Noncommutative geometric means (Q816388)

This is an excellent article on the problems met in extending the notion of means to positive definite matrices. It should be read by all interested in means. A reasonable set of conditions that a mean \(M\) should satisfy are: (i) \(M(A_1, \dots,M_n)\) is invariant under any permutation of the matrices \(A_i, i\leq i\leq n\); (ii) \(M\) is increasing and continuous in each variable \(A_i, 1\leq i\leq n\); (iii) \(M(X^*A_1X, \dots,X^*A_nX)=X^*M(A_1, \ldots,A_n)X\), \(X\) an invertible matrix. This justifies the well-known but non-intuitive definition of the geometric mean \(G(A,B) = A^{1/2}(A^{-1/2}BA^{-1/2})^{1/2}A^{1/2}= \sqrt{AB}\) when the matrices commute. Several definitions of \(G(A,B,C)\) lead to an interesting discussion but only one definition appears to satify all of (i)--(iii). The definition of Sagae and Tanabe \[ G(A,B,C)=A^{1/2}\bigl(A^{-1/2}B^{1/2}(B^{-1/2}CB^{-1/2})^{1/2}B^{1/2}A^{-1/2} \bigr)^{2/3} A^{1/2} \] seems to be another possibility that is not discussed.