Gram determinants of real binary tensors (Q1638008)

Let \(T\) be real tensor of shape \(2\times \cdots \times2\) (\(n\) factors). Then \(T\) has \(n\) flattenings where the \(i\)-th flattening \(T_{i}\) is defined as a \(2\times2^{n-1}\) matrix with first row consisting of all \(2^{n-1}\) entries of \(T\) whose \(i\)-th index is \(1\) and whose second row has the entries in corresponding order with their \(i\)-th index \(2\). Set \(d_{i}:=\det(T_{i} T_{i}^{\ast})\) and consider the mapping \(\mathcal{G}:\mathbb{R}^{2} \times\dots \times\mathbb{R}^{2}\rightarrow\mathbb{R}^{n}\) given by \(T\longmapsto(d_{1},\dots,d_{n})\). Let \(\mathcal{B}\) be the unit ball consisting of all \(T\) with the sum of the squares of their entries at most \(1\). The author defines the set \(\mathcal{G(B)}\) to be the Gram locus of the tensors and answers some questions raised in [\textit{W. Hackbusch} and \textit{A. Uschmajew}, Numer. Math. 135, No. 3, 875--894 (2017; Zbl 1364.15018)] related to the geometry of this image. She shows that \(\mathcal{G(B)}\) is not convex, but its convex hull is given by the inequalities \(0\leq d_{i}\leq\frac{1}{4}\) and \(d_{i}\leq\sum_{j\neq i}d_{j}\) for all \(i\). Now define \[ Q_{1}:=\prod_{i=1}^{n}(\sum_{j\neq i}d_{j}-d_{i}), \qquad Q_{2}:=\frac {1}{2}\prod\left( \varepsilon_{1}\sqrt{d_{1}}+\dots+\varepsilon_{n}\sqrt{d_{n} }\right) , \] where the second product has \(2^{n}\) factors with \(\varepsilon_{i}=\pm1\). For \(n=3\), the author shows that \(\mathcal{G(B)}\) consists of the points inside \([0,\frac{1}{4}]^{3}\) satisfying \(Q_{1}\geq Q_{2}\) together with those satisfying \(Q_{1}<Q_{2}\) for which \((d_{i}-d_{j})^{2}+\frac{1}{2}(d_{i} +d_{j})\leq\frac{3}{16}\) for all \(i\neq j\). She conjectures that for \(n\geq4\) the points in \(\mathcal{G(B)}\) are just those in \([0,\frac{1}{4}]^{n}\) satisfying \(Q_{1}\geq Q_{2}\).