Georgia Institute of Technology College of Sciences

We study the nonparametric regression model (X1 , Y1 ), ...(Xn , Yn ) , where (Xi )i≥1 is the deterministic design and (Yi )i≥1 is a sequence of real random variables. Assume that the density of Yi is known and can be written as g (., f (Xi )) , which depends on a regression function f at the point Xi . The function f is assumed smooth, i.e. belonging to a Hoelder ball or a Nikol’ski ball. The aim is to estimate the regression function from the observations for two error risks (pointwise and global estimations) and to find the optimal estimator (in the sense of rates of convergence) for each density g . We are particularly interested in the study of irregular models, i.e. when the Fisher information does not exist (for example, when the density g is discontinuous like the uniform density). In this case, the rate of convergence can be improved with the use of nolinear estimators like Maximum likelihood or bayesian estimators. We use the locally parametric approach to construct a new local version of bayesian estimators. Under some conditions on the likelihood of the model, we propose an adaptive procedure based on the so-called Lepski’s method (adaptive selection of the bandwidth) which allows us to construct an optimal adaptive bayesian estimator. We apply this theory to several models like multiplicative uniform model, shifted exponential model, alpha model, inhomogeous Poisson model and Gaussian model