- #1
mxbob468
- 49
- 0
I'm taking my first Markov processes class and I'm a little confused. This is in relation to a HW problem that I've been wrestling with for the past couple of hours. The task is to find all of the stationary distributions of a Markov process (finite state space).
I know that if my process were irreducible then it would have a unique distribution but it's not; there are 3 communication classes. First question: am I correct in guessing that there should be 3 distinct stationary distributions? One for each class?
I know that I can blunt-force-trauma compute the stationary distributions by raising the transition matrix to some large power and inspecting. I know this is a bad idea in general and this only works because it's a 5x5 matrix. I know in the case wherein the stationary distribution would be unique the alternative is to find all the eigenvectors corresponding to eigenvalue 1. Second question: is this also true for my case, wherein the distributions aren't unique?
I went ahead and computed these eigenvectors and for the longest time was pulling my hair out because they didn't match the rows in the asymptotic transition matrix. Then I remembered degenerate eigenvalues => degenerate eigenspace => linear combinations of eigenvectors (long time since I've had a linear algebra class). So now I can match the eigenvectors computed by MATLAB to the rows in the asymptotic transition matrix. Third question (maybe related to first): if I could not compute the asymptotic transition matrix how would I know whether I've found all the stationary distributions of the process? How do I even know I've found the right ones? Ie those that match the rows in the asymp transition matrix.
Thanks for any help anyone can give me
I know that if my process were irreducible then it would have a unique distribution but it's not; there are 3 communication classes. First question: am I correct in guessing that there should be 3 distinct stationary distributions? One for each class?
I know that I can blunt-force-trauma compute the stationary distributions by raising the transition matrix to some large power and inspecting. I know this is a bad idea in general and this only works because it's a 5x5 matrix. I know in the case wherein the stationary distribution would be unique the alternative is to find all the eigenvectors corresponding to eigenvalue 1. Second question: is this also true for my case, wherein the distributions aren't unique?
I went ahead and computed these eigenvectors and for the longest time was pulling my hair out because they didn't match the rows in the asymptotic transition matrix. Then I remembered degenerate eigenvalues => degenerate eigenspace => linear combinations of eigenvectors (long time since I've had a linear algebra class). So now I can match the eigenvectors computed by MATLAB to the rows in the asymptotic transition matrix. Third question (maybe related to first): if I could not compute the asymptotic transition matrix how would I know whether I've found all the stationary distributions of the process? How do I even know I've found the right ones? Ie those that match the rows in the asymp transition matrix.
Thanks for any help anyone can give me