Conditional splitting probabilities for hidden-state inference in drift-diffusive processes

Abstract: Splitting probabilities quantify the likelihood of particular outcomes out of a set of mutually-exclusive possibilities for stochastic processes and play a central role in first-passage problems. For two-dimensional Markov processes ${X(t),Y(t)}{t\in T}$, a joint analogue of the splitting probabilities can be defined, which captures the likelihood that the variable $X(t)$, having been initialised at $x_0 \in \mathbb{L}$, exits $\mathbb{L}$ for the first time via either of the interval boundaries \emph{and} that the variable $Y(t)$, initialised at $y_0$, is given by $y{\rm exit}$ at the time of exit. We compute such joint splitting probabilities for two classes of processes: processes where $X(t)$ is Brownian motion and $Y(t)$ is a decoupled internal state, and unidirectionally coupled processes where $X(t)$ is drift-diffusive and depends on $Y(t)$, while $Y(t)$ evolves independently. For the first class we obtain generic expressions in terms of the eigensystem of the Fokker-Planck operator for the $Y$ dynamics, while for the second we carry out explicit derivations for three paradigmatic cases (run-and-tumble motion, diffusion in an intermittent piecewise-linear potential and diffusion with stochastic resetting). Drawing on Bayes' theorem, we subsequently introduce the related notion of conditional splitting probabilities, defined as the posterior likelihoods of the internal state $Y$ \emph{given} that the observable degree of freedom $X$ has undergone a specific exit event. After computing these conditional splitting probabilities, we propose a simple scheme that leverages them to partially infer the assumedly hidden state $Y(t)$ from point-wise detection events.