Asymptotic locations of bounded and unbounded eigenvalues of sample correlation matrices of certain factor models -- application to a components retention rule

TL;DR

This paper analyzes the asymptotic behavior of eigenvalues of sample correlation matrices in factor models, extending results to unbounded cases, and discusses implications for component retention rules in high-dimensional data.

Contribution

It extends spectral analysis of correlation matrices to unbounded models and links eigenvalue behavior to component retention rules in high-dimensional factor models.

Findings

Eigenvalues of correlation matrices exhibit specific asymptotic behaviors.

Broken-stick rule effectively estimates the number of significant components.

Results apply to unbounded correlation matrices in high-dimensional settings.

Abstract

Let the dimension $N$ of data and the sample size $T$ tend to $\infty$ with $N/T \to c > 0$. The spectral properties of a sample correlation matrix $\mathbf{C}$ and a sample covariance matrix $\mathbf{S}$ are asymptotically equal whenever the population correlation matrix $\mathbf{R}$ is bounded (El Karoui 2009). We demonstrate this also for general linear models for unbounded $\mathbf{R}$, by examining the behavior of the singular values of multiplicatively perturbed matrices. By this, we establish: Given a factor model of an idiosyncratic noise variance $\sigma^2$ and a rank-$r$ factor loading matrix $\mathbf{L}$ which rows all have common Euclidean norm $L$. Then, the $k$th largest eigenvalues $\lambda_k$ $(1\le k\le N)$ of $\mathbf{C}$ satisfy almost surely: (1) $\lambda_r$ diverges, (2) $\lambda_k/s_k^2\to1/(L^2 + \sigma^2)$ $(1 \le k \le r)$ for the $k$th largest singular value $s_k$ of $\mathbf{L}$, and (3) $\lambda_{r + 1}\to(1-\rho)(1+\sqrt{c})^2$ for $\rho := L^2/(L^2 + \sigma^2)$. Whenever $s_r$ is much larger than $\sqrt{\log N}$, then broken-stick rule (Frontier 1976, Jackson 1993), which estimates $\mathrm{rank}\, \mathbf{L}$ by a random partition (Holst 1980) of $[0,\,1]$, tends to $r$ (a.s.). We also provide a natural factor model where the rule tends to "essential rank" of $\mathbf{L}$ (a.s.) which is smaller than $\mathrm{rank}\, \mathbf{L}$.