High-dimensional covariance matrices in elliptical distributions with application to spherical test

Jiang Hu, Weiming Li, Zhi Liu, Zhou Wang

TL;DR

This paper extends Random Matrix Theory (RMT) to the "elliptical" family of distributions, which better models real-world data with heavy tails and non-linear dependencies. The authors derive a new Central Limit Theorem (CLT) for the eigenvalues of high-dimensional sample covariance matrices and use it to build more robust statistical tests for data sphericity.

What Problem Does This Paper Address?

In high-dimensional statistics, we often study the "spectrum" (the set of eigenvalues) of sample covariance matrices to understand the underlying structure of data. For decades, the gold standard has been the "linear transform" model, which assumes data points are linear combinations of independent random variables.

However, real-world data—especially in finance and biology—rarely follows this rule. It often exhibits heavy tails or complex, non-linear correlations that the linear model can't capture. Elliptical distributions (like the multivariate $t$-distribution) are a much better fit for these scenarios, but they break the independence assumptions required by classical RMT. This paper fills that gap by establishing how the eigenvalues behave when the data comes from an elliptical population.

Methods

The authors represent the data using a stochastic form: $x = \xi A u$, where $\xi$ is a random scalar (the radius), $A$ is a transformation matrix, and $u$ is a random direction on a unit sphere.

The core of their work is a new Central Limit Theorem (CLT) for Linear Spectral Statistics (LSS). To achieve this, they:

  1. Define a specific class of elliptical distributions where the radius $\xi$ satisfies a fourth-moment condition characterized by a parameter $\tau$.
  2. Use the Stieltjes transform, a mathematical tool that translates the distribution of eigenvalues into a complex-valued function.
  3. Employ Martingale Decomposition, adapting it to handle the non-linear dependencies introduced by the radius and direction components.
  4. Derive explicit formulas for the limiting mean and covariance of the eigenvalue statistics, showing exactly how the "elliptical-ness" (the $\tau$ parameter) shifts the results compared to standard normal data.

Datasets

  • Simulated Data: The authors used extensive Monte Carlo simulations (10,000 replications) to verify their theoretical CLT. They tested various distributions for the radius $\xi$, including Gamma and Beta distributions, to represent different elliptical populations.
  • S&P 500 Stock Returns: For their empirical study, they used weekly closing prices of S&P 500 stocks from the first half of 2013 (124 trading days). They categorized these into 11 sectors based on the North American Industry Classification System (NAICS).

Key Insights and Ideas

  • The Tug-of-War between Radius and Direction: The authors discovered that the non-linear dependency in elliptical data comes from two sources: the radius $\xi$ and the direction $u$. In a standard normal distribution, the effects of these two components perfectly cancel each other out. In other elliptical distributions, they don't, which is why classical RMT formulas fail.
  • The Power of $\tau$: A single constant, $\tau$, effectively captures the impact of the elliptical distribution on the spectral fluctuations. This simplifies the transition from "normal" RMT to "elliptical" RMT.
  • Spatial-Sign Transformation: To create a "sphericity test" that works regardless of the specific distribution of the radius, the authors used "spatial-sign" samples (normalizing each data vector to unit length). This effectively "strips away" the influence of the radius $\xi$, making the test more universal.

Key Results

  • Theoretical Proof: The paper provides a rigorous proof of the CLT for LSS under elliptical distributions (Theorem 2.2).
  • Sphericity Testing: They developed a "John-type" test ($T_m$) that combines information from multiple moments of the eigenvalues.
  • Financial Application: Applying their test to S&P 500 data, they found:
  • Stocks within the same sector are strongly correlated and fail the sphericity test (p-values $< 10^{-9}$).
  • Stocks chosen across different sectors are much closer to being "spherical" (uncorrelated), though they still show some evidence of high-dimensional structure.

Limitations and Future Work

  • Moment Constraints: The current theory requires the radius to have a finite fourth moment ($E(\xi^4)$). This excludes extremely heavy-tailed distributions, such as a multivariate Student-$t$ with very low degrees of freedom (e.g., $v \leq 4$).
  • Spiked Models: While the paper discusses "spiked" covariance models (where a few eigenvalues are much larger than the rest), the theoretical power of these tests under elliptical distributions is still an area for deeper exploration.
  • Complex Data: The current work focuses on real-valued elliptical distributions; extending this to complex-valued data (common in signal processing) would be a natural next step.

Summary updated

Abstract

This paper discusses fluctuations of linear spectral statistics of high-dimensional sample covariance matrices when the underlying population follows an elliptical distribution. Such population often possesses high order correlations among their coordinates, which have great impact on the asymptotic behaviors of linear spectral statistics. Taking such kind of dependency into consideration, we establish a new central limit theorem for the linear spectral statistics in this paper for a class of elliptical populations. This general theoretical result has wide applications and, as an example, it is then applied to test the sphericity of elliptical populations.
BibTeX
@article{hu2018high,
  author = {Jiang Hu and Weiming Li and Zhi Liu and Zhou Wang},
  title = {{High-dimensional covariance matrices in elliptical distributions with application to spherical test}},
  journal = {The Annals of Statistics},
  year = {2018},
  doi = {10.1214/18-aos1699},
}

Loading related content...

Figure