2
$\begingroup$

It is well known that bivariate normal pdf can be written in terms of univariate pdfs: $$ \frac{1}{2\pi\sigma_x\sigma_y\sqrt{1-\rho^2}}\exp\left(-\frac{(\xi_1^2+\xi_2^2-2\rho \xi_1\xi_2)}{2(1-\rho^2)}\right)=\frac{1}{\sigma_x\sigma_y\sqrt{1-\rho^2}}\phi\left(\frac{\xi_1-\rho\xi_2}{\sqrt{1-\rho^2}}\right)\phi\left(\xi_2\right) $$ Is there a similar result for bivariate Student t-distribution, that is, can a bivariate Student t-distribution be written in terms of univariate Student t-densities?

$\endgroup$
1
  • $\begingroup$ where did you see the last equality derived, and what is $\phi$? $\endgroup$
    – develarist
    Dec 7, 2020 at 1:33

1 Answer 1

2
$\begingroup$

There is a mistake in the expressions of the Normal bivariate pdf decomposition. As a general result, the conditional distribution of the first component of a Normal bivariate vector given the second component is $$X_1\mid X_2=\xi_2 \ \sim\ \mathcal{N}\left(\mu_1+\frac{\sigma_1}{\sigma_2}\rho(\xi_2 - \mu_2),\, (1-\rho^2)\sigma_1^2\right).$$ Hence, assuming $\mu_1=\mu_2=0$, the conditional density of $X_1$ given $X_2=\xi_2$ is $$\frac{1}{\sigma_1\sqrt{1-\rho^2}}\,\phi\left(\frac{(\xi_1-\frac{\sigma_1}{\sigma_2}\rho\xi_2)^2}{\sigma_1\sqrt{1-\rho^2}}\right)$$ and the equality should be \begin{align}\frac{1}{2\pi\sigma_1\sigma_2\sqrt{1-\rho^2}}\exp\left(-\frac{\sigma_1^{-2}\xi_1^2+\sigma_2^{-2}\xi_2^2-2\rho \xi_1\xi_2/\sigma_1\sigma_2}{2}\right)\\=\frac{1}{\sigma_1\sigma_2\sqrt{1-\rho^2}}\phi\left(\frac{(\xi_1-\frac{\sigma_1}{\sigma_2}\rho\xi_2)^2}{\sigma_1\sqrt{1-\rho^2}}\right)\phi\left(\xi_2/\sigma_2\right)\end{align} with $\phi(\cdot)$ denoting the standard Normal pdf.

Reading through this paper, if $(\boldsymbol X_1,\boldsymbol X_2)$ is distributed from a $p$ dimensional multivariate Student's $\mathfrak{t}$ distribution $\mathfrak{t}_{p}(\nu,\boldsymbol\mu,\boldsymbol\Sigma)$ [LaTeX copied from Wikipedia] $$ \frac{\Gamma\left(\frac{\nu+p}{2}\right)}{(\nu\pi)^{\frac{p}{2}}\Gamma(\frac{\nu}{2})\left|{\boldsymbol\Sigma}\right|^{1/2}}\left[1+\frac{1}{\nu}({\mathbf x}-{\boldsymbol\mu})^{\rm T}{\boldsymbol\Sigma}^{-1}({\mathbf x}-{\boldsymbol\mu})\right]^{-\frac{\nu+p}{2}} $$ then both the marginal and the conditional distributions of $\boldsymbol X_1$ given $\boldsymbol X_2$ are also $p_1$ dimensional multivariate Student's $\mathfrak{t}$ distributions: $$\boldsymbol X_1 \sim \mathfrak{t}_{p_1}(\nu,\boldsymbol\mu_1,\boldsymbol\Sigma_{11})$$ and \begin{align}\boldsymbol X_1|\boldsymbol X_2 \sim \mathfrak{t}_{p_1}\big(&\nu+p_2,\boldsymbol\mu_1+\boldsymbol\Sigma_{12}\boldsymbol\Sigma^{-1}_{22}(\boldsymbol X_2−\boldsymbol \mu_2),\\&\dfrac{\nu+(\boldsymbol X_2-\boldsymbol\mu_2)^\text{T}\boldsymbol \Sigma^{−1}_{22}(\boldsymbol X_2-\boldsymbol\mu_2)}{\nu+p_2}\boldsymbol \Sigma_{11|2}\Big) \end{align} where$$\boldsymbol \Sigma_{11|2}=\boldsymbol \Sigma_{11}−\boldsymbol \Sigma_{21}\boldsymbol \Sigma^{−1}_{22}\boldsymbol \Sigma_{12}$$ This is easily proved by using the demarginalisation of the Student's $\mathfrak{t}$ as a mixture of a Normal variate with a chi-squared variate: $$\boldsymbol X|q\sim\mathcal N_p(\boldsymbol\mu,\boldsymbol\Sigma/q),\qquad q\sim\chi^2_\nu/\nu$$

$\endgroup$
3
  • 1
    $\begingroup$ Thank you! I derived it for the bivariate case: $ t_{\nu,2}(x,y;0,C) = \frac{1}{\sqrt{(1-\rho^2)\frac{x^2+\nu}{\nu+1}}} \times t_{\nu,1}(x)\times t_{\nu+1,1}\left(\frac{y-\rho x}{\sqrt{(1-\rho^2)\frac{x^2+\nu}{\nu+1}}}\right). $ $\endgroup$
    – Alex
    May 22, 2020 at 4:39
  • $\begingroup$ Wikipedia also says $\frac{\Gamma\left(\frac{\nu+2}{2}\right)}{(\nu\pi)^{\frac{2}{2}}\Gamma(\frac{\nu}{2})}=\frac{1}{2\pi}$, which means that the bivariate t-distribution is $$\frac{1}{2\pi \left|{\boldsymbol\Sigma}\right|^{1/2}}\left[1+\frac{1}{\nu}({\mathbf x}-{\boldsymbol\mu})^{\top}{\boldsymbol\Sigma}^{-1}({\mathbf x}-{\boldsymbol\mu})\right]^{-\frac{\nu+p}{2}}$$ do any sources derive/confirm this to be the bivariate t-distribution? $\endgroup$
    – develarist
    Dec 7, 2020 at 1:29
  • 1
    $\begingroup$ @Dev Yes, the Wikipedia article on the multivariate t distribution gives this formula with $p=2,$ of course. $\endgroup$
    – whuber
    Dec 7, 2020 at 22:45

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.