Multinomial approximations for nonparametric experiments which minimize the maximal loss of Fisher information (Q1114266)

Let \(\omega =\{A_ 1,...,A_ k\}\) be a partition of the real line into k intervals such that each \(A_ i\) has measure 1/k under the measure \(H=h\lambda_ 1\), \(\lambda_ 1=\) Lebesgue measure. Given a probability measure \(P_ 0\), denote, for any \(g\in L^ 2(P_ 0)\), by \(\pi_{\omega}g\) the conditional expectation of g with respect to the \(\sigma\)-field generated by \(\omega\). The authors derive an expansion of \(\| g-\pi_{\omega}g\|^ 2_ 2\) and then find a minimax solution for the problem \[ \inf_{h}\sup_{g}\| g-\pi_{\omega}g\|^ 2_ 2, \] where the sup extends over a class of functions serving as asymptotic directions for sequences of local alternatives. The quantity \(\| g-\pi_{\omega}g\|^ 2_ 2\) may be interpreted as the loss of Fisher-information when replacing a statistical model by a model obtained by grouping.