Rare Event Simulation via the Single Big Jump Principle

23 concepts13 papers2 proposals

In stochastic systems, rare events—such as an insurance company going bankrupt, a telecommunications network dropping packets, or a financial portfolio experiencing catastrophic loss—are notoriously difficult to simulate. If we rely on standard Monte Carlo methods, we might have to simulate billions of trajectories before witnessing a single failure, making risk estimation computationally intractable.

To solve this, researchers use importance sampling, a technique that changes the underlying probability distribution to make the rare event happen more frequently, and then corrects the bias using a likelihood ratio. However, how we change the distribution depends entirely on the nature of the system's randomness.

For decades, the gold standard for rare-event simulation was designed for "light-tailed" systems (like those governed by Gaussian or exponential distributions). In a light-tailed world, a rare event occurs through a conspiracy of small shocks—a long sequence of slightly larger-than-average increments that steadily accumulate over time. To simulate this efficiently, you shift the mean of every single step (a technique known as exponential twisting).

But when a system is heavy-tailed (governed by subexponential or regularly varying distributions, like Pareto or log-normal distributions), this logic completely breaks down. In a heavy-tailed world, rare events happen due to the single big jump principle.

The Intuition: Conspiracy vs. Catastrophe

Imagine a random walk trying to cross a high threshold.

  • Light-tailed scenario: The most likely path to the threshold is a steady, upward climb where every step contributes a small fraction of the total distance.
  • Heavy-tailed scenario: The most likely path is for the system to drift normally around its expected value, and then suddenly experience a single, massive, catastrophic outlier that clears the threshold in one go.

If you apply light-tailed importance sampling (shifting every step) to a heavy-tailed process, you force the simulation to look like a steady climb. But in a heavy-tailed universe, a steady climb is exponentially less likely than a single giant leap. As a result, the sampler explores the wrong region of the path space, causing the likelihood ratios to fluctuate wildly and leading to unstable estimators with massive or even infinite variance. As noted by Lam, efficient simulation in these systems requires an entirely different design principle: the sampling distribution must be biased to produce extreme values in only a small number of components, rather than a collective deviation across all components.

This single big jump phenomenon is not just a heuristic; it is a fundamental physical property of heavy-tailed systems. For example, in communication networks, Lu et al. demonstrated in Asymptotic Blocking Probabilities in Loss Networks with Subexponential Demands that network congestion under heavy-tailed demands is almost entirely driven by the arrival of a single request that exceeds available capacity, rather than the cumulative effect of many average-sized requests. Even in continuous-time settings like Langevin diffusions, Bazhba et al. showed in Large deviations asymptotics for unbounded additive functionals of diffusion processes that when a system's growth outpaces its restorative drift, it undergoes a sharp transition: rare empirical averages cease to be driven by steady deviations and are instead triggered by a single massive excursion.

Designing Single-Jump Samplers

To achieve strong efficiency—meaning the computational time required to achieve a target precision does not grow as the event becomes rarer—an importance sampling algorithm must mimic the true physical mechanism of the tail. It must force exactly one big jump.

Mixture-Based Sampling

One elegant way to implement this is through mixture models. Instead of altering every step, the algorithm samples increments from a mixture distribution. With high probability, it samples from the nominal (original) distribution, keeping the system's behavior ordinary. But with a small, carefully calibrated probability, it samples from a heavy-tailed distribution specifically biased to produce a massive outlier.

Liu pioneered this in State-dependent importance sampling for regularly varying random walks, using Lyapunov-type inequalities to prove that forcing single big jump trajectories guarantees bounded relative error. Similarly, Blanchet et al. combined this mixture-based approach with the cross-entropy method in Efficient rare event simulation for heavy-tailed systems via cross entropy, creating a flexible framework that automatically learns where and when to inject the large jump to accurately capture tail behavior.

State-Dependent vs. State-Independent Conditioning

How do we know when the jump should happen? Early breakthrough algorithms relied on state-dependent conditioning. In Efficient rare-event simulation for the maximum of heavy-tailed random walks, Glynn developed an algorithm that dynamically updates the sampling distribution at each step based on the current position of the random walk. By conditioning the remaining increments on the eventual occurrence of a single large jump, the estimator achieves asymptotically vanishing relative error—meaning the simulation actually becomes more precise the rarer the event gets.

While powerful, state-dependent algorithms can be mathematically complex to implement. To circumvent this, Murthy et al. introduced a state-independent approach in State-independent Importance Sampling for Random Walks with Regularly Varying Increments. The authors proved that one can achieve strong efficiency by mathematically decomposing the rare event probability into two distinct parts: a dominant component (where the event is triggered by a single massive outlier) and a residual component (consisting of smaller, nominal increments). By treating these components separately, the simulation can target the single-jump mechanism without needing to recalculate system states at every step.

Selective Twisting

Another perspective on this mechanism involves modifying the classical concept of hazard rate twisting. If a rare event is a function of a sum of heavy-tailed variables, modifying all variables introduces statistical noise that obscures the true driver of the event.

In An Improved Hazard Rate Twisting Approach for the Statistic of the Sum of Subexponential Variates (Extended Version), Kammoun et al. demonstrated that by selectively twisting only the most dominant components of the sum while leaving the non-dominant variables alone, one can achieve variance reduction of several orders of magnitude. This matches the strategy proposed by Rojas‐Nandayapa in Efficient simulation of tail probabilities of sums of dependent random variables, which shifts the simulation measure to focus exclusively on scenarios where at least one variable in a dependent system exceeds a high threshold, exploiting the fact that aggregate extremes are driven by isolated individual extremes.

Beyond a Single Dimension

As the field has matured, researchers have extended the single big jump principle to more complex, higher-dimensional environments.

In multidimensional risk and ruin problems, a single jump must not only be large, but it must also point in the right direction across a multi-variable space. Liu extended the single-jump framework to these complex settings in Efficient importance sampling in ruin problems for multidimensional regularly varying random walks and Total variation approximations and conditional limit theorems for multivariate regularly varying random walks conditioned on ruin. By mixing nominal multi-line business dynamics with a specific threshold-crossing jump vector, the author constructed tractable proxy processes that converge in total variation, mapping out how multidimensional systems behave on their way to ruin.

Finally, while a single catastrophic jump dominates classical subexponential tails, certain complex boundaries or compound processes can be triggered by a small number of discrete, moderately large jumps. To handle this, Chen et al. developed a universal importance sampling technique in Efficient Rare-Event Simulation for Multiple Jump Events in Regularly Varying Random Walks and Compound Poisson Processes. Their method constructs an auxiliary set that guides the simulation by conditioning on paths that experience a precise, small number of larger increments, extending the core intuition of the big jump principle to its logical limit.

Key Foundational Work

State-dependent importance sampling for regularly varying random walks (2008) — contributed 3 concept elements to this direction.
Efficient rare-event simulation for the maximum of heavy-tailed random walks (2008) — contributed 2 concept elements to this direction.
Asymptotic Blocking Probabilities in Loss Networks with Subexponential Demands (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_single_big_jump_principle_18af594b,
  title = {single big jump principle},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/single-big-jump-principle/rdr_18af594b93fb48cfac55d3e8e3d1a1ef},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_single_big_jump_principle_18af594b_v1,
  title = {single big jump principle},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/single-big-jump-principle/rdr_18af594b93fb48cfac55d3e8e3d1a1ef/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake