Using the bootstrap to estimate mean squared error and select smoothing parameter in nonparametric problems (Q756327)

The paper is concerned with the clever idea of using bootstrap samples of essentially smaller size \(n_ 1\) than the size of the original sample n. More precisely, the author considers a bootstrap version \(f^*(\cdot | n_ 1,h_ 1)\) of the kernel density estimate \(\hat f(\cdot | n,h)\) and proves, in particular, that quantities like \[ (1)\quad E[\hat f(x| n_ 1,h_ 1)-f(x)]^ p,\text{ and } (2)\quad E[\hat f^*(x| n_ 1,h_ 1)-\hat f(x| n,h)]^ p \] (which obviously take account of both variance and bias of \(\hat f\) and \(\hat f^*)\) are close to each other for \(n_ 1<cn^{1-\delta}.\) The same statment concerning integral (in x) versions of (1) and (2) is proved. Some emphasis is on the problem of bootstrap estimation of a bias.