Stability and statistical inference for semidiscrete optimal transport maps

Ritwik Sadhu, Ziv Goldfeld, Kengo Kato

Abstract

Optimal Transport (OT) maps, specifically the Brenier map, are fundamental to modern statistical machine learning, yet a rigorous framework for their statistical inference has remained elusive. This paper addresses this gap in the semidiscrete setting—where a known continuous reference distribution is mapped to an unknown discrete target distribution. We derive limit distributions for $L^p$-estimation errors and linear functionals, establish the consistency of the nonparametric bootstrap, and prove that the empirical OT map is asymptotically efficient within a dual Hölder space. Our results provide the theoretical foundation necessary for constructing confidence sets and performing hypothesis testing on OT maps.

Introduction

The Brenier map represents the most efficient transformation between two probability measures. In the semidiscrete case, a continuous source (such as a Gaussian or uniform distribution) is mapped to a finite set of data points, partitioning the source domain into regions known as Laguerre cells. While researchers have established rates of convergence for estimating these maps, the field has lacked a theory for statistical inference. Without such a theory, practitioners cannot quantify the uncertainty of an estimated map or construct reliable confidence intervals.

This research provides a comprehensive mathematical framework to solve this problem. The challenge is significant because the OT map is defined implicitly through a complex optimization problem. Small perturbations in the target data can cause non-linear shifts in the boundaries of the Laguerre cells, making the map's statistical properties difficult to track using classical methods.

Methodological Framework: Stability and Differentiability

The authors analyze the stability of the OT map by examining its relationship with the "dual potential vector." A key innovation is the application of Hadamard directional differentiability. Because the $L^p$-error of the map is not a linear function of the data, standard calculus is insufficient. By employing a generalized version of the derivative, the authors describe how the map evolves as the empirical target distribution converges toward the truth.

The geometry of the problem is centered on the behavior of Laguerre cells. In a semidiscrete map, every point within a specific cell of the continuous source is mapped to a single point in the discrete target. The authors identify a "super-consistency" paradox: at any fixed point strictly inside a cell, the empirical map will eventually match the true map exactly as the sample size grows. Consequently, "pointwise" inference is trivial and provides little information. The meaningful statistical noise is concentrated at the boundaries (faces) of these cells. The paper provides detailed stability estimates for how these boundaries shift, which is essential for deriving the limit theorems.

Furthermore, the authors address the "weak limit" problem. In standard $L^2$ spaces, the empirical OT map does not possess a non-trivial weak limit due to its discontinuous nature. To resolve this, the authors prove a Central Limit Theorem (CLT) in the dual of an α-Hölder space, a functional space capable of accommodating the map's inherent discontinuities.

Statistical Inference and Limit Laws

The study reveals a critical distinction between the types of errors being measured:

  1. $L^p$-Estimation Error: The authors derive a non-Gaussian limit distribution for the $L^p$-error (for $p \in [1, \infty)$). This is a significant departure from the typical Gaussian (bell curve) limits found in classical statistics, and the authors provide an explicit density for this limit.
  2. Linear Functionals: For linear functionals of the map, the authors establish asymptotic normality, meaning these measures do follow a Gaussian distribution in the limit.

For practitioners, the most vital contribution is the proof of bootstrap consistency. The nonparametric bootstrap—a technique of resampling from observed data to estimate a sampling distribution—is shown to be mathematically valid for these maps. This allows for the practical construction of confidence intervals even when the underlying derivatives are non-linear.

Empirical Validation and Applications

The theoretical results were validated through 10,000 Monte Carlo repetitions using synthetic environments, such as uniform distributions on unit squares or balls mapped to random discrete points. The simulations confirmed that bootstrap-based confidence intervals for $L^1$-error and tail correlations achieve the correct coverage probabilities, following the theoretical 45-degree line in probability plots.

The paper also demonstrates the utility of these limit theorems in finance. They provide a framework for inference on the maximum tail correlation, a risk measure for multivariate assets that relies on the geometry of the OT map to identify extreme co-movements in risky assets.

Conclusion and Future Work

This work bridges the gap between optimal transport theory and statistical practice. By proving that the empirical OT map is asymptotically efficient in dual Hölder spaces, the authors show that this estimator makes the best possible use of available data.

While this framework is a major step forward, it is currently limited to the semidiscrete case with a known reference distribution. Future research is required to extend these results to the "continuous-to-continuous" case, where the regularity of the map is much harder to control, and to scenarios where both the source and target distributions must be estimated from data.

Summary updated

Abstract

We study statistical inference for the optimal transport (OT) map (also known as the Brenier map) from a known absolutely continuous reference distribution onto an unknown finitely discrete target distribution. We derive limit distributions for the $L^p$-error with arbitrary $p \in [1,\infty)$ and for linear functionals of the empirical OT map, together with their moment convergence. The former has a non-Gaussian limit, whose explicit density is derived, while the latter attains asymptotic normality. For both cases, we also establish consistency of the nonparametric bootstrap. The derivation of our limit theorems relies on new stability estimates of functionals of the OT map with respect to the dual potential vector, which may be of independent interest. We also discuss applications of our limit theorems to the construction of confidence sets for the OT map and inference for a maximum tail correlation. Finally, we show that, while the empirical OT map does not possess nontrivial weak limits in the $L^2$ space, it satisfies a central limit theorem in a dual Hölder space, and the Gaussian limit law attains the asymptotic efficiency bound.
BibTeX
@article{sadhu2024stability,
  author = {Ritwik Sadhu and Ziv Goldfeld and Kengo Kato},
  title = {{Stability and statistical inference for semidiscrete optimal transport maps}},
  journal = {The Annals of Applied Probability},
  year = {2024},
  doi = {10.1214/24-aap2104},
}
Figure