Updated 12 September 2026

Gumbel distribution

Take the maximum of n independent, identically distributed normal or exponential draws. With suitable centering and scaling as n grows, its distribution approaches Gumbel. Other parent distributions can have different extreme-value limits.

What to notice

• Right-skewed. The mode sits at $\mu$ but the mean leans right by $\beta\gamma$, where $\gamma$ is the Euler-Mascheroni constant ≈ 0.5772.
• Matched-variance Normal is wrong on both sides. The dashed curve has the same mean and variance as the Gumbel but misses badly — it’s too light in the right tail and too heavy in the left.
• Closed-form CDF. Unusually clean for an extreme-value distribution:

Inverting gives the sampler this demo uses.

Extreme-value theory

Fisher and Tippett proved that the renormalised maximum of $n$ iid draws — when a non-degenerate limit exists — must be one of three types: Gumbel, Fréchet, or reverse-Weibull. Normal and exponential parents lie in the Gumbel domain; Pareto parents give Fréchet, and uniform parents give reverse-Weibull. These examples are not a complete classification: boundedness or the informal label “heavy-tailed” alone does not identify every domain of attraction. These three collapse into the generalised extreme value (GEV) family.

Annual environmental maxima are a common motivation for extreme-value models. Choosing Gumbel rather than a more general GEV model requires checking the tail assumptions; a return-period calculation is only as reliable as its model and stationarity assumptions.

Derive one maximum distribution before taking a limit

Let m independent observations each have Exponential(rate 1) distribution. A maximum M is at most t exactly when all m observations are at most t, so P(M≤t)=(1−e⁻ᵗ)ᵐ for t≥0. Independence is what turns that joint probability into a product.

For Y=M−log(m), this becomes P(Y≤y)=(1−e⁻ʸ/m)ᵐ wherever y≥−log(m), and zero below that support. As m grows it approaches exp(−e⁻ʸ), the standard maximum-type Gumbel CDF. At y=0, m=10 gives 0.9¹⁰≈0.348678, while m=100 gives 0.99¹⁰⁰≈0.366032; the limit is e⁻¹≈0.367879.

The experiment draws every observation in each block before taking its maximum. Increasing the number of blocks makes its empirical CDF less noisy. Increasing observations per block changes the distribution being approximated. These controls answer different questions.

A different parent gives a different limit

For m independent Uniform(0,1) draws, P(M≤t)=tᵐ on [0,1]. Normalize using Y=m(M−1). Its CDF is (1+y/m)ᵐ on [−m,0], approaching exp(y) for y≤0 and one above zero. This reverse-Weibull example has an upper endpoint at zero. It cannot be the unbounded Gumbel distribution.

Model an application without promising its tail

NIST’s extreme-wind overview discusses Gumbel among several candidate extreme-value families and compares estimation methods. It is a documented application context, not evidence that every wind series follows Gumbel. Dependence, nonstationary weather, block choice, and sparse extreme observations can all affect a fitted return level.

A return period is the reciprocal of a model exceedance probability. A “100-year” annual threshold does not schedule an event once per century. The earlier 50-year calculation additionally assumes independent years and unchanged risk; it is not guaranteed by the Gumbel name.

References

NIST: extreme value type I distinguishes the maximum and minimum conventions. The finite-block calculations above derive the two specific examples directly from their parent CDFs. Compare exponential waiting times, uniform bounds, and Pareto tails before generalizing a maximum limit.