5
$\begingroup$

I am trying to understand how my professor (see derivation below) has derived the multivariate marginal distribution of a subvector of $\theta_j$´s from a Dirichlet distribution. I understand everything really, until the third line. I understand that he wants the integral to represent a unnormalised beta integral and hence he have to divide $\theta^{\alpha_{k-1}-1}_{k-1}$ and $(1-\theta_1 ... - \theta_{k-1})^{\alpha_{k}-1}$ by their sum, i.e $\theta^{\alpha_{k-1}-1}_{k-1} + (1-\theta_1 ... - \theta_{k-1})^{\alpha_{k}-1}$

If I understand correctly, the sum is equal to $(1-\theta_1 ... - \theta_{k-1})^{\alpha_{k-2}+\alpha_{k-1}-2}$, but how do we get that this is in fact the sum?

enter image description here

$\endgroup$
1
  • $\begingroup$ Thanks for the edit @Xi'an $\endgroup$ Nov 10, 2020 at 13:08

1 Answer 1

7
$\begingroup$

Since$$p(\theta_1,\ldots,\theta_{k-1})\propto (1-\theta_1-\cdots-\theta_{k-1})^{\alpha_{k}-1}\prod_{i=1}^{k-1} \theta_i^{\alpha_i-1}$$over the $\mathbb R^k$-simplex, $$\mathfrak S = \left\{(\theta_1,\ldots,\theta_{k-1})\in\mathbb R^{k-1}_+\,;\,\sum_{i=1}^{k-1} \theta_i\le 1\right\}$$ integrating out $\theta_{k-1}$ produces the marginal density of $(\theta_1,\ldots,\theta_{k-2})$: \begin{align} p(\theta_1,\ldots,\theta_{k-2})&\propto \int_0^{1-\theta_1-\cdots-\theta_{k-2}} (1-\theta_1-\cdots-\theta_{k-1})^{\alpha_k-1}\prod_{i=1}^{k-1} \theta_i^{\alpha_i-1}\,\text d\theta_{k-1}\\ &= \prod_{i=1}^{k-2} \theta_i^{\alpha_i-1}\, \int_0^{1-\theta_1-\cdots-\theta_{k-2}} (1-\theta_1-\cdots-\theta_{k-1})^{\alpha_{k}-1}\,\theta_{k-1}^{\alpha_{k-1}-1}\,\text d\theta_{k-1}\\ \end{align} Given the upper bound on $\theta_{k-1}$ found in the integral, one can rather naturally consider the change of variable $$\theta_{k-1}=(1-\theta_1-\cdots-\theta_{k-2})\eta$$ leading to \begin{align} p(\theta_1,\ldots,\theta_{k-2}) &= \prod_{i=1}^{k-2} \theta_i^{\alpha_i-1}\, \int_0^1 (1-\theta_1-\cdots-\theta_{k-2})^{\alpha_{k-1}-1}(1-\eta)^{\alpha_{k-1}-1}\\ &\qquad\times(1-\theta_1-\cdots-\theta_{k-2})^{\alpha_{k-1}-1}\eta^{\alpha_{k-1}-1}\, (1-\theta_1-\cdots-\theta_{k-2})\text d\eta\\ &=\prod_{i=1}^{k-2} \theta_i^{\alpha_i-1}\,(1-\theta_1-\cdots-\theta_{k-2})^{\alpha_{k-1}+\alpha_k-1}\, \underbrace{\int_0^1 (1-\eta)^{\alpha_{k}-1}\eta^{\alpha_{k-1}-1}\,\text d\eta}_\text{constant in $\theta$}\\ &\propto\prod_{i=1}^{k-2} \theta_i^{\alpha_i-1}\,(1-\theta_1-\cdots-\theta_{k-2})^{\alpha_{k-1}+\alpha_k-1}\, \end{align}

$\endgroup$
1
  • 1
    $\begingroup$ +1, I wonder if there's a typo at the first line after "leading to..."; shouldn't it be $(1 - \eta)^{\alpha_{k}-1}$ rather than $(1 - \eta)^{\alpha_{k-1}-1}$? $\endgroup$ Apr 26, 2022 at 6:29

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.