Kraus matrices and matrix convex functions (Q2151572)

Let \(f\) be a real-valued \(C^2\) function on an open interval \(I\) of the real line. Then \(f\) is called matrix convex of order \(n\) if \(f(tA+(1-t)B)\leq tf(A)+(1-t)f(B)\) for all \(n\times n\) Hermitian matrices \(A, B\) with \(\sigma(A), \sigma(B)\subset I\) and for all \(t\in[0, 1]\), where \(\sigma(A)\) is the set of all eigenvalues of \(A\) and \(A\leq B\) means \(B-A\) is positive semidefinite. The purpose of this article is to show the gap between orders \(n\) and \(n+1\) for matrix convex functions. More precisely, for any \(n\geq2\) and \(\epsilon>0\), the authors show that the function \(f(x)=x\sqrt{x}-\epsilon x^2\) defined on \((0,1)\) is matrix convex of order \(n\) if \(\epsilon\leq\frac{1}{2n+1}\) and is not matrix convex of order \(n+1\) if \(\epsilon>\frac{1}{2n+2}\). In particular, the function \(f(x)=x\sqrt{x}-\frac{1}{2n+1}x^2\) is matrix convex of order \(n\) but not of order \(n+1\).