Remarks on extremal problems in nonparametric curve estimation (Q1293841): Difference between revisions

The author investigates for \(\beta>0\) the following variational problems: \[ \bigl\|x(\cdot) \bigr\|_2 \to\inf\text{ subject to }x (0)=1, \;x(\cdot)\in \text{Lip} (\beta,1), \tag{*} \] where \(\text{Lip} (\beta,1)\) contains all continuous functions satisfying with \(m=[\beta]\) that: \[ \bigl|x^{(m)} (t)-x^{(m)}(t') \bigr|\leq 1\cdot |t-t'|^{\beta -m}\text{ for real }t,t'. \] This problem is related to minimax problems in nonparametric curve estimation. Explicit solutions for (*) are known for \(\beta\leq 1\) and \(\beta =2\) [see \textit{A.P. Korostelev}, Theory Probab. Appl. 38, No. 4, 737-743 (1993); translation from Teor. Veroyatn. Primen. 38, No. 4, 875-882 (1993; Zbl 0819.62034); \textit{A.T. Fuller}, Int. J. Control 35, 575-604 (1982; Zbl 0481.93069); \textit{L.H. Zhao}, Ann. Stat. 25, No. 2, 745-755 (1997; Zbl 0879.62033)]. The main result of the present paper shows that a solution \(x_\beta(\cdot)\) of the problem (*) (which is known to have compact support) satisfies \[ \int x_\beta(t) dt/ \bigl\|x_\beta (\cdot)\bigr \|^2_2= 2\beta/(2 \beta+1) \text{ for } \beta>0. \] This can be used to verify that optimal kernels providing asymptotically efficient kernel estimators do exist and to calculate exact constants in the asymptotic minimax risk. For the case \(\beta=2\) a discussion of the optimal and related kernels is given.