On linearized versions of matrix inequalities (Q6173912)

The authors prove linearized versions of the Aleksandrov-Fenchel and Brunn-Minkowski inequalities for positive semidefinite matrices. In order to present the results some definitions are needed. Given \(n\ge 1\) and arbitrary \(n\times n\) matrices \(A_1,\cdots,A_n\), denote by \(A_j^{(i)}\) the \(i\)-th column of the matrix \(A_j\). The mixed discriminant is defined as \[ \mathrm{D}(A_1,\cdots,A_n):=\frac{1}{n{!}}\sum_{\sigma \in S_n}\mathrm{det}\Big(A_{\sigma(1)}^{(1)},\cdots,A_{\sigma(n)}^{(n)}\Big), \] where \(S_n\) denotes the symmetric group of permutations of \(\{1,\cdots,n\}\). The Aleksandrov-Fenchel inequality says the following: Given \(A,B,C,A_2,\cdots,A_n\) real \(n\times n\) symmetric matrices, where \(A,A_2,\cdots,A_n\) are positive definite and \(C\) is positive semidefinite, then \(\mathrm{D}(C,A_2,\cdots,A_n)\ge 0\) and the equality holds if and only if \(C=0\). Further, there holds \[ \mathrm{D}(A,B,A_3,\cdots,A_n)^2\ge \mathrm{D}(A,A,A_3,\cdots,A_n)\mathrm{D}(B,B,A_3,\cdots,A_n), \] with equality if and only if \(B=\lambda A\) for \(\lambda \in \mathbb{R}\). The Brunn-Minkowski inequality is the following: Let \(A,B\) be positive definite \(n\times n\) matrices, then \[ \mathrm{det}((1-\lambda)A+\lambda B)^{1/n}\ge (1-\lambda)\mathrm{det}(A)^{1/n}+\lambda\mathrm{det}(B)^{1/n}, \] for any \(\lambda\in [0,1]\). A long introductory section includes known results concerning the mixed discriminant, followed by a section creating the necessary tools for proving the desired results. The main tool is the notion of projection of a matrix onto a linear subspace defined as follows. Let \(Q\) be a positive semidefinite \(n\times n\) matrix and let \(L\subset \mathbb{R}^n\) be a linear subspace of \(\mathbb{R}^n\) of dimension \(1\le k\le n\). Let \(q:\mathbb{R}^n \to \mathbb{R}\) be the quadratic form \(x \mapsto x^TQx\), then the projection of the matrix \(Q\) on \(L\) is the positive semidefinite \(k\times k\) matrix of the restriction of \(q\) to \(L\) and will be denoted by \(Q|L\). The main results of the paper are given by the following theorems. Theorem. Let \(C_3\cdots,C_n,C\) be symmetric positive semidefinite \(n\times n\) matrices, and let \(L=u^\perp\) be the \((n-1)\)-dimensional subspace orthogonal to the vector \(u\). If \(\mathrm{D}^{n-1}(C|L,C_3|L,\cdots,C_n|L)>0\) (\(\mathrm{D}^{n-1}\) is the mixed discriminant of \((n-1)\times (n-1)\) matrices), then \[ 2\mathrm{D}(A,C,C_3,\cdots,C_n)\ge \mathrm{D}(A,A,C_3,\cdots,C_n)+\mathrm{D}(C,C,C_3,\cdots,C_n), \] for every positive semidefinite \(n\times n\) matrix \(A\) with \(A|L=C|L\), the Aleksandrov-Fenchel inequality can be linearized. The next theorem deals with the linearized version of Brunn-Minkowski inequality. Theorem. Let \(u\) be a vector belonging to the unit sphere in \(\mathbb{R}^n\), and let \(L=u^\perp\) be the \((n-1)\)-dimensional subspace orthogonal to the vector \(u\). Let \(A, B\) be symmetric positive semidefinite \(n\times n\) matrices such that \(A|L=B|L\). Then \[ \mathrm{det}((1-\lambda)A+\lambda B)\ge (1-\lambda)\mathrm{det}(A)+\lambda \mathrm{det}(B), \] for every \(\lambda \in [0,1]\). Moreover, the equality holds if and only if there exists a symmetric positive semidefinite matrix \(R\) of rank at most \(1\), such that \(B=A+R\). These two results are only a selection of those contained in this interesting paper.