Partitioning points by parallel planes (Q1827723)

A hyperplane is an \((n-1)\)-flat in \(\mathbb{R}^n\) that does not need to contain the origin. Let \(P(N,k,n)\) be the maximum number of ways in which \(N\) points in \(\mathbb{R}^n\) can be partitioned by \(k\) parallel hyperplanes, none of which contains any of the \(N\) points. The author shows that \(P(N,k,n)\leq \sum^{n+k-1}_{i=0} {Nk-1\choose i}\). This is an improvement on a result of \textit{S. Olafsson} and \textit{Y. S. Abu-Mostafa} [IEEE Trans. Pattern Anal. Mach. Intell. 10, No. 2, 277--281 (1988; Zbl 0709.94512)]. An earlier claimed improvement by the same authors is shown to be incorrect. Applications are pointed out in the theory of artificial neural networks.