On adaptive estimation in stationary ARMA processes (Q1821447)

The paper deals with adaptive construction of locally asymptotically minimax (LAM) estimators for stationary ARMA processes with independent and identically, but not necessarily normally distributed innovations. First the local asymptotic normality (LAN) for this model is proved using the sufficient conditions for LAN given by \textit{G. G. Roussas} [Z. Wahrscheinlichkeitstheor. Verw. Geb. 47, 31-46 (1979; Zbl 0377.62011)]. Then a construction of LAM estimators is suggested if only \(\sqrt{n}\)- consistent initial estimators are available. As these estimates can depend on the distribution of the innovations strongly adaptive estimators are suggested finally which are optimal in LAM sense for a wide class of symmetric innovation distributions. The kernel estimators for the score function -f'/2f (f is the density of the innovation distribution) are used for this purpose. The suggested construction of the adaptive estimators generalizes the results of \textit{R. Beran} [Ann. Inst. Stat. Math. 28, 77-89 (1976; Zbl 0362.62093)] and \textit{P. J. Bickel} [Ann. Stat. 10, 647-671 (1982; Zbl 0489.62033)]. A comparative simulation study concludes the paper.