Refined methods for the identifiability of tensors (Q478578)

The aim of the paper is to determine when a general tensor on \(\mathbb{C}^{a_1+1}\otimes \dots\otimes \mathbb{C}^{a_q+1}\) of rank \(k\) is identifiable, i.e., when it can be expressed in an unique way as the sum of \(k\) decomposable tensor (and no less than \(k\)). It is known that if \[ k> k_c = \frac {\prod_{i=1}^{q}(a_i+1)}{1+\sum_{i=1}^{q}a_i }, \] then tensors of rank \(k\) are never identifiable, while if \(k=k_c\), the only cases for which the decomposition is unique are known (\(q=3\), \(a_1=1\), \(a_2=a_3\)). Hence the problem is studied here for \(k<k_c\). In this hypothesis, the most ``unbalanced cases'', are described first, i.e. the cases when \(a_q\) is much bigger than the other \(a_i\)'s (suppose that the \(a_i\) are in increasing order), namely \[ a_q \geq \prod_{i=1}^{q-1}(a_i+1) - \sum_{i=1}^{q-1}a_i . \] For those unbalanced cases a list of the ones for which the decomposition is not unique is given (see Corollary 8.4). When \(a_q\) is not that big, the authors' conjecture is that the decomposition is always unique, with a few known exceptions; uniqueness is proved in several cases, as: \(a_1=\dots = a_q =1\), \(k\leq 0,9997k_c\); \(a_1=\dots = a_q =2\), \(k\leq 0,998k_c\); moreover, the conjecture is proved for \(\prod_{i=1}^{q}(a_i+1)\leq 100\). The induction technique used to obtain these results is a generalization of procedures due to Strassen in the case \(q=3\). An algorithm is given to study this problem, which is based on the concept of ``weakly defective'' secant varieties.