minimizing kullback-leiber for normal-gamma

23 Views Asked by At

Considering two continuous probability distributions q(x) and p(x), The Kullback leiber Divergence is defined as the measure of the information lost when q(x) is used to approximate p(x). \begin{equation*} KL (p(x)|| q(x)) = \int_{R} p(x) \ln \frac{p(x)}{q(x)} dx \end{equation*}

A distributions in the exponential family can be written as : \begin{equation*} p_\theta(x) = \frac{1}{Z(\theta)} exp(\theta^T \phi (x)) \end{equation*} where $\phi(x)$ is the vector of the natural statistics of x and $Z(\theta) = \int_ exp(\theta^T \phi(x)) dx$ .

My goal is to prove that for the special case of the normal-gamma distribution. $$(X,T) - NormalGamma(\mu, \lambda,\alpha,\beta)$$ with pdf: $$ f(x,t;\mu,\lambda,\alpha,\beta) = \frac{\beta^\alpha \sqrt{\lambda}}{\Gamma(\alpha)\sqrt{2\pi}} t^{\alpha-\frac{1}{2}} e^{-\beta t} \exp\left(-\frac{\lambda t(x-\mu)^2}{2} \right). $$ we have the following theorem: the distribution $p_{\theta*}$ which minimises the Kullback-Leibler divergence for the normal inverse gamma distribution with probability distribution $p_\theta$ is given by : \begin{equation*} E_{p_{\theta^*}(x)}[\phi(x)] = E_{p_{(x)}}[\phi(x)] \end{equation*}

This will imply finding $\theta^*$ such that $\nabla_\theta f(\theta^*) =0$ Then I will need to find the hessian to proof that $\theta^*$ is actually the minimum. My problem is that I was not able to realize those derivation successfully. Could anyone help me please?