Characterization of positively correlated squared Gaussian processes (Q2447334)

The authors solve a conjecture raised by \textit{S. N. Evans} [J. Multivariate Anal. 36, No.~2, 199--203 (1991; Zbl 0719.62063)] who characterized the positively correlated squared Gaussian vectors. They extend this characterization from squared Gaussian vectors to permanental vectors. As additional results, they obtain several equivalent formulations of the property of infinite divisibility for squared Gaussian processes. Evans conjectured that, given a centered Gaussian vector \(\eta = (\eta_i)_{1 \leq i \leq n}\), the squared central Gaussian vector \(\eta^2=(\eta^2_i)_{1 \leq i \leq n}\) is ``positively correlated'' iff there exists a function \(\sigma\) from \(\{ 1 \leq i \leq n \}\) into \(\{ -1, 1 \}\) such that \((\sigma(i)\eta_i)_{1\leq i \leq n}\) is positively correlated. The present authors prove the following. Theorem. A squared central Gaussian vector is ``positively correlated'' if and only if it is infinitely divisible. There are several necessary and sufficient conditions for a squared centered Gaussian vector to be infinitely divisible. The authors use \textit{D. Vere-Jones}' characterization [N. Z. J. Math. 26, No.~1, 125--149 (1997; Zbl 0879.15003)] of infinitely divisible squared Gaussian vectors to establish three other equivalent necessary and sufficient conditions for a squared centered Gaussian process with continuous covariance to be ``positively correlated''. They raise two questions: (1) Which are the centered Gaussian processes with a covariance equal to a Green function? (2) Which are the centered Gaussian processes \(\eta\) such that the law of \(\eta^2\) under \(\mathbb{E}[\eta_a\eta_b, \cdot]\) is a positive measure? An answer to the first question has been given in [\textit{N. Eisenbaum} and \textit{H. Kaspi}, Ann. Probab. 34, No. 2, 728--742 (2006; Zbl 1102.60031)]. The authors state a corollary which provides three alternative formulations of this answer. One of them is a solution to the Evans conjecture for processes. Another one answers also to the second question. Once the question of the characterization of ``positively correlated'' squared centered Gaussian processes is solved, the authors suggest the same question for shifted Gaussian processes. In particular, given a centered Gaussian process \((\eta_x)_{x\in E}\) and a real number \(r\), when is the process \(((\eta_x +r)^2)_{x\in E}\) ``positively correlated''? The authors find a sufficient condition for the realization of that property: the infinite divisibility of \(((\eta_x + r)^2)_{x\in E}\). But here, the characterization, in terms of the covariance of \(\eta\), of that condition for a fixed \(r\) is not known. A characterization obtained by assuming that the set \(E\) contains more than two elements is used to enunciate another sufficient condition for \(((\eta_x + r)^2)_{x\in E}\) to be positively correlated for every \(r\).