Subexponentiality

Statement¶

Let $X_1,X_2,\dots$ be iid nonnegative random variables with distribution $F$ and unbounded right support, so $\bar F(x)=\mathbb P(X>x)>0$ for every finite $x$. The distribution is subexponential when, for two independent copies,

Equivalently, for each fixed integer $n\ge1$, with $S_n=X_1+\cdots+X_n$ and $M_n=\max_{1\le i\le n}X_i$,

Equivalently,

This is the one-big-jump principle: for a fixed number of summands, a very large sum is usually caused by one summand being very large, not by all summands moving moderately together.

The maximum-tail asymptotic $\mathbb P(M_n>x)\sim n\bar F(x)$ is not the distinctive part of the theorem: it follows for every iid distribution with an unbounded right tail and fixed $n$. The subexponential content is that the sum tail has the same first-order asymptotic as the maximum tail.

Regularly varying survival tails with positive exponent, $\bar F\in RV_{-\alpha}$ for $\alpha>0$, are a standard subexponential class. We cite that theorem and use the exact Pareto family as a checkable example.

Here $F$ is the common CDF, $n$ is fixed while $x$ grows, and $\alpha>0$ and $x_m>0$ are the exponent and cutoff in the Pareto example. The notation $a(x)\sim b(x)$ means $a(x)/b(x)\to1$; $o(b(x))$ denotes a term negligible after division by $b(x)$.

One-big-jump intuition¶

Subexponential does not mean “lighter than exponential,” and it is unrelated to the concentration-theory phrase “sub-exponential random variable” for a light-tailed $\psi_1$ condition. In this context it means the opposite: the tail is heavy enough that convolution barely changes the leading-order tail probability.

For two iid summands, a sufficiently large sum is about twice as likely as one summand exceeding the same threshold. For $n$ fixed summands, the leading factor is $n$.

This is the mathematical version of a recurring Incerto warning. In thin-tail settings, a large aggregate deviation is often the result of many small coordinated deviations. In subexponential settings, the aggregate tail is usually dominated by a single large term. That changes how sums, ruin events, insurance losses, and sample moments should be read.

What we can prove directly¶

The statement collects several equivalent forms and one important sufficient class. We prove the elementary maximum-tail asymptotic and an exact Pareto two-summand check. The full equivalence theory and the theorem that regularly varying tails are subexponential are cited, because their proofs need more convolution-tail machinery than we develop here.

The theorem that regularly varying tails are subexponential goes back to Chistyakov’s convolution-tail criterion for sums of independent positive random variables 1964. Standard modern references with proofs are Embrechts, Klueppelberg, and Mikosch, Appendix A3 1997, and Foss, Korshunov, and Zachary, Chapter 3 2013. We do not reprove the full theorem, but the exact Pareto case gives the right calibration.

Universal maximum tail¶

For any distribution with unbounded right tail, $\bar F(x)\to0$ as $x\to\infty$. For fixed $n$, iid independence gives

Therefore

Set $u=\bar F(x)$. Since $u\to0$ and $n$ is fixed, the binomial expansion gives

and therefore

Substituting back,

This maximum formula only uses iid independence, fixed $n$, and $\bar F(x)\to0$. For nonnegative variables, $M_n>x$ implies $S_n=X_1+\cdots+X_n>x$, so it is a lower bound for the sum tail. What is special about subexponential tails is the matching upper bound:

Those are the configurations where several observations are moderately large but none exceeds $x$. Subexponentiality says this residual event is negligible at the scale of one tail probability.

Pareto upper-bound check¶

For the Pareto tail $\bar F(x)=(x_m/x)^\alpha$ with $x\ge x_m$, the missing upper-bound step can be checked directly for $n=2$. Let

Choose $h=x_m(x/x_m)^\gamma$ with $1/2<\gamma<1$. This keeps the units consistent and gives $h/x\to0$ and $h\to\infty$.

The residual event $A_x$ is the region in $[0,x]^2$ above $X_1+X_2=x$. It is covered by two thin strips and a region where both coordinates exceed $h$. The following table replaces the geometric plot.

Outside the cover region, a point has neither coordinate in $(x-h,x]$ and not both coordinates above $h$. Thus at least one coordinate is $\le h$, and the other is $\le x-h$, so $X_1+X_2\le x$. Therefore any point in $A_x$ must lie in the cover: either one summand lies in $(x-h,x]$, or both summands exceed $h$. Hence

For the first term,

because $h/x\to0$. For the second term, independence gives

because $\gamma>1/2$. Thus $\mathbb P(A_x)=o(\bar F(x))$. Since $\{S_2>x\}$ is the disjoint union of $\{M_2>x\}$ and $A_x$,

The regularly varying theorem cited above generalizes this Pareto calculation: the maximum formula is universal, while the negligible residual event is the heavy-tail property.

A useful comparison table is:

The Tail Class Catalog gives a broader comparison. The lognormal and stretched Weibull entries are cited examples, not consequences of the Pareto proof.

The exponential is an exact nonexample. If $X_i\sim\operatorname{Exp}(\lambda)$, then the universal maximum formula still holds:

But the sum tail is much larger:

so

not 1.

The fixed-$n$ assumption is load-bearing. If $n$ grows with $x$, additional large-deviation regimes can appear.

A static bound on the residual tail¶

The proof supplies a finite-threshold bound. Put $t=x/x_m$ and take $\alpha=1.5$, $\gamma=0.75$. For sufficiently large $t$ with $x-h\ge x_m$,

Together with $\mathbb P(M_2>x)=2\bar F(x)-\bar F(x)^2$, this bounds the sum tail around its leading term. It illustrates how an asymptotic statement can require large thresholds even when the model is exactly Pareto. These are conservative analytic bounds, not measured convolution probabilities.

The fixed-$n$, large-threshold limit differs from the growing-sample max-to-sum ratio. A finite-mean Pareto law is subexponential even though $M_n/S_n\to0$ almost surely as $n\to\infty$.

Caveats¶

• The definition above is for nonnegative iid summands. Two-sided or dependent data need extra assumptions before the one-big-jump reading is valid.

• The equivalence for $n$ summands holds for fixed $n$. It is not a uniform statement over arbitrary growing horizons.

• Subexponentiality is a tail property. It does not say the body of the distribution is Pareto, nor does it choose an empirical threshold.

• A simulation ratio near one is only a diagnostic. A durable mathematical claim needs a theorem, a checked distributional assumption, or a citation.

References¶

• Embrechts, Klueppelberg, and Mikosch, Modelling Extremal Events 1997.

• Bingham, Goldie, and Teugels, Regular Variation 1987.

• Resnick, Heavy-Tail Phenomena 2007.

• Taleb, Statistical Consequences of Fat Tails 2020.

Backlinks¶

• Depends on: Regular Variation.

• Related catalog: Tail Class Catalog.

• Related: Max-to-Sum Ratio.

• Related to: LLN Failure Under Infinite Mean.

Source and adaptation¶

Adapted from incerto-wiki, content/concepts/theorems/subexponentiality.md, revision 9717c9c (2026-09-13 Batch 2 import). Copyright (c) 2023 xshi19. Licensed under MIT. Links, notation, and qualifications were adapted for this site; executable figures and simulations were replaced with static calculations. No upstream execution or formal-proof verification is claimed for this adaptation.