Asymptotically minimax estimation of a function with jumps (Q1385005)

The author considers the asymptotic behavior of the minimax risk of nonparametric regression estimators in the Gaussian white noise model, where the risk is defined with respect to \(L_2\)-loss and the unknown function is assumed to be smooth except for an unknown although finite number of jumps, i.e. the underlying function class is a piecewise Sobolev ''ellipsoid''. It is known that the convergence rate of the optimal estimators is not affected by a finite number of discontinuities. In the present paper it is proved that even the same asymptotic optimal constant, known as Pinsker's constant, can be attained. The author does not assume that the number and locations of the jumps are known; based on an estimator of the jump points she proposes a projection-type estimator.