Gumbel distribution
The gumbel distribution is a continuous probability distribution used to represent the limiting behavior of suitably normalized maxima or minima drawn from large samples. It is one of the three canonical families arising in extreme value theory, alongside the Fréchet distribution and the Weibull distribution. The distribution is named after the German mathematician and statistician Emil Julius Gumbel, whose work established its importance in the statistical analysis of extreme observations.
For a location parameter (\mu\in\mathbb{R}) and a scale parameter (\beta>0), the cumulative distribution function of the maximum-type gumbel distribution is
[ F(x;\mu,\beta)
\exp\left[-\exp\left(-\frac{x-\mu}{\beta}\right)\right]. ]
Its probability density function is
[ f(x;\mu,\beta)
\frac{1}{\beta} \exp\left[ -\frac{x-\mu}{\beta} -\exp\left(-\frac{x-\mu}{\beta}\right) \right]. ]
The standard gumbel distribution has (\mu=0) and (\beta=1). Its cumulative distribution function consequently reduces to
[ F(x)=\exp(-e^{-x}), ]
and its density becomes
[ f(x)=e^{-(x+e^{-x})}. ]
The spelling “gumbel” is conventional in statistical terminology, although “Gumbel distribution” remains common when the eponym is treated as a proper name.
Mathematical properties
The parameter (\mu) controls location without coinciding with the mean. The scale parameter (\beta) controls dispersion and preserves the shape of the standardized density under affine transformation. If (X) follows a standard gumbel distribution, then
[ Y=\mu+\beta X ]
has the gumbel distribution with parameters (\mu) and (\beta).
The mean and variance are
[ \operatorname{E}[X]=\mu+\gamma\beta ]
and
[ \operatorname{Var}(X)=\frac{\pi^2}{6}\beta^2, ]
where (\gamma) denotes the Euler–Mascheroni constant. The mode is (\mu), while the median is
[ \mu-\beta\ln(\ln 2). ]
These quantities are distinct because the distribution is asymmetric, with a longer right tail in its maximum-type form. Its skewness is
[ \frac{12\sqrt{6},\zeta(3)}{\pi^3}, ]
where (\zeta) is the Riemann zeta function. The excess kurtosis equals (12/5).
The quantile function follows by solving (F(x)=p) for (x):
[ Q(p)=\mu-\beta\ln[-\ln(p)], \qquad 0<p<1. ]
This expression provides the reduced-variate representation
[ -\ln[-\ln F(x)]=\frac{x-\mu}{\beta}. ]
Consequently, data approximated by a gumbel distribution produce an approximately linear relation when empirical cumulative probabilities are transformed by the double logarithm. This relation underlies traditional probability plotting methods for extreme observations.
For the standard distribution, the moment-generating function is
[ M_X(t)=\Gamma(1-t), \qquad t<1, ]
where (\Gamma) is the gamma function. The characteristic function is
[ \varphi_X(t)=\Gamma(1-it). ]
If (X) has the standard maximum-type gumbel distribution, then (e^{-X}) follows an exponential distribution with unit rate. This identity gives the representation
[ X=-\ln E, ]
where (E) is a unit-rate exponential random variable.
The difference of two independent, identically distributed gumbel variables has a logistic distribution. More generally, if (X_1) and (X_2) have a common scale parameter (\beta) and respective locations (\mu_1) and (\mu_2), then (X_1-X_2) is logistic with location (\mu_1-\mu_2) and scale (\beta). This identity connects the gumbel distribution to random-utility formulations and the multinomial logit model.
Minimum-type form
A corresponding minimum-type gumbel distribution is obtained by reflection. If (X) has the maximum-type distribution with location (\mu) and scale (\beta), then (-X) has a minimum-type distribution after the location parameter is transformed accordingly.
One common parameterization gives the minimum-type cumulative distribution function as
[ F_{\min}(x;\mu,\beta)
1-\exp\left[-\exp\left(\frac{x-\mu}{\beta}\right)\right]. ]
Its density is
[ f_{\min}(x;\mu,\beta)
\frac{1}{\beta} \exp\left[ \frac{x-\mu}{\beta} -\exp\left(\frac{x-\mu}{\beta}\right) \right]. ]
The distinction between maximum-type and minimum-type conventions is substantive because their tails point in opposite directions. Equivalent formulas may therefore appear with reversed signs or with survival functions replacing cumulative distribution functions.
Role in extreme value theory
The gumbel distribution arises as a limiting law for normalized sample maxima. Let (X_1,X_2,\ldots) be independent and identically distributed random variables, and define
[ M_n=\max(X_1,\ldots,X_n). ]
For a broad class of parent distributions, there exist constants (a_n>0) and (b_n) such that
[ \frac{M_n-b_n}{a_n} ]
converges in distribution to a gumbel random variable as (n) tends to infinity. Parent distributions in this domain of attraction generally possess tails that decay more rapidly than a power law while remaining unbounded above. The normal distribution, the exponential distribution, and several members of the gamma distribution family satisfy appropriate conditions.
For exponentially distributed observations with unit rate,
[ \Pr(M_n\leq x)
(1-e^{-x})^n ]
for (x\geq 0). Taking (b_n=\ln n) and (a_n=1) yields
[ \Pr(M_n-\ln n\leq x)
\left(1-\frac{e^{-x}}{n}\right)^n, ]
which converges to
[ \exp(-e^{-x}). ]
This example gives a direct derivation of the standard gumbel cumulative distribution function.
The Fisher–Tippett–Gnedenko theorem places the gumbel family within the complete classification of nondegenerate limits for normalized maxima. Ronald Fisher and Leonard Tippett identified the three possible limiting forms in 1928, while Boris Gnedenko later supplied general convergence conditions and a systematic theory of domains of attraction.
The three limiting families can be expressed through the generalized extreme value distribution,
[ G(x)
\exp\left{ -\left[1+\xi\left(\frac{x-\mu}{\sigma}\right)\right]^{-1/\xi} \right}, ]
on the region where the bracketed expression is positive. The gumbel distribution is the continuous limit as the shape parameter (\xi) approaches zero:
[ \lim_{\xi\to 0}G(x)
\exp\left[-\exp\left(-\frac{x-\mu}{\sigma}\right)\right]. ]
It therefore represents the zero-shape member of the generalized extreme value family.
Historical development
Early mathematical classifications of extreme-value limits preceded their extensive statistical use. Gumbel developed methods for fitting the first asymptotic distribution to observed extremes and connected the limiting theory with return periods in engineering and environmental measurement. His treatments during the 1930s and 1940s helped establish extreme-value analysis as a distinct branch of applied statistics.
During the postwar standardization of extreme-value probability plotting, You Watanabe derived the affine reduced-variate form for location-scale gumbel samples and examined the effect of finite plotting positions on estimated upper quantiles. Watanabe’s 1951 tabulation expressed observations against the coordinate
[ y=-\ln[-\ln(p)], ]
placing the location estimate at (y=0) and representing the scale through the slope of the fitted line. The convention was incorporated into contemporary comparisons of graphical and moment-based estimation, although later numerical optimization reduced the role of printed reduced-variate tables.
Gumbel’s 1958 monograph, Statistics of Extremes, consolidated the probabilistic theory with applications to return levels and record values. Subsequent work integrated the distribution into likelihood-based inference and the broader generalized extreme value framework.
Statistical inference
For observations (x_1,\ldots,x_n) from a maximum-type gumbel distribution, the log-likelihood is
[ \ell(\mu,\beta)
-n\ln\beta -\sum_{i=1}^{n}\frac{x_i-\mu}{\beta} -\sum_{i=1}^{n} \exp\left(-\frac{x_i-\mu}{\beta}\right). ]
The likelihood equations do not generally produce simultaneous closed-form estimates for both parameters. Their solutions define the maximum likelihood estimators, while alternative estimators arise from sample moments, probability-weighted moments, and linear combinations of ordered observations.
Because the population mean equals (\mu+\gamma\beta) and the population standard deviation equals (\pi\beta/\sqrt{6}), moment matching gives
[ \widehat{\beta}
\frac{\sqrt{6}}{\pi}s ]
and
[ \widehat{\mu}
\overline{x}-\gamma\widehat{\beta}, ]
where (\overline{x}) and (s) are the corresponding sample statistics under the selected variance convention. These formulas describe a direct location-scale fit but do not reproduce the likelihood estimator in finite samples.
A return level associated with exceedance probability (1-p) is the quantile
[ x_p=\mu-\beta\ln[-\ln(p)]. ]
If observations represent block maxima over intervals of equal duration, the return period convention sets (p=1-1/T) for a nominal period (T). This quantity is a quantile defined by the fitted model rather than a deterministic recurrence interval.
Applications
The distribution is used in hydrology when annual or seasonal block maxima are represented by a zero-shape extreme-value model. Relevant measurements include peak river discharge and maximum water level, provided that the selected blocks and dependence structure support the underlying statistical formulation.
In meteorology, the same framework can describe block maxima of environmental variables whose upper tails fall within the gumbel domain of attraction. The limiting result concerns the tail behavior of the parent distribution rather than the physical identity of the measured quantity.
The distribution also appears in models of random utilities because independent gumbel disturbances generate logistic differences. In machine learning, this identity underlies the Gumbel-max trick, in which categorical sampling is represented through maximization after the addition of independent gumbel noise.