Abstract
This paper characterizes the large deviation behavior of multivariate Gaussian extremes on closed convex sets in $d$-dimensional Euclidean space. We prove that these rare events are governed by a unique "dominating point" located on the boundary of the target set. This point is identified as the optimizer of a specific convex quadratic programming problem, highlighting a "collusion" between the Gaussian covariance structure and the geometry of the convex set. Our results extend to polyhedral sets and Gaussian mixture models, offering new insights for rare-event probability estimation and stochastic optimization.
Introduction
In fields such as reliability engineering, finance, and machine learning, quantifying the probability of "worst-case scenarios" is paramount. Traditional Extreme Value Theory (EVT) typically focuses on the limiting distribution of the maximum of a sequence of random variables. However, EVT often fails to provide precise estimates for the probability of "extreme extremes"—rare events where the maximum falls into an atypical region far from the mean. This paper addresses this gap by analyzing the large deviations of multivariate Gaussian extremes on closed convex sets, where the interplay between variable correlation and set geometry becomes critical.
The Dominating Point and Quadratic Programming
The core finding of this research is the existence of a "dominating point" ($x^*$) that characterizes the asymptotic behavior of rare events. As the probability of the maximum falling into a rare set $C$ vanishes, the probability mass concentrates almost entirely around $x^*$, which lies on the boundary of $C$.
The dominating point is not necessarily the point in $C$ with the minimum Euclidean distance to the origin. Instead, it is the solution to the following constrained optimization problem: $$x^* = \arg \min_{x \in C} \langle x, \Sigma^{-1} x \rangle$$ where $\Sigma$ is the covariance matrix of the Gaussian vectors. This quadratic programming (QP) formulation reveals how the internal dependence structure of the Gaussian variables "colludes" with the geometry of the set $C$ to define the most likely path to a rare event.
Methodology: Restricted Large Deviation Principle
The authors establish a restricted Large Deviation Principle (rLDP) for the componentwise maximum $M_n$. This principle provides a rate function $J(x^*)$ that describes the exponential decay of the rare-event probability as the sample size $n$ increases and the set $C$ is scaled. The scaling regime employed is $O(\sqrt{\log n})$, which aligns with the natural tail behavior of Gaussian distributions.
To rigorously prove the lower bounds of these probabilities, the authors utilize sophisticated geometric constructions, including "truncated hoop-cone segments." These constructions demonstrate that the probability mass does not "leak" away from the dominating point in the limit. A significant theoretical insight is the asymptotic equivalence between the probability of the componentwise maximum entering a set and the probability of at least one individual sample entering that set, which simplifies the treatment of complex multivariate extremes.
Extensions and Applications
The theoretical framework is specialized for polyhedral sets—regions defined by linear inequalities—which are common in engineering and optimization. In these cases, the dominating point is even more readily characterized. Furthermore, the authors extend the main result to Gaussian Mixture Models (GMMs). They find that the rare-event behavior of a mixture is dominated by the single component that can most easily "reach" the target set, a manifestation of the "principle of the largest term."
The identification of $x^*$ has significant practical implications for rare-event simulation. In Monte Carlo methods, the dominating point can be used to center a new importance sampling distribution. By shifting the sampling measure toward $x^*$, the rare event occurs much more frequently in simulations, drastically reducing the variance of the estimates and enabling efficient computation in high-dimensional spaces.
Conclusion
By quantifying the large deviations of Gaussian extremes through dominating points and quadratic programming, this work provides a robust framework for analyzing rare events in correlated systems. The results bridge the gap between geometry and probability, offering powerful tools for stochastic optimization and reliability analysis.
Summary updated