Max-Stability and Block-Maximum Scaling

Why Max-Stability Matters

In block-maxima analysis, each block contributes one observation: the largest value seen in that block. A useful model should describe how the distribution changes when the block size changes. The generalized extreme-value (GEV) family has a particularly convenient property: the maximum of $m$ independent observations from a GEV distribution is still GEV, with the same shape parameter and explicitly changed location and scale.

This property is called max-stability. It explains why GEV models can be used at different block sizes and gives an exact scaling rule, rather than merely an approximation. The independence and common distribution assumptions matter: the result below concerns independent, identically distributed observations from one GEV law.

GEV Distribution and Max-Stability

Definition: For location $\mu\in\mathbb{R}$, scale $\sigma>0$, and shape $\xi\in\mathbb{R}$, the GEV distribution has cumulative distribution function $$ H_{\mu,\sigma,\xi}(x)= \begin{cases} \exp\!\left\{-\left[1+\xi(x-\mu)/\sigma\right]^{-1/\xi}\right\}, & \xi\ne 0,\quad 1+\xi(x-\mu)/\sigma>0,\\[4pt] \exp\!\left\{-\exp\!\left[-(x-\mu)/\sigma\right]\right\}, & \xi=0,\quad x\in\mathbb{R}. \end{cases} $$ For $\xi\ne 0$, the CDF is zero where $1+\xi(x-\mu)/\sigma\leq 0$ if $\xi>0$, and one there if $\xi<0$. The support has a finite upper endpoint when $\xi<0$, and a finite lower endpoint when $\xi>0$.

Let $X_1,\ldots,X_m$ be independent and identically distributed with this CDF, and let $M_m=\max(X_1,\ldots,X_m)$. Independence gives the starting point:

P(M_m\leq x)=P(X_1\leq x,\ldots,X_m\leq x) =H_{\mu,\sigma,\xi}(x)^m.

Max-stability says that this maximum can be transformed by a positive scale and a location shift to recover the original GEV distribution.

Theorem: If $X_1,\ldots,X_m$ are independent and identically distributed as $\operatorname{GEV}(\mu,\sigma,\xi)$, for an integer $m\geq1$, define $$ b_m= \begin{cases} m^\xi,&\xi\ne0,\\ 1,&\xi=0, \end{cases} \qquad a_m= \begin{cases} \dfrac{\sigma}{\xi}(m^\xi-1),&\xi\ne0,\\[5pt] \sigma\log m,&\xi=0. \end{cases} $$ Then $M_m\sim\operatorname{GEV}(\mu+a_m,b_m\sigma,\xi)$. Equivalently, $$ \frac{M_m-d_m}{b_m}\sim\operatorname{GEV}(\mu,\sigma,\xi), \qquad d_m=a_m+(1-b_m)\mu. $$

The two forms express the same result. The maximum itself has location $\mu+a_m$ and scale $b_m\sigma$. To return to location $\mu$ after dividing by $b_m$, the centering constant must be $d_m$, not generally $a_m$. In particular, subtracting $a_m$ alone and dividing by $b_m$ gives location $(\mu+a_m-a_m)/b_m=\mu/b_m$, which need not equal $\mu$.

Deriving the Block-Maximum Parameters

First suppose $\xi\ne0$. Write $A(x)=1+\xi(x-\mu)/\sigma$. On the support, the GEV CDF raised to the $m$th power is

H_{\mu,\sigma,\xi}(x)^m =\exp\{-m A(x)^{-1/\xi}\}.

Choose $b_m=m^\xi$ and $\mu_m=\mu+a_m$, where $a_m=\sigma(m^\xi-1)/\xi$. The quantity inside the GEV power for parameters $(\mu_m,b_m\sigma,\xi)$ becomes

1+\frac{\xi(x-\mu_m)}{b_m\sigma} =m^{-\xi}\left[1+\frac{\xi(x-\mu)}{\sigma}\right] =m^{-\xi}A(x).

Because $m^{-\xi}>0$, this transformed quantity is positive exactly where $A(x)$ is positive. On that support,

\left[m^{-\xi}A(x)\right]^{-1/\xi} =m A(x)^{-1/\xi}.

Thus the CDF with location $\mu_m$ and scale $b_m\sigma$ equals $H_{\mu,\sigma,\xi}(x)^m$. For $\xi>0$, both CDFs are zero below their common lower endpoint; for $\xi<0$, both are one above their common upper endpoint. Thus the equality holds for every $x$. This proves the stated distribution of $M_m$ when $\xi\ne0$.

When $\xi=0$, the same conclusion follows directly from the Gumbel CDF:

H_{\mu,\sigma,0}(x)^m =\exp\!\left\{-m\exp\!\left[-\frac{x-\mu}{\sigma}\right]\right\} =\exp\!\left\{-\exp\!\left[-\frac{x-(\mu+\sigma\log m)}{\sigma}\right]\right\}.

Therefore $M_m\sim\operatorname{GEV}(\mu+\sigma\log m,\sigma,0)$, as claimed. Finally, if $M_m$ has location $\mu+a_m$ and scale $b_m\sigma$, applying the affine transformation $(M_m-d_m)/b_m$ gives location $(\mu+a_m-d_m)/b_m=\mu$, scale $b_m\sigma/b_m=\sigma$, and unchanged shape $\xi$. This proves the normalization statement as well. $\square$

Worked Scaling Calculations

Worked Example: Gumbel Blocks of Size Ten

Suppose $X_i\sim\operatorname{GEV}(2,3,0)$ independently, and take blocks of size $m=10$. For the Gumbel case, $a_m=\sigma\log m$ and $b_m=1$. Therefore

a_{10}=3\log 10,\qquad b_{10}=1, \qquad M_{10}\sim\operatorname{GEV}(2+3\log 10,3,0).

The normalized maximum is $M_{10}-3\log 10$, and it has the original $\operatorname{GEV}(2,3,0)$ distribution. For an exact CDF check, take the threshold $2$ for the normalized maximum. It corresponds to the threshold $2+3\log 10$ for $M_{10}$. The single-observation CDF at that threshold is

H_{2,3,0}(2+3\log 10) =\exp\{-\exp(-\log 10)\} =\exp(-1/10).

Consequently, $$ P(M_{10}-3\log 10\leq2) =\left[\exp(-1/10)\right]^{10} =\exp(-1), $$ which is exactly $H_{2,3,0}(2)$. No rounded decimal approximation is needed for this equality.

Worked Example: Fréchet-Type Shape and a Four-Observation Block

Let $X_i\sim\operatorname{GEV}(1,2,0.5)$ independently and let $m=4$. Since $4^{0.5}=2$,

b_4=2,\qquad a_4=\frac{2}{0.5}(2-1)=4.

Thus $M_4\sim\operatorname{GEV}(5,4,0.5)$. The centering constant that returns the original parameters after scaling is

d_4=a_4+(1-b_4)\mu=4+(1-2)(1)=3.

It follows that $(M_4-3)/2\sim\operatorname{GEV}(1,2,0.5)$. Check at normalized threshold $1$, which corresponds to $M_4\leq 3+2(1)=5$. The original CDF there is

H_{1,2,0.5}(5) =\exp\!\left\{-\left[1+0.5(5-1)/2\right]^{-2}\right\} =\exp(-1/4).

Raising this to the fourth power gives $P(M_4\leq5)=\exp(-1)$. The transformed GEV CDF at $1$ is also $\exp(-1)$, verifying the normalization. If one had instead used $(M_4-a_4)/b_4$, its location would be $1/2$, not $1$.

Worked Example: Reverse-Weibull Shape and a Finite Endpoint

Suppose $X_i\sim\operatorname{GEV}(10,2,-0.5)$ independently, with block size $m=9$. Here $9^{-0.5}=1/3$, so

b_9=\frac13,\qquad a_9=\frac{2}{-0.5}\left(\frac13-1\right)=\frac83.

The block maximum therefore has location $10+8/3=38/3$, scale $2/3$, and shape $-0.5$. Its upper endpoint is

\frac{38}{3}-\frac{2/3}{-0.5} =\frac{38}{3}+\frac43 =14.

This agrees with the endpoint of each observation, $10-2/(-0.5)=14$: taking a maximum does not move a finite upper endpoint. As a direct CDF check at $x=12$, the original support term is $1-0.5(12-10)/2=0.5$, so

H_{10,2,-0.5}(12)=\exp\{-(0.5)^2\}=\exp(-1/4), \qquad P(M_9\leq12)=\exp(-9/4).

Using the block parameters gives support term $1-0.5(12-38/3)/(2/3)=1.5$. Its power is $-1/\xi=2$, hence $H_{38/3,2/3,-0.5}(12)=\exp\{-(1.5)^2\}=\exp(-9/4)$, as required.

Consistency When Blocks Are Combined

The scaling rules agree whether a sample is treated as one large block or as smaller blocks combined together. For $\xi\ne0$, the formulas imply

b_m b_k=m^\xi k^\xi=(mk)^\xi=b_{mk}, \qquad a_m+b_m a_k =\frac{\sigma}{\xi}\{m^\xi-1+m^\xi(k^\xi-1)\} =a_{mk}.

For $\xi=0$, $b_m=b_k=b_{mk}=1$, and

a_m+b_m a_k=\sigma\log m+\sigma\log k=\sigma\log(mk)=a_{mk}.

To interpret these identities, first take maxima within $m$-observation blocks. Those maxima have location $\mu+a_m$, scale $b_m\sigma$, and shape $\xi$. Taking the maximum of $k$ such independent block maxima applies the same rule with scale $b_m\sigma$: the additional location shift is $b_m a_k$, and the additional scale multiplier is $b_k$. The resulting location shift is $a_m+b_m a_k=a_{mk}$, and the scale multiplier is $b_m b_k=b_{mk}$. This matches taking the maximum of all $mk$ original observations at once.

Common Mistake and Practical Interpretation

A common mistake is to report $(M_m-a_m)/b_m$ as having the original location $\mu$. That is not generally true: the maximum has location $\mu+a_m$, so the location after that transformation is $\mu/b_m$. The correct normalization is

\frac{M_m-\{a_m+(1-b_m)\mu\}}{b_m} =\mu+\frac{M_m-\mu-a_m}{b_m} \sim\operatorname{GEV}(\mu,\sigma,\xi).

The simpler expression $(M_m-a_m)/b_m$ works only when $\mu=0$ or $b_m=1$. The distinction is especially relevant when the shape is nonzero, because then $b_m=m^\xi$ generally differs from one.

The unchanged shape parameter means that block size affects the location and scale but not the GEV tail type. For $\xi>0$, larger blocks increase the scale; for $\xi<0$, the scale multiplier is less than one, while the finite upper endpoint remains fixed. For $\xi=0$, scale stays constant and location increases by $\sigma\log m$. These are exact consequences of the GEV model, not claims that arbitrary data have an exact GEV law.

Check Your Understanding

Use the formulas and the support conditions above to answer the following.

For a GEV distribution with $\xi\ne0$, derive the block scale multiplier $b_m$ by matching the support term in $H(x)^m$.
For $\operatorname{GEV}(0,2,0.25)$ and block size $16$, calculate $a_m$, $b_m$, and the block-maximum location and scale.
Why does $(M_m-a_m)/b_m$ generally fail to have location $\mu$? State the correct centering constant.
For $\xi<0$, explain algebraically why the upper endpoint of the block maximum is the same as the upper endpoint of one observation.
Show that the location and scale transformations for block sizes $m$ and $k$ compose to the transformations for block size $mk$ when $\xi=0$.