What Is a Lightly-Tailed Distribution?

A lightly-tailed distribution, more often written light-tailed distribution, is a probability distribution whose upper tail decays at least as fast as an exponential, so that extremely large values are rare in a strong and quantifiable sense. The standard criterion is the existence of the moment generating function: a random variable X is light-tailed if the expectation of exp(sX) is finite for some positive s. An equivalent statement is that the tail probability P(X > x) is bounded above by a constant times exp(-cx) for some c greater than zero and all sufficiently large x. Distributions that fail this condition are called heavy-tailed, and the exponential distribution sits at the edge of the light-tailed class.

The distinction is not a stylistic one. Whether a model is light-tailed determines which limit theorems apply, which bounds are available, how quickly simulation converges, and whether rare events are driven by many moderate deviations or by a single catastrophic one.

Formal Criteria and Equivalent Characterizations

Several conditions identify the class, and they are not all equivalent in the strictest sense, which is a common source of confusion. Finiteness of the moment generating function in a neighborhood of the origin is the usual definition and immediately implies that all moments are finite. A tail that is eventually bounded by a decaying exponential gives the same conclusion. A hazard rate bounded away from zero is a third route, and it is convenient in practice because the hazard function can be estimated directly from data; a proposed test of tail weight based on the hazard rate uses exactly this property to distinguish the two regimes from a finite sample. Sub-Gaussian and sub-exponential random variables form nested subclasses inside the light-tailed family, defined by explicit bounds on the growth of the moment generating function, and these are the classes assumed in most concentration inequalities.

Examples and the Exponential Boundary

Bounded random variables are trivially light-tailed, as are the Gaussian, gamma, chi-squared, Poisson, and binomial distributions, and the Weibull distribution when its shape parameter is at least one. The exponential distribution itself is light-tailed under the standard definition but marks the frontier: any tail heavier than exp(-cx) for every positive c fails the test. Pareto, lognormal, Cauchy, Student's t with few degrees of freedom, and Weibull with shape parameter below one are all heavy-tailed. Concentration bounds for conditional value at risk estimation, derived separately for the cases of light-tailed and heavy-tailed distributions, show how the same estimation problem changes character across the divide, which is a useful check on intuition built entirely on Gaussian models.

Consequences for Analysis and Engineering

Light tails make sharp tools available. Chernoff and Bernstein bounds give exponentially decaying error probabilities for sums of independent light-tailed variables, which is the foundation of large deviations theory and of most sample complexity results in statistical learning. In queueing theory, a light-tailed service time distribution yields waiting times whose tails also decay exponentially, so a modest buffer bounds the loss probability. Heavy-tailed service times destroy that conclusion and produce subexponential waiting time tails instead, which is why the assumption is stated explicitly in results such as work on tail-optimal scheduling in the light-tailed M/G/1 queue. Under maxima, light-tailed distributions typically fall in the Gumbel domain of attraction in extreme value theory. Rare event simulation is also easier, since importance sampling with exponential tilting is effective precisely when the moment generating function exists.

Applications

Light-tailed distributions are assumed or tested in a range of settings, including:

• Queueing and buffer sizing in communication networks
• Concentration bounds and generalization analysis in machine learning
• Reliability modeling and failure time analysis
• Rare event simulation and importance sampling
• Risk modeling in insurance and finance, where the assumption is often rejected
• Signal detection and channel capacity analysis under noise models