No label defined (Q4037076)

[For the entire collection see Zbl 0754.00005.] Let \(X_ 1,\dots,X_ n\) be a random sample from a distribution with the density function \(f(x)\). When the density is only known to have a certain degree of smoothness, it is of interest to establish the best precision with which all such densities can be estimated. There exists considerable literature on the best obtainable convergence rates of density estimators. The first results of this type have been obtained, as far as we know, by \textit{N. N. Chentsov} [Soviet. Math., Doklady 3, 1559-1562 (1963); translation from Doklady Akad. Nauk SSSR 147, 45-48 (1962; Zbl 0133.118); Statistical decision rules and optimal inference (1982; Zbl 0484.62008)] who established the optimal rates for the integrated mean square error using orthogonal series expansions, reduction of the nonparametric risk to the parametric risk of estimation of a large number of Fourier coefficients, and a Cramér-Rao type inequality. Unfortunately, the densities covered by Chentsov's results have to be bounded away from zero, which obviously rules out many important distributions. We obtain the asymptotic lower bound for the integral risk of nonparametric density estimates using a method similar to Chentsov's: we reduce the nonparametric risk to the parametric risk of a strategically chosen family with a large number of parameters and then apply a version of the Hájek - Le Cam lower bound.