Optimal Transport (OT) provides a powerful geometric framework for comparing probability distributions by measuring the minimum cost of "moving" one distribution to match another. While the resulting Wasserstein distance is highly discriminative and captures the underlying geometry of data, it suffers from a severe "curse of dimensionality." Calculating the exact Wasserstein distance between two $d$-dimensional distributions with $n$ samples typically requires solving a linear programming problem with $O(n^3 \log n)$ complexity, making it impractical for modern deep learning.
To bypass this bottleneck, researchers have converged on a clever workaround: Slicing. By projecting high-dimensional distributions onto one-dimensional lines, the complex transport problem collapses into a simple sorting operation that can be solved in $O(n \log n)$ time.
The Efficiency of One Dimension
The core intuition behind the Sliced-Wasserstein (SW) distance is that optimal transport is "trivial" in one dimension. As explained by Soheil Kolouri, Matthew Thorpe, and Gustavo K. Rohde in their foundational tutorial, the optimal way to move mass between two 1D distributions is simply to match their sorted values (or quantiles).
The SW distance leverages this by:
- Projecting the high-dimensional data onto a set of random 1D lines (slices).
- Computing the 1D Wasserstein distance for each slice via sorting.
- Averaging these 1D distances to approximate the global high-dimensional distance.
This mechanism transforms a heavy optimization problem into a series of fast sorting tasks. In Sliced-Wasserstein Autoencoder: An Embarrassingly Simple Generative Model, Soheil Kolouri and colleagues demonstrated that this approach provides smooth gradients for optimization, allowing generative models to shape their latent spaces without the instability of adversarial training.
Beyond Random Projections
While early methods relied on a large number of random projections to approximate the true geometry, recent work focuses on making these "slices" more informative. Sliced Wasserstein Generative Models introduced the idea of learned projections, where the model identifies the specific directions that best distinguish two distributions, significantly reducing the number of slices needed.
Other researchers have extended slicing to non-Euclidean geometries:
- Spherical Slicing: To compare distributions on hyperspheres (common in geospatial data or normalized embeddings), Clément Bonet and Nicolas Courty developed Spherical Sliced-Wasserstein (SSW), which projects data onto "great circles" rather than straight lines.
- Manifold Slicing: For brain signal analysis, Clément Bonet and Lucas Drumetz introduced SPDSW, which slices distributions of Symmetric Positive Definite (SPD) matrices by projecting them onto one-dimensional geodesics.
- Tree Slicing: Hoang V. Tran and Tan Minh Nguyen proposed Tree-Sliced Entropy Partial Transport, which uses tree structures instead of lines to capture more complex hierarchical relationships in the data.
From Distances to Embeddings and Kernels
The utility of slicing extends beyond just measuring distance; it allows for the creation of fixed-length representations for complex, variable-sized data like point clouds or documents.
In Word Mover's Embedding, Lingfei Wu and Ian En-Hsu Yen use a random features approach to turn expensive alignment-based distances into fast document vectors. Similarly, Tal Amir and Nadav Dym introduced the Fourier Sliced-Wasserstein Embedding, which maps multisets into Euclidean space while provably preserving their geometric structure.
For tasks like classification and regression on distributions, slicing enables the construction of "universal kernels." Mathieu Carrière and Marco Cuturi utilized this in Sliced Wasserstein Kernel for Persistence Diagrams to integrate topological data analysis into standard machine learning pipelines. More recently, Dimitri Meunier and Massimiliano Pontil proved in Distribution Regression with Sliced Wasserstein Kernels that these kernels can approximate any continuous function on distributions, providing a robust alternative to Maximum Mean Discrepancy (MMD).
Recovering the Transport Map
A traditional criticism of sliced methods is that they provide a distance value but lose the "transport plan"—the map showing where each point in the source distribution should go. Xinran Liu and Ashkan Shahbazi addressed this in Expected Sliced Transport Plans by "lifting" the 1D transport solutions back into high-dimensional space. By averaging these lifted plans, they recover meaningful correspondences between distributions, making sliced transport useful for tasks like domain adaptation and image morphing where the mapping itself is the goal.