3
$\begingroup$

In the Wikipedia entry for "Inverse-Wishart distribution" (current revision) there is a formula for a Wishart distribution conditioned on a conjugate prior ( $\Psi,\nu$) in which the original Wishart scale matrix $\Sigma$ has been integrated out to yield the conditional distribution

$$ p(X\mid\Psi,\nu) = \frac { |\Psi|^{\nu/2}\Gamma_p( \frac {\nu + n}{2})}{\pi^{np/2} |\Psi+X^TX|^{(\nu+n)/2} \Gamma_p (\frac{\nu}{2})} $$
where $A_{n\times n} $ is the observed data $X^TX $.

This doesn't appear to be a valid Wishart distribution - which I think should be something more like

$$ p(X) = \frac { |\Psi|^{\nu/2}\Gamma_p( \frac {\nu + n}{2})}{\pi^{np/2} |\Psi+A|^{(\nu+n)/2} \Gamma_p (\frac{\nu}{2})} \exp \Big [-\operatorname{tr} \Big ((\Psi +\Sigma)^{-1}A \Big) \Big].$$ So am I misunderstanding this and the formula is a partial result for illustration purposes only?

$\endgroup$

1 Answer 1

2
$\begingroup$

The distribution shown is not a Wishart distribution, it is a multivariate t-distribution. This is the marginal distribution of $X$ after integrating the covariance matrix out of the Gaussian sampling model. The expression may not be correct (I haven't checked carefully, but it doesn't look quite right). A good reference for the derivation is section 1.4.3 of the book Markov Chain Monte Carlo: Stochastic Simulation for Bayesian Inference by Gamerman and Lopes.

$\endgroup$
2
  • $\begingroup$ The equation also differs from Wikipedia's own definition of a multivariate t distribution so it needs to be clarified. My problem is in comms and has known zero means everywhere: only the covariance part is of interest. I had referred to Murphy's "Conjugate Bayesian analysis of the Gaussian distribution" previously. $\endgroup$
    – eulogy
    May 9, 2019 at 10:44
  • $\begingroup$ The density given appears to have a typo (not clear if it intends to integrate over the prior or the posterior), but it is closer than it looks after applying the matrix determinant lemma to $|A + \Psi|$, to get $(1 + X^t \Psi^{-1} X)|\Psi|$, which is the more familiar t-distribution form. $\endgroup$
    – R Hahn
    May 9, 2019 at 15:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.