Geometric Characteristics of Wasserstein Metric on SPD(n)

Yihao Luo, Shiqiang Zhang, Yueqi Cao, Huafei SunarXiv 2020 · DOI · arXiv

Abstract

Wasserstein distance, especially among symmetric positive-definite matrices, has broad and deep influences on development of artificial intelligence (AI) and other branches of computer science. A natural idea is to describe the geometry of $SPD\left(n\right)$ as a Riemannian manifold endowed with the Wasserstein metric. In this paper, by involving the fiber bundle, we obtain explicit expressions for some locally geometric quantities, including geodesics, exponential maps, the Riemannian connection, Jacobi fields and curvatures. Furthermore, we discuss the behaviour of geodesics and prove that the manifold is globally geodesic convex with non-negative curvatures but no conjugate pair and cut locus. According to arithmetic estimates, we find curvatures can be controlled by the minimal eigenvalue.

Introduction

Symmetric Positive-Definite (SPD) matrices are fundamental in modern data science, serving as covariance matrices in statistics and diffusion tensors in medical imaging. To compare these matrices, the Wasserstein distance (or Earth Mover's Distance) is often employed due to its effectiveness in measuring differences between probability distributions. While the distance metric itself is well-understood, the underlying Riemannian geometry of the space $SPD(n)$ under this metric has remained partially obscured. This paper provides a complete Riemannian characterization, deriving explicit formulas for geodesics, curvatures, and connections.

Mathematical Framework

The authors treat $SPD(n)$ as a Riemannian manifold by utilizing the framework of Riemannian submersions and fiber bundles. Specifically, they represent $SPD(n)$ as a quotient of the general linear group $GL(n)$ by the orthogonal group $O(n)$.

A central technical tool in this characterization is the Sylvester equation. Most geometric quantities in this space are defined through the solution to the matrix equation $A\Gamma_A[X] + \Gamma_A[X]A = X$. To facilitate practical computation, the authors provide an efficient algorithm based on eigenvalue decomposition to solve for the Sylvester operator $\Gamma_A[X]$.

Extracted table from page 5 of Geometric Characteristics of Wasserstein Metric on SPD(n)

By "lifting" vectors from the SPD manifold to the larger $GL(n)$ space—where the geometry is simpler—the authors perform calculations and project the results back to derive Wasserstein-specific formulas.

Geometric Insights

The study reveals several unique properties of the Wasserstein metric on $SPD(n)$:

  • The Wasserstein Correction: The exponential map consists of a standard Euclidean term plus a "Wasserstein correction" term, which accounts for the manifold's intrinsic curvature.
  • Geodesic Convexity: The manifold is proven to be globally geodesically convex. Any two SPD matrices can be joined by a unique minimal geodesic that remains within the set of positive-definite matrices, a property crucial for optimization.
  • Absence of Conjugate Points: Despite having non-negative curvature—a condition that typically leads to converging geodesics—$SPD(n)$ under the Wasserstein metric contains no conjugate pairs or cut loci. Geodesics in this space tend to diverge or reach the manifold's boundary before they can ever cross again.
  • Curvature Control: Sectional curvature is strictly non-negative and bounded by a value inversely proportional to the minimal eigenvalue ($\lambda_{min}$) of the matrix. This implies that the space becomes highly curved only as a matrix approaches singularity.

Key Results

The paper offers explicit, closed-form expressions for several key geometric quantities:

  1. Explicit Geodesics: A closed-form solution for the path $\gamma(t)$ between any two matrices $A_1$ and $A_2$.
  2. Wasserstein Radius: The "radius" of the manifold at point $A$ is defined as $r(A) = \sqrt{\lambda_{min}(A)}/2$, representing the maximum distance a geodesic can travel before hitting the boundary.
  3. Curvature Tensors: Succinct expressions for the Riemannian connection, curvature tensor, and scalar curvature using the Sylvester operator.

Discussion and Future Work

While the manifold is geodesically convex, it is not complete; geodesics can exit the manifold by reaching the boundary of positive-definiteness. This incompleteness presents challenges for certain global analysis techniques. Future work involves determining the exact degree of symmetry for this manifold and translating these theoretical insights into practical data processing algorithms for real-world applications.

Summary updated

Abstract

Wasserstein distance, especially among symmetric positive-definite matrices, has broad and deep influences on development of artificial intelligence (AI) and other branches of computer science. A natural idea is to describe the geometry of $SPD\left(n\right)$ as a Riemannian manifold endowed with the Wasserstein metric. In this paper, by involving the fiber bundle, we obtain explicit expressions for some locally geometric quantities, including geodesics, exponential maps, the Riemannian connection, Jacobi fields and curvatures. Furthermore, we discuss the behaviour of geodesics and prove that the manifold is globally geodesic convex with non-negative curvatures but no conjugate pair and cut locus. According to arithmetic estimates, we find curvatures can be controlled by the minimal eigenvalue.
BibTeX
@misc{luo2020geometric,
  author = {Yihao Luo and Shiqiang Zhang and Yueqi Cao and Huafei Sun},
  title = {{Geometric Characteristics of Wasserstein Metric on SPD(n)}},
  year = {2020},
  eprint = {2012.07106},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2012.07106},
}
Figure