I've been reading Norris's book on Markov Chains, and there's one part of his proof of convergence of a Markov chains to equilibrium that leaves me a bit confused. He starts with $X_n$ and $Y_n$, with $X_0 \sim \lambda$ and $Y_0\sim \pi$, the invariant measure. He then introduces the coupling time $T$, at which both chains first reach a given state $b$, and $W_n = (X_n, Y_n)$, the joint process. Next, he introduces a piecewise defined $Z_n$ process, and proceeds to show that it is Markov:enter image description here

It's his proof that $Z_n$ is Markov I'm stuck on. It seems like there are some details left out here.

  • $\begingroup$ What part did you understand up until? $\endgroup$ – spaceisdarkgreen Mar 22 '17 at 1:22
  • $\begingroup$ I found the "By symmetry, we can replace..." line a bit confusing, but I kind of buy it. I was mostly troubled by the last sentence, that $Z_n$ is Markov, but now see that, if $W_n'$ really is Markov on $(\mu, \tilde P)$, then its components must also be Markov because this is just a decoupled Markov Chain on the product space. $\endgroup$ – user2379888 Mar 22 '17 at 12:40
  • $\begingroup$ Yeah, that's right on the last point. $\endgroup$ – spaceisdarkgreen Mar 22 '17 at 23:43

You have the Markov process $(X_n,Y_n)\sim \mathrm{markov}(\mu,\tilde P).$ $(Z_n,Z'_n)$ must be statistically identical to it by the symmetry argument given. Basically, you wait till the time when $(X_n,Y_n)=(b,b)$ and then pull a switcheroo. Since the transition matrix is symmetrical and the state is symmetrical, you've done nothing. Thus $(Z_n,Z'_n)\sim \mathrm{markov}(\mu,\tilde P)$ which means $Z_n\sim\mathrm{markov}(\lambda, P)$ and $Z'_n\sim\mathrm{markov}(\pi, P)$

  • $\begingroup$ Just to be clear, you can make the switch at the stopping time because, at $T$, you know $X_{T} = Y_{T} = b$, and they then propagate, independently, with the same transition matrix, $P$, is that right? $\endgroup$ – user2379888 Mar 22 '17 at 12:41
  • $\begingroup$ @user2379888 intuitively, yes. The reason for constructing the pair is that, from the perspective of the markov chain on $I\times I$ with initial distribution $(\lambda,\pi)$ and transition matrix $\tilde P = P\otimes P$, you are waiting till it's in state $(b,b),$ doing nothing, and then carrying on. Thus $(Z_n,Z_n')$ is the same markov chain as $(X_n,Y_n).$ $\endgroup$ – spaceisdarkgreen Mar 22 '17 at 23:34

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.