It is my understanding that if we have an improper prior, we may either get a proper or a posterior distribution. My question is if there are cases where a proper prior distribution may, when combined with the likelihood function, result in an improper posterior. Are there any trivial cases? Thanks!


1 Answer 1


I think that everything you need is contained in the several answers for this related but slighly different question (Does the Bayesian posterior need to be a proper distribution?). Here is a proposal of summary for your specific question:

From the sketch of proof https://stats.stackexchange.com/a/89157/14346: if the prior is proper and $p(x|\theta)$ none degenerate then the set of observations for which the posterior is improper is a Lesbegue null set. In other words: $$ \mbox{proper prior} \Rightarrow \mbox{proper posterior almost everywhere}. $$ So, there exists some observations $x$ (but a null set) leading to an improper $p(\theta|x)$. An example of such $x$ for a given $p(\theta|x)$ is given here https://stats.stackexchange.com/a/89150/14346.

Nevertheless, (as pointed out in the comment of the example below) density functions are equivalent (roughly, can be treated as the same) if they only differ on a null set. So the specific observations leading to improperness are complelety arbitrary depending on which specific forms of the density you choose from the equivalent sets.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.