2025-11-10T03:05:00.236892

Stein's method for Fréchet approximation: a regularly varying functions approach

Paul, Guillaume, Yvik
We develop a variant of Stein's method of comparison of generators to bound the Kolmogorov, total variation, and Wasserstein-1 distances between distributions on the real line. Our discrepancy is expressed in terms of the ratio of reverse hazard rates; it therefore remains tractable even when density derivatives are intractable. Our main application concerns the approximation of normalized extremes by Fréchet laws. In this setting, the new discrepancy provides a quantitative measure of distributional proximity in terms of the average regular variation at infinity of the underlying cumulative distribution function. We illustrate the approach through explicit computations for maxima of Pareto, Cauchy, and Burr~XII distributions.
academic

Stein's method for Fréchet approximation: a regularly varying functions approach

Basic Information

  • Paper ID: 2510.14016
  • Title: Stein's method for Fréchet approximation: a regularly varying functions approach
  • Authors: Paul Mansanarez, Guillaume Poly, Yvik Swan
  • Classification: math.PR (Probability Theory)
  • Publication Date: October 17, 2025
  • Paper Link: https://arxiv.org/abs/2510.14016

Abstract

This paper develops a variant of Stein's method for comparing generator operators to bound the Kolmogorov distance, total variation distance, and Wasserstein-1 distance between distributions on the real line. The discrepancy of the method is expressed in terms of the ratio of reverse hazard rates, thereby maintaining tractability even when density derivatives are difficult to handle. The primary application focuses on approximating normalized extremes by the Fréchet law. In this setting, the new discrepancy provides a quantitative measure of distributional proximity based on the average regular variation of the underlying cumulative distribution function at infinity. The method is illustrated through explicit calculations on the maxima of Pareto, Cauchy, and Burr XII distributions.

Research Background and Motivation

Problem Background

  1. Quantifying approximation in extreme value theory: In extreme value theory, the Fréchet distribution is one of three generalized extreme value (GEV) distributions and represents the unique possible limit of appropriately normalized maxima of independent identically distributed random variables. Quantifying the speed and precision of this convergence is an important theoretical and practical problem.
  2. Limitations of existing methods:
    • Traditional Stein's method relies on comparing score functions (derivatives of log-likelihood ratios of densities), which are difficult to handle in the GEV setting
    • Existing convergence rate results are technically demanding and difficult to apply
    • Lack of quantitative tools linking distributional proximity to regular variation theory
  3. Research motivation:
    • Develop a new variant of Stein's method that avoids computing density derivatives
    • Provide more tractable bounds in GEV approximation
    • Establish quantitative connections between distributional proximity and regular variation

Core Contributions

  1. New Stein discrepancy measure: Proposes a Stein discrepancy based on the ratio of reverse hazard rates, avoiding the computational complexity of density derivatives
  2. Unified bounds for three distances: Provides unified upper bounds for Kolmogorov distance, total variation distance, and Wasserstein-1 distance
  3. Quantitative characterization of regular variation: Links the Stein discrepancy to regular variation of the cumulative distribution function at infinity, providing quantitative convergence characterization
  4. Concrete application examples: Demonstrates the practical utility of the method through explicit calculations for Pareto, Cauchy, and Burr XII distributions

Methodology Details

Problem Definition

Given two cumulative distribution functions P and Q satisfying Assumption 0, the goal is to bound the probabilistic distances between them:

  • Kolmogorov distance: Kol(P,Q)\text{Kol}(P,Q)
  • Total variation distance: TV(P,Q)\text{TV}(P,Q)
  • Wasserstein-1 distance: Wass(P,Q)\text{Wass}(P,Q)

Core Assumptions

Assumption 0: There exist initial mass f0≥0f_0 \geq 0, left endpoint cF∈[−∞,+∞)c_F \in [-\infty,+\infty), and density-like function f:R→R+f: \mathbb{R} \to \mathbb{R}_+ such that the cumulative distribution function is:

0, & x < c_F \\ f_0 + \int_{c_F}^x f(u) du, & x \geq c_F \end{cases}$$ ### Stein Discrepancy Definition For distributions P and Q satisfying Assumption 0, define the reverse hazard rate: $$r_F(x) = \frac{f(x)}{F(x)}I_{(c_F,+\infty)}(x) = (\log F)'(x)I_{(c_F,+\infty)}(x)$$ **Primary Stein discrepancy**: $$\Delta(Q | P) := \int_{c_Q}^{+\infty} \left|1 - \frac{r_P(x)}{r_Q(x)}\right| q(x) dx$$ **Weighted Stein discrepancy** (for Wasserstein distance): $$\Delta_w(Q | P) := \int_{c_Q}^{+\infty} x \left|1 - \frac{r_P(x)}{r_Q(x)}\right| q(x) dx$$ ### Main Theoretical Results **Theorem 1.1** (Bounds for Kolmogorov and total variation distances): $$\text{Kol}(P,Q) \leq \Delta(Q | P) + q_0$$ $$\text{TV}(P,Q) \leq 2\Delta(Q | P) + q_0$$ **Theorem 1.2** (Bound for Wasserstein-1 distance): $$\text{Wass}(P,Q) \leq 2\mu\Delta(Q | P) + 3\Delta_w(Q | P)$$ ### Three-Step Framework of Stein's Method 1. **Construction of Stein operator**: Define the reparameterized operator $$A^p\phi := \left(\frac{\bar{P}}{p}\phi' + \phi\right)I_{(c_P,+\infty)}$$ 2. **Solution of Stein equation**: For a given test function $h$, solve the ODE $$\frac{\bar{P}}{p}\phi' + \phi = h - P(h)$$ 3. **Transfer principle**: Establish distance bounds by exploiting the discrepancy between operators across different distributions ## Experimental Setup ### Theoretical Verification The paper validates the method's effectiveness through the following distributions: 1. **Pareto distribution**: $F(x) = (1-x^{-\alpha})I_{(1,\infty)}(x)$ 2. **Cauchy distribution**: Standard Cauchy distribution 3. **Burr XII distribution**: $F(x) = 1-(1+x^\alpha)^{-\tau}$, $x > 0$ 4. **Log-logistic distribution**: $F(x) = \frac{x^\alpha}{1+x^\alpha}I_{(0,\infty)}(x)$ ### Evaluation Metrics - Precise constants in convergence rates - Numerical computation verification - Comparison with known theoretical results ## Experimental Results ### Results for Pareto Distribution For the Pareto distribution, exact convergence rates are obtained: $$\Delta(F_n | \Phi_\alpha) = \frac{1}{n+1}$$ This is in complete agreement with known convergence rates for Kolmogorov and total variation distances. ### Results for Cauchy Distribution For the Cauchy distribution, through integral transformation: $$\Delta(\Phi_1 | F_n) = \mathbb{E}[|1-R(E_n)|]$$ where $E_n \sim \text{Exp}(n)$, and an $O(1/n)$ convergence rate is established. ### Results for Burr XII Distribution Numerical calculations demonstrate a convergence rate of $O(n^{-1/\tau})$, consistent with theoretical expectations. ### Comparison Between Fréchet Distributions When both P and Q are Fréchet distributions, explicit expressions for the discrepancy are provided, involving Gamma functions and incomplete Gamma functions. ## Related Work ### Stein's Method in Extreme Value Theory - **Feidt (2013)**: Stein's method for multivariate extremes - **Bartholmé & Swan (2013)**: Generator operator comparison strategy - **Costacèque-Cecchi (2024)**: Framework based on Markov semigroups - **Kusumoto & Takeuchi (2020)**: Convergence rates via Stein equations ### Regular Variation Theory - **Karamata theory**: Foundational theory of regularly varying functions - **Attraction to extreme value domains**: Characterization of convergence conditions via regular variation ## Conclusions and Discussion ### Main Conclusions 1. **Method effectiveness**: The new Stein discrepancy provides effective bounds on distributional distances 2. **Computational advantages**: Avoiding density derivatives makes the method more applicable to complex distributions 3. **Theoretical connections**: Establishes quantitative links between distributional proximity and regular variation ### Limitations 1. **Assumption constraints**: Requires satisfaction of Assumption 0, limiting the scope of applicability 2. **Support conditions**: Imposes requirements on the relative positions of distribution supports 3. **Technical conditions**: Some results require additional boundedness conditions ### Future Directions 1. **Extension to multivariate case**: Generalization to multivariate extreme value distributions 2. **Other GEV distributions**: Application to Gumbel and Weibull distributions 3. **Higher-order analysis**: Investigation of finer asymptotic expansions ## In-Depth Evaluation ### Strengths 1. **Theoretical innovation**: Proposes a new Stein discrepancy based on reverse hazard rates, circumventing technical difficulties of traditional methods 2. **Unified framework**: Provides unified treatment for three important probabilistic distances 3. **Practical operability**: Demonstrates computational feasibility through concrete examples 4. **Theoretical depth**: Deeply integrates Stein's method with regular variation theory ### Weaknesses 1. **Assumption limitations**: Assumption 0 imposes strong restrictions on distribution form, potentially excluding some practical applications 2. **Bound tightness**: Insufficient discussion of the tightness of the obtained bounds 3. **Numerical experiments**: Lacks large-scale numerical experiments to verify practical performance of the method ### Impact 1. **Theoretical contribution**: Provides new tools for quantifying approximation in extreme value theory 2. **Methodological value**: The new variant of Stein's method may inspire other applications 3. **Practical value**: Potential applications in risk management, insurance, and related fields ### Applicable Scenarios - Distributional approximation in extreme value statistics - Tail risk assessment in risk management - Extreme event modeling in insurance actuarial science - VaR and ES estimation in finance ## References The paper cites classical literature in extreme value theory, regular variation theory, and Stein's method, including: - Bingham, Goldie & Teugels (1987): Regular Variation - de Haan & Ferreira (2006): Extreme Value Theory - Resnick (2008): Extreme Values, Regular Variation, and Point Processes - Ley, Reinert & Swan (2017): Stein's Method for Comparison of Univariate Distributions --- This paper makes important theoretical innovations by successfully applying Stein's method to the problem of extreme value distribution approximation, providing new analytical tools for the field. The practical utility of the method is verified through calculations on multiple classical distributions, demonstrating good application prospects.