Sharp minimax adaptation over Sobolev ellipsoids in nonparametric testing (Q1684148)

The authors study minimax results in testing the following hypothesis: The sequence of observations is given by the sum of a deterministic signal sequence and Gaussian noise with mean 0. Under the null hypothesis, the signal is equal to 0, while under the alternative, the \(l_2\)-norm is bigger than a \(\rho\). Both the variance of the noise and the lower bound \(\rho\) under the alternative converge to 0 as \(n\) goes to infinity. Additionally, it is assumed that the signal belongs to a ball with radius \(M\) with respect to a Sobolev type norm with smoothness parameter \(\beta\). The authors show that the adaption over the radius \(M\) is not possible. In contrast, they prove that the adaptation is possible if the assumption on the decay of the lower bound \(\rho\) for the signal is relaxed. Furthermore, the authors show that sharp minimax adaptation simultaneously over the radius \(M\) and the smoothness parameter \(\beta\) is possible.