Central Limit Theorems for Empirical Optimal Transport

This direction is split into other directions: statistical inference for optimal transport. Its last version stays readable and citable here.

Version 1 of 128 papers46 conceptsfrozen 2026-10-04

Optimal Transport (OT) has become a cornerstone of modern machine learning, providing a geometrically meaningful way to compare probability distributions. However, for years, its practical use in statistics was hampered by a "curse of dimensionality": the number of samples required to estimate the true transport cost grows exponentially with the dimension of the data.

Recent breakthroughs have shifted the focus from the slow convergence of the expected cost to the surprisingly fast convergence of its fluctuations. By establishing Central Limit Theorems (CLTs) for empirical transport costs, researchers have unlocked the ability to perform rigorous statistical inference—such as hypothesis testing and confidence interval construction—even in high-dimensional spaces.

The Statistical Barrier: Mean vs. Fluctuation

The traditional difficulty with Optimal Transport is that the empirical Wasserstein distance $W_p(\hat{\mu}_n, \nu)$ converges to the true distance $W_p(\mu, \nu)$ at a rate of $O(n^{-1/d})$. In 10 dimensions, this means you might need millions of samples just to get a decent estimate of the mean cost.

However, as Eustasio del Barrio and Jean-Michel Loubes demonstrated in Central Limit Theorem for empirical transportation cost in general dimension, there is a silver lining. While the bias (the difference between the expected empirical cost and the true cost) is subject to the curse of dimensionality, the fluctuations of the empirical cost around its own mean follow a standard Gaussian distribution at the parametric rate of $O(n^{-1/2})$, regardless of the dimension.

This distinction is crucial. It means that while we might not know the exact "true" distance easily, we can precisely quantify the uncertainty of our estimate. This allows for the construction of p-values and confidence intervals, enabling tasks like the fairness assessment proposed in A Central Limit Theorem for $L_p$ transportation cost with applications to Fairness Assessment in Machine Learning.

The Mechanism: Hadamard Differentiability

Why do these complex, non-linear transport costs behave like simple Gaussian averages? The answer lies in the Functional Delta Method.

In classical statistics, if a sequence of random variables converges to a Gaussian, any smooth function of those variables also converges to a (different) Gaussian. Researchers like Alberto González-Sanz and Jean-Michel Loubes have extended this to the space of probability measures. In A Central Limit Theorem for Semidiscrete Wasserstein Distances, they show that the optimal transport cost is Hadamard differentiable—a specific form of functional smoothness.

By treating the transport cost as a differentiable functional of the underlying measures, the fluctuations of the empirical measure (which are Gaussian by the standard CLT) propagate through the "derivative" of the transport map to produce Gaussian fluctuations in the cost. This linearization technique allows researchers to bypass the geometric complexity of the transport plan and treat the cost as an M-estimator, a generalization of maximum likelihood estimation.

Regularization as a Statistical Smoother

To make these CLTs more robust and computationally efficient, researchers often turn to regularized versions of OT.

  1. Entropic Regularization: By adding an entropy penalty, the transport problem becomes "smoother." Gonzalo Mena and Jonathan Weed proved in Statistical bounds for entropic optimal transport: sample complexity and the central limit theorem that entropic OT completely avoids the curse of dimensionality, converging at the $n^{-1/2}$ rate even in the mean.
  2. Gaussian Smoothing: In Smooth p-Wasserstein Distance: Structure, Empirical Approximation, and Statistical Applications, Sloan Nietert and colleagues show that convolving distributions with a small amount of Gaussian noise provides similar parametric convergence rates, making it a powerful tool for generative model evaluation.

These regularized distances are not just approximations; they are statistical objects in their own right with predictable Gaussian limits, as detailed in Weak limits of entropy regularized Optimal Transport; potentials, plans and divergences.

From Costs to Potentials and Plans

The distributional limit theory extends beyond the scalar cost to the actual "machinery" of the transport: the dual potentials (the functions that define the price of moving mass) and the transport plans (the map of where mass goes).

In Central Limit Theorems for Semidiscrete Wasserstein Distances, the authors establish that in semidiscrete settings—where one distribution is continuous and the other is a finite set of points—the optimal potentials themselves converge to a multivariate Gaussian vector. This is a profound result for practitioners: it means we can perform hypothesis testing not just on how much two distributions differ, but on where and how they differ.

Applications in Generative Modeling and Fairness

This theoretical framework has immediate practical consequences. In Reforming Generative Autoencoders via Goodness-of-Fit Hypothesis Testing, Aaron Palmer and Jinbo Bi replace the unstable adversarial training of GANs with a classical Shapiro-Wilk normality test. They justify this by showing that such tests are asymptotically equivalent to Wasserstein-based distances, providing a theoretically grounded objective for latent space regularization.

Similarly, the ability to calculate asymptotic variance for transport costs allows for "Fairness Assessment." By testing if the Wasserstein distance between different demographic groups exceeds a statistically significant threshold, researchers can provide a mathematically rigorous audit of machine learning models, as explored by Paula Gordaliza and her collaborators.

By bridging the gap between the geometry of Monge-Kantorovich theory and the rigor of frequentist statistics, this line of research transforms Optimal Transport from a descriptive tool into a predictive, inferential framework.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_central_limit_theorems_for_empirical_opt_3b58380a,
  title = {Central Limit Theorems for Empirical Optimal Transport},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/central-limit-theorems-for-empirical-optimal-transport/rdr_3b58380aaf374bd9bf92edcbe374fb51},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_central_limit_theorems_for_empirical_opt_3b58380a_v1,
  title = {Central Limit Theorems for Empirical Optimal Transport},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/central-limit-theorems-for-empirical-optimal-transport/rdr_3b58380aaf374bd9bf92edcbe374fb51/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}