39
0

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

Abstract

Let the dimension NN of data and the sample size TT tend to \infty with N/Tc>0N/T \to c > 0. The spectral properties of a sample correlation matrix C\mathbf{C} and a sample covariance matrix S\mathbf{S} are asymptotically equal whenever the population correlation matrix R\mathbf{R} is bounded (El Karoui 2009). We demonstrate this also for general linear models for unbounded R\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 σ2\sigma^2 and a rank-rr factor loading matrix L\mathbf{L} which rows all have common Euclidean norm LL. Then, the kkth largest eigenvalues λk\lambda_k (1kN)(1\le k\le N) of C\mathbf{C} satisfy almost surely: (1) λr\lambda_r diverges, (2) λk/sk21/(L2+σ2)\lambda_k/s_k^2\to1/(L^2 + \sigma^2) (1kr)(1 \le k \le r) for the kkth largest singular value sks_k of L\mathbf{L}, and (3) λr+1(1ρ)(1+c)2\lambda_{r + 1}\to(1-\rho)(1+\sqrt{c})^2 for ρ:=L2/(L2+σ2)\rho := L^2/(L^2 + \sigma^2). Whenever srs_r is much larger than logN\sqrt{\log N}, then broken-stick rule (Frontier 1976, Jackson 1993), which estimates rankL\mathrm{rank}\, \mathbf{L} by a random partition (Holst 1980) of [0,1][0,\,1], tends to rr (a.s.). We also provide a natural factor model where the rule tends to "essential rank" of L\mathbf{L} (a.s.) which is smaller than rankL\mathrm{rank}\, \mathbf{L}.

View on arXiv
Comments on this paper