On the geometry of geometric rank (Q2166203): Difference between revisions

Several definitions of rank for tensors (= multidimensional matrices) are introduced in the literature, e.g. in order to determine the complexity of the process that the tensor encodes. For a \(3\)-tensor \(T\in \mathbb C^a\otimes \mathbb C^b \otimes \mathbb C^c\) the authors study the geometric rank, which is defined as follows. The tensor \(T\) determines matrices of type \(b\times c\) (the slices), which generate a subspace \(\Sigma^A\) in the space of matrices. The geometric rank \(GR(T)\) measures the intersection of \(\Sigma^A\) with spaces of matrices of low rank. Indeed, if \(\Sigma^A_j\) is the set of matrices of rank \(\leq \min\{b,c\}-j\) in \(\Sigma^A\), then \(GR(T)\) is defined as \(a+\min\{b,c\}-1-\max_j(\dim \Sigma^A_j+j)\). The geometric rank of \(T\), which gives an upper bound for the slice rank, also determines a lower bound for the border subrank, which is an important invariant in algebraic complexity theory, as it measures the minimal computational cost of the approximations of \(T\). Since the geometric rank is much easier to compute than the border subrank, then it provides a valuable information on the complexity of procedures, as e.g. in the case of matrix multiplication. In the paper under review, the authors use methods of multilinear algebra and the study of group actions on secant varieties to Segre varieties to classify tensors of geometric rank \(2\). In addition, the authors give partial results on the description of tensors of geometric rank \(3\), analyze the geometry of tensors with degenerate geometric rank, and give examples of the computation of the geometric rank of some specific tensors, mainly related with the tensor of matrix multiplication.