On the optimal reconstruction of partially observed functional data (Q2196241)

The aim of the paper is to develop a reconstruction procedure for recovering functions from their partial observations. Assume that \(X_1,\dots,X_n\) is an identical distributed, weakly dependent sample of continuous random functions and denote by \(X_i^{O}\) the observed parts of \(X_i\) and by \(X_i^{M}\) the missing parts of \(X_i\). To reconstruct the unobserved missing parts knowing the partial observations, the linear reconstruction model \(X_i^M = L(X_i^O) + Z_i\) is considered. The goal is to identify the optimal linear reconstruction operator \(L: {\mathbb{L}}^2(O) \to{\mathbb{L}}^2(M)\) that minimizes the mean squared error loss \(\mathbb{E}[(X_i^M(u)-L(X_i^O)(u))^2]\), for \(u\in M\). In the second section, the reconstruction operator \(\mathcal{L}(X_i^O)\) is introduced and an optimal linear reconstruction result is obtained. Estimators for \(\mathcal{L}\) from real data and estimating procedures are developed in the third section and asymptotic results in the fourth section. An iterative reconstruction algorithm is described in the fifth section and the performance of the considered reconstruction operators in the part devoted to the simulation study. Some applications are shown in the last part of the article. Proofs and additional theoretical results are reported to be contained in a supplementary paper, \url{doi:10.1214/19-AOS1864SUPP}.