1
$\begingroup$

Consider a Markov chain $X_n$ taking values in finite countable set $\mathcal{X}$ with transition matrix $P$. Consider a function $f:\mathcal{X}\to\mathcal{Y}$ inducing a partition $\mathcal{Y}=\{\mathcal{X}_{[i]}\}$. Then, the process $Y_n = f(X_n)$ is Markov if, given any two elements of the partition $\mathcal{X_{[k]}}$ and $\mathcal{X_{[\ell]}}$ it holds

$$\sum_{z\in\mathcal{X_{[\ell]}}}P(x\to z)=\sum_{z\in\mathcal{X}_{[\ell]}}P(x'\to z),\ \text{for all}\ x,x'\in \mathcal{X}_{[k]}.$$

I would like to know if there exists a comprehensive reference on the properties preserved under this type of reduction/coarsening of the state space, which is usually called lumping or projection.

I am particularly interested to this in relationship to particle systems with finite state space and coupling techniques for such systems, but if there's anything more general or on different type of spaces that would be welcome as well.

$\endgroup$
2
  • $\begingroup$ These are commonly called hidden Markov models. I'm not an expert on such models but there should be a vast literature. $\endgroup$ Feb 16 '18 at 22:16
  • $\begingroup$ I believe that the hidden Markov model terminology is typically used for applications (simulation, inference and parameter estimates), whereas within probability theory they are referred to as lumped or projected chains instead (and the questions are of course different: is the model still Markov would be the prototypical one). $\endgroup$
    – Three Diag
    Feb 19 '18 at 13:09
1
$\begingroup$

The question of what properties are preserved has attracted a lot of attention in dynamics. I think the state of the art is Mark Piraino, see 'projection of gibbs states for holder potentials.'

Of you care only about starting with a Markov system then probably work of Chazottes and Ugalde is more direct.

Sorry for answering your question two years late...

$\endgroup$
1
  • $\begingroup$ I will take it :) $\endgroup$
    – Three Diag
    Jan 17 '20 at 12:37

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.