Let the probability density function of $(Y_1, \ldots, Y_k)$ be a Dirichlet distribution parameterized by $\alpha_1 , \ldots, \alpha_k$, that is

$$ f(y_1, \ldots, y_k|\alpha_1, \ldots, \alpha_k) = \text{Dir}(\alpha_1, \ldots, \alpha_k) $$

Let $D(y_1, \ldots, y_k|\alpha_1, \ldots, \alpha_k)$ denote the distribution function of $(Y_1, \ldots, Y_k)$,

and $X$ is a discrete random variable with p.f.

$$ P(X=j|Y_1, \ldots, Y_k) = Y_j \ \text{ a.s } \text{ for } j = 1, \ldots, k $$

What I want to do is to calculate the following joint distribution of $X$ and $Y_1, \ldots, Y_k$:

$$ P(X=j, Y_1 \leq z_1, \ldots, Y_k \leq z_k) $$

My way is using p.d.f. of the joint distribution above:

\begin{align} &P(X=j, Y_1 \leq z_1, \ldots, Y_k \leq z_k) =\\ &\ \ \ \ \ \ \ P(X=j)P(Y_1 \leq z_1, \ldots, Y_k \leq z_k|X=j) \\ &\ \ \ \ \ \ = P(X=j)\int_0^{z_1}\cdots\int_0^{z_k}\frac{P(X=j|y_1, \ldots, y_k)f(y_1, \ldots, y_k)}{P(X=j)}dy_1 \cdots dy_k \\ &\ \ \ \ \ \ = \int_0^{z_1}\cdots\int_0^{z_k}y_jdD(y_1, \ldots, y_k|\alpha_1, \ldots, \alpha_k) \end{align}

However, I'm looking for another solution by factorizing the joint distribution as

\begin{align} &P(X=j, Y_1 \leq z_1, \ldots, Y_k \leq z_k) =\\ &\ \ \ \ \ \ \ P(X=j|Y_1 \leq z_1, \ldots, Y_k \leq z_k)P(Y_1 \leq z_1, \ldots, Y_k \leq z_k) \end{align}

How can I deal with this factorization?


Let $A=P[X=j,Y_1\leqslant y_1, \ldots, Y_k\leqslant y_k]$. Then $$ A=E[P[X=j|Y_1, \ldots, Y_k]; Y_1\leqslant y_1, \ldots, Y_k\leqslant y_k]=E[Y_j; Y_1\leqslant y_1, \ldots, Y_k\leqslant y_k], $$ hence $$ A=\int z_j\,\mathbf 1_{z_1\leqslant y_1,\ldots,z_k\leqslant y_k}\,\mathrm{Dir}(z_1,\ldots,z_k)\,\mathrm dz. $$ Note that there is no simple formula for $P[X=j|Y_1\leqslant y_1, \ldots, Y_k\leqslant y_k]$.

  • $\begingroup$ Could you say something for the advantage of using expectation rather than the integration of p.d.f? $\endgroup$ – fishiwhj Aug 9 '13 at 10:27
  • $\begingroup$ Not much (basically because I do not understand your comment)--except that the manipulations in your post (factoring P(X=j), reversing Bayes formula (apparently), simplifying by P(X=j)) are somewhat odd. $\endgroup$ – Did Aug 9 '13 at 10:44
  • $\begingroup$ OK...Please forgive my poor English, by the way, is there something wrong(or non-standard) with my solution? $\endgroup$ – fishiwhj Aug 10 '13 at 4:22

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.