Large Deviation Theory (LDT) is the mathematical toolkit used to study "rare events"—those one-in-a-million occurrences like market crashes or massive network failures. Traditionally, LDT relies on "light-tailed" distributions where extreme values are exponentially unlikely. This paper breaks new ground by providing a rigorous framework for processes with Weibull-type increments, a class of "heavy-tailed" distributions where rare events happen not through a series of small unfortunate steps, but through a few massive jumps.
TL;DR
The authors establish a Large Deviation Principle (LDP) for Lévy processes and random walks with Weibull-type jump distributions. They prove that while the standard $J_1$ topology is too restrictive to support a full LDP, a more flexible topology ($M_1'$) allows for a complete description of how these rare sample paths behave.
What Problem Does This Paper Address?
In classical probability, rare events for sums of random variables usually follow the "Principle of Small Deviations": if you end up far from the mean, it’s likely because every step was slightly biased. However, for heavy-tailed distributions, the "Principle of One Big Jump" takes over—rare events are usually caused by a single, catastrophic outlier.
Weibull distributions (with tail probabilities like $e^{-x^\alpha}$) sit in a tricky middle ground. When the shape parameter $\alpha < 1$, they are heavy-tailed, but they don't always follow the "one big jump" rule. If $\alpha$ is close to 1, a rare event might be caused by several large jumps.
The mathematical challenge is that standard LDT tools (like Cramér’s Theorem) require the existence of a moment-generating function, which heavy tails lack. Previous research was often limited to one-dimensional distributions or specific, less useful path spaces. This paper seeks a sample-path LDP, which describes the entire trajectory of the process, not just its final value.
Methods
The authors employ a "divide and conquer" strategy to handle the complexity of Lévy processes:
- Decomposition: They split the process into two parts: the $k$ largest jumps and a "remainder" term containing all smaller fluctuations.
- Poisson Representation: They use Poisson random measures to represent the jumps, allowing them to treat the largest jumps as a step function.
- Bryc’s Inverse Varadhan Lemma: This tool is used to establish the LDP for the $k$ largest jumps.
- Concentration Bounds: To ensure the "remainder" term doesn't ruin the result, they adapt specialized concentration inequalities (specifically from Jelenković and Momčilović) to prove the remainder is "exponentially negligible."
- Topological Analysis: A significant portion of the work involves the Skorokhod $J_1$ and $M_1'$ topologies. The $J_1$ topology is the standard for paths with jumps, but it is very "stiff"—it doesn't like it when jump times are slightly different. The $M_1'$ topology is more "forgiving," allowing jumps that occur very close together to merge in the limit.
Datasets
Not discussed in the paper. This is a theoretical probability paper; the results are proven mathematically rather than evaluated on empirical data.
Key Insights and Ideas
- The $J_1$ Failure: The most striking insight is that a full LDP is impossible in the standard $J_1$ topology. The authors provide a clever counterexample: a set of paths where jumps are forced to happen at specific times. Because $J_1$ cannot "merge" jumps that occur at nearly the same time, the mathematical bounds for the LDP break down.
- The $M_1'$ Solution: By moving to the $M_1'$ topology, the authors allow for "jump merging." This matches the physical intuition that a rare event might be caused by several large jumps occurring in rapid succession.
- The Rate Function: The resulting rate function $I(\xi)$ is essentially the sum of the $\alpha$-powers of the jump sizes: $\sum (\Delta \xi)^\alpha$. This looks like an $L^\alpha$ "norm," reflecting that the "cost" of a rare path is the sum of the costs of its individual big jumps.
- Quasi-Variational Representation: The rate function can be viewed as the solution to a minimization problem, which is useful for calculating the "most likely" path to a rare event.
Key Results
- Extended LDP in $J_1$: They prove an "extended" version of the LDP for both Lévy processes and random walks in the $J_1$ topology, which provides upper and lower bounds even if a full LDP doesn't exist.
- Full LDP in $M_1'$: They successfully establish a full LDP in the $M_1'$ topology, providing a complete characterization of sample-path large deviations.
- Ruin Probability Application: The authors apply their theory to an insurance ruin problem where large claims are reinsured (capped). They show how to calculate the probability that a company goes bankrupt due to a specific combination of "moderate-sized" large claims.
Limitations and Future Work
- Topology Constraints: The non-existence of the LDP in $J_1$ is a fundamental limitation of the topology itself, not the authors' method. Users must be comfortable working in the more complex $M_1'$ space.
- Non-Negative Jumps: While the main theorems focus on processes with positive jumps, the authors suggest extensions to two-sided processes are possible but more technically involved.
- Future Directions: The authors mention that these results can be extended to Markov random walks, where the jump distribution depends on an underlying state. This would have significant implications for analyzing more complex systems like queueing networks or financial models with regime switching.
Summary updated