To use Optimal Transport (OT) in statistics, we need to know more than just the distance between two samples; we need to know how that distance fluctuates due to randomness. If we calculate the Wasserstein distance ($W_2$) between a finite sample and a target distribution, will that value follow a Bell curve? Can we use standard tools like the bootstrap to calculate confidence intervals?
Work by Eustasio del Barrio, Jean-Michel Loubes, and their collaborators has established a rigorous distributional limit theory to answer these questions. They show that while the empirical transport cost often converges to a Gaussian distribution, it behaves fundamentally differently when comparing a distribution to itself or when the "optimal plan" for moving mass is not unique.
The Stability of Optimal Potentials
The core intuition behind the Central Limit Theorem (CLT) for optimal transport lies in the stability of "potentials." In the dual formulation of the $W_2$ problem, we look for a function (a potential) that maximizes the difference in expectations between two distributions.
As shown in Central limit theorems for empirical transportation cost in general dimension, if the underlying distributions are distinct and the optimal potential is unique and stable, the fluctuations of the empirical cost around its mean are driven by the fluctuations of the data itself. Because the data points are independent and identically distributed, their collective influence on the cost converges to a Gaussian distribution. This provides the foundation for statistical inference: if you are comparing two different populations, you can generally expect a standard normal limit for your test statistic.
The Degeneracy of the Null Hypothesis
A major complication arises in "goodness-of-fit" testing—the scenario where we compare a sample to its own generating distribution. In this case, the true distance is zero.
When the two distributions are identical, the first-order fluctuations that usually lead to a Gaussian limit vanish. As Eustasio del Barrio and Jean-Michel Loubes explain in their 2017 arXiv paper, this creates a mathematical singularity. The limiting variance drops to zero, and the standard CLT fails. Instead, the limit becomes non-Gaussian, requiring a more complex "second-order" analysis.
This distinction is critical for practitioners. You cannot use the same Gaussian approximation to test if a sample matches a model (the null case) that you would use to estimate the distance between two different models.
Semidiscrete Transport and the Curse of Dimensionality
In many modern applications, such as resource allocation or generative modeling, we compare a continuous distribution to a discrete one (a set of points). This is known as semidiscrete optimal transport.
In Central Limit Theorems for Semidiscrete Wasserstein Distances, Alberto González-Sanz and colleagues demonstrate a surprising advantage of this setting: it avoids the "curse of dimensionality." While general OT costs often converge very slowly in high dimensions, semidiscrete costs maintain a standard $1/\sqrt{n}$ convergence rate regardless of the dimension $d$. This makes semidiscrete OT a robust tool for high-dimensional data analysis.
When the Bootstrap Fails
Statistical software often relies on the "bootstrap"—resampling data to estimate uncertainty—when the underlying distribution is too complex to calculate analytically. However, the research by del Barrio and González-Sanz warns that the bootstrap can be inconsistent for optimal transport.
The problem occurs when the optimal transport potential is not unique. If there are multiple "best" ways to move mass from one distribution to another, the empirical cost doesn't converge to a single Gaussian. Instead, it converges to the supremum of a Gaussian process.
Think of this as a "competition" between different optimal paths. The randomness of the sample might make one path look slightly better than another, and the final cost is always the result of picking the maximum (the supremum) of these fluctuating possibilities. Because the "max" operation is non-linear, it warps the distribution away from a Bell curve. Standard bootstrap techniques cannot track this non-linear "switching" between optimal paths, leading to unreliable confidence intervals.
A Unified Framework for Inference
The recent review Distributional Limit Theory for Optimal Transport by Clément Lalanne and the team synthesizes these findings into a roadmap for practical use. By understanding whether a problem falls into the Gaussian regime (distinct distributions, unique potentials) or the non-Gaussian regime (identical distributions or non-unique potentials), researchers can design valid tests for AI fairness, model validation, and high-dimensional hypothesis testing.