Bounded-rank tensors are defined in bounded degree (Q2443215): Difference between revisions

The notion of tensor rank has revealed to be of crucial importance in many areas (not only in multilinear algebra or algebraic geometry, but also for applications to algebraic statistics, phylogenetic and complexity theory), and it has been the center of many study in the last twenty years. One fundamental question is to find equations that characterize tensors of given degree; of course for matrices we have that matrices of rank at most \(k\) are determined by polynomials of degree \(k+1\) in their entries (their \((k+1)\times(k+1)\) minors); but there is nothing similar for tensor rank in general. What is considered in this paper is \textit{border rank}, an important notion which can be defined like this: consider the decomposable (rank 1) \((n_1\times \ldots \times n_p)\)-tensors; they are parameterized by the Segre variety \(S\) (the embedding of \(\mathbb{P}^{n_1}\times \ldots \mathbb{P}^{n_p}\) into \(\mathbb{P}^{N}\), \(N=\prod ^p_{i=1}n_i -1\)); if one considers the secant varieties \(\sigma_k(S)\), given by the closure of the union of all \((k-1)\)-spaces spanned by \(k\) points on \(S\), we have that all tensors of rank at most \(k\) are parameterized by points of \(\sigma_k(S)\) and those of rank \(k\) form an open dense set of \(\sigma_k(S)\). The tensors parameterized by points on \(\sigma_k(S)\) are said to have\textit{border rank} equal to \(k\). Hence all tensor of border rank \(k\) can be approximated with tensors of rank \(k\) (hence the name ``border rank''). The remarkable result proved in this paper is that there is a bound \(d=d(k)\) such that all tensors of border rank \(k\) are defined by the vanishing of polynomials of degree at most \(d\), whatever their dimension \(p\) and sizes \(n_i\) can be (i.e. the secant varieties \(\sigma_k(S)\) can be defined, at least set-theoretically. by polynomials of degre at most \(d(k)\)). The main idea to achieve this result is to consider ``infinite-dimensional'' tensors (via a limit on the tensor powers of a vector space) and to exploit their symmetries.