Optimal transport (OT) and Wasserstein distances have become foundational tools in machine learning for comparing probability distributions. By incorporating the underlying geometry of the data space, OT captures structural differences that traditional metrics like Kullback-Leibler divergence completely miss.
However, applying OT to real-world data introduces a fundamental statistical challenge. In practice, we rarely have access to the true underlying distributions; instead, we must estimate them using finite samples (empirical distributions). This transition from continuous distributions to empirical samples exposes two major hurdles: the curse of dimensionality (where sample complexity scales exponentially with the dimension of the space) and the lack of statistical inference tools (such as confidence intervals and hypothesis tests) for transport-based metrics.
To resolve these challenges, a major research direction has emerged focused on establishing Central Limit Theorems (CLTs) and limit distributions for empirical optimal transport. By proving that empirical transport costs and maps converge asymptotically to Gaussian processes, researchers have unlocked the mathematical machinery required for rigorous statistical inference in high-dimensional settings.
The Mechanism of Asymptotic Normality
In classical frequentist statistics, the sample mean satisfies a CLT because it is a linear combination of independent random variables. The empirical Wasserstein distance, by contrast, is highly non-linear and computed via a complex optimization problem. Why, then, should its fluctuations around the true cost follow a Gaussian distribution?
The answer lies in Kantorovich duality. The optimal transport cost can be rewritten as a supremum over a constrained class of functions known as dual potentials. In Central limit theorems for empirical transportation cost in general dimension, Eustasio del Barrio and Jean-Michel Loubes demonstrated that when the underlying target distributions differ, the optimal dual potentials are unique and stable under small perturbations.
Because these potentials are stable, a small change in the empirical distribution results in a predictable, approximately linear change in the transport cost. This linearization allows researchers to apply the functional delta method—a tool that propagates the weak convergence of empirical processes (which naturally converge to Gaussian limits via the Donsker theorem) through the optimal transport functional. As a result, the scaled empirical transportation cost converges to a Gaussian distribution in arbitrary dimensions. This foundational behavior holds true even under minimal assumptions and general cost functions, as shown by Eustasio del Barrio, Alberto González-Sanz, and Jean-Michel Loubes in Central limit theorems for general transportation costs.
Bypassing the Curse of Dimensionality
While unregularized Wasserstein distances satisfy a CLT when the two distributions are strictly different, they still suffer from the curse of dimensionality when estimating distances to a target distribution: the estimation error scales as $O(n^{-1/d})$, requiring exponentially more samples as the dimension $d$ grows. To achieve a practical parametric convergence rate of $O(n^{-1/2})$, researchers alter the transport problem using three primary mechanisms: entropic regularization, smoothing, and projection.
Empirical Data (High-Dimensional Spikes)
│
┌──────────┼──────────┐
▼ ▼ ▼
[Entropic] [Smooth] [Sliced]
Blur plans Blur data Project to 1D
│ │ │
└──────────┼──────────┘
▼
Stable Dual Potentials & Maps
│
▼
Parametric Rate (1/√n) & CLT
Entropic Regularization
By adding an entropic penalty to the optimal transport objective, the transport plan is forced to become more diffuse. In Statistical bounds for entropic optimal transport: sample complexity and the central limit theorem, Gonzalo Mena and Jonathan Weed proved that this penalty completely bypasses the curse of dimensionality, establishing a $1/\sqrt{n}$ convergence rate and an accompanying CLT for unbounded, high-dimensional data.
Building on this, Ziv Goldfeld et al. derived the exact asymptotic distributions for the underlying entropic transport maps and dual potentials themselves in Limit theorems for entropic optimal transport maps and Sinkhorn divergence. This work unlocked the limiting distribution under the null hypothesis, providing a mathematical foundation for valid independence testing.
Gaussian Smoothing
Instead of regularizing the transport plan, one can regularize the data itself by convolving the distributions with a Gaussian kernel. In Smooth p-Wasserstein Distance: Structure, Empirical Approximation, and Statistical Applications, Sloan Nietert, Ziv Goldfeld, and Kengo Kato demonstrated that the smoothed $p$-Wasserstein distance achieves the parametric $n^{-1/2}$ rate regardless of dimension.
By treating the smoothed distance as a functional on dual Sobolev spaces, Ziv Goldfeld et al. established a comprehensive limit distribution theory in Limit distribution theory for smooth p-Wasserstein distances, enabling tasks like calculating confidence intervals for high-dimensional generative models.
Sliced Projections
Another way to avoid high-dimensional sample complexity is to project the distributions onto one-dimensional lines, where optimal transport is trivial to compute and statistically efficient. In Distributional Convergence of the Sliced Wasserstein Process, Jiaqi Xi, Jonathan Niles-Weed, and Jonathan Weed proved that the empirical process representing all possible one-dimensional projections converges to a tight Gaussian process. This functional CLT allows researchers to derive the statistical properties of many different sliced distance variants simultaneously.
Expanding the Geometric Domain
The development of these limit theorems has extended beyond standard continuous-to-continuous transport problems into more complex structural domains:
- Semi-Discrete Settings: In many applications, such as resource allocation or clustering, a continuous distribution is transported to a discrete one. Eustasio del Barrio, Alberto González-Sanz, and Jean-Michel Loubes established CLTs for this domain in Central limit theorems for semi-discrete Wasserstein distances, proving that the estimation error converges to the supremum of a Gaussian process. Further refining this, Ritwik Sadhu, Ziv Goldfeld, and Kengo Kato quantified how quickly entropic semi-discrete maps converge to true Brenier maps in Approximation rates of entropic maps in semidiscrete optimal transport, validating the use of entropic estimators in dual Lipschitz spaces.
- Infinite-Dimensional Spaces: When analyzing populations of curves or functions modeled as Gaussian processes, the underlying space itself becomes infinite-dimensional. Aasa Feragen and Anton Mallasto proved in Learning from uncertain curves: The 2-Wasserstein metric for Gaussian processes that finite-dimensional approximations of these infinite-dimensional metrics converge reliably to their true continuous values as sampling resolution increases, making functional data analysis tractable.
Applications to Statistical Inference
Establishing these Gaussian limits transitions optimal transport from a purely geometric optimization framework into a rigorous statistical paradigm.
Goodness-of-Fit and Normality Testing
In generative modeling, evaluating whether a latent space matches a target distribution (like a standard Gaussian) traditionally required training unstable adversarial discriminator networks. In Reforming Generative Autoencoders via Goodness-of-Fit Hypothesis Testing, Aaron Palmer, Dipak K. Dey, and Jinbo Bi leveraged the asymptotic equivalence between certain classical frequentist normality tests (like the Shapiro-Wilk test) and the $L_2$-Wasserstein distance. This connection provides an objective, non-adversarial measure of when a latent space has reached normality.
Analyzing Deformed Distributions
When datasets undergo geometric warping, shifting, or stretching, standard statistical tests fail to recognize their shared underlying structure. By proving that empirical variations of Wasserstein distances converge to a Gaussian process, the authors of Central limit theorem and bootstrap procedure for Wasserstein's variations with an application to structural relationships between distributions developed a formal bootstrap procedure to test for structural relationships between deformed probability distributions.
High-Dimensional Probability Approximations
Finally, the geometric properties of optimal transport can be inverted to improve classical statistical theory itself. In Strong Gaussian Approximation for the Sum of Random Vectors, Nazar Buzun, Nikolay Shvetsov, and Dmitry V. Dylov utilized Wasserstein distances to measure the proximity between a sum of independent random vectors and a target Gaussian distribution. By calculating the minimum work required to transform one distribution into another, they derived tighter convergence bounds than traditional Fourier-based methods, making high-dimensional learning theory more practical when features vastly outnumber observations.
Key Foundational Work
Reforming Generative Autoencoders via Goodness-of-Fit Hypothesis Testing (2018) — contributed 2 concept elements to this direction.
Learning from uncertain curves: The 2-Wasserstein metric for Gaussian processes (2017) — contributed 2 concept elements to this direction.