This paper investigates optimal linear prediction for a random function in an infinite-dimensional Hilbert space. We analyze the mean square prediction error (MSPE) associated with a linear predictor, revealing that non-unique solutions that minimize the MSPE generally exist, and consistent estimation is often impossible even if a unique solution exists. However, this paper shows that it is still feasible to construct an asymptotically optimal linear operator, for which the empirical MSPE approaches the minimal achievable level. Remarkably, standard post-dimension reduction estimators, widely employed in the literature, serve as such estimators under minimal conditions. This finding affirms the use of standard post-dimension reduction estimators as a way to achieve the minimum MSPE without requiring a careful examination of various technical conditions commonly required in functional linear models.
翻译:暂无翻译