I have a set of independent random walkers on a graph that produce a distribution of the nodes in that graph. I would like to test whether the random walkers have produced a distribution which is consistent.

I thought about separating the results from the random walker into various sets, eg 4 separate groups, and using the distribution that these 4 iid groups make from the graph, use Gelman-Rubin diagnostics link to def.to test for convergence of those distributions.

I cannot use Gelman and Rubin's potential scale reduction factor directly on the independent walkers, because they are not let to run till they reach ergodic steady states. I suppose that the reasoning behind using Gelman and Rubin still holds when comparing within $W$, and between variance $R$ of distributions from groups of these independent walkers because even a markov chain run till it is supposed to reached ergodicity is a distribution.

Is my approach valid and why? are there any references for it?


When performing Markov chain Monte Carlo, you never know when and if the ergodic steady state is reached. Gelman-Rubin statistic is a test for reaching this state.

An important assumption for the Gelman-Rubin statistic is overdispersion of the initial states. This is needed to guarantee that the between-chain variance approaches the variance of the posterior from above. Make sure that this assumption is satisfied.

| cite | improve this answer | |
  • 1
    $\begingroup$ can the use of multiple iid walkers be used in aggregation to approximate that state. The search is for testing the locality of a discrete distribution's consistency for the iid walkers $\endgroup$ – Vass Sep 2 '13 at 10:47

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.