Local Adaptivity of Kernel Estimates with Plug-in Local Bandwidth Selectors
The authors consider the white noise model \(dY(t)=m(t) dt + \sigma dW(t)\), \(0\leq t\leq 1\), where \(W(t)\), \(t\geq 0\), is a Brownian motion and \(m\) is an unknown (regression) function. They study the asymptotic behaviour as \(\sigma\to 0\) of the kernel estimates \({\hat m}_{h}(x)= \int K_{h}(x-t) dY(t)\), where \(K_{h}(x)=h^{-1}K(x/h)\) is the rescaled kernel and \(h\) denotes the bandwidth of the kernel estimate \({\hat m}_{h}.\) The authors consider global rates of convergence over different function classes: Sobolev classes, bounded variation function classes, classes of convex functions and classes of monotone functions. The authors consider plug-in bandwidth selectors \({\bar h}(x)\) which depend only on \({\tilde m}_{g}(x)\), where \({\tilde m}_{g}(x)\) is a kernel estimate with kernel \(L_{g}(x)=g^{-1}L(x/g)\) and bandwidth \(g.\) \(K\) and \(L\) are twice continuously differentiable, positive and symmetric with compact support. For example, it is proved that under some conditions the following theorem holds true: Theorem 2. The estimate \(\hat m\) has a Gaussian distribution with mean \[ m_{\hat h(x)}(x)+({\tilde m}_{g}{}(x)- m_{L,g}{}(x))\biggl[\int K_{\hat h(x)} L_{g}{}\biggr]\cdot\biggl[\int (L_{g}{})^2 \biggr]^{-1} \] and variance \[ \sigma^2\biggl\{\int K^2_{\hat h(x)}- \biggl[\int K_{\hat h(x)}L_{g}{}\biggr]^2 \biggl[\int (L_{g}{})^2\biggr]^{-1} \biggl\}. \]
This page was built for publication: Local Adaptivity of Kernel Estimates with Plug-in Local Bandwidth Selectors
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4255147)
