DRAFT This draft is under development — DO NOT CITE OR SHARE.

38  Gumbel Distribution

The Gumbel distribution is the limiting distribution of the maximum of many independent random variables from light-tailed distributions. It is the cornerstone of extreme-value statistics, used wherever one needs to model the probability of record-breaking events: annual flood peaks, maximum daily temperatures, or largest structural loads.

Formally, the random variate \(X\) defined for all of \(\mathbb{R}\), is said to have a Gumbel Distribution (i.e. \(X \sim \text{Gumbel}(\mu, \beta)\)) with location parameter \(\mu \in \mathbb{R}\) and scale parameter \(\beta > 0\). The Euler-Mascheroni constant is \(\gamma \approx 0.5772156649\).

38.1 Probability Density Function

\[ f(x) = \frac{1}{\beta}\exp\!\bigl(-z - e^{-z}\bigr), \quad z = \frac{x - \mu}{\beta} \]

The figure below shows examples of the Gumbel Probability Density Function for different parameter combinations.

Code 
dgumbel <- function(x, mu, beta) {
  z <- (x - mu) / beta
  exp(-z - exp(-z)) / beta
}

par(mfrow = c(2, 2))
x <- seq(-4, 12, length = 500)

plot(x, dgumbel(x, 0, 1), type = "l", lwd = 2, col = "blue",
     xlab = "x", ylab = "f(x)", main = expression(paste(mu == 0, ",  ", beta == 1)))

plot(x, dgumbel(x, 0, 2), type = "l", lwd = 2, col = "blue",
     xlab = "x", ylab = "f(x)", main = expression(paste(mu == 0, ",  ", beta == 2)))

plot(x, dgumbel(x, 2, 1), type = "l", lwd = 2, col = "blue",
     xlab = "x", ylab = "f(x)", main = expression(paste(mu == 2, ",  ", beta == 1)))

plot(x, dgumbel(x, 5, 3), type = "l", lwd = 2, col = "blue",
     xlab = "x", ylab = "f(x)", main = expression(paste(mu == 5, ",  ", beta == 3)))

par(mfrow = c(1, 1))
Figure 38.1: Gumbel Probability Density Function for various parameter combinations

38.2 Purpose

The Gumbel distribution belongs to the generalized extreme value (GEV) family and specifically describes the limiting distribution of block maxima from distributions with exponentially bounded tails (Normal, Exponential, Gamma). Its right-skewed shape with a mode below the mean is characteristic of maximum-type data. Common applications include:

  • Hydrology: annual maximum flood discharge, maximum daily rainfall
  • Structural engineering: maximum wind load, maximum snow depth for design standards
  • Climatology: maximum daily temperature records, extreme precipitation events
  • Financial risk: maximum portfolio loss within a given time horizon
  • Reliability engineering: maximum stress or load applied to a component

Relation to the discrete setting. The Gumbel distribution is the continuous limit of the maximum of \(n\) i.i.d. Geometric or Poisson variates after normalization — the Fisher-Tippett-Gnedenko theorem applies in the discrete domain too. For the maximum of Geometric\((p)\) variates as \(n\to\infty\), the normalized maximum approaches a Gumbel distribution.

38.3 Distribution Function

\[ F(x) = \exp\!\bigl(-e^{-(x-\mu)/\beta}\bigr) \]

The figure below shows the Gumbel Distribution Function for \(\mu = 0\) and \(\beta = 1\).

Code 
pgumbel <- function(x, mu, beta) {
  exp(-exp(-(x - mu) / beta))
}

x <- seq(-4, 10, length = 500)
plot(x, pgumbel(x, 0, 1), type = "l", lwd = 2, col = "blue",
     xlab = "x", ylab = "F(x)", main = "Gumbel Distribution Function",
     sub = expression(paste(mu == 0, ",  ", beta == 1)))
Figure 38.2: Gumbel Distribution Function (location = 0, scale = 1)

38.4 Moment Generating Function

\[ M_X(t) = \Gamma(1 - \beta t)\,e^{\mu t}, \quad t < \frac{1}{\beta} \]

where \(\gamma \approx 0.5772156649\) is the Euler-Mascheroni constant and \(\Gamma(\cdot)\) is the gamma function.

38.5 1st Uncentered Moment

\[ \mu_1' = \mu + \gamma\beta \]

where \(\gamma \approx 0.5772156649\) is the Euler-Mascheroni constant.

38.6 2nd Uncentered Moment

\[ \mu_2' = (\mu + \gamma\beta)^2 + \frac{\pi^2\beta^2}{6} \]

38.7 3rd Uncentered Moment

Obtained via the third derivative of the MGF; involves the polygamma function \(\psi''(1) = -2\zeta(3)\) where \(\zeta(3) \approx 1.20206\).

38.8 4th Uncentered Moment

Obtained via the fourth derivative of the MGF.

38.9 2nd Centered Moment

\[ \mu_2 = \frac{\pi^2\beta^2}{6} \]

38.10 3rd Centered Moment

\[ \mu_3 = 2\zeta(3)\beta^3, \quad \zeta(3) \approx 1.20206 \]

38.11 4th Centered Moment

\[ \mu_4 = \frac{27}{5}\,\mu_2^2 = \frac{27}{5}\cdot\frac{\pi^4\beta^4}{36} \]

38.12 Expected Value

\[ \text{E}(X) = \mu + \gamma\beta, \quad \gamma \approx 0.5772156649 \]

38.13 Variance

\[ \text{V}(X) = \frac{\pi^2\beta^2}{6} \]

38.14 Median

\[ \text{Med}(X) = \mu - \beta\ln(\ln 2) \]

38.15 Mode

\[ \text{Mo}(X) = \mu \]

38.16 Coefficient of Skewness

\[ g_1 = \frac{12\sqrt{6}\,\zeta(3)}{\pi^3} \approx 1.1395 \]

This is a fixed universal constant, independent of \(\mu\) and \(\beta\).

38.17 Coefficient of Kurtosis

\[ g_2 = \frac{27}{5} = 5.4 \]

This is also a fixed universal constant. The excess kurtosis is \(g_2 - 3 = 2.4\).

38.18 Parameter Estimation

Method-of-moments estimates from mean and variance:

\[ \tilde\beta = \frac{s\sqrt{6}}{\pi}, \qquad \tilde\mu = \bar x - \gamma\tilde\beta \]

where \(s\) is the sample standard deviation and \(\gamma \approx 0.5772\).

set.seed(42)
mu_true <- 10; beta_true <- 5
u <- runif(100)
x_obs <- mu_true - beta_true * log(-log(u))

# Method of moments
gamma_em <- 0.5772156649
beta_hat <- sd(x_obs) * sqrt(6) / pi
mu_hat   <- mean(x_obs) - gamma_em * beta_hat
cat("MoM mu:  ", round(mu_hat, 4), "\n")
cat("MoM beta:", round(beta_hat, 4), "\n")
cat("True mu:", mu_true, "  True beta:", beta_true, "\n")
MoM mu:   10.2689 
MoM beta: 5.3249 
True mu: 10   True beta: 5

38.19 R Module

38.19.1 RFC

The Gumbel Distribution module is available in RFC under the menu “Distributions / Gumbel Distribution”.

38.19.3 R Code

The following code demonstrates Gumbel probability calculations:

mu <- 10; beta <- 5
gamma_em <- 0.5772156649

dgumbel <- function(x, mu, beta) {
  z <- (x - mu) / beta
  exp(-z - exp(-z)) / beta
}
pgumbel <- function(x, mu, beta) exp(-exp(-(x - mu) / beta))

# P(X <= 18)
pgumbel(18, mu, beta)

# P(X > 18)
1 - pgumbel(18, mu, beta)

# Mean and mode
cat("Mean:", mu + gamma_em * beta, "\n")
cat("Mode:", mu, "\n")
cat("Median:", mu - beta * log(log(2)), "\n")
[1] 0.8171795
[1] 0.1828205
Mean: 12.88608 
Mode: 10 
Median: 11.83256

38.20 Example

Annual maximum wind gust speeds (m/s) at a weather station are modeled as \(X \sim \text{Gumbel}(\mu = 10, \beta = 5)\). We compute the probability that the annual maximum gust exceeds 18 m/s.

mu <- 10; beta <- 5
gamma_em <- 0.5772156649
pgumbel <- function(x, mu, beta) exp(-exp(-(x - mu) / beta))

# P(max gust > 18 m/s)
cat("P(max gust > 18 m/s):", 1 - pgumbel(18, mu, beta), "\n")

# Mean annual maximum
cat("Mean max gust (m/s):", mu + gamma_em * beta, "\n")

# 100-year return level: 99th percentile
x_100yr <- mu - beta * log(-log(0.99))
cat("100-year return level (m/s):", round(x_100yr, 2), "\n")
P(max gust > 18 m/s): 0.1828205 
Mean max gust (m/s): 12.88608 
100-year return level (m/s): 33
Interactive Shiny app (click to load).
Open in new tab

38.21 Random Number Generator

Gumbel random variates are generated via the inverse-CDF method. Since \(F(x) = \exp(-e^{-(x-\mu)/\beta})\), solving for \(X\) gives:

\[ X = \mu - \beta\ln(-\ln U) \sim \text{Gumbel}(\mu, \beta) \quad \text{when } U \sim \text{U}(0,1) \]

set.seed(123)
n <- 1000; mu <- 0; beta <- 1
gamma_em <- 0.5772156649

# Inverse-transform method
u <- runif(n)
x_inv <- mu - beta * log(-log(u))

cat("Simulated mean:", round(mean(x_inv), 4), "\n")
cat("Theoretical mean:", mu + gamma_em * beta, "\n")
cat("Simulated var:", round(var(x_inv), 4), "\n")
cat("Theoretical var:", round(pi^2 * beta^2 / 6, 4), "\n")
Simulated mean: 0.561 
Theoretical mean: 0.5772157 
Simulated var: 1.5925 
Theoretical var: 1.6449
Interactive Shiny app (click to load).
Open in new tab

38.22 Property 1: Fisher-Tippett-Gnedenko Theorem

The Gumbel distribution is one of only three possible limiting distributions for normalized block maxima (Fisher-Tippett-Gnedenko theorem). It arises from distributions with exponentially bounded tails (Normal, Gamma, Exponential, Lognormal). This is the continuous-data counterpart of the classical central limit theorem for extreme values.

38.23 Property 2: Universal Shape Constants

Both the skewness \(g_1 \approx 1.1395\) and kurtosis \(g_2 = 5.4\) are universal constants — they do not depend on \(\mu\) or \(\beta\). Any data with these shape properties is consistent with the Gumbel model.

38.24 Property 3: Relationship to Weibull via Minimum

If \(X \sim \text{Gumbel}(\mu, \beta)\) then \(-X\) follows the minimum (reversed) Gumbel distribution. Furthermore, if \(Y = e^{-X/\beta}\) then \(Y\) follows a Weibull distribution, linking the two extreme-value families.