Contribution to the bandwidth choice for kernel density estimates (Q2271691)

The authors propose a new technique of bandwidth selection for univariate kernel density estimates \(\hat f_{h,K}\) as solutions to the equation \[ h=(2nk)^{-1}\left (V(K)/\int(\text{bias}\;\hat f_{h,K})^2dx\right), \] where \(n\) is the sample size, \(k\) is the order of the kernel \(K\), \(V(K)=\int K^2(x)dx\), and \[ \text{bias}\;\hat f_{h,K}=\int \hat f_{h,K}(x-hy)K(y)dy-\hat f_{h,K}(x) \] is an estimate of the bias of \(\hat f_{h,K}\). The equation is solved iteratively using the Steffensen acceleration method. Results of simulations are presented.