Scalable Optimal Transport via Sliced Wasserstein Projections

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2478 papers521 conceptsfrozen 2026-10-04

Optimal Transport (OT) provides a mathematically rigorous way to measure the distance between probability distributions by calculating the minimum "effort" required to transform one into another. While the resulting Wasserstein distance is a gold standard for tasks like generative modeling and domain adaptation, it suffers from a severe computational bottleneck: calculating it exactly scales cubically ($O(n^3)$) with the number of data points.

To bypass this "curse of dimensionality," researchers have converged on a powerful alternative: the Sliced Wasserstein (SW) distance. By projecting high-dimensional data onto one-dimensional lines, we can transform a complex optimization problem into a series of simple sorting tasks.

The Intuition: Slicing the Radon Transform

The core idea behind Sliced Wasserstein distances relies on a property of the Radon transform: a high-dimensional probability distribution is uniquely determined by the collection of all its one-dimensional projections.

In high dimensions, finding the optimal coupling between two distributions is a massive linear programming problem. However, in one dimension, the optimal transport map is trivial. If you have two sets of points on a line, the "cheapest" way to move one set to the other is simply to sort both sets and match them in order. This sorting operation takes only $O(n \log n)$ time.

As noted by Keanu Sisouk, Julie Delon, and Julien Tierny in A User's Guide to Sampling Strategies for Sliced Optimal Transport, the SW distance is computed by averaging these 1D Wasserstein distances across many different projection directions. This turns a single $O(n^3)$ problem into many $O(n \log n)$ problems, which is significantly faster and more scalable for modern deep learning.

Beyond Random Projections

The simplest version of SW uses random directions sampled uniformly from a unit sphere. However, random directions are often "blind" to the actual geometry of the data, requiring thousands of projections to achieve an accurate estimate. Recent research has focused on making these projections "smarter" or more "adaptive."

Energy-Based and Markovian Slicing

Khai Nguyen and Nhat Ho have pioneered several methods to find more informative directions. In Energy-Based Sliced Wasserstein Distance, they propose a parameter-free approach where the distribution of projection directions is itself an energy-based model. This forces the metric to automatically focus on directions where the distributions are most distinct.

In Markovian Sliced Wasserstein Distances, Nguyen, Ren, and Ho move away from independent projections entirely. By using a Markov chain to generate sequences of directions, the algorithm can "explore" the space of projections more effectively, leading to faster convergence in generative tasks.

Random-Path Projections

Another innovation is the Random-Path Projecting Direction method. Instead of projecting onto arbitrary lines, this approach projects data along paths defined by the differences between samples from the two distributions. This ensures that the "slices" are always relevant to the actual gap between the datasets, providing a more discriminative signal for training diffusion models and gradient flows.

Theoretical Guarantees and Generalization

While slicing is a heuristic for speed, it is backed by increasingly rigorous theory. Ruben Ohana and Kimia Nadjahi have applied PAC-Bayesian theory to prove that adaptive SW distances—those where directions are learned from data—remain statistically robust. Their work provides generalization bounds that guarantee these learned metrics won't overfit to specific samples, a crucial property for high-dimensional representation learning.

Alternative Paths to Scalability

While slicing is the dominant paradigm, other researchers explore different structural approximations to the Wasserstein distance:

Impact Across the AI Landscape

The ability to compute geometric distances at scale has unlocked new capabilities in diverse fields:

  • Generative Modeling: Zhiwu Huang demonstrated that Sliced Wasserstein GANs can generate high-resolution video by matching distributions in 3D space, a task previously impossible with exact OT.
  • Fairness: Silvia Chiappa and colleagues at DeepMind use OT to nudge model predictions toward fair distributions, treating bias as a distribution-matching problem.
  • Climate Science: In Validating Climate Models with Spherical Convolutional Wasserstein Distance, Robert C. Garrett uses these metrics to identify regional biases in global climate simulations, accounting for spatial geometry that standard statistical checks miss.

By reducing the computational cost of "moving earth," these slicing and projection strategies have transformed Optimal Transport from a theoretical curiosity into a practical engine for high-dimensional machine learning.

Papers in this version

Showing 60 of 478 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 440 papers
  2. Version 1 · 2026-10-04 · 478 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_wasserstein_distance_scalability_8f2a744f,
  title = {Wasserstein distance scalability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/wasserstein-distance-scalability/rdr_8f2a744f59984801aaba9af95440167c},
  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_scalable_optimal_transport_via_sliced_wa_8f2a744f_v1,
  title = {Scalable Optimal Transport via Sliced Wasserstein Projections},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scalable-optimal-transport-via-sliced-wasserstein-projections/rdr_8f2a744f59984801aaba9af95440167c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}