Unique decomposition for a polynomial of low rank

(1)

HAL Id: inria-00613049

https://hal.inria.fr/inria-00613049

Preprint submitted on 2 Aug 2011

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of

sci-entific research documents, whether they are

pub-lished or not. The documents may come from

teaching and research institutions in France or

abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est

destinée au dépôt et à la diffusion de documents

scientifiques de niveau recherche, publiés ou non,

émanant des établissements d’enseignement et de

recherche français ou étrangers, des laboratoires

publics ou privés.

Unique decomposition for a polynomial of low rank

Edoardo Ballico, Alessandra Bernardi

To cite this version:

Edoardo Ballico, Alessandra Bernardi. Unique decomposition for a polynomial of low rank. 2011.

�inria-00613049�

(2)

UNIQUE DECOMPOSITION FOR A POLYNOMIAL OF LOW RANK

EDOARDO BALLICO, ALESSANDRA BERNARDI

Abstract

Let F be a homogeneous polynomial of degree d in m + 1 variables defined over an algebraically closed field of characteristic 0 and suppose that F belongs to the s-th secant variety of the d-uple Veronese embedding of Pm

into P(m+dd )−1 _{but that its minimal decomposition as a sum of d-th}

powers of linear forms requires more than s addenda. We show that if s ≤ d then F can be uniquely written as F = Md

1+ · · · + Mtd+ Q, where M1, . . . , Mtare linear forms with t ≤ (d − 1)/2,

and Q a binary form such that Q =Pq

i=1l d−di

i mi with li’s linear forms and mi’s forms of degree

di such thatP(di+ 1) = s − t.

Introduction

In this paper we will always work with an algebraically closed field K of characteristic 0. Let Xm,d⊂ PN, with m ≥ 1, d ≥ 2 and N := m+d_m − 1, be the classical Veronese variety obtained as

the image of the d-uple Veronese embedding νd : Pm→ PN. The s-th secant variety σs(Xm,d) of

the Veronese variety Xm,d is the Zariski closure in PN of the union of all linear spans hP1, . . . , Psi

with P1, . . . , Ps ∈ Xm,d. For any point P ∈ PN, we indicate with sbr(P ) = s the minimum

integer s such that P ∈ σs(Xm,d). This integer is called the symmetric border rank of P . Since

Pm' P(K[x0, . . . , xm]1) ' P(V∗), with V an (m + 1)-dimensional vector space over K, the generic

element belonging to σs(Xm,d) is the projective class of a form (a symmetric tensor) of type:

(1) F = Ld₁+ · · · + Ld_r, (T = v⊗d₁ + · · · + v⊗d_r ).

The decomposition of a homogeneous polynomial that combines a minimum number of terms and that involves a minimum number of variables is a problem that is having a lot of attentions not only form classical Algebraic Geometry ([1], [7], [5], [6], [9]), but also from applications like Computational Complexity ([8]) and Signal Processing ([10]).

At the Workshop on Tensor Decompositions and Applications (September 13–17, 2010, Monop-oli, Bari, Italy), A. Bernardi presented the following result:

([2], Corollary 1) Let F ∈ K[x0, . . . , xm]dbe such that sbr(F )+sr(F ) ≤ 2d+1 and sbr(F ) < sr(F ).

Then there are an integer t ≥ 0, linear forms L1, L2, M1, . . . , Mt ∈ K[x0, . . . , xm]1, and a form

Q ∈ K[L1, L2]dsuch that F = Q+M1d+· · ·+Mtd, t ≤ sbr(F )+sr(F )−d−2, and sr(F ) = sr(Q)+t.

Moreover t, M1, . . . , Mtand the linear span of L1, L2 are uniquely determined by F .

In terms of tensors it can be translates as follows:

1991 Mathematics Subject Classification. 15A21, 15A69, 14N15.

Key words and phrases. Waring problem, Polynomial decomposition, Symmetric rank, Symmetric tensors, Veronese varieties, Secant varieties.

The authors were partially supported by CIRM of FBK Trento (Italy), Project Galaad of INRIA Sophia Antipolis M´editerran´ee (France), Marie Curie: Promoting science (FP7-PEOPLE-2009-IEF), MIUR and GNSAGA of INdAM (Italy).

(3)

2 EDOARDO BALLICO, ALESSANDRA BERNARDI

([2], Corollary 2) Let T ∈ SdV∗ be such that sbr(T ) + sr(T ) ≤ 2d + 1 and sbr(T ) < sr(T ). Then there are an integer t ≥ 0, vectors v1, v2, w1, . . . , wt ∈ S1V∗, and a symmetric tensor

v ∈ Sd(hv1, v2i) such that T = v +w⊗d1 +· · ·+w ⊗d

t , t ≤ sbr(T )+sr(T )−d−2, and sr(T ) = sr(v)+t.

Moreover t, w1, . . . , wtand hv1, v2i are uniquely determined by T .

The natural questions that arose at that workshop from applied people, were about the possi-ble uniqueness of the binary form Q in [2], Corollary 1 (ie. the vector v in [2], Corollary 2) and a bound on the number t of linear forms (ie. rank 1 symmetric tensors). We are finally able to give the most possible complete answer to this question. The main result of this paper is the following. Theorem 1. Let P ∈ PN _{with N =} m+d

d − 1. Suppose that:

sbr(P ) < sr(P ) and sbr(P ) + sr(P ) ≤ 2d + 1.

Let S ⊂ Xm,d be a 0-dimensional reduced subscheme that realizes the symmetric rank of P , and

let Z ⊂ Xm,d be a 0-dimensional non-reduced subscheme such that P ∈ hZi and deg Z ≤ sbr(P ).

There is a unique rational normal curve Cd⊂ Xm,d such that Cd∩ (S ∪ Z) ≥ d + 2. Then, for all

points P ∈ PN as above we have that:

S = S1t S2, Z = Z1t S2,

where S1= S ∩ Cd, Z1= Z ∩ Cd and S2= (S ∩ Z) \ S1.

Moreover Cd, S2 and Z are unique, deg(Z) = sbr(P ), deg(Z1) + deg(S1) = d + 2, Z1∩ S1 = ∅

and Z is a the unique zero-dimensional subscheme N of Xm,d such that deg(N ) ≤ sbr(P ) and

P ∈ hN i.

In the language of polynomials, Theorem 1 can be rephrased as follows.

Corollary 1. Let F ∈ K[x0, . . . , xm]d be such that sbr(F ) + sr(F ) ≤ 2d + 1 and sbr(F ) < sr(F ).

Then there are an integer 0 ≤ t ≤ (d − 1)/2, linear forms L1, L2, M1, . . . , Mt∈ K[x0, . . . , xm]1,

and a form Q ∈ K[L1, L2]d such that F = Q + M1d+ · · · + Mtd, t ≤ sbr(F ) + sr(F ) − d − 2, and

sr(F ) = sr(Q) + t.

Moreover the line hL1, L2i, the forms M1, . . . , Mt and Q such that Q = Pq_i=1ld−di imi with li’s

linear forms and mi’s forms of degree di such that P(di+ 1) = s − t, are uniquely determined by

F .

An analogous corollary can be stated for symmetric tensors.

Corollary 2. Let T ∈ Sd_V∗ _{be such that sbr(T ) + sr(T ) ≤ 2d + 1 and sbr(T ) < sr(T ). Then}

there are an integer 0 ≤ t ≤ (d − 1)/2, vectors v1, v2, w1, . . . , wt∈ S1V∗, and a symmetric tensor

v ∈ Sd_(hv

1, v2i) such that T = v +w⊗d1 +· · ·+w ⊗d

t , t ≤ sbr(T )+sr(T )−d−2, and sr(T ) = sr(v)+t.

Moreover the line hv1, v2i, the vectors v1, . . . , vtand the tensor v such that v =Pq_i=1u⊗(d−di ) i ⊗ zi

with ui ∈ hv1, v2i and zi ∈ Sdi(hv1, v2i) such thatP(di+ 1) = s − t, are uniquely determined by

T .

Moreover in Theorem 2 and in Corollary 4, by introducing the notion of linearly general po-sition of a scheme (Definition 1), we can also extend to a geometric description the condition for the uniqueness of the scheme Z of Theorem 1. We can rephrase their contents in therms of homogeneous polynomials and symmetric tensors in the following Corollary.

Corollary 3. Fix integers m ≥ 2 and d ≥ 4. Fix F an homogeneous polynomial in m + 1 variables of degree d (T ∈ Sd

(4)

any smoothable zero-dimensional scheme such that νd(Z) computes sbr(F ) (sbr(T )). Assume that

Z is in linearly general position. Then Z is the unique scheme computing sbr(P ) (sbr(F )). 1. Proofs

The existence of such a scheme Z was known from [3] and [4] (see Remark 1 of [2]). Lemma 1. Fix integers m ≥ 2 and d ≥ 2, a line ` ⊂ Pm

and any finite set E ⊂ Pm_{\ ` such that}

](E) ≤ d. Then dim(hνd(E)i) = ](E) − 1 and hνd(`)i ∩ hνd(E)i = ∅.

Proof. Since h0_{(` ∪ E, O}

`∪E(d)) = d + 1 + ](E), to get both statements it is sufficient to prove

h1_(I

`∪E(d)) = 0. Let H ⊂ Pm be a general hyperplane containing `. Since E is finite and H is

general, we have H ∩ E = ∅. Hence the residual exact sequence of the scheme ` ∪ E with respect to the hyperplane H is the following exact sequence on Pm_:

(2) 0 → IE(d − 1) → I`∪E(d) → I`,H(d) → 0

Since h1_(I

E(d − 1)) = h1(H, I`,H(d)) = 0, we get the lemma.

Proof of Theorem 1. All the statements are contained in [2], Theorem 1, except the unique-ness of Z, that deg(Z1) + deg(S1) = d + 2 and that Z1∩ S1 = ∅. Let ` ⊂ Pm be the line such

that νd(`) = Cd. Take Z, S, Zi, Si ⊂ Pm, i = 1, 2, such that νd(Z) = Z, νd(S) = S, νd(Zi) = Zi,

and νd(Si) = Si. Assume the existence of another subscheme Z0 ⊂ Xm,d such that P ∈ hνd(Z0)i

and deg(Z0) ≤ sbr(P ). Set Z10 := Z0 ∩ Cd. The proof of [2], Theorem 1, gives Z0 = Z10 t S2.

Since Cd is a smooth curve, Z1 ∪ Z10 ⊂ Cd, S2 ∩ ` = ∅, and Z ∪ Z0 = (Z1 ∪ Z10) t S2, the

schemes Z and Z0 are curvilinear. Hence all subschemes of Z and Z0 are smoothable. Hence any subscheme of either Z or Z0 _{may be used to compute the border rank of some point of P}N_.

Since deg(` ∩ (Z ∪ S)) ≥ d + 2, νd((Z ∪ S) ∩ `) spans hCdi. Lemma 1 implies hCdi ∩ hS2i = ∅.

Since P ∈ hS1∪ S2i and ](S) = sr(P ), we have P /∈ hAi for any A $ S. Therefore we get that

h{P } ∪ S2i ∩ hS1i is a unique point. Call P1 this point. Similarly, hZ1i ∩ hS2i is a unique point

and we call it P2. Since hCdi ∩ hS2i = ∅, the set hCdi ∩ h{P } ∪ S2i is at most one point. Since

Pi ∈ hCdi ∩ h{P } ∪ S2i, i = 1, 2, we have P1= P2 and {P1} = hCdi ∩ h{P } ∪ S2i. Since P1= P2

and P1 ∈ hS1i and P2 ∈ hZ10i, we have P1 ∈ hZ10i ∩ hS1i. Take any E ⊆ Z1 such that P1 ∈ hEi.

Since P ∈ h{P1} ∪ S2i ⊆ hE ∪ S2i and P /∈ hU i for any U ( Z, we get E ∪ S2= Z. Hence E = Z1.

Therefore Z1computes sbr(P1) with respect to Cd. Similarly, Z10 computes sr(P2) with respect to

the same rational normal curve Cd. Since P1= P2, we have Z10 = Z1(as for all curves we get the

uniqueness of Z1 for sbr(P1) ≤ b(d + 2)/2c). Since sbr(P1) 6= sr(P1), a theorem of Sylvester gives

sbr(P1) + sr(P1) = d + 2, i.e. deg(Z1) + deg(S1) = d + 2.

Definition 1. A scheme Z ⊂ Pm is said to be in linearly general position if for every linear subspace R $ Pmwe have deg(R ∩ Z) ≤ dim(R) + 1.

Notice that the next theorem is false if either d = 2 or m = 1. Theorem 2. Fix integers m ≥ 2 and d ≥ 4. Fix P ∈ PN

Let Z ⊂ Pm _{be any smoothable}

zero-dimensional scheme such that P ∈ hνd(Z)i and P /∈ hνd(Z0)i. Assume deg(Z) ≤ d and that Z is

in linearly general position. Then Z is the unique scheme Z0 ⊂ Pm _{such that deg(Z}0_{) ≤ d and}

P ∈ hνd(Z0)i. Moreover νd(Z) computes sbr(P ).

Proof. The existence of a scheme computing sbr(P ) follows from =[2], Remark ++= and the assumption “ sbr(P ) ≤ d ”. Fix any scheme Z0 ⊂ Pm _{such that Z}0 _{6= Z, deg(Z}0_{) ≤ d, P ∈}

hνd(Z0)i, and P /∈ hνd(Z00)i for any Z00( Z0. Assume Z0 6= Z. Since deg(Z ∪ Z0) ≤ 2d + 1 and h1

(Pm_{, I}

Z∪Z0(d)) > 0 ([2], Lemma 1), there is a line D ⊂ Pmsuch that deg(D ∩ (Z ∪ Z0)) ≥ d + 2.

(5)

4 EDOARDO BALLICO, ALESSANDRA BERNARDI

Hence deg(Z0) = d. Since deg(Z0) = d, we get Z0 ⊂ D. Hence P ∈ hνd(D)i. Hence sr(P ) = d.

As for all curves we get sbr(P ) ≤ b(d + 2)/2c. Since deg(Z0) = d, we assumed deg(Z0) ≤ sbr(P ), contradicting the assumption d ≥ 4. Corollary 4. Fix integers m ≥ 2 and d ≥ 4. Fix P ∈ PN

such that sbr(P ) ≤ d. Let Z ⊂ Pm _be

any smoothable zero-dimensional scheme such that νd(Z) computes sbr(P ). Assume that Z is in

linearly general position. Then Z is the unique scheme computing sbr(P ). References

[1] J. Alexander, A. Hirschowitz. Polynomial interpolation in several variables. J. Algebraic Geom. 4 (1995), no. 2, 201–222.

[2] E. Ballico, A. Bernardi, Decomposition of homogeneous polynomials with low rank. Math. Z. DOI 10.1007/s00209-011-0907-6.

[3] A. Bernardi, A. Gimigliano, M. Id`a. Computing symmetric rank for symmetric tensors. J. Symbolic. Comput. 46 (2011), 34–55.

[4] J. Buczy´nski, A. Ginensky, J. M. Landsberg. Determinantal equations for secant varieties and the Eisenbud-Koh-Stillman conjecture. arXiv:1007.0192 [math.AG].

[5] L. Chiantini, C. Ciliberto. Weakly defective varieties. Trans. Amer. Math. Soc. 454 (2002), no. 1, 151–178. [6] C. Ciliberto, M. Mella, F. Russo. Varieties with one apparent double point. J. Algebraic Geom. 13 (2004), no.

3, 475–512.

[7] A. Iarrobino, V. Kanev. Power sums, Gorenstein algebras, and determinantal loci. Lecture Notes in Mathe-matics, vol. 1721, Springer-Verlag, Berlin, 1999, Appendix C by Iarrobino and Steven L. Kleiman.

[8] L.-H. Lim, V. De Silva. Tensor rank and the ill-posedness of the best low-rank approximation problem. SIAM J. Matrix Anal. 30 (2008), no. 3, 1084–1127.

[9] J. M. Landsberg, Z. Teitler. On the ranks and border ranks of symmetric tensors. Found. Comput. Math. 10, (2010) no. 3, 339–366.

[10] R. C. Vaughan, T. D. Wooley. Waring’s problem: a survey, Number theory for the millennium. III (Urbana, IL, 2000), A K Peters, Natick, MA, (2002), 301–340.

Dept. of Mathematics, University of Trento, 38123 Povo (TN), Italy

GALAAD, INRIA M´editerran´ee, BP 93, 06902 Sophia Antipolis, France. E-mail address: ballico@science.unitn.it, alessandra.bernardi@inria.fr