Bandwidth choice for differentiation (Q1095530)

The r-th derivative of a smooth function \(f^{(r)}(x)\) is observed with error on a discrete set of points. A nonparametric linear estimate of this derivative \(f^{(r)}(x,\lambda)\) is given, where \(\lambda\) is a smoothing parameter. The task is to choose \(\lambda^*_ n\) as to minimize the mean integrated square error \[ MISE_ n(\lambda)=E\int [f^{(r)}(x)-f^{(r)}(x,\lambda)]^ 2dx. \] A nearly unbiased estimate of \(MISE_ n(\lambda)\) is proposed. A parameter \({\bar \lambda}_ n\) for minimizing this estimate is considered. It is proven, for the special case of a tapered Fourier series estimate, that under certain conditions \({\bar \lambda}_ n\) is a consistent estimate of \(\lambda^*_ n\). The asymptotic behavior of \({\bar \lambda}_ n\) is investigated.