Index: The Book of Statistical ProofsStatistical ModelsUnivariate normal dataBayesian linear regression with known covariance ▷ Posterior distribution

Theorem: Let

\[\label{eq:GLM} y = X \beta + \varepsilon, \; \varepsilon \sim \mathcal{N}(0, \Sigma)\]

be a linear regression model with measured $n \times 1$ data vector $y$, known $n \times p$ design matrix $X$ and known $n \times n$ covariance matrix $\Sigma$ as well as unknown $p \times 1$ regression coefficients $\beta$. Moreover, assume a multivariate normal distribution over the model parameter $\beta$:

\[\label{eq:GLM-N-prior} p(\beta) = \mathcal{N}(\beta; \mu_0, \Sigma_0) \; .\]

Then, the posterior distribution is also a multivariate normal distribution

\[\label{eq:GLM-N-post} p(\beta|y) = \mathcal{N}(\beta; \mu_n, \Sigma_n)\]

and the posterior hyperparameters are given by

\[\label{eq:GLM-N-post-par} \begin{split} \mu_n &= \Sigma_n (X^\mathrm{T} \Sigma^{-1} y + \Sigma_0^{-1} \mu_0) \\ \Sigma_n &= \left( X^\mathrm{T} \Sigma^{-1} X + \Sigma_0^{-1} \right)^{-1} \; . \end{split}\]

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

\[\label{eq:GLM-N-BT} p(\beta|y) = \frac{p(y|\beta) \, p(\beta)}{p(y)} \; .\]

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

\[\label{eq:GLM-N-post-JL} p(\beta|y) \propto p(y|\beta) \, p(\beta) = p(y,\beta) \; .\]

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

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

Combining the likelihood function \eqref{eq:GLM-LF} with the prior distribution \eqref{eq:GLM-N-prior} using the probability density function of the multivariate normal distribution, the joint likelihood of the model is given by

\[\label{eq:GLM-N-JL-s1} \begin{split} p(y,\beta) = \; & p(y|\beta) \, p(\beta) \\ = \; & \sqrt{\frac{1}{(2 \pi)^n |\Sigma|}} \, \exp\left[ -\frac{1}{2} (y-X\beta)^\mathrm{T} \Sigma^{-1} (y-X\beta) \right] \cdot \\ \; & \sqrt{\frac{1}{(2 \pi)^p |\Sigma_0|}} \, \exp\left[ -\frac{1}{2} (\beta-\mu_0)^\mathrm{T} \Sigma_0^{-1} (\beta-\mu_0) \right] \; . \end{split}\]

Collecting identical variables gives:

\[\label{eq:GLM-N-JL-s2} \begin{split} p(y,\beta) = \; & \sqrt{\frac{1}{(2 \pi)^{n+p} |\Sigma| |\Sigma_0|}} \cdot \\ & \exp\left[ -\frac{1}{2} \left( (y-X\beta)^\mathrm{T} \Sigma^{-1} (y-X\beta) + (\beta-\mu_0)^\mathrm{T} \Sigma_0^{-1} (\beta-\mu_0) \right) \right] \; . \end{split}\]

Expanding the products in the exponent gives:

\[\label{eq:GLM-N-JL-s3} \begin{split} p(y,\beta) = \; & \sqrt{\frac{1}{(2 \pi)^{n+p} |\Sigma| |\Sigma_0|}} \cdot \\ & \exp\left[ -\frac{1}{2} \left( y^\mathrm{T} \Sigma^{-1} y - y^\mathrm{T} \Sigma^{-1} X \beta - \beta^\mathrm{T} X^\mathrm{T} \Sigma^{-1} y + \beta^\mathrm{T} X^\mathrm{T} \Sigma^{-1} X \beta + \right. \right. \\ & \hphantom{\exp \left[ -\frac{1}{2} \right.} \; \left. \left. \beta^\mathrm{T} \Sigma_0^{-1} \beta - \beta^\mathrm{T} \Sigma_0^{-1} \mu_0 - \mu_0^\mathrm{T} \Sigma_0^{-1} \beta + \mu_0^\mathrm{T} \Sigma_0^{-1} \mu_0 \right) \right] \; . \end{split}\]

Regrouping the terms in the exponent gives:

\[\label{eq:GLM-N-JL-s4} \begin{split} p(y,\beta) = \; & \sqrt{\frac{1}{(2 \pi)^{n+p} |\Sigma| |\Sigma_0|}} \cdot \\ & \exp\left[ -\frac{1}{2} \left( \beta^\mathrm{T} [ X^\mathrm{T} \Sigma^{-1} X + \Sigma_0^{-1} ] \beta - 2 \beta^\mathrm{T} [X^\mathrm{T} \Sigma^{-1} y + \Sigma_0^{-1} \mu_0] + \right. \right. \\ & \hphantom{\exp \left[ -\frac{1}{2} \right.} \; \left. \left. y^\mathrm{T} \Sigma^{-1} y + \mu_0^\mathrm{T} \Sigma_0^{-1} \mu_0 \right) \right] \; . \end{split}\]

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

\[\label{eq:GLM-N-JL-s5} \begin{split} p(y,\beta) = \; & \sqrt{\frac{1}{(2 \pi)^{n+p} |\Sigma| |\Sigma_0|}} \cdot \\ & \exp\left[ -\frac{1}{2} \left( (\beta-\mu_n)^\mathrm{T} \Sigma_n^{-1} (\beta-\mu_n) + (y^\mathrm{T} \Sigma^{-1} y + \mu_0^\mathrm{T} \Sigma_0^{-1} \mu_0 - \mu_n^\mathrm{T} \Sigma_n^{-1} \mu_n) \right) \right] \end{split}\]

with the posterior hyperparameters

\[\label{eq:GLM-N-post-par-qed} \begin{split} \mu_n &= \Sigma_n (X^\mathrm{T} \Sigma^{-1} y + \Sigma_0^{-1} \mu_0) \\ \Sigma_n &= \left( X^\mathrm{T} \Sigma^{-1} X + \Sigma_0^{-1} \right)^{-1} \; . \end{split}\]

Ergo, the joint likelihood is proportional to

\[\label{eq:GLM-N-JL-s6} p(y,\beta) \propto \exp\left[ -\frac{1}{2} (\beta-\mu_n)^\mathrm{T} \Sigma_n^{-1} (\beta-\mu_n) \right] \; ,\]

such that the posterior distribution over $\beta$ is given by

\[\label{eq:GLM-N-post-qed} p(\beta|y) = \mathcal{N}(\beta; \mu_n, \Sigma_n)\]

with the posterior hyperparameters given in \eqref{eq:GLM-N-post-par-qed}.

Sources:

Metadata: ID: P433 | shortcut: blrkc-post | author: JoramSoch | date: 2024-01-19, 08:51.