Abstract
This research establishes a rigorous mathematical framework for understanding sample-path large deviations in heavy-tailed stochastic processes, specifically Lévy processes and random walks with regularly varying increments. While traditional heavy-tail theory often relies on the "Principle of a Single Big Jump," this work addresses scenarios where multiple large jumps are required for a rare event to occur. By utilizing the framework of $\mathbb{M}$-convergence on the Skorokhod space, the authors derive sharp asymptotic estimates and conditional limit theorems, while demonstrating that while a weak large deviation principle holds, a full large deviation principle does not.
Introduction
In the study of rare events within probability theory, Large Deviation Theory (LDT) typically characterizes how extreme fluctuations occur and quantifies their likelihood. For "light-tailed" processes, such as Brownian motion, rare events generally result from the accumulation of many small, coordinated shocks. Conversely, in "heavy-tailed" systems—prevalent in finance, insurance, and network traffic—the conventional wisdom is the "Principle of a Single Big Jump," which posits that a rare event is most likely caused by a single massive outlier.
However, many complex systems require more than one extreme event to reach a critical state. For instance, a process might need to exceed a threshold and maintain that level, or hit multiple distant targets sequentially. Previous research often handled these "multiple jump" scenarios through ad hoc methods. This paper provides a systematic, unified framework to estimate the probabilities of rare path-sets that require $j$ jumps to occur, moving beyond the single-jump paradigm.
Methodological Framework: $\mathbb{M}$-convergence
Classical LDT often fails in heavy-tailed settings because the probability mass can "leak" to infinity too easily, violating the requirement of exponential tightness necessary for a full Large Deviation Principle (LDP). To circumvent this, the authors employ $\mathbb{M}$-convergence, a modern refinement of the theory of regular variation designed for measures on metric spaces—specifically the Skorokhod space $D([0, 1], \mathbb{R})$ of càdlàg paths (right-continuous with left limits).
The methodology involves several key steps:
- Scaling: The analysis focuses on the scaled process $\bar{X}_n(t) = X(nt)/n$. As $n \to \infty$, the goal is to determine the decay rate of $P(\bar{X}_n \in A)$ for a given set of paths $A$.
- Decomposition: The Lévy process is decomposed into a "big jump" component and a "residual" component (comprising small jumps and Brownian motion). The authors prove that the residual component becomes negligible in the large-deviation limit.
- Continuous Mapping: By establishing $\mathbb{M}$-convergence for the jump processes, the authors utilize a mapping theorem to derive results for various path functionals, such as the maximum, the integral, or first-passage times.
- Two-Sided Analysis: The theory is extended to processes with both positive and negative jumps, introducing combinatorial complexity in identifying the most likely sequence of upward and downward movements.
Key Insights and Results
The "Jump Counter" Rate Function
In classical LDT, the "cost" of a rare event is typically represented by an integral of a rate function. In this heavy-tailed framework, the cost is essentially an integer $J(A)$, representing the minimum number of jumps required for the path to enter the set $A$. The probability of the event decays at a polynomial rate proportional to the tail of the jump distribution raised to the power of $J(A)$.
Sharp Asymptotics
The authors provide a precise limit theorem: $$\frac{P(\bar{X}_n \in A)}{(n \nu[n, \infty))^{J(A)}} \to C_J(A)$$ where $\nu$ is the Lévy measure and $C_J(A)$ is a constant related to the measure of the set of paths with exactly $J(A)$ jumps. This provides a significantly more detailed estimate than the logarithmic limits found in standard LDT.
Sample-Path Behavior and Conditional Limits
The research demonstrates that if a rare event $A$ occurs, the process is most likely to behave like a step function with exactly $J(A)$ jumps. Between these jumps, the scaled process remains essentially constant. The authors prove conditional limit theorems showing that, given $\bar{X}_n \in A$, the entire path converges in distribution to a specific limit process composed of $J(A)$ discrete jumps at specific times and with specific magnitudes.
Failure of the Full LDP
A significant theoretical finding is that while a weak LDP holds with a logarithmic speed, a full LDP does not. This distinction is crucial for researchers; it confirms that $\mathbb{M}$-convergence is the more natural tool for heavy tails, as it focuses on the tails of the distribution while ignoring the "origin" (small fluctuations) that prevents the satisfaction of classical compactness requirements.
Conclusion and Future Directions
This work bridges the gap between continuous-time Lévy processes and discrete-time random walks, showing that the same large-deviation logic applies to both under heavy-tailed increments. While the framework is robust, it does require the set $A$ to be "bounded away" from paths with fewer than $J(A)$ jumps. Future research may focus on cases where the boundary of $A$ is more complex, or extend these one-dimensional results to high-dimensional or infinite-dimensional heavy-tailed processes. Additionally, calculating $J(A)$ for complex sets remains a variational challenge, potentially linked to impulse control problems in stochastic optimization.
Summary updated