Scalable Optimal Transport Plans via Slicing and Projections

This direction is merged into another direction: optimal transport scalability (computational complexity of optimal transport). Its last version stays readable and citable here.

Version 1 of 117 papers17 conceptsfrozen 2026-10-04

Optimal Transport (OT) is often described as the "gold standard" for comparing probability distributions. Unlike simple divergence measures that tell you if two distributions differ, OT tells you how they differ by finding the most efficient way to move mass from one to the other. This "how" is encoded in the transport plan—a mapping that assigns every point in a source distribution to a destination in the target.

However, calculating this plan in high-dimensional space is notoriously slow, typically scaling cubically with the number of data points. While "Sliced" Wasserstein (SW) distances solved the speed problem by projecting data onto one-dimensional lines, they introduced an "identity crisis": they provide a scalar distance but discard the transport plan. Recent research has focused on bridging this gap, developing methods that retain the $O(n \log n)$ speed of slicing while recovering the high-dimensional mapping required for tasks like image color transfer and domain adaptation.

The Distance vs. Map Dilemma

In one dimension, Optimal Transport is trivial. You simply sort the points in both distributions and match them by rank. This efficiency is the core of Sliced Wasserstein distances: project high-dimensional data onto a random line, solve the 1D problem instantly, and average the results over many lines.

As Eloi Tanguy and Laëtitia Chapel point out in Sliced Optimal Transport Plans, the problem is that these 1D solutions don't naturally "stitch" back together into a coherent high-dimensional map. If you move a point according to its rank on one line, and then according to its rank on a different line, the resulting trajectories often conflict or fail to represent a valid global transformation. Without a functional map, the metric is useless for generative modeling or point cloud registration, where the goal is to actually transform the data, not just measure it.

Lifting 1D Solutions to High Dimensions

To solve this, researchers have moved beyond simple averaging toward "lifting" mechanisms that reconstruct a global plan from local projections.

One prominent approach is Expected Sliced Transport (EST), introduced by Xinran Liu, Matthew Thorpe, and colleagues. In Expected Sliced Transport Plans, they propose a framework that treats the high-dimensional transport plan as the expectation of 1D plans. By mathematically "lifting" these 1D assignments back into the original space, they create a valid transport map that maintains the computational benefits of slicing.

Similarly, Guillaume Mahey and Nicolas Courty developed min-SWGG (Fast Optimal Transport through Sliced Wasserstein Generalized Geodesics). Instead of averaging over many random directions, they search for a specific projection that acts as a reliable proxy for the full Wasserstein distance. This method leverages a closed-form solution for transport between a distribution and a line, providing a functional mapping that is particularly effective for high-resolution image colorization.

Beyond Lines: Subspaces and Trees

While lines are the simplest projection, they aren't always the most informative. Other researchers have looked at more complex structures to capture the geometry of the data.

Neural and Generative Approximations

For truly massive datasets, even optimized projections can be slow. This has led to the rise of neural transport solvers. Yujia Xie and Tuo Zhao introduced SPOT (On Scalable and Efficient Computation of Large Scale Optimal Transport), which treats the transport plan as a "push-forward" of a latent distribution. By using a minimax optimization similar to a GAN, they can recover the transport mapping for high-dimensional data.

A different neural approach involves restricting the search space to specific mathematical functions. Amirhossein Taghvaei and Amin Jalali utilize Input-Convex Neural Networks (ICNNs) to approximate the 2-Wasserstein distance (2-Wasserstein Approximation via Restricted Convex Potentials). Because the optimal transport map for the 2-Wasserstein distance is the gradient of a convex function, ICNNs provide a natural, bias-free way to parameterize and learn the mapping directly.

Why This Works

The shared intuition across these papers is that high-dimensional transport is often redundant. Data usually sits on lower-dimensional manifolds, meaning we don't need to solve the full $N \times N$ coupling matrix to find a good map. By projecting the problem into 1D lines, trees, or subspaces—or by constraining the solution to convex neural potentials—we can find "good enough" transport plans that are computationally feasible while remaining functionally useful for downstream applications.

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_scalable_optimal_transport_plans_via_sli_24c8b866,
  title = {Scalable Optimal Transport Plans via Slicing and Projections},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scalable-optimal-transport-plans-via-slicing-and-projections/rdr_24c8b866fb434cb9bce275c17dd4aa12},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_scalable_optimal_transport_plans_via_sli_24c8b866_v1,
  title = {Scalable Optimal Transport Plans via Slicing and Projections},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scalable-optimal-transport-plans-via-slicing-and-projections/rdr_24c8b866fb434cb9bce275c17dd4aa12/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}