Mutual characterizations of the gamma and the generalized inverse Gaussian laws by constancy of regression (Q2767481)

This paper contains interesting results on the mutual characterization of the generalized inverse Gaussian (GIG) and gamma law distributions through the constancy of regression. Let \(X\) and \(Y\) be non-negative, independent random variables, and let \(U=1/(X+Y)\) and \(V=1/X-1/(X+Y)\). According to a known result of \textit{H. Matsumoto} and \textit{M. Yor} [Nagoya Math. J. 162, 65-86 (2001; Zbl 0983.60075)], if \(X\) follows a GIG law, \(Y\) is a gamma distribution, and \(X\) and \(Y\) are independent, then \(U\) and \(V\) defined above are also independent with the distributions GIG and gamma, respectively. The converse of this result, relying on the independence property of \(U\) and \(V\), and being a simultaneous characterization of the GIG and gamma laws, has been proved by \textit{G. Letac} and \textit{J. Wesołowski} [Ann. Probab. 28, No. 3, 1371-1383 (2000; Zbl 1010.62010)].NEWLINENEWLINENEWLINEThe present paper generalizes the result of Letac and Wesołowski, replacing the independence condition on \(U\) and \(V\) by the constancy of regression of \(V\) on \(U\), under suitable moment conditions. Theorem 1 (Section 2) proves that if \(E(X)\) and \(E(1/X)\) are finite, \(Y\) is a gamma distribution, and the regression of \(V\) on \(U\) is constant, then \(X\) follows a GIG distribution. A parallel result is obtained (Theorem 2) if in Theorem 1 the properties of \(X\) and \(Y\) are interchanged. Section 3 considers a dual problem for the constancy of regression of \(1/V\) on \(U\). Thus the price to be paid for relaxing the independence condition on \(X\) and \(Y\) is that the distribution law characterization of the two random variables is not simultaneous; it is necessary to assume one law for the first variable in order to obtain the other law for the second variable.