Discrimination of hypotheses for Gaussian measures, and a geometrical characterization of Gaussian distribution (Q790509)

Let X be a linear space of real functions, \({\mathcal B}^ a \sigma\)-algebra generated by cylindrical sets of X, \(P_ 0\) and \(P_ a\) Gaussian measures on (X,\({\mathcal B})\) with correlation function B and the means O and \(a\in X\), respectively. Denote by \(M=\{a: P_ a\sim P_ 0\}\) the linear space of admissible shifts of the measure \(P_ 0\). Consider the problem of the minimax discrimination of two complex hypotheses for a single observation \(x\in X: H_ 0: P=P_ a\), \(a\in A\subset M\); \(H_ 1: P=P_ b\), \(b\in B\subset M\), \(A\cap B=\emptyset.\) A nonrandomized criterion for verifying the hypotheses \(H_ 0\) and \(H_ 1\) is defined using the critical domain \({\mathcal D}\in {\mathcal B}:\) \(x\in {\mathcal D}\Rightarrow H_ 1\), \(x\not\in {\mathcal D}\Rightarrow H_ 0\). Besides, the probabilities of errors of type I and type II are, respectively, equal to \(\alpha_{{\mathcal D}}=\sup_{a\in A}P_ a({\mathcal D})\), \(\beta_{{\mathcal D}}=\sup_{b\in B}P_ b({\mathcal D}^ c)\), \({\mathcal D}^ c=X\backslash {\mathcal D}\). In the paper under review the quantity \(\beta(\alpha,A,B)=\inf_{{\mathcal D}:\alpha_{{\mathcal D}}\leq \alpha}\beta_{{\mathcal D}}\) is found and the corresponding optimal domain is defined. Further, an inequality is also proved for Gaussian measures and some of its applications are given.