Index: The Book of Statistical ProofsStatistical Models ▷ Univariate normal data ▷ Bayesian linear regression ▷ Posterior distribution

Theorem: Let

$\label{eq:GLM} y = X \beta + \varepsilon, \; \varepsilon \sim \mathcal{N}(0, \sigma^2 V)$

be a linear regression model with measured $n \times 1$ data vector $y$, known $n \times p$ design matrix $X$, known $n \times n$ covariance structure $V$ as well as unknown $p \times 1$ regression coefficients $\beta$ and unknown noise variance $\sigma^2$. Moreover, assume a normal-gamma prior distribution over the model parameters $\beta$ and $\tau = 1/\sigma^2$:

$\label{eq:GLM-NG-prior} p(\beta,\tau) = \mathcal{N}(\beta; \mu_0, (\tau \Lambda_0)^{-1}) \cdot \mathrm{Gam}(\tau; a_0, b_0) \; .$

Then, the posterior distribution is also a normal-gamma distribution

$\label{eq:GLM-NG-post} p(\beta,\tau|y) = \mathcal{N}(\beta; \mu_n, (\tau \Lambda_n)^{-1}) \cdot \mathrm{Gam}(\tau; a_n, b_n)$

and the posterior hyperparameters are given by

$\label{eq:GLM-NG-post-par} \begin{split} \mu_n &= \Lambda_n^{-1} (X^\mathrm{T} P y + \Lambda_0 \mu_0) \\ \Lambda_n &= X^\mathrm{T} P X + \Lambda_0 \\ a_n &= a_0 + \frac{n}{2} \\ b_n &= b_0 + \frac{1}{2} (y^\mathrm{T} P y + \mu_0^\mathrm{T} \Lambda_0 \mu_0 - \mu_n^\mathrm{T} \Lambda_n \mu_n) \; . \end{split}$

Proof: According to Bayes’ theorem, the posterior distribution is given by

$\label{eq:GLM-NG-BT} p(\beta,\tau|y) = \frac{p(y|\beta,\tau) \, p(\beta,\tau)}{p(y)} \; .$

Since $p(y)$ is just a normalization factor, the posterior is proportional to the numerator:

$\label{eq:GLM-NG-post-JL} p(\beta,\tau|y) \propto p(y|\beta,\tau) \, p(\beta,\tau) = p(y,\beta,\tau) \; .$

Equation \eqref{eq:GLM} implies the following likelihood function

$\label{eq:GLM-LF-class} p(y|\beta,\sigma^2) = \mathcal{N}(y; X \beta, \sigma^2 V) = \sqrt{\frac{1}{(2 \pi)^n |\sigma^2 V|}} \, \exp\left[ -\frac{1}{2 \sigma^2} (y-X\beta)^\mathrm{T} V^{-1} (y-X\beta) \right]$

which, for mathematical convenience, can also be parametrized as

$\label{eq:GLM-LF-Bayes} p(y|\beta,\tau) = \mathcal{N}(y; X \beta, (\tau P)^{-1}) = \sqrt{\frac{|\tau P|}{(2 \pi)^n}} \, \exp\left[ -\frac{\tau}{2} (y-X\beta)^\mathrm{T} P (y-X\beta) \right]$

using the noise precision $\tau = 1/\sigma^2$ and the $n \times n$ precision matrix $P = V^{-1}$.

Combining the likelihood function \eqref{eq:GLM-LF-Bayes} with the prior distribution \eqref{eq:GLM-NG-prior}, the joint likelihood of the model is given by

$\label{eq:GLM-NG-JL-s1} \begin{split} p(y,\beta,\tau) = \; & p(y|\beta,\tau) \, p(\beta,\tau) \\ = \; & \sqrt{\frac{|\tau P|}{(2 \pi)^n}} \, \exp\left[ -\frac{\tau}{2} (y-X\beta)^\mathrm{T} P (y-X\beta) \right] \cdot \\ & \sqrt{\frac{|\tau \Lambda_0|}{(2 \pi)^p}} \, \exp\left[ -\frac{\tau}{2} (\beta-\mu_0)^\mathrm{T} \Lambda_0 (\beta-\mu_0) \right] \cdot \\ & \frac{ {b_0}^{a_0}}{\Gamma(a_0)} \, \tau^{a_0-1} \exp[-b_0 \tau] \; . \end{split}$

Collecting identical variables gives:

$\label{eq:GLM-NG-JL-s2} \begin{split} p(y,\beta,\tau) = \; & \sqrt{\frac{\tau^{n+p}}{(2 \pi)^{n+p}} |P| |\Lambda_0|} \, \frac{ {b_0}^{a_0}}{\Gamma(a_0)} \, \tau^{a_0-1} \exp[-b_0 \tau] \cdot \\ & \exp\left[ -\frac{\tau}{2} \left( (y-X\beta)^\mathrm{T} P (y-X\beta) + (\beta-\mu_0)^\mathrm{T} \Lambda_0 (\beta-\mu_0) \right) \right] \; . \end{split}$

Expanding the products in the exponent gives:

$\label{eq:GLM-NG-JL-s3} \begin{split} p(y,\beta,\tau) = \; & \sqrt{\frac{\tau^{n+p}}{(2 \pi)^{n+p}} |P| |\Lambda_0|} \, \frac{ {b_0}^{a_0}}{\Gamma(a_0)} \, \tau^{a_0-1} \exp[-b_0 \tau] \cdot \\ & \exp\left[ -\frac{\tau}{2} \left( y^\mathrm{T} P y - y^\mathrm{T} P X \beta - \beta^\mathrm{T} X^\mathrm{T} P y + \beta^\mathrm{T} X^\mathrm{T} P X \beta + \right. \right. \\ & \hphantom{\exp \left[ -\frac{\tau}{2} \right.} \; \left. \left. \beta^\mathrm{T} \Lambda_0 \beta - \beta^\mathrm{T} \Lambda_0 \mu_0 - \mu_0^\mathrm{T} \Lambda_0 \beta + \mu_0^\mathrm{T} \Lambda_0 \mu_0 \right) \right] \; . \end{split}$

Completing the square over $\beta$, we finally have

$\label{eq:GLM-NG-JL-s4} \begin{split} p(y,\beta,\tau) = \; & \sqrt{\frac{\tau^{n+p}}{(2 \pi)^{n+p}} |P| |\Lambda_0|} \, \frac{ {b_0}^{a_0}}{\Gamma(a_0)} \, \tau^{a_0-1} \exp[-b_0 \tau] \cdot \\ & \exp\left[ -\frac{\tau}{2} \left( (\beta-\mu_n)^\mathrm{T} \Lambda_n (\beta-\mu_n) + (y^\mathrm{T} P y + \mu_0^\mathrm{T} \Lambda_0 \mu_0 - \mu_n^\mathrm{T} \Lambda_n \mu_n) \right) \right] \end{split}$

with the posterior hyperparameters

$\label{eq:GLM-NG-post-beta-par} \begin{split} \mu_n &= \Lambda_n^{-1} (X^\mathrm{T} P y + \Lambda_0 \mu_0) \\ \Lambda_n &= X^\mathrm{T} P X + \Lambda_0 \; . \end{split}$

Ergo, the joint likelihood is proportional to

$\label{eq:GLM-NG-JL-s5} p(y,\beta,\tau) \propto \tau^{p/2} \cdot \exp\left[ -\frac{\tau}{2} (\beta-\mu_n)^\mathrm{T} \Lambda_n (\beta-\mu_n) \right] \cdot \tau^{a_n-1} \cdot \exp\left[ -b_n \tau \right]$

with the posterior hyperparameters

$\label{eq:GLM-NG-post-tau-par} \begin{split} a_n &= a_0 + \frac{n}{2} \\ b_n &= b_0 + \frac{1}{2} (y^\mathrm{T} P y + \mu_0^\mathrm{T} \Lambda_0 \mu_0 - \mu_n^\mathrm{T} \Lambda_n \mu_n) \; . \end{split}$

From the term in \eqref{eq:GLM-NG-JL-s5}, we can isolate the posterior distribution over $\beta$ given $\tau$:

$\label{eq:GLM-NG-post-beta} p(\beta|\tau,y) = \mathcal{N}(\beta; \mu_n, (\tau \Lambda_n)^{-1}) \; .$

From the remaining term, we can isolate the posterior distribution over $\tau$:

$\label{eq:GLM-NG-post-tau} p(\tau|y) = \mathrm{Gam}(\tau; a_n, b_n) \; .$

Together, \eqref{eq:GLM-NG-post-beta} and \eqref{eq:GLM-NG-post-tau} constitute the joint posterior distribution of $\beta$ and $\tau$.

Sources:

Metadata: ID: P10 | shortcut: blr-post | author: JoramSoch | date: 2020-01-03, 17:53.