18
$\begingroup$

This is a very simple question but I can't find the derivation anywhere on the internet or in a book. I would like to see the derivation of how one Bayesian updates a multivariate normal distribution. For example: imagine that

$$ \begin{array}{rcl} \mathbb{P}({\bf x}|{\bf μ},{\bf Σ}) & = & N({\bf \mu}, {\bf \Sigma}) \\ \mathbb{P}({\bf \mu}) &= & N({\bf \mu_0}, {\bf \Sigma_0})\,. \end{array} $$

After observing a set of ${\bf x_1 ... x_n}$, I would like to compute $\mathbb{P}({\bf \mu | x_1 ... x_n})$. I know that the answer is $\mathbb{P}({\bf \mu | x_1 ... x_n}) = N({\bf \mu_n}, {\bf \Sigma_n})$ where

$$ \begin{array}{rcl} \bf \mu_n &=& \displaystyle\Sigma_0 \left(\Sigma_0 + \frac{1}{n}\Sigma\right)^{-1}\left(\frac{1}{n}\sum_{i=1}^{n}{\bf x_i}\right) + \frac{1}{n}\Sigma\left(\Sigma_0+\frac{1}{n}\Sigma\right)^{-1}\mu_0 \\ \bf \Sigma_n & =&\displaystyle \Sigma_0\left(\Sigma_0 + \frac{1}{n}\Sigma\right)^{-1}\frac{1}{n}\Sigma \end{array}$$

I am looking for the derivation of this result with all the intermediate matrix algebra.

Any help is much appreciated.

$\endgroup$
  • 2
    $\begingroup$ It is also solved in our book Bayesian Core, Chap. 3, Section 3.2, pages 54-57 with what we think is detailed matrix algebra! $\endgroup$ – Xi'an May 18 '12 at 20:10
  • 1
    $\begingroup$ The OP said it was not a homework problem and even explained why he asked it and how he wants to use the answer. Why not post it for others? I understand why we do not want to provide a homework problem solving service but this is taking it a little too far. $\endgroup$ – Michael R. Chernick May 18 '12 at 23:11
  • 3
    $\begingroup$ @Alex: Sorry, wrong link, I meant Bayesian Core. Note that we also posted solutions to all problems on arXiv. So posting a complete solution here would not hurt! $\endgroup$ – Xi'an May 19 '12 at 7:56
  • 1
    $\begingroup$ I have deleted the portion of the comments that amount to a private exchange between individuals with an arrangement to share a private answer to the question. That sort of thing is abusing this site, which is all about public questions and public answers. $\endgroup$ – whuber May 19 '12 at 14:20
  • 1
    $\begingroup$ Just as an FYI, the derivation is in Pattern Classification by Duda, Hart and Stork. However, I was having difficulty following some of their steps which only matters to me. If this was simply homework one could just write down exactly what they have. $\endgroup$ – Alex May 19 '12 at 16:46
6
$\begingroup$

With the distributions on our random vectors:

$\mathbf x_i | \mathbf \mu \sim N(\mu , \mathbf \Sigma)$

$\mathbf \mu \sim N(\mathbf \mu_0, \mathbf \Sigma_0)$

By Bayes's rule the posterior distribution looks like:

$p(\mu| \{\mathbf x_i\}) \propto p(\mu) \prod_{i=1}^N p(\mathbf x_i | \mu)$

So:

$\ln p(\mu| \{\mathbf x_i\}) = -\frac{1}{2}\sum_{i=1}^N(\mathbf x_i - \mu)'\mathbf \Sigma^{-1}(\mathbf x_i - \mu) -\frac{1}{2}(\mu - \mu_0)'\mathbf \Sigma_0^{-1}(\mu - \mu_0) + const$

$ = -\frac{1}{2} N \mu' \mathbf \Sigma^{-1} \mu + \sum_{i=1}^N \mu' \mathbf \Sigma^{-1} \mathbf x_i -\frac{1}{2} \mu' \mathbf \Sigma_0^{-1} \mu + \mu' \mathbf \Sigma_0^{-1} \mu_0 + const$

$ = -\frac{1}{2} \mu' (N \mathbf \Sigma^{-1} + \mathbf \Sigma_0^{-1}) \mu + \mu' (\mathbf \Sigma_0^{-1} \mu_0 + \mathbf \Sigma^{-1} \sum_{i=1}^N \mathbf x_i) + const$

$= -\frac{1}{2}(\mu - (N \mathbf \Sigma^{-1} + \mathbf \Sigma_0^{-1})^{-1}(\mathbf \Sigma_0^{-1} \mu_0 + \mathbf \Sigma^{-1} \sum_{i=1}^N \mathbf x_i))' (N \mathbf \Sigma^{-1} + \mathbf \Sigma_0^{-1}) (\mu - (N \mathbf \Sigma^{-1} + \mathbf \Sigma_0^{-1})^{-1}(\mathbf \Sigma_0^{-1} \mu_0 + \mathbf \Sigma^{-1} \sum_{i=1}^N \mathbf x_i)) + const$

Which is the log density of a Gaussian:

$\mu| \{\mathbf x_i\} \sim N((N \mathbf \Sigma^{-1} + \mathbf \Sigma_0^{-1})^{-1}(\mathbf \Sigma_0^{-1} \mu_0 + \mathbf \Sigma^{-1} \sum_{i=1}^N \mathbf x_i), (N \mathbf \Sigma^{-1} + \mathbf \Sigma_0^{-1})^{-1})$

Using the Woodbury identity on our expression for the covariance matrix:

$(N \mathbf \Sigma^{-1} + \mathbf \Sigma_0^{-1})^{-1} = \mathbf \Sigma(\frac{1}{N} \mathbf \Sigma + \mathbf \Sigma_0)^{-1} \frac{1}{N} \mathbf \Sigma_0$

Which provides the covariance matrix in the form the OP wanted. Using this expression (and its symmetry) further in the expression for the mean we have:

$\mathbf \Sigma(\frac{1}{N} \mathbf \Sigma + \mathbf \Sigma_0)^{-1} \frac{1}{N} \mathbf \Sigma_0 \mathbf \Sigma_0^{-1} \mu_0 + \frac{1}{N} \mathbf \Sigma_0(\frac{1}{N} \mathbf \Sigma + \mathbf \Sigma_0)^{-1} \mathbf \Sigma \mathbf \Sigma^{-1} \sum_{i=1}^N \mathbf x_i$

$= \mathbf \Sigma(\frac{1}{N} \mathbf \Sigma + \mathbf \Sigma_0)^{-1} \frac{1}{N} \mu_0 + \mathbf \Sigma_0(\frac{1}{N} \mathbf \Sigma + \mathbf \Sigma_0)^{-1} \sum_{i=1}^N (\frac{1}{N} \mathbf x_i)$

Which is the form required by the OP for the mean.

$\endgroup$

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.