On the maximum likelihood estimator for the Generalized Extreme-Value\n distribution

Axel Bücher, Johan Segers

Abstract

The vanilla method in univariate extreme-value theory consists of fitting the three-parameter Generalized Extreme-Value (GEV) distribution to a sample of block maxima. Despite claims to the contrary, the asymptotic normality of the maximum likelihood estimator has never been established. In this paper, a formal proof is given using a general result on the maximum likelihood estimator for parametric families that are differentiable in quadratic mean but whose supports depend on the parameter. An interesting side result concerns the (lack of) differentiability in quadratic mean of the GEV family.

Introduction

The Generalized Extreme-Value (GEV) distribution is the cornerstone of univariate extreme-value theory. It is the standard model for analyzing the tails of probability distributions, used extensively to predict rare but impactful events such as 100-year floods, catastrophic financial crashes, or record-breaking temperatures. To estimate the parameters of this distribution, researchers and practitioners almost universally rely on the Maximum Likelihood Estimator (MLE).

In standard statistical models, we assume that as the sample size grows, the MLE is "well-behaved"—meaning it converges to the true value and its error follows a predictable bell curve (asymptotic normality). However, for the GEV distribution, this assumption rested on shaky ground. While influential papers in the 1980s claimed to have proven this behavior, subsequent analysis revealed significant logical gaps, particularly regarding the uniformity of convergence across the parameter space. This paper provides the long-missing formal proof, establishing a rigorous mathematical foundation for one of the most common tools in statistics.

The Challenge of Non-Regularity

The primary difficulty in proving the asymptotic normality of the GEV MLE lies in the distribution's "non-regular" nature. In most common distributions, the range of possible values (the support) is fixed. In the GEV family, however, the support depends on the parameters themselves.

When the parameters change, the "edge" of the distribution moves. This makes standard calculus-based proofs—which rely on the smoothness of the log-likelihood function—unreliable. If the parameters shift such that a data point falls outside the predicted support, the log-likelihood can suddenly jump to negative infinity. This lack of regularity requires a more sophisticated mathematical approach than the standard Taylor expansions used in textbook MLE proofs.

Methodology: A New Theoretical Framework

Rather than simply patching the existing GEV proofs, the authors develop a general framework for handling distributions with parameter-dependent support. They introduce Proposition 2.1, which provides sufficient conditions for the asymptotic normality of MLEs in these complex cases.

To achieve this, the authors utilize Empirical Process Theory, specifically the "Donsker" machinery. The proof relies on three key technical pillars:

  1. Differentiability in Quadratic Mean (DQM): Instead of relying on standard derivatives, which may not exist at the boundary of the support, the authors use DQM. This is a more robust form of differentiability that is better suited for distributions with shifting boundaries.
  2. Lipschitz Conditions: They establish that the log-likelihood function does not change too violently as parameters shift, provided the parameters remain within a certain distance from the boundary of the support.
  3. Bracketing Integrals: This technique allows the authors to control the complexity of the functions under study, ensuring that the average behavior of the likelihood function across a sample is representative of the underlying distribution.

Key Insights and Results

The $\gamma > -1/2$ Threshold

The paper confirms a critical threshold in extreme-value theory: the MLE's "nice" behavior only exists when the shape parameter $\gamma$ is greater than $-1/2$. The authors show that the GEV family is Differentiable in Quadratic Mean if and only if this condition is met. If $\gamma \le -1/2$, the distribution's tail is either so "short" or the boundary is so "sharp" that the MLE converges at a different rate and follows a non-normal distribution.

Formal Proof of Normality

The central result of the paper is the formal proof that for $\gamma_0 > -1/2$, the MLE $\hat{\theta}_n$ satisfies: $$\sqrt{n}(\hat{\theta}_n - \theta_0) \to N(0, I_{\theta_0}^{-1})$$ where $I_{\theta_0}$ is the Fisher Information Matrix. This result confirms that for the vast majority of practical applications, the standard errors and confidence intervals used by statisticians are mathematically valid.

A Template for Future Research

Beyond the GEV distribution, the authors' general framework (Proposition 2.1) serves as a template for other "non-regular" models. This includes the Generalized Pareto distribution and the three-parameter Weibull family, both of which face similar challenges regarding parameter-dependent support.

Conclusion and Limitations

While this paper resolves a decades-old uncertainty in extreme-value theory, it also defines the limits of standard MLE application. The proof explicitly excludes cases where $\gamma \le -1/2$. In these "very non-regular" scenarios, which occasionally appear in specialized fields, the MLE may converge much faster (at a rate of $n$ rather than $\sqrt{n}$) and follow a non-Gaussian limiting distribution.

For the broader scientific community, however, this work provides the necessary theoretical validation for the GEV MLE, ensuring that our models for extreme risks are built on a solid mathematical base.

Summary updated

Abstract

The vanilla method in univariate extreme-value theory consists of fitting the\nthree-parameter Generalized Extreme-Value (GEV) distribution to a sample of\nblock maxima. Despite claims to the contrary, the asymptotic normality of the\nmaximum likelihood estimator has never been established. In this paper, a\nformal proof is given using a general result on the maximum likelihood\nestimator for parametric families that are differentiable in quadratic mean but\nwhose supports depend on the parameter. An interesting side result concerns the\n(lack of) differentiability in quadratic mean of the GEV family.\n
BibTeX
@inproceedings{bucher2017maximum,
  author = {Axel Bücher and Johan Segers},
  title = {{On the maximum likelihood estimator for the Generalized Extreme-Value\textbackslash\{\}n distribution}},
  booktitle = {Extremes},
  year = {2017},
  doi = {10.1007/s10687-017-0292-6},
}

Loading related content...

Figure