A general approach to Bahadur-Kiefer representations for \(M\)-estimators (Q1376534)

Let \(X_1,X_2,\dots\) be independent, identically distributed random variables (r.v.s) defined on the probability space \((S,{\mathcal S},P)\), and let \(X\) be a r.v. distributed as the \(X_i\)s and whose probability law depends on a parameter \(\theta\) in an open set \(\Theta\) in \({\mathfrak R}^m\). Let \(h\) be a function defined on \(S\times\Theta\) into \({\mathfrak R}^d\), and let \(\theta_0\) be a parameter point characterized by the relation \(Eh(X,\theta_0)=0\). Then, roughly speaking, an \(M\)-estimate of \(\theta_0\) is a statistic \(\theta_n= \theta_n (X_1, \dots,X_n)\) such that \(n^{-1}\sum^n_{i=1} h(X_i,\theta_n)\) is close to 0 in some sense. The authors consider general classes of \(M\)-estimates and study the asymptotic behavior of their first order expansion. They obtain as special cases several results previously established, including the Bahadur-Kiefer representation for quantiles. Under suitable regularity conditions, results obtained by the authors are of the following nature. With \(a_n= (n/2\log \log n)^{1/2}\) and \(H'(\theta_0)\) denoting the derivative, it is shown that \(a^2_n(H_n(\theta_0) +H'(\theta_0)(\theta_n -\theta_0))\) is relatively compact and its set of limit points is also characterized. Here \[ H(\theta)=Eh(X,\theta) \text{ and }H_n(\theta)=n^{-1}\sum^n_{i=1} h(X_i, \theta). \] Also, the asymptotic distribution of \(n(H_n(\theta_0) +H'(\theta_0) (\theta_n- \theta_0))\) is determined. Finally, it is shown that, with probability one, \[ \biggl\{a_n^{3/2} \bigl(H_n (\theta_0)+ H'(\theta_0) (\theta_n -\theta_0) \bigr)\biggr\},\;n=1,2, \dots \] is almost surely relatively compact and its limit set is identified.