Mean matrices and infinite divisibility (Q884408)

It is well-known that if \(A=[a_{ij}]\) is a positive semidefinite matrix, then its ``Hadamard powers'' \(A^{\circ n}=[a_{ij}^n]\) are positive semidefinite. With this in mind, a positive semidefinite matrix with nonnegative entries is called \textit{infinitely divisible} if for every real \(r\geq0\) the matrix \([a_{ij}^r]\) is positive semidefinite. These matrices have been studied by \textit{R. A. Horn} [Z. Wahrscheinlichkeitstheor. Verw. Geb. 8, 219--230 (1967; Zbl 0314.60017); Trans. Am. Math. Soc. 136, 269--286 (1969; Zbl 0177.05003)]. A continuous function \(m:\mathbb{R}^+\times\mathbb{R}^+\to\mathbb{R^+}\) is called a \textit{mean} if it is symmetric, increasing in both variables, \(m(\alpha a,\alpha b)=\alpha\,m(a,b)\) for every \(\alpha>0\), and \(\min(a,b)\leq m(a,b)\leq\max(a,b)\). Given a mean \(m\) and positive numbers \(\lambda_1<\lambda_2<\cdots<\lambda_n\), the authors consider matrices of the form \[ M=[m(\lambda_i,\lambda_j)], \quad W=\left[\frac1{m(\lambda_i,\lambda_j)}\right]. \] It is shown that at least one of the matrices \(M\) and \(W\) (and in some cases both) are infinitely divisible for several well-known means (among them: logarithmic mean, Heinz means, binomial means, Lehmer means, power difference means, Stolarsky means). The matrix \(W\) is shown to be \textbf{not} infinitely divisible when \(m\) happens to be one of the Heron means. The analysis of infinite divisibility is done in most cases by considering infinite divisibility of certain functions. Positive definiteness of functions has been used by the authors in the past to obtain inequalities for norms of operators. As a particular consequence of the techniques used in the paper, a refinement of an inequality of \textit{F. Hiai} and \textit{H. Kosaki} [J. Funct. Anal. 163, No. 2, 300--323 (1999; Zbl 0929.47009)] is obtained.