Asymptotic distribution of regression M-estimators (Q5945257)

We consider the following linear regression model: \[ Y_i=Z_i' \theta_0+ U_i,\quad i=1,\dots,n, \] where \(\{U_i\}_{i=1}^\infty\) is a sequence of \(\mathbb{R}^m\)-valued i.i.d. r.v.'s; \(\{Z_i\}_{i=1}^\infty\) is a sequence of i.i.d. \(d\times m\) random matrices; and \(\theta_0\) is a \(d\)-dimensional parameter to be estimated. Given a function \(\rho:\mathbb{R}^m \to\mathbb{R}\), we define a robust estimator \(\widehat\theta_n\) as a value such that \[ n^{-1}\sum^n_{i=1} \rho(Y_i-Z_i'\widehat \theta_n)=\inf_{\theta \in\mathbb{R}^d} n^{-1}\sum^n_{i=1} \rho(Y_i-Z_i' \theta). \] We study the convergence in distribution of \(a_n(\widehat \theta_n-\theta_0)\) in different situations, where \(\{a_n\}\) is a sequence of real numbers depending on \(\rho\) and on the distributions of \(Z_i\) and \(U_i\). As a particular case, we consider the case \(\rho(x)=|x|^p\). In this case, we show that if \(E[\|Z\|^p+ \|Z\|^2] <\infty\), either \(p>1/2\) or \(m \geq 2\), and some other regularity conditions hold, then \(n^{1/} (\widehat \theta_n -\theta_0)\) converges in distribution to a normal limit. For \(m=1\) and \(p=1/2\), \(n^{1/2} (\log n)^{-1/2}(\widehat \theta_n-\theta_0)\) converges in distribution to a normal limit. For \(m=1\) and \(1/2> p>0\), \(n^{1/(3-2p)}(\widehat \theta_n- \theta_0)\) converges in distribution.