Additive models in high dimensions

Abstract

Additive decompositions are established tools in nonparametric statistics and effectively address the curse of dimensionality. For the analysis of the approximation properties of additive decompositions, we introduce a novel framework which includes the number of variables as an ingredient in the definition of the smoothness of the underlying functions. This approach is motivated by the effect of concentration of measure in high dimensional spaces. Using the resulting smoothness conditions, convergence of the additive decompositions is established. Several examples confirm the error rates predicted by our error bounds. Explicit expressions for optimal additive decompositions (in an $L_2$ sense) are given which can be seen as a generalisation of multivariate Taylor polynomials where the monomials are replaced by higher order interactions. The results can be applied to the numerical approximation of functions with hundreds of variables.