Introduction
Optimal Transport (OT) and Extreme Value Theory (EVT) are two pillars of modern statistical analysis, yet they have historically operated in different mathematical realms. Optimal transport provides a framework for finding the most efficient way to transform one probability distribution into another, while Extreme Value Theory focuses on the behavior of distributions in their extreme "tail" regions. This paper bridges the gap between these fields by investigating how optimal transport plans behave when applied to heavy-tailed, regularly varying distributions.
By extending classical results in OT to a class of infinite measures, the authors demonstrate that the transport map between two heavy-tailed distributions converges to a well-defined limit. This limit possesses specific scaling properties, offering a rigorous multivariate generalization of the "matching of quantiles" often used in one-dimensional statistics.
The Problem: Multivariate Quantiles and Infinite Measures
In univariate statistics, matching the tails of two distributions is a trivial task: one simply aligns their quantiles. However, in multivariate settings ($d > 1$), the concept of a "quantile" is not uniquely defined. Optimal transport offers a compelling alternative by defining a map that minimizes a specific cost (usually the squared Euclidean distance, or $L_2$ cost) to reshape one measure into another.
Applying OT to extreme values introduces a significant technical hurdle. EVT often involves "zooming in" on the tails of a distribution, a process that results in an intensity measure. These intensity measures are often infinite—they have infinite total mass because they are concentrated near the origin or at infinity—though they remain finite on sets bounded away from the origin. Standard OT theory, such as the landmark work by Robert McCann, is primarily designed for finite probability measures. The authors sought to determine if an optimal transport plan could be defined for these infinite measures and whether the transport maps of a sequence of probability measures would converge as they approach their heavy-tailed limits.
Methodological Framework
To address these questions, the paper develops a robust theoretical framework involving several key components:
1. Extending McCann’s Theorem
The researchers generalized the existence and uniqueness results for $L_2$ optimal transport to the space $\mathcal{M}_0(\mathbb{R}^d)$. This space consists of Borel measures that are finite on any set bounded away from the origin but may have infinite mass at the origin itself. This extension is crucial for handling the intensity measures that arise in the study of regularly varying distributions.
2. Graphical Convergence
Because transport maps can be multi-valued (specifically, they are the subdifferentials of convex functions), the authors utilized graphical convergence. Based on Painlevé–Kuratowski set convergence, this approach allows for a stable definition of convergence for the "graphs" of these maps. This ensures that as a sequence of measures converges, their corresponding transport maps also converge in a meaningful way.
3. Regular Variation and Tail Limits
The study applies these stability results to "regularly varying" measures—distributions whose tails decay according to power laws. By rescaling these measures, the authors prove the existence of a "tail limit" for the transport plan, showing that the process of taking a limit and the process of finding an optimal map are commutative.
Key Insights and Results
The Geometry of the Rare
The research reveals that mass movement in extreme regions is governed by a predictable geometry. If the source and target distributions are regularly varying, the optimal transport map inherits this regularity. This suggests that the "optimal" way to move rare events is not stochastic or arbitrary but follows a structured path determined by the tail indices of the distributions.
Homogeneity of the Limit Map
A central finding is the homogeneity of the limit transport map $\nabla \bar{\psi}$. The authors discovered that if the input is scaled by a factor $\lambda$, the output scales by $\lambda^{\alpha_1/\alpha_2}$, where $\alpha_1$ and $\alpha_2$ represent the tail indices of the source and target distributions, respectively. This provides the mathematical foundation for multivariate quantile matching.
Stability and Existence (Theorems 3.1, 4.2, & 5.1)
- Theorems 3.1 & 4.4: Establish that for measures in $\mathcal{M}_0(\mathbb{R}^d)$, there exists a unique (up to null sets) gradient of a convex function that pushes one measure to the other.
- Theorem 4.2: Confirms that if a sequence of measures converges in the $\mathcal{M}_0$ topology, their optimal coupling measures and transport maps also converge.
- Theorem 5.1: Demonstrates that for regularly varying distributions, the rescaled transport plan converges to an "operator regularly varying" limit measure $\bar{\pi}$.
Limitations and Future Outlook
While the paper establishes a rigorous foundation, it focuses exclusively on the quadratic ($L_2$) cost function. Future research is needed to determine if these tail properties hold for other costs, such as $L_1$ or general $L_p$ metrics. Additionally, the uniqueness of the transport map requires the source measure to be "well-behaved" (not charging sets of Hausdorff dimension $d-1$), which excludes highly singular or discrete distributions.
The implications for applied fields are significant. These "tail transport maps" could eventually be used to calibrate multivariate extreme value models, improve risk management in finance and insurance, and provide a standardized way to define multivariate quantiles in high-dimensional data analysis.
Summary updated