Asymptotic sufficiency. I: Regular cases (Q799042)

Higher order asymptotic sufficiency is investigated in the classical parametric iid case: the parameter is finite-dimensional, the support of the distributions does not depend on the parameter, and the log- likelihood is a smooth function of the parameter. For the concept of higher order asymptotic sufficiency see \textit{J. K. Ghosh} and \textit{K. Subramanyam}, Sankhyā, Ser. A 36, 325-358 (1974; Zbl 0321.62035) or \textit{R. Michel}, ibid., Ser. A 40, 76-84 (1978; Zbl 0414.62009). It is shown that the classical statistics containing a sequence of asymptotic maximum likelihood estimators and derivatives of the log- likelihood function up to order k, are asymptotically sufficient of order \(O(n^{-k/2})\). This refines results of Michel and of \textit{T. Suzuki [Osaka J. Math. 15, 575-588 (1978; Zbl 0416.62003) and ibid. 17, 245-252 (1980; Zbl 0441.62028)]. It has been shown by \textit{J. Pfanzagl}} [Metrika 19, 89-97 (1972; Zbl 0253.62016)] that this order is optimal in the case \(k=1\). It seems to be impossible to improve the order \(O(n^{-k/2})\) for \(k\neq 1\).