The necessity of strongly subadditive capacities for Neyman-Pearson minimax tests (Q1098183)

The main result is the following converse of \textit{P. J. Huber} and \textit{V. Strassen}'s Neyman-Pearson lemma for minimax tests [Ann. Stat. 1, 251- 263 (1973; Zbl 0259.62008)]: Let \({\mathcal P}\) be a weakly compact set of Borel probability measures on a Polish space X having the property that, for every Borel probability measure Q on X, there exist some \(P\in {\mathcal P}\) and some version g of dQ/dP such that, for arbitrary \(\alpha\in [0,1]\), the Neyman-Pearson test of level \(\alpha\) between P and Q determined by g is a minimax test of the same level between \({\mathcal P}\) and \(\{\) \(Q\}\). Then the upper envelope \(c\) of \({\mathcal P}\) is a strongly subadditive capacity, and \({\mathcal P}\) is the convex set of all Borel probability measures on X dominated by \(c.\) For X finite, this result is due to Huber and Strassen, whereas a modified version for Polish spaces involving all finite sub-\(\sigma\)- algebras of the Borel \(\sigma\)-algebra was proved by \textit{T. Bednarski} [ibid. 10, 226-232 (1982; Zbl 0496.62004)].