Smoothness in Bayesian Non-parametric Regression with Wavelets

Consider the classical nonparametric regression problem y_i = f(t_i) + _ii = 1,...,n where t_i = i/n, and _i are i.i.d. zero mean normal with variance ². The aim is to estimate the true function f which is assumed to belong to the smoothness class described by the Besov space B _pq ^q . These are functions belonging to L_p with derivatives up to order s, in L_p sense. The parameter q controls a further finer degree of smoothness. In a Bayesian setting, a prior on B _pq ^q is chosen following Abramovich, Sapatinas and Silverman (1998). We show that the optimal Bayesian estimator of f is then also a.s. in B _pq ^q if the loss function is chosen to be the Besov norm of B _pq ^q . Because it is impossible to compute this optimal Bayesian estimator analytically, we propose a stochastic algorithm based on an approximation of the Bayesian risk and simulated annealing. Some simulations are presented to show that the algorithm performs well and that the new estimator is competitive when compared to the more standard posterior mean.