Proof: Log model evidence for the univariate Gaussian
Theorem: Let
\[\label{eq:ug} m: \; y = \left\lbrace y_1, \ldots, y_n \right\rbrace, \quad y_i \sim \mathcal{N}(\mu, \sigma^2), \quad i = 1, \ldots, n\]be a univariate Gaussian data set with unknown mean $\mu$ and unknown variance $\sigma^2$. Moreover, assume a normal-gamma prior distribution over the model parameters $\mu$ and $\tau = 1/\sigma^2$:
\[\label{eq:UG-NG-prior} p(\mu,\tau) = \mathcal{N}(\mu; \mu_0, (\tau \lambda_0)^{-1}) \cdot \mathrm{Gam}(\tau; a_0, b_0) \; .\]Then, the log model evidence for this model is
\[\label{eq:UG-NG-LME} \log p(y|m) = - \frac{n}{2} \log (2 \pi) + \frac{1}{2} \log \frac{\lambda_0}{\lambda_n} + \log \Gamma(a_n) - \log \Gamma(a_0) + a_0 \log b_0 - a_n \log b_n\]where the posterior hyperparameters are given by
\[\label{eq:UG-NG-post-par} \begin{split} \mu_n &= \frac{\lambda_0 \mu_0 + n \bar{y}}{\lambda_0 + n} \\ \lambda_n &= \lambda_0 + n \\ a_n &= a_0 + \frac{n}{2} \\ b_n &= b_0 + \frac{1}{2} (y^\mathrm{T} y + \lambda_0 \mu_0^2 - \lambda_n \mu_n^2) \; . \end{split}\]Proof: According to the law of marginal probability, the model evidence for this model is:
\[\label{eq:UG-NG-ME-s1} p(y|m) = \iint p(y|\mu,\tau) \, p(\mu,\tau) \, \mathrm{d}\mu \, \mathrm{d}\tau \; .\]According to the law of conditional probability, the integrand is equivalent to the joint likelihood:
\[\label{eq:UG-NG-ME-s2} p(y|m) = \iint p(y,\mu,\tau) \, \mathrm{d}\mu \, \mathrm{d}\tau \; .\]Equation \eqref{eq:ug} implies the following likelihood function
\[\label{eq:UG-LF-class} \begin{split} p(y|\mu,\sigma^2) &= \prod_{i=1}^{n} \mathcal{N}(y_i; \mu, \sigma^2) \\ &= \prod_{i=1}^{n} \frac{1}{\sqrt{2 \pi} \sigma} \cdot \exp\left[ -\frac{1}{2} \left( \frac{y_i-\mu}{\sigma} \right)^2 \right] \\ &= \frac{1}{(\sqrt{2 \pi \sigma^2})^n} \cdot \exp\left[ -\frac{1}{2 \sigma^2} \sum_{i=1}^{n} \left( y_i-\mu \right)^2 \right] \end{split}\]which, for mathematical convenience, can also be parametrized as
\[\label{eq:UG-LF-Bayes} \begin{split} p(y|\mu,\tau) &= \prod_{i=1}^{n} \mathcal{N}(y_i; \mu, \tau^{-1}) \\ &= \prod_{i=1}^{n} \sqrt{\frac{\tau}{2 \pi}} \cdot \exp\left[ -\frac{\tau}{2} \left( y_i-\mu \right)^2 \right] \\ &= \left( \sqrt{\frac{\tau}{2 \pi}} \right)^n \cdot \exp\left[ -\frac{\tau}{2} \sum_{i=1}^{n} \left( y_i-\mu \right)^2 \right] \end{split}\]using the inverse variance or precision $\tau = 1/\sigma^2$.
When deriving the posterior distribution $p(\mu,\tau|y)$, the joint likelihood $p(y,\mu,\tau)$ is obtained as
Using the probability density function of the normal distribution, we can rewrite this as
\[\label{eq:UG-NG-LME-s2} \begin{split} p(y,\mu,\tau) = \; & \sqrt{\frac{\tau^{n}}{(2 \pi)^{n}}} \sqrt{\frac{\tau \lambda_0}{2 \pi}} \sqrt{\frac{2 \pi}{\tau \lambda_n}} \, \frac{ {b_0}^{a_0}}{\Gamma(a_0)} \, \tau^{a_0-1} \exp[-b_0 \tau] \cdot \\ & \mathcal{N}(\mu; \mu_n, (\tau \lambda_n)^{-1}) \, \exp\left[ -\frac{\tau}{2} \left( y^\mathrm{T} y + \lambda_0 \mu_0^2 - \lambda_n \mu_n^2 \right) \right] \; . \end{split}\]Now, $\mu$ can be integrated out easily:
\[\label{eq:UG-NG-LME-s3} \begin{split} \int p(y,\mu,\tau) \, \mathrm{d}\mu = \; & \sqrt{\frac{1}{(2 \pi)^{n}}} \sqrt{\frac{\lambda_0}{\lambda_n}} \, \frac{ {b_0}^{a_0}}{\Gamma(a_0)} \, \tau^{a_0+n/2-1} \exp[-b_0 \tau] \cdot \\ & \exp\left[ -\frac{\tau}{2} \left( y^\mathrm{T} y + \lambda_0 \mu_0^2 - \lambda_n \mu_n^2 \right) \right] \; . \end{split}\]Using the probability density function of the gamma distribution, we can rewrite this as
\[\label{eq:UG-NG-LME-s4} \int p(y,\mu,\tau) \, \mathrm{d}\mu = \sqrt{\frac{1}{(2 \pi)^{n}}} \sqrt{\frac{\lambda_0}{\lambda_n}} \, \frac{ {b_0}^{a_0}}{\Gamma(a_0)} \, \frac{\Gamma(a_n)}{ {b_n}^{a_n}} \, \mathrm{Gam}(\tau; a_n, b_n) \; .\]Finally, $\tau$ can also be integrated out:
\[\label{eq:UG-NG-LME-s5} \iint p(y,\mu,\tau) \, \mathrm{d}\mu \, \mathrm{d}\tau = \sqrt{\frac{1}{(2 \pi)^{n}}} \sqrt{\frac{\lambda_0}{\lambda_n}} \, \frac{\Gamma(a_n)}{\Gamma(a_0)} \, \frac{ {b_0}^{a_0}}{ {b_n}^{a_n}} \; .\]Thus, the log model evidence of this model is given by
\[\label{eq:UG-NG-LME-s6} \log p(y|m) = - \frac{n}{2} \log (2 \pi) + \frac{1}{2} \log \frac{\lambda_0}{\lambda_n} + \log \Gamma(a_n) - \log \Gamma(a_0) + a_0 \log b_0 - a_n \log b_n \; .\]- Bishop CM (2006): "Bayesian linear regression"; in: Pattern Recognition for Machine Learning, pp. 152-161, ex. 3.23, eq. 3.118; URL: http://users.isr.ist.utl.pt/~wurmd/Livros/school/Bishop%20-%20Pattern%20Recognition%20And%20Machine%20Learning%20-%20Springer%20%202006.pdf.
Metadata: ID: P203 | shortcut: ug-lme | author: JoramSoch | date: 2021-03-03, 10:25.