Abstract
This paper establishes a Large Deviation Principle (LDP) for sequences of static Schrödinger bridges, specifically focusing on the small-noise limit where the scale parameter decreases toward zero. By extending existing LDP results to cover a broader range of reference processes, the authors provide a theoretical foundation for the convergence of Schrödinger bridges to limiting optimal transport plans. The analysis particularly addresses Schrödinger bridges associated with reflected diffusions on bounded convex domains, solving an open problem regarding entropic optimal transport with non-fixed cost functions in constrained settings.
Introduction
In generative modeling, Schrödinger bridges (SB) serve as a robust framework for transforming one probability distribution into another. Mathematically, this framework is deeply connected to Optimal Transport (OT). When entropy (noise) is added to an OT problem to improve computational feasibility, it becomes Entropic Optimal Transport (EOT). A critical question in this field is how these EOT solutions behave as the noise parameter ($\epsilon$) vanishes: do they converge to the "perfect" OT solution, and at what rate?
While previous research established this convergence for unconstrained Brownian motion, many real-world generative tasks involve constrained data (e.g., pixel values restricted to a [0, 1] range). Such constraints require "reflected" diffusions. This paper provides the first rigorous proof of the Large Deviation Principle for these constrained cases, quantifying the "speed limit" of convergence as noise approaches zero.
Theoretical Framework and Methodology
The authors bridge the gap between stochastic analysis and geometric measure theory through several key methodological steps:
Generalizing the Large Deviation Principle
The study extends the LDP for fixed-cost EOT to a more flexible setting. In this generalized framework, the "cost" function itself is allowed to change as the noise scales, provided the sequence of cost functions converges uniformly to a limit. This allows for a much wider variety of reference processes beyond standard Brownian motion.
Skorokhod Maps and Neumann Heat Kernels
To account for boundary constraints, the authors employ Skorokhod maps—mathematical tools that describe paths "pushed back" into a domain upon contact with its boundary. They perform a detailed analysis of the transition densities of reflected Brownian motion, known as Neumann heat kernels. A central challenge was proving that the "cost" of moving between two points in a reflected system converges uniformly to the standard squared-distance cost ($|x-y|^2/2$) used in classical OT as noise decreases.
Cyclical Monotonicity
The authors utilize the geometric property of cyclical monotonicity to characterize the support of the limiting measures. This ensures that the mathematical results remain consistent with the physical intuition of the "most efficient" or least-cost movement between distributions.
Key Results
- Theorem 4.2 (Weak-type LDP): The core theoretical contribution, establishing an LDP for sequences of EOT plans where the cost functions exhibit uniform convergence. This provides a way to quantify the exponential decay of the probability that a noisy Schrödinger bridge deviates from the optimal transport plan.
- Theorem 5.2 (Uniform Convergence in Convex Domains): The authors prove that for any open, bounded, convex domain, the cost function associated with reflected Brownian motion converges uniformly to the standard quadratic cost. This demonstrates that the influence of the boundary is effectively "tamed" in the small-noise limit.
- Validation of Reflected Models: The paper provides the first rigorous proof that "Reflected Diffusion Models," which have gained popularity for generating data on manifolds or constrained domains, are indeed approximating optimal transport plans as their noise is reduced.
Discussion and Future Work
While this work provides a significant theoretical advancement for the static case (the coupling of start and end points), the authors note that the dynamic case—describing the full path of the particles—remains a challenge. Proving a dynamic LDP for reflected processes is mathematically more complex because the reflection prevents the use of standard Gaussian properties.
Additionally, the current proofs require the domain to be convex. Future research is needed to extend these results to non-convex domains (such as manifolds with holes) and to explore more complex dynamics, including SDEs with jumps or oblique reflections, which are relevant to advanced physics and engineering simulations.
Summary updated