I assume the following model for a sample $y_1 \in \mathbb{R}^2$ of size $1$ with bivariate Gaussian likelihood and independent bivariate Gaussian and inverse-Wishart prior for the mean and variance parameters, respectively. That is

$$ y_1 \mid \mu, \Sigma \sim \mbox{Normal}(y_1 \mid\texttt{mean}=\mu, \texttt{Var}=\Sigma) \\ \mu \sim \mbox{Normal}(\mu \mid\texttt{mean}=\mu_\mu, \texttt{Var}=\Sigma_\mu)\\ \Sigma \sim \mbox{Inverse-Wishart}(\Sigma\mid\texttt{df}=\nu_\Sigma,\texttt{scale}=S_\Sigma) $$

I want to derive the analytical expression of the marginal likelihood of $y_1$, that is $$ p(y_1)=\iint p(y_1 \mid \mu, \Sigma) \, p(\mu) \, p(\Sigma) \mathrm{d}\mu \mathrm{d}\Sigma, $$ where, with an abuse of notation, $p(x)$ denotes the density of $x$ evaluated in the point $x$.

Remark: I know that if I assume a conjugate Normal-Inverse-Wishart prior such result is well-known and coincide with a Generalized-T distribution evaluated in $y_1$. See for instance Wikipedia Conjugate Prior. However, I do not want to change my prior.

My attempt: Relying on well-known conjugacy results (see Wikipedia Conjugate Prior) I know how to solve one of the two integral fixing either $\mu$ or $\Sigma$, but I cannot solve (that is find a tractable analytical expression in terms of well-known functions) the second integral.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.