Abstract
We extend Robert McCann’s treatment of the existence and uniqueness of an optimal transport map between two probability measures on a Euclidean space to a class of possibly infinite measures, finite outside neighbourhoods of the origin. For convergent sequences of pairs of such measures, we study the stability of the multivalued transport maps and associated quantile contours, defined as the images of spheres under these maps. The measures involved in the coupling are not required to be absolutely continuous, there are no restrictions on their supports, and no moment assumptions are needed. Weakly convergent sequences of probability measures forming a special case, our results apply to the recently introduced Monge–Kantorovich depth contours. The set-up involving infinite limit measures is applied to regularly varying probability measures: we derive tail limits of transport maps and of quantile contours defined with respect to a judiciously chosen spherical reference measure. Examples are discussed in detail.
Introduction
In univariate statistics, quantiles provide a robust framework for understanding the distribution of data. However, extending this concept to multivariate settings is challenging because Euclidean space lacks a natural, total ordering. Recent developments in optimal transport have introduced Monge–Kantorovich (M-K) quantile contours as a solution. By mapping a simple, symmetric reference distribution (such as a uniform ball) onto a target distribution, the images of concentric spheres under the optimal transport map serve as multivariate quantiles.
Despite their promise, the theoretical foundations of M-K contours have historically relied on restrictive assumptions, including the existence of finite moments (e.g., variance) and the absolute continuity of the distributions. This paper addresses these limitations, extending the theory to handle heavy-tailed distributions and providing a rigorous mathematical description of how these contours behave in the "tails"—the extreme limits of the distribution.
Theoretical Framework: Beyond Probability Measures
The authors build upon Robert McCann’s theory of displacement convexity, extending it to a broader class of measures denoted as $\mathcal{M}_0$. These measures are allowed to be infinite at the origin while remaining finite elsewhere. This extension is critical for studying regular variation, the mathematical language used to describe heavy-tailed phenomena where extreme events occur with higher frequency than in Gaussian models.
Graphical Convergence and Stability
A key technical challenge in this extension is handling distributions that are not "smooth" or those that contain gaps and atoms. To maintain stability, the authors utilize graphical convergence of subdifferentials rather than simple pointwise convergence. This approach allows for the treatment of multi-valued transport maps. The paper proves that if a sequence of distributions converges, their corresponding M-K quantile contours also converge according to the Hausdorff metric, a standard measure of distance between geometric shapes. Notably, this stability is achieved without requiring finite moments or absolute continuity, relying instead on the geometric property of cyclical monotonicity.
Tail Limits and Homogeneity
One of the paper's primary contributions is the derivation of tail limits for regularly varying distributions. The authors demonstrate that when a transport map is correctly scaled, it converges to a limit map $\bar{\psi}$ that is positively homogeneous. Mathematically, this is expressed as $\nabla\bar{\psi}(\lambda x) = \lambda \nabla\bar{\psi}(x)$.
This homogeneity implies that at extreme limits, the quantile contours settle into a fixed geometric shape that simply scales upward as one moves further into the tail. This provides a rigorous way to define the "shape" of extreme risk in multivariate data.
Geometric Insights and Examples
The research highlights how the limiting shape of a quantile contour is intrinsically linked to the underlying structure of the data:
- Elliptical Distributions: For generalizations of the multivariate normal distribution, the extreme contours converge to ellipsoids.
- Factor Models: In systems driven by a finite number of independent heavy-tailed shocks (linear or max-linear models), the extreme contours are shown to be polytopes. These polyhedral shapes reflect the directional nature of the underlying shocks, providing a visual representation of how extremes propagate through the system.
Conclusion and Future Directions
By allowing measures to be infinite at the origin, the authors provide a framework to rigorously define the intensity of extreme events. While this paper establishes the population-level theory, several avenues for future work remain. The practical calculation of these limit shapes requires solving complex Monge–Ampère partial differential equations, potentially necessitating numerical methods from computational fluid dynamics. Furthermore, extending these results to handle finite data samples (random measures) is a vital next step for the statistical application of transport-based depth contours.
Summary updated