Canonical correlation coefficients of high-dimensional Gaussian vectors: finite rank case (Q1731774)

The relationship between two sets of $p$ and $q$ dimensional random variables $\mathbf{x}$ and $\mathbf{y}$ is investigated. Here $p$ and $q$ are proportional to the sample size $n$. Here $n$ independent observations of Gaussian vector $\mathbf{z}=(\mathbf{x}',\mathbf{y}')'$ are assumed to be known. $\Sigma_{\mathbf{x}\mathbf{y}}$ is assumed to enjoy the finite rank $k$, namely there are $k$ nonzero canonical correlation coefficients, whose squares are denoted by $r_1\geq\cdots\geq r_k0$. The authors assume $(p+q)/n\to y\in (0,1)$ and $p/q\not\to 1$ and study the sample counterparts of $r_i,i=1,\dots,k$, i.e., the largest $k$ eigenvalues $\lambda_1\geq\cdots\geq \lambda_k$ of the sample canonical correlation matrix $S_{\mathbf{x}\mathbf{x}}^{-1}S_{\mathbf{x}\mathbf{y}}S_{\mathbf{y}\mathbf{y}}^{-1}S_{\mathbf{y}\mathbf{x}}$. The main result of this paper is the demonstration of the existence of a threshold $r_c\in(0,1)$ such that for each $i\in\{1,\dots,k\},\, r_i\leq r_c,$ eigenvalues $\lambda_i$ converge almost surely to the right edge of the limiting spectral distribution of the sample canonical correlation matrix.