0
$\begingroup$

Given $n$ normally distributed observations $f(X|\sigma_X)=\mathcal{N}(\mu_X, \sigma_X^2)$ and assuming a uniform prior on $\log(\sigma_X)$ and known $\mu_x$, I'm trying to find marginal posterior distribution for $\sigma^2_X$. I know how to approach for finding marginal posterior distribution for $\sigma_X$, but not $\sigma^2_X$ and even in that case, I don't find any nice form for the posterior, so I was not sure if my approach is correct.

Here is my try:

$$P(\sigma_X|X)=P(X|\sigma_X)P(\sigma_X)$$

$$f(log(\sigma_X))\propto 1 \qquad \text{therefore, }\qquad f(\sigma_X)\propto\frac{1}{\sigma_X}$$

$$P(\sigma_X|X)\propto \frac{1}{\sigma_X}\prod_i^n\frac{1}{\sqrt{2\pi\sigma^2_X}}\exp(-\frac{(x_i-\mu_X)^2}{2\sigma_X^2})$$

Which seems to me that is no particular distribution form. My questions are:

1) Is my interpretation of the uniform prior on log scale correct?

2) Does this final form represent any particular distribution function?

3) Is there a way to find marginal posterior distribution for $\sigma^2_X$ or should I assume that this was a typo and we only can find marginal posterior distribution for $\sigma_X$?

$\endgroup$
  • $\begingroup$ You need to integrate out $\mu_X$ $\endgroup$ – Xi'an Sep 30 '18 at 15:17
  • $\begingroup$ In general, yes. But $\mu_X$ is a known value here. $\endgroup$ – Blade Sep 30 '18 at 15:29
1
$\begingroup$

In the case where $\mu$ is known, there is no "marginal" posterior distribution, only a posterior distribution, and you already have it. To see this, start out by changing your prior to be on $\sigma^2$; as it happens, making it uniform on $\log \sigma^2$ gives you $f(\sigma^2) \propto 1/\sigma$, as before. (The answer to your first question is yes, you are interpreting the uniform prior on the log scale correctly.) Your posterior is then $P(\sigma^2|X)$ with exactly the same functional form as you have already derived; perform the product multiplication from $1$ to $n$ to get:

$$P(\sigma^2|X,\mu) \propto (\sigma^2)^{-(n+1)/2}\exp\left(-{\sum(x_i-\mu)^2 \over 2\sigma^2}\right)$$

which is simply a rewritten version of what you already have.

As for what distribution this is, if you look at the functional form, it looks a lot like that of a Gamma distribution, just with the variable of interest in the denominator everywhere instead of the numerator. This leads us to the inverse gamma distribution (Wikipedia link.)

$\endgroup$

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.