Let's say that $(X_1, \dots, X_4) \sim Dirichlet(\alpha_1, \dots, \alpha_4)$ with $\sum_{i=1}^4 X_i=1$. I want to find the distribution of $(X_1, X_2)$.

I know the marginal distributions of the $X_i$ (answered in other questions on this site) but it doesn't seem that this has been answered.

Denote the full joint pdf by $f_0$ so that $f_0(x_1, x_2, x_3, x_4) = \frac{1}{B(\vec \alpha)} \prod_{i=1}^4 x_i^{\alpha_i-1}$. I'll denote the joint pdf of $(X_1, X_2)$ by $g$.

First of all, since $X_4 = 1 - X_1 - X_2 - X_3$ it seems that I should only consider $X_1$, $X_2$, and $X_3$ and therefore I'll only need to do a single integration.

This means that I'll be working with $$ f(x_1, x_2, x_3) := \frac{x_1^{\alpha_1 - 1} x_2^{\alpha_2 - 1} x_3^{\alpha_3 - 1}(1-x_1-x_2-x_3)^{\alpha_4-1}}{B(\vec \alpha)}. $$

From this it follows that $$ g(x_1, x_2) = \int_{x_3} f(x_1, x_2, x_3) dx_3 $$

$$ = \frac{x_1^{\alpha_1-1} x_2^{\alpha_2-1}}{B(\vec \alpha)} \int_{x_3} x_3^{\alpha_3-1} (1-x_1-x_2-x_3)^{\alpha_4-1} dx_3. $$

My questions:

  1. What are the limits of integration here?

  2. How do I actually do this integral? I assume that I need to shoehorn it into a beta integral but I don't see how.

For the limits of integration, certainly $x_3 \geq 0$, but I don't know what the upper bound is. Would it just be $0 \leq x_3 \leq 1 - x_1 - x_2$?

As for the actual integral, if I'm correct about $0 \leq x_3 \leq 1 - x_1 - x_2$ then I just need to be able to do $$ I = \int_{0}^{1-k} t^{\alpha-1}(1-k-t)^{\beta-1} dt $$ where $k = x_1 + x_2$, $\alpha = \alpha_3$, and $\beta = \alpha_4$. This looks really close to an incomplete beta but not quite.

  • $\begingroup$ If I am not mistaking there is a factor $2$ missing in the integral. You lift $(x_1, x_2, x_3)$ to $\mathbb R^4$ with a function $u(x_1,x_2,x_3) = (x_1,x_2,x_3,1-x_1-x_2-x_3)$. So when you integrate the Dirichlet distribution on $\mathbb R^3$, which maps to the 3 simplex a 3 dimensional manifold embedded into $\mathbb R^4$, you have to account for the distortion done by $u$. This is a constant factor $\sqrt{Du^TDu}=2$ in the integral over $f(u(x_1,x_2,x_3))$ where $Du$ is the Jacobian matrix of $u$. Please do not mind if I am wrong, but I intend to do something very similar and the answer prov $\endgroup$ – katosh Sep 5 '18 at 16:29

By manipulating the integral you can obtain the beta function. For ease of notation, let $1-k = x$, then $I$ can be written as \begin{align} I &= \int_{0}^{x} t^{\alpha-1}(x-t)^{\beta-1}dt \\ &= x^{\beta-1} \int_{0}^{x} t^{\alpha-1} \left(1-\frac{t}{x}\right)^{\beta-1}dt. \end{align} Let $t/x = u$ then $dt = x du$. If $t=x$ then $u=1$, hence the above integral reduces to $$\int_{0}^{1} x^{\alpha-1} u^{\alpha-1} (1-u)^{\beta-1} x du.$$ Therefore, \begin{align} I &= x^{\alpha+\beta-1} \int_{0}^{1} u^{\alpha-1}(1-u)^{\beta-1}du \\ &= x^{\alpha+\beta-1} \beta(\alpha,\beta). \end{align}

  • $\begingroup$ Thanks a lot! Regarding my question (1), did I get the limits of integration correct? $\endgroup$ – alfalfa Apr 6 '16 at 22:00
  • $\begingroup$ Regarding the definition of the Dirichlet distribution, I think the bounds for $x_3$ are indeed correct. $\endgroup$ – Cavents Apr 6 '16 at 22:06

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.