Local minimizers of the Crouzeix ratio: a nonsmooth optimization case study (Q2070327)

For an \(n \times n\) complex matrix \(A\), the field of values of \(A\) is defined as \(W(A) = \{ v^* A v : v \in\mathbf{C}^n, \| v \|_2 = 1 \}\), and if \(p\) is a polynomial, the Crouzeix ratio is defined to be \[ f(p,A) = \frac{\| p \|_{W(A)}}{\| p(A) \|_2}. \] Consider the optimization problem \[ \min \{ f(p,A): p \text{ is a polynomial and } A \text{ is a square matrix} \}. \] This is a nonconvex, nonsmooth optimization problem: at some pairs \((p, A)\), \(f\) is not differentiable. A famous conjecture by \textit{M. Crouzeix} [Integral Equations Oper. Theory 48, No. 4, 461--477 (2004; Zbl 1069.47004)] in matrix theory postulates that the globally minimal value of \(f\) is 0.5. Making use of a heavy-tailed distribution, the author proves for the first time that the Crouzeix ratio has many other locally minimal values between 0.5 and 1. The cases \(n = 2\), \(m = 3\) and \(n = 3\), \(m = 3\) are studied in detail (\(n\) is the size of the matrices and \(m\) is the degree of the polynomials). The extensive computations in the paper strongly support Crouzeix's conjecture. He presents results for nearly half a million optimization runs computing the Crouzeix ratio for about 250 million pairs \((p, A)\). The smallest locally minimal value is 0.5.