Monotonic properties of the least squares mean (Q641850): Difference between revisions

Let \(\mathbb{P=P}_{m}\) be the cone of \(m\times m\) positive definite Hermitian matrices with the Loewner ordering (\(A\leq B \iff B-A\) is semidefinite). The classical paper of \textit{F. Kubo} and \textit{T. Ando} [Math. Ann. 246, 205--224 (1980; Zbl 0412.47013)] has led to an extensive study of the geometric mean of pairs of elements of \(\mathbb{P}\), and it has been shown that many properties such as the arithmetic-geometric mean inequality hold in \(\mathbb{P}\). More recently several authors (see \textit{M. Moakher} [SIAM J. Matrix Anal. Appl. 26, No.~3, 735--747 (2005; Zbl 1079.47021)]; \textit{R. Bhatia} and \textit{J. Holbrook} [Linear Algebra Appl. 413, No.~2--3, 594--618 (2006; Zbl 1088.15022)]) have observed that a natural way to go from the two variable case to geometric means of \(n\) matrices is to use the trace metric on \(\mathbb{P}\). The trace metric \(\delta\) arises from the natural Riemann structure on \(\mathbb{P}\) and is defined by \(\delta(A,B):=\) \(\left( \sum_{i=1}^{m}\log^{2}\lambda_{i}\right) ^{1/2}\) where the sum is over the eigenvalues \(\lambda_{i}\) of \(A^{-1}B\). For example, the least squares mean \(\mathfrak{G}_{n}(A_{1},\dots ,A_{n})\) of \(n\) matrices in \(\mathbb{P}\) may be defined as the unique \(X\in\mathbb{P}\) for which \(\sum_{i=1}^{n}\delta ^{2}(X,A_{i})\) is minimum, and this is equal to the usual two variable geometric mean when \(n=2\) (alternative generalizations have also been proposed). Bhatia and Holbrook conjectured, but could not prove, that \(\mathfrak{G}_{n}\) is monotonic in the sense: \(\mathfrak{G}_{n}(A_{1} ,\dots, A_{n})\leq\mathfrak{G}_{n}(B_{1},\dots, B_{n})\) when \(A_{i}\leq B_{i}\) for each \(i\). The authors prove this conjecture for the more general case of weighted least squares means and show moreover that other desirable properties of extended geometric means hold for these means. They carry out their investigations in the general context of globally nonpositively curved metric spaces and use probabilistic methods.