The multivariate t distribution seems to be defined as a "ratio" of a vector of normal random variables and a single gamma (or chi-squared) random variable (independent from the vector of normal).

Thus, even if the normal rvs are independent (i.e. generated from a multivariate normal with a diagonal covariance matrix) this definition seems to, nevertheless, imply that the random variables within a vector draw from a multivariate t are not independent since the same gamma rv is used to scale all normal rvs.

What is then a joint distribution of independent t-distributed rvs? Also, why in the definition of the multivariate t, we use only one gamma rv to construct a vector of t rvs from a vector of normal, instead of using a vector of gamma rvs so that each normal rv is scaled by a different gamma rv?


Let $\mathbf{X} \sim N_p(\mu, \Sigma)$ be a multivariate Normal random variable, and $Y \sim \chi^2_{\nu}$ be a $\chi^2$ random variable with $\nu$ degrees of freedom. Then the random variable $\mathbf{X}/\sqrt{Y/\nu}$ follows a Multivariate $t$ distribution with $\nu$ degrees of freedom. The pdf of this distribution is

$$f(\mathbf{x}) = \dfrac{\Gamma((\nu+p)/2)}{\Gamma(\nu/2) \nu^{p/2} \pi^{p/2} |\Sigma|^{1/2} } \left[1 + \dfrac{(\mathbf{x} - \mu)^T \Sigma^{-1} (\mathbf{x}-\mu)}{\nu} \right]^{-(v+p)/2} \,.$$

When $\Sigma$ is a diagonal matrix, the components of $\mathbf{X}$ are independent normal, but note that the pdf of the resulting multivariate $t$ distribution does not decompose into the product of the marginals, since the pdf is:

$$f(\mathbf{x}) = \dfrac{\Gamma((\nu+p)/2)}{\Gamma(\nu/2) \nu^{p/2} \pi^{p/2} |\Sigma|^{1/2} } \left[1 + \dfrac{1}{\nu}\sum_{i=1}^{p}\dfrac{(\mathbf{x} - \mu)^2 }{\sigma^2_i} \right]^{-(v+p)/2} \,.$$

Thus, independence of the underlying normal random variables, does not imply independence of the resulting $t$ variables. Also not that, the since the $\Sigma$ matrix is a diagonal matrix, the $t$ marginal variables are uncorrelated (but not independent).

So the joint distribution of uncorrelated $t$-random variables is above. As for the joint distribution of uncorrelated and independent $t$ random variables, that would just be the product of marginals.

As for using separate $\chi^2$ distributions, for a diagonal $\Sigma$ matrix that should correspond to independent univariate $t$ distributions. I am not quite sure what happens for a non-diagonal $\Sigma$ in this case.

  • $\begingroup$ Thank you for your answer. However, the first 4 paragraphs of your answer are already described in the OP; and the question is precisely about a pdf of an rv constructed using a multivariate normal and a vector of (independent) χ2, to which your reply is "I am not quite sure what happens for a non-diagonal Σ in this case." So, I cannot accept it as an answer, although I will up-vote it. Thanks again. $\endgroup$ – Confounded Mar 13 '18 at 16:11
  • $\begingroup$ @Confounded I agree I haven't completely answered your question. I will argue that the first 4 paragraphs of the answer are not entirely described in the question. The purpose of the first 4 paragraphs is to explain the delicate difference here between dependence and correlations. This is not described in the original question. $\endgroup$ – Greenparker Mar 13 '18 at 16:16

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.