Optimal bounds for aggregation of affine estimators (Q1747732)

The author investigates the problem of aggregating a collection of affine estimators in a heteroscedastic regression model with Gaussian additive noise. The estimators are not independent of the data used for the aggregation. This leads to the question whether well-known minimax rates which hold for deterministic estimators remain valid or whether the variances of the estimators and their dependence increase the price to pay for aggregation because of statistical complexity. It is proven that for affine estimators, the minimax rate of aggregation is unchanged. The author defines a penalized procedure over the complex hull of the estimators and shows that it achieves sharp oracle inequalities in deviation for aggregation of affine estimators. The penalty term is inspired from the \(Q\)-aggregation procedure. Moreover, it is shown that the penalized aggregate is robust to non-Gaussian noise and to a certain variance misspecification.