The Euclidean distance degree of orthogonally invariant matrix varieties (Q1678646)

Let \(\mathcal{V}\) be a real algebraic variety of \(\mathbb{R}^{n}\) and consider the least square minimization problem of finding, for a given \(y\in \mathbb{R}^{n}\), all solutions of the problem \[ \min_{x\in \mathcal{V}}\;\sum_{i=1}^{n}(y_{i}-x_{i})^{2}. \] In the article [\textit{J. Draisma} et al., Found. Comput. Math. 16, No. 1, 99--149 (2016; Zbl 1370.51020)], the authors, viewing \(\mathcal{V}\) as a complex variety in \(\mathbb{C}^{n}\), they considered the complex critical points of the above function that lie on the regular part of \(\mathcal{V}\). They showed that the number of such points is constant for all \(y\) lying in an open dense subset of \(\mathbb{R}^{n}\). This number, denoted \(\mathrm{EDdegree}(\mathcal{V})\), was called the Euclidean distance degree of \(\mathcal{V}\) and measures the algebraic complexity of the above optimization problem. The effective computation of \(\mathrm{EDdegree}(\mathcal{V})\) is not an easy task though, since this number may change considerably under a linear transformation of \(\mathcal{V}\). In the current work, the authors study the Euclidean distance degree for orthogonally invariant matrix varieties \(\mathcal{M}\) in \(\mathbb{R}^{n\times m},\) that is, varieties invariant under the action of the orthogonal groups \(\mathcal{O}(n)\) and \(\mathcal{O}(m)\), in the sense that \(\mathcal{O}(n)\mathcal{MO}(m)=\mathcal{M}\). Membership in \(\mathcal{M}\) of a matrix \(M\) is fully determined by the singular values \(\sigma (M)=(\sigma _{1}(M),\ldots ,\sigma _{n}(M))\) of \(M\), and \(\mathcal{M}\) can be seen as the lift, from \(\mathbb{R}^{n}\) to \(\mathbb{R}^{n\times m}\) of an absolutely symmetric set \(S\subset \mathbb{R}^{n}\) via the inverse of the singular value map \(\sigma :\mathbb{R}^{n\times m}\rightarrow \mathbb{R}^{n}\). The main result of the paper is to establish the formula \[ \mathrm{EDdegree}(\mathcal{M})=\mathrm{EDdegree}(\mathcal{S}), \] which correponds to a transfer principle, according to the terminology of \textit{A. Daniilidis} et al. [J. Anal. Math. 128, 369--397 (2016; Zbl 1337.53010)]. This formula allows to compute effectively the Euclidean distance degree of several important instances of orthogonally invariant matrix varieties, such as the rank variety of matrices \(n\times m\) of rank at most \(r\), or the essential variety \(\mathcal{E}=\{E\in \mathbb{R}^{3\times 3}:\sigma _{1}(M)=\sigma _{2}(M),\;\sigma _{3}(M)=0\}\).