Decomposition of homogeneous polynomials with low rank (Q444154)

Let \(F\) be a homogeneous polynomial in many variables, of degree \(d\). The authors study the representations of \(F\) as a sum of powers of linear forms \(F=L_1^d+\dots + L_k^d\). The minimal \(k\) for which such decomposition holds is the \textit{symmetric rank} of \(F\). If \(F\) is identified with a symmetric tensor, the problem corresponds to find its decompositions in terms of symmetric tensors of rank one. From this point of view, the problem is relevant for applications to Signal Processing and Complexity Theory. One initial problem concerns the uniqueness of the decomposition. A typical case where uniqueness does not hold is when \(F\) can be approximated by polynomials of smaller rank \(k'< k\). The authors study in details this last situation, in the case where the ranks are small, i.e. \(k+k'\leq 2d+1\). With geometric methods, they prove that such polynomials \(F\) have an interesting decomposition as a sum \(Q+M_1+\dots +M_t\), where \(t\leq k+k'-d-2\), the \(M_i\)'s are uniquely determined and \(Q\) belongs to a subring generated by two linear forms, whose span is uniquely determined by \(F\). The authors also give a criterion for detecting when a specific polynomial of small rank \(k\leq d\) has a unique decomposition as a sum of \(k\) powers.