Bounds on the number of real solutions to polynomial equations
aa r X i v : . [ m a t h . AG ] O c t BOUNDS ON THE NUMBER OF REAL SOLUTIONS TOPOLYNOMIAL EQUATIONS
DANIEL J. BATES, FR´ED´ERIC BIHAN, AND FRANK SOTTILE
Abstract.
We use Gale duality for complete intersections and adapt the proof of thefewnomial bound for positive solutions to obtain the bound e + 34 2( k ) n k for the number of non-zero real solutions to a system of n polynomials in n variableshaving n + k +1 monomials whose exponent vectors generate a subgroup of Z n of oddindex. This bound only exceeds the bound for positive solutions by the constant factor( e + 3) / ( e + 3) and it is asymptotically sharp for k fixed and n large. Introduction
In [3], the sharp bound of 2 n +1 was obtained for the number of non-zero real solutions toa system of n polynomial equations in n variables having n +2 monomials whose exponentsaffinely span the lattice Z n . In [4], the sharp bound of n +1 was given for the positivesolutions to such a system of equations. This last bound was generalized in [7], whichshowed that the number of positive solutions to a system of n polynomial equations in n variables having n + k +1 monomials was less than e + 34 2( k ) n k , which is asymptotically sharp for k fixed and n large [5]. This dramatically improvedKhovanskii’s fewnomial bound [8] of 2( n + k )( n + 1) n + k .We give a bound for all non-zero real solutions. Under the assumption that the exponentvectors W span a subgroup of Z n of odd index, we show that the number of non-degeneratenon-zero real solutions to a system of polynomials with support W is less than(1) e + 34 2( k ) n k . The novelty is that this bound exceeds the bound for solutions in the positive orthant bya fixed constant factor ( e + 3) / ( e + 3), rather than by a factor of 2 n , which is the numberof orthants. By the construction in [5], it is asymptotically sharp for k fixed and n large.We follow the outline of [7]—we use Gale duality for real complete intersections [6]and then bound the number of solutions to the dual system of master functions. The Mathematics Subject Classification.
Key words and phrases. sparse polynomial system, hyperplane arrangement, fewnomial.Bates and Sottile supported by the Institute for Mathematics and its Applications.Sottile supported by the NSF CAREER grant DMS-0538734. key idea is that including solutions in all chambers in a complement of an arrangementof hyperplanes in RP k , rather than in just one chamber as in [7], does not increase ourestimate on the number of solutions very much. This was discovered while implementinga numerical continuation algorithm for computing the positive solutions to a system ofpolynomials [1]. That algorithm was improved by this discovery to one which finds allreal solutions. It does so without computing complex solutions and is based on [7] and theresults of this paper. Its complexity depends on (1), and not on the number of complexsolutions.We state our main theorem in Section 1 and then use Gale duality to reduce it to astatement about systems of master functions, which we prove in Section 2.1. Gale duality for systems of sparse polynomials
Let W = { w = 0 , w , . . . , w n + k } ⊂ Z n be a collection of n + k +1 integer vectors ( |W| = n + k +1), which correspond to monomials in variables x , . . . , x n . A (Laurent) polynomial f with support W is a real linear combination of monomials with exponents from W ,(2) f ( x , . . . , x n ) = n + k X i =0 c i x w i with c i ∈ R . A system with support W is a system of polynomial equations(3) f ( x , . . . , x n ) = f ( x , . . . , x n ) = · · · = f n ( x , . . . , x n ) = 0 , where each polynomial f i has support W . Since multiplying every polynomial in (3) by amonomial x α does not change the set of non-zero solutions but translates W by the vector α , we see that it was no loss of generality to assume that 0 ∈ W .The system (3) has infinitely many solutions if W does not span R n . We say that W spans Z n mod Z -linear span of W is a subgroup of Z n of odd index. Theorem 1.
Suppose that W spans Z n mod and |W| = n + k +1 . Then there are fewerthan (1) non-degenerate non-zero real solutions to a sparse system (3) with support W . The importance of this bound for the number of real solutions is that it has a completelydifferent character than Kouchnirenko’s bound for the number of complex solutions.
Proposition 2 (Kouchnirenko [2]) . The number of non-degenerate solutions in ( C × ) n toa system (3) with support W is no more than n !vol(conv( W )) . Here, vol(conv( W )) is the Euclidean volume of the convex hull of W .Perturbing coefficients of the polynomials in (3) so that they define a complete intersec-tion in ( C × ) n can only increase the number of non-degenerate solutions. Thus it sufficesto prove Theorem 1 under this assumption. Such a complete intersection is equivalent toa complete intersection of master functions in a hyperplane complement [6].Let R n + k have coordinates z , . . . , z n + k . A polynomial (2) with support W is the pull-back Φ ∗W (Λ) of the degree 1 polynomial Λ := c + c z + · · · + c n + k z n + k along the mapΦ W : ( R × ) n ∋ x ( x w i | i = 1 , . . . , n + k ) ∈ R n + k . If we let Λ , . . . , Λ n be the degree 1 polynomials which pull back to the polynomials inthe system (3), then they cut out an affine subspace L of R n + k of dimension k . OUNDS ON THE NUMBER OF REAL SOLUTIONS TO POLYNOMIAL EQUATIONS 3
Let { p i | i = 1 , . . . , n + k } be degree 1 polynomials on R k which induce an isomorphismbetween R k and L ,Ψ p : R k ∋ y ( p ( y ) , . . . , p n + k ( y )) ∈ L ⊂ R n + k . Let
A ⊂ R k be the arrangement of hyperplanes defined by the vanishing of the p i ( y ). Thisis the pullback along Ψ p of the coordinate hyperplanes of R n + k .The image Φ W (( R × ) n ) inside of the torus ( R × ) n + k has equations z β = z β = · · · = z β k = 1 , where the weights { β , . . . , β k } form a basis for the Z -submodule of Z n + k of linear relationsamong the vectors W . To these data, we associate a system of master functions on thecomplement M A of the arrangement A of R k ,(4) p ( y ) β = p ( y ) β = · · · = p ( y ) β k = 1 . Here, if β = ( b , . . . , b n + k ) then p β := p ( y ) b · · · p n + k ( y ) b n + k .A basic result of [6] is that if W spans Z n modulo 2 and either of the systems (3)or (4) defines a complete intersection, then the other defines a complete intersection andthe maps Φ W and Ψ p induce isomorphisms between the two solution sets, as analyticsubschemes of ( R × ) n and M A . Since we assumed that the system (3) is general, thesehypotheses hold and the arrangement is essential in that the polynomials p i span thespace of all degree 1 polynomials on R k . Theorem 3.
A system (4) of master functions in the complement of an essential arrange-ment of n + k hyperplanes in R k has at most (1) non-degenerate real solutions. We actually prove a bound for a more general system than (4), namely for p ( z ) β = p ( z ) β = · · · = p ( z ) β k = 1 . We write this more general system as(5) | p ( z ) | β = | p ( z ) | β = · · · = | p ( z ) | β k = 1 . In a system of this form we may have real number weights β i ∈ R n + k . We give thestrongest form of our theorem. Theorem 4.
A system of the form (5) with real weights β i in the complement of anessential arrangement of n + k hyperplanes in R k has at most (1) non-degenerate realsolutions. Proof of Theorem 4
We follow [7] with minor, but important, modifications. Perturbing the polynomials p i ( y ) and the weights β j will not decrease the number of non-degenerate real solutions in M A . This enables us to make the following assumptions.The arrangement A + ⊂ RP k , where we add the hyperplane at infinity, is general inthat every j hyperplanes of A + meet in a ( k − j ) dimensional linear subspace, called a codimension j face of A . If B is the matrix whose columns are the weights β , . . . , β k ,then the entries of B are rational numbers and no minor of B vanishes. This last technical DANIEL J. BATES, FR´ED´ERIC BIHAN, AND FRANK SOTTILE condition as well as the freedom to further perturb the β j and the p i are necessary for theresults in [7, Section 3] upon which we rely.For functions f , . . . , f j on M A , let V ( f , . . . , f j ) be the subvariety they define. Supposethat β j = ( b ,j , . . . , b n + k,j ). For each j = 1 , . . . , k , define ψ j ( y ) := n + k X i =1 b i,j log | p i ( y ) | . Then (5) is equivalent to ψ ( y ) = · · · = ψ k ( y ) = 0. Inductively define Γ k , Γ k − , . . . , Γ byΓ j := Jac( ψ , . . . , ψ j , Γ j +1 , . . . , Γ k ) , the Jacobian determinant of ψ , . . . , ψ j , Γ j +1 , . . . , Γ k . Set C j := V ( ψ , . . . , ψ j − , Γ j +1 , . . . , Γ k ) , which is a curve in M A .Let ♭ ( C ) be the number of unbounded components of a curve C ⊂ M A . We have theestimate from [7], which is a consequence of the Khovanskii-Rolle Theorem,(6) | V ( ψ , . . . , ψ k ) | ≤ ♭ ( C k ) + · · · + ♭ ( C ) + | V (Γ , . . . , Γ k ) | . Here, | S | is the cardinalty of the set S . We estimate these quantities. Lemma 5. (1) | V (Γ , . . . , Γ k ) | ≤ k ) n k . (2) C j is a smooth curve and ♭ ( C j ) ≤
12 2( k − j ) n k − j (cid:0) n + k +1 j (cid:1) · j ≤
12 2( k ) n k · j − j ! . Proof of Theorem 4.
By (6) and Lemma 5, we have | V ( ψ , . . . , ψ k ) | ≤ k ) n k (cid:16) k X j =1 j j ! (cid:17) < k ) n k · e + 34 . (cid:3) Proof of Lemma 5.
The bound (1) is from Lemma 3.4 of [7]. Statements analogous to(2) for e C j , the restriction of C j to a single chamber (connected component) of M A , wereestablished in Lemma 3.4 and the proof of Lemma 3.5 in [7]:(7) ♭ ( e C j ) ≤
12 2( k − j ) n k − j (cid:0) n + k +1 j (cid:1) ≤
12 2( k ) n k · j − j ! . The bound we claim for ♭ ( C j ) has an extra factor of 2 j . A priori we would expect tomultiply this bound (7) by the number of chambers of M A to obtain a bound for ♭ ( C j ),but the correct factor is only 2 j .We work in RP k and use the extended hyperplane arrangement A + , as we will needpoints in the closure of C j in RP k . The first inequality in (7) for ♭ ( e C j ) arises as each OUNDS ON THE NUMBER OF REAL SOLUTIONS TO POLYNOMIAL EQUATIONS 5 unbounded component of e C j meets A + in two distinct points (this accounts for the factor ) which are points of codimension j faces where the polynomials F i ( y ) := Γ k − i ( y ) · (cid:16) n + k Y i =1 p i ( y ) (cid:17) i for i = 0 , . . . , k − j − F i is a polynomial of degree 2 i n .)The genericity of the weights and the linear polynomials p i ( y ) imply that these pointswill lie on faces of codimension j but not of higher codimension. The factor 2( k − j ) n k − j isthe B´ezout number of the system F = · · · = F k − j − on a given codimension j plane, andthere are exactly (cid:0) n + k +1 j (cid:1) codimension j faces of A + .At each of these points, C j will have one branch in each chamber of M A incident onthat point. Since the hyperplane arrangement A + is general there will be exactly 2 j suchchambers. (cid:3) References [1] D.J. Bates and F. Sottile,
Khovanskii-Rolle continuation for real solutions , 2007, in preparation.[2] D. Bernstein, A. Kouchnirenko, and A. Khovanskii,
Newton polytopes , Usp. Math. Nauk. (1976),no. 3, 201–202, (in Russian).[3] B. Bertrand, F. Bihan, and F. Sottile, Polynomial systems with few real zeroes , Math. Z. (2006),no. 2, 361–385.[4] F. Bihan,
Polynomial systems supported on circuits and dessins d’enfants , Journal of the LondonMathematical Society (2007), no. 1, 116–132.[5] F. Bihan, J.M. Rojas, and F. Sottile, Sharpness of fewnomial bounds and the number of components ofa fewnomial hypersurface , Algorithms in Algebraic Geometry (Alicia Dickenstein, Frank-Olaf Schreyer,and Andrew J. Sommese, eds.), IMA Volumes in Mathematics and its Applications, vol. 146, SpringerNew York, 2007, pp. 15–20.[6] F. Bihan and F. Sottile,
Gale duality for complete intersections , 2007, Annales de l’Institut Fourier,to appear.[7] ,
New fewnomial upper bounds from Gale dual polynomial systems , Moscow MathematicalJournal (2007), no. 3, 387–407.[8] A.G. Khovanskii, A class of systems of transcendental equations , Dokl. Akad. Nauk. SSSR (1980),no. 4, 804–807.
Institute for Mathematics and its Applications, University of Minnesota, 114 LindHall, 207 Church Street S.E., Minneapolis, MN 55455-0436, USA
E-mail address : [email protected] URL : Laboratoire de Math´ematiques, Universit´e de Savoie, 73376 Le Bourget-du-Lac Cedex,France
E-mail address : [email protected]
URL : Department of Mathematics, Texas A&M University, College Station, Texas 77843,USA
E-mail address : [email protected] URL ::