7
$\begingroup$

Let $X_i\sim\text{Gamma}(\alpha,p_i),i=1,2,...,n+1$ be independent random variables.

Define $Z_1=\sum_{i=1}^{n+1}X_i$ and $Z_i=\frac{X_i}{\sum_{j=1}^iX_j},\quad i=2,3,...,n+1$. Then show that $Z_1,Z_2,...,Z_{n+1}$ are independently distributed.

The joint density of $(X_1,...,X_{n+1})$ is given by

$$f_{\bf X}(x_1,...,x_{n+1})=\left[\frac{\alpha^{\sum_{i=1}^{n+1}p_i}}{\prod_{i=1}^{n+1}\Gamma(p_i)}\exp\left(-\alpha\sum_{i=1}^{n+1}x_i\right)\prod_{i=1}^{n+1}x_i^{p_i-1}\right]\mathbf I_{x_i>0}\quad,\alpha>0,p_i>0$$

We transform $\mathbf X=(X_1,\cdots,X_{n+1})\mapsto\mathbf Z=(Z_1,\cdots,Z_{n+1})$ such that

$Z_1=\sum_{i=1}^{n+1}X_i$ and $Z_i=\frac{X_i}{\sum_{j=1}^iX_j},\quad i=2,3,...,n+1$

$\implies x_{n+1}=z_1z_{n+1},$

$\qquad x_n=z_1z_n(1-z_{n+1}),$

$\qquad x_{n-1}=z_1z_{n-1}(1-z_n)(1-x_{n+1}),$

$\qquad\vdots$

$\qquad x_3=z_1z_3\prod_{j=4}^{n+1}(1-z_j)$

$\qquad x_2=z_1z_2\prod_{j=3}^{n+1}(1-z_j)$

$\qquad x_1=z_1\prod_{j=2}^{n+1}(1-z_j)$, where $0<z_1<\infty$ and $0<z_i<1,\quad i=2,3,\cdots,n+1$

The Jacobian of the transformation is $J=\dfrac{\partial(x_1,...,x_{n+1})}{\partial(z_1,...,z_{n+1})}=\det\begin{pmatrix}\frac{\partial x_1}{\partial z_1} &\cdots& \frac{\partial x_1}{\partial z_{n+1}} \\ &\ddots\\\frac{\partial x_{n+1}}{\partial z_1} &\cdots& \frac{\partial x_{n+1}}{\partial z_{n+1}}\\\end{pmatrix}$

Performing the operation $R_1'=\sum_1^{n+1}R_i$, we get $J$ as the determinant of

$\begin{pmatrix}1&0&0&\cdots&0&0 \\z_2\prod_3^{n+1}(1-z_j)& z_1\prod_3^{n+1}(1-z_j) \\ z_3\prod_4^{n+1}(1-z_j)&0& z_1\prod_4^{n+1}(1-z_j)\\&&&\ddots\\z_n(1-z_{n+1})&0&0&\cdots&z_1(1-z_{n+1})& -z_1z_n \\z_{n+1}&0&0 &\cdots&0& z_1\\\end{pmatrix}$

which equals $z_1^n(1-z_3)(1-z_4)^2...(1-z_n)^{n-2}(1-z_{n+1})^{n-1}$.

After some simplification we get the joint density of $\bf Z$ as

$f_{\bf Z}(z_1,...,z_{n+1})=\prod_{i=1}^{n+1}f_{Z_i}(z_i)$

where $Z_1\sim \text{Gamma}(\alpha,\sum_1^{n+1}p_i),$

$\qquad Z_2\sim \text{Beta}_1(p_1,p_2),$

$\qquad Z_3\sim \text{Beta}_1(p_3,p_1+p_2),$

$\qquad \vdots$

$\qquad Z_{n+1}\sim \text{Beta}_1(p_{n+1},\sum_1^np_i)$,

with $0<z_1<\infty$ and $0<z_i<1,\quad i=2,3,\cdots,n+1$,

$\alpha>0$ and $p_i>0$ for $i=1,2,...,n+1$.


Needless to say, finding the inverse solutions $x_i$'s and evaluating the Jacobian was cumbersome and time consuming. Besides getting the job done, it also determines the distributions of the $Z_i$'s.

Is there any simpler way to show just the independence of the $Z_i$'s?

$\endgroup$
  • $\begingroup$ I meant $...(1-z_{n+1})$ in the rhs of $x_{n-1}$. $\endgroup$ – StubbornAtom Dec 9 '17 at 19:39
  • $\begingroup$ The only way I could think of simplifying this is using induction, i.e. starting with the n=1 case, and finding $Z_1$ and $Z_2$ are independent, then using the independence of $X_i$ with the $Z_j$ ($j<i$) to just add in the $X_i$ one by one as needed. $\endgroup$ – aleshing Dec 9 '17 at 19:41
  • $\begingroup$ Related with the Dirichlet distribution. $\endgroup$ – StubbornAtom Dec 24 '17 at 19:48
1
$\begingroup$

I'll prove the equivalent statement.

Let $n\ge 1$, $X_k\sim \Gamma(\alpha,p_k)$, $k=1,\dots,n+1$, are independent. Denote $S_k = X_1+\dots+X_k$, $k=1,\dots,n+1$; $R_k = \frac{S_k}{S_{k+1}}$, $k=1,\dots,n$. Then $R_1$, $R_2$, $\dots$, $R_{n}$, and $S_{n+1}$ are independent and $S_{n+1}\sim \Gamma(\alpha,p_1+\dots + p_{n+1})$.

Remark In OP notation, $Z_1=S_{n+1}$, $Z_k = 1 - R_{k-1}$, $k=2,\dots,n+1$.

Proof. $n=1$ is easy (and well known).

$n-1\Rightarrow n$.

By induction hypothesis and independence, the vectors $\mathbf{R} = (R_1,\dots,R_{n-1})$ and $(S_{n},X_{n+1})$ are independent. Therefore, the vectors $\mathbf{R}$ and $\big(\frac{S_{n}}{S_{n}+X_{n+1}},S_{n}+X_{n+1}\big) = (R_n,S_{n+1})$ are independent. Both vectors have independent components: $\mathbf R$ by induction hypotheses, $(R_n,S_{n+1})$ by induction base. Therefore, their components are independent random variables.


It is no problem to include the distribution of $R_n$ to the statement, the proof won't change. The distribution of $S_n$ is already there: I need it for saying that the independence of $R_n$ and $S_{n+1}$ follows from the induction base.

$\endgroup$

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.