Adaptive estimates of linear functionals (Q1326282): Difference between revisions

Let \(\ell_ 2\) be the space of square summable sequences. Consider the white noise model \(Y_ i = \theta_ i + \varepsilon_ i/ \sqrt{n}\), \(i = 1,2,\dots\), where the \(\varepsilon_ i\) are independent Gaussian errors and \(\theta = (\theta_ 1,\theta_ 2,\dots)\in \ell_ 2\). Here \(\Phi\) is a closed, convex symmetric subset of \(\ell_ 2\). The modulus of continuity \(w(\varepsilon)\) is given by \[ w(\varepsilon) = \sup \{2T(\theta): \| \theta\| \leq \varepsilon/2,\quad \theta \in \Phi\}, \] where \(\| \theta\|\) is the usual Hilbert space norm and \(T\) is a linear functional. Let \(\Phi_ t \subset \Phi_ s\) if \(a \leq s \leq t \leq b\). Let \(w_ t\) \((\varepsilon)\) be the modulus of continuity associated with \(\Phi_ t\). If \(0 < q_ s < q_ t < 1\) whenever \(s < t\) and if \[ \varliminf_{n \to \infty} n^{q_ b} \sup_{\Phi_ b} E(T_ n - T(\theta))^ 2 < \infty, \] then for all \(t < b\), \[ \varliminf_{n \to \infty} (n/\log n)^{q_ t} \sup_{\Phi_ t} E(\widehat{T}_ n - T(\theta))^ 2 > 0. \] Estimates of linear functionals \(T(\theta)\) are given which satisfy the above condition and attain the given lower bounds. Moreover, a sequence of quasi-optimal estimators is given which only suffer the penalty of the logarithm term on a small subset of \(\Phi_ t\).