# Chaitanya's Random Pages

## June 10, 2012

### Outline proof of the extreme value theorem in statistics

Filed under: mathematics — ckrao @ 9:15 am

Recently I read a proof in [1] of the main theorem of extreme value statistics: the Fisher–Tippett–Gnedenko theorem. In this post I give an outline.

Here we are interested in the maximum of many independent and identically distributed random variables $X_i$ with distribution function $F$. Let $x^* = \sup \{x: F(x) < 1\}$ which may be infinite. Then as $n \rightarrow \infty$,

${\rm Pr}(\max(X_1, X_2, \ldots X_n) \leq x) = F^n(x)$ tends to 0 if $x < x^*$ and 1 if $x > x^*$.

Therefore $\displaystyle \max(X_1, X_2, \ldots, X_n)$ converges in probability to $x^*$ as $n \rightarrow \infty$.

In order to avoid this degenerate limiting distribution for all extreme value distributions, it is necessary to normalise the distribution.

To this end, suppose there exist real numbers $a_n >0, b_n$ such that

$\displaystyle \frac{\max(X_1, X_2, \ldots, X_n) - b_n}{a_n}$ approaches a non-degenerate limiting distribution.

In other words, there exists a distribution function $G(x)$ such that

$\displaystyle \lim_{n \rightarrow \infty} F^n(a_n x + b_n) = G(x).$

Taking logarithms of both sides, this is equivalent to

$\displaystyle \lim_{n \rightarrow \infty} n \log F(a_n x + b_n) = \log G(x)$

This requires $F(a_n x + b_n) \rightarrow 1$ as $n \rightarrow \infty$. Using $\log x \approx x -1$ for $x$ close to 1, the above is also equivalent to

$\displaystyle \lim_{n \rightarrow \infty} \frac{1}{n(1 - F(a_n x + b_n) } = \frac{1}{- \log G(x)}. \quad \quad (1)$

Next we use the following definition.

A non-decreasing function $f$ has left-continuous inverse $f^{\leftarrow}$ defined by

$\displaystyle f^{\leftarrow}(x) := \inf\{y: f(y) > x\}.$

One can use this definition to prove

Lemma 1: If $f_n(x) \rightarrow g(x)$ for non-decreasing functions $f_n$ then for each continuity point $x$ we have $f_n^{\leftarrow}(x) \rightarrow g^{\leftarrow}(x)$.

Next we claim that (1) is equivalent to

$\displaystyle \frac{U(nx) - b_n}{a_n} \rightarrow G^{\leftarrow}\left(e^{-1/x}\right) \quad \quad (2)$

where $U$ is the left-continuous inverse of $\frac{1}{1-F}$. To see this, let $V(x) = \frac{1}{1 - F(x)}$. Then by the definition of $U$, $U(x) = \inf \{ y: V(y) \geq x \}$. Then for any $n \in \mathbb{N}$,

$U(nx) = \inf \left\{ y: \frac{1}{n(1-F(y))} \geq x \right\}$

and so

$\begin{array}{lcl} \frac{U(nx) - b_n}{a_n} &=& \inf \left\{\frac{y - b_n}{a_n}: \frac{1}{n(1 - F(y))} \geq x \right\}\\&=& \inf \left\{z: \frac{1}{n(1 - F(a_n z + b_n))} \geq x \right\}\end{array}$

By (1) and the lemma as $n \rightarrow \infty$ this tends to

$\begin{array}{lcl} & & \inf \left\{ z: \frac{1}{-\log G(z)} \geq x \right\}\\& = & \inf \left\{ z: \log G(z) \geq \frac{-1}{x} \right\} \\ &=& \inf \left\{ z: G(z) \geq e^{-1/x} \right\}\\ &=& G^{\leftarrow} \left( e^{-1/x} \right), \end{array}$

proving the claim. We can also write

$\displaystyle \lim_{t \rightarrow \infty} \frac{U(tx) - b(t)}{a(t)} = G^{\leftarrow}\left(e^{-1/x}\right) =: D(x) \quad \quad (3)$

where $a(t) := a_{[t]}, b(t) := b_{[t]}$ and $[t]$ denotes the integer part of $t$.

We are now ready to prove the main theorem of extreme value theory.

Theorem (Fisher, Tippet, Gnedenko):

$\displaystyle G_{\gamma}(ax + b) = \exp\left( -(1 + \gamma x)^{-1/\gamma} \right), \quad 1 + \gamma x > 0, \quad \gamma \in \mathbb{R}$

where the right side is equal to its limiting value $\exp \left( - e^{-x}\right)$ if $\gamma = 0$.

Proof:

This will involve numerous substitutions but the main idea is to arrive at a differential equation that can be solved to obtain the above. Suppose $1$ is a continuity point of $D$. Then for any continuity point $x > 0$,

$\displaystyle \lim_{t \rightarrow \infty} \frac{U(tx) - U(t)}{a(t)} = D(x) - D(1) =: E(x). \quad \quad (4)$

We can write

$\displaystyle \frac{ U(txy) - U(t)}{a(t)} = \frac{U(txy) - U(ty)}{a(ty)} \frac{a(ty)}{a(t)} + \frac{U(ty) - U(t)}{a(t)}.$

The claim is that both $\frac{a(ty)}{a(t)}$ and $\frac{U(ty) - U(t)}{a(t)}$ have limits as $t \rightarrow \infty$. If they had more than one limit point, say $A_1, A_2, B_2, B_2$ then for $i = 1, 2$ (4) in the limit $t \rightarrow \infty$ gives us

$E(xy) = E(x) A_i + B_i$.

Subtracting gives $E(x) (A_1 - A_2) = B_2 - B_1$ which implies $A_1 = A_2, B_2 = B_1$ as we know $E(x)$ is non-constant (since we seek a non-degenerate solution).

We conclude that

$\displaystyle E(xy) = E(x) A(y) + E(y). \quad \quad (5)$

This is a functional equation that we wish to solve. We let $s:= \log x, t := \log y, H(x) := E(e^x)$ to obtain

$\displaystyle H(t+ s) = H(s) A(e^t) + H(t),$

which using $H(0) = E(1) = 0$ implies

$\displaystyle \frac{H(t+s) - H(t)}{s} = \frac{H(s) - H(0)}{s} A(e^t). \quad \quad (6)$

Since $H$ is monotone (following from the monotonicity of $D$), it is differentiable at some $t$. By (6) it is differentiable at all $t$. Indeed from (6) we obtain

$\displaystyle H'(t) = H'(0) A(e^t) \neq 0. \quad \quad (7)$

Let $Q(t) = H(t)/H'(0)$. Then $Q(0) = 0, Q'(0) = 1$.

From (6) and (7),

$\displaystyle Q(t+s) - Q(t) = Q(s) A(e^t) = Q(s) H'(t)/H'(0) = Q(s) Q'(t). \quad \quad (8)$

Similarly, $Q(s+t) - Q(s) = Q(t) Q'(s)$ and upon subtraction from (8) we obtain $Q(s) - Q(s)Q'(t) = Q(t) - Q(t)Q'(s)$ from which

$\displaystyle \frac{Q(s)}{s} (Q'(t) - 1) = Q(t) \frac{Q'(s) - 1}{s}.$

Taking the limit as $s \rightarrow 0$ and using $Q'(0) = H'(0)/H'(0) = 1$ gives the following differential equation for $Q$.

$\displaystyle Q'(t) -1 = Q(t) Q''(0), Q(0) = 0, Q'(0) = 1 \quad \quad (9)$

To solve (9), differentiate both sides with respect to $t$: from $Q''(t) = Q'(t) Q''(0)$ we see that

$\displaystyle (\log Q')' (t) = Q''(0) =: \gamma$.

Hence $Q'(t) = e^{\gamma t}$ and since $Q(0) = 0$, $Q(t) = \int_0^t e^{\gamma s}\ ds$.

Recalling $Q(t) = H(t)/H'(0)$, this leads to $H(t) = H'(0) \frac{e^{\gamma t} - 1}{\gamma}$.

Recalling $H(x) := E(e^x) := D(e^x) - D(1)$ this means

$D(t) = D(1) + H'(0) \frac{t^{\gamma} - 1}{\gamma}.$

Taking the left-continuous inverse of both sides,

$D^{\leftarrow}(x) = \left(1 + \gamma \frac{x - D(1)}{H'(0)} \right)^{1/\gamma}.$

Since $D(x) := G^{\leftarrow}\left(e^{-1/x}\right)$, $D^{\leftarrow}(x) = \frac{-1}{\log G(x)}$.

Hence

$\displaystyle G(x) = e^{-1/D^{\leftarrow}(x) }.$

In other words,

$\displaystyle G(H'(0) y + D(1) ) = \exp \left(- (1 + \gamma y)^{-1/\gamma} \right),$

where $\gamma = Q''(0) = H''(0)/H(0)$.

If 1 is not a continuity point, we repeat the above proof with $U(t)$ replaced by $U(tx_0)$ where $x_0$ is a continuity point. This completes the proof.

This generalised extreme value distribution encaptures three distributions depending on the nature of the tail of the original distribution $X_i$:

• Type I – $\gamma = 0$: Gumbel (double exponential) distribution (exponential tail – e.g. normal or exponential distribution)
• Type II – $\gamma > 0$: Fréchet distribution (polynomial tail – e.g. power law distribution to model extreme flood levels or high incomes)
• Type III – $\gamma < 0$: reverse-Weibull distribution (finite maximum – e.g. uniform distribution)

Sample density functions are plotted below for specific values of $\gamma$.

#### Reference

[1] L. De Haan, A. Ferreira, Extreme Value Theory: An Introduction, Springer, 2006.