Big Jump Principle in Heavy-Tailed Processes

• Big Jump Principle is a concept that defines rare large deviations in heavy-tailed distributions as typically caused by one exceptionally large event.
• It applies across diverse settings—from IID sums to correlated processes and renewal models—using rigorous probabilistic methods and rate approaches.
• This principle challenges classical large deviation theory and informs practical insights in risk management, statistical physics, and network analysis.

The Big Jump Principle (BJP) is a central concept in the theory of rare events in stochastic processes with heavy-tailed or subexponential distributions. It asserts that, in such systems, the realization of a large deviation—such as a sum, displacement, or aggregate observable taking an atypically large value—is typically achieved by a single, exceptionally large event (“big jump”) rather than by the coherent accumulation of many moderate-sized fluctuations. This principle has far-reaching implications in mathematics, physics, random graph theory, and complex systems, and admits rigorous generalization to a variety of correlated, structured, or disordered settings.

1. Mathematical Formulation of the Big Jump Principle

Let $\{X_i\}_{i=1}^n$ be i.i.d. nonnegative random variables with common distribution $F$, and tail $\overline{F}(x) = 1 - F(x)$. The classic subexponentiality condition reads: $$\lim_{x\to\infty} \frac{\overline{F^{*2}}(x)}{\overline{F}(x)} = 2,$$ where $F^{*2}$ is the convolution of $F$ with itself. The Big Jump Principle states: $$\mathbb{P}\left(\sum_{i=1}^n X_i > x\right) \sim n\,\mathbb{P}(X_1 > x), \qquad x\to\infty,$$ and, more strongly,

$$\lim_{K\to\infty} \liminf_{x\to\infty} \mathbb{P}\left(\max_{1\leq i \leq n} X_i > x - K\,\Big|\,\sum_{i=1}^n X_i > x\right) = 1.$$

That is, conditional on the sum being large, exactly one summand is anomalously large, while all others are negligible at that scale (Xu et al., 2014, Xu et al., 2014, Vezzani et al., 2018).

This principle extends to varied heavy-tailed classes—including subexponential, regularly varying, and certain light-tailed distributions—and survives under weak dependencies and structural deformations.

2. Extensions Beyond IID Sums

2.1 Structured and Correlated Processes

The insights of the BJP persist in more complex settings:

• Renewal and pinning models: In generalized pinning models, the partial localization regime is a direct manifestation of the big-jump regime, wherein the rare event is realized by a single macroscopic excursion (“big gap” in renewal points) (Giacomin et al., 2020).
• Correlated increments: For random walks with kernel-induced dependence (e.g., autoregressive or Ornstein-Uhlenbeck), the big jump propagates through kernels, leading to a renormalized (memory-amplified) big jump, but the tail is still controlled by the largest “effective” increment, possibly clustered over a block of steps (Höll et al., 2021).
• Tree-indexed and branching processes: For random walks indexed by trees, the maximum displacement is asymptotically determined by the largest single increment, provided the tree height is sufficiently small relative to its size; similar criteria govern large deviations in branching random walks and Galton–Watson trees (Maillard, 2015, Stonner, 17 Mar 2026, Zhao, 6 Sep 2025).

2.2 Physical Stochastic Processes

• Continuous-time random walks and Lévy walks: The far tails of position (displacement) distributions are controlled by one long trapping time or a single long ballistic flight, as shown in disordered transport, glassy systems, and contaminant migration models (Wang et al., 2019, Burioni et al., 2019, Bassanoni et al., 2024).
• Generalized Lévy walks: In processes where displacement within a step follows a non-universal microscopic law, the rare-event tail encodes explicit dependence on this law and is always determined by the “single big step” (Vezzani et al., 2019).
• Continuous-path settings: In time-integrated observables of processes such as the Ornstein–Uhlenbeck velocity process, the BJP characterizes the anomalous, subexponential decay of the large-deviation rate function: the tail is generated by a single excursion rather than accumulation over time (Bassanoni et al., 13 Jan 2025).

3. Generalizations: Few-Big-Jumps and High Dimensions

Moving beyond one-dimensional settings, for large deviations of sums in high dimensions or under finite cutoffs, one encounters the “fewest-big-jumps principle.” In such regimes, the optimal strategy for exceeding a large threshold is to realize the excess using the smallest number of summands possible, constrained by the system's geometry (Tuchel et al., 1 Feb 2026, Kerriou et al., 2022).

Key results:

• High-dimensional sums: For stretched-exponential or Weibull-tailed random vectors in $\mathbb{R}^k$, the large deviation of the sum into a remote orthant is realized by at most $k$ entries having large values—the “few-big-jumps” principle (Tuchel et al., 1 Feb 2026).
• Random graphs and condensation: In graphs with heavy-tailed degree distributions and bounded support (e.g., geometric random graphs with radii capped at system size $F$0), rare excesses in total out-degree are concentrated in a finite set of vertices (exact number determined by the excess), with remaining vertices obeying the law of large numbers (Kerriou et al., 2022).

4. Analytical Frameworks and Rate Methods

The BJP admits a unified rate-formalism for continuous- or discrete-time processes: $F$1 where $F$2 is the rate of new “attempts” at a big jump at time $F$3, $F$4 is the distribution for jump or step magnitudes, and $F$5 encapsulates kinematic or dynamical constraints. This approach is used to derive rare-event tails for:

• Biased CTRWs (superdiffusive tails, infinite densities distinguishing between ordinary and stationary cases) (Wang et al., 2019).
• Lévy walks and generalized models (explicit scaling functions $F$6 with non-analytic features tied to dynamics) (Vezzani et al., 2019).
• Quenched-disorder models and memory effects (Lévy–Lorentz gas, walks with reflection/memory) (Bassanoni et al., 2024, Vezzani et al., 2018).

5. Structural and Class Properties

The Big Jump Principle characterizes a class $F$7 of distributions with the one-big-jump property (also called the PSBJ class): $F$8 where $F$9 is the largest among $\overline{F}(x) = 1 - F(x)$0 i.i.d. variables given their sum exceeds $\overline{F}(x) = 1 - F(x)$1 (Xu et al., 2014, Xu et al., 2014). This class:

• Contains all subexponential and convolution-equivalent distributions, but is strictly larger (including, via Esscher transforms, light-tailed laws that fail convolution-equivalence).
• Is closed under weak tail-equivalence, but not under certain transformations.
• Exhibits a rich structure, intersecting but strictly extending classic heavy-tailed classes S, L, D, and OS.

6. Physical, Mathematical, and Applied Implications

• Superdiffusion and anomalous kinetics: In many physical systems (glassy effects, porous media) actual transport or displacement is fundamentally determined by rare single-trap or single-excursion events, not by many small increments (Wang et al., 2019).
• Condensation and localization: Pinning models and random graphs exhibit condensation phases or partial localization where a macroscopic portion of mass or degree concentrates in one or few sites—the big jump regime (Giacomin et al., 2020, Kerriou et al., 2022).
• Statistical diagnostics and risk: Empirical identification of the big jump regime (e.g., comparing $\overline{F}(x) = 1 - F(x)$2 and $\overline{F}(x) = 1 - F(x)$3) is essential for risk management, as risk tails are controlled by individual extremes, not by moderate fluctuations (Vezzani et al., 2018).
• Breakdown of classical large deviations: In subexponential cases, classical Cramér-type large deviation theory fails; the BJP and its rate formalism provide the correct asymptotics. For faster-than-exponential tails, multiple moderate sums regain dominance.

7. Phase Transitions, Corrections, and Open Directions

• Transitions and refinements: The BJP underlies condensation phenomena, dynamical phase transitions (e.g., in pinning and Ornstein-Uhlenbeck observables), and admits a perturbative correction theory beyond leading order, bridging Gaussian and big-jump regimes (Bassanoni et al., 2 Mar 2026).
• Boundary and cluster expansions: At critical tail exponents (boundary index one), refined asymptotics and countable sum closures clarify the behavior at phase frontiers (Zhao, 6 Sep 2025).
• Open questions: General characterization and extension of the BJP in dependent, nonstationary, multiscale, or multivariate environments remain active areas of research, with implications for both probability theory and physics.

Representative Table: Key Manifestations of the Big Jump Principle

Domain	Typical Observable/Tail	Mechanism
IID Sums	$\overline{F}(x) = 1 - F(x)$4	Single summand $\overline{F}(x) = 1 - F(x)$5
Pinning/renewal	Longest gap in path	Single macroscopic excursion
CTRW/Lévy walk	Far tail of displacement	Single long wait or flight
Correlated walk	Weighted sum tail	Clustered, renormalized big jump
High-dimensional sum	Large deviation in $\overline{F}(x) = 1 - F(x)$6	At most $\overline{F}(x) = 1 - F(x)$7 big jumps
Random graphs	Unusually large out-degree	Few vertices condense excess

The universality and robustness of the Big Jump Principle position it as a foundational paradigm for understanding the emergence of rare, extreme events in systems with heavy-tailed statistics, with broad applicability from statistical physics to random graph theory and network science (Xu et al., 2014, Xu et al., 2014, Vezzani et al., 2018, Giacomin et al., 2020, Wang et al., 2019, Tuchel et al., 1 Feb 2026, Kerriou et al., 2022, Bassanoni et al., 2024, Stonner, 17 Mar 2026, Zhao, 6 Sep 2025, Bassanoni et al., 2 Mar 2026, Bassanoni et al., 13 Jan 2025).