scientific article; zbMATH DE number 3901857

From MaRDI portal
Publication:3680113





Let \({\mathcal F}\) be a family of continuous distribution functions and suppose n i.i.d. \(F\in {\mathcal F}\) real r.v.'s \(X_ 1, X_ 2,..., X_ n\) are observed sequentially with the object to choose the largest. Neither recall nor future knowledge is allowed and one choice must be made. Denote: \(L_ n=\max (X_ 1, X_ 2,..., X_ n)\) and \({\mathcal S}_ n\) the class of all stopping rules \(\sigma\) (n) for the X's such that \(\sigma\) (n)\(\leq n.\) There are three cases possible: (1) The no information case, when \({\mathcal F}\) contains all c.d.f.'s (the secretary problem). In this case the asymptotic probability for the optimal s.r. \(\rho\) (n) is equal to \(\lim_{n\to \infty}P(X_{\rho (n)}=L_ n)=e^{-1}\). (ii) The full information case, when \({\mathcal F}\) contains a single c.d.f. Then the optimal s.r.'s \(\sigma\) (n) gives the asymptotic probability \(\alpha\) \(\doteq 0,5802\). (iii) The partial information case, when \({\mathcal F}\) is neither the class of all c.d.f.'s nor a single c.d.f. \textit{S. M. Samuels} [J. Am. Stat. Assoc. 76, 188-197 (1981; Zbl 0459.62070)] has shown that for the family \({\mathcal F}\) of all uniform c.d.f.'s the minimax s.r.'s are the same as in the no information case. The author [Ann. Stat. 8, 1171-1174 (1980; Zbl 0459.62071)] has proved that for \({\mathcal F}\) containing all uniform U[\(\theta\)-1/2, \(\theta\) \(+1/2]\) c.d.f.'s the minimax s.r.'s \(\tau\) (n) give \(\lim_{n\to \infty}P(X_{\tau (n)}=L_ n)\doteq 0,4352.\) In the paper, for an arbitrary location, scale and location-scale parameter family \({\mathcal F}\), sufficient conditions are given for the existence of s.r.'s \(\tau\) (n) having asymptotic probability of choosing the largest \(\alpha\), i.e. as in the full information case.











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3680113)