Regression function estimation as a partly inverse problem (Q778881)

This article develops nonparametric regression function estimation by a projection method. The regression problem is treated as a partially inverse problem. The procedure can be applied to non-compactly supported bases. A support constraint is eliminated by introducing a new selection procedure including a cutoff for the underlying matrix inversion. The eigenvalues of the matrix which must be inverted play a role in the definition of the collection of models. The dimension of the projection space is chosen in a random set. From the bias-variance decomposition, upper and lower rates of the estimator on specific Sobolev spaces are obtained.