Abstract: Sliced Inverse Regression is a method for reducing the dimension of the explanatory variables x in non-parametric regression problems. Li (1991) discussed a version of this method which begins with a partition of the range of y into slices so that the conditional covariance matrix of x given y can be estimated by the sample covariance matrix within each slice. After that the mean of the conditional covariance matrix is estimated by averaging the sample covariance matrices over all slices. Hsing and Carroll (1992) have derived the asymptotic properties of this procedure for the special case where each slice contains only two observations. In this paper we consider the case that each slice contains an arbitrary but fixed number of yi and more generally the case when the number of yi per slice goes to infinity. The asymptotic properties of the associated eigenvalues and eigenvectors are also obtained.
Key words and phrases: Asymptotics, sliced inverse regression, dimension reduction, eigenvalues and eigenvectors.