Solving bihomogeneous polynomial systems with a

zero-dimensional projection
Matı́as Bender∗ Laurent Busé† Carles Checa ‡
Elias Tsigaridas§
arXiv:2502.07048v1 [math.AC] 10 Feb 2025

February 12, 2025

Abstract
We study bihomogeneous systems defining, non-zero dimensional, biprojective vari-
eties for which the projection onto the first group of variables results in a finite set of
points. To compute (with) the 0-dimensional projection and the corresponding quotient
ring, we introduce linear maps that greatly extend the classical multiplication maps for
zero-dimensional systems, but are not those associated to the elimination ideal; we
also call them multiplication maps. We construct them using linear algebra on the
restriction of the ideal to a carefully chosen bidegree or, if available, from an arbitrary
Gröbner basis. The multiplication maps allow us to compute the elimination ideal of the
projection, by generalizing FGLM algorithm to bihomogenous, non-zero dimensional,
varieties. We also study their properties, like their minimal polynomials and the mul-
tiplicities of their eigenvalues, and show that we can use the eigenvalues to compute
numerical approximations of the zero-dimensional projection. Finally, we establish a
single exponential complexity bound for computing multiplication maps and Gröbner
bases, that we express in terms of the bidegrees of the generators of the corresponding
bihomogeneous ideal.

1 Introduction
Solving systems of polynomial equations is a common problem in (applied) mathematics
and engineering. Most available computational algebra tools focus on systems that have a
finite number of solutions, that is, zero-dimensional polynomial systems. We present new
mathematical and algorithmic tools to study (and to compute the solutions of) non-zero
dimensional bihomogeneous systems for which the projection onto the first group of variables
is zero dimensional. We introduce a novel technique to recover efficiently the points in the
zero dimensional projection.
∗
Inria & CMAP, CNRS, École Polytechnique, IP Paris, Palaisseau, France, [email protected]
†
Université Côte d’Azur, Inria, Sophia Antipolis, France. [email protected]
‡
University of Copenhagen, Copenhagen, Denmark. [email protected]
§
Inria Paris & IMJ-PRG, Sorbonne Université, Paris, France. [email protected]

1
 Let us introduce the setting of the problem and our techniques by considering different
algebraic formulations and solving strategies of a well-known problem from linear algebra:
the computation of eigenvalues of a matrix. Needless to say that we do not promote com-
puting in this way eigenvalues, but the problem serves as an (excellent) working example
of our formulation.

Computing eigenvalues Let k be an infinite field and A ∈ km×m a matrix. We can

consider the computation of the eigenvalues of A as an elimination problem. In particular,
if we consider the variety

V := {(x, y) ∈ P1 × Pm−1 : (x0 A − x1 Id) · y ⊤ = 0} ⊂ P1 × Pm−1 , (1)

then the eigenvalues correspond to computing the projection π(V ) ⊂ P1 . An obstacle of this
formulation is that even though the projection of π(V ) consists of a finite number of points,
the variety V can be much more complicated, e.g., it might contain irreducible components
of different (possibly high) dimensions. Therefore, it is not easy to transform the problem
into the task of solving a zero dimensional polynomial system, for which our algorithmic
arsenal is enormous, e.g., [1, 8, 22, 26, 27, 30, 34, 35]; the bibliography on polynomial
systems solving is huge and the cited references are just the tip of the iceberg.
Another formulation consists in computing the eigenvalues using the characteristic poly-
nomial of A. For this we consider the determinantal ideal associated to the equation
det(x0 A − x1 Id) = 0, which we solve to obtain the eigenvalues. However, this approach in-
troduces new obstacles. On the one hand, we have to compute symbolically, that is exactly,
the determinant of a matrix; this is a hard and unstable computation. On the other hand,
the solution set defined by the determinantal ideal might be more complicated than the
original one; indeed, this simple determinantal system corresponding to the characteristic
polynomial gives us the solutions, in our case the eigenvalues, with a higher multiplicity
(algebraic multiplicity) than their multiplicity in π(V ) (geometric multiplicity).
Fortunately, if we choose to compute the eigenvalues using the minimal polynomial of
A, many of the previous obstacles disappear. Indeed, the minimal polynomial has the
eigenvalues as roots with the ”correct” multiplicity and it generates the elimination ideal
of the eigenvalue problem, which corresponds to π(V ). However, even if we employ the
minimal polynomial to compute the eigenvalues, if we opt to compute it using elimination
theory, then we face the difficulty that we are forced to compute with the polynomials of
the original ideal at high enough degrees in the x’s and y’s; the degree of x’s being at
least the degree of the minimal polynomial of A. Hence, computing the elimination ideal,
in our case the minimal polynomial, in this way is suboptimal. To overcome this obstacle,
(multidimensional) linear recurrence relations can be used [28].
Our approach is inspired by the computation of the minimal polynomial. We use linear
recurrence relations to compute polynomials that encode the (zero-dimensional) projection
of non-zero-dimensional bihomogeneous polynomial systems; alas not always the elimination
ideal. Indeed, for the eigenvalue problem (1), our algorithm computes the elimination ideal
of π(V (I)) by computing the minimal polynomial of A using linear recurrence relations.

2
Problem statement Consider the bigraded ring R := k[x, y] where deg(xi ) = (1, 0)
and deg(yj ) = (0, 1). An ideal I ⊂ R is bihomogeneous if it is generated by homogeneous
bigraded elements in R. Bihomogeneous polynomial systems are the algebraic counter-part
of subschemes and subvarieties of Pn × Pm . We focus on bihomogeneous ideals that define
biprojective varieties whose projection to Pn is a finite set of points. Our objective is to
recover these points. As in the case of eigenvalues, where understanding the difference of
geometric and algebraic multiplicity involves extra computations, that is, to distinguish
eigenvectors from generalized eigenvectors, we only focus on recovering the points of the
projection and not their multiplicity. More formally, for a bihomogeneous ideal I ⊂ R, we
denote by V (I) ⊂ Pn × Pm its corresponding biprojective variety. Let π : Pn × Pm → Pn
denote the projection on the first set of variables. We only consider bihomogeneous ideals
where π(V (I)) is finite. Our goal is to recover the points in π(V (I)), which, in our case,
translates to

• Obtain a Gröbner basis for an (affine) ideal J ⊂ k[x1 , . . . , xn ] such that its variety agrees
with π(V (I)).

• Compute (approximate) the coordinates of the points in π(V (I)).

Context and contributions Bihomogeneous polynomial systems and their correspond-

ing varieties in Pn × Pm appear commonly in theory and practice when we work with
parametrized systems, e.g., [32], since elimination theory is simpler in this setting. From a
more theoretical point of view, we can consider them as the first generalization of projec-
tive varieties and as such they have received a significant amount of attention in the last
years [6, 11, 15, 24, 29]. The bihomogeneous varieties we consider correspond to parame-
terized polynomial systems, where there is only a finite number of admissible parameters.
This formulation contains as special cases most of the generalizations of the eigenvalue
problems [39, 40]. Even more, for a special class of such systems, we know explicit bounds
for the bitsize of the solutions [10].
A key tool for solving zero-dimensional polynomial systems are multiplication maps
(and their associated matrices), as they encode all the information from the quotient ring
by a given ideal; see [19, Ch. 2]. A typical approach for solving consists in computing a
Gröbner basis of the corresponding ideal (usually using a graded order), that in turn we
use to construct the multiplication maps. Then, the FGLM algorithm [25] employs these
maps to produce the Gröbner basis of the ideal with respect to a lexicographical order.
Moreover, from the eigenvalues and the eigenvectors of the multiplication maps we can
compute numerical approximations of the solutions [31].
There are many different (numerical) linear algebra-based algorithms to construct the
multiplication maps, e.g., [35, 36], which we could say that they lead to a paradigm known
as degroebnerization of system solving [16]. However, until now, to our knowledge, their
use was restricted to zero dimensional systems.
We introduce a novel approach to compute π(V (I)) by generalizing the use of multipli-
cation maps on the non-zero dimensional ideal I. In particular, we construct matrices that

3
we use to recover the points in π(V (I)) from their eigenvalues and to compute a Gröbner
basis of an ideal defining these points; for the latter, we employ (a generalization of) FGLM.
These matrices are bigger than the multiplication maps of the associated elimination ideal,
but we can compute them directly from the ideal I, without computing elimination ide-
als. In this way, we generalize the classical pipeline for solving polynomial systems using
Gröbner bases that forces us to use a particular order to compute an elimination ideal, and
instead, we can use any order to compute these multiplication matrices.
To construct our matrices, we consider the restriction of our ideal to a specific bidegree,
we call it admissible; see Def. 2.1. We can verify when a bidegree is admissible by performing
some rank computations on certain (Macaulay-like) matrices. Admissible bidegrees are
closely related to the Castelnuovo-Mumford regularity of certain modules [18] and also
to the multigraded Castelnuovo-Mumford regularity [6, 33]. We emphasize that admissible
degrees, in same cases, allow us to compute π(V (I)) by working with matrices corresponding
to bidegrees smaller than the Castelnuovo-Mumford regularity of the elimination ideal. For
example, in the case of the eigenvalue problem, even though (1, 1) is always an admissible
bidegree, the Castelnuovo-Mumford regularity of the elimination ideal is the degree of the
minimal polynomial of the matrix.
Geometrically, the ideal at an admissible degree induces a zero dimensional scheme that
contains π(V (I)). Even though we never miss points of π(V (I)), in our computations, they
might appear with higher (local) multiplicity and also additional points might be present.
However, if we consider a higher admissible degree, then we obtain a better ”approximation”
of π(V (I)). Eventually, by increasing the bidegree, we can precisely encode the points in
π(V (I)) and their appropriate multiplicities. Nevertheless, computing additional points is
not a problem for our approach because we also introduce a tool to verify if they belong to
π(V (I)).
We compute upper bounds for the admissible bidegrees. When the variety V (I) defines
a finite number of points, we show that the multihomogeneous Macaulay bound [7] is an
admissible degree. We were unable to prove (or disprove) that such bound holds in the
case of positive dimensional fibers. However, we deduce a (loose) bound using generalized
Koszul complexes (Theorem 4.2) for the general case with no further assumptions on the
geometry of V (I). Using these bounds, we deduce that the arithmetic complexity of our
algorithm for recovering the points in π(V (I)) is single exponential (Thm 4.5).
Finally, we relate the admissible bidegrees with the minimal elements appearing in the
Gröbner bases of the bihomogeneous ideals under study, building on regulatity and bigeneric
initial ideals [6].

Organization of the paper In Section 2, we introduce the admissible bidegrees and

study their properties. In Section 3, we present the construction of the multiplication maps
and their main properties; notably, we show how to use them to recover a Gröbner basis of
the elimination ideal (see Theorem 3.6). In Section 3.1, we study the eigenstructure of these
maps, generalizing known results for multiplication maps of zero-dimensional systems. In
Section 4, we present upper bounds for the admissible bidegrees and derive single exponen-

4
tial bounds to solve our problem. Moreover, we discuss how to decide when a candidate
point belongs to the projection. Finally, in Section 5, we compare the region of admissible
bidegrees with the the minimal generators of certain bigeneric initial ideals.

Acknowledgments MB and ET were partially funded by the project SOAP, a public

grant from the Fondation Mathématique Jacques Hadamard. CC has been supported by
the European Union under the Grant Agreement number 101044561, POSALG. Views and
opinions expressed are those of the authors only and do not necessarily reflect those of
the European Union or European Research Council (ERC). We thank Marc Chardin for
discussions on the bounds appearing in Section 4.

Notation Let k be a field and k̄ be its algebraic closure; and Pn the projective space
of dimension n ∈ N over k̄. Let R = k[x0 , . . . , xn , y0 , . . . , ym ] be the standard Z2 -graded
ring, with deg(xi ) = (1, 0), for 1 ≤ i ≤ n, and deg(yj ) = (0, 1), for 0 ≤ j ≤ m. Also
RX := k[x0 , . . . , xn ]. For a homogeneous h ∈ RX , the ring RX,h is the localization of RX
by he multiplicative set generated by h. Let mx := hx0 , . . . , xn i, my := hy0 , . . . , ym i be the
irrelevant ideals of Pn and Pm , respectively.
For a bihomogeneous polynomial f ∈ R, its bidegree is

deg(f ) = (degx (f ), degy (f )).

For a bihomogeheous ideal I ⊂ R, Ia,b denotes the vector space of homogeneous polynomials
of degree (a, b). Moreover, the Hilbert function of R/I at degree (a, b) ∈ Z2 is HFR/I (a, b) :=
dimk (R/I)a,b . We denote by HPR/I the Hilbert polynomial of R/I, i.e., a polynomial
such that HFR/I (a, b) = HPR/I (a, b) for a, b ≫ 0. We define the RX -module (R/I)∗,b :=
L
a (R/I)a,b . The local cohomology modules of I (or R/I) with respect to mx will be denoted
as Hmi x (I).
Finally, π(V (I)) is the set of points in the projection onto the first factor of V (I) ⊂
Pn × Pm . For each ξ ∈ Pn , pξ ⊂ R denotes the homogeneous ideal of polynomials that
vanish at this point.

2 Admissible bidegrees
To generalize multiplication matrices for zero-dimensional projections, we first present con-
ditions for their existence in terms of the Hilbert function of I.

Definition 2.1. Let I be a bihomogeneous ideal. A bidegree (a, b) ∈ Z2 is an admissible

bidegree for I, if I is generated in degrees lower or equal (component-wise) to (a, b) and if
there exists a polynomial h of bidegree (1, 0) such that

HFR/I (a, b) = HFR/I (a + 1, b) and HFR/(I,h) (a, b) = 0.

In this case, we call h an admissible linear form for (a, b).

5
Remark 2.2. Admissible bidegrees exist, if and only if, π(V (I)) is a zero-dimensional (or
empty) variety. This follows from the fact that the intersection with V (h) is empty, if and
only if, the projection of this variety to Pn is a finite set of points.
The following theorem shows that, given an admissible bidegree, the bidegrees with
bigger first coordinate and the same second coordinate are also admissible.
Theorem 2.3. Let I be a bihomogeneous ideal. If (a, b) is an admissible bidegree with an
admissible linear form h, then (a′ , b) is also an admissible bidegree, with the same admissible
linear form h, for every a′ ≥ a.
Before proving this theorem, we need a technical lemma that will be also used in the
sequel.
Lemma 2.4. Consider a bihomogeneous ideal I ⊂ R and a homogeneous polynomial g ∈ R
of degree (k, 0). Then, for any (a, b) ∈ Z2≥0 , any two of the following three conditions implies
the third:
i) I(a,b) = (I : g)(a,b) .

ii) HFR/(I,g) (a + k, b) = 0.

iii) HFR/I (a, b) = HFR/I (a + k, b).

Proof. Consider the following short exact sequence of vector spaces induced by multiplica-
tion by g
×g
0 → (R/(I : g))(a,b) −−→ (R/I)(a+k,b) → (R/(I, g))(a+k,b) → 0
As the alternate sum of the dimensions of the vector spaces appearing in this sequence
must be equal to zero, we deduce that if (I : g)(a,b) = I(a,b) then conditions ii) and iii) are
equivalent. To prove that ii) + iii) imply i), the previous exact sequence implies

HFR/(I:g) (a, b) − HFR/I (a, b) = HFR/I (a + k, b) − HFR/I (a, b) − HFR/(I,g) (a + k, b) = 0.

Hence, (I : g)(a,b) = I(a,b) since (I : g) ⊇ I.

Proof of Theorem 2.3. Let (a, b) be an admissible bidegree and h a corresponding admis-
sible linear form.
First, we analyze the syzygies of the ideal (I, h). Since by assumption HFR/(I,h) (a, b) =
0, we deduce that the graded RX -module M b := (R/(I, h))∗,b (i.e. the component of degree
b with respect to the y’s) satisfies (M b )a = 0, and so (M b )a′ = 0 for every a′ ≥ a. As
a consequence, the Castelnuovo-Mumford regularity of M b is < a. Let us call b2 (M b ) the
maximum degree shift that appears in a minimal finite free RX resolution of M b (i.e. b2 (M b )
is the maximum second Betti number of M b ).
By property of the Castelnuovo-Mumford regularity (see for instance [17, §1]), b2 (M b ) −
2 < a, hence b2 (M b ) ≤ a + 1. It turns out that b2 (M b ) is also an upper bound for the degree
at which syzygies of (I, h) in degree b with respect to the y’s are generated (this follows

6
from the definition of M b ). That is, any syzygy of (I, h) of degree (a′ , b) with a′ ≥ a + 1
can be RX -generated from syzygies of (I, h) of degree (a + 1, b).
Now, as explained in the proof [5, Lemma 1.9], which requires that I is generated in
degree ≤ (a, b), there is a correspondence between syzygies of (I, h) in degree (a′ + 1, b)
and elements in (I : h) of degree (a′ , b). Thus, we deduce that for every a′ ≥ a, (I : h)(a′ ,b)
is generated by (I : h)(a,b) . But by our assumption, (I : h)(a,b) = I(a,b) (apply Lemma 2.4
with k = 1), so it follows that (I : h)(a′ ,b) = I(a′ ,b) , which concludes the proof (apply again
Lemma 2.4 with k = a′ − a).

Remark 2.5. The previous proof relies on a deeper result: assuming the Hilbert polynomial
of the RX -module (R/I)∗,b is a constant, its Castelnuovo-Mumford regularity is a, if and
only if, (a+1, b) is an admissible bidegree of I and (a, b) is not. This follows mutatis
mutandis from [5, Lemma 1.8].
We conclude that I is saturated with respect to mx at the admissible bidegrees.
Corollary 2.6. If (a, b) is an admissible bidegree and h an admissible linear form, then

Ia,b = (I : m∞ ∞
x )a,b = (I : h )a,b .

Proof. For every k > 0 and a′ ≥ a, we have the inclusion Ia′ ,b ⊆ (I : mkx )a′ ,b ⊆ (I : hk )a′ ,b .
As (a′ , b) is admissible and h is an admissible form, by Lemma 2.4, (I : h)a′ ,b = Ia′ ,b . The
proof follows from the identity (I : h2 ) = ((I : h) : h).

To prove that generic linear forms lead to admissible forms, we need to characterize the
latter. For this, we introduce the colon ideal Jb given by the annihilator of the RX -module
(R/I)∗,b . We first show how we can characterize admissible linear forms using Jb and later
we explain the relation between the variety defined by Jb and π(V (I)).
Definition 2.7. Given b ≥ 0, we define the ideal Jb ⊂ RX as

Jb := (I : mby ) ∩ RX = AnnRX ((R/I)∗,b ).

Moreover, let πb (V (I)) := V (Jb ) ⊂ Pn be the vanishing set of Jb .

Lemma 2.8. Consider a linear form g ∈ RX and an admissible bidegree (a, b). We have
that (I : g)a,b = Ia,b , if and only if, g is a non-zero divisor of RX /(Jb : m∞
x ).

Proof. Note that, using Corollary 2.6, we know that I is saturated with respect to mx at
degree (a, b). In particular, Jb is saturated with respect to mx at degree a as q ∈ (Jb : m∞ x )a,0
β ∞ β
implies y q ∈ (I : mx )a,b = Ia,b for every monomial y of degree b and so q ∈ Jb . By [5,
Lemma 1.4], if g is not a zero divisor in R/(Jb : m∞ ∞ ∞
x ), then ((Jb : mx ) : g)a′ ,0 6= (Jb : mx )a′ ,0
for big enough a′ ≥ a. Hence, (I : g)a′ ,b 6= Ia′ ,b . As (a, b) is admissible, (I : g)a,b 6= Ia,b , by
contrapositive of Theorem 2.3.
Conversely, if (I : g)a,b = Ia,b , by Corollary 2.6, g is an admissible form. As colon ideals
commute, (Jb : g)a′ ,0 = (Jb )a′ ,0 , for every a′ ≥ a, by [5, Lemma 1.4], g is not a zero-divisor
in RX /(Jb : mx ∞ ).

7
 We obtain an effective criterion for admissibility of bidegrees.

Corollary 2.9. Assume that V (I) is not empty. Let (a, b) be a bidegree such that
HFR/I (a, b) = HFR/I (a + 1, b). Then, (a, b) is admissible , if and only if, HFR/(I,h′ ) (a, b) = 0
for every h′ in a open Zariski subset of R1,0 , that is, for a generic linear form h′ ∈ R1,0 .

Proof. As V (I) is not empty, (Jb : m∞ x ) 6= RX . The proof follows from the fact that,
as (Jb : m∞x ) =
6 R X and k is infinite, any generic form g ∈ RX is not a zero-divisor in
∞
RX /(Jb : mx ); see [5, pg. 4].

Example 2.10. Consider the ideal I ⊂ k[x0 , x1 , x2 , y0 , y1 , y2 ], where:

I = h2x0 − x1 − x2 , y0 y2 − y1 y2 − y22 , x1 y2 − x2 y2 , y02 − y12 − 2y1 y2 − y22 ,

x1 y0 − x1 y1 − x2 y2 , x21 − x1 x2 , x1 y12 − x2 y12 i.

Note that (2, 2) is already an admissible bidegree as

HFR/I (2, 2) = HFR/I (3, 2) HFR/(I,h) (2, 2) = 0 ,

for general h. The only point in π(V (I)) is ξ = [1 : 1 : 1].

Remark 2.11. The Castelnuovo-Mumford regularity of Jb and the admissible bidegrees of

I are not obviously related. Consider the eigenvalue problem from the introduction; the
degree (1, 1) is admissible, but the elimination ideal J1 is generated in the degree of the
minimal polynomial of the matrix, generally bigger than 1.

2.1 Geometry and admissible bidegrees

The classical way of constructing a geometric object from a module is to consider its anni-
hilator. For more details on this construction, see [21, Ch. 3]. In our setting, we construct
πb (V (I)) using an admissible bidegree which corresponds to a variety that contains π(V (I))
and, possibly, a finite set of extra points.
Notice that for b sufficiently big, Jb is exactly the elimination ideal of I. However, Jb
might not coincide with the elimination ideal (I : m∞ y ) ∩ RX for some specific b. As we
mentioned, there may be points in πb (V (I)) that are not in the projection. Algebraically,
this happens if there are associated primes of I that correspond to points in Pn but contain
the ideal my . Alternatively, the projective scheme defined by (I : m∞ y ) ∩ RX and Jb might
have the same points, but the multiplicities in the latter can be higher.

Lemma 2.12. Let I ⊂ R be a bihomogeneous ideal and let (a, b) ∈ Z2 an admissible

bidegree.

i) For every b′ ≥ b, we have:

π(V (I)) ⊆ πb′ (V (I)) ⊆ πb (V (I)).

8
 ii) The variety πb (V (I)) is zero-dimensional.

Proof. The inclusions in i) follow from the fact that

Jb ⊆ Jb′ ⊆ (I : m∞
y ) ∩ RX .

Similarly to what we used in the proof of Theorem 2.3, the Hilbert polynomial of (R/I)∗,b
is constant. Thus, its support consists of a finite number of points [3, Theorem 11.1]. As Jb
is the annihilator of (R/I)∗,b , using Hilbert polynomial, it is supported in a finite number
of points. By [38, Lemma 10.40.5], the vanishing locus of the annihilator of a finite module
coincides with its support. Thus, πb (V (I)) is finite.

Example 2.13 (Cont. Example 2.10). Note that both (2, 2) and (2, 3) are admissible bide-
grees. However, while we have π3 (V (I)) = π(V (I)), π2 (V (I)) = {[1 : 1 : 1], [1 : 0 : 2]}. To
understand why this happens, consider the following primary decomposition of I, which is
also minimal and unique,

I = hy0 − y1 − y2 , x1 − x2 , x0 − x2 i ∩ hy02 , y12 , y2 , x1 , 2x0 − x2 i ∩ hy2 , y0 + y1 , x2 , x1 , x0 i

The associated prime of the second primary component contains my and it cannot be an
associated prime of (I : m∞ 2
y ). However, this primary component does not contain my but
3
my , so it contributes with the point [1 : 0 : 2] to π2 (V (I)), but not to π3 (V (I)). Thus, it
will not appear in πb (V (I)) for b ≫ 0 (in this case, b = 3).

3 Multiplication maps
In this section, we describe the generalization of multiplication maps to the setting of zero-
dimensional projections. Throughout this section, we fix an admissible bidegree (a, b) and
a set B of elements of degree (a, b) which form a basis for the vector space (R/I)a,b . For
each g ∈ R(k,0) , we define the map

m̄g : (R/I)a,b → (R/I)a+k,b [f ] 7→ m̄g ([f ]) := [gf ].

We can characterize when a multiplication map is invertible.

Lemma 3.1. Let g be a polynomial of degree (k, 0). The map m̄g is invertible, if and only
if, (I, g)a+k,b = Ra+k,b . In particular, if h is an admissible form, the map m̄hk is invertible
for any k ≥ 0.

Proof. The first statement is trivial. For the second one, observe that as (a, b) is admissible,
HFR/I (a + k, b) = HFR/I (a, b) and, by Corollary 2.6, Ia,b = (I : hk )a,b . The proof follows
from Lemma 2.4.

Multiplication maps commute in the following way.

9
Lemma 3.2. Let h be an admissible linear form for I and consider homogeneous f, g ∈ RX
such that deg(f ) = s and deg(g) = t. Then, we have the following identities,

m̄−1
hs+t
◦ m̄f g = m̄−1 −1 −1 −1
hs ◦ m̄f ◦ m̄ht ◦ m̄g = m̄ht ◦ m̄g ◦ m̄hs ◦ m̄f

Proof. We will prove the first equality of the lemma, the other one follows by swapping f
and g. First, by Lemma 3.1, the map m̄hk is invertible for any k ≥ 0. For every q ∈ Ra,b and
homogeneous f ∈ RX , we have that f q ≡ hk (m̄−1 hk
◦ m̄f )(q) mod Ia+k,b , where k = deg f .
Hence, for every q ∈ Ia,b ,

hs+t (m̄−1
hs+t
◦ m̄g f )(q) ≡ g f q ≡ g hs (m̄−1 t s −1 −1
hs ◦ m̄f (q)) ≡ h h (m̄ht ◦ m̄g ◦ m̄hs ◦ m̄f )(q) mod I.

By Lemma 2.4,(I : hs+t )a,b = Ia,b . Hence, we have that, for any q ∈ Ra,b , (m̄−1
hs+t
◦ m̄g f )(q) ≡
−1 −1
(m̄ht ◦ m̄g ◦ m̄hs ◦ m̄f )(q) mod Ia,b , and so the first equality follows.

In what follows, fix an admissible linear form h and let (RX,h )0 be the zero-degree part
of the localization of RX by the multiplicative set generated by h.
g
Definition 3.3. Consider hk
∈ (RX,h )0 , where k = deg g. We define the multiplication
map by hgk as
m g := m̄−1
hk
◦ m̄g .
hk

Let M g be the matrix representing this map in the basis B.

hk

Remark 3.4. By Lemma 3.2, the map mg′ does not depend on the rational function that
we choose as the representative of g′ . Note that the matrix of the multiplication map
can be constructed either using the Schur complement of the Macaulay matrix at degree
(a + k, b) [9] or using a precomputed Gröbner basis with respect to an arbitrary order.
For the eigenvalue problem that we considered in the introduction, the computation of the
multiplication map by xx01 in the appropriate monomial basis recovers the matrix A.

Lemma 3.5. The map (2) is a commutative ring homomorphism,

→ k[mg′ : g ′ ∈ (RX,h )0 ] g ′ 7→ mg′

(RX,h )0 − (2)

Proof. The map is a ring homomorphism as, for any f ′ , g′ ∈ (RX,h )0 , we have that m1 = Id
and mg′ f ′ = mg′ ◦ mf ′ = mf ′ ◦ mg′ (Lemma 3.2) and mg′ +f ′ = mg′ + mf ′ . To prove this
last point, let f ′ = hfs and g ′ = hgt , for homogenoeus f, g ∈ RX . Assume with no loss of
generality that s ≥ t. Hence,

mf ′ + mg′ = mf ′ + m ht−s ◦ mg′ = mf ′ + m ht−s +g′

ht−s ht

= m̄−1 −1
hs ◦ m̄f + m̄hs ◦ m̄hs−t g
= m̄−1
hs ◦ (m̄f + m̄hs−t g ) = mf ′ +g ′

where the second to last equality holds by linearity of m̄−1

hs .

10
 Our next theorem relates the commutative ring formed by the multiplication maps and
a localization of the ideal Jb .
Theorem 3.6. The kernel of the ring homomorphism defined in Equation (2) corresponds
to the zero-degree part of the localization of the ideal Jb (Theorem 2.7) in RX,h , i.e., the
ideal (Jb ⊗RX RX,h )0 ⊆ (RX,h )0 .
Proof. Let f ′ ∈ (RX,h )0 . We have that mf ′ = 0 if and only if m̄f = 0, where f ∈ RX is any
polynomial such that f ′ = hfk . Equivalently, if and only if f q ∈ Ia+k,b , for any q ∈ Ra,b .
If m̄f = 0, we have that f ha y β ∈ Ia+k,b , for any y β ∈ R0,b , so f ha ∈ (I : mby )a+k,0 .
Hence, hfk ∈ (Jb ⊗RX RX,h )0 .
If f ∈ (Jb ⊗RX RX,h )0 , f hω y β ∈ Iω+k,b for big enough ω ≥ 0 and arbitrary y β ∈ R0,b . So,
for any xα ∈ Ra,0 , we have that f hω xα y β ∈ Ia+k+ω,b . As (a, b) is admissible, by Corollary
2.6, (Ia+k,b : hω ) = Ia+k,b . Hence, f xα y β ∈ Ia+k,b for any xα y β ∈ Ra,b , and so m̄f = 0.

If h is an admissible form, it cannot vanish at any point of πb (V (I)), as RX /Jb is

zero-dimensional and h is not a zero-divisor in the saturation; see Lemma 2.8. Hence,
the projective scheme of Jb ⊆ RX is isomorphic to the affine one of (Jb ⊗RX RX,h )0 ⊆
(RX,h )0 . Therefore, to compute πb (V (I)), we will use the localization of Jb at h. As a
consequence of Theorem 3.6, we can recover a Gröbner bases for the ideal (Jb ⊗RX RX,h )0
by finding polynomial relations between the multiplication matrices as in FGLM [25]. For
this, we assume with no loss of generality that x0 is an admissible linear form in I. When
the characteristic of the field is big enough, this is guaranteed after a generic change of
coordinates on the x variables.
Corollary 3.7. Assume h = x0 is an admissible linear form. From the multiplication
matrices M x1 , . . . , M xn , the algorithm FGLM [25] recovers any Gröbner basis of (Jb ⊗RX
x0 x0
RX,x0 )0 ⊂ k[ xx01 , . . . , xxn0 ] by computing linear relations between the multiplication maps.
Example 3.8 (Cont. Example 2.13). At the admissible bidegree (2, 2), the monomials
{x20 y02 , x20 y0 y1 , x20 y0 y2 , x20 y12 } form a basis for (R/I)2,2 the multiplication matrices by z1 := xx10
and z2 := xx20 are:
   
1 0 0 0 1 0 0 0
0 1 0 0 0 1 0 0
Mz1 =
0
 Mz2 = 
0 1 0 0 0 1 0
0 1 1 0 0 −1 −1 2

A lexicographical Gröbner basis of the ideal (J2 ⊗RX RX,x0 )0 is

hz1 + z2 − 2, z22 − 3z2 + 2i.

At degree (2, 4), the corresponding multiplication matrices for z1 and z2 are of size 5 × 5
but the Gröbner basis is hz1 + z2 − 2, z2 − 1i. This is coherent with the stabilization of
π4 (V (I)) = π(V (I)).

11
3.1 Eigenvalues of multiplication maps
On the top of computing a Gröbner bases for our ideal, we can also recover information from
πb (V (I)) by computing the eigenvalues of the multiplication maps. We start this section
by characterizing the eigenvalues of these maps and then we study their multiplicities.
Theorem 3.9. Let I ⊂ R be a bihomogeneous ideal and let (a, b) ∈ Z2 an admissible
bidegree. For every g ′ ∈ (RX,h )0 , there are µξ > 0, for ξ ∈ πb (V (I)), such that the
characteristic polynomial of mg′ is
Y  µξ
CharPolmg′ (λ) = λ − g′ (ξ) . (3)
ξ∈πb (V (I))

Proof. Observe that λ is an eigenvalue, if and only, if 0 = mg′ (f ′ ) − λf ′ = mg′ −λ (f ′ ), for

some f ′ ∈ (RX,h )0 . If we write g′ = hgk , the previous holds if and only if m̄g−λhk has a
non-trivial kernel. By Lemma 3.1, this is equivalent to the fact that
(I, g − λhk )a+k,b 6= Ra+k,b . (4)
Using the same argument as in Lemma 2.8, which corresponds to the case where k = 1, (4)
is equivalent to the fact that g − λhk is a zero-divisor on RX /(Jb : m∞x ), that is, g − λh
k
k
vanishes at ξ ∈ πb (V (I)). This means that g(ξ) − λh (ξ) = 0, so λ is as claimed.

Example 3.10 (Cont. Example 3.8). Consider the admissible bidegree (3, 2) and the ad-
missible form h = x0 + x1 . If we construct the multiplication map with respect to xh1 at this
degree, its eigenvalues are 0 and 1, which correspond to the two minimal primes that do
not contain mx . However, the same construction at admissible bidegree (2, 4), leads to only
one eigenvalue equal to 1, which corresponds to the minimal prime that does not contain
either mx or my .
Once we have identified the eigenvalues of mg , we would like to know the multiplicity
of each of the zeros of the characteristic polynomial, i.e. find the values of the exponents
µξ in (3). In the zero-dimensional case, these values are exactly the algebraic multiplicities
of each of the points ξ ∈ V (I), i.e. length of the modules (R/I)pξ . In our case, we will
find an analogue of such multiplicities by considering the length of the localization of the
RX -module (R/I)∗,b at each of the primes pξ for ξ ∈ πb (V (I)).
Definition 3.11. Let I ⊂ R be a bihomogeneous ideal and let (a, b) ∈ Z2 an admissible
bidegree. Consider the RX -module Mb := (R/I)∗,b . Then, for every ξ ∈ πb (V (I)), we
consider:
Mb,ξ := Mb ⊗RX (RX )pξ ,
where (RX )pξ is the localization of RX at the prime ideal pξ .
Lemma 3.12. Let I be a bihomogeneous ideal and let (a, b) ∈ Z2 an admissible bidegree.
Then, the natural map defines an isomorphism:
∼
=
Y
(Mb )a −
→ Mb,ξ (5)
ξ∈πb (V (I))

12
Proof. Consider the exact sequence of local cohomology of Mb (see [21, Theorem A4.1]):

M

0−→ Hm0 x Mb − → Mb −→ fb (ν)) −
Γ(Pn , M → Hm1 x Mb −→0 (6)
v∈Z

where M fb is the coherent sheaf associated to the module Mb and L n f

ν∈Z Γ(P , Mb (ν)) are
its global sections.
As Mb is supported in a finite number of points, the global sections of M fb (ν) satisfy
n f ∼ Q
Γ(P , Mb (ν)) = ξ∈πb (V (I)) Mb,ξ for any ν ∈ Z. Thus, taking the graded pieces of degree a
in (6), we have:
Y
0−→ Hm0 x (R/I)a,b −
→ (R/I)a,b −
→ → Hm1 x (R/I)a,b −
Mb,ξ − →0 (7)
ξ∈πb (V (I))

where the central arrow is the natural map. This follows from the fact that Hmi x (Mb )a ∼ =
Hmi x (R/I)a,b for every i ≥ 0, which follows by definition of local cohomology from Cech
complexes; see for instance [18, Lemma 3.7]. Thus, it is enough to show that Hm0 x (R/I)a,b =
Hm1 x (R/I)a,b = 0.

Firstly, we have Hm0 x (R/I)a,b = (I : m∞x )/I a,b = 0 by Corollary 2.6. Moreover, as we
have HFR/I (a, b) = HFR/I (a + k, b) for all k > 0, the Hilbert function of Mb at degree a
coincides with its Hilbert polynomial. Thus, we can use Grothendieck-Serre formula [12,
§4.4] to deduce that Hm1 x ((R/I)∗,b )a = 0.

As the prime ideal pξ define points in RX , the localizations (RX )pξ are local rings with
residue field k. Therefore, the modules Mb,ξ have a structure of k-vector space whose di-
mension is equal to their length. Our next theorem characterizes the algebraic multiplicities
of the multiplication matrices.

Theorem 3.13 (Cont. of Theorem 3.9). The exponents of characteristic polynomial of

mg′ in Equation (3) are µξ = length(Mb,ξ ), for ξ ∈ πb (V (I)).

Proof. The proof follows the same argument as in [19, §2, Prop. 2.7]. Multiplication by g′
and the isomorphism in Lemma 3.12 induce a commutative diagram:
∼
= Q
(Mb )a Mb,ξ
ξ∈πb (V (I))
×g ′
×g ′
∼
= Q
(Mb )a Mb,ξ
ξ∈πb (V (I))

Thus, the multiplication map mg′ has a block structure, where each block has dimension
equal to the length of Mb,ξ and corresponds to the map:
×g ′
Mb,ξ −−→ Mb,ξ .

13
Note that pξ is the only associated prime in Mb,ξ . So, if g ′ = hgk , then, using the same
argument as in Theorem 3.9, g − λhk is a nonzero divisor in Mb,ξ for some λ ∈ k, if and
only if, g − λhk ∈ pξ . Therefore, λ = hgk (ξ) which is the only possible eigenvalue in this
block.

4 Complexity bounds
We provide bounds for the admissible bidegrees (a, b). These lead to an upper bound for
the complexity of constructing multiplication maps and computing Gröbner bases in our
setting.
Theorem 4.1. (Macaulay bound for finitely many points) Given an ideal I = hf1 , . . . , fl i ⊂
R with deg(fi ) = (ai , bi ) and such that V (I) ⊂PPn × Pm defines a finite number of points.
Then, every bidegree (a, b) such that (a, b) ≥ li=1 (ai , bi ) − (n, m), (a, b) is an admissible
bidegree for I.
Proof. The proof uses standard techniques, based on the comparison of two spectral se-
quences associated to the Koszul-Cech bicomplexes of I and of (I, h), where h ∈ R1,0 ; see
[15, Prop. 3.15] for the details. The first step is to deduce from these spectral sequences
that all the local cohomology Pmodules of both R/I and R/(I, h) vanish at bidegrees which
are higher but not equal than li=1 (ai , bi ) + (1, 0) − (n + 1, m + 1). From there, applying the
Grothendieck-Serre formula [12, § 4.4], we deduce that for all such bidegree (a, b) the Hilbert
functions of both R/I and R/(I, h) are equal to their respective Hilbert polynomials, hence
are both constants, which implies that (a, b) is admissible.

The approach proposed in this paper, together with the previous bound, can be further
optimized to compute Gröbner bases for zero-dimensional ideals in product of projective
spaces; see [7].
Our proof only works when V (I) is zero-dimensional. We performed several experiments
in more general contexts, that is, when V (I) is non-zero-dimensional, and in all of them the
Macaulay bound was an admissible bidegree. Unfortunately, we were not able yet to prove
that the bound holds for arbitrary systems, neither to find a counter-example. Therefore,
in what follows, we present a drastically more pessimistic bound to obtain an admissible
bidegree.
In the next theorem, given ℓ ∈ N, let Nℓ := HFk[y0,...,ym ] (ℓ), that is, the number of
monomials of degree ℓ in the y’s.
Theorem 4.2. Given an ideal I = hf1 , . . . , fl i ⊂ R with deg(fi ) = (ai , bi ) and
Pl such that
n
π(V (I)) ⊂ P defines a finite number of points. Then, if (a, b) is such that b ≥ ( i=1 bi )−m
P
and a ≥ ( li=1 ai Nb−bi ) − n, it is an admissible bidegree for I.
Proof. For any b ∈ N, we define the graded RX -module Mb := (R/I)∗,b . Since h is of
bidegree (1, 0), it is independent on the y’s and we have
Mb /hMb ≃ (R/(I, h))(∗,b) .

14
In particular, applying Theorem 4.1 we get that P the Hilbert function of the RX -module
Mb /hMb in degree a is equal to zero if (a, b) ≥ li=1 (ai , bi ) − (n, m). This implies that for
all such b the Hilbert polynomial of Mb is a constant, which means that Mb is geometrically
supported on finitely many points of Pn . Now, it remains to estimate a degree (in the x’s)
at which we are sure that this Hilbert function stabilizes at a constant value.
For that purpose, we define the map
f1 ,...,f
φb : E := ⊕li=1 RX (−ai )Nb−bi −−−−→
l
F := RX Nb ,
which is a finite free presentation of the graded RX -module Mb . We denote by e, respectively
f , the rank of the free module E, respectively F and we consider the generalized Koszul
complex associated to φb which is of the following form (see [37, Ap. C.2]):
φ
0 → Se−f −1 (F ∗ ) ⊗ ∧e E → · · · → S0 (F ∗ ) ⊗ ∧f +1 E → E −→
b
F,
where S(−) and ∧(−) stand for the symmetric and exterior algebras, respectively. An
important property of this complex is that its homology is annihilated by the 0th Fitting
ideal of φb , i.e. the ideal generated by the maximal minors (of size f ) of φb . As a consequence,
the homology of the complex is supported on the support of Mb which are finitely many
points. Now, to estimate the Hilbert function of Mb , which is the cokernel of φb , we proceed
with the comparison of the two spectral sequences associated to bicomplex obtained from
the above generalized Koszul complex by replacing each term by its Cech complex (this is
similar to what we did in the proof of Theorem 4.1, the Koszul complex being replaced by a
generalized Koszul complex). It follows that the local cohomology modules of Mb (actually,
only the local cohomology modules Hm0 x (Mb ) and Hm1 x (Mb ) are nonzero) vanish at all degree
ν such that (∧e E)ν+n = 0, i.e.,
 Xl 
ν≥ ai Nb−bi − n.
i=1

Therefore, the Hilbert function of Mb at those degree ν is a constant, which concludes the
proof (use Theorem 4.1 for the bound for the stabilization of the Hilbert function of R/(I, h)
where h is an admissible linear form).

Example 4.3 (Cont. Example 3.8). In order to illustrate how much room for improvement
the Theorem 4.2 leaves, in this example (2, 2) is already an admissible bidegree, but the
upper bound is (83, 6).
P
Proposition 4.4. If b ≥ ( li=1 bi ) − m, then πb (V (I)) = π(V (I)) as a finite set of points.
Proof. This is a classical result. As already observed in the proof of Lemma 2.12, we have
inclusions
Jb = AnnRX ((R/I)∗,b ) ⊂ Jb+1 = AnnRX ((R/I)∗,b+1 )
for all b. It is well known (and easy to check) that this increasing sequence of annihilators
stabilizes as soon as Hm0 y (R/I)(∗,b) = 0, and that this latter equality holds for all b ≥
P
( li=1 bi ) − m (see for instance [14, Prop. 3.18 and Theorem 3.29]).

15
 As a consequence, we deduce a single exponential bound for the number of arithmetic
operations for computing a Gröbner basis of Jb . We remark that, if the characteristic of the
field is big enough, then x0 is always an admissible form after performing a generic linear
change of coordinates.
Theorem 4.5. Let I = hf1 , . . . , fl i ⊂ R be a bihomogeneous ideal such that π(V (I)) is
finite and x0 is an admissible linear form. The number of arithmetic operations needed to
recover a Gröbner basis for (Jb ⊗RX RX,x0 )0 ⊂ k[ xx10 , . . . , xxn0 ], which describes the variety
π(V (I)), is upper bounded by
O(l4(n+m+nm) A4n B 4(m+mn) ),
where (A, B) = (maxj degx fj , maxj degy fj ).
Proof. Let (a, b) be the admissible bidegree from Theorem 4.2. Then, the number of
n+m+nm An B m+mn
rows/columns of the multiplication map Mxi /x0 is at most S := l n! (m!)n+1 . Hence,
4
we can construct each matrix in O(S ) arithmetic operations. We recover the Gröbner basis
using Corollary 3.7 in O(nS 4 ) arithmetic operations; the complexity bound follows from
the classical analysis of FGLM [25], taking into account that we work with linear relations
of matrices and not vectors.
Better complexity estimations and faster algorithms can be obtained by using random-
ized approaches in FGLM; see e.g. [23].

How to tell apart π(V (I)) from πb (V (I))

The new methods proposed might compute the points in π(V (I)) and some extra ones.
In this section, we study how to tell them apart. For this, given ξ ∈ Pn , we consider the
homogenoeus ideal Iξ ⊂ RY := k̄[y0 , . . . , ym ] obtained by specializing the variables x’s to ξ.
Proposition 4.6 ([14, Corollary 3.30]). Given an ideal I = hf1 , . . . , fl i ⊂ R with deg(fi ) =
(ai , bi ). Given ξ ∈ Pn , we have that ξ ∈
/ π(V (I)), if and only if, (Iξ )b = (RY )b for b =
Pl
i=1 bi − m.
To apply the Proposition 4.6, we need to take special considerations according to the
computation paradigm in which we work.

Symbolic paradigm. Using Corollary 3.7, we can compute a Gröbner bases bases
and from it a rational univariate representation of πb (V (I)) [20] given by polynomials
h0 (t), . . . , hn (t), r(t) ∈ k[t], with r(t) square-free, defining the superset of π(V (I)),
{[h0 (ξ) : · · · : hn (ξ)] ∈ Pn : ξ ∈ k̄, r(t) = 0}. Let r(t) = r1 (t) · · · rk (t) be the irreducible
factorization of r(t) over k[t]. For each i, consider
Wi := {[h0 (ξ) : h1 (ξ) : · · · : hn (ξ)] ∈ Pn : ξ ∈ k̄ and ri (t) = 0}.
Consider the homomorphism φ : R → Qi := k[y0 , . . . , ym ][t]/ri (t), sending each xi 7→
φ(xi ) := hi (t) ∈ Qi and yj 7→ yj . Prop. 4.6 implies that Wi ⊂ π(V (I)) if and only if we
have the equality of (k[t]/ri (t))-vector spaces (φ(I))b = (Qi )b .

16
Numerical paradigm. If ξ¯ is an approximation of ξ ∈ Pn and we want to verify if
ξ ∈ π(V (I)), we need to check numerically if (Iξ̄ )b = (RY )b . This can be done by computing
the numerical rank of the Macaulay matrix associated to (Iξ̄ ) at degree b. This kind of check
has been employed in the context of Computer Aided Design [13] and similar considerations
apply to our context.

5 Admissible bidegrees and Gröbner bases

In the previous sections, we have claimed that an arbitrary Gröbner basis of I leads us to
admissible degrees and thus to the construction of our multiplication maps. Indeed, we can
do this by exploiting the Gröbner basis to verify if a bidegree is admissible. For example,
we can use the shape of the Gröbner basis to deduce the behavior of the Hilbert function
[21, Theorem 15.3] or to construct smaller Macaulay matrices that allow us to perform rank
computations efficiently [4].
However, there is a particular degree reverse lexicographical monomial order from which
it is easier to deduce the admissible degrees. This order satisfies the following inequalities,

y n > · · · > y 0 > xn > · · · > x0 . (8)

If we work in generic coordinates, the variable x0 is an admissible form. Hence, checking

the saturation of the ideal with respect to x0 or updating the Gröbner bases of the ideal
when we add x0 can be done with no additional computational cost; see [5, Section 2].
We had recently shown that, in generic coordinates, the structure of the Gröbner basis of
a bihomogeneous ideal with respect to this order depends on some cohomological properties
of the ideal. In what follows, we will relate them to the admissible degrees, to show how the
admissible degrees coincide with certain degrees of the minimal elements in the Gröbner
bases with respect to this order in generic coordinates, that is, with minimal generators
of the bigeneric initial ideal [2]. This illustrates the connection between the result of this
paper and the computation of Gröbner bases.

Definition 5.1. Let I ⊂ S be a bihomogeneous ideal. The partial regularity region, xreg(I),
is a region of bidegrees (a, b) ∈ Z2 , such that for all i ≥ 1 and (a′ , b′ ) ≥ (a − i + 1, b):

Hmi x (I)(a′ ,b′ ) = 0.

The region xreg(I) relates to the admissible degrees as follow.

Theorem 5.2. Let I ⊂ R be a bihomogeneous ideal such that π(V (I)) is finite. Then, for
(a, b) ∈ N2 , the following are equivalent:

i) (a, b′ ) is an admissible bidegree for all b′ ≥ b.

ii) (a, b) ∈ xreg(I) and HFR/(I,h) (a, b) = 0.

17
Proof. From [6, Theorem 4.7], we deduce that if (a, b) ∈ xreg(I), then (I : h)(a′ ,b′ ) = I(a′ ,b′ )
for every (a′ , b′ ) ≥ (a, b) and h generic. From this, we deduce that ii) implies i).
Conversely, using Theorem 2.3, we deduce that if (a, b′ ) is an admissible bidegree for
b ≥ b, then (I : h)a′ ,b′ = Ia′ ,b′ for any (a′ , b′ ) ≥ (a, b). Together with the fact that
′

(I, h)a,b = Ra,b and [6, Theorem 4.7], this implies that (a, b) ∈ xreg(I).

As a consequence, we relate admissible bidegrees and the generators of the bigeneric

initial ideal, that is, the ideal of leading monomials of the elements in a minimal Gröbner
basis of I in generic coordinates, which we will denote as bigin(I).

Corollary 5.3. Let (a, b) ∈ Z2 such that (a, b′ ) is an admissible bidegree for all b′ ≥ b.
Then, there is no minimal generator of bigin(I) of degree (a + 1, b). Moreover, if there is
b′ ≥ b such that (a − 1, b′ ) is not an admissible bidegree and HF(I,h) (a − 1, b) = 0, then there
is a minimal generator of bigin(I) of degree (a, b0 ) for some b0 ≤ b.

Proof. Follows from Theorem 5.2 and [6, Theorem 5.3, Thm 5.4].

Example 5.4 (Cont. of Example 4.3). If we compute the bigeneric initial ideal of I with
respect to the reverse lexicographical order with the relative order being Equation (8), the
bidegrees of its generators are:

{(0, 1), (0, 2), (1, 1), (2, 0), (2, 1), (3, 1)}.

Every bidegree bigger than (2, 2) is admissible, as the Hilbert function stabilizes to the
Hilbert polynomial.

Remark 5.5. The order defined in Equation (8) is not the one we should use to eliminate
the y variables, but the ones in x. Further work is needed to understand this phenomena.

References
[1] J. Abbott, A. M. Bigatti, E. Palezzato, and L. Robbiano. Computing and using min-
imal polynomials. Journal of Symbolic Computation, 100:137–163, 2020. Symbolic
Computation and Satisfiability Checking.

[2] A. Aramova, K. Crona, and E. De Negri. Bigeneric initial ideals, diagonal subalgebras
and bigraded hilbert functions. Journal of Pure and Applied Algebra, 150(3):215–235,
2000.

[3] M. F. Atiyah and J. G. MacDonald. Introduction to Commutative Algebra. Addison-

Wesley, Reading, Massachusetts, 1969.

[4] M. Bardet, J.-C. Faugère, and B. Salvy. On the complexity of the f5 gröbner basis
algorithm. Journal of Symbolic Computation, 70:49–70, 2015.

18
 [5] D. Bayer and M. Stillman. A criterion for detecting m-regularity. Inventiones Mathe-
maticae, 87(1):1–11, Feb. 1987.

[6] M. Bender, L. Busé, C. Checa, and E. Tsigaridas. Multigraded castelnuovo-mumford

regularity and gröbner bases, 2024.

[7] M. R. Bender, J.-C. Faugère, and E. Tsigaridas. Towards mixed gröbner basis al-
gorithms: the multihomogeneous and sparse case. ISSAC - 43rd Intern. Symp. on
Symbolic & Algebraic Computation, 2018.

[8] M. R. Bender and S. Telen. Toric eigenvalue methods for solving sparse polynomial
systems. Mathematics of Computation, 91:2397–2429, 2022.

[9] M. R. Bender and S. Telen. Yet another eigenvalue algorithm for solving polynomial
systems, 2022.

[10] W. D. Brownawell and C. K. Yap. Lower bounds for zero-dimensional projections. In

Proceedings of the 2009 International Symposium on Symbolic and Algebraic Compu-
tation, ISSAC ’09, page 79–86. Association for Computing Machinery, 2009.

[11] J. Bruce, L. C. Heller, and M. Sayrafi. Characterizing multigraded regularity on prod-

ucts of projective spaces, 2021.

[12] W. Bruns and H. Herzog. Cohen-Macaulay Rings. Cambridge Studies in Advanced

Mathematics. Cambridge University Press, 1998.

[13] L. Busé. Implicit matrix representations of rational bézier curves and surfaces.
Computer-Aided Design, 46:14–24, 2014.

[14] L. Busé, F. Catanese, and E. Postinghel. Algebraic curves and surfaces: a history of
shapes, volume 4 of SISSA Springer Series. Springer, 2023.

[15] L. Busé, M. Chardin, and N. Nemati. Multigraded sylvester forms, duality and elimi-
nation matrices. Journal of Algebra, 609:514–546, 2022.

[16] M. Ceria, S. Lundqvist, and T. Mora. Degröbnerization: a political manifesto. Appli-

cable Algebra in Engineering, Communication and Computing, 33:1–49, 11 2022.

[17] M. Chardin. Some results and questions on castelnuovo–mumford regularity. In

I. Peeva, editor, Syzygies and Hilbert Functions, chapter 1, pages 1–40. Chapman and
Hall/CRC, 2007.

[18] M. Chardin and R. Holanda. Multigraded tor and local cohomology, 2022.

[19] D. Cox, J. Little, and D. O’Shea. Using Algebraic Geometry, volume 185. 2015.

[20] A. Demin, F. Rouillier, and J. Ruiz. Reading rational univariate representations on

lexicographic groebner bases. arXiv preprint arXiv:2402.07141, 2024.

19
[21] D. Eisenbud. Commutative Algebra: With a View Toward Algebraic Geometry. Grad-
uate Texts in Mathematics. Springer-Verlag GmbH, 1995.

[22] D. Eisenbud, C. Huneke, and W. Vasconcelos. Direct methods for primary decompo-
sition. Inventiones Mathematicae, 110:207–235, 12 1992.

[23] J.-C. Faugère and C. Mou. Sparse fglm algorithms. Journal of Symbolic Computation,
80:538–569, 2017.

[24] J.-C. Faugère, M. Safey El Din, and P.-J. Spaenlehauer. Gröbner bases of bihomoge-
neous ideals generated by polynomials of bidegree (1,1): Algorithms and complexity.
Journal of Symbolic Computation, 46(4):406–437, 2011.

[25] J.-C. Faugère, P. Gianni, D. Lazard, and T. Mora. Efficient computation of zero-
dimensional gröbner bases by change of ordering. Journal of Symbolic Computation,
16(4):329–344, 1993.

[26] P. Gianni, B. Trager, and G. Zacharias. Gröbner bases and primary decomposition of
polynomial ideals. Journal of Symbolic Computation, 6(2-3):149–167, 1988.

[27] J. v. d. Hoeven and G. Lecerf. On the complexity exponent of polynomial system

solving. Foundations of Comp. Math., 21:1–57, 2021.

[28] S. G. Hyun, V. Neiger, and E. Schost. Algorithms for linearly recurrent sequences of
truncated polynomials. In Proceedings of the 2021 International Symposium on Sym-
bolic and Algebraic Computation, ISSAC ’21, page 201–208. Association for Computing
Machinery, 2021.

[29] H. T. Hà and A. Van Tuyl. The regularity of points in multi-projective spaces. Journal
of Pure and Applied Algebra, 187(1):153–167, 2004.

[30] T. Krick and A. Logar. An algorithm for the computation of the radical of an ideal in
the ring of polynomials. volume 539, pages 195–205, 10 1991.

[31] D. Lazard. Gröbner-bases, gaussian elimination and resolution of systems of algebraic

equations. In Proceedings of the European Computer Algebra Conference on Computer
Algebra, EUROCAL ’83, page 146–156, Berlin, Heidelberg, 1983. Springer-Verlag.

[32] D. Lazard and F. Rouillier. Solving parametric polynomial systems. Journal of Sym-
bolic Computation, 42(6):636–667, 2007.

[33] D. Maclagan and G. G. Smith. Multigraded castelnuovo-mumford regularity. Journal

für die reine und angewandte Mathematik, 2004(571):179–212, 2004.

[34] M. Marinari, T. Mora, and H. M. Moeller. Gröbner duality and multiplicities in

polynomial system solving. International Symposium on Symbolic and Computation -
ISSAC, Proceedings, 06 1996.

20
[35] B. Mourrain, S. Telen, and M. Van Barel. Truncated normal forms for solving polyno-
mial systems: Generalized and efficient algorithms. Journal of Symbolic Computation,
102:63–85, 2021.

[36] B. Mourrain and P. Trebuchet. Border basis representation of a general quotient al-
gebra. In International Conference on Symbolic and Algebraic Computation (ISSAC),
pages 265–272, Grenoble, France, July 2012. ACM Press.

[37] D. G. Northcott. Finite free resolutions. Cambridge University Press, Cambridge,

1976. Cambridge Tracts in Mathematics, No. 71.

[38] T. Stacks Project Authors. Stacks Project. https://stacks.math.columbia.edu,

2018.

[39] C. Vermeersch and B. De Moor. Two complementary block macaulay matrix algorithms
to solve multiparameter eigenvalue problems. Linear Algebra and its Applications, 654,
09 2022.

[40] C. Vermeersch and B. De Moor. Two double recursive block macaulay matrix algo-
rithms to solve multiparameter eigenvalue problems. IEEE Control Systems Letters,
7:1–1, 01 2022.

21