Influence functionals for time series (with discussion) (Q1085911): Difference between revisions

In this very interesting paper influence functionals of parameter estimates in time series models are introduced and their properties studied. The authors consider a contaminated observations process of the form \(y_ t^{\gamma}=(1-z_ t^{\gamma})x_ t+z_ t^{\gamma}w_ t\), \(t=1,2,...\), \(0\leq \gamma \leq 1\), where \(x_ t\) is a core process (usually Gaussian), \(w_ t\) is a contaminating process and \(z_ t^{\gamma}\) is a zero-one process with \(P(z_ t^{\gamma}=1)=\gamma +o(\gamma)\). If \(T(\mu_ y^{\gamma})\) denotes the functional representation of a given estimate where the measures \(\mu_ y^{\gamma}\) for \(y_ t^{\gamma}\) are in an appropriate subset of the family of stationary and ergodic measures on \(({\mathbb{R}}^{\infty},\beta^{\infty})\), then the influence functional IF is a derivative of T along ''arcs'' traced by \(\mu_ y^{\gamma}\) as \(\gamma\to 0\). More precisely \[ IF(\mu_{\omega},T,\{\mu_ y^{\gamma}\})=\lim_{\gamma \to 0}(T(\mu_ y^{\gamma})-T(\mu_ y^ 0))/\gamma \] provided the limit exists. In other sections the authors deal with the following questions and problems: 1) the connection to Hampel's definition of influence curve is thoroughly discussed and a simple relationship established; 2) specific results for generalized M-estimates and RA estimates of first order autoregressive and moving average models are derived and definition of gross-error sensitivity for them introduced; 3) a class of generalized RA estimates based on IF is described and optimality properties of them are given; 4) some possible generalizations of the IF through applications to the white-noise test statistic and to spectral density estimation are sketched. The paper is supplemented by a very ample discussion.