2
$\begingroup$

I need to compute the probability of sampling a specific vector from any multinomial generated by a Dirichlet parametrization.

Let's call the vector $\mathbf{x} = \{x_i\}, i \in K$

I need to compute $$ p(\mathbf{x}|\alpha) = \int Mult_{PMF}(x, y) \, Dirichlet_{PDF}(y|\alpha)\;dy$$

Multinomial PMF is $$Mult_{PMF}(\mathbf{x}, \mathbf{p}) = B(\mathbf{x} +1) \prod_{i=1}^K p_i^{x_i} = \frac{\Gamma(\sum_{i=1}^K x_i +1)}{\prod_{i=1}^K \Gamma(x_i +1)} \prod_{i=1}^K p_i^{x_i}$$

Dirichlet PDF is $$Dirichlet_{PDF}(\mathbf{x}, \mathbf{\alpha}) = B(\mathbf{\alpha}) \prod_{i=1}^K x_i^{\alpha_i - 1} = \frac{\Gamma(\sum_{i=1}^K \alpha_i )}{\prod_{i=1}^K \Gamma(\alpha_i )} \prod_{i=1}^K x_i^{\alpha_i - 1}$$

This is what I derived

$$ \begin{align} p(\mathbf{x}|\alpha) & = \int Mult_{PMF}(x, y) \, Dirichlet_{PDF}(y|\alpha) \; dy \\ & = \int B(\mathbf{x} + 1) \prod_{i=1}^{K}y_i^{x_i} \; B(\alpha) \prod_{i=1}^K y_i^{\alpha_i-1} \; dy \\ & = B(\mathbf{x} + 1) B(\alpha) \int \prod_{i=1}^{K}y_i^{x_i} \; \prod_{i=1}^K y_i^{\alpha_i-1} \; dy \\ & = B(\mathbf{x} + 1) B(\alpha) \int \prod_{i=1}^{K}y_i^{x_i} y_i^{\alpha_i-1} \; dy \\ & = B(\mathbf{x} + 1) B(\alpha) \int \prod_{i=1}^{K}y_i^{x_i + \alpha_i-1} \; dy \\ \end{align} $$

At thi point I try to bring the part inside the integral in the form of the Dirichlet

$$ \begin{align} p(\mathbf{x}|\alpha) & = \frac{B(\mathbf{x} + 1) B(\alpha)}{B(\mathbf{x} + \alpha)} \int B(\mathbf{x} + \alpha) \prod_{i=1}^{K}y_i^{x_i + \alpha_i-1} \; dy \\ & = \frac{B(\mathbf{x} + 1) B(\alpha)}{B(\mathbf{x} + \alpha)} \int Dirichlet_{PDF}(y, \mathbf{x} + \mathbf{\alpha}) \; dy \\ & = \frac{B(\mathbf{x} + 1) B(\alpha)}{B(\mathbf{x} + \alpha)} \end{align}$$

I'm not sure I did all the steps correctly, as the result I obtained is pretty different from what I see online (more specifically at How to derive the Dirichlet-multinomial? )

In my result I have an additional $ B(x+1)$.

Is my result correct?

Why does it differ from the one showed in the other answer?

$\endgroup$
0
$\begingroup$

It ends up that I was comparing it with the Dirichlet-Categorical distribution instead of the Dirichlet-Multinomial distribution.

$$ \begin{align} \frac{B(\mathbf{x} + 1) B(\alpha)}{B(\mathbf{x} + \alpha)} & = \frac{\Gamma(\sum_i^K x_i + 1)}{\prod_i^K \Gamma(x_i + 1)} \frac{\Gamma(\sum_i^K \alpha_i)}{\prod_i^K \Gamma(\alpha_i)} \frac{\prod_i^K \Gamma(\alpha_i + x_i)}{\Gamma(\sum_i^K \alpha_i + x_i)}\\ &\\ & = \frac{(\sum_i^K x_i)!}{\prod_i^K x_i!} \frac{\Gamma(\sum_i^K \alpha_i)}{\Gamma(\sum_i^K \alpha_i + x_i)} \prod_i^K \frac{\Gamma(\alpha_i + x_i)}{\Gamma(\alpha_i)} \end{align} $$

$\endgroup$

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.