--- title: Big Jump Phenomenon in Heavy-Tailed Systems url: https://www.emergentmind.com/topics/big-jump-phenomenon type: topic --- # Big Jump Phenomenon in Heavy-Tailed Systems The big jump phenomenon is a family of asymptotic principles asserting that a rare large fluctuation is realized predominantly by one exceptional increment, one exceptional excursion, or, in truncated settings, by the minimal number of exceptional increments compatible with the constraint. In its classical form for sums of i.i.d. heavy-tailed random variables, it appears as max–sum equivalence or as asymptotics of the form \( \mathbb{P}(S_n>x)\sim n\,\mathbb{P}(X>x) \); in modern work it has been extended to branching random walks, tree-indexed walks, correlated and multivariate sums, continuous-path models analyzed through excursions, and finite-speed jump processes. A different but related usage occurs in non-local spectral theory, where “big jumps” denote the large-distance part of a jump kernel and are quantified by a jump-rate functional rather than by a one-summand domination principle [2303.12505] [1804.02932] [2503.22899]. ## 1. Classical formulation for sums, maxima, and large deviations For i.i.d. non-negative random variables \(X_i\) with partial sums \(S_n=X_1+\cdots+X_n\), Beck–Blath–Scheutzow’s class \(\mathcal{J}\) formalizes the principle of a single big jump by requiring that, for all \(n\ge2\), \[ \lim_{K\to\infty}\liminf_{x\to\infty} \mathbb{P}\!\left(X_{n,1}>x-K \,\middle|\, \sum_{i=1}^n X_i>x\right)=1, \] equivalently, the second largest summand remains tight under conditioning on a large sum [1406.2754]. In classical subexponential theory, this corresponds to the statement that large deviations of the sum are caused by one summand carrying essentially all of the excess. For sums in the domain of attraction of an \(\alpha\)-stable law with \(\alpha\in(0,2)\), the one-big-jump scenario is universal at genuine large-deviation scales. If \(x_n/a_n\to\infty\), then \[ \mathbb{P}(S_n-b_n\ge x_n)\sim n\,\mathbb{P}(\xi>x_n), \] and, conditionally on the event \(S_n-b_n\ge x_n\), the remaining \(n-1\) variables after removing the largest one are asymptotically i.i.d. with the original law [2303.12505]. This is the standard heavy-tailed regime in which collective Gaussian behavior is asymptotically negligible. The borderline case \(\alpha=2\), including infinite-variance laws in the normal domain of attraction, is structurally different. There the large deviation may be realized either by a collective Gaussian mechanism or by one big jump, and the paper identifies a transition scale of order \(a_n\sqrt{|\log q(a_n)|}\), where \(q(x)=x^2\overline F(x)/\sigma^2(x)\) [2303.12505]. For integral large deviations, the asymptotic probability is governed by the competition between a Gaussian tail term and \(n\overline F(x_n)\); for local large deviations, the competition is between a Gaussian density term and \(n\mathbb{P}(\xi=x_n)\), and the threshold is slightly larger than in the integral case [2303.12505]. This distinction makes precise that “one big jump” is not universal even within heavy-tail asymptotics; it depends on both the deviation scale and the observable. From the perspective of extremes, the same idea underlies large-threshold asymptotics for running maxima of jump processes. For stochastic symmetric jump processes with power-law jump tails, the large-\(X\) tail of the maximum \(M_T\) is asymptotically the probability that at least one jump exceeds \(X\), with the model dependence entering through an effective number of jump attempts \(N_{\rm eff}(T)\) [2404.19406]. In this sense, the classical one-big-jump principle is both a sum principle and an extreme-value principle. ## 2. Distribution classes and structural formulations The class \(\mathcal{J}\) was introduced precisely to capture the conditional structure of single-big-jump events. It sits strictly beyond classical subexponentiality. The inclusion relations recalled in the literature are \[ \mathcal{S}\cup\mathcal{D}\subset \mathcal{J}\subset \mathcal{OS}, \qquad \mathcal{S}=\mathcal{J}\cap\mathcal{L}, \] so every subexponential or dominatedly varying distribution belongs to \(\mathcal{J}\), but \(\mathcal{J}\) is larger than either class [1406.2754]. The heavy-tailed part of \(\mathcal{J}\) is nonetheless constrained: if \(F\in\mathcal{J}\cap\mathcal{K}\), then \(F\) is strongly heavy-tailed, in the sense that for every \(\lambda>0\), \[ e^{\lambda x}\,\overline F(x)\to\infty . \] This excludes tails that are heavy only in an integrated sense but too thin on large subsequences [1406.2754]. A central structural result is that \((\mathcal{J}\cap\mathcal{K})\setminus(\mathcal{L}\cup\mathcal{D})\) is non-empty, and that this region contains distributions both weakly tail equivalent to subexponential laws and not weakly tail equivalent to any subexponential law [1406.2754]. Thus the big jump phenomenon does not reduce to long-tailedness, dominated variation, or subexponentiality. It is a conditional asymptotic property of sums, not merely a regularity property of the tail. The light-tailed side is equally nontrivial. The transformation \[ G(x)=\mathbf{1}(x<0)+e^{-\gamma x}F(x)\mathbf{1}(x\ge0) \] maps a heavy-tailed law \(F\) to a light-tailed law \(G\) and preserves enough asymptotic structure to transport single-big-jump behavior [1411.1625]. Using this transform, one can construct light-tailed distributions \(G\in\mathcal{J}\) that do not belong to any convolution-equivalent class \(\mathcal{S}(\beta)\), and even laws in \(\mathcal{J}\cap\mathcal{K}^c\) that are not weakly tail equivalent to any convolution-equivalent distribution [1411.1625]. This gives a negative answer to the conjecture that every light-tailed distribution in \(\mathcal{J}\) must be convolution equivalent. A common misconception is therefore false: the big jump principle does not coincide with convolution equivalence on the light-tailed side. Taken together, these results suggest that the big jump phenomenon is best viewed as a structural property of conditioned rare events rather than as a synonym for any single tail class. ## 3. Branching and tree-indexed systems In branching systems, the one-big-jump mechanism is lifted from a single path to a random collection of genealogical paths. For a branching random walk on \(\mathbb{R}\), with random measure \[ Z_n=\sum_{|u|=n} e^{-V_u}\delta_{V_u}, \] Biggins’ martingale \(W_n=Z_n(\mathbb{R})\) converges to a non-degenerate limit \(W\) under assumptions (A1)–(A3), and the associated random walk \(S_n\) is defined by the many-to-one identity \[ \mathbb{E}[Z_n(B)]=\mathbb{P}(S_n\in B). \] If the increment law \(F(dx)=\mathbb{P}(S_1\in dx)\) has regularly varying right tail \(F(t,\infty)\sim t^{-p}\ell(t)\) with \(p>2\), then for \(t_n\ge a\sigma\sqrt{n\log n}\) with \(a>\sqrt{p-2}\), \[ \frac{Z_n(nc+t_n)}{n\mathbb{P}(S_1>t_n)} \xrightarrow{L^1} W, \] and in a stronger regime the convergence holds in \(L^q\) for every \(q<2-1/\gamma\) [2603.16316]. This is an exact branching-random-walk analogue of Nagaev’s asymptotics for heavy-tailed sums. The probabilistic content is explicit: the weighted tail mass \(Z_n((nc+t_n,\infty))\) is asymptotically generated by lineages containing exactly one large displacement, while contributions from lineages with no large jump or at least two large jumps are negligible in expectation [2603.16316]. The random factor \(W\) records the random effective population size under the tilted weights \(e^{-V_u}\); the large deviation scale itself remains the same \(n\mathbb{P}(S_1>t_n)\) as for a single path. A closely related phenomenon appears for tree-indexed random walks. If \(T\) is a finite rooted tree, \(S_v\) is the sum of i.i.d. increments along the root-to-\(v\) path, \(M^S=\max_v S_v\), and \(M^X\) is the maximum increment over the edges, then under regularly varying increment tails and a geometric condition \(D>(X)\) relating tree height \(H\) and size \(V\), \[ \frac{M_n^S}{M_n^X}\to 1 \quad\text{in probability}, \] and similarly for leaf maxima and absolute maxima [1505.04949]. Here \(D\) is a dimension parameter controlling \(H\lesssim V^{1/D}\), while \((X)\) encodes whether path sums are governed by Gaussian or heavy-tail scaling. For conditioned critical Galton–Watson trees in the domain of attraction of an \(\alpha_T\)-stable law, \(D=\alpha_T/(\alpha_T-1)\), and the same criterion yields big-jump dominance of the maximal displacement [1505.04949]. The branching and tree-indexed results show that the phenomenon survives in random geometries with many competing paths. What changes is not the single-jump mechanism itself, but the normalization: martingale limits, tree height, and size enter as multiplicative or geometric modifiers. ## 4. Correlated, truncated, and multivariate extensions The one-big-jump principle is not confined to independent scalar sums. For heavy-tailed triangular arrays with truncation at scale \(n\), the relevant principle becomes the fewest-big-jumps principle. If \(W_1^{(n)},\dots,W_n^{(n)}\) are i.i.d. nonnegative variables truncated near \(n\), and one conditions the sum \(S_n\) to exceed its law-of-large-numbers value by \(\rho n\), then the minimal number of macroscopic summands needed is \(k\), where \(k-1<\rho1\), the rare event is realized by the minimal number of jumps permitted by the moving cut-off. For heavy-tailed random walks with correlated increments, the mechanism persists but the effect of the big jump propagates through the memory kernel. If \[ \tilde\delta_i=\sum_{j=1}^i M_{i-j}\delta_j, \qquad \tilde x_N=\sum_{i=1}^N \tilde\delta_i = \sum_{k=1}^N W_{N-k}\delta_k, \] with heavy-tailed i.i.d. noise \(\delta_k\), then the maximum correlated increment has the same tail as the maximum of the i.i.d. noise, but the sum tail is renormalized: \[ \mathbb{P}(\tilde x_N>z)\sim \frac{\tilde\gamma_N}{N}\,\mathbb{P}(\tilde\delta_{\max}>z), \] and, conditional on the big jump occurring at time \(b\), \[ \mathbb{P}(\tilde x_N>z\mid b)\sim (W_{N-b})^\alpha \mathbb{P}(\tilde\delta_{\max}>z\mid b) . \] For the exponential kernel \(M_{i-j}=m^{\,i-j}\), corresponding to an AR(1) model or discretized Ornstein–Uhlenbeck process with heavy-tailed noise, the amplification factor saturates at \(1/(1-m)\) [2106.14222]. The big jump therefore determines not only the large value but also the subsequent correlated tail geometry. The multivariate setting replaces scalar thresholds by rare sets \(xA\subset\mathbb{R}_+^d\). Following Samorodnitsky and Sun, one defines \[ Y_A:=\sup\{u>0:X\in uA\}, \qquad \overline F_A(x)=\mathbb{P}(X\in xA), \] for \(A\) in the class \(\mathcal R\) of open, increasing, convex sets avoiding the origin [2410.10292]. Multivariate analogues \(C_A,D_A,L_A,S_A\) are then defined by requiring the one-dimensional law of \(Y_A\) to be consistently varying, dominatedly varying, long-tailed, or subexponential. Under quasi asymptotic independence, tail asymptotic independence, or regression dependence, the finite-sum asymptotic takes the form \[ \mathbb{P}(S_n\in xA)\sim \sum_{i=1}^n \mathbb{P}(X^{(i)}\in xA), \] and for a random number \(N\) of i.i.d. vectors, \[ \mathbb{P}(S_N\in xA)\sim \mathbb{E}[N]\mathbb{P}(X\in xA) \] under suitable moment conditions [2410.10292]. This extends the one-big-jump principle from real sums to rare events in cones of \(\mathbb{R}_+^d\), and it shows that the correct multidimensional analogue is set-based rather than coordinatewise. ## 5. Continuous-path, finite-speed, and compact-state models In physical modeling, the big jump principle is often formulated as a rate-and-path approximation for the tail of a propagating process. For generalized Lévy walks with step durations \(t_i\) drawn from \[ \lambda(t)=\frac{\tau_0^\alpha}{t^{1+\alpha}}, \qquad t>\tau_0, \] and intra-step motion \[ r(T)=r(T_i)+c_i\,t_i^{\nu-\eta}(T-T_i)^\eta, \] the bulk density has a model-dependent scale \(\ell(T)\), but the rare-event tail \(B(R,T)\) for \(R\gg \ell(T)\) is computed from the big-jump formula \[ B(R,T)=\int_0^T dT_w \int_0^\infty dt\, p_{\rm tot}(t,T_w)\,{\cal P}(R|T,t,T_w), \] where \(p_{\rm tot}(t,T_w)=n_R(T_w)\lambda(t)\) is the rate to attempt a jump of duration \(t\) at time \(T_w\) [1908.10975]. The resulting tails are non-universal and non-analytic, depend on \(\alpha,\nu,\eta\), and the principle can fail when one single step cannot exceed the bulk scale [1908.10975]. An earlier formulation generalized the same principle to Lévy walks, laser cooling, scattering on heterogeneous structures, and Lévy walks with memory, emphasizing that the relevant rare-event scale is often ballistic or super-ballistic rather than the central-limit scale [1804.02932]. For extreme-value statistics of Lévy flights, Lévy walks, and the Lévy–Lorentz gas, the same logic yields asymptotics for maxima. In the big-jump regime, \[ \mathbb{P}(M_T>X)\sim \frac12 N_{\rm eff}(T)\,\mathbb{P}(\text{single jump}>X), \] with \(N_{\rm eff}(T)\) equal to the number of steps, the effective number of flights, or the effective number of large-gap crossings, depending on the model [2404.19406]. For the Lévy–Lorentz gas, the quenched disordered lattice induces memory effects, so the tail of the maximum differs from the tail of the position, because after entering a large gap the walker reflects inside that gap rather than simply moving away [2404.19406]. The big jump phenomenon can also be present in compact-state Markov processes. For the \(q\)-Ornstein–Uhlenbeck process with \(q\in(-1,1)\), the state space is \[ \Big[-\frac{2}{\sqrt{1-q}},\,\frac{2}{\sqrt{1-q}}\Big], \] and one studies jumps from an \(\epsilon\)-neighborhood of the left endpoint to an \(\epsilon\)-neighborhood of the right endpoint. If \(A^{(q)}(\epsilon)\) is the event that such a jump occurs in \((0,1]\), then \[ \mathbb{P}(A^{(q)}(\epsilon))\sim \alpha_q \epsilon^3, \] and the number of such jumps in the enlarged interval \((0,\epsilon^{-3}]\) converges in law: \[ N^{(q)}\big((0,\epsilon^{-3}],\epsilon\big)\Rightarrow \mathrm{Pois}(\alpha_q) . \] Thus domain-crossing jumps occur with positive probability for each fixed \(q\in(-1,1)\), but on the endpoint scale they become rare and asymptotically Poissonian [1603.09685]. A continuous-path analogue arises for the Ornstein–Uhlenbeck functional \[ A=\int_0^T v^n(t)\,dt . \] Using an excursion decomposition between zero crossings of \(v\), the observable is mapped to a continuous-time random walk in excursion areas. For \(n>2\), the first-passage excursion area has stretched-exponential tail, and the rare-event asymptotic becomes \[ P(A,T)\asymp T\exp\left\{-\frac{\gamma^{\frac{n+2}{n}}}{\sigma^2} c_n A^{2/n}\right\}, \] so the large-\(A\) behavior is dominated by the largest excursion rather than by a collective tilt of all excursions [2501.07704]. This gives a probabilistic interpretation of anomalous dynamical scaling and the associated dynamical phase transition in terms of a single-excursion big jump. ## 6. Alternative meanings, limits of validity, and conceptual scope Not every use of “big jump” in current probability theory refers to one-summand domination of a large deviation. In regular Dirichlet forms, the jump part of the form is split into small and big components through a scale-dependent threshold \(F_r(x,y)\), and the quantity \[ M_r^{(2)}:=\operatorname*{ess\,sup}_{x\in E}\int_{d(x,y)>F_r(x,y)} J(x,dy) \] is the big jump rate at scale \(r\) [2503.22899]. This enters upper bounds for the bottom of the essential spectrum, \[ \lambda_e \le \inf_{r>0}\frac{\mu_r^2}{4}M_r^{(1)} + 2M_r^{(2)} \] in the infinite-volume case, and analogously with \(\nu_r\) in finite volume [2503.22899]. Here “big jump” denotes the large-distance part of the jump kernel and its spectral contribution, not a rare-event realization mechanism for sums. The validity of a one-big-jump interpretation is itself regime-dependent. In the normal domain of attraction, the Gaussian and one-big-jump scenarios coexist and exchange dominance across a threshold scale [2303.12505]. In generalized Lévy walks, the principle can fail when a single step cannot exceed the characteristic bulk scale, in which case large displacements must be produced by many steps coherently aligned [1908.10975]. In heavy-tailed branching random walks, by contrast, Cramér tilts are unavailable at the relevant scale because \(m(\theta)=\infty\) for all \(\theta<1\), and this is precisely why one big jump dominates rather than a smooth collective deviation [2603.16316]. A second limit case is truncation. Under a moving cut-off, the correct mechanism is no longer “one big jump” in general but “fewest big jumps,” meaning the minimal number of macroscopic terms needed to meet the constraint [2206.14627]. A third is correlation: for correlated heavy-tailed increments, the rare event is still triggered by one large innovation, but the observed sum depends on the big jump and the following increments generated by the memory kernel [2106.14222]. These distinctions indicate that the phrase does not denote a single theorem or a single asymptotic formula. Rather, it names a recurring mechanism: rare events in heavy-tailed or non-local systems are often organized by one exceptional jump, one exceptional excursion, or, when constraints intervene, by the minimal number of such exceptional events. The modern literature shows both the robustness of this mechanism and the precision with which its breakdowns can be characterized. Source: https://www.emergentmind.com/topics/big-jump-phenomenon