This is a reference request for the following result. Let $X$ be a random matrix following the matrix variate $t$-distribution $T_{p,m}(\nu, M, U, V)$ (as defined in Wikipedia). Then $$ \frac{\det(U)}{\det\bigl(U + (X-M)V^{-1}{(X-M)}'\bigr)} $$ has the same distribution as the product $B_1 \ldots B_m$ of $m$ independent random variables $B_i \sim \textrm{Beta}\left(\frac{\nu+m-i}{2}, \frac{p}{2}\right)$.

This result is checked by simulations in R below.

This result is given in Geisser's article Bayesian estimation in multivariate analysis but it is a bit hidden: it is given in a Bayesian context where a particular matrix variate $t$-distribution appears without name, and the proof is only sketched.

I'm looking for a reference which precisely states this result. It is not given in Gupta & Nagar's book Matrix variate distributions.

library(LaplacesDemon) # to simulate Wishart distribution

# simulates matrix normal N_{p,m}(M, U ⊗ V)
rmatrixnormal <- function(M, U, V){
  Z <- matrix(rnorm(nrow(M)*ncol(M)), nrow(M), ncol(M))
  M + t(chol(U)) %*% Z %*% chol(V)

# parameters of the matrix t-distribution
p <- 5; m <- 3
U <- rwishart(p, diag(p)) # arbitray U matrix
V <- rwishart(m, diag(m)) # arbitrary V matrix
Vinv <- solve(V) # inverse V
nu <- 6 # degrees of freedom

# simulations of matrix t-distribution
nsims <- 25000
sims <- numeric(nsims)
for(i in 1:nsims){
  Psi <- rinvwishart(nu+m-1, V)
  X <- rmatrixnormal(matrix(0,p,m), U, Psi)
  sims[i] <- det(U)/det(U + X %*% Vinv %*% t(X))

# simulations of product of m Beta distributions
k <- (nu+m-1)/2
sims2 <- rbeta(nsims, k, p/2)*rbeta(nsims, k-0.5, p/2)*rbeta(nsims, k-1, p/2)

# compare cumulative distribution functions
curve(ecdf(sims)(x), from=0, to=1)
curve(ecdf(sims2)(x), col="red", add=TRUE, lty=2, lwd=3)

enter image description here


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.