Universally Consistent Regression Function Estimation Using Hierarchial B-Splines

Estimation of multivariate regression functions from i.i.d. data is considered. We construct estimates by empiricalL2-error minimization over data-dependent spaces of polynomial spline functions. For univariate regression function estimation these spaces are spline spaces with data-dependent knot sequences. In the multivariate case, we use so-called hierarchical spline spaces which are defined as linear span of tensor product B-splines with nested knot sequences. The knot sequences of the chosen B-splines depend locally on the data.  We show the strongL2-consistency of the estimators without any condition on the underlying distribution. The estimators are similar to histogram regression estimators using data-dependent partitions and partitioning regression estimators based on local polynomial fits. The main difference is that the estimators considered here are smooth functions, which seems to be desirable especially in the case that the regression function to be estimated is smooth.