On extremes for Gaussian subordination

Basic Information

• Paper ID: 2510.10578
• Title: On extremes for Gaussian subordination
• Authors: Shuyang Bai (University of Georgia), Marie-Christine Düker (FAU Erlangen-Nürnberg)
• Classification: math.PR math.ST stat.TH
• Publication Date: October 14, 2025
• Paper Link: https://arxiv.org/abs/2510.10578

Abstract

This paper investigates extreme value theory for processes obtained by applying transformations to stationary Gaussian processes (termed Gaussian subordination processes). The main contributions include: first, an improvement of the Sly and Heyde (2008) methodology, allowing the covariance decay of the underlying Gaussian process to be slower than any polynomial rate, nearly matching the Berman condition; second, an extension of the theory to multivariate settings, where both the subordination process and the underlying Gaussian process can be vector-valued with finite-dimensional transformations; establishment of weak convergence of point processes constructed from subordinated Gaussian processes, from which multivariate extreme value limit theorems are derived.

Research Background and Motivation

Problem Background

1. Core Problem: Investigation of extreme value behavior of Gaussian subordination processes, specifically the extreme value properties of processes $Y_k = G(X_k)$ obtained by applying measurable transformations $G$ to stationary Gaussian processes $\{X_k\}$.
2. Significance:
   • Gaussian subordination models hold important positions in time series analysis, capable of flexibly generating various marginal distributions (including heavy-tailed distributions)
   • Widely applied in extreme value analysis in finance, insurance, meteorology, and other fields
   • Enables simultaneous modeling of temporal and cross-sectional dependence
3. Limitations of Existing Methods:
   • Results by Sly and Heyde (2008) require covariance decay at polynomial rates
   • Lack of systematic theory for multivariate cases
   • Incomplete treatment of infinite-dimensional transformations
4. Research Motivation:
   • Extend classical results by Berman (1964) on extreme value theory for Gaussian processes to subordination processes
   • Handle Gaussian processes with long-memory properties (slowly decaying covariance)
   • Establish a complete multivariate extreme value theory framework

Core Contributions

1. Improved Covariance Decay Conditions: Generalization of the Sly-Heyde method to covariance decay conditions $\gamma(n) = o(1/\log(n))$, nearly achieving optimality of the Berman condition
2. Multivariate Extreme Value Theory: Establishment of complete extreme value theory for multivariate Gaussian subordination processes, including characterization of multivariate extreme value indices
3. Point Process Convergence Theorem: Proof of weak convergence of point processes constructed from Gaussian subordination processes (Theorem 2.2)
4. Extreme Value Independence Results: Demonstration that any bivariate random vector transformed from jointly Gaussian vectors with non-perfect correlation is always extremally independent
5. m-Extremal Dependence Concept: Introduction and development of m-extremal dependence, generalizing classical m-dependence
6. Infinite-Dimensional Extension: Extension of finite-dimensional transformation results to infinite-dimensional settings through approximation arguments
7. Application Examples: Establishment of limit theorems for multivariate moving maximum processes driven by subordinated Gaussian processes

Methodology Details

Basic Model Setup

Gaussian Subordination Process Model: Let $\{X_k = (X_{k,1}, \ldots, X_{k,d_0})'\}_{k \in \mathbb{Z}}$ be a $d_0$-dimensional stationary Gaussian process. Define a $(m+1)$-moving window multivariate Gaussian subordination model:

$Y_{k,i} = G_i(X_k, X_{k-1}, \ldots, X_{k-m}), \quad k \in \mathbb{Z}, i = 1, \ldots, d$

where $G_i: \mathbb{R}^{d_0 \times (m+1)} \to \mathbb{R}$ is a measurable function.

Underlying Gaussian Process Representation: Assume the latent process $\{X_k\}$ has a causal linear representation: $X_k = \sum_{l=0}^{\infty} \Psi_l \varepsilon_{k-l}$

where $\{\varepsilon_j\}$ are i.i.d. standard Gaussian vectors, and the matrix sequence $\{\Psi_l\}$ satisfies the key condition: $\psi_{ij,l} = o(l^{-1/2}\log(l)^{-1}), \quad \text{as } l \to \infty$

Core Technical Innovations

1. Application of Hypercontractivity Utilization of hypercontractivity of Gaussian Hilbert spaces and the Mehler transform: $M_a(X) = \sum_{k=0}^{\infty} a^k \pi_k(X)$

where $\pi_k$ is the projection onto the $k$-th order Wiener chaos.

2. Control of Canonical Correlations For Gaussian Hilbert subspaces $H_1, H_2$, their canonical correlation is: $r(H_1, H_2) = \sup_{X_1 \in H_1, X_2 \in H_2} \text{Corr}(X_1, X_2)$

Key inequality (Proposition 3.3): $E|X_1 X_2| \leq \|X_1\|_{1+\rho} \|X_2\|_{1+\rho}$

3. Point Process Method Construction of point processes: $\xi_n = \sum_{j=1}^{\infty} \delta_{(Y_n(j), j(r+p)/n)}$

where $Y_n(j)$ is the transformation of data blocks, $r$ is the block length, and $p \geq m$ is the spacing.

Main Theorems

Theorem 2.1 (Multivariate Extreme Value Limit Theorem): Under appropriate conditions, $\lim_{n \to \infty} P(M_n \leq u_n(\tau)) = G(\tau)^{\theta(\tau)}$

where $M_n = (M_{n,1}, \ldots, M_{n,d})$ is the componentwise maximum, and $\theta(\tau)$ is the multivariate extreme value index.

Theorem 2.2 (Point Process Convergence): $\xi_n \Rightarrow \xi$ where $\xi$ is a Poisson point process with mean measure $\mu_r = \Lambda_{r,p} \times \text{Leb}$.

Experimental Setup

Theoretical Verification

The paper is primarily theoretical work, with results verified through:

1. Multivariate Moving Maximum (M4) Process Application: $Y_{k,i} = \bigvee_{r=0}^{\infty} \bigvee_{j=1}^d a_{ij,r} W_{k-r,j}$
   where $W_{k,i} = h(X_{k,i})$, $P(W_{k,i} > u) = u^{-\alpha}L(u)$ (regular variation)
2. Covariance Decay Condition Verification: Proof that condition $\psi_{ij,l} = o(l^{-1/2}\log(l)^{-1})$ implies the Berman condition $\Gamma(h) = o(1/\log(h))$

Technical Verification Methods

Lemma 4.1 (Approximation Strategy): For infinite-dimensional transformations, through truncation approximation:

• Condition (4.6): Boundary continuity condition
• Condition (4.7): Approximation error control

Experimental Results

Main Theoretical Results

1. Extreme Value Index Formula: For M4 processes, an explicit multivariate extreme value index is obtained: $\theta(\tau) = \frac{\sum_{j=1}^d \bigvee_{r \in \mathbb{Z}} \left(\bigvee_{i=1}^d \frac{a_{ij,r}^{\alpha}\tau_i}{A_i}\right)}{\sum_{j=1}^d \sum_{r \in \mathbb{Z}} \left(\bigvee_{i=1}^d \frac{a_{ij,r}^{\alpha}\tau_i}{A_i}\right)}$

2. Extreme Value Independence: Proof that for jointly Gaussian vectors with canonical correlation $\rho < 1$, any transformation preserves extreme value independence: $P(Y_1 > x, Y_2 > x) \leq P(Y_1 > x)^{2/(1+|\rho|)}$

3. m-Extremal Dependence Lower Bound: $\theta_m(\tau) \geq \frac{1}{m+1}$

Technical Innovation Verification

Effectiveness of Hypercontractivity: Lemma 4.4 proves the key variance control in point process convergence: $\limsup_{n \to \infty} E\left[\sum_{i<j} Z_n(i)Z_n(j)\right] \leq \frac{1}{2}\Lambda_{r,p}(\Delta)^2\text{Leb}(T)^2$

Related Work

Classical Extreme Value Theory

• Berman (1964): Foundational work on extreme value theory for Gaussian processes
• Leadbetter et al. (1983): Systematic development of extreme value theory for stationary sequences
• De Haan & Resnick (1977): Multivariate extreme value theory

Gaussian Subordination Processes

• Sly & Heyde (2008): Pioneering systematic study of extremes for Gaussian subordination processes
• Davis (1983): Early results on monotone transformations
• Kulik & Soulier (2012): Applications in stochastic volatility models

Advantages of This Work

1. Weaker Decay Conditions: Relaxation from polynomial decay to logarithmic decay
2. Complete Multivariate Theory: Systematic multivariate extreme value index theory
3. Improved Technical Methods: More refined application of hypercontractivity

Conclusions and Discussion

Main Conclusions

1. Successful extension of extreme value theory for Gaussian subordination processes to long-memory cases
2. Establishment of a complete multivariate extreme value theory framework
3. Demonstration that extreme value independence of Gaussian transformations is a universal phenomenon
4. Provision of practical analytical tools for multivariate moving maximum models

Limitations

1. Model Restrictions: Still limited to finite-dimensional cross-sectional dependence
2. Technical Conditions: Requires linear representation of underlying Gaussian processes
3. Computational Complexity: Calculation of multivariate extreme value indices may be complex

Future Directions

1. Extension to infinite-dimensional cross-sectional dependence
2. Investigation of non-Gaussian subordination processes
3. Development of statistical inference methods

In-Depth Evaluation

Strengths

1. Theoretical Depth: Significant theoretical breakthrough in extreme value theory
2. Technical Innovation: Clever application of hypercontractivity and canonical correlations
3. Practical Value: Provides tools for extreme value analysis of long-memory time series
4. Writing Quality: Rigorous argumentation and clear structure

Weaknesses

1. Application Scope: Primarily theoretical work with limited practical application verification
2. Computational Methods: Lacks concrete numerical computation algorithms
3. Statistical Inference: Does not address parameter estimation and related statistical problems

Impact

1. Academic Contribution: Advances extreme value theory development under complex dependence structures
2. Methodological Value: Provided techniques applicable to other related problems
3. Application Prospects: Potential applications in financial risk, insurance, climate, and other fields

Applicable Scenarios

1. Financial Risk Management: Modeling extreme market events
2. Insurance Actuarial Science: Assessment of catastrophic risks
3. Climate Research: Statistical analysis of extreme weather events
4. Engineering Reliability: Modeling system extreme failures

References

The paper cites important literature in the field, including:

• Leadbetter, Lindgren, and Rootzén (1983): Classical textbook on extreme value theory
• Sly and Heyde (2008): Pioneering work on extremes for Gaussian subordination processes
• Janson (1997): Theory of Gaussian Hilbert spaces
• Kulik and Soulier (2020): Modern theory of heavy-tailed time series