Dimension Independent Optimal Transport via Regularization

This direction is merged into another direction: dimension-independent convergence (dimension-independent sample complexity). Its last version stays readable and citable here.

Version 1 of 133 papers41 conceptsfrozen 2026-10-04

Optimal Transport (OT) provides a powerful geometric framework for comparing probability distributions by measuring the minimum "work" required to transform one into another. However, for years, its application to high-dimensional machine learning was stifled by the curse of dimensionality. In its pure form, the sample complexity of the Wasserstein distance scales as $n^{-1/d}$, meaning that as the number of dimensions $d$ increases, the number of samples $n$ required to estimate the distance grows exponentially.

A recent surge of research has transformed OT from a theoretical curiosity into a practical tool for high-dimensional data by introducing regularization. By smoothing the transport problem or imposing structural constraints, researchers have achieved a "parametric" convergence rate of $n^{-1/2}$—a rate that is entirely independent of the data's dimensionality.

The Intuition of Entropic Smoothing

The most influential breakthrough in this area is the use of entropic regularization. In Learning Generative Models with Sinkhorn Divergences, Aude Genevay and Marco Cuturi demonstrate that adding an entropy term to the OT objective does more than just make the optimization faster via the Sinkhorn algorithm; it fundamentally changes the statistical nature of the problem.

Standard OT looks for a "hard" assignment between points, which is highly sensitive to the specific locations of samples in high-dimensional space. Entropic OT instead looks for a "soft" coupling. This smoothing allows the resulting Sinkhorn Divergence to interpolate between the geometry-aware Wasserstein distance and the statistically efficient Maximum Mean Discrepancy (MMD).

As proven in Sample Complexity of Sinkhorn divergences and Statistical bounds for entropic optimal transport by Gonzalo Mena and Jonathan Weed, this regularization allows the empirical distance to converge at a rate of $1/\sqrt{n}$. This means that even in thousands of dimensions, we can estimate the distance between distributions as reliably as we can estimate a simple mean.

Factored Couplings and Low Rank Structure

While entropy smooths the transport plan, another approach to beating the curse of dimensionality is to restrict the "rank" of the transport. In Statistical Optimal Transport via Factored Couplings, Aden Forrow, Philippe Rigollet, and their colleagues propose that complex high-dimensional distributions often move in a small number of independent "pieces" or along low-dimensional subspaces.

By imposing a structural rank constraint on the transport plan—known as FactoredOT—the estimation error is decoupled from the ambient dimensionality. This approach is particularly effective for biological data, such as single-cell RNA sequencing, where the "features" (genes) are numerous, but the underlying biological transitions are governed by a few latent factors. Like entropic OT, this method recovers the $n^{-1/2}$ convergence rate, ensuring stable estimation in spaces where standard OT would fail.

Scaling to Generative Modeling

The theoretical stability of regularized OT has paved the way for new generative model architectures. Traditional Generative Adversarial Networks (GANs) often suffer from training instability due to their minimax (saddle-point) optimization. Alexander Korotin and Evgeny Burnaev in Wasserstein-2 Generative Networks leverage Input Convex Neural Networks (ICNNs) to learn optimal transport mappings directly.

Because these networks are constrained to be convex, the researchers can bypass the unstable minimax game and instead solve a simpler, non-minimax optimization. This approach scales to thousands of dimensions without the "blurring" effect sometimes seen in entropic methods, maintaining the precision of the learned distribution while benefiting from the convergence guarantees of the OT framework.

Beyond Simple Distances

The implications of dimension-independent OT extend into broader statistical tasks:

By shifting the focus from "exact" transport to "regularized" transport, this field has successfully bridged the gap between the elegant geometry of Monge and Kantorovich and the practical requirements of modern, high-dimensional machine learning.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dimension_independent_optimal_transport__a4647263,
  title = {Dimension Independent Optimal Transport via Regularization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dimension-independent-optimal-transport-via-regularization/rdr_a4647263cbb84d1c986379f1b1ccdfbc},
  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_dimension_independent_optimal_transport__a4647263_v1,
  title = {Dimension Independent Optimal Transport via Regularization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dimension-independent-optimal-transport-via-regularization/rdr_a4647263cbb84d1c986379f1b1ccdfbc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}