Dimension reduction in regression without matrix inversion