Additive regression with Hilbertian responses (Q2215755)

The paper under review deals with the problem of regression analysis with non-Euclidean data. The authors develop the building block of structured nonparametric regression for Hilbertian responses. To this end they present a theory that generalizes the conventional Lebesgue integration to the case of Hilbert-space-valued maps. Namely, the notion of Bochner integration is detailed. Based on the Bochner integral theory, the authors introduce a technique of estimating the additive model (while the main ideas may be adapted to other structured models) that consists of solving a system of integral equations, expressed in terms of Bochner integrals, and an associated backfitting algorithm. Next, the existence of the estimator that solves the system of equations and the convergence of the algorithm are established. These results include both asymptotic and nonasymptotic versions. Moreover, a complete theory for the rates of convergence of the estimators of the component maps and their asymptotic distributions is presented. Finally, the validity of authors' approach is demonstrated in numerical examples (that include the cases of functional, density-valued and simplex-valued responses).