No label defined (Q4778786)

Let \(X\) and \(Y\) be two independent random vectors with the density functions \(f(x |\theta_1)\) and \(f(x |\theta_2)\). For a vector \(a\) and a number \(c\) let the linear discriminant rule \(x\in\pi_1\;\text{if} a'x\leq c\), and \(x\in\pi_2\;\text{if} a'x>c\), for classification of objects from the populations \(\pi_1\) and \(\pi_2\) be given. Let \(P(\pi_1 |\pi_2)=F_1(c)\) and \(P(\pi_2 |\pi_1)=1-F_2(c)\) be the conditional probabilities of correct classification. The curve \(x=F_1(c),\;y=1-F_2(c),\;c\in(-\infty,\infty)\), is called the relative operating characteristic (ROC) curve of the discriminant. Let \(D(a)\) denote the area under the ROC curve. The area \(D(a)\) may be used as a qualitative characteristic of the discriminant. In this paper probabilities \(P(a'Y>a'X)\) are found and properties of the corresponding ROC curve are studied in the case of multi-dimensional Student and uniform distributions. Estimates for the probability \(P(b'Y>a'X)\) are proposed and their properties are investigated. For the entire collection see [Zbl 0979.00030].