Introduction
In modern machine learning, datasets are increasingly represented as probability distributions rather than simple vectors. This is particularly true in fields like natural language processing (word distributions), biology (cell populations), and computer vision. The natural geometric framework for these distributions is the Wasserstein space. While the Wasserstein distance is a powerful tool for comparing individual distributions, comparing datasets of distributions (distributions of distributions) is computationally prohibitive.
To address this, researchers often employ "slicing" techniques—projecting high-dimensional data onto one-dimensional lines to simplify calculations. However, the curved geometry of Wasserstein space makes defining "straight lines" or "flat hyperplanes" non-trivial. This paper introduces the Busemann function to the Wasserstein space, providing a principled geometric tool to define these structures and create efficient "sliced" distances.
Geometric Foundations: Busemann Functions and Geodesic Rays
The Busemann function is a classical tool in differential geometry used to define hyperplanes in non-Euclidean spaces. In this work, the authors bridge the gap between geometric theory and optimal transport (OT) through three main contributions:
- Characterizing Geodesic Rays: Identifying conditions under which a path in Wasserstein space can be extended to infinity. For 1D distributions, this occurs when the difference between quantile functions is non-decreasing.
- Deriving Closed-Forms: Proving that for 1D distributions and Gaussian measures, the Busemann function has a closed-form expression. This eliminates the need for expensive optimization, allowing for direct calculation.
- Novel Slicing Schemes: Using Busemann functions as projection operators, the authors define two new distances: SWB1DG (1D Gaussian-based) and SWBG (higher-dimensional Gaussian-based).
Experimental Results
Correlation with Optimal Transport Dataset Distance (OTDD)
A key metric for any sliced distance is how well it approximates the "gold standard" OTDD. On the CIFAR10 dataset, the proposed Busemann-based distances demonstrated significantly higher correlation with OTDD than previous methods like SOTDD.
Gradient Flows and Convergence
The authors also tested their methods on "gradient flows"—the process of morphing one distribution into another. Using the synthetic "3 Rings" dataset, they visualized how quickly and accurately different distances could drive a source distribution toward a target.
Quantitatively, the Busemann-based flows reached lower error rates in fewer iterations. This efficiency was also reflected in transfer learning tasks (e.g., MNIST to USPS), where Busemann distances improved classifier accuracy in low-data (k-shot) regimes. For instance, moving from MNIST to USPS with 10 samples per class, SWB1DG achieved 77.8% accuracy, outperforming the baseline.
Conclusion and Future Work
This work provides a robust theoretical and practical framework for applying Busemann functions to the Wasserstein space. By turning complex geometric limits into simple closed-form expressions, the authors have made high-dimensional dataset comparison significantly more accessible. Future research will focus on scaling these methods to even larger datasets and exploring their use in Wasserstein-space Principal Component Analysis (PCA).
Summary updated