On the solution of an optimal recovery problem and its applications in nonparametric regression (Q1269541): Difference between revisions

The author investigates an optimization problem which arises in various nonparametric statistical settings with Hölder function classes. Let \(g^*(t)= g^*_\beta(t)\) be the solution of the following optimization problem: \[ g(0)\to \sup,\;g\in\Sigma (\beta,1),\;\| g\|_2\leq 1,\tag{1} \] where \[ \| g\|_2= \left(\int^\infty_{-\infty} g^2(t)dt \right)^{1/2}, \quad \text{and} \quad \Sigma (\beta,1)= \Sigma_{[-\infty, \infty]} (\beta,1). \] Actually, being easier to be studied, the paper concentrates on the solution \(x_*(t)=x_*^\beta(t)\) of the dual optimization problem: \[ \| x \|_2 \to \inf,\;x\in\Sigma (\beta,1),\;x(0)=1. \tag{2} \] For \(\beta\leq 1\) the solution to the problem (2) is known. The case of integer \(\beta\) is examined and its relationship to the optimal control problem is pointed out. The main goal of the present paper is to prove the following important result: the compactness of the support of \(x_*\) for any \(\beta>1\). The application of the compactness is illustrated via regression estimation with the minimax Bahadur risk.