Statistical analysis of latent generalized correlation matrix estimation in transelliptical distribution (Q502854)

Covariance and correlation matrices play a central role in multivariate analysis. For this reason, an efficient estimation of the covariance/correlation matrix is a fundamental step in conducting many multivariate methods. The conventional method in estimating large correlation matrices focuses on the use of Pearson's sample correlation matrix, which has various good properties under normal models. If the Gaussian assumption is correct, accurate estimation can be expected, otherwise, the obtained result may be misleading. However, in practise the Gaussian assumption is not realistic for many real-world applications. To relax the Gaussian assumption, the authors [``Scale-invariant sparse PCA on high-dimensional meta-elliptical data'', J. Am. Stat. Assoc. 109, No. 505, 275--287 (2014; \url{doi:10.1080/01621459.2013.844699})] use the transelliptical family for modeling and analyzing complex and noisy data. The transelliptical family, which contains many well-known distributions as special cases (Gaussian, \(t\), rank-deficient Gaussian, Cauchy, Kotz etc.), assumes that, after unspecified increasing marginal transformations, the data are elliptically distributed. In this context, the authors [loc. cit.] exploited a transformed version of the Kendall's tau sample correlation matrix to estimate the latent Pearson's correlation matrix. They showed, under the transelliptical distribution and without any moment constraint, that a transformed Kendall's tau sample correlation matrix approximates the latent Pearson's correlation in a parametric rate. In this paper, motivated by [loc. cit.], the authors study, in high dimensions, the theoretical properties of the Kendall's tau sample correlation matrix and its transformed version for estimating the population Kendall's tau correlation matrix and the latent Pearson's correlation matrix, respectively, under both spectral and restricted spectral norms. The role of ``effective rank'' and of the first time presented ``sign sub-Gaussian condition'' are highlighted in quantifying the rate of convergence with regard to the spectral and restricted spectral norm, respectively. In both cases, there is no need for any moment condition. Summarizing, this paper, different from the existing ones, gives a more general study on the convergence of the Kendall's tau matrix itself and provides more insights into rank-based statistics. The new theories developed have significantly new contribution to high-dimensional robust statistics literature and broad implications, since they can be applied to the study of factor model, robust regression etc.