Approximate solutions of a general stochastic velocity-jump process subject to discrete-time noisy observations

A Ceccarelli, AP Browning, RE Baker


Advances in experimental techniques allow the collection of high-space-and-time resolution data that track individual motile entities over time. This poses the question of how to use these data to efficiently and effectively calibrate motion models. However, typical mathematical models often overlook the inherent aspects of data collection, such as the discreteness and the experimental noise of the measured locations. In this paper, we focus on velocity-jump models suitable to describe single-agent motion in one spatial dimension, characterised by successive Markovian transitions between a finite network of n states, each with a specified velocity and a fixed rate of switching to every other state. Since the problem of finding the exact distributions of discrete-time noisy data is generally intractable, we derive a series of approximations for the data distributions and compare them to in-silico data generated by the models using four example network structures. These comparisons suggest that the approximations are accurate given sufficiently infrequent state switching, or equivalently, a sufficiently high data collection frequency. Moreover, for infrequent switching, the PDFs comparisons highlight the importance of accounting for the correlation between subsequent measured locations, due to the likely permanence in the state visited in the previous measurement. The approximate distributions computed can be used for fast parameter inference and model selection between a range of velocity-jump models using single-agent tracking data.