No label defined (Q4935656)

Recall that the random variable \(X\) has an inverse Gaussian distribution with parameters \(\mu> 0\) and \(\lambda> 0\), \(X\sim\text{IG}(\mu, \lambda)\), if it is absolutely continuous with density \[ f(x)= \biggl( \frac{\lambda} {2\pi x^3} \biggr)^{1/2} \exp \biggl[- \frac{\lambda} {2\mu^2} \frac{(x-\mu)^2} {x} \biggr], \qquad x> 0. \] (Of course, \(f(x)= 0\) for \(x\leq 0\).) It can be shown that all moments \(\mu_k= E[X^k]\), \(k= 1,2,\dots\), are finite. For any real \(r\) the power \(X^r\) is a well defined random variable and clearly all moments of \(X^r\) are also finite. We discuss here the following question: For what values of \(r\) is the distribution of \(X^r\), denoted by \({\mathcal L}(X^r)\), uniquely determined by its moment sequence? This is exactly the classical Stieltjes moment problem (for distributions concentrated on \((0, \infty)\)). It is well known that the IG distribution itself, i.e. \({\mathcal L}(X)\) is unique in terms of the moment sequence, or equivalently, that the IG distribution is M-determinate. (M stands for ``moment''.) This follows from the fact that the moment generating function \(E [\exp (tX)]\) exists in a neighborhood of the origin 0. One of the main results is: If \(-2\leq r\leq 2\), then the distribution \({\mathcal L}(X^r)\) is M-determinate, while for \(r< -2\) and \(r> 2\), \({\mathcal L}(X^r)\) is M-indeterminate. The proof is based on the so-called Krein-Lin techniques. Then we analyze the following interesting property of the moments of \(X\sim \text{IG} (\mu,\lambda)\) (for simplicity taking \(\mu= 1\)): \(E[X^{-k}]= E[X^{k+1}]\), \(k= 1,2,\dots\). It turns out that this property together with few other ones is not a characterization property of the IG distribution. An example is described in detail. Some additional remarks are made at the end of the paper, in particular, we give the exact formulation of two basic results in the Hamburger moment problem (for distributions on the whole real line).