If $\theta$ is a ($k$-dimensional) Dirichlet distribution, the sufficient statistics are $\log\theta_i, i = 1,\ldots, k$. It can be shown that if the Dirichlet has parameter $\alpha = (\alpha_1, \ldots, \alpha_k) > 0$, then the expected sufficient statistics are $$\mathbb{E}(\log\theta_i) = \Psi(\alpha_i) - \Psi\left(\sum_{i=1}^k \alpha_i\right)$$ where $\Psi$ is the digamma function.

I would like to compute the space of all realizable mean values $\mathbb{E}(\log\theta)$, i.e. $$\mathcal{M}= \{\mu\in\mathbb{R}^k\,|\,\mu = \mathbb{E_\alpha}(\log\theta) \text{ for some }\alpha\} = \left\{\Psi(\alpha)-\Psi\left(\sum_{i=1}^k\alpha_i\right) \, |\, \alpha \in \mathbb{R}_{>0}^k \right\}$$ Clearly $\mathcal{M} \subseteq \mathbb{R}_{<0}^k$ (because $\Psi$ is increasing), but apart from this however, I am not able to compute the image. Does anyone know what this set is?


Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.