Asymptotic behavior of M-estimators in continuous-time non-linear regression with long-range dependent errors (Q4780939)

Nonlinear regression models with independent or weakly dependent errors were considered by many authors. First results on nonlinear regression with long-range dependent errors were obtained by \textit{P.M. Robinson} and \textit{F.J. Hidalgo} [Ann. Stat. 25, No. 1, 77-104 (1997; Zbl 0870.62072)]. They proposed conditions under which some estimates of a parameter of nonlinear regression with errors in discrete-time models are consistent. The asymptotic distribution of \(M\)-estimators in nonlinear regression models with discrete time and long-range dependent noise processes were proposed by \textit{H.L. Koul} [Lect. Notes Stat. 115, 272-290 (1996)]. This paper discusses the asymptotic distributions of a class of \(M\)-estimators of a vector parameter of continuous-time nonlinear regression models with Gaussian and non-Gaussian long-range dependent noise. It has very close relations with other papers of the present authors [see, e.g., Theory Probab. Math. Stat. 63, 65-85 (2001; Zbl 0988.62015)]. A more precise description of the problem looks as follows. Let \(\xi(t)\) be a measurable, mean-square continuous Gaussian process with \(E\xi(t)=0\) and covariance function \[ B(t)=(1+t^2)^{-\alpha/2},\quad 0<\alpha<1,\quad t\in R^1. \] Let \(\Theta\subset R^q\) be a bounded open set, and let \(\Theta^{c}\) be the closure of \(\Theta\) in \( R^q\). The authors consider the regression model \(X(t)=g(t,\theta)+G(\xi(t)),\;t\in [0,T],\) where \(g(t,\theta):[0,\infty)\times\Theta^{c}\to R^1\) is a measurable function, \(\theta=(\theta_1,\dots,\theta_q)\in \Theta\); the random process \(\xi(t),\;t\in R^1\), has the spectral representation \(\xi(t)=\int_Re^{it\lambda}\sqrt{f_{\alpha}(\lambda)}W(d\lambda)\), where \(W(d\lambda)\) is a Gaussian white noise random measure and \(f_{\alpha}(\lambda)\) is the spectral density of the process. The nonlinear Borel function \(G:R^1\to R^1\) is such that \({\mathbf E}G(\xi(0))=0\). The problem is to estimate the unknown parameter from observations of the process \(X(t)\) for \(t\in[0,T]\) as \(T\to\infty\). To estimate \(\theta\), the \(M\)-estimator generated by a smooth function \(\rho(x)\colon R\to R\) is used.