Register to reply

Metropolis Algorithm - Why does it work?

by Pi-Bond
Tags: algorithm, metropolis, work
Share this thread:
Pi-Bond
#19
Dec6-12, 10:19 AM
P: 304
Ok, I understand this better now.

However I'm having trouble with the concept of "state" here. For a conventional (discrete time) Markov Chain, we have a row vector which tells us the probability of being in a state. So the size of the row vector is the number of states.

I'm wondering what is the analogous vector for the Markov Chain we construct in the Metropolis Algorithm. Is it an infinite row vector, as there are infinitely many points in the parameter space? Then the transition matrix would also be infinite then.
bpet
#20
Dec7-12, 10:41 PM
P: 523
Quote Quote by Pi-Bond View Post
Ok, I understand this better now.

However I'm having trouble with the concept of "state" here. For a conventional (discrete time) Markov Chain, we have a row vector which tells us the probability of being in a state. So the size of the row vector is the number of states.

I'm wondering what is the analogous vector for the Markov Chain we construct in the Metropolis Algorithm. Is it an infinite row vector, as there are infinitely many points in the parameter space? Then the transition matrix would also be infinite then.
In a way, yes, though infinite vectors can be tricky to compute with and there's a few technical considerations in going to infinite number of states (especially if there is a continuum of lilypads) :)

If you have the time, some courses on measure theory and stochastic processes would certainly be worth your while, but the gist of it is that the same arguments for Metropolis-Hastings apply as long as the jump "density" is defined and non-zero between every pair of distinct lilypads.

HTH
Pi-Bond
#21
Dec11-12, 03:36 PM
P: 304
Alright, thanks for the information.

Final query (I think)- should a Markov Chain be aperiodic if its stationary distribution is to be explored by the Metropolis Algorithm? (and why?)
chris ball
#22
Dec24-12, 02:27 PM
P: 3
I recommend the third edition of the book Understanding Probability (Cambridge University Press, 2012) by Henk Tijms. The author provides a very clear explanation of Markov chain Monte Carlo simulation. The explanation of the Metropolis algorithm allows one to understand how to construct an algorithm of their own. Nice illustrations are given.


Register to reply

Related Discussions
Metropolis Algorithm and integration volume General Physics 1
Metropolis-Hastings algorithm Set Theory, Logic, Probability, Statistics 1
Using metropolis algorithm for 2D ising model Set Theory, Logic, Probability, Statistics 0
Metropolis-Monte Carlo Algorithm Atomic, Solid State, Comp. Physics 2
Metropolis-Hastings algorithm Set Theory, Logic, Probability, Statistics 2