Spherical optimization with complex variables for computing US-eigenpairs (Q2374370)

Let \(\mathcal{S}\) be an order \(d\), \(n\)-dimensional complex symmetric tensor and \(x\) a vector on the unit sphere of \(\mathbb{C}^n\) such that for some \(\lambda\in\mathbb{R}\), \(\mathcal{S}^*x^{d-1}=\lambda x^*\) (the star is complex conjugate), then \((\lambda,x)\) is called a unitary symmetric (US) eigenvalue. Computing this US-pair can be formulated in several ways: as a spherical optimization problem of a real-valued function depending on complex variables or as a best rank-one approximation problem for the tensor. In this paper the optimization approach is followed. This requires an adaptation of the definition of complex derivatives (Wirtinger calculus) and an analysis of first- and second-order optimality conditions. This general theory is applied to the optimization of \(f(x)=|\mathcal{S}^* x^d| +\alpha\|x\|^{2d}\) for \(x\) on the unit sphere and \(\alpha>0\). Convexity is proved for values of \(\alpha\) large enough. It is proved that a simple fix-point iteration will converge to the eigenvalue and if there are finitely many US-eigenvectors, then also one of them will be found.