Nonstandard problems of robust identification of regression models (Q1355322)

Our main concern is to give some new formulations of the problems of robust identification when the level of a priori information is low, and to construct, for some of them, procedures which guarantee perfect identification. In \S 1, we briefly introduce \(M\)-estimates, minimal-contrast (MC-) estimates, and joint estimates of the regression parameter and the functional of noise distribution. In \S 2, we give the scheme of constructing a procedure of estimating the regression parameter with guaranteed accuracy in the case where the noise distribution belongs to a union of several semi-canonical families (for each of those families there is a procedure of \(M\)-estimation with guaranteed accuracy, e.g., the asymptotically minimax one). In \S 3 extremal (minimax) problems are considered which arise in the robust identification if the traditional unbiasedness condition is absent. An upper bound for the accuracy of estimates of the regression parameter is found in the case where the noise distribution belongs to a union of semi-canonical families. In \S 4, some procedures of joint estimation of the regression parameter and a finite-dimensional functional of the noise distribution are analyzed.