Bootstrapping the mean integrated squared error (Q2365602)

Let \(X_ 1,\dots,X_ n\) be a sequence of i.i.d. real random variables from an unknown density \(f\). For a given kernel \(K\) and a bandwidth \(h>0\), denote with \(f_ h\) the associated Parzen-Rosenblatt estimator of \(f\). There exists a huge literature on how to choose \(h\) in an optimal way. In particular, the integrated squared error (ISE) and its mean (MISE) yield criteria functions which cannot be optimized in practice and therefore need to be replaced by proper substitutes. In the present paper the author suggests to pick a pilot estimator \(f_ g\) and then to draw a bootstrap sample \(X^*_ 1,\dots,X^*_ n\) from \(f_ g\). This leads to bootstrap versions of ISE and MISE which may be minimized in practice by some \(h^*\), say. Several large sample results are provided which compare the minimizers of ISE and MISE with their bootstrap analogues.