Implicit Max-Stable Laws

• Implicit max-stable laws are defined through a loss function that selects the maximizing observation, generalizing classical coordinatewise maxima.
• They leverage regular variation and explicit spectral representations to characterize limit laws and stochastic models under functional maximization.
• The framework extends to infinite-dimensional settings and order statistics, offering practical applications in simulation and multivariate extremes analysis.

Implicit max-stable laws generalize the concept of max-stability in extreme value theory, accounting for random vectors selected via a maximizing functional rather than coordinatewise maxima. In implicit extreme value theory, the "extremal" event is defined through a loss function $f:\mathbb R^d \to [0,\infty)$, and focus is placed on those observations in a sequence that maximize $f(X_i)$. This framework yields a rich class of limit laws—called implicit max-stable laws—exhibiting max-stability with respect to a generalized $\vee_f$ operation. The theory encompasses multivariate and function-valued settings, supports explicit stochastic and spectral representations, and provides deep connections to regular variation, Poisson point processes, and max-stable random sup-measures (Scheffler et al., 2014, Kremer, 2019, Mai et al., 2018).

1. Foundations and Definitions

The implicit max-stable framework is built upon the analysis of the sample element $X_{k(n)}$ from a sequence of i.i.d. $\mathbb R^d$-valued random vectors $X_1,\ldots,X_n$, chosen such that $f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$. The law of $X_{k(n)}$ has density

$$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$

where $G(y) = P(f(X) \leq y)$. An implicit extreme value limit arises if there exist normalizing constants $f(X_i)$0 such that $f(X_i)$1 converges in distribution to a nondegenerate limit $f(X_i)$2, called an $f(X_i)$3-implicit extreme value law (Scheffler et al., 2014).

A random vector $f(X_i)$4 is $f(X_i)$5-implicit max-stable if for any $f(X_i)$6 there exists $f(X_i)$7 such that for i.i.d. $f(X_i)$8,

$f(X_i)$9

The classical case is recovered when $\vee_f$0 is a projection or a norm, but implicit max-stability is considerably more general and encompasses arbitrary 1-homogeneous, continuous or measurable loss functionals.

2. Regular Variation and Spectral Structure

Structural analysis employs regular variation on cones. Let $\vee_f$1 and assume $\vee_f$2 is regularly varying on $\vee_f$3 with index $\vee_f$4 and Radon measure $\vee_f$5, i.e.,

$\vee_f$6

on $\vee_f$7, with $\vee_f$8. This structure supports a polar (radial-angular) decomposition: for a continuous 1-homogeneous $\vee_f$9 (often $X_{k(n)}$0), define $X_{k(n)}$1. Then

$X_{k(n)}$2

with finite spectral measure $X_{k(n)}$3 (Scheffler et al., 2014).

Regular variation confirms that the event $X_{k(n)}$4 asymptotically has a probability governed by the measure $X_{k(n)}$5, establishing precise control on the asymptotics of implicit maxima.

3. Limit Theorems and Law Characterization

Under mild conditions on $X_{k(n)}$6 and $X_{k(n)}$7, the normalized implicit maximum converges:

$X_{k(n)}$8

where the nondegenerate limit $X_{k(n)}$9 has density $\mathbb R^d$0, with $\mathbb R^d$1. This family of limit laws inherits many key properties of max-stable laws, but in the context of the maximization via $\mathbb R^d$2 rather than via coordinatewise maxima.

The limit law admits an explicit stochastic representation:

$\mathbb R^d$3

with $\mathbb R^d$4 standard $\mathbb R^d$5-Fréchet ($\mathbb R^d$6), $\mathbb R^d$7 on $\mathbb R^d$8, and tilting function $\mathbb R^d$9. The spectral measure $X_1,\ldots,X_n$0 is normalized (Scheffler et al., 2014).

Uniqueness is established: every $X_1,\ldots,X_n$1-implicit max-stable law arises as the $X_1,\ldots,X_n$2-implicit extreme value law for some homogeneous $X_1,\ldots,X_n$3; equivalence of implicit max-stability and extreme value limit laws is established under continuity and homogeneity of $X_1,\ldots,X_n$4.

4. Extremal Integrals and Sup-Measures

The implicit max-stable property extends naturally to infinite-dimensional settings via sup-measures and extremal integrals (Kremer, 2019). For a continuous, 1-homogeneous $X_1,\ldots,X_n$5 on $X_1,\ldots,X_n$6, and a $X_1,\ldots,X_n$7-finite measure space $X_1,\ldots,X_n$8, an $X_1,\ldots,X_n$9-implicit sup-measure $f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$0 satisfies:

• Disjointness: $f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$1 are independent for disjoint $f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$2.
• $f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$3-additivity: $f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$4, with $f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$5 maximizing $f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$6.
• Margins: $f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$7.

Define the implicit extremal integral for $f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$8 as

$f(X_{k(n)}) = \max_{1\leq i\leq n} f(X_i)$9

with $X_{k(n)}$0. The resulting stochastic process $X_{k(n)}$1 is $X_{k(n)}$2-implicit max-stable, and finite-dimensional marginals satisfy:

$X_{k(n)}$3

Properties of independence, max-linearity under $X_{k(n)}$4, and monotonicity with respect to $X_{k(n)}$5 are inherited, distinguishing implicit extremal integrals from classical theory (Kremer, 2019).

5. Implicit Order Statistics and Point Process Limits

The implicit theory supports the analysis of order statistics: the $X_{k(n)}$6 largest arguments under $X_{k(n)}$7 (the $X_{k(n)}$8-largest order statistics) have joint asymptotic distributions described by Poisson point processes. Specifically,

$X_{k(n)}$9

converges to a Poisson process on $$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$0 with intensity $$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$1. The joint law of $$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$2 top implicit order statistics converges to that of $$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$3 for $$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$4, where $$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$5 is an ordered sequence of unit Poisson points on $$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$6, and $$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$7 are i.i.d. from the normalized spectral measure (Scheffler et al., 2014). This generalizes Gnedenko's classical order statistics results to the implicit framework.

6. Illustrative Models and Applications

Examples of implicit max-stable laws span Pareto–Dirichlet models, classical multivariate extremes, and hidden regular variation:

• The Pareto–Dirichlet model results when $$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$8 are independent Pareto-distributed components and $$P\{X_{k(n)} \in dx\} = n \, G(f(x))^{n-1} \, P_X(dx),$$9 acts as a harmonic mean; explicit Dirichlet-type limit laws arise (Scheffler et al., 2014).
• For elliptical losses, $G(y) = P(f(X) \leq y)$0, the framework recovers Gaussian copula features in the implicit extremes.
• For $G(y) = P(f(X) \leq y)$1 constructed from a bivariate Gaussian copula with Pareto margins, hidden regular variation yields nontrivial implicit extreme value distributions corresponding to the loss $G(y) = P(f(X) \leq y)$2.

Applications of implicit max-stable laws arise in simulation of max-stable random sequences and copulas (Mai et al., 2018). Exchangeable max-stable sequences can be constructed using strongly infinitely divisible time processes, and their extreme-value copulas can be simulated exactly via spectral (Pickands) representations and algorithms such as the Dombry–Schlather procedure.

7. Relationship to Classical and Boolean Extreme Value Theory

Implicit max-stable laws subsume the classical max-stable laws as a special case. When $G(y) = P(f(X) \leq y)$3 (scalar case) or $G(y) = P(f(X) \leq y)$4, all the implicit theory reduces to known results for the scalar or max-norm case. Boolean extreme value theory, by contrast, yields max-stable laws corresponding to Dagum (log–logistic) distributions, with only a single nondegenerate fixed point family for nonnegative variables—a stark contrast with the spectrum of Fréchet, Gumbel, and Weibull families in classical theory (Vargas et al., 2017).

The theory of implicit max-stable laws, through its generality and deep connection to functional analysis, spectral theory, and Poisson point process limits, forms a modern cornerstone of extreme value analysis under loss-driven or functional maximization criteria (Scheffler et al., 2014, Kremer, 2019, Mai et al., 2018, Vargas et al., 2017).