Rank of a Hadamard product (Q2174506)

Given two \(n\times n\) matrices \(A, B\), the Hadamard product, \(A\circ B =[a_{ij}b_{ij}]\) of \(A\) and \(B\) behaves very differently from the usual matrix product \(AB\). For example, \(A\circ B = B\circ A\) but \(AB\not=BA\); if \(A\) and \(B\) are positive semidefinite, \(A\circ B\) is positive semidefinite, but \(AB\) is in general not (though \(AB\) is diagonalizable with nonnegative eigenvalues); \(A\circ B\) could be nonsingular even if both \(A, B\) are both singular when \(n\ge 3\), but \(AB\) is singular if \(A\) or \(B\) is singular. Let \(A\) and \(B\) be \(n\times n\) positive semidefinite matrices with no zero main diagonal entries. The authors show that \(A\circ B\) is positive definite if there is \(p\in \{1, \dots, n\}\) such that \(\operatorname{rank} B > n- p\) and every principal minor of \(A\) of size \(p\) is positive. This explains why \(A, B\) could be singular while their Hadamard product is not. Indeed, they prove that if \(A, B\) have no zero main diagonal entries and \[\max\{s_A+r_B, r_A+k_B\} >n, \] then \(A\circ B\) is positive definite, where \(r_A\) is the usual rank of \(A\) and \(k_A\) is the Kruskal rank of \(A\), i.e., \(k_A\) is the largest positive integer \(k\) such that any \(k\) distinct columns of \(A\) are linearly independent (convention: if \(A\) has a zero column, then \(k_A =0\)). They further show that the lower bound \(n\) is sharp by giving an example. The usual matrix exponential \(e^A\) for any square matrix \(A\) is nonsingular in view of the well known formula \(\det e^{A} = e^{\mathrm{tr}\, A}\). Now, the authors prove that the Hadamard exponential of a positive semidefinite matrix \(A\): \[ e^{\circ A} = [e^{a_{ij}}] = E+A+\frac 1{2!}A^{\circ 2} +\frac 1{3!}A^{\circ 3}+\cdots \] is positive definite if and only if no two columns of \(A\) are equal, where \(E\) is the \(n\times n\) all-ones matrix. This provides a new understanding of the singular behaviour of \(e^{\circ A}\).