Let there be two Markov processes on the same state space (which is countably infinite), but different transition matrices, denoted by $P_{1}$ and $P_{2}$. Assume positive recurrence, irreducibility and aperiodicity for both of them, so that a stationary distribution exists for both of them, denote them by $\pi_{1},\pi_{2}$ resp. Under what criteria do we have $\pi_{1} = \pi_{2}$. Obviously we should have $\pi_{1}P_{1} = \pi_{2}P_{2}$. What other stronger conditions are implied by $\pi_{1} = \pi_{2}$.
$\begingroup$
$\endgroup$
3
-
1$\begingroup$ Not sure if much is implied. Consider a finite state space (of size n), and consider the Markov Chain given by random walks on a (non-bipartite) d-regular graph on n vertices. Regardless of the structure of the graph (and the value d), the stationary distribution is uniform. This is a very large set of graphs, and there isn't much structural similarity between these graphs. $\endgroup$– SeshCommented Apr 14, 2013 at 4:18
-
1$\begingroup$ Under the conditions you list, and with $\pi$ given, I think that $P$ can be described by specifying for each unordered pair $\lbrace i,j\rbrace$ a transition rate $P_{ij}$. The reverse rate $P_{ji}$ is determined by detailed balanced. But for each pair one rate is completely free and independent of all other rates so long as you satisfy ergodicity ($(P^n)_{ij}>0$ for some $n$). $\endgroup$– Yoav KallusCommented Apr 14, 2013 at 4:51
-
1$\begingroup$ mathoverflow.net/questions/14729 $\endgroup$– Steve HuntsmanCommented Apr 14, 2013 at 13:36
Add a comment
|