I have been learning variational bayesian mixture model. Appearantly, beta / dirichlet distribution are used as prior distribution. Why use these two distribution as priori and what is the principle behind this?

  • $\begingroup$ They are used for convenience reason. $\endgroup$ – Xi'an Jan 7 '19 at 10:20

A prior distribution represents the prior believe concerning how likely different values of the parameter are. So at its core, a beta or Dirichlet ( [pedantic mode on] called after Johann Dirichlet, so writen with a capital letter [pedantic mode off] ) distribution is chosen by someone because that person thinks that it represents her or his believes. If your believes differ than you choose a different distribution.

However, not all distributions are equally sensible. The mixture parameter is a proportion that must remain between 0 and 1, so a normal distribution would not be the most sensible prior distribution for that parameter because that would assign a positive probility to impossible values (less than 0 or more than 1). A beta distirbution is constrained between 0 and 1 and fairly flexible, so you can represent a lot of different believes with that one distribution. This makes it a convenient choice. A Dirichlet distribution also ensures that the parameters add up to 1, which is typically required, though its correlation structure is somewhat restrictive. In that case it is a convenient choice.

  • $\begingroup$ Thanks for explaining. I thought beta / Dirichlet is used as priori not just because they are convenient (ranging from 0 to 1 representing probability), but also there should be some properties of these distribution that logically / methmatically represent our prior lelieve, right? $\endgroup$ – unicorn Jan 8 '19 at 1:21

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.