Distributions Provided by mniw

Martin Lysy

2024-09-19

Wishart Distribution

The Wishart distribution on a random positive-definite matrix ${\boldsymbol{X}}_{q\times q}$ is is denoted ${\boldsymbol{X}}\sim \operatorname{Wish}({\boldsymbol{\Psi}}, \nu)$, and defined as ${\boldsymbol{X}}= ({\boldsymbol{L}}{\boldsymbol{Z}})({\boldsymbol{L}}{\boldsymbol{Z}})'$, where:

• ${\boldsymbol{\Psi}}_{q\times q} = {\boldsymbol{L}}{\boldsymbol{L}}'$ is the positive-definite matrix scale parameter,

• $\nu > q$ is the shape parameter,

• ${\boldsymbol{Z}}_{q\times q}$ is a random lower-triangular matrix with elements

  $$Z_{ij} \begin{cases} \overset{\;\textrm{iid}\;}{\sim}\operatorname{Normal}(0,1) & i < j \\ \overset{\:\textrm{ind}\:}{\sim}\chi^2_{(\nu-i+1)} & i = j \\ = 0 & i > j. \end{cases}$$

The log-density of the Wishart distribution is

$$\log p({\boldsymbol{X}}\mid {\boldsymbol{\Psi}}, \nu) = -\textstyle{\frac{1}{2}} \left[\mathrm{tr}({\boldsymbol{\Psi}}^{-1} {\boldsymbol{X}}) + (q+1-\nu)\log |{\boldsymbol{X}}| + \nu \log |{\boldsymbol{\Psi}}| + \nu q \log(2) + 2 \log \Gamma_q(\textstyle{\frac{\nu }{2}})\right],$$

where $\Gamma_n(x)$ is the multivariate Gamma function defined as

$$\Gamma_n(x) = \pi^{n(n-1)/4} \prod_{j=1}^n \Gamma\big(x + \textstyle{\frac{1}{2}} (1-j)\big).$$

Inverse-Wishart Distribution

The Inverse-Wishart distribution ${\boldsymbol{X}}\sim \operatorname{InvWish}({\boldsymbol{\Psi}}, \nu)$ is defined as ${\boldsymbol{X}}^{-1} \sim \operatorname{Wish}({\boldsymbol{\Psi}}^{-1}, \nu)$. Its log-density is given by

$$\log p({\boldsymbol{X}}\mid {\boldsymbol{\Psi}}, \nu) = -\textstyle{\frac{1}{2}} \left[\mathrm{tr}({\boldsymbol{\Psi}}{\boldsymbol{X}}^{-1}) + (\nu+q+1) \log |{\boldsymbol{X}}| - \nu \log |{\boldsymbol{\Psi}}| + \nu q \log(2) + 2 \log \Gamma_q(\textstyle{\frac{\nu }{2}})\right].$$

Properties

If ${\boldsymbol{X}}_{q\times q} \sim \operatorname{Wish}({\boldsymbol{\Psi}},\nu)$, the for a nonzero vector ${\boldsymbol{a}}\in \mathbb R^q$ we have

$$\frac{{\boldsymbol{a}}'{\boldsymbol{X}}{\boldsymbol{a}}}{{\boldsymbol{a}}'{\boldsymbol{\Psi}}{\boldsymbol{a}}} \sim \chi^2_{(\nu)}.$$

Matrix-Normal Distribution

The Matrix-Normal distribution on a random matrix ${\boldsymbol{X}}_{p \times q}$ is denoted ${\boldsymbol{X}}\sim \operatorname{MatNorm}({\boldsymbol{\Lambda}}, {\boldsymbol{\Sigma}}_R, {\boldsymbol{\Sigma}}_C)$, and defined as ${\boldsymbol{X}}= {\boldsymbol{L}}{\boldsymbol{Z}}{\boldsymbol{U}}+ {\boldsymbol{\Lambda}}$, where:

• ${\boldsymbol{\Lambda}}_{p \times q}$ is the mean matrix parameter,
• ${{\boldsymbol{\Sigma}}_R}_{p \times p} = {\boldsymbol{L}}{\boldsymbol{L}}'$ is the row-variance matrix parameter,
• ${{\boldsymbol{\Sigma}}_C}_{q \times q} = {\boldsymbol{U}}'{\boldsymbol{U}}$ is the column-variance matrix parameter,
• ${\boldsymbol{Z}}_{q\times q}$ is a random matrix with $Z_{ij} \overset{\;\textrm{iid}\;}{\sim}\operatorname{Normal}(0,1)$.

The log-density of the Matrix-Normal distribution is

$$\log p({\boldsymbol{X}}\mid {\boldsymbol{\Lambda}}, {\boldsymbol{\Sigma}}_R, {\boldsymbol{\Sigma}}_C) = -\textstyle{\frac{1}{2}} \left[\mathrm{tr}\big({\boldsymbol{\Sigma}}_C^{-1}({\boldsymbol{X}}-{\boldsymbol{\Lambda}})'{\boldsymbol{\Sigma}}_R^{-1}({\boldsymbol{X}}-{\boldsymbol{\Lambda}})\big) + \nu q \log(2\pi) + \nu \log |{\boldsymbol{\Sigma}}_C| + q \log |{\boldsymbol{\Sigma}}_R|\right].$$

Properties

If ${\boldsymbol{X}}_{p \times q} \sim \operatorname{MatNorm}({\boldsymbol{\Lambda}}, {\boldsymbol{\Sigma}}_R, {\boldsymbol{\Sigma}}_C)$, then for nonzero vectors ${\boldsymbol{a}}\in \mathbb R^p$ and ${\boldsymbol{b}}\in \mathbb R^q$ we have

$${\boldsymbol{a}}' {\boldsymbol{X}}{\boldsymbol{b}}\sim \operatorname{Normal}({\boldsymbol{a}}' {\boldsymbol{\Lambda}}{\boldsymbol{b}}, {\boldsymbol{a}}'{\boldsymbol{\Sigma}}_R{\boldsymbol{a}}\cdot {\boldsymbol{b}}'{\boldsymbol{\Sigma}}_C{\boldsymbol{b}}).$$

Matrix-Normal Inverse-Wishart Distribution

The Matrix-Normal Inverse-Wishart Distribution on a random matrix ${\boldsymbol{X}}_{p \times q}$ and random positive-definite matrix ${\boldsymbol{V}}_{q\times q}$ is denoted $({\boldsymbol{X}},{\boldsymbol{V}}) \sim \operatorname{MNIW}({\boldsymbol{\Lambda}}, {\boldsymbol{\Sigma}}, {\boldsymbol{\Psi}}, \nu)$, and defined as

$$\begin{aligned} {\boldsymbol{X}}\mid {\boldsymbol{V}}& \sim \operatorname{MatNorm}({\boldsymbol{\Lambda}}, {\boldsymbol{\Sigma}}, {\boldsymbol{V}}) \\ {\boldsymbol{V}}& \sim \operatorname{InvWish}({\boldsymbol{\Psi}}, \nu). \end{aligned}$$

Properties

The MNIX distribution is conjugate prior for the multivariable response regression model

$${\boldsymbol{Y}}_{n \times q} \sim \operatorname{MatNorm}({\boldsymbol{X}}_{n\times p} {\boldsymbol{\beta}}_{p \times q}, {\boldsymbol{V}}, {\boldsymbol{\Sigma}}).$$

That is, if $({\boldsymbol{\beta}}, {\boldsymbol{\Sigma}}) \sim \operatorname{MNIW}({\boldsymbol{\Lambda}}, {\boldsymbol{\Omega}}^{-1}, {\boldsymbol{\Psi}}, \nu)$, then

$${\boldsymbol{\beta}}, {\boldsymbol{\Sigma}}\mid {\boldsymbol{Y}}\sim \operatorname{MNIW}(\hat {\boldsymbol{\Lambda}}, \hat {\boldsymbol{\Omega}}^{-1}, \hat {\boldsymbol{\Psi}}, \hat \nu),$$

where

$$\begin{aligned} \hat {\boldsymbol{\Omega}}& = {\boldsymbol{X}}'{\boldsymbol{V}}^{-1}{\boldsymbol{X}}+ {\boldsymbol{\Omega}} & \hat {\boldsymbol{\Psi}}& = {\boldsymbol{\Psi}}+ {\boldsymbol{Y}}'{\boldsymbol{V}}^{-1}{\boldsymbol{Y}}+ {\boldsymbol{\Lambda}}'{\boldsymbol{\Omega}}{\boldsymbol{\Lambda}}- \hat {\boldsymbol{\Lambda}}'\hat {\boldsymbol{\Omega}}\hat {\boldsymbol{\Lambda}} \\ \hat {\boldsymbol{\Lambda}}& = \hat {\boldsymbol{\Omega}}^{-1}({\boldsymbol{X}}'{\boldsymbol{V}}^{-1}{\boldsymbol{Y}}+ {\boldsymbol{\Omega}}{\boldsymbol{\Lambda}}) & \hat \nu & = \nu + n. \end{aligned}$$

Matrix-t Distribution

The Matrix-$t$ distribution on a random matrix ${\boldsymbol{X}}_{p \times q}$ is denoted ${\boldsymbol{X}}\sim \operatorname{MatT}({\boldsymbol{\Lambda}}, {\boldsymbol{\Sigma}}_R, {\boldsymbol{\Sigma}}_C, \nu)$, and defined as the marginal distribution of ${\boldsymbol{X}}$ for $({\boldsymbol{X}}, {\boldsymbol{V}}) \sim \operatorname{MNIW}({\boldsymbol{\Lambda}}, {\boldsymbol{\Sigma}}_R, {\boldsymbol{\Sigma}}_C, \nu)$. Its log-density is given by

$$\begin{aligned} \log p({\boldsymbol{X}}\mid {\boldsymbol{\Lambda}}, {\boldsymbol{\Sigma}}_R, {\boldsymbol{\Sigma}}_C, \nu) & = -\textstyle{\frac{1}{2}} \Big[(\nu+p+q-1)\log | I + {\boldsymbol{\Sigma}}_R^{-1}({\boldsymbol{X}}-{\boldsymbol{\Lambda}}){\boldsymbol{\Sigma}}_C^{-1}({\boldsymbol{X}}-{\boldsymbol{\Lambda}})'| \\ & \phantom{= -\textstyle{\frac{1}{2}} \Big[} + q \log |{\boldsymbol{\Sigma}}_R| + p \log |{\boldsymbol{\Sigma}}_C| \\ & \phantom{= -\textstyle{\frac{1}{2}} \Big[} + pq \log(\pi) - \log \Gamma_q(\textstyle{\frac{\nu+p+q-1}{2}}) + \log \Gamma_q(\textstyle{\frac{\nu+q-1}{2}})\Big]. \end{aligned}$$

Properties

If ${\boldsymbol{X}}_{p\times q} \sim \operatorname{MatT}({\boldsymbol{\Lambda}}, {\boldsymbol{\Sigma}}_R, {\boldsymbol{\Sigma}}_C, \nu)$, then for nonzero vectors ${\boldsymbol{a}}\in \mathbb R^p$ and ${\boldsymbol{b}}\in \mathbb R^q$ we have

$$\frac{{\boldsymbol{a}}'{\boldsymbol{X}}{\boldsymbol{b}}- \mu}{\sigma} \sim t_{(\nu -q + 1)},$$

where $$\mu = {\boldsymbol{a}}'{\boldsymbol{\Lambda}}{\boldsymbol{b}}, \qquad \sigma^2 = \frac{{\boldsymbol{a}}'{\boldsymbol{\Sigma}}_R{\boldsymbol{a}}\cdot {\boldsymbol{b}}'{\boldsymbol{\Sigma}}_C{\boldsymbol{b}}}{\nu - q + 1}.$$

Random-Effects Normal Distribution

Consider the multivariate normal distribution on $q$-dimensional vectors ${\boldsymbol{x}}$ and ${\boldsymbol{\mu}}$ given by

$$\begin{aligned} {\boldsymbol{x}}\mid {\boldsymbol{\mu}}& \sim \operatorname{Normal}({\boldsymbol{\mu}}, {\boldsymbol{V}}) \\ {\boldsymbol{\mu}}& \sim \operatorname{Normal}({\boldsymbol{\lambda}}, {\boldsymbol{\Sigma}}). \end{aligned}$$

The random-effects normal distribution is defined as the posterior distribution ${\boldsymbol{\mu}}\sim p({\boldsymbol{\mu}}\mid {\boldsymbol{x}})$, which is given by

$${\boldsymbol{\mu}}\mid {\boldsymbol{x}}\sim \operatorname{Normal}\big({\boldsymbol{G}}({\boldsymbol{x}}-{\boldsymbol{\lambda}}) + {\boldsymbol{\lambda}}, {\boldsymbol{G}}{\boldsymbol{V}}\big), \qquad {\boldsymbol{G}}= {\boldsymbol{\Sigma}}({\boldsymbol{V}}+ {\boldsymbol{\Sigma}})^{-1}.$$

The notation for this distribution is ${\boldsymbol{\mu}}\sim \operatorname{RxNorm}({\boldsymbol{x}}, {\boldsymbol{V}}, {\boldsymbol{\lambda}}, {\boldsymbol{\Sigma}})$.

Hierarchical Normal-Normal Model

The hierarchical Normal-Normal model is defined as

$$\begin{aligned} {\boldsymbol{y}}_i \mid {\boldsymbol{\mu}}_i, {\boldsymbol{\beta}}, {\boldsymbol{\Sigma}}& \overset{\:\textrm{ind}\:}{\sim}\operatorname{Normal}({\boldsymbol{\mu}}_i, {\boldsymbol{V}}_i) \\ {\boldsymbol{\mu}}_i \mid {\boldsymbol{\beta}}, {\boldsymbol{\Sigma}}& \overset{\;\textrm{iid}\;}{\sim}\operatorname{Normal}({\boldsymbol{x}}_i'{\boldsymbol{\beta}}, {\boldsymbol{\Sigma}}) \\ ({\boldsymbol{\beta}}, {\boldsymbol{\Sigma}}) & \sim \operatorname{MNIW}({\boldsymbol{\Lambda}}, {\boldsymbol{\Omega}}^{-1}, {\boldsymbol{\Psi}}, \nu), \end{aligned}$$

where:

• ${{\boldsymbol{y}}_i}_{q\times 1}$ is the response vector for subject $i$,
• ${{\boldsymbol{\mu}}_i}_{q\times 1}$ is the random effect for subject $i$,
• ${{\boldsymbol{V}}_i}_{q\times q}$ is the error variance for subject $i$,
• ${{\boldsymbol{x}}_i}_{p\times 1}$ is the covariate vector for subject $i$,
• ${\boldsymbol{\beta}}_{p \times q}$ is the random-effects coefficient matrix,
• ${\boldsymbol{\Sigma}}_{q \times q}$ is the random-effects error variance.

Let ${\boldsymbol{Y}}_{n\times q} = ({\boldsymbol{y}}_{1},\ldots,{\boldsymbol{y}}_{n})$, ${\boldsymbol{X}}_{n\times p} = ({\boldsymbol{x}}_{1},\ldots,{\boldsymbol{x}}_{n})$, and ${\boldsymbol{\Theta}}_{n \times q} = ({\boldsymbol{\mu}}_{1},\ldots,{\boldsymbol{\mu}}_{n})$. If interest lies in the posterior distribution $p({\boldsymbol{\Theta}}, {\boldsymbol{\beta}}, {\boldsymbol{\Sigma}}\mid {\boldsymbol{Y}}, {\boldsymbol{X}})$, then a Gibbs sampler can be used to cycle through the following conditional distributions:

$$\begin{aligned} {\boldsymbol{\mu}}_i \mid {\boldsymbol{\beta}}, {\boldsymbol{\Sigma}}, {\boldsymbol{Y}}, {\boldsymbol{X}}& \overset{\:\textrm{ind}\:}{\sim}\operatorname{RxNorm}({\boldsymbol{y}}_i, {\boldsymbol{V}}_i, {\boldsymbol{x}}_i'{\boldsymbol{\beta}}, {\boldsymbol{\Sigma}}) \\ {\boldsymbol{\beta}}, {\boldsymbol{\Sigma}}\mid {\boldsymbol{\Theta}}, {\boldsymbol{Y}}, {\boldsymbol{X}}& \sim \operatorname{MNIW}(\hat {\boldsymbol{\Lambda}}, \hat {\boldsymbol{\Omega}}^{-1}, \hat {\boldsymbol{\Psi}}, \hat \nu), \end{aligned}$$

where $\hat {\boldsymbol{\Lambda}}$, $\hat {\boldsymbol{\Omega}}$, $\hat {\boldsymbol{\Psi}}$, and $\hat \nu$ are obtained from the MNIW conjugate posterior formula with ${\boldsymbol{Y}}\gets {\boldsymbol{\Theta}}$.