The RDS-free M–P law for large-dimensional sample covariance matrices
Random Matrices: Theory and Applications · 30.10.2026
DOI: 10.1142/s2010326326500188 ↗Kurz gefasst
Fachartikel in der Zeitschrift Random Matrices: Theory and Applications, online erschienen am 30.10.2026. Verfasst von Wenya Luo und 4 weiteren Autor:innen. Gefördert von: National Natural Science Foundation of China, Natural Science Foundation of Jilin Province, Natural Science Foundation of Zhejiang Province.
Abstract (Originalsprache)
This paper develops an RDS-free framework for the spectral analysis of large sample covariance matrices. In the classical large-dimensional (CLD) framework, the ratio of the dimension [Formula: see text] to the sample size [Formula: see text], referred to as the ratio of dimensionality to sample size (RDS), is usually assumed to converge to a positive finite constant [Formula: see text]. Although this assumption is mathematically convenient, the limiting value [Formula: see text] is generally not observable in applications, where only one finite pair [Formula: see text] is available. In practice, formulas involving the limiting ratio [Formula: see text] are therefore often implemented by replacing [Formula: see text] with the finite-sample ratio [Formula: see text]. Motivated by this finite-sample viewpoint, we compare the empirical spectral distribution (ESD) of the sample covariance matrix directly with the [Formula: see text]-parametrized Marčenko–Pastur (M–P) law. The resulting formulation avoids imposing the classical assumption [Formula: see text] and allows the approximation to be studied under explicit size conditions on [Formula: see text] and [Formula: see text], rather than under a limiting condition on [Formula: see text]. We obtain and use an explicit Stieltjes transform representation for the [Formula: see text]-parametrized M–P law and establish quadratic-form concentration bounds adapted to the growth regimes considered in this paper. Using these tools, we prove that, under the moment assumptions and growth conditions stated in the main theorems, the ESD is asymptotically close, in the Lévy distance, to the [Formula: see text]-parametrized M–P law. This provides a theoretical justification for using the finite-sample ratio [Formula: see text] in first-order global spectral approximations without requiring [Formula: see text] itself to converge to a prescribed limiting value. We further study the nonzero eigenvalues after a [Formula: see text]-dependent centering and scaling transformation, and obtain a semicircle-type consequence for the transformed spectral distribution. Simulation studies are presented to illustrate the finite-sample behavior of the proposed RDS-free approximation.
Förderung
- National Natural Science Foundation of China
- Natural Science Foundation of Jilin Province
- Natural Science Foundation of Zhejiang Province
Zitieren
APA
Luo, W., Wang, J., Bai, Z., Niu, Z., & Wong, W. (2026). The RDS-free M–P law for large-dimensional sample covariance matrices. Random Matrices: Theory and Applications. https://doi.org/10.1142/s2010326326500188
Vancouver
Luo W, Wang J, Bai Z, Niu Z, Wong W. The RDS-free M–P law for large-dimensional sample covariance matrices. Random Matrices: Theory and Applications. 2026. doi:10.1142/s2010326326500188
BibTeX
@article{Luo2026,
author = {Luo, Wenya and Wang, Jiaying and Bai, Zhidong and Niu, Zhenzhen and Wong, Wing-Keung},
title = {The RDS-free M–P law for large-dimensional sample covariance matrices},
journal = {Random Matrices: Theory and Applications},
year = {2026},
doi = {10.1142/s2010326326500188}
}
EUVIMED