Uniform approximation in classical weak convergence theory

Viktor Bengs, Hajo HolzmannOpenReview 2019 · DOI · OpenReview

Abstract: A common statistical task lies in showing asymptotic normality of certain statistics. In many of these situations, classical textbook results on weak convergence theory suffice for the problem at hand. However, there are quite some scenarios where stronger results are needed in order to establish an asymptotic normal approximation uniformly over a family of probability measures. In this note we collect some results in this direction. We restrict ourselves to weak convergence in $\mathbb R^d$ with continuous limit measures.

Introduction

In foundational statistics, weak convergence (or convergence in distribution) serves as the bedrock for asymptotic inference. The Central Limit Theorem (CLT), for instance, guarantees that as the sample size $n$ increases, the distribution of a sample mean approaches a Normal distribution. However, in modern econometrics and high-dimensional statistics, "pointwise" convergence—where the limit is established for a single, fixed distribution—is often insufficient.

Researchers frequently require uniform convergence to ensure that statistical approximations remain valid across an entire family of probability measures. This is essential for constructing robust confidence intervals and hypothesis tests that maintain their nominal coverage or size across a range of potential models. While these uniform results are often invoked in high-level research, they are frequently scattered across specialized literature or treated as folk theorems. This paper systematically organizes and proves uniform counterparts to the classical tools of weak convergence theory.

Defining Uniform Weak Convergence

The core objective is to upgrade standard convergence results to a uniform setting. We consider a family of probability measures indexed by a parameter $\vartheta \in \Theta$. The authors define uniform convergence in distribution by requiring the supremum of the difference between the sequence of distributions and the limit to vanish across the entire parameter set:

$$\sup_{\vartheta \in \Theta} |F_{X_n^\vartheta}(x) - F_{X^\vartheta}(x)| \to 0$$

To maintain mathematical tractability, the scope is restricted to random vectors in $\mathbb{R}^d$ where the limit measures are continuous (lacking point masses or jumps). A critical technical requirement for these results is uniform absolute continuity. This condition ensures that the family of measures is sufficiently well-behaved, preventing mass from concentrating on infinitesimal sets in a manner that varies erratically across $\Theta$.

The Uniform Toolkit

The paper establishes a comprehensive suite of theorems that mirror the classical "textbook" results used in asymptotic analysis:

1. The Uniform Portmanteau Theorem

Often described as the "Swiss Army Knife" of convergence theory, the Portmanteau Theorem relates the convergence of integrals of functions to the convergence of probabilities of sets. The uniform version (Theorem 2.1) provides seven equivalent conditions for uniform weak convergence, including uniform convergence over Lipschitz continuous functions and sets with "thin" boundaries.

2. Continuous Mapping and Slutzky’s Theorems

These results are vital for the algebraic manipulation of limits. The Uniform Continuous Mapping Theorem ensures that if a sequence of random variables $X_n$ converges to $X$ uniformly, then $f(X_n)$ converges to $f(X)$ uniformly, provided $f$ is continuous. Similarly, the Uniform Slutzky’s Theorem allows researchers to handle sums and products of sequences where one component converges to a constant uniformly.

3. Uniform Lévy’s Continuity Theorem

This theorem connects the uniform convergence of distribution functions to the uniform convergence of their characteristic functions (Fourier transforms). This provides a powerful analytical path for proving convergence by working in the frequency domain.

The Lindeberg-Feller Theorem

The culmination of this theoretical framework is the Uniform Lindeberg-Feller Theorem. In classical theory, the Lindeberg-Feller condition is the general requirement for a sum of independent (but not necessarily identically distributed) variables to converge to a Normal distribution.

The uniform version established here provides the specific conditions under which this asymptotic normality holds reliably across a family of measures. This result is the primary engine for proving uniform asymptotic normality in complex statistical models, allowing for valid inference even when the underlying distribution is not fixed.

Scope and Limitations

While providing a robust toolkit, the results are bounded by specific theoretical constraints:

  • Dimensionality: The proofs are tailored for $\mathbb{R}^d$. While extensions to abstract metric spaces or function spaces are theoretically possible, they are beyond the scope of this specific note.
  • Continuity Requirement: The reliance on continuous limit measures is fundamental. Handling discrete limits or limits with atoms would introduce significant complexity regarding boundary cases and the relationship between CDFs and measures.
  • Absolute Continuity: The theorems assume uniform absolute continuity across the parameter space. In certain statistical edge cases, such as parameters on the boundary of a space, this condition may be violated, necessitating further refinements.

By consolidating these results, this work provides a rigorous foundation for researchers to move beyond pointwise asymptotics toward more robust, uniform statistical theory.

Summary updated

Abstract

A common statistical task lies in showing asymptotic normality of certain statistics. In many of these situations, classical textbook results on weak convergence theory suffice for the problem at hand. However, there are quite some scenarios where stronger results are needed in order to establish an asymptotic normal approximation uniformly over a family of probability measures. In this note we collect some results in this direction. We restrict ourselves to weak convergence in $R^d$ with continuous limit measures
BibTeX
@misc{bengs2019uniform,
  author = {Viktor Bengs and Hajo Holzmann},
  title = {{Uniform approximation in classical weak convergence theory}},
  year = {2019},
  eprint = {1903.09864},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1903.09864},
  howpublished = {OpenReview},
}

Loading related content...

Figure