Parabolic optimal transport equations on manifolds

Young‐Heon Kim, Jeffrey Streets, Micah Warren

Abstract

We study a parabolic equation for finding solutions to the optimal transport problem on compact Riemannian manifolds with general cost functions. We show that if the cost satisfies the strong MTW condition and the stay-away singularity property, then the solution to the parabolic flow with any appropriate initial condition exists for all time and it converges exponentially to the solution to the optimal transportation problem. Such results hold in particular, on the sphere for the distance squared cost of the round metric and for the far-field reflector antenna cost, among others.

Introduction

Optimal Transport (OT) theory traditionally focuses on finding the most efficient map to redistribute mass between two distributions, $\rho$ and $\bar{\rho}$, given a cost function $c$. In the context of Riemannian manifolds, this problem typically culminates in solving a Monge-Ampère type equation—a fully nonlinear elliptic partial differential equation (PDE). While these equations are central to geometric analysis, they are notoriously difficult to solve directly.

This research, published in 2010, shifts the paradigm from static elliptic equations to dynamic parabolic flows. By introducing a time-dependent evolution for the potential function, the authors provide a robust framework for finding optimal maps through a process of continuous deformation, akin to the use of Ricci flow in other areas of differential geometry.

The Parabolic Flow Framework

The core contribution of the paper is the proposal of a fully nonlinear parabolic equation that evolves an initial potential function $u$ toward the unique solution of the OT problem. The flow is defined as:

$$\frac{\partial u}{\partial t} = \ln \det(u_{ij} + c_{ij}) - \ln \rho + \ln \bar{\rho} - \ln \det |c_{i\bar{j}}|$$

This equation acts as a relaxation algorithm. Starting from a suitable initial guess, the potential function "smooths out" over time, guided by the geometry of the cost function and the mass distributions.

Geometric Stability Conditions

To ensure that this flow converges to the desired optimal map without encountering singularities, the authors identify two critical structural requirements:

  1. The Strong MTW Condition: Named after Ma, Trudinger, and Wang, this condition involves the fourth derivatives of the cost function. It serves as a regularity guarantee, ensuring the geometry of the cost function prevents the transport map from developing "wrinkles" or discontinuities that would impede convergence.
  2. Stay-Away Singularity Property: This property ensures the transport map maintains a safe distance from the "cut-locus"—regions where the cost function loses smoothness (e.g., antipodal points on a sphere).

Analytical Methodology

The authors employ several sophisticated tools from geometric analysis to prove the stability and convergence of the flow:

  • Tensor Maximum Principle: This is used to maintain control over the second derivatives of the potential function throughout the evolution, preventing the equation from becoming degenerate.
  • Linearization and Laplace-Beltrami Operator: A significant insight of the paper is that the linearization of the flow operator corresponds to the Laplace-Beltrami operator of a specific, conformally deformed Riemannian metric. This allows the authors to leverage established results from linear heat equation theory.
  • Li-Yau Harnack Inequality: By deriving a version of this powerful inequality for their nonlinear setting, the authors prove that the solution converges to the stationary state at an exponential rate.
  • Dimension Augmentation: For 2D manifolds, the authors utilize a technique of taking a product with a circle ($S^1$) to "fake" a third dimension, solving the problem in the higher-dimensional space before projecting the results back down.

Key Results and Applications

The paper establishes three primary results:

  1. Global Existence: Under the MTW and stay-away conditions, the flow $u(x, t)$ exists for all $t \in [0, \infty)$ and remains smooth.
  2. Exponential Convergence: The potential function $u(x, t)$ converges to the optimal transport potential $u_\infty$ at a rate of $e^{-\beta t}$.
  3. Broad Applicability: The authors demonstrate that these conditions are met for several physically and geometrically significant problems, including the distance-squared cost on a round sphere, the far-field reflector antenna problem, and costs on complex projective spaces ($\mathbb{CP}^n$).

Conclusion and Limitations

While this work provides a powerful "algorithmic" foundation for Optimal Transport on manifolds, it is bounded by the stringency of its conditions. The reliance on the "strong" MTW condition limits its immediate application to manifolds where only the "weak" condition holds. Furthermore, the current framework is designed for closed manifolds; extending these results to manifolds with boundaries requires addressing complex oblique boundary conditions, a direction for future research.

Summary updated

Abstract

We study a parabolic equation for finding solutions to the optimal transport problem on compact Riemannian manifolds with general cost functions. We show that if the cost satisfies the strong MTW condition and the stay-away singularity property, then the solution to the parabolic flow with any appropriate initial condition exists for all time and it converges exponentially to the solution to the optimal transportation problem. Such results hold in particular, on the sphere for the distance squared cost of the round metric and for the far-field reflector antenna cost, among others.
BibTeX
@inproceedings{kim2011parabolic,
  author = {Young‐Heon Kim and Jeffrey Streets and Micah Warren},
  title = {{Parabolic optimal transport equations on manifolds}},
  booktitle = {International Mathematics Research Notices},
  year = {2011},
  doi = {10.1093/imrn/rnr188},
}
Figure