Posted: November 23 2004
by Maria Beltran
Hi,

Does anybody know whether the actual Fisher matrix of the posterior distribution can be calculated using CosmoMC?
Since this distribution is not exactly a multivariate gaussian, its Fisher matrix should be different from the covariance matrix, shouldn't it?
Thanks!

Maria

Re: CosmoMC: Fisher Matrix

Posted: November 23 2004
by Antony Lewis
In recent CosmoMC versions you can set

estimate_propose_matrix = T

in the .ini file. This will find the approx best fit point (by conjugate gradients) and then work out the inverse Hessian about this point by taking numerical derivatives. It outputs the result to a file called

file_root.local_invhessian

The program will then continue with a normal MCMC run, using the inverse Hessian for the proposal density matrix. If you just want the Hessian you can kill the process after the local_invhessian file is produced.