Some limit theorems on the eigenvectors of large dimensional sample covariance matrices (Q802194)

Let \(\{v_{ij}\}\), \(i,j=1,2,...\), be i.i.d. standardized random variables with E \(v^ 4_{11}=3\) and \(E| v_{11}|^ m\leq m^{\alpha m}\), \(m\geq 2\), for some \(\alpha >0\). Let \(M_ n=s_ n^{- 1}V_ nV^ T_ n\), where \(V_ n=(v_{ij})\), \(i=1,...,n\); \(j=1,...,s_ n\); and \(n/s_ n\to y>0\). Let \(\{f_ i\}\) be a sequence of functions on \([0,\infty)\), analytic at 0, \(f_ i(0)=0\), with radius of convergence greater than \((1+\sqrt{y})^ 2.\) The result of the paper says that under the above assumptions for any sequence of unit vectors \(\{x_ n\}\), \(x_ n\in R^ n\), the sequence \(\{\sqrt{n/2}x^ T_ nf_ i(M_ n)x_ n-n^{-1}tr f_ i(M_ n)\}^{\infty}_{i=1}\) converges in distribution in \(R^{\infty}\) as \(n\to \infty\) to \(\{\int^{(1+\sqrt{y})^ 2}_{(1-\sqrt{y})^ 2}f_ i(x)dW^ 0(F_ y(x))\}^{\infty}_{i=1}\), where \(W^ 0\) is the Brownian bridge and \(F_ y(x)\) is the a.s. limit of the sequence of the empirical distribution functions of the eigenvalues of \(M_ n\). This result shows that some limit behaviour of eigenvectors of \(M_ n\) does not depend on the normality of \(v_{ij}\). However, the condition \(Ev^ 4_{11}=3\) is essential. The paper continues previous investigations of the author [SIAM J. Math. Anal. 12, 274-281 (1981; Zbl 0465.60017)].