Entropic Selection Principle for Monge's Optimal Transport
Establishes the unique support structure of the entropic optimal transport plan in high dimensions with Euclidean cost, supported on transport rays, via relative entropy minimization.
Key Findings
Methodology
This work analyzes the small regularization limit of entropic optimal transport in Rd with Euclidean cost, leveraging the geometric structure of transport rays. By applying local coordinate transformations and multi-scale expansions, the authors derive the asymptotic form of the Radon-Nikodym derivatives of the coupling, showing they behave like multivariate Gaussian distributions supported on transport rays. The core variational principle minimizes a relative entropy functional within each ray, leading to a unique characterization of the limit coupling supported precisely on these rays. The approach combines geometric measure theory, convex analysis, and asymptotic expansions, establishing the support and uniqueness of the limiting plan.
Key Results
- In dimensions d>1 with absolutely continuous marginals supported on disjoint compact sets, the limit of entropic couplings concentrates on transport rays. The limiting plan ̃µ_{opt} supports on these rays and minimizes a relative entropy functional, with density proportional to \((2\pi\|x-y\|\)^{(d-1)/2} e^{c\|x-y\|}f(x)g(y)\).
- Numerical simulations confirm the support structure and the Gaussian-like behavior of the Radon-Nikodym derivatives near transport rays, across various geometric configurations, validating the theoretical predictions.
- The analysis reveals that as \(\varepsilon o 0\), the entropic optimal plan increasingly aligns with transport rays, with the local behavior governed by the asymptotic expansion of the potentials, leading to a stable and unique limit supported on these geometric structures.
Significance
This research advances the theoretical understanding of the zero-regularization limit in high-dimensional optimal transport, especially for non-smooth, non-strictly convex costs. It provides a rigorous geometric and variational foundation for the support and uniqueness of the limiting plan, bridging a gap in the literature. The results have profound implications for computational algorithms, geometric analysis, and statistical inference, enabling more precise modeling of resource allocation, distribution matching, and data transport in complex spaces. The methodology also opens pathways to analyze other non-smooth cost functions and dynamic Schrödinger bridges, broadening the scope of entropic regularization theory.
Technical Contribution
The paper introduces a novel combination of geometric measure theory, local coordinate analysis, and asymptotic expansions to characterize the zero-regularization limit of entropic optimal transport in high dimensions. It rigorously proves that the limit coupling is supported on transport rays and minimizes a relative entropy functional with a specific density form, establishing the support's support and uniqueness. The approach extends classical localization techniques to the entropic setting, providing a detailed variational principle that uniquely determines the limit. This framework can be adapted to other non-smooth costs and geometric settings, representing a significant step forward in the theoretical understanding of entropic regularization.
Novelty
This work is the first to rigorously characterize the support and uniqueness of the zero-regularization limit of entropic optimal transport in high dimensions with Euclidean cost. Unlike prior results limited to one dimension or discrete measures, it leverages the geometric structure of transport rays and local expansions to establish a complete variational principle. The innovative combination of asymptotic analysis, geometric localization, and relative entropy minimization provides a new paradigm for understanding the support and structure of optimal couplings as regularization vanishes.
Limitations
- The analysis assumes absolutely continuous marginals with smooth densities supported on disjoint compact sets, limiting applicability to more general measures or singular supports.
- The derivation relies on the geometric structure of transport rays, which may be difficult to explicitly compute or approximate in complex or high-dimensional spaces.
- The results are primarily established for Euclidean distance costs; extending to other non-smooth or non-distance costs remains an open challenge.
Future Work
Future research will focus on relaxing regularity assumptions, extending the framework to measures with singularities, and exploring dynamic Schrödinger bridges. Developing efficient numerical schemes to approximate the support structure in complex geometries and applying these insights to machine learning tasks such as distribution alignment and generative modeling are promising directions. Additionally, investigating other non-smooth cost functions and their zero-regularization limits will broaden the theoretical landscape.
AI Executive Summary
Optimal transport theory has profoundly influenced mathematics and its applications, providing a framework for transforming one probability distribution into another with minimal cost. Classical formulations, such as Monge and Kantorovich problems, have been extensively studied, with solutions well-understood under smooth and regular conditions. However, real-world problems often involve high-dimensional, non-smooth costs like Euclidean distances, where the structure of optimal plans becomes complex and less understood.
To address computational challenges, entropic regularization was introduced, smoothing the problem and enabling scalable algorithms like Sinkhorn iterations. Yet, as the regularization parameter tends to zero, the behavior of the resulting couplings—particularly their support and uniqueness—remains elusive in high dimensions. This paper makes a significant breakthrough by rigorously characterizing the zero-regularization limit in the Euclidean setting, revealing that the limiting optimal plan concentrates on transport rays—geometric structures along which mass is transported.
The authors employ a combination of geometric measure theory, local coordinate transformations, and asymptotic expansions of the entropic potentials. They demonstrate that, supported on these rays, the limit coupling uniquely minimizes a relative entropy functional, with densities approximating multivariate Gaussian distributions. This result not only confirms the support structure but also guarantees the plan’s uniqueness, resolving a longstanding open problem.
Numerical simulations validate the theoretical predictions across various geometric configurations, showing the convergence of the entropic plan to a structure supported on transport rays. The findings have broad implications for high-dimensional data analysis, resource allocation, and geometric inference, providing a rigorous foundation for future algorithmic and theoretical developments. Moving forward, extending these results to measures with less regularity and other cost functions will further deepen our understanding of entropic optimal transport in complex settings.
Deep Dive
Abstract
We investigate the small regularization limit of entropic optimal transport when the cost function is the Euclidean distance in dimensions $d > 1$, and the marginal measures are absolutely continuous with respect to the Lebesgue measure. Our results establish that the limiting optimal transport plan is supported on transport rays. Furthermore, within each transport ray, the limiting transport plan uniquely minimizes a relative entropy functional with respect to specific reference measures supported on the rays. This provides a complete and unique characterization of the limiting transport plan. While similar results have been obtained for $d = 1$ in \cite{Marino} and for discrete measures in \cite{peyré2020computationaloptimaltransport}, this work resolves the previously open case in higher dimensions $d>1.$