Euclidean distance degree in manifold optimization (Q6911542)

The paper deals with theoretical aspects of manifold optimization that can be of interest also in practical situation. In particular, the author focus on the so-called Euclidean Distance Degree (EDD) of some classes of manifolds (actually complex algebraic varieties), namely flag, Grassmann and Stiefel manifolds. The EDD has been defined in the seminal paper by \textit{J. Draisma} et al. [Found. Comput. Math. 16, No. 1, 99--149 (2016; Zbl 1370.51020)].\N\NThe EDD of a complex variety $X$ in $C^n$ measures the complexity of $X$ in the following sense: it corresponds to the number of (complex) critical points of the distance function to a generic element $p$ in $C^n$ restricted to $X$. In this paper $X$ is a variety in the space of complex matrices of size $m\times n$. In matrix optimization the number of (real) critical points of the Frobenius norm is an important generic invariant of a matrix manifold: the EDD is an upper bound on this number and in some situations can be computed in closed form.\N\NThe flag manifold is the variety of nested subspaces of some fixed increasing dimensions of some finite-dimensional vector space. Section 3 of the paper hosts the computation of its EDD, based on the so-called isospectral model of the flag manifold. This first gives a lower bound on the EDD, which is then proved to be sharp.\N\NThe Grassmannian is the variety of subspaces of fixed dimension of some linear space of finite dimension. Its EDD is computed in section 4 together with more general Schubert varieties. Moreover the critical points are characterized explicitly (Cor. 4.2). Also the best approximation problem on a Schubert variety is treated and solved (the projection is given explicitly).\N\NThe Stiefel manifold is the variety of orthonormal $k$-frames in some linear space of finite dimension. In section 5 its EDD is computed and proved to be independent on the number of rows of the matrices (but exponential in the number of columns).