Asymptotic distributions of functions of the eigenvalues of sample covariance matrix and canonical correlation matrix in multivariate time series (Q1822435)

A p-dimensional random vector X(t) is generated linearly by \(X(t)=\sum^{\infty}_{j=0}G(j)e(t-j)\), where the G(j) are \(p\times q\) matrices and the e(n) are q-dimensional zero-mean random vectors with \[ E\{e(n)e(m)'\}=\delta (m,n)K,\quad K\quad non\sin gular,\quad \sum^{\infty}_{j=0}tr G(j) KG(j)'<\infty. \] With a stretch X(1),...,X(n) let \(C(0)=n^{- 1}\sum^{n}_{t=1}X(t)X(t)'=(C(\alpha,\beta))\), and let \(l_ 1\geq l_ 2\geq...\geq l_ p\) be the corresponding eigenvalues forming the diagonal matrix L, so that \(C(0)=BLB'\), where B is orthogonal. The asymptotic joint normality of \(\sqrt{n}(L-\Lambda)\) and \(\sqrt{n}(B- \Psi)\) is proved, where \(\Psi \Lambda \Psi '=\Gamma (0)=EC(0)\), and the diagonal matrix \(\Lambda\) has diagonal elements \(\lambda_ 1>...>\lambda_ p\). The cases of X(t) Gaussian and non-Gaussian are considered, and also the case where the q smallest population roots are equal, together with a test of such hypothesis. By considering the \(\Gamma (j)=E\{X(t)X(t+j)'\}\) covariance matrices, and assuming that \(\sum^{\infty}_{u=-\infty}| u| \| \Gamma (u)\| <\infty\), an asymptotic expansion for the joint distribution of \(L_ 1,...,L_ k\) is obtained, where \[ L_ j=\sqrt{n}\{T_ j(l_ 1,...,l_ p)-T_ j(\lambda_ 1,...,\lambda_ p)\}, \] the \(T_ j\) being analytic functions. After this, the authors derive the asymptotic expansion for certain functions of the sample canonical correlations in multivariate time series. Applications of some of the results in signal processing are discussed.