Frechet Limits for Heavy-Tailed Models

Why Heavy Tails Produce Fréchet Limits

A sample maximum is determined by observations far out in the upper tail. When that tail decreases as a power of the threshold, unusually large observations remain more likely than they would under an exponential-tail model. The maximum therefore grows at a polynomial rate with sample size, and its properly scaled distribution has a Fréchet limit.

The tail exponent controls both parts of this description: it determines the rate at which the maximum grows and the shape of its limiting distribution. It also determines the positive shape parameter when the limit is written in generalized extreme-value (GEV) form.

Definition: Let $X_1,X_2,\ldots$ be independent and identically distributed, nonnegative random variables, and write $M_n=\max(X_1,\ldots,X_n)$. Suppose there are constants $c>0$ and $\alpha>0$ such that $$ \frac{P(X_1>t)}{c t^{-\alpha}}\longrightarrow 1 \qquad\text{as }t\longrightarrow\infty. $$ This is a Pareto-type upper-tail condition. The number $\alpha$ is the tail exponent. Set $b_n=(cn)^{1/\alpha}$. The standard Fréchet distribution with exponent $\alpha$ has cumulative distribution function $$ H_\alpha(x)= \begin{cases} 0, & x\leq 0,\\ \exp(-x^{-\alpha}), & x>0. \end{cases} $$

The normalization $b_n$ is chosen so that an observation of order $b_n$ has a tail probability of order $1/n$. In fact, for each fixed $x>0$, the tail condition gives $nP(X_1>b_nx)$ a nonzero finite limit. This is the scale at which a sample of $n$ observations has a meaningful chance of containing an extreme value.

Deriving the Fréchet Limit

Theorem: Under the Pareto-type tail condition above, $M_n/b_n$ converges in distribution to the standard Fréchet distribution $H_\alpha$. In particular, for every $x>0$, $$ P(M_n/b_n\leq x)\longrightarrow \exp(-x^{-\alpha}). $$

Proof. Fix $x>0$, and put $u_n=P(X_1>b_nx)$. Since $b_nx$ tends to infinity, the tail condition implies $$ n u_n \longrightarrow n c(b_nx)^{-\alpha} = n c(cn)^{-1}x^{-\alpha} = x^{-\alpha}. $$ In this calculation the right-hand expression is the asymptotic equivalent of $n u_n$, so its limit is $x^{-\alpha}$. In particular, $u_n\to 0$. Independence gives $$ P(M_n/b_n\leq x) = P(X_1\leq b_nx,\ldots,X_n\leq b_nx) = (1-u_n)^n. $$ For $0\leq u<1$, the logarithm satisfies $\log(1-u)=-u+O(u^2)$ as $u\to0$. Here $nu_n\to x^{-\alpha}$, and $nu_n^2=(nu_n)u_n\to0$. Therefore $$ n\log(1-u_n)=-nu_n+O(nu_n^2)\longrightarrow -x^{-\alpha}, $$ so $(1-u_n)^n\to\exp(-x^{-\alpha})$. For $x<0$, nonnegativity gives $P(M_n/b_n\leq x)=0$. At $x=0$, the tail condition ensures $P(X_1>0)>0$, so $P(M_n=0)=P(X_1=0)^n\to0$. These limits give the stated distribution function at every continuity point of $H_\alpha$. $\square$

This argument makes the role of independence explicit: it turns the probability that one observation stays below a threshold into the $n$th power of that probability for the maximum. The tail condition supplies the limiting size of the single-observation exceedance probability.

Worked Example: A Pareto Maximum at a Fixed Sample Size

Suppose $X_1,\ldots,X_{100}$ are independent Pareto random variables with scale $2$ and exponent $\alpha=2$, so $$ P(X_i>t)=\left(\frac{2}{t}\right)^2,\qquad t\geq 2. $$ Find the exact probability that the maximum exceeds $40$, and compare it with the Fréchet limiting approximation.

Here the tail constant is $c=2^2=4$. Thus $b_n=(4n)^{1/2}=2\sqrt n$, and $b_{100}=20$. The threshold $40$ is twice this normalization, so its normalized value is $x=40/20=2$.

For one observation, $P(X_i>40)=(2/40)^2=1/400$. Independence gives the exact finite-sample probability $$ P(M_{100}>40)=1-\left(1-\frac{1}{400}\right)^{100} =1-\left(\frac{399}{400}\right)^{100} \approx 0.2214430. $$ The Fréchet limit instead gives $$ 1-H_2(2)=1-\exp(-2^{-2})=1-\exp(-1/4)\approx 0.2211992. $$ These values are close but not identical: the first is exact for $n=100$, while the second is the limiting probability as the sample size grows with the threshold kept at twice $b_n$.

Connection to the Positive GEV Shape

The Fréchet limit uses a positive variable $x$ and the form $\exp(-x^{-\alpha})$. The usual GEV convention instead writes the distribution in terms of a location and scale transformation, with the shape parameter denoted by $\xi$. To connect the forms, transform the normalized maximum by setting $$ Z_n=\frac{M_n-b_n}{b_n/\alpha} =\alpha\left(\frac{M_n}{b_n}-1\right). $$ If $z>-\alpha$, then $Z_n\leq z$ is equivalent to $M_n/b_n\leq 1+z/\alpha$. The Fréchet limit consequently gives $$ P(Z_n\leq z)\longrightarrow \exp\left[-(1+z/\alpha)^{-\alpha}\right]. $$

Result: In the GEV convention $$ G(z)=\exp\left[-(1+\xi z)^{-1/\xi}\right], \qquad 1+\xi z>0, $$ a Pareto-type tail with exponent $\alpha$ has positive shape parameter $\xi=1/\alpha$. Under the normalization used here, the corresponding GEV location is $0$ and scale is $1$. The support condition is $z>-\alpha$; the CDF is zero for $z\leq-\alpha$.

The reciprocal matters. A smaller tail exponent means a heavier tail, and it produces a larger positive GEV shape parameter. For example, exponent $\alpha=2$ corresponds to $\xi=1/2$, not $\xi=2$. The GEV shape describes the limiting distribution after location and scale normalization; it is not itself the power appearing in the survival probability.

Worked Example: A Lomax Tail and Its Normalization

Let $X_i$ have the Lomax survival function $P(X_i>t)=(1+t/3)^{-4}$ for $t\geq0$, independently across observations. Identify the normalization, the limiting distribution, and the limiting probability that the maximum exceeds twice the normalization.

As $t\to\infty$, $$ P(X_i>t) =(1+t/3)^{-4} =\left(\frac{3}{t+3}\right)^4 \sim 3^4t^{-4} =81t^{-4}. $$ Thus $c=81$ and $\alpha=4$. The normalization is $$ b_n=(81n)^{1/4}=3n^{1/4}. $$ For each $x>0$, $P(M_n/b_n\leq x)\to\exp(-x^{-4})$. At twice the normalization, $x=2$, so $$ P(M_n>2b_n)\longrightarrow 1-\exp(-2^{-4}) =1-\exp(-1/16) \approx 0.0605869. $$ The GEV shape associated with this limit is $\xi=1/4$. This example shows why the asymptotic tail constant must be included in $b_n$: using only $n^{1/4}$ would leave the limiting distribution on a different scale.

Worked Example: A Heavier Pareto Tail

Suppose $X_i$ are independent Pareto random variables with scale $2$ and exponent $\alpha=3/2$. Find the normalization and the limiting exceedance probability at four times that normalization.

The survival function is $(2/t)^{3/2}$ for $t\geq2$, so $c=2^{3/2}=2\sqrt{2}$. Therefore $$ b_n=(2\sqrt{2}\,n)^{2/3}=2n^{2/3}, $$ since $(2\sqrt{2})^{2/3}=2$. The threshold $4b_n$ has normalized value $x=4$. Hence $$ P(M_n>4b_n)\longrightarrow 1-\exp(-4^{-3/2}) =1-\exp(-1/8) \approx 0.1175031. $$ The corresponding GEV shape is $\xi=1/(3/2)=2/3$. Compared with the previous example’s exponent $4$, this smaller tail exponent yields a larger GEV shape, reflecting the heavier upper tail.

Using the Limit Carefully

For a threshold proportional to the normalization, the general limiting exceedance probability is $$ P(M_n>b_nx)\longrightarrow 1-\exp(-x^{-\alpha}),\qquad x>0. $$ This is useful when comparing extreme risks across sample sizes or when a direct finite-sample calculation is unwieldy. It is an approximation for large $n$, not an identity at a particular finite $n$. For an exact Pareto model, the finite-sample probability can often be calculated directly, as in the first example.

The result depends on the stated assumptions. The observations are independent and identically distributed, the variables are nonnegative, and their upper tail has the stated asymptotic form with positive $c$ and $\alpha$. Changing the tail model, dependence structure, or normalization can change the limit or invalidate this derivation. In applications, a fitted heavy-tail model should therefore be accompanied by a check that its tail behavior is plausible over the range relevant to the largest observations.

Common Mistakes

Using the exponent as the GEV shape. The shape is $\xi=1/\alpha$, the reciprocal of the Pareto tail exponent.
Dropping the tail constant. If $P(X>t)\sim ct^{-\alpha}$, the normalization is $b_n=(cn)^{1/\alpha}$, not necessarily $n^{1/\alpha}$.
Treating a limit as an exact finite-sample probability. The Fréchet formula describes a limit as $n\to\infty$. At finite $n$, use the exact distribution when available, or label the limit as an approximation.
Applying the formula at an unscaled threshold without checking its size. The stated limit concerns thresholds $b_nx$ for fixed $x>0$, not every fixed threshold as $n$ grows.

Key takeaway: If the upper tail is asymptotic to $ct^{-\alpha}$, scale the maximum by $(cn)^{1/\alpha}$. The normalized maximum has CDF $\exp(-x^{-\alpha})$ for $x>0$, and the associated positive GEV shape is the reciprocal tail exponent, $\xi=1/\alpha$.

Check Your Understanding

Use the definitions and derivations above to answer the following questions.

If $P(X>t)\sim 5t^{-2}$, what normalization $b_n$ gives the standard Fréchet limit?
For a Pareto-type tail with exponent $\alpha=3$, what is the positive GEV shape parameter?
Under a Fréchet limit with exponent $\alpha=2$, what is the limiting probability that $M_n$ exceeds $3b_n$?
Why does the derivation of the maximum distribution use the $n$th power of a one-observation probability?
In the Lomax example, why is the tail constant $81$, even though the scale parameter is $3$?