1
$\begingroup$

In Gupta & Nagar's book Matrix variate distributions, the noncentral Beta type I(B) distribution with parameters $a$, $b$ and noncentrality parameter $\Theta$ is defined by $U={(S_1+S_2)}^{-\frac12}S_1{(S_1+S_2)}^{-\frac12}$ where $S_1 \sim W(2a, I, \Theta)$ is independent of $S_2\sim W(2b, I)$ (Wishart distributions), and the noncentral Beta type II(B) distribution is defined by $V={(S_2)}^{-\frac12}S_1{(S_2)}^{-\frac12}$ where $S_1$ and $S_2$ as above.

Then it is claimed that $V \sim U{(I-U)}^{-1}$. However my simulations show that this distributional equality does not hold. I've found that it holds when $\Theta = \textrm{diag}(\theta, \ldots, \theta)$. Otherwise I get $\det(V) \sim \det\bigl(U{(I-U)}^{-1}\bigr)$ and $\textrm{tr}(V) \sim \textrm{tr}\bigl(U{(I-U)}^{-1}\bigr)$ but not $V \sim U{(I-U)}^{-1}$.

Is it true that that $V \sim U{(I-U)}^{-1}$? That would mean I do something wrong in my simulations but I'm pretty sure my simulations are correct, and I'm rather wondering why the equality holds when $\Theta = \textrm{diag}(\theta, \ldots, \theta)$.

Note

With my simulations again, I've found that $V \sim U{(I-U)}^{-1}$ if we define $U = S_1^\frac12{(S_1+S_2)}^{-1}S_1^\frac12$ and $V = S_1^\frac12 S_2^{-1}S_1^\frac12$. These are other proposed definitions of the matrix Beta distributions.

Update

I think I misread. In fact the book is a bit confusing regarding this point. There are two Beta type I(B) distributions in the book: the one defined above and another one which was originally defined by Asoo. The book claims that $V \sim U{(I-U)}^{-1}$ when $U$ has the Asoo distribution, there's no such claim for the $U$ defined as above. So the real question is: why this equality holds when $\Theta = \textrm{diag}(\theta, \ldots, \theta)$?

$\endgroup$
0
$\begingroup$

The density of $U$ is (see Gupta & Nagar page 189) $$ f(U) = C_f {\det(U)}^{a - \frac12(p+1)} {\det(I_p-U)}^{b - \frac12(p+1)} \\ \int_{S>0} {\det(S)}^{a+b-\frac12(p+1)} \exp\left(-\mathrm{tr}\left(\frac{S}{2}\right)\right) {}_0\!F_1\left(a, \frac{1}{4}\Theta_1 S^\frac12 US^\frac12\right) \mathrm{d}S \\ = K_f {\det(U)}^{a - \frac12(p+1)} {\det(I_p-U)}^{b - \frac12(p+1)} \\ \int_{S>0} {\det(S)}^{a+b-\frac12(p+1)} \exp\bigl(-\mathrm{tr}(S)\bigr) {}_0\!F_1\left(a, \frac{1}{2}\Theta_1 S^\frac12 US^\frac12\right) \mathrm{d}S $$ (where $C_f$ and $K_f$ are constants).

The density of $V$ is (see Gupta & Nagar page 192) $$ g(V) = C_g {\det(V)}^{a - \frac12(p+1)} {\det(I+V)}^{\frac12 (a+b)} {}_1\!F_1\left(a+b, a, \frac12 \Theta_1 V{(I+V)}^{-1}\right). $$ Make the change of variables $U = V{(I_p+V)}^{-1}$, so $V = U{(I_p-U)}^{-1}$ and $I_p + V = {(I_p-U)}^{-1}$. The Jacobian is ${\det(I_p -U)}^{-(p+1)}$. We find that the density of $U$ is $$ h(U) = C_g {\det(U)}^{a - \frac12(p+1)} {\det(I_p-U)}^{b - \frac12(p+1)} {}_1\!F_1\left(a+b, a, \frac12 \Theta_1 U\right). $$

The integral in $f(U)$ does not simplify to something nice. Hence the density $f(U)$ and $h(U)$ are not the same.

But when $\Theta_1$ is scalar, by integral involving hypergeometric function of matrix argument, $$ f(U) = K_f {\det(U)}^{a - \frac12(p+1)} {\det(I_p-U)}^{b - \frac12(p+1)} \\ \int_{S>0} {\det(S)}^{a+b-\frac12(p+1)} \exp\bigl(-\mathrm{tr}(S)\bigr) {}_0\!F_1\left(a, \frac{1}{2}\Theta_1 S U\right) \mathrm{d}S \\ = K_f\Gamma_p(a+b) {\det(U)}^{a - \frac12(p+1)} {\det(I_p-U)}^{b - \frac12(p+1)} {}_1\!F_1\left(a+b, a, \frac12 \Theta_1 U\right), $$ the same as $h(U)$.

$\endgroup$

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.