What is the ML estimate of the parameter $e_i$ for the Dirichlet-Multinomial (Polya) distribution defined below?

$p(\mathbf{x}|\mathbf{e}) = \frac{N!}{\prod_i^d x_i!}\frac{\Gamma(A)}{\Gamma(N+A)}\prod_i^d\frac{\Gamma(x_i+e_i)}{\Gamma(e_i)}$ ,

where $\mathbf{x}$ is a vector of $d$ observations (i.e., how many times we observed the $i^{th}$ value among $d$) and $\mathbf{e}$ is a vector of $d$ prior expectations (i.e., parameters of Dirichlet distribution), with $N=\sum_ix_i$ and $A=\sum_ie_i$.


1 Answer 1


If you just have a single observation $\mathbf{x}=(x_1,\dots,x_d)$, the likelihood would be maximised by sending each $e_i$ to infinity while keeping $e_i/\sum_{i=1}^d e_i = x_i/\sum_{i=1}^d x_i$ in which case the distribution simplifies to an ordinary multinomial distribution.


Your Answer

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

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