No label defined (Q4213628)

Let \((X_{0},X_{1},...)\) be a time series with state space \(\Xi_{0},\Xi_{1},...\), resp. This work deals with the robust estimation of a real parameter of the conditional density of \(X_{n}\) given \(X_{1},...,X_{n-1}\). Formally, let \(\Theta \) be an open subset of \(\mathbb R^{1}\); for each \(n=1,2,...\), each \(i=1,...,n\), each \(x_{0},x_{1},...,x_{n-1}\) in \(\Xi_{0},\Xi_{1},...,\Xi_{n-1}\), resp, and each \(\theta \in \Theta \) let \(f_{i}^{n}(. ,\theta | (x_{0},x_{1},...,x_{i-1}))\) be a density function over \(\Xi_{i}\). The family formed by such functions is the assumed model nominal system, i.e., once observed \((X_{0},X_{1},...,X_{n-1},X_{n})= (x_{0},x_{1},...,x_{n-1},x_{n})\), the aim is to estimate the parameter \(\theta \in \Theta\) assuming that the density \(X_{i}\) conditioned on \( (x_{0},x_{1},...,x_{i-1})\) is \(f_{i}^{n}(. , \theta | (x_{0},x_{1},...,x_{i-1}))\). Robust estimation in this case consists in finding estimators alternative to the maximum likelihood ones that are as effective as these when the nominal model is verified, and which are effective even when the nominal model is not valid. In order to pose the problem precisely, it is necessary to define a neighbourhood of the nominal system which will be assumed as the alternative model: for each \(n=1,2,...\), each \(i=1,...,n\), each \(x_{0},x_{1},...,x_{n-1}\) in \(\Xi_{0},\Xi_{1},...,\Xi_{n-1}\), resp, and each \(\theta \in \Theta\) let \(H_{i}^{n}(. 0, \theta | (x_{0},x_{1},...,x_{i-1}))\) be a function over \(\Xi_{i}\), and consider \( H=(H_{i}^{n})_{i=1,...,n; n\geq 1}\). The density function over \(\Xi_{i}\) is therefore defined as \[ f_{i}^{n,H}(. ,\theta | (x_{0},x_{1},...,x_{i-1}))= \] \[ =f_{i}^{n}(. , \theta | (x_{0},x_{1},...,x_{i-1}))[1+n^{-1/2}H_{i}^{n}(. , \theta | (x_{0},x_{1},...,x_{i-1}))]. \] The sequence \(H\) must belong to a set of functions satisfying convenient regularity properties. The proposed robust estimators are defined as follows: For each \(n=1,2,...\), each \(i=1,...,n\), each \(x_{0},x_{1},...,x_{n-1}\) in \(\Xi_{0},\Xi_{1},...,\Xi_{n-1}\), resp., and each \(\theta \in \Theta\) let \(\psi_{i}^{n}(. , \theta | (x_{0},x_{1},...,x_{i-1}))\) be a function over \(\Xi_{i}\), with similar properties as those of the score functions used in the theory of M-estimators for the location model. Let \(\Psi =(\psi_{i}^{n})_{i=1,...,n}\), \(n\geq 1\); the CLAN-estimator (consistent linear asymptotically normal) is any sequence \(T^{\Psi }=(T_{n}^{\Psi })_{n\geq 1}\) of random variables such that: (i) for each \(n\), \(T_{n}^{\Psi}\) is a function of \((X_{0},X_{1},...,X_{n-1},X_{n})\); (ii) \(\psi_{i}^{n}(X_{i},\theta | T_{n}^{\Psi }(X_{0},X_{1},...,X_{i-1}))@>P_{\theta}^{n}>>0\), \(n\rightarrow \infty\); where \(P_{\theta }^{n}\) is the probability over \(\Xi_{0}\times\Xi_{1}\times...\times\Xi_{n}\) induced by the conditional densities \((f_{i}^{n}(.,\theta))_{i=1,...,n}\); (iii) under \((P_{\theta }^{n})_{n\geq 1}\) one has that \[ \mathcal{L}(\sqrt{n}(T_{n}^{\Psi }-\theta))\rightarrow \mathcal{N}(0,\Gamma ^{\Psi}(\theta)/(\gamma^{\Psi}(\theta))^{2}), \] with \(0<\Gamma^{\Psi}(\theta),\;\gamma^{\Psi}(\theta)<\infty\), for every \(\theta \in \Theta\). With convenient properties of H, (ii) and (iii) are extended to \( (P_{\theta }^{n,H})_{n\geq 1}\), for every \(\theta\in\Theta\), where \( P_{\theta}^{n,H}\) is the probability over \(\Xi_{0}\times\Xi_{1}\times...\times\Xi_{n}\) induced by the conditional densities \((f_{i}^{n,H}(.,\theta))_{i=1,...,n}\). In this fashion, it is possible to deduce an explicit expression for the asymptotic mean square under \((P_{\theta}^{n,H})_{n\geq 1}\): \[ D(\Psi,H;\theta):=\lim_{a\rightarrow\infty}\lim_{n\rightarrow\infty}E_{\theta}^{n,H}[(\min(a,n(T_{n}^{\Psi}-\theta)^{2})]. \] Regarding this quantity, the minimax optimisation problem for the class of CLAN estimators is solved, obtaining as optimal score functions, with respect to the parameter \(\theta\), the Huber \(\psi\) functions (trimmed maximum likelihood scores). A great deal of the work is devoted to showing that the proposed model of conditional densities extends several particular models already studied in robustness: the case of independent, non-identically distributed observations, stationary ergodic Markov chains, stationary ergodic AR(1) and MA(1) models, innovation contamination and replacement models [\textit{D.R. Martin} and \textit{V.J. Yohai}, Ann. Stat. 14, 781-818 (1986; Zbl 0608.62042)], \(ARMA(p,q)\) models with parameters that are functions of a single real parameter. The work is mathematically demanding, and hard to read for professionals not possessing good working knowledge in measure theory, stochastic processes, etc. Future works providing computational implementations of these ideas will allow their assessment with real data and/or with Monte Carlo experiences.