30. Variance of the Standard Gumbel Distribution
Settings
- Standard Gumbel (Type I extreme value) with CDF and PDF $$F(\epsilon) = \exp\!\left(-e^{-\epsilon}\right), \qquad f(\epsilon) = \exp\!\left(-\epsilon - e^{-\epsilon}\right), \qquad \epsilon \in (-\infty, +\infty) \tag{1}$$
- Gamma $\Gamma(z) = \int_0^{+\infty} u^{z-1} e^{-u}\, du$; Digamma $\psi(z) \equiv (\ln\Gamma)'(z) = \dfrac{\Gamma'(z)}{\Gamma(z)}$; Trigamma $\psi'(z) \equiv (\ln\Gamma)''(z)$. Their series representations are derived in Stage 3.
- MGF $M_X(t) = \mathbb{E}[e^{tX}]$, CGF $K_X(t) \equiv \ln M_X(t)$. Term-by-term expansion (valid when MGF exists near $0$): $$M_X(t) = \mathbb{E}\!\left[\sum_{k=0}^\infty \dfrac{(tX)^{k}}{k!}\right] = \sum_{k=0}^\infty \dfrac{t^{k}}{k!}\,\mathbb{E}[X^{k}] \quad\Longrightarrow\quad M_X^{(k)}(0) = \mathbb{E}[X^{k}]$$Chain rule then quotient rule on $\ln M_X$, evaluated at $0$ using $M_X(0) = 1$: $$K_X'(t) = \dfrac{M_X'(t)}{M_X(t)}, \qquad K_X''(t) = \dfrac{M_X''(t)\,M_X(t) - (M_X'(t))^{2}}{M_X(t)^{2}}$$$$\Longrightarrow\quad K_X'(0) = \mathbb{E}[X], \qquad K_X''(0) = \mathbb{E}[X^{2}] - (\mathbb{E}[X])^{2} = \operatorname{Var}(X) \tag{2}$$
- Objective: derive $\operatorname{Var}(\epsilon) = \dfrac{\pi^2}{6}$ via the MGF. Strategy runs three stages:
1. Close the Gumbel MGF into a Gamma function $\Gamma(1-t)$
2. Log-differentiate to translate mean and variance into polygamma values at $z = 1$
3. Evaluate those polygammas via the Weierstrass product, exposing the Trigamma → Basel link.
Proof
- **Stage 1** — Close the MGF into $\Gamma(1-t)$. Plug the Gumbel PDF into $M_\epsilon(t) = \mathbb{E}[e^{t\epsilon}]$ and merge the exponents linear in $\epsilon$ ($e^{t\epsilon}\cdot e^{-\epsilon} = e^{-(1-t)\epsilon}$): $$M_\epsilon(t) = \int_{-\infty}^{+\infty} e^{-(1-t)\epsilon}\, e^{-e^{-\epsilon}}\, d\epsilon \tag{3}$$
- Substitute $u = e^{-\epsilon}$: $\epsilon = -\ln u$, $d\epsilon = -\dfrac{du}{u}$, $e^{-(1-t)\epsilon} = u^{1-t}$, and the map is monotone decreasing so limits $(-\infty, +\infty) \to (+\infty, 0)$. The sign from $d\epsilon$ cancels the limit reversal and $\dfrac{u^{1-t}}{u} = u^{-t}$: $$M_\epsilon(t) = \int_0^{+\infty} u^{-t}\, e^{-u}\, du \tag{4}$$
- Match against the Gamma integral at $z = 1 - t$: $${\color{Orange} M_\epsilon(t) = \Gamma(1 - t)}, \qquad t < 1 \tag{5}$$Convergence range: $\Gamma(z) = \int_0^{+\infty} u^{z-1} e^{-u}\, du$ needs both endpoints finite — check each, then combine.
1. At $u \to 0^+$: $e^{-u} \in [e^{-1}, 1]$ on $[0,1]$, so $u^{z-1}$ alone decides integrability. **[[31. p-Integral Test for Improper Integrals|p-integral test]]** with $p = z - 1$ gives $$\int_0^1 u^{z-1} e^{-u}\, du \asymp \int_0^1 u^{z-1}\, du = \left.\dfrac{u^{z}}{z}\right|_0^1 = \dfrac{1}{z} < \infty \iff z > 0$$
2. At $u \to +\infty$: $e^{-u}$ dominates any polynomial for all $z$. **Cauchy ratio test** on log-derivative $g(u) = \ln\!\big(u^{z-1} e^{-u}\big)$ gives $g'(u) = \dfrac{z-1}{u} - 1 \to -1$, so $u^{z-1} e^{-u} \le C\, e^{-u/2}$ for $u \ge M$: $$\int_M^{+\infty}\! u^{z-1} e^{-u}\, du \le C\int_M^{+\infty}\! e^{-u/2}\, du < \infty \quad \forall z \in \mathbb{R}$$
3. Combine: $\Gamma(z) < \infty \iff z > 0$, so $z = 1 - t > 0 \iff t < 1$; moment derivatives only need a $t = 0$ neighborhood.
- **Stage 2** — Translate MGF into polygamma moments. Take logs of (5): $K(t) = \ln M_\epsilon(t) = \ln\Gamma(1-t)$. Chain rule with $\psi = \dfrac{\Gamma'}{\Gamma}$ (the extra $-1$ from $(1-t)'$ absorbs the outer minus at the second derivative): $$K'(t) = -\psi(1-t), \qquad K''(t) = \psi'(1-t) \tag{6}$$Evaluate at $t = 0$ and read moments off the CGF identities in (2): $$\mathbb{E}[\epsilon] = K'(0) = -\psi(1), \qquad \operatorname{Var}(\epsilon) = K''(0) = \psi'(1)$$The task reduces to two polygamma values at $z = 1$.
- **Stage 3** — Evaluate polygammas at $z = 1$ via Weierstrass product. Both polygamma values come from $\dfrac{1}{\Gamma(z)} = z\, e^{\gamma z}\prod_{n=1}^\infty\!\left(1 + \dfrac{z}{n}\right) e^{-\frac{z}{n}}$. Taking $-\ln$ of both sides gives $$\ln\Gamma(z) = -\ln z - \gamma z - \sum_{n=1}^\infty\!\left[\ln\!\left(1 + \dfrac{z}{n}\right) - \dfrac{z}{n}\right]$$Differentiate once for Digamma, twice for Trigamma, then evaluate at $z = 1$.
1. Digamma $\psi(z) = (\ln\Gamma)'(z)$. Term-by-term derivative of the log-series (using $\dfrac{d}{dz}\ln\!\left(1 + \dfrac{z}{n}\right) = \dfrac{1}{n+z}$): $$\psi(z) = -\dfrac{1}{z} - \gamma - \sum_{n=1}^\infty\!\left(\dfrac{1}{n+z} - \dfrac{1}{n}\right)$$At $z = 1$ the sum telescopes to $1$: $$\psi(1) = -\gamma - 1 + 1 = -\gamma \;\;\Longrightarrow\;\; \mathbb{E}[\epsilon] = \gamma$$This is the **Euler–Mascheroni** constant; it has no closed form — the series is the definition.
2. Trigamma $\psi'(z) = (\ln\Gamma)''(z)$. Differentiate the Digamma series once more (using $\dfrac{d}{dz}\dfrac{1}{n+z} = -\dfrac{1}{(n+z)^{2}}$): $$\psi'(z) = \dfrac{1}{z^{2}} + \sum_{n=1}^\infty \dfrac{1}{(n+z)^{2}} = \sum_{n=0}^\infty \dfrac{1}{(z+n)^{2}}$$At $z = 1$ this becomes the **Basel sum**: $$\psi'(1) = \sum_{k=1}^\infty \dfrac{1}{k^{2}} = \zeta(2) = \dfrac{\pi^{2}}{6} \;\;\Longrightarrow\;\; \operatorname{Var}(\epsilon) = \dfrac{\pi^{2}}{6} \tag{7}$$
Conclusion
The standard Gumbel has mean $\gamma$ and variance $\dfrac{\pi^2}{6}$, both read off the first two derivatives of $\ln\Gamma(1-t)$ at zero. Q.E.D.

