Infinitely divisible Wald's couples. Examples linked with the Euler gamma and the Riemann zeta functions. (Q2485446)

The starting point of these investigations are pairs of infinitely divisible laws on the real line related to the \(\Gamma\)-function, cf. \textit{P. Hartman} [Ann. Sc. Norm. Super., Pisa, Cl. Sci., IV. Ser. 3, 267--287 (1976; Zbl 0386.34016)]. Pairs of real random variables \((X,H)\), with \(H\geq 0\) with Laplace transforms satisfying the equation \(E(e^{\lambda X}) E(e^{-\frac{1}{2}\lambda^2 H}) = 1\) for \(\lambda >0\) are called couples de Wald. The key result of Section I, Théorème I.1.1, provides the existence of infinitely divisible solutions of this equation defined by Lévy measures \(c\) on the half line with finite first moments. By construction, \[ E(e^{-\frac{1}{2}\lambda^2 H}) = \exp\left(-\int_0^{\infty} (e^{-\lambda x} + \lambda x -1) c(dx)\right)\text{ for }\lambda >0. \] In fact, \(X\) is centred with finite first absolute moment and \(H\) is nonnegative, hence a subordinator, and self-decomposable. It is shown that these so defined random variables resp. their distributions are closely related to various other examples which had been investigated in a different context. Section I provides properties of \((X,H)\), depending on the particular Lévy measure \(c\) and further examples. E.g., stopped Brownian motions had been studied by the authors, \textit{B. de Meyer} and \textit{P. Vallois} [Rev. Mat. Iberoam. 18, 541--586 (2002; Zbl 1055.60078)]. This are pairs \((B_T,T)\) where \(B_t\) is a Brownian motion and \(T\) is an independent stopping time. In fact, it turns out (Proposition I.5.1) that such pairs are infinitely divisible iff \(T\) is constant. A further class of examples is related to distributions investigated by \textit{E. Lukacs} [Theor. Probab. Appl. 13, 116--127 (1968) and Teor. Veroyatn. Primen. 13, 114--125 (1968; Zbl 0201.20701)]: To given \(c\) and the corresponding Wald pair \((X,H)\) the authors construct via subordination a further pair of random variables \((X,\widehat{H}) \) satisfying the equation \(E(e^{\lambda X})E(e^{i \lambda \widehat{H}}) = 1\) which is crucial in the above mentioned investigations of E. Lukacs. Section II is concerned with constructions of new families of pairs \((X_{\alpha}, H_{\alpha})\) and \((X_{\alpha, \beta}, H_{\alpha, \beta})\) defined by \(c\): The Lévy measures \(c\) are replaced by measures with exponential densities w.r.t. \(c\). These constructions are closely related to the Esscher transformation of Lévy processes (defined by exponential martingales). Hence these families of random variables are called Esscher families. The construction relies on functions \(F\) -- defined by \(c\) and the densities \(e^{-\alpha x}\) -- with crucial properties. In the sequel, in Sections III--V, special cases of \(c\) and \(F\) and corresponding Esscher families are investigated: The Lévy measures \(c(dx) = k_0 dx/x^{3-\mu}\) lead to \(\gamma\)-distributed random variables, in particular, the case \(\mu = 1\) leads to an example investigated by \textit{G. Letac} [Adv. Appl. Probab. 17, 911--912 (1985; Zbl 0579.60013)]. Section IV is concerned with examples where \(F\) is related to the \(\Gamma\)-function, in Section V \(F\) is defined by Riemann's zeta function and further related functions. The appendix contains some possible generalizations \((X, H^{(\mu)})\) satisfying similar equations (depending on the real parameter \(\mu\)).