Inferring potential landscapes: A Schr\"odinger bridge approach to Maximum Caliber

Abstract

Schr\"odinger bridges have emerged as an enabling framework for unveiling the stochastic dynamics of systems based on marginal observations at different points in time. The terminology "bridge'' refers to a probability law that suitably interpolates such marginals. The theory plays a pivotal role in a variety of contemporary developments in machine learning, stochastic control, thermodynamics, and biology, to name a few, impacting disciplines such as single-cell genomics, meteorology, and robotics. In this work, we extend Schr\"odinger's paradigm of bridges to account for integral constraints along paths, in a way akin to Maximum Caliber - a Maximum Entropy principle applied in a dynamic context. The Maximum Caliber principle has proven useful to infer the dynamics of complex systems e.g., that model gene circuits and protein folding. We unify these two problems via a maximum likelihood formulation to reconcile stochastic dynamics with ensemble-path data. A variety of data types can be encompassed, ranging from distribution moments to average currents along paths. The framework enables inference of time-varying potential landscapes that drive the process. The resulting forces can be interpreted as the optimal control that drives the system in a way that abides by specified integral constraints. Analogous results are presented in a discrete-time, discrete-space setting and specialized to steady-state dynamics. We finish by illustrating the practical applicability of the framework through paradigmatic examples, such as that of bit erasure or protein folding. In doing so, we highlight the strengths of the proposed framework, namely, the generality of the theory, the ease of computation, and the ability to interpret results in terms of system dynamics. This is in contrast to Maximum-Caliber problems where the focus is typically on updating a probability law on paths.