I often read that Pitman-Yor Process has power-law properties. Let's say I am interested in modelling English word's distribution (which follows power-law). Using CRP metaphor, words come and get assigned to tables using CRP probabilities. Now I want to draw samples from this CRP to show that it actually has powe-law properties. How would I conduct such sampling?

What I thought that would be correct, is to draw a random number from a uniform U[0,1] and multiply it by (number_of_customers_in_restaurant-1+concentration), and on each table use (number_of_customers_on_table-d) to pick the table to sample from. Does this seem correct?

Ed.1) I think this is not correct, because what I proposed looks like sampling from a Multinomial Distribution. What is missing here, is the probability that comes from the base. So, I guess instead of picking a table proportional to (number_of_customers_on_table-d), I think I should use (number_of_customers_on_table-d + (concentration-dk)P(base)).

Ed.2) I found the answer. Check my answer in below.


1 Answer 1


I found exactly what I was looking for: Yee Whye Teh explains the sampling process in his paper "A Hierarchical Bayesian Model based on Pitman-Yor Process", Section 2.


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.