Beta-Bernstein Smoothing for Regression Curves with Compact Support (Q4255155)

This paper proposes a variable kernel method of nonparametric curve estimation for functions with compact support which generalize the classical Bernstein polynomial approach and do not suffer from problems of boundary bias. The authors introduce estimators of the type \(m(x)=\sum_{j=0}^n y_j w_j(x)\) with weights \(w_j(x)\) formed by splitting the beta density function in such a way that \(w_j(x)= \int_0^1 K_{\lambda x+1, \lambda (1-x)+1}(t)f_j(t) dt\), where \(K_{a,b}\) denotes the density of the beta(a,b)-distribution and \(f_j(t)={n \choose j} t^j(1-t)^{n-j}\). The shapes of the kernels are symmetric in the middle of the support and become more asymmetric near the boundary, but never put any weight outside the data support interval and thus avoid boundary bias. It is proved that the proposed estimator has the optimal mean integrated squared error equivalent to that of the standard kernel estimates with unbounded support. Some empirical results are presented.