Condition of constancy of the regression for monomials of independent random variables (Q805112)

A monomial of the random variables (r.v.s.) \(X_ 1,...,X_ n\) is the r.v. \(M=X_ 1^{\alpha_ 1}...X_ n^{\alpha_ n}\), where \(\alpha_ 1,...,\alpha_ n\) are real numbers. A positive r.v. X is lognormally distributed if the r.v. log X is normally distributed. If \(X_ 1,...,X_ n\) are independent positive, possibly nonidentically distributed r.v.s, then the theorem of Darmois-Skitovich on independence of linear forms immediately leads to the following result: independence of the monomials \(M_ 1=M_ 1^{\alpha_ 1}...X_ n^{\alpha_ n}\) and \(M_ 2=X_ 1^{\beta_ 1}...X_ n^{\beta_ n}\) implies lognormality of the r.v.s \(X_ j\) effectively entering both monomials. Let us now consider the problem of constancy of regression of one monomial on one or several other monomials. The corresponding regression variants of the Darmois-Skitovich theorem do not contribute anything new in this context, except raising the hope that in an appropriately posed problem the lognormal distribution will play the same role for monomials that the normal distribution plays for linear forms. In analytical terms, the problem of constancy of the regression of one linear form on other linear forms leads to a differential equation for characteristic functions whereas the same problem for monomials leads to a difference equation for Mellin transforms. Therefore, in order to separate out the lognormal distribution, we have to impose the condition of constancy of the regression not for one monomial, as with linear forms, but for a pair of monomials in a ``general position''. We consider the general case of nonidentically distributed r.v.s and arbitrary (monomial) regressors. Furthermore, we examine a special case when the original r.v.s. are identically distributed and the regressor is their product (or geometric mean).