Mean Square Error for Uniform-Kernel Estimate and Nearest Neighbor Estimate of Probability Density Function

Abstract:

      In this paper we consider the Mean Square Error (MSE) of two uaual estimates of density function f(x) at a point x: The uniform kernel estimate f_n(x) and the NN estimate fn(x). we- show that when f is differentiable for sufficiently high order at x. these MSE can be expanded in a form E(f_n(x)-f(x))²=A₁(x)n^-4/5 +A₂(x)n^-1+A₃(x)n^-6/5+…;E(f_n(x)-f(x))²=B₁(x)n^-4/5 +B₂(x)n^-1+B₃(x)n^-6/5+… And if we suitably choose the parameters in f_n and f_n to make A₁(x) and B₁(x)to assume its minimunm value, then we also, have A₂(x) =B₂(x) but A₃(X) differs form B₃(X). This result shows that while the two estimates are not identical with respect to MSE. each one can be superior to the other in various special cases.