Abstract
This article presents a new class of distances between arbitrary nonnegative Radon measures inspired by optimal transport (OT). These distances are defined by two equivalent alternative formulations: a dynamic formulation defining the distance as a geodesic distance over the space of measures, and a static "Kantorovich" formulation where the distance is the minimum of an optimization problem over pairs of couplings describing the transfer (transport, creation, and destruction) of mass. Both formulations are convex optimization problems, and the ability to switch between them is a crucial property of the models. A primary example is the Wasserstein-Fisher-Rao metric, which benefits from this unified framework.
Introduction
Classical Optimal Transport (OT) is a powerful tool for comparing probability distributions, but it possesses a significant constraint: it strictly requires the two distributions to have the exact same total mass. In many real-world applications—such as comparing images with different brightness levels or tracking cell populations that grow and die—this mass-conservation requirement is a dealbreaker.
To address this, researchers developed "Unbalanced" Optimal Transport (UOT), which allows mass to be transported, created, or destroyed. However, the field has historically been split between two perspectives: the dynamic formulation (viewing mass as a fluid flowing through time) and the static formulation (viewing mass as a "coupling" or matching between points). This paper provides the mathematical "glue" to prove these two views are equivalent, allowing researchers to choose the formulation best suited for their specific application.
The Two Faces of Unbalanced OT
The authors introduce a generalized framework that encompasses both perspectives through convex optimization:
1. The Dynamic Formulation
The authors generalize the Benamou-Brenier formula, a staple of classical OT. In the unbalanced setting, the standard continuity equation (which preserves mass) is augmented with a source term to account for growth and decay. The resulting cost function penalizes two distinct behaviors: the velocity of the mass transport and the rate at which mass is created or destroyed.
2. The Static Formulation
The paper proposes a novel "Kantorovich-like" problem using semi-couplings. Unlike standard OT, which uses a single plan to move mass from source to target, this framework uses a pair of measures $(\gamma_0, \gamma_1)$. These measures describe how much mass leaves a location $x$ and how much (potentially different) amount of mass arrives at location $y$. This decoupling is a clever way to handle mass change while maintaining the convexity of the optimization problem.
The Equivalence Theorem
The core contribution of the work is Theorem 4.3, which proves that the minimal cost of the dynamic "fluid" path is identical to the minimal cost of the static "semi-coupling" matching, provided the cost functions are chosen correctly. This equivalence ensures that the geometric properties of the dynamic formulation are preserved in the more computationally tractable static formulation.
Key Insights and Geometric Interpretations
The Wasserstein-Fisher-Rao (WFR) Metric
The paper highlights the WFR metric as a prototypical example of this framework. WFR interpolates between the Wasserstein distance (pure transport) and the Fisher-Rao/Hellinger distance (pure mass change). By applying their framework, the authors show that WFR—initially defined dynamically—automatically gains a static Kantorovich formulation.
Lifting to the Cone
A beautiful geometric interpretation offered by the authors is the concept of lifting metrics to a "cone" manifold. In this view, the base space (e.g., $\mathbb{R}^d$) is the base of a cone, and the "tip" represents zero mass. Mass destruction is interpreted geometrically as transporting mass to the tip of the cone. This provides a rigorous manifold-based foundation for unbalanced distances.
Results and Convergence
Beyond the equivalence theorem, the authors derive dual formulations for these problems. In the unbalanced case, the dual variables (the "potentials") must satisfy a Hamilton-Jacobi equation that includes a zeroth-order term, reflecting the cost of mass creation and destruction.
Furthermore, the authors prove the limit behavior of these metrics through $\Gamma$-convergence. They demonstrate that as the penalty for mass change goes to infinity, Unbalanced OT (specifically the WFR metric) converges back to standard Optimal Transport between normalized measures. This confirms that the new framework is a true, consistent generalization of classical theory.
Limitations and Future Directions
While the paper establishes a robust theoretical foundation, it notes a few areas for further exploration:
- Domain Geometry: The proof for the equivalence theorem currently requires the domain to be "star-shaped." While this covers most practical cases (like boxes or balls), it leaves a gap for complex, non-convex domains.
- Computational Implementation: The paper focuses on the mathematical structure rather than numerical solvers. However, the static formulation's structure suggests promising directions for new, efficient algorithms.
- Regularity: Further research is needed to explore the specific regularity of the optimal dual potentials in various high-dimensional applications.
Summary updated