Principal components in linear mixed models with general bulk
share
We study the outlier eigenvalues and eigenvectors in variance components estimates for high-dimensional mixed effects linear models using a free probability approach. We quantify the almost-sure limits of these eigenvalues and their eigenvector alignments, under a general bulk-plus-spikes assumption for the population covariances of the random effects, extending previous results in the identity-plus-spikes setting. Our analysis develops two tools in free probability and random matrix theory which are of independent interest---strong asymptotic freeness of GOE and deterministic matrices, and a method of proving deterministic anisotropic approximations to resolvents using free deterministic equivalents. Statistically, our results quantify bias and aliasing effects for the leading principal components of variance components estimates in modern high-dimensional applications.
READ FULL TEXT