Gumbel distribution (Type I extreme value distribution)

Definition and common formulas

In its standard parameterization the Gumbel distribution has two parameters: a location parameter μ and a positive scale parameter β. For z = (x − μ)/β the cumulative distribution function is F(x) = exp(−e^(−z)), and the probability density function is f(x) = (1/β) exp(−z − e^(−z)). With these formulas the distribution is skewed to the right and has a single mode at x = μ. For minima one often uses the mirrored form or negates observations before fitting. Further context is available in introductory statistics sources.

Key properties

• Mean: μ + γβ, where γ is the Euler–Mascheroni constant (≈ 0.5772).
• Variance: (π^2/6) β^2; the distribution therefore scales with β.
• Skewness is fixed and does not depend on μ or β; kurtosis is also a constant.
• It is the limiting distribution for properly normalized maxima of many light-tailed parent distributions (extreme value theory).

Origin and relations

The name honors Emil J. Gumbel, who contributed to extreme value theory in the early 20th century. The Gumbel family appears naturally via the Fisher–Tippett–Gnedenko theorem as the limit law for maxima in many independent, identically distributed samples after affine normalization. It is a special case of the generalized extreme value (GEV) distribution obtained when the GEV shape parameter tends to zero. Although sometimes called a "double exponential" in older texts, that phrase is also used for the Laplace distribution, so the terms are not interchangeable.

Applications and examples

Because it models rare but high-impact events, the Gumbel distribution is widely used in hydrology (annual flood peaks), meteorology (extreme temperatures or wind speeds), civil engineering (design for extreme loads), and risk management. In machine learning the Gumbel distribution underlies the Gumbel-max trick, a sampling method for discrete distributions useful in simulations and variational approximations. Practical use typically involves fitting μ and β by maximum likelihood or probability plotting.

Estimation, simulation and notable facts

Parameter estimation can be done by maximum likelihood, method of moments, or specialized plotting techniques for extremes. Simulating Gumbel variates is straightforward: if U is uniform on (0,1) then X = μ − β ln(−ln U) has the Gumbel(μ,β) law. For more advanced discussion and references consult further reading.