Welcome to MathOverflow

Question

4

1

Target distribution is multimodal, 24 dimensions, continuous state space. For MCMC integration (MH sampler) I use a manually tuned proposal distribution.

When I measure the convergence rate numerically I observe only a polynomial convergence rate.

I have a complete freedom of choice for the proposal distribution. Are there practically useful conditions to quickly check if a particular proposal distribution leads to the geometric convergence rate?

edited Dec 2 at 23:56

asked Nov 30 at 17:03

Anton
81●5

an12 · Answer 1 · 2012-12-03 06:35:30Z UTC

1

I guess this is a difficult topic. Maybe this recent manuscript could be helpful

Johnson and Geyer (2012). Variable Transformation to Obtain Geometric Ergodicity in the Random-walk Metropolis Algorithm.

answered Dec 3 at 6:35

an12
735●2●11

	Thank you for the suggestion. An interesting reading indeed. The only question I have is that it seems like the transformation can help with heavy tails. However it is not clear for me how it helps with reachability of distinct modes. Thanks, Anton – Anton Dec 7 at 11:21
	@Anton. That is definitely an interesting (and not easy) question. It is not clear to me either. – an12 Dec 8 at 5:05

Practical way to check for geometric convergence

Remember to vote up questions/answers you find interesting or helpful (requires 15 reputation points)

1 Answer

You can accept an answer to one of your own questions by clicking the check mark next to it. This awards 15 reputation points to the person who answered and 2 reputation points to you.