Dismiss Notice
Join Physics Forums Today!
The friendliest, high quality science and math community on the planet! Everyone who loves science is here!

Stationary distribution of Markov chain

  1. Apr 8, 2008 #1
    Hi all,

    I'm given a Markov chain [itex]Q_k[/itex], k>0 with stationary transition probabilities. The state is space uncountable.
    What I want to show is that the chain is asymptotically stationary, that is it converges in distribution to some random variable Q.

    All I have at hand is an k-independent upper bound for [itex]P(|Q_k|>x)[/itex] for all x in the state space (and some indecomposability assumptions on the state space).

    Is this enough to conclude convergence of the chain?

    Thanks for any help.

  2. jcsd
  3. Apr 8, 2008 #2
    I don't have much experience with uncountable state-space Markov chains, but the irreducibility conditions will typically imply that any stationary distribution must be unique. Also, you'll need some ergodicity condition to ensure that the chain actually converges to the stationary distribution, even if it isn't necessary to show that the stationary distribution exists.
  4. Apr 8, 2008 #3
    Thanks for your reply quadrophonics.

    Any more insights on the topic?
  5. Apr 8, 2008 #4
    Well, it seems that the basic result should be about checking the "return time" for a given state; in countable chains, there's a necessary and sufficient condition for the existance of a stationary distribution, where the expected return time for every state needs to be finite (and then the value of the stationary distribution for that state is the inverse of the return time). I'm not sure what the uncountable extension of that result would be: it seems that you can't ever expect to get back to *exactly* the same point in the state space (since it's a set of measure 0). Presumably some condition relating to the probability of returning within some [itex]\delta[/itex] of a given point in the state space?
  6. Apr 8, 2008 #5
    Thanks again.

    So irreducibility ensures existence of a stationary distribution and ergodicity convergence to this stationary distribution?

    Is ergodicity the same as non-periodicity....? I should read some more on the topic.

    The theory of Harris chains seems to be a good starting point.

    So i found two properties related to continous state space Markov chains

    The first [itex]\phi[/itex] -irreducibility:

    Here P_n are the n-step transition probabilities.

    The second one is a-periodicity, that is there do not exist subsets X_1 , ... , X_d of the state space such that the chain jumps with probability one from X_1 to X_2, from X-2 to X_3 .. and from X_d back to X_1.

    Under these two conditions, the Markov chain is (reportedly) asymptotically stationary.

    So I have to think about it ..:smile:
    Last edited: Apr 8, 2008
  7. Apr 8, 2008 #6
    Well, I think all that irreducibility buys you is the assurance that IF there is a stationary distribution, it's unique. If you have a reducible chain, there are non-communicating subsets of the state space, each of which can have their own stationary distribution. Irreducibility ensures that this can't happen, but doesn't guarantee that there is a stationary distribution. I'm trying to think of a counterexample, but having trouble... it may be that for finite state spaces you also get the existance of the stationary distribution, but I have some recollection it's not generally the case.

    But, yeah, ergodicity guarantees that you eventually visit every state, and so will eventually fall into the stationary distribution.

    I believe it implies both aperiodicity and "positive recurrence," which means that there a positive probability of returning to any state after you visit it once. Contrast this with a "transient" state, where you might start out but never return to.

    And, yeah, I believe that Harris Chains is the right term to search for. If you can find any good references, let me know.
Share this great discussion with others via Reddit, Google+, Twitter, or Facebook