The paper develops a mathematical framework for updating beliefs about hidden system states as new observations arrive over time, a process known as Bayesian filtering. The classical examples include tracking a moving object using noisy sensor readings, or estimating economic conditions from imperfect measurements. What makes this work distinctive is that it handles "feedback" models, where the system's behavior and the observations themselves can depend on the history of everything seen so far, not just the current moment. This is more realistic than the standard assumption that the future depends only on the present, but it requires more careful mathematical machinery to handle correctly.
The authors build their framework using a branch of mathematics called category theory, specifically working with objects called Markov kernels, which are mathematical objects that describe probabilistic relationships between spaces. Rather than relying on visual tools like string diagrams or requiring probability densities to exist, they develop a clean algebraic notation that feels natural to anyone familiar with standard probability. The central idea is to systematically track how probability laws depend on parameters, and to show that conditioning, which is the act of updating beliefs given new information, behaves consistently within this framework. They prove a property called Bayesian coherence, which guarantees that the order and manner in which you apply conditioning does not lead to contradictions.
As a payoff, the paper derives the Bayesian filter, prediction equations, and smoothing algorithms for these general feedback models under mild technical assumptions. Smoothing refers to going back and refining earlier estimates once more data has been collected. When the framework is specialized to linear Gaussian systems, familiar and practically important algorithms fall out naturally: the Kalman filter for real-time state estimation, and the Rauch-Tung-Striebel smoother for retrospective refinement. Crucially, these results are obtained in a strictly broader setting than previously covered by categorical approaches, accommodating the feedback structure that arises in many real applications.