Pickands-Balkema-de Haan Theorem

Statement¶

Let $F$ be a distribution function with finite or infinite right endpoint $x_F=\sup\{x:F(x)<1\}$ . For a high threshold $u<x_F$ , define the conditional excess distribution

F_u(y)=\mathbb P(X-u\le y\mid X>u) =\frac{F(u+y)-F(u)}{1-F(u)}, \qquad 0\le y<x_F-u.

(1)

If $F$ is in the maximum domain of attraction of an extreme-value law with extreme-value index $\xi$ , then there is a positive scale function $\beta(u)$ such that

\sup_{0\le y<x_F-u} \left|F_u(y)-G_{\xi,\beta(u)}(y)\right| \to 0, \qquad u\uparrow x_F.

(2)

The comparison is over the excess support, using the full GPD CDF: for $\xi<0$ , set $G_{\xi,\beta}(y)=1$ at and beyond $-\beta/\xi$ . The generalized Pareto term has scale $\beta>0$ and is

G_{\xi,\beta}(y) = 1-\left(1+\frac{\xi y}{\beta}\right)^{-1/\xi}, \qquad \xi\ne0,

(3)

on the support where $1+\xi y/\beta>0$ , and

G_{0,\beta}(y)=1-e^{-y/\beta}

(4)

for $\xi=0$ . The theorem is the asymptotic justification for the peaks-over-threshold model: sufficiently high threshold exceedances are modeled by a generalized Pareto distribution.

$X$ denotes a draw from $F$ , $u$ the threshold, $x_F$ its right endpoint, $\xi$ the extreme-value index, and $\beta(u)>0$ the excess scale. The maximum-domain-of-attraction assumption means that for iid draws, some $a_n>0$ and $b_n$ give $(M_n-b_n)/a_n\Rightarrow H_{\xi,0,1}$ . Here $M_n$ is the sample maximum and $\Rightarrow$ is convergence in distribution. The GEV page defines $H$ . See Shared notation.

Threshold intuition¶

Block-maxima theory says that properly normalized maxima have only a small number of possible limiting shapes, collected in the Generalized Extreme-Value Distribution. The Pickands-Balkema-de Haan theorem says the matching threshold view has the same discipline: once we condition on being far enough into the tail, the remaining excess has an approximately Generalized Pareto Distribution.

For fat-tail work, the case $\xi>0$ is the main bridge. It corresponds to a regularly varying right tail with exponent $\alpha=1/\xi$ . The theorem does not say every high observation is generated by a clean Pareto law. It says the excess distribution over moving high thresholds has a universal limit shape under the same domain-of-attraction assumptions used in extreme-value theory.

Exact Pareto excess special case¶

The full Pickands-Balkema-de Haan theorem is cited here, not proved. The calculation below proves the exact Pareto special case: once the threshold is above the lower cutoff, the excess distribution is already generalized Pareto, not merely asymptotically close to it.

The full theorem is a classical result of Balkema and de Haan and, independently, Pickands. Its proof uses the equivalence between convergence of normalized maxima and convergence of normalized threshold excesses. We record the statement and use exact Pareto algebra as a checkable special case.

If $X$ is Pareto Type I with lower cutoff $x_m$ and tail exponent $\alpha$ , then for $u\ge x_m$ ,

\mathbb P(X-u>y\mid X>u) = \frac{\mathbb P(X>u+y)}{\mathbb P(X>u)} = \left(1+\frac{y}{u}\right)^{-\alpha}.

(5)

This is exactly a generalized Pareto survival function with

\xi=\frac1\alpha, \qquad \beta(u)=\frac{u}{\alpha}.

(6)

Thus exact Pareto tails do not merely approach the generalized Pareto form; they have it at every threshold above $x_m$ .

Exact threshold calibration¶

For $\alpha=1.7=17/10$ , the population excess parameters are $\xi=10/17$ and $\beta(u)=10u/17$ . The scaled excess $Z=(X-u)/u$ , conditional on $X>u$ , has survival $(1+z)^{-1.7}$ .

Threshold $u\ge x_m$	Shape $\xi$	Scale $\beta(u)$	$\mathbb P(Z>1\mid X>u)$
$x_m$	$10/17$	$10x_m/17$	2^-1.7
$2x_m$	$10/17$	$20x_m/17$	2^-1.7
$4x_m$	$10/17$	$40x_m/17$	2^-1.7

These are exact population identities, not fitted estimates. In finite data, estimated shape and scale fluctuate, and real tails may approach the GPD family slowly. Threshold selection examines that modeling choice; the theorem supplies no finite-sample cutoff.

Caveats¶

The theorem is asymptotic in the threshold. It does not choose the threshold for a finite dataset.
Dependence, seasonality, volatility clustering, rounding, truncation, and mixtures can all distort threshold exceedances.
A generalized Pareto fit is a tail model, not a proof that the full distribution is Pareto.
When $\xi\ge1$ , the fitted generalized Pareto model has no finite mean. When $\xi\ge1/2$ , it has no finite variance. These moment boundaries are often the practical reason the shape estimate matters.
Raising the threshold can reduce tail-approximation bias but leaves fewer exceedances; it does not guarantee a monotone improvement in a fitted model.

References¶

Balkema and de Haan, “Residual Life Time at Great Age” 1974.
Pickands, “Statistical Inference Using Extreme Order Statistics” 1975.
Embrechts, Klueppelberg, and Mikosch, Modelling Extremal Events 1997.
Coles, An Introduction to Statistical Modeling of Extreme Values 2001.

Backlinks¶

Depends on: Regular Variation, Extreme-Value Index, Generalized Pareto Distribution, Generalized Extreme-Value Distribution, and Shared notation.
Used by: Mean Excess Function and S&P 500 Tail Diagnostics.

References¶

Balkema, A. A., & de Haan, L. (1974). Residual Life Time at Great Age. The Annals of Probability, 2(5), 792–804. 10.1214/aop/1176996548
Pickands, J. (1975). Statistical Inference Using Extreme Order Statistics. The Annals of Statistics, 3(1), 119–131. 10.1214/aos/1176343003
Embrechts, P., Klüppelberg, C., & Mikosch, T. (1997). Modelling Extremal Events. Springer. 10.1007/978-3-642-33483-2
Coles, S. (2001). An Introduction to Statistical Modeling of Extreme Values. Springer. 10.1007/978-1-4471-3675-0