Linearity of Cartan and Wasserstein means (Q6184914)

From MaRDI portal
scientific article; zbMATH DE number 7796455
Language Label Description Also known as
English
Linearity of Cartan and Wasserstein means
scientific article; zbMATH DE number 7796455

    Statements

    Linearity of Cartan and Wasserstein means (English)
    0 references
    0 references
    0 references
    0 references
    29 January 2024
    0 references
    The authors consider the cone of positive definite Hermitian matrices as a Cartan-Hadamard-Riemannian manifold endowed with the metrics: \[ \langle X,Y\rangle_A=\mathrm{tr} ( A^{-1}XA^{-1}Y) , \] \[ d(X,Y)=\left[\mathrm{tr}(X+Y)-2\mathrm{tr}(X^{\frac 12}YX^{\frac 12})^{\frac 12}\right]^{\frac 12}. \] For these metrics the means of matrices \(A\), \(B\) are, respectively, equal to \[ A\sharp_tB=A^{\frac 12}\left(A^{-\frac 12}BA^{-\frac 12}\right)^tA^{\frac 12}, \quad t\in[0,1], \] and \[ A\diamond_tB=(1-t)^2A+t^2B+t(1-t)\left[(AB)\frac 12+(BA)^{\frac 12}\right], \quad t\in[0,1]. \] In most cases the geodesic segment \(t\mapsto A\sharp_tB\) (resp., \(t\mapsto A\diamond_tB\)), is not linear, i.e., except of \(t=0,1\) it lies outside the linear span of \(A\), \(B\). It is a natural problem then to find out when the equations \[ A\sharp_tB=xA+yB, \qquad A\diamond_tB=xA+yB, \] have solutions \((x,y)\in\mathbb R^2\). The first main result claims that if equation \(A\diamond_tB=xA+yB\) has a solution, then \[ A\diamond_tB=(1-t)(1-at)A+t(1-b+bt)B \] for \(a\), \(b\) satisfying conditions \(a<1\), \(b<1\), \((1-a)(1-b)<1\). The second main result states that equation \(A\sharp_tB=xA+yB\) has a solution if and only if the matrix \(A^{-1}B\) has exactly two eigenvalues \(a\), \(b\), and in this case there holds \[ A\sharp_tB=\frac{ab^t-ba^t}{a-b}A+\frac{a^t-b^t}{a-b}B. \] One of the consequences of the above result is that \[ A\sharp B=xA+yB \quad\textrm{iff}\quad (xA+yB)^{-1}=yA^{-1}+xB^{-1}, \] with \(x,y>0,\; xy<1/4\). Moreover, the authors discuss the linearity problem for the case when matrices \(A\), \(B\) commute. It is solvable if and only if \(A^{-1}B\) has exactly two eigenvalues.
    0 references
    positive definite matrix
    0 references
    geometric and Wasserstein mean
    0 references
    Ostrowski-Taussky inequality
    0 references
    matrix with positive real part
    0 references
    Lyapunov equation
    0 references

    Identifiers

    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references