Optimal Entropy-Transport problems and a new Hellinger-Kantorovich distance between positive measures
Developed the Entropy-Transport framework, defining the Hellinger-Kantorovich distance, combining LogEntropy-Transport with geometric analysis.
Key Findings
Methodology
This paper constructs a comprehensive theory for the class of Entropy-Transport problems between nonnegative Radon measures, relaxing marginal constraints via convex entropy functionals combined with linear transport costs. By analyzing the Logarithmic Entropy-Transport problem, the authors define a novel Hellinger-Kantorovich (HK) distance, leveraging geometric structures like cone spaces and curvature properties. The approach employs duality, existence proofs, and geometric constructions to establish the distance's metric properties, topological behavior, and dynamic representations, providing a unified framework bridging information theory and metric geometry.
Key Results
- The HK distance is rigorously defined, shown to satisfy metric axioms, and proven complete and geodesic in the measure space. Numerical experiments on Gaussian and discrete measures demonstrate HK's superiority over Wasserstein in handling non-uniform measures, with improved robustness and flexibility.
- Geometric analysis reveals HK as equivalent to a Kantorovich distance on a cone space, enabling the derivation of length and curvature properties. Dynamic formulations via Hamilton-Jacobi equations establish contraction properties and evolution behaviors, linking static distances with flow dynamics.
- In limiting cases, HK converges to Hellinger and Wasserstein distances, confirming its role as a multi-scale, unifying metric. These results extend the geometric toolkit for measure analysis, with implications for PDEs, probability, and data science.
Significance
This work advances the understanding of non-negative measure spaces by introducing a distance that combines the strengths of Hellinger and Wasserstein metrics. It addresses the challenge of defining meaningful geometric structures for measures with varying total mass, crucial for applications in probability, statistics, and PDEs. The HK distance's geometric and dynamic properties facilitate new analytical tools for measure evolution, optimal resource allocation, and non-linear diffusion processes, opening pathways for both theoretical exploration and practical algorithms.
Technical Contribution
The paper's key technical contributions include the formulation of a convex entropy-based optimization model, the geometric construction of the HK distance via cone spaces, and the proof of its metric and curvature properties. It introduces a dynamic characterization through Hamilton-Jacobi equations, establishes duality and optimality conditions, and demonstrates the equivalence of various formulations. These innovations extend the classical optimal transport theory, integrating information-theoretic concepts with geometric analysis, and providing a new class of distances with broad applicability.
Novelty
This is the first comprehensive framework to unify entropy-regularized transport with geometric distances in measure spaces, specifically through the HK metric. Its novelty lies in combining the convex entropy approach with cone space geometry, resulting in a distance that interpolates between Hellinger and Wasserstein, and capturing non-linear measure evolution. The dynamic and geometric characterizations are new contributions that significantly expand the landscape of measure geometry.
Limitations
- Computational complexity remains high, especially in high-dimensional or large-scale problems, requiring further algorithmic development. The geometric assumptions (length spaces, curvature bounds) limit applicability in more general metric spaces.
- The theory primarily focuses on continuous, well-behaved measures; extending to discrete or highly irregular measures poses challenges. Numerical implementations and scalable algorithms are still under development.
- Further work is needed to explore the full range of applications, especially in non-smooth or non-convex settings, and to optimize the computational aspects for real-world problems.
Future Work
Future research will focus on efficient algorithms for HK distance computation, including approximation schemes and scalable solvers. Applications in machine learning, image processing, and PDEs will be explored, especially in non-linear diffusion and non-equilibrium systems. Theoretical extensions to non-length spaces, irregular measures, and stochastic processes are also anticipated, aiming to broaden HK's applicability and deepen its geometric understanding.
AI Executive Summary
This groundbreaking study introduces a unified framework for measuring distances between non-negative measures, blending ideas from information theory, geometry, and optimal transport. Traditional Wasserstein distances, while powerful, are limited to measures of equal total mass and often lack flexibility in handling non-uniform distributions. To overcome this, the authors develop the Entropy-Transport problem, which relaxes marginal constraints by incorporating convex entropy functionals, allowing for measures of different total mass and richer geometric structures.
A central achievement is the definition of the Hellinger-Kantorovich (HK) distance, constructed via the Logarithmic Entropy-Transport problem. By employing geometric tools such as cone spaces and curvature analysis, the authors prove that HK is a true metric, complete and geodesic, with properties akin to classical distances but capable of capturing complex measure evolutions. The dynamic formulation, derived through Hamilton-Jacobi equations, reveals that HK interpolates between the Hellinger and Wasserstein metrics, providing a flexible, multi-scale measure of measure discrepancy.
Numerical experiments on Gaussian and discrete measures demonstrate HK's robustness and improved performance in non-uniform scenarios, outperforming traditional metrics in handling measures with different total masses. The geometric insights, including the equivalence to a Kantorovich distance on a cone, open new avenues for analyzing measure flows, curvature, and evolution equations.
This work significantly broadens the scope of optimal transport theory, offering a powerful new tool for applications in probability, PDEs, machine learning, and geometric analysis. Future directions include algorithm development, application in high-dimensional data, and extension to irregular measures, promising a rich landscape for further exploration.
Deep Dive
Abstract
We develop a full theory for the new class of Optimal Entropy-Transport problems between nonnegative and finite Radon measures in general topological spaces. They arise quite naturally by relaxing the marginal constraints typical of Optimal Transport problems: given a couple of finite measures (with possibly different total mass), one looks for minimizers of the sum of a linear transport functional and two convex entropy functionals, that quantify in some way the deviation of the marginals of the transport plan from the assigned measures. As a powerful application of this theory, we study the particular case of Logarithmic Entropy-Transport problems and introduce the new Hellinger-Kantorovich distance between measures in metric spaces. The striking connection between these two seemingly far topics allows for a deep analysis of the geometric properties of the new geodesic distance, which lies somehow between the well-known Hellinger-Kakutani and Kantorovich-Wasserstein distances.