Linear spectral sets and their extremal varieties (Q2361400)

The authors give a unifying approach to diverse distance problems. They define a vector space \({\mathcal V} \subseteq \mathbb{R}^k\) to be \textit{compatible} with the convex cone \({\mathcal C}_k := \{(a_1,a_2,\dots, a_k) : 0\leq a_1 \leq a_2 \leq \dots \leq a_k \} \subseteq \mathbb{R}^k\) if the orthogonal projection of \({\mathcal C}_k\) into \(\mathcal V\) is contained in \({\mathcal C}_k\). \({\mathcal G}_{\mathcal V} := \{ a_k\mathfrak{x}_k + a_{k-1}\mathfrak{x}_{k-1} + \dots + a_1\mathfrak{x}_1 \in \bigwedge^2 \mathbb{R}^n : \exists \mathfrak{y}_i, \mathfrak{z}_i \text{ such that } \mathfrak{x}_i = \mathfrak{y}_i \wedge \mathfrak{z}_i \text{ and } \{ \mathfrak{y}_1, \mathfrak{z}_1, \dots , \mathfrak{y}_k, \mathfrak{z}_k \} \text{ is orthonormal}, (a_1,a_2,\dots , a_k) \in {\mathcal V} \}\). The Grassmann variety \(G_2(\mathbb{R}^n)\) as well as the variety of \(G_2^\lambda(\mathbb{R}^n)\) of skew-symmetric matrices of rank at most \(2(\lambda-1)\) arise as such linear spectral sets \({\mathcal G}_{\mathcal V}\) for suitable chosen \({\mathcal V}\). Calculations of the least distance between a 2-tensor and \({\mathcal G}_{\mathcal V}\) in the projective space \(\mathbf{P}(\bigwedge^2 \mathbb{R}^n )\) whose underlying vector space is \(\bigwedge^2 \mathbb{R}^n \) are carried out. This extends results given for \(G_2(\mathbb{R}^n) \) in [\textit{C. Eckart} and \textit{G. Young}, Psychometrika 1, 211--218 (1936; JFM 62.1075.02); \textit{X. Jiang} et al., Math. Program. 127, No. 1 (B), 203--244 (2011; Zbl 1210.90142); \textit{J. Leventides} et al., Linear Algebra Appl. 461, 139--162 (2014; Zbl 1298.93119)], and for the extremal variety of \(G_2(\mathbb{R}^n)\) in [\textit{J. Leventides} et al., J. Optim. Theory Appl. 169, No. 1, 1--16 (2016; Zbl 1342.51010)]; the Eckart-Young theorem is given as a corollary. The set of points in \(\mathbf{P}(\bigwedge^2 \mathbb{R}^n )\) with maximum distance from \({\mathcal G}_{\mathcal V}\) with respect to the gap metric is called the \textit{extremal variety} of \({\mathcal G}_{\mathcal V}\) and denoted \(\mathrm{Extr}({\mathcal G}_{\mathcal V})\) by the authors. Among others, they prove \(\mathrm{Extr}(G_2^\lambda(\mathbb{R}^n)) = \mathrm{Extr}(G_2(\mathbb{R}^n))\) and \(\mathrm{Extr}(\mathrm{Extr}(G_2(\mathbb{R}^n))) = G_2\mathbb{R}^n \).