# Reference on continuous-time finite state filtering

Problem:
I'm working in reliability field and have seen papers written on the topic like process of failures when systems are functioning under unobservable (or observable) Markov-like environment, i.e. probability to fail is dependent on the state of environment. This state is described as discrete-state discrete-time homogeneous Markov chain. In mathematical notation it would be like this:
$P\left [ X_{t}=k | Y_{t}=i \right ]=\pi _{k}\left ( i \right )$;
where $X _{t}$ is a binary random failure process with possible states 0 (failed) and 1 (working). And $Y _{t}$ is a Markov process at the moment $t$ being in the $i ^{th}$ state.

My question:

Is it possible (or even reasonable) to extend mentioned model? For example, from discrete-state to continuous-state Markov model? Is there any literature (I havent found yet) about continuous stochastic conditional processes. I suppose it is not so trivial, because for continuouse stochastic process statements like $Y_{t}=i$ are meaningless.

• It is easy to extend: for each $t$ there is a conditional law of $X_t$ given $Y_t$. Feb 28, 2012 at 15:36
• @ Stéphane Laurent Could recommend some literature (papers) on this the conditional stochastic processes topic. Because all I can find is just short notes in textbooks and no deeper analysis. I would be greatfull for it. Feb 29, 2012 at 9:44
• The formula $P(X_t=k\mid Y_t=i)=\pi_k(i)$ can be reformulated as $P(X_t=k\mid Y_t)=\pi_k(Y_t)$, and this formula makes perfectly good sense, no matter what the distribution of $Y_t$. See en.wikipedia.org/wiki/Conditional_expectation#Formal_definition and en.wikipedia.org/wiki/…. Mar 15, 2012 at 20:33

What you're looking for is the Wonham Filter. In this setting, $$(X_t)_{t\geq 0}$$ is a time homogeneous Markov chain defined on a probability space $$(\Omega,\mathcal{F},\mathbb{P})$$ taking values in a finite state space $$S\triangleq \left\{s_n\right\}_{n=1}^N$$ governed by the coupled system: \begin{aligned} p_t^n & \triangleq \mathbb{P}(X_t \in s_n)\\ \frac{dp_t}{dt} & = Q p_t,\\ Y_t & = \int_0^t X_s\beta^{\top}ds + W_t \end{aligned} for some Brownian motion defined on the same probability space as $$(X_t)_{t\geq 0}$$ and where $$\beta\in \mathbb{R}^N$$ and $$Q$$ is an $$N\times N$$ matrix (typically called the "Q-Matrix") and satisfies: $$\sum_{i=1}^N q_{i,j}=0, \mbox{ and }q_{i,j}\geq 0 \mbox{ for }i\neq j.$$