On optimal linear prediction

TL;DR

This paper establishes conditions under which optimal linear predictors and model reductions minimize mean square prediction error, linking statistical models, chemometric algorithms, and quantum perspectives.

Contribution

It demonstrates the optimality of certain model reduction methods, like partial least squares, within a linear prediction framework, and connects these to quantum mechanical concepts.

Findings

Optimal model reduction corresponds to partial least squares under specific assumptions.

Partial least squares predictors outperform other predictors under certain conditions.

Quantum mechanical analogy provides a new perspective on model reduction.

Abstract

The main purpose of this article is to prove that, under certain assumptions in a linear prediction setting, optimal methods based upon model reduction and even an optimal predictor can be provided. The optimality is formulated in terms of the expected mean square prediction error. The optimal model reduction turns out, under a certain assumption, to correspond to the statistical model for partial least squares discussed by the author elsewhere, and under a certain specific condition, the partial least squares predictors is proved to be good compared to all other predictors. It is also proved in this article that the situation with two different model reductions can be fit into a quantum mechanical setting. Thus, the article contains a synthesis of three cultures: mathematical statistics as a basis, algorithms introduced by chemometricians and used very much by applied scientists as a background, and finally, notions from quantum foundation as an alternative point of view.