4
$\begingroup$

After struggling with auto-Poisson model (a.k.a. Random Markov Network with conditional Poisson distributions) trying to force Gibbs sampler to obtain discrete sample of the network (since I know conditional distributions) I realized that I have never seen discrete version of this sampler.

I wonder what is the state-of-the-art regarding sampling from discrete distribution (not confined to Bayesian posterior sampling)? Is there something like Gibbs sampler to sample from conditional discrete distributions?

$\endgroup$
  • 3
    $\begingroup$ Gibbs sampling can be used for discrete distributions as well, under appropriate considerations. This in mentioned, for instance, in the wikipedia entry, "Implementation" section. $\endgroup$ – user10525 Jul 24 '12 at 12:22
3
$\begingroup$

This is a very puzzling question: the Gibbs sampler draws its name from Gibbs random fields where it was originally used by the Geman brothers to generate realisations from such discrete models. For instance the Ising model is made of a vector of random variables each taking values in $\{-1,1\}$...

So indeed the Gibbs can be used for discrete models. And I do not understand why you have to force your "Gibbs sampler to obtain discrete samples". If you are using the proper full conditionals they should be on a discrete support.

$\endgroup$
  • 1
    $\begingroup$ Thank you Xi'an. I was just a bit confused and had some misunderstanding about Markov random fields. But know, thanks to your answer here and comment at my another question, all this seems to make perfect sense. $\endgroup$ – Tomas Jul 25 '12 at 8:46

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.