FLAIR · project
Stochastic Control Liaisons: Richard Sinkhorn Meets Gaspard Monge on a Schrödinger Bridge
This survey develops a common view of optimal transport, Schrödinger bridges, and stochastic control. Each asks how a distribution can evolve between prescribed endpoints, but the objectives differ in how they balance transport cost, uncertainty, and deviation from reference dynamics.
From transport to stochastic control
A Schrödinger bridge identifies a likely stochastic evolution consistent with endpoint distributions. Entropy regularization connects this problem to optimal transport; in an appropriate zero-noise limit, the transport problem is recovered. A control formulation turns the evolution of probability distributions into a dynamic optimization problem.
Why Sinkhorn iterations appear
The survey explains how Fortet’s iterative method and its discrete counterpart, Sinkhorn scaling, enforce the distributional constraints. It treats both continuous and discrete settings and traces the historical development of these connections. Applications ranging from guidance and ensemble control to sensing and network routing illustrate why this mathematical framework is useful beyond a single algorithm or application area.
Publication
Stochastic Control Liaisons: Richard Sinkhorn Meets Gaspard Monge on a Schrödinger Bridge
Yongxin Chen, Tryphon T. Georgiou, and Michele Pavon
SIAM Review, 63(2), 249–313, 2021.