Subspace perspective on canonical correlation analysis: dimension reduction and minimax rates (Q2278668)

This paper studies theoretical properties of canonical correlation analysis, focusing on estimation of the canonical variables. Two losses functions are proposed based on the principal angles between the linear spans determined by the sample canonical variates and those determined by the population correspondents. These loss functions are compared with other loss functions in the literature. Non-asymptotic uniform sharp upper and lower bounds for estimating canonical correlation coefficients under the new loss functions proposed are established. The upper bounds illustrate how the population canonical correlation coefficients affect the estimation accuracy. It is shown that when the leading canonical correlation is close to 1, the estimation of canonical variates can be significantly accurate even with small eigen-gaps. Minimax lower bounds under a family of stringent and localized parameter spaces are also derived. The theoretical findings are illustrated by numerical simulations.