When I try to calculate the Information matrix of a multivariate
Gaussian of dimension d, I arrive at

I(Sigma) = 0.5 * D' * kron( Sigma^(-1), Sigma^(-1) ) * D

where D is the duplication matrix, Sigma is the covariance matrix
of the Gaussian, kron is the kronecker product.

(I hope I do not make a mistake in my calculation... but I am not sure.)

I am wondering if the integral

int     D' * kron( Sigma^(-1), Sigma^(-1) ) * D     ds

is finite or not, where 

vec Sigma = D s

To make the matter worse, I do not know how to state the integration
limit, since the integration should be over values of s that make
Sigma positive definite.

My guess is that the integral is infinite, since Jeffery's prior is
usually improper. If my guess is correct, does close form solution
exist for the integral if the range of the integration is restricted to 
a bounded subset? The form of the subset is not important, since I
am trying to modify the prior to become "proper".

Any pointers or references are welcome, as this is the first time
I handle matrix parameter.
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to