One Big Jump Principle

• The One Big Jump Principle is a heavy-tailed asymptotic concept where extreme outcomes are typically caused by one exceptional value.
• It shows that in subexponential regimes, the tail of a sum is asymptotically equivalent to the maximum summand, revealing a condensation phenomenon.
• The principle extends to various settings—including multivariate deviations, moving cutoffs, and correlated increments—demonstrating broad applicability in stochastic modeling.

The One Big Jump Principle is a heavy-tail asymptotic principle asserting that certain rare large values of sums, additive functionals, or extremal observables are typically realized by a single exceptional contribution rather than by the coherent accumulation of many moderate ones. In its classical one-dimensional form, for i.i.d. summands $X_1,\dots,X_n$ with a subexponential distribution and $S_n=\sum_{i=1}^n X_i$, the far-right tail satisfies

$$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$

and, at density level, $f_{S_n}(x)\sim n f(x)$. The interpretation is that the event $S_n\approx x$ is asymptotically produced by one summand of order $x$, with $n$ equivalent locations for that exceptional fluctuation (Bassanoni et al., 2 Mar 2026). Although this principle originated in the study of sums of independent heavy-tailed random variables, subsequent work has extended, modified, or contradicted it in settings involving stretched-exponential tails, multivariate deviations, moving cutoffs, correlated increments, branching structures, excursion decompositions, and constrained persistence events (Tuchel et al., 1 Feb 2026, Höll et al., 2021, Kerriou et al., 2022).

1. Classical formulation and probabilistic meaning

The standard formulation concerns sums of i.i.d. nonnegative or centered real-valued random variables in a heavy-tailed or subexponential regime. In the one-dimensional stretched-exponential setting, if $Y$ satisfies

$\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$

with $\alpha\in(0,1)$, then Nagaev-type results imply that for every $S_n=\sum_{i=1}^n X_i$0,

$S_n=\sum_{i=1}^n X_i$1

This is the classical one-big-jump principle: on the logarithmic scale, the deviation of the sum is caused by one exceptional summand (Tuchel et al., 1 Feb 2026).

A closely related asymptotic, standard in the subexponential literature, is

$S_n=\sum_{i=1}^n X_i$2

In this regime, the upper tail of the sum is asymptotically the same as the upper tail of the maximum. The mechanism is that simultaneous occurrence of two or more extreme summands is asymptotically negligible relative to the event that exactly one summand is extreme (Tuchel et al., 1 Feb 2026, Höll et al., 2021).

A precise abstract formulation of the “principle of a single big jump” also appears in the distributional class $S_n=\sum_{i=1}^n X_i$3. For i.i.d. nonnegative $S_n=\sum_{i=1}^n X_i$4, with $S_n=\sum_{i=1}^n X_i$5 and $S_n=\sum_{i=1}^n X_i$6 the largest order statistic, $S_n=\sum_{i=1}^n X_i$7 if for all $S_n=\sum_{i=1}^n X_i$8,

$S_n=\sum_{i=1}^n X_i$9

Equivalently, conditional on the rare event $$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$0, asymptotically one summand lies within bounded distance of the full threshold, while the remaining summands contribute only a bounded amount (Xu et al., 2014, Xu et al., 2014).

2. Tail classes and asymptotic regimes

The principle is most naturally associated with subexponential or heavy-tailed distributions, but the precise form depends strongly on the tail class. For symmetric stretched-exponential densities

$$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$1

the far tail of the $$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$2-fold convolution $$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$3 obeys

$$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$4

which is the density-level version of the principle. In this regime one summand forms a condensate, while the remaining $$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$5 variables produce only lower-order fluctuations (Bassanoni et al., 2 Mar 2026).

The same paper emphasizes that this description is strictly asymptotic and does not capture the crossover from the Gaussian core to the condensed tail. For $$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$6, one has the CLT regime

$$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$7

whereas for fixed $$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$8 and $$\mathbb P(S_n>x)\sim n\,\mathbb P(X>x),\qquad x\to\infty,$$9, the one-big-jump asymptotic applies. Between these regimes lies a moderate-deviation region in which the classical principle alone is insufficient (Bassanoni et al., 2 Mar 2026).

By contrast, for thin-tailed distributions the principle fails. In stretched-exponential Weibull laws with parameter $f_{S_n}(x)\sim n f(x)$0, the big-jump formulas underestimate rare-event probabilities by many orders of magnitude, and the correct asymptotics are of ordinary large-deviation type,

$f_{S_n}(x)\sim n f(x)$1

with a collective many-small-deviation mechanism rather than a single dominant event (Burioni et al., 2019).

The class $f_{S_n}(x)\sim n f(x)$2 shows that one-big-jump behavior is broader than classical subexponentiality but still constrained. Heavy-tailed distributions in $f_{S_n}(x)\sim n f(x)$3 are automatically strongly heavy-tailed, meaning

$f_{S_n}(x)\sim n f(x)$4

Thus $f_{S_n}(x)\sim n f(x)$5, so heavy-tailed single-big-jump laws cannot have arbitrarily irregular tails (Xu et al., 2014). At the same time, $f_{S_n}(x)\sim n f(x)$6 is strictly larger than $f_{S_n}(x)\sim n f(x)$7 and larger than the light-tailed convolution-equivalent classes $f_{S_n}(x)\sim n f(x)$8; there exist light-tailed members of $f_{S_n}(x)\sim n f(x)$9 not even weakly tail equivalent to any convolution-equivalent distribution (Xu et al., 2014).

3. Beyond leading order: perturbative and large-deviation refinements

A recent development replaces the leading statement $S_n\approx x$0 by a systematic asymptotic expansion around the condensed configuration. For stretched-exponential summands, the convolution integral can be rewritten by isolating one macroscopic jump and scaling the remaining $S_n\approx x$1 variables as $S_n\approx x$2, yielding an integral dominated by $S_n\approx x$3 nonanalytic cusps of a geometric function $S_n\approx x$4, each cusp representing one symmetry-related condensed configuration (Bassanoni et al., 2 Mar 2026).

For general fixed $S_n\approx x$5, this yields the asymptotic structure

$S_n\approx x$6

where

$S_n\approx x$7

The zeroth-order approximation is precisely the classical one-big-jump formula,

$S_n\approx x$8

and the next correction is controlled by

$S_n\approx x$9

so that

$x$0

Probabilistically, $x$1 is the condensate contribution, while $x$2 captures the collective fluctuation cloud of the non-condensed background (Bassanoni et al., 2 Mar 2026).

This perturbative construction yields the crossover scale

$x$3

with anomalous exponents

$x$4

The resulting approximate rate function,

$x$5

matches the large-$x$6 expansion of the exact large-deviation rate function

$x$7

which describes competition between a condensate of size $x$8 and a Gaussian background of size $x$9 (Bassanoni et al., 2 Mar 2026).

This line of work explicitly positions itself as complementary to Edgeworth theory: Edgeworth expansions correct the Gaussian core in large $n$0, whereas the perturbative big-jump expansion corrects the condensed tail in large $n$1 (Bassanoni et al., 2 Mar 2026).

4. Generalizations, deformations, and failures of the principle

Multivariate deviations: few big jumps

In higher dimensions, the one-dimensional principle does not survive unchanged. For i.i.d. centered vectors in $n$2 with stretched-exponential tails of rate $n$3, the large-deviation rate for coordinatewise exceedance is

$n$4

The geometric argument behind this variational problem shows that at most $n$5 nonzero summands are needed. The correct multivariate analogue is therefore a few-big-jumps principle, not a literal one-big-jump principle: a deviation in $n$6 is typically realized by at most $n$7 large vectors, with one jump remaining optimal only in special geometries (Tuchel et al., 1 Feb 2026).

The paper gives explicit counterexamples in $n$8 where the target $n$9 is more cheaply achieved by two large vectors, one supporting each coordinate, than by a single vector satisfying both constraints. In one dimension, the variational rate reduces to $Y$0, recovering the classical principle (Tuchel et al., 1 Feb 2026).

Moving cutoff: fewest big jumps

A different deformation arises when summands are truncated at the same scale as the deviation. For triangular arrays $Y$1 with $Y$2 and

$Y$3

the event $Y$4 cannot be produced by one summand once $Y$5, because each variable is bounded by $Y$6. The dominant mechanism becomes the fewest-big-jumps principle: if $Y$7, then exactly

$Y$8

macroscopic summands realize the excess (Kerriou et al., 2022).

The main local asymptotic is

$Y$9

and the conditional structure theorem states that, under this event, exactly $\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$0 summands are of order $\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$1, their sum produces the entire excess $\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$2, and the remaining $\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$3 variables still obey a law of large numbers (Kerriou et al., 2022). The classical one-big-jump case is recovered only for $\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$4.

Correlated increments: one jump plus memory trail

For random walks with correlated increments

$\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$5

with heavy-tailed IID innovations $\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$6, the rare event is still initiated by one large innovation, but correlation propagates its effect to later increments. The conditional principle becomes

$\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$7

while the unconditional version is

$\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$8

Thus the principle survives, but the observable extreme is the big jump together with its memory-generated aftereffects (Höll et al., 2021).

5. Physical and stochastic-process realizations

Continuous-time random walks and propagators

In CTRWs with stretched-exponential jump density $\log \PP(Y\ge t)\sim -c\,t^\alpha \qquad (t\to\infty),$9 and finite-mean waiting-time law $\alpha\in(0,1)$0, the propagator

$\alpha\in(0,1)$1

inherits a one-big-jump structure. At leading order,

$\alpha\in(0,1)$2

meaning the tail is the probability of one large jump times the mean number of opportunities to realize it. Incorporating the perturbative correction to $\alpha\in(0,1)$3 yields an expansion in powers of $\alpha\in(0,1)$4, with coefficients determined by moments $\alpha\in(0,1)$5 of the renewal count (Bassanoni et al., 2 Mar 2026).

This connects to an earlier physical “rate formalism” for rare displacements in transport processes. There the far-tail density $\alpha\in(0,1)$6 is written as

$\alpha\in(0,1)$7

where $\alpha\in(0,1)$8 is the rate of opportunities to realize a new microscopic event, $\alpha\in(0,1)$9 is its tail density, and $S_n=\sum_{i=1}^n X_i$00 is the kinematic propagator from one single event to the observed rare displacement (Burioni et al., 2019, Vezzani et al., 2018).

Lévy flights, Lévy walks, and Lévy-Lorentz gas

For Lévy flights with jump law

$S_n=\sum_{i=1}^n X_i$01

the far tail of the running maximum is

$S_n=\sum_{i=1}^n X_i$02

which is the direct extreme-value analogue of the one-big-jump principle (Bassanoni et al., 2024).

For Lévy walks with finite velocity $S_n=\sum_{i=1}^n X_i$03, the maximum tail becomes

$S_n=\sum_{i=1}^n X_i$04

with one term corresponding to a big flight still ongoing at time $S_n=\sum_{i=1}^n X_i$05 and the other to a big flight completed earlier. In infinite-density scaling form,

$S_n=\sum_{i=1}^n X_i$06

showing that finite velocity imposes a ballistic cutoff and splits the tail by path topology (Bassanoni et al., 2024, Vezzani et al., 2018).

For the Lévy-Lorentz gas, the big jump is a quenched large gap in the disorder landscape rather than an IID flight. The maximum still arises from one exceptional structural element, but topology-induced memory modifies both the effective number of opportunities and the post-jump dynamics. The resulting asymptotics depend on $S_n=\sum_{i=1}^n X_i$07 and differ sharply from those of Lévy walks because the walker can reflect repeatedly inside the same long gap (Bassanoni et al., 2024).

Continuous correlated observables and excursions

A major continuous-path realization appears in Ornstein–Uhlenbeck dynamics for the time-integrated observable

$S_n=\sum_{i=1}^n X_i$08

Using the excursion decomposition between zero crossings, one writes

$S_n=\sum_{i=1}^n X_i$09

where $S_n=\sum_{i=1}^n X_i$10 is the area under one excursion and $S_n=\sum_{i=1}^n X_i$11 its duration. For $S_n=\sum_{i=1}^n X_i$12, the single-excursion area density has stretched-exponential tail

$S_n=\sum_{i=1}^n X_i$13

with $S_n=\sum_{i=1}^n X_i$14, hence subexponential. Therefore

$S_n=\sum_{i=1}^n X_i$15

so a rare large value of the time-integrated observable is realized by one exceptional excursion area (Bassanoni et al., 13 Jan 2025). This gives a physical interpretation of the anomalous large-deviation scaling previously derived by instanton methods.

Trapping models and lagging plumes

In biased CTRWs with heavy-tailed waiting times

$S_n=\sum_{i=1}^n X_i$16

the one big jump is not a large spatial displacement but one exceptionally long trapping time. In the rare-event sector far behind the mean drift,

$S_n=\sum_{i=1}^n X_i$17

so the position deficit is asymptotically determined by the largest waiting time. The resulting rare tail of the propagator and the rare tail of the maximum trapping time share the same infinite-density form, and these rare sluggish trajectories control the anomalous growth of the mean square displacement (Wang et al., 2019).

6. Exceptions, competing principles, and boundary cases

The one-big-jump principle is not universal. Several papers make clear that the correct asymptotic mechanism can be qualitatively different depending on geometry, constraints, and recursion structure.

A sharp counterexample is the persistence problem for sample averages of a heavy-tailed random walk in a compact convex set $S_n=\sum_{i=1}^n X_i$18 with $S_n=\sum_{i=1}^n X_i$19. The event

$S_n=\sum_{i=1}^n X_i$20

is not realized by one catastrophic jump. Instead, it requires a sequence of large jumps whose number grows like

$S_n=\sum_{i=1}^n X_i$21

where $S_n=\sum_{i=1}^n X_i$22 is a geometric radial expansion factor of $S_n=\sum_{i=1}^n X_i$23. The survival probability then satisfies

$S_n=\sum_{i=1}^n X_i$24

and the authors describe this as a principle of infinitely many big jumps (Bhattacharya et al., 2019). The reason is that one jump cannot sustain the sample average in a set bounded away from the origin across all times up to $S_n=\sum_{i=1}^n X_i$25; repeated rejuvenation is necessary.

In branching settings, further variants arise. For tree-indexed random walks on a finite rooted tree, the maximum displacement can obey a one-big-jump principle provided the largest single edge increment dominates branchwise fluctuation scales. If $S_n=\sum_{i=1}^n X_i$26 is the number of non-root vertices, $S_n=\sum_{i=1}^n X_i$27 the height, and $S_n=\sum_{i=1}^n X_i$28 the edge law, the paper gives a sufficient criterion in terms of a geometric exponent $S_n=\sum_{i=1}^n X_i$29 and a distribution-dependent threshold $S_n=\sum_{i=1}^n X_i$30: if $S_n=\sum_{i=1}^n X_i$31 with $S_n=\sum_{i=1}^n X_i$32, then

$S_n=\sum_{i=1}^n X_i$33

and the tail of the maximal displacement is asymptotically the tail of the largest edge jump (Maillard, 2015).

A different branching analogue appears for weighted branching random walks. There, for

$S_n=\sum_{i=1}^n X_i$34

one has

$S_n=\sum_{i=1}^n X_i$35

in the heavy-tail Nagaev regime, with $S_n=\sum_{i=1}^n X_i$36 the additive martingale limit. The paper interprets this as a principle of one big jump along the ancestral lineage of those particles contributing to the upper tail mass (Stonner, 17 Mar 2026).

Finally, delicate boundary-index cases still exhibit a single-big-jump structure but require refined control. In a branching-process fixed-point equation

$S_n=\sum_{i=1}^n X_i$37

with

$S_n=\sum_{i=1}^n X_i$38

the stationary tail obeys

$S_n=\sum_{i=1}^n X_i$39

with only negligible logarithmic corrections of order $S_n=\sum_{i=1}^n X_i$40. The dominant rare event is still effectively one large immigration shock, geometrically amplified by branching, and the paper terms this a Boundary Principle of a Single Big Jump (Zhao, 6 Sep 2025).

7. Conceptual significance and recurring structures

Across these varied formulations, several structural motifs recur.

The first is condensation: a rare event localizes onto one degree of freedom, one excursion, one gap, one lineage increment, one waiting time, or one structural cluster. In stochastic-process language, the principle explains why far tails often lie outside the scope of Gaussian or Cramér-type large deviations (Bassanoni et al., 2 Mar 2026, Bassanoni et al., 13 Jan 2025).

The second is competition with collective fluctuations. In many models the rare event can be realized either democratically, by many moderate contributions, or condensately, by one exceptional contribution. The one-big-jump principle identifies the condensed side of that competition. The multivariate few-big-jumps principle, the moving-cutoff fewest-big-jumps principle, and the infinitely-many-big-jumps persistence mechanism are all ways this competition can be reshaped by geometry or constraints (Tuchel et al., 1 Feb 2026, Kerriou et al., 2022, Bhattacharya et al., 2019).

The third is nontrivial crossover structure. The leading asymptotic $S_n=\sum_{i=1}^n X_i$41 is often not enough in applications. The perturbative expansion “beyond the big jump” demonstrates that the condensed regime itself has an internal fluctuation theory, with explicit higher-order corrections and anomalous crossover exponents (Bassanoni et al., 2 Mar 2026).

The fourth is robustness beyond IID sums. Physical transport models, correlated random walks, excursion decompositions, branching recursions, and quenched media all admit one-big-jump descriptions once the correct microscopic unit of condensation is identified. That unit need not be a literal jump length; it may be an excursion area, a trapping interval, a gap in a random medium, or an ancestral displacement (Höll et al., 2021, Bassanoni et al., 13 Jan 2025, Wang et al., 2019, Bassanoni et al., 2024).

The fifth is fragility under smoothing mechanisms. Disorder in generalized pinning models destroys the flat constrained-free-energy branch on which the big-jump phase depends, thereby removing the partially localized phase altogether (Giacomin et al., 2020). This suggests that whether one-big-jump structure survives under perturbation depends sensitively on whether the perturbation preserves a zero-cost branch for underconstrained configurations.

In modern usage, therefore, the “One Big Jump Principle” refers less to a single rigid theorem than to a family of asymptotic mechanisms. The classical statement remains the canonical prototype, but current research treats it as one member of a broader taxonomy that includes systematic corrections, multivariate few-jump analogues, cutoff-constrained minimal-jump analogues, correlated and branching variants, boundary-case refinements, and explicit counterexamples where condensation requires more than one extreme contribution (Bassanoni et al., 2 Mar 2026, Tuchel et al., 1 Feb 2026, Kerriou et al., 2022, Höll et al., 2021, Bhattacharya et al., 2019).