Cramér Exponential-Moment Condition

Statement¶

A random variable $X$ satisfies a right-tail Cramér condition when there is some $\theta>0$ such that

\mathbb E[e^{\theta X}]<\infty.

(1)

For two-sided large-deviation statements, one usually asks for the moment generating function $M(\theta)=\mathbb E[e^{\theta X}]$ to be finite in an open neighborhood of 0. Equivalently, the cumulant generating function $\Lambda(\theta)=\log \mathbb E[e^{\theta X}]$ is finite near zero, and the Cramér rate function is

I(x)=\sup_\theta\{\theta x-\Lambda(\theta)\}.

(2)

We use the one-sided version when the question is upper-tail concentration; the classical two-sided Cramér large-deviation theorem requires stronger mgf control.

The condition gives the elementary Chernoff bound

\mathbb P(X>x)\le e^{-\theta x}\mathbb E[e^{\theta X}],

(3)

so the upper-tail probability is bounded above by an exponentially decaying function. Equivalently, $\bar F(x)=O(e^{-\theta x})$ for that value of $\theta$ .

Here $\bar F(x)=\mathbb P(X>x)$ is the survival function, $\mathbb E$ is expectation, and $\theta$ is the exponential-tilting parameter. All symbols needed here are defined locally; see also Shared notation.

Thin-tail intuition¶

The Cramér condition is a thin-tail gate. If exponential moments exist, then exponential tilting, Chernoff bounds, and classical large-deviation rates have room to operate. If no positive exponential moment exists, those tools may give a false sense of security.

This is the clean contrast with Pareto-type fat tails. A regularly varying tail can have many finite ordinary moments, but multiplying by $e^{\theta X}$ eventually overwhelms every power-law decay. Failure of the Cramér condition does not require infinite variance: it can also occur when both the mean and variance are finite.

Chernoff bound and examples¶

We prove the displayed Chernoff bound from Markov’s inequality. The right-tail and two-sided Cramér conditions are definitions or standard large-deviation hypotheses; the larger Cramér theorem is cited rather than proved here. The normal, exponential, and Pareto cases below are direct checks of the condition.

The Chernoff bound follows from Markov’s inequality applied to the nonnegative random variable $e^{\theta X}$ :

\mathbb P(X>x) = \mathbb P(e^{\theta X}>e^{\theta x}) \le e^{-\theta x}\mathbb E[e^{\theta X}].

(4)

A standard normal random variable satisfies the two-sided version because $\mathbb E[e^{\theta X}]=e^{\theta^2/2}$ for all real $\theta$ . An exponential random variable with rate $\lambda>0$ has mgf $\lambda/(\lambda-\theta)$ for $\theta<\lambda$ , so it also satisfies the two-sided neighborhood condition. Its positive exponential moments exist exactly for $0<\theta<\lambda$ .

For a Pareto Type I random variable,

\mathbb E[e^{\theta X}] = \int_{x_m}^{\infty} e^{\theta x}\alpha x_m^\alpha x^{-(\alpha+1)}\,dx.

(5)

For every $\theta>0$ , the exponential factor dominates the polynomial decay, so the integral diverges. Thus Pareto tails fail the right-tail Cramér condition for every positive $\theta$ .

Truncated exponential moment check¶

For an exponential variable of rate $\lambda$ and $c\ge0$ ,

\mathbb E[e^{\theta X}\mathbf 1_{\{X\le c\}}] =\frac{\lambda}{\lambda-\theta} \left(1-e^{-(\lambda-\theta)c}\right), \qquad 0<\theta<\lambda.

(6)

For example, $\lambda=1$ and $\theta=0.12$ give a finite limit $1/0.88\approx1.13636$ as $c\to\infty$ .

For Pareto Type I with $\alpha>0$ , every $\theta>0$ instead gives a lower bound that diverges. For $c\ge2x_m$ , retain only the interval $(c/2,c]$ :

\begin{aligned} \mathbb E[e^{\theta X}\mathbf 1_{\{X\le c\}}] &\ge e^{\theta c/2}\mathbb P(c/2<X\le c)\\ &=(2^\alpha-1)\left(\frac{x_m}{c}\right)^\alpha e^{\theta c/2}\longrightarrow\infty. \end{aligned}

(7)

These are population calculations, independent of simulation. Even a Pareto law with $\alpha=2.5$ , finite mean, and finite variance has no positive exponential moment.

Caveats¶

The condition is direction-specific unless the moment generating function is finite around both sides of zero.
Failure of the Cramér condition does not by itself prove regular variation or subexponentiality. A lognormal distribution has every positive polynomial moment finite but no positive exponential moment, so “all moments finite” is not enough for the Cramér condition.
Empirical samples cannot prove that an exponential moment exists; a finite sample always has a finite empirical exponential average.
Large-deviation results need more than the one-line condition here. This page records the gate and the thin-tail contrast, not the full theorem.

References¶

Feller, An Introduction to Probability Theory and Its Applications, Vol. II 1971, 2nd ed., Wiley.
Embrechts, Klueppelberg, and Mikosch, Modelling Extremal Events 1997.
Taleb, Statistical Consequences of Fat Tails, third edition (2025), arXiv:2001.10488v4, ch. 1, pp. 3–5: motivation for distinguishing tail assumptions. The elementary Chernoff and model calculations are proved here; this locator is not a citation for the general large-deviation theorem.

Backlinks¶

Depends on: Regular Variation.
Used by: Subexponentiality and thin-tail contrast pages.

References¶

Embrechts, P., Klüppelberg, C., & Mikosch, T. (1997). Modelling Extremal Events. Springer. 10.1007/978-3-642-33483-2