## Decomposition of quantics in sums of powers of linear forms (1996)

Venue: | Signal Processing |

Citations: | 68 - 18 self |

### BibTeX

@ARTICLE{Comon96decompositionof,

author = {P. Comon and B. Mourrain},

title = {Decomposition of quantics in sums of powers of linear forms},

journal = {Signal Processing},

year = {1996},

volume = {53},

pages = {93--107}

}

### Years of Citing Articles

### OpenURL

### Abstract

Symmetric tensors of order larger than two arise more and more often in signal and image processing and automatic control, because of the recent complementary use of High-Order Statistics (HOS). However, very few special purpose tools are at disposal for manipulating such objects in engineering problems. In this paper, the decomposition of a symmetric tensor into a sum of simpler ones is focused on, and links with the theory of homogeneous polynomials in several variables (i.e. quantics) are pointed out. This decomposition may be seen as a formal extension of the Eigen Value Decomposition (EVD), known for symmetric matrices. By reviewing the state of the art, quite surprising statements are emphasized, that explain why the problem is much more complicated in the tensor case than in the matrix case. Very few theoretical results can be applied in practice, even for cubics or quartics, because proofs are not constructive. Nevertheless in the binary case, we have more freedom to devise numerical algorithms. Keywords. Tensors, Polynomials, Diagonalization, EVD, High-Order Statistics, Cumulants. 1

### Citations

1569 |
Practical Optimization
- Gill, Murray, et al.
(Show Context)
Citation Context ...s a C-code that evaluates these matrices. This C-code is then linked with a general-purpose minimization algorithm (developed by J. Grimm, SAFIR Project). One could also use a more so sticated method =-=[11]-=-. We illustrate this method on the following polynomial: x1 4 + x1 3 x2 + x1 3 x3 + x1 2 x2 2 + x1 2 x2 x3 + x1 2 x3 2 + x1 x2 3 + x1 x2 2 x3 + x1 x2 x3 2 + x1 x3 3 + x2 4 + x2 3 x3 + x2 2 x3 2 + x2 x... |

1468 |
Independent component analysis, a new concept
- Comon
- 1994
(Show Context)
Citation Context ...osition to third order tensors is proposed. In [5], a canonical decomposition of tensors of T (n�2k) into n k powers is suggested, based on the assimilation of tensors to linear operators, whereas in =-=[6]-=- an approximate decomposition is described that always yields n powers. See also [7] for a discussion of the two latterapproaches. 2.4 Application in array processing Decompositions of quadrics in sum... |

390 |
The Classical Groups. Their Invariants and Representations
- Weyl
- 1946
(Show Context)
Citation Context ...rs as famous as Gauss, Kronecker, Noether, Cayley, Weyl, Hilbert, or Dieudonne, have contributed to this eld. At that time, a homogeneous polynomial of degree d in n variables was called a n-ary d-ic =-=[27]-=- [23]. For d =2� 3� 4� 5� 6:: the adopted terminology was the quadric, the cubic, the quartic, the quintic, the sextic... The same terminology will be subsequently retained. The goal of this paper is ... |

122 |
Algorithms: An Introduction to Computational Algebraic Geometry and Commutative Algebra, 3rd Edition
- Cox, Little, et al.
- 2007
(Show Context)
Citation Context ..., a polynomial vanishes on L2 (resp. L2) if and only if it vanishes by substitutions (a i 1 + a i 2) for i. These relations can be computed by elimination techniques (for instance using Grobner Bases =-=[8]-=-). This technique extends naturally to sums of any powers in the following way. The relations satis ed by the points in Lr are the polynomials ofIK [ i] \Ir. They give information only on the closure,... |

98 |
Plane Algebraic Curves
- Brieskorn, Knörrer
- 1986
(Show Context)
Citation Context ...nical expression as expected, as shows the small reconstruction error, err. Let's take now another example with p(x� y) =(2x + y) 5 +(;2x + y) 5 +(3x +3y) 5 . > p=convd([2,1],5)+convd([-2,1],5)+convd(=-=[3,3]-=-,5) p = 243 1375 2430 2510 1215 245 > [lambda,q]=binarydec2(p)� err = 7.52827e-13 > [lambda,q] = 1.0000 2.0000 1.0000 1.0000 -2.0000 1.0000 243.0000 1.0000 1.0000 The polynomial is reconstructed corre... |

58 | Super-symmetric decomposition of the fourth-order cumulant tensor. Blind identification of more sources than sensors
- Cardoso
- 1991
(Show Context)
Citation Context ...ges. Other approaches have been also proposed recently but are not discussed subsequently. In[14] for instance, an extension of the Singular Value Decomposition to third order tensors is proposed. In =-=[5]-=-, a canonical decomposition of tensors of T (n�2k) into n k powers is suggested, based on the assimilation of tensors to linear operators, whereas in [6] an approximate decomposition is described that... |

58 |
The invariant theory of binary forms
- Kung, Rota
- 1984
(Show Context)
Citation Context ...xpression, it is clear that because of the symmetry of G, some terms appear several times. Actually, there is another way of writing polynomials of F(n� d) by resorting to a standard compact notation =-=[12]-=- [10] [20], widely used in invariant theory. Let IN be the set of integers f0� 1� 2�::g, and J(n) the subset f1� 2�::ng. A multi-index of size n is a vector of n indices, i 2 IN n . By convention, if ... |

57 |
Sums of even powers of real linear forms
- Reznick
(Show Context)
Citation Context ... it is clear that because of the symmetry of G, some terms appear several times. Actually, there is another way of writing polynomials of F(n� d) by resorting to a standard compact notation [12] [10] =-=[20]-=-, widely used in invariant theory. Let IN be the set of integers f0� 1� 2�::g, and J(n) the subset f1� 2�::ng. A multi-index of size n is a vector of n indices, i 2 IN n . By convention, if a 2 IR n a... |

54 |
Lessons Introductory to the Modern Higher Algebra
- Salmon
- 1885
(Show Context)
Citation Context ... famous as Gauss, Kronecker, Noether, Cayley, Weyl, Hilbert, or Dieudonne, have contributed to this eld. At that time, a homogeneous polynomial of degree d in n variables was called a n-ary d-ic [27] =-=[23]-=-. For d =2� 3� 4� 5� 6:: the adopted terminology was the quadric, the cubic, the quartic, the quintic, the sextic... The same terminology will be subsequently retained. The goal of this paper is (i) t... |

38 |
Optimality of High Resolution array processing using the eigensystem Approach
- Bienvenu, Kopp
- 1983
(Show Context)
Citation Context ...l processing. The principle consists of approximating a symmetric matrix (the covariance of the observations) by another of lower rank, allowing to partition the space into signal and noise subspaces =-=[2]-=-. Now in order to apply the same principle to higher order tensors, one would like to approximate a tensor of F(n� d)by another of lower width, with our terminology. More precisely, the linear statist... |

35 |
Apolarity and canonical forms for homogeneous polynomials
- Ehrenborg, Rota
- 1993
(Show Context)
Citation Context ...sion, it is clear that because of the symmetry of G, some terms appear several times. Actually, there is another way of writing polynomials of F(n� d) by resorting to a standard compact notation [12] =-=[10]-=- [20], widely used in invariant theory. Let IN be the set of integers f0� 1� 2�::g, and J(n) the subset f1� 2�::ng. A multi-index of size n is a vector of n indices, i 2 IN n . By convention, if a 2 I... |

30 |
Invariant theory, old and new
- Dieudonne, Carrell
- 1970
(Show Context)
Citation Context ...iables. The methods used at that time were in some way, quite e cient [27] [23] [19]. Then, came the modern theory of algebraic geometry, which gave avery theoretical and general setting for this eld =-=[9]-=- [18]. Our discussion borrows results from both frameworks. 2.3 Statement of the problem It is known that symmetric matrices can be diagonalized by achange of coordinates, and that there are in nitely... |

27 |
Multichannel ARMA processes
- Swami, Giannakis
- 1994
(Show Context)
Citation Context ...nsors, but not matrices. Of course for simplicity, most algorithms taking advantage of HOS resort only to slices or contracted forms of those tensors, that can be stored in matrices (see for instance =-=[25]-=- and references therein). But it should be borne in mind that information is discarded when proceeding this way, and symmetry is broken. 2.2 Homogeneous polynomials As a rst obvious statement, it can ... |

26 |
Basic Algebraic Geometry
- Shafarevitch
- 1974
(Show Context)
Citation Context ... a nite number of points. This numberofpoints (by de nition) is the degree of the map . As it is de ned by D polynomials of degree d, a rough bound on the degree is d D (according to Bezout's theorem =-=[24]-=-). 3.4 Calculation of a decomposition To date, constructive algorithms for calculating a decomposition into powers of linear forms exist only for cubics and binary forms. 1 We discovered recently a wo... |

23 |
Tensor Methods
- McCullagh
- 1987
(Show Context)
Citation Context ... and cumulants are more than tensors, since they satisfy the multilinearity property even under non invertible transformations: matrix A de ning the transform may be rank de cient or rectangular. See =-=[15]-=- [17] for more details. As a consequence, when statistics of order larger than two (HOS) are utilized, the appropriate framework is no longer linear algebra anymore, but multilinear algebra, and the t... |

22 |
La méthode d’Horace éclatée: Application à l’interpolation en degré quatre
- Alexander, Hirschowitz
- 1992
(Show Context)
Citation Context ... algorithms for calculating a decomposition into powers of linear forms exist only for cubics and binary forms. 1 We discovered recently a work borrowing tools from another area of algebraic geometry =-=[1]-=-. Although it was addressing a di erent question (interpolation), the results presented allow to compute the generic width in a di erent way, especially for d>4. 104s3.4.1 Completion of the cube The r... |

19 |
Apolarity and canonical forms
- WAKEFORD
- 1918
(Show Context)
Citation Context ... linear forms if and only if there exist linear forms L1�:::�LN such that the rank of (x1L d;1 1 �:::�xnL d;1 1 , x1L d;1 2 �:::�xnL d;1 2 ,. . . , x1L d;1 N �:::�xnL d;1 N ) is equal to D(n� d). See =-=[26]-=-, [13], [21], [10] for more details. Another way to formulate this theorem is to say that there exist linear forms L1�:::�LN such that there is no polynomial of degree d orthogonal (for the scalar pro... |

11 |
Multipolynomial resultants and linear algebra
- MANOCHA, CANNY
- 1992
(Show Context)
Citation Context ...ry form in sums of powers. We have implemented the Matlab program listed in gure 5. We show now the results obtained with some examples. Take p(x� y) =(2x + y) 4 +(x +4y) 4 : > p=convd([2,1],4)+convd(=-=[1,4]-=-,4) p = 17 48 120 264 257 > [lambda,q]=binarydec2(p) err = 6.70011e-14 lambda = 1.0000 256.0000 q = 2.0000 1.0000 0.2500 1.0000 So we nd the original canonical expression as expected, as shows the sma... |

10 |
iIndependent component analysis - a new concept?,j
- Comon
- 1994
(Show Context)
Citation Context ...sition to third order tensors is proposed. In [5], a canonical decomposition of tensors of T (n; 2k) into n k powers is suggested, based on the assimilation of tensors to linear operators, whereas in =-=[6]-=- an approximate decomposition is described that always yields n powers. See also [7] for a discussion of the two latter approaches. 2.4 Application in array processing Decompositions of quadrics in su... |

8 |
A singular value decomposition for higher-order tensors
- Lathauwer, Moor, et al.
- 1993
(Show Context)
Citation Context ...arge is the di culty, and to give an idea of the state of the art, as accurately as is possible in a few pages. Other approaches have been also proposed recently but are not discussed subsequently. In=-=[14]-=- for instance, an extension of the Singular Value Decomposition to third order tensors is proposed. In [5], a canonical decomposition of tensors of T (n�2k) into n k powers is suggested, based on the ... |

8 | Second-order versus fourth-order MUSIC algorithms. an asymptotical statistical performance analysis
- MOULINES, CARDOSO
- 1991
(Show Context)
Citation Context ...As will be pointed out in section 3.2, there is in fact an upper bound to r (namely the width of the approximating tensor), depending on n and on the order d of the statistics to be used. Contrary to =-=[16]-=-, the array manifold is not required, to be able to detect and estimate the sources sj, aswellastoidentify the source vectors Aj. If one desires to perform localization, the array manifold can be used... |

5 |
The one-hundred anniversary of the death of invariant theory
- PARSHALL
- 1990
(Show Context)
Citation Context ...es. The methods used at that time were in some way, quite e cient [27] [23] [19]. Then, came the modern theory of algebraic geometry, which gave avery theoretical and general setting for this eld [9] =-=[18]-=-. Our discussion borrows results from both frameworks. 2.3 Statement of the problem It is known that symmetric matrices can be diagonalized by achange of coordinates, and that there are in nitely many... |

4 |
Eigenvalue decomposition of a cumulant tensor with applications
- COMON, CARDOSO
(Show Context)
Citation Context ...sors of T (n�2k) into n k powers is suggested, based on the assimilation of tensors to linear operators, whereas in [6] an approximate decomposition is described that always yields n powers. See also =-=[7]-=- for a discussion of the two latterapproaches. 2.4 Application in array processing Decompositions of quadrics in sums of powers are already used in antenna processing, and related areas in signal proc... |

3 |
Kanonische formen
- LASKER
- 1904
(Show Context)
Citation Context ...r forms if and only if there exist linear forms L1�:::�LN such that the rank of (x1L d;1 1 �:::�xnL d;1 1 , x1L d;1 2 �:::�xnL d;1 2 ,. . . , x1L d;1 N �:::�xnL d;1 N ) is equal to D(n� d). See [26], =-=[13]-=-, [21], [10] for more details. Another way to formulate this theorem is to say that there exist linear forms L1�:::�LN such that there is no polynomial of degree d orthogonal (for the scalar product d... |

2 |
Suns of powers of complex linear forms
- Reznick
- 1992
(Show Context)
Citation Context ...N). If nN were smaller than D then the image would lie in an hypersurface and would not be dense. But these bounds are clearly too loose to be really useful. 101sIt has been shown recently by Reznick =-=[21]-=-that 8p 2F(n� d)� !(p) n + d ; 2 d ; 1 � (9) which isamuch tighter bound. Moreover, this bound holds true when p ranges in the whole space F(n� d), and not only in a dense subset (i.e. the inequality ... |

1 | Foundation for Intelligent Physical Agents, “FIPA Communicative Act Library Specification,” http://www.fipa.org/repository/cas.html - SALMON - 1879 |

1 |
iMultipolynomial resultants and linear algebraj
- CANNY, MONACHA
- 1992
(Show Context)
Citation Context ...m in sums of powers. We have implemented the Matlab program listed in ��gure . We show now the results obtained with some examples. Take p(x; y) = (2 x + y) 4 + (x + 4 y) 4 : ? p=convd([2,1],4)+co=-=nvd([1,4],4) -=-p = 17 48 120 264 257 ? [lambda,q]=binarydec2(p) err = 6.70011e-14 lambda = 1.0000 256.0000 q = 2.0000 1.0000 0.2500 1.0000 So we ��nd the original canonical expression as expected, as shows the s... |

1 |
iInvariant theory, old and newj
- DIEUDONN, CARREL
- 1968
(Show Context)
Citation Context ...es. The methods used at that time were in some way, quite eOEcient [27] [23] [19]. Then, came the modern theory of algebraic geometry, which gave a very theoretical and general setting for this ��=-=eld [9] [18-=-]. Our discussion borrows results from both frameworks. 2.3 Statement of the problem It is known that symmetric matrices can be diagonalized by a change of coordinates, and that there are in��nite... |

1 |
iApolarity and canonical forms for homogeneous polynomialsj
- EHRENBORG, ROTA
- 1993
(Show Context)
Citation Context ...sion, it is clear that because of the symmetry of G, some terms appear several times. Actually, there is another way of writing polynomials of F(n; d) by resorting to a standard compact notation [12] =-=[10]-=- [20], widely used in invariant theory. Let IN be the set of integers f0; 1; 2; ::g, and J(n) the subset f1; 2; ::ng. A multi-index of size n is a vector of n indices, i 2 IN n . By convention, if a 2... |

1 |
iThe invariant theory of binary formsj
- KUNG, ROTA
- 1984
(Show Context)
Citation Context ...xpression, it is clear that because of the symmetry of G, some terms appear several times. Actually, there is another way of writing polynomials of F(n; d) by resorting to a standard compact notation =-=[12]-=- [10] [20], widely used in invariant theory. Let IN be the set of integers f0; 1; 2; ::g, and J(n) the subset f1; 2; ::ng. A multi-index of size n is a vector of n indices, i 2 IN n . By convention, i... |

1 |
iKanonische formenj, Mathematische Annalen
- LASKER
- 1904
(Show Context)
Citation Context ... LN such that the rank of (x 1 L d\Gamma1 1 ; : : : ; xn L d\Gamma1 1 , x 1 L d\Gamma1 2 ; : : : ; xn L d\Gamma1 2 ,. . . , x 1 L d\Gamma1 N ; : : : ; xn L d\Gamma1 N ) is equal to D(n; d). See [26], =-=[13]-=-, [21], [10] for more details. Another way to formulate this theorem is to say that there exist linear forms L 1 ; : : : ; LN such that there is no polynomial of degree d orthogonal (for the scalar pr... |

1 |
iSums of even powers of real linear formsj
- REZNICK
- 1992
(Show Context)
Citation Context ... it is clear that because of the symmetry of G, some terms appear several times. Actually, there is another way of writing polynomials of F(n; d) by resorting to a standard compact notation [12] [10] =-=[20]-=-, widely used in invariant theory. Let IN be the set of integers f0; 1; 2; ::g, and J(n) the subset f1; 2; ::ng. A multi-index of size n is a vector of n indices, i 2 IN n . By convention, if a 2 IR n... |

1 |
iSums of powers of complex linear formsj
- REZNICK
- 1992
(Show Context)
Citation Context ...s nN ). If nN were smaller than D then the image would lie in an hypersurface and would not be dense. But these bounds are clearly too loose to be really useful. It has been shown recently by Reznick =-=[21]-=- that 8p 2 F(n; d); !(p)si n + d \Gamma 2 d \Gamma 1 j ; (9) which is a much tighter bound. Moreover, this bound holds true when p ranges in the whole space F(n; d), and not only in a dense subset (i.... |

1 |
iMultichannel ARMA processesj
- SWAMI, GIANNAKIS, et al.
- 1994
(Show Context)
Citation Context ...ct tensors, not matrices. Of course for simplicity, most algorithms taking advantage of HOS resort only to slices or contracted forms of those tensors that can be stored in matrices (see for instance =-=[25] and-=- references therein). But it should be borne in mind that information is discarded when proceeding this way. 2.2 Homogeneous polynomials As a ��rst obvious statement, it can be pointed out that th... |

1 |
iApolarity and canonical formsj
- WAKEFORD
- 1918
(Show Context)
Citation Context ... : : ; LN such that the rank of (x 1 L d\Gamma1 1 ; : : : ; xn L d\Gamma1 1 , x 1 L d\Gamma1 2 ; : : : ; xn L d\Gamma1 2 ,. . . , x 1 L d\Gamma1 N ; : : : ; xn L d\Gamma1 N ) is equal to D(n; d). See =-=[26]-=-, [13], [21], [10] for more details. Another way to formulate this theorem is to say that there exist linear forms L 1 ; : : : ; LN such that there is no polynomial of degree d orthogonal (for the sca... |

1 | KN "ORRER, Plane Algebraic Curves, Birkha"user Verlag - BRIESKORN, H - 1986 |

1 | Kanonische formen", Mathematische Annalen - LASKER - 1904 |