Central Limit Theorem for Linear Spectral Statistics of Large Dimensional Kendall's Rank Correlation Matrices and its Applications

Zeng Li, Qinwen Wang, Runze LiThe Annals of Statistics 2019 · arXiv

Introduction

In high-dimensional multivariate statistics, a fundamental challenge is determining whether the components of a large-dimensional vector are independent. As the dimension $p$ grows relative to the sample size $n$, classical statistical tools often fail. The standard approach typically relies on the Pearson correlation matrix and the framework of Random Matrix Theory (RMT). However, Pearson-based methods are sensitive to the underlying distribution of the data, generally requiring finite fourth moments to ensure the validity of the Central Limit Theorem (CLT) for their eigenvalues.

In fields such as finance, insurance, and environmental science, data often exhibit heavy-tailed or highly skewed distributions (e.g., Cauchy or Pareto distributions) where such moment conditions are not met. This paper addresses this gap by establishing a CLT for the linear spectral statistics (LSS) of Kendall’s rank correlation matrices, providing a robust, moment-free framework for high-dimensional independence testing.

The Challenge of Heavy-Tailed Data

The Pearson correlation coefficient measures linear relationships but is highly susceptible to outliers. In the "Marchenko-Pastur regime," where $p/n \to c \in (0, \infty)$, the spectral behavior of the sample covariance matrix is well-understood only under strict distributional assumptions. When data are heavy-tailed, the sample covariance matrix does not converge to the Marchenko-Pastur law, rendering standard RMT-based tests unreliable.

Kendall’s $\tau$ offers a distribution-free alternative. Because it is based on the ranks of the data rather than their raw values, it is invariant to monotonic transformations and remains valid even for distributions without a defined variance. However, analyzing the eigenvalues of a matrix of Kendall’s $\tau$ coefficients is mathematically complex because the entries are U-statistics, creating non-linear dependencies between the elements of the matrix.

Methodology: Bridging U-Statistics and RMT

The authors overcome the dependency challenges of Kendall’s rank correlation matrices by integrating the theory of U-statistics with the spectral analysis of large random matrices. Their approach utilizes several key technical innovations:

  • Hoeffding Decomposition: The Kendall’s $\tau$ U-statistic is decomposed into a linear part and a degenerate part. This allows the authors to approximate the complex matrix structure with a more manageable sum of independent components.
  • Stieltjes Transform: This fundamental tool in RMT is used to derive the limiting spectral distribution. By analyzing the transform, the authors can map the behavior of the eigenvalues as $p$ and $n$ diverge.
  • New Quadratic Form Lemmas: Because the transformed data matrix does not have strictly independent rows in the traditional sense, the authors developed new lemmas to calculate the expectations of products of quadratic forms, a necessary step for establishing the CLT.

The Central Limit Theorem for LSS

The core contribution of this research is the formal establishment of the CLT for the Linear Spectral Statistics (LSS) of the Kendall’s rank correlation matrix. Unlike previous literature, this theorem requires no moment conditions and only assumes that the underlying components are independent and continuous.

The authors derived explicit formulas for the limiting mean and covariance functions of the LSS. A notable finding is that the "population" version of the Kendall matrix possesses a unique spectral signature—with specific eigenvalue structures at $(n+1)/3$ and $1/3$—which differs from the identity matrix structure typically encountered in Pearson-based RMT.

Applications: High-Dimensional Independence Testing

Leveraging the CLT, the paper proposes three new nonparametric test statistics to evaluate the hypothesis of independence in high dimensions:

  1. $Q_{\tau,2}$: Based on the Frobenius norm, effectively measuring the sum of squares of the correlations.
  2. $Q_{\tau,4}$: Based on the fourth spectral moment.
  3. $Q_{\tau,\log}$: Based on "entropy loss," utilizing the log-determinant of the matrix.

These tests provide a robust mechanism for detecting dependencies without needing to know the underlying distribution of the data.

Numerical Performance and Robustness

Through extensive numerical simulations, the authors compared their Kendall-based tests against traditional Pearson and Spearman-based methods. The simulations utilized various scenarios, including skewed Gamma distributions and heavy-tailed Cauchy and $t$-distributions, across both linear (Toeplitz) and nonlinear correlation structures.

Key findings include: * Robustness: In Cauchy distribution cases where Pearson tests failed entirely, the Kendall-based tests maintained accurate Type I error rates (approximately 5%). * Power: The $Q_{\tau,2}$ and $Q_{\tau,\log}$ tests demonstrated the highest power in detecting dependencies across a wide range of mixed-distribution scenarios. * Nonlinear Detection: The rank-based approach proved superior in identifying dependencies in models where variables were related through nonlinear terms that typically trip up linear measures.

Conclusion and Future Work

This paper provides a significant theoretical advancement by proving that the spectral limits of Kendall’s rank correlation matrices can be analyzed without the restrictive moment conditions required by classical RMT. This makes Kendall-based LSS a powerful tool for analyzing "messy" real-world data.

While the tests are robust, the authors note that computational complexity is higher than Pearson-based methods, and tests based on extreme eigenvalues (Tracy-Widom limits) may converge slowly in smaller samples. Future research may extend this framework to "spiked models," where a small number of components exhibit high correlation, a common occurrence in signal processing and genetic studies.

Summary updated

Abstract

This paper is concerned with the limiting spectral behaviors of large dimensional Kendall's rank correlation matrices generated by samples with independent and continuous components. We do not require the components to be identically distributed, and do not need any moment conditions, which is very different from the assumptions imposed in the literature of random matrix theory. The statistical setting in this paper covers a wide range of highly skewed and heavy-tailed distributions. We establish the central limit theorem (CLT) for the linear spectral statistics of the Kendall's rank correlation matrices under the Marchenko-Pastur asymptotic regime, in which the dimension diverges to infinity proportionally with the sample size. We further propose three nonparametric procedures for high dimensional independent test and their limiting null distributions are derived by implementing this CLT. Our numerical comparisons demonstrate the robustness and superiority of our proposed test statistics under various mixed and heavy-tailed cases.
BibTeX
@article{li2019central,
  author = {Zeng Li and Qinwen Wang and Runze Li},
  title = {{Central Limit Theorem for Linear Spectral Statistics of Large Dimensional Kendall's Rank Correlation Matrices and its Applications}},
  journal = {The Annals of Statistics},
  year = {2019},
}
Figure