1
$\begingroup$

For the Chinese restaurant process, as used in Dirichlet Process mixture models, we have a prior that data point i belongs to cluster j, where c is an indicator. n represents the total number of data points that are assigned to cluster j, where subscript -i refers to all observations excpet i.

enter image description here

As we take k to infinity, (which represents taking the number of components in the mixture model to infinity), the expression transforms to

enter image description here

when that cluster already exists (which I understand).

For the case that the cluster doesn't already exist, the prior transforms to

enter image description here

This is where my problem lies - how does alpha feature in this expression if (alpha/k) tends to 0 as k tends to infinity?

Thanks in advance

edit: link to paper (pages 3-4) https://www.seas.harvard.edu/courses/cs281/papers/rasmussen-1999a.pdf

$\endgroup$
1
$\begingroup$

This wasn't obvious to me either at first glance. Looking at the first expression, the prior probability that $c_i=c$ where $c$ has no other occupants is $\frac{\alpha/K}{n-1+\alpha}$. Let $J$ be the number of occupied clusters. Then the total prior probability of being in any unoccupied cluster is just $$\frac{\alpha/K}{n-1+\alpha}(K-J)$$ Holding J constant and taking the limit, $$\lim_{K \rightarrow \infty} \frac{\alpha/K}{n-1+\alpha}(K-J)=\frac{\alpha}{n-1+\alpha}\lim_{K\rightarrow \infty}\frac{K-J}{K}=\frac{\alpha}{n-1+\alpha}$$

$\endgroup$

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.