1: \begin{abstract}

2: We present an algorithm for decomposing a symmetric tensor, of dimension n

3: and order d as a sum of rank-1 symmetric tensors, extending the algorithm of

4: Sylvester devised in 1886 for binary forms. 

5: 

6: We recall the correspondence between the decomposition of a homogeneous

7: polynomial in n variables of total degree d as a sum of powers of linear

8: forms (Waring's problem), incidence properties on secant varieties of the

9: Veronese Variety and the representation of linear forms as a

10: linear combination of evaluations at distinct points. Then we reformulate

11: Sylvester's approach from the dual point of view.

12: 

13: Exploiting this duality, we propose necessary and sufficient conditions for

14: the existence of such a decomposition of a given rank, using the properties of

15: Hankel (and quasi-Hankel) matrices, derived from multivariate polynomials and

16: normal form computations. This leads to the resolution of polynomial

17: equations of small degree in non-generic cases.

18: 

19: We propose a new algorithm for symmetric tensor decomposition, based on

20: this characterization and on linear algebra computations with these Hankel

21: matrices. 

22: 

23: The impact of this contribution is two-fold. First it permits an efficient

24: computation of the decomposition of any tensor of sub-generic rank, as

25: opposed to widely used iterative algorithms with unproved global convergence

26: (e.g. Alternate Least Squares or gradient descents). Second, it gives tools

27: for understanding uniqueness conditions, and for detecting the rank.

28: \end{abstract}

29: