Ideals Generated by Quadratic Polynomials

arXiv:1106.0839v1 [math.AC] 4 Jun 2011

Tigran Ananyan and Melvin Hochster∗

September 24, 2018

Abstract
Let R be a polynomial ring in N variables over an arbitrary field K
and let I be an ideal of R generated by n polynomials of degree at most
2. We show that there is a bound on the projective dimension of R/I that
depends only on n, and not on N . The proof depends on showing that if K
is infinite and n is a positive integer, there exists a positive integer C(n),
independent of N , such that any n forms of degree at most 2 in R are
contained in a subring of R generated over K by at most t ≤ C(n) forms
G1 , . . . , Gt of degree 1 or 2 such that G1 , . . . , Gt is a regular sequence in
R. C(n) is asymptotic to 2n2n .

1 Introduction
Throughout this paper let R denote a polynomial ring over an arbitrary field
K: say R = K[x1 , . . . , xN ]. We will denote the projective dimension of the R-
module M over R by pd(M ). The following conjecture was posed by M. Stillman
in [10].
Conjecture 1. There is an upper bound, independent of N , on pd(R/I), where
I is any ideal of R generated by n homogeneous polynomials of given degrees
d1 , . . . , dn .
One motivation for proving this conjecture comes from its equivalence to the
following open question (the proof of equivalence due to Caviglia is given in [4]):
Conjecture 2. There is a bound on the Castelnuovo-Mumford regularity of
an ideal in polynomial ring that depends only on the number of its minimal
generators and the degrees of those generators.
Another motivation for Conjecture 1 comes from certain results about ideals
generated by three homogeneous polynomials. A construction of Burch ([2])
in the local case, extended to the global case by Kohn ([7]), shows that there
exist three-generated ideals of arbitrarily large projective dimension. However
∗ The second author was partially supported by a grant from the National Science Foun-

dation (DMS–0901145).

1
in the polynomial ring case, as the projective dimension grows, the degrees of
the generators are growing as well, thus motivating Conjecture 1 (see [4] for
analysis of growth of degrees in Burch’s construction).

Conjecture 1 clearly holds when n ≤ 2 or when all of the di = 1, but even

the simplest next case of n = 3 and d1 = d2 = d3 = 2 requires non-trivial
arguments. D. Eisenbud and C. Huneke proved that if I is generated by three
quadratic forms, then pd(R/J) ≤ 4. B. Engheta showed the existence of a
bound on the projective dimension for three cubic forms (see [5], [6]); he has
shown that pd(R/I) ≤ 36 (even though in all known examples of three cubics
pd(R/I) ≤ 5). In general, the projective dimension can grow relatively fast as
one increases the number of generators and the degrees (see [1], [3], and [9] for
specific examples). In this paper the authors prove Conjecture 1 for the case
when all of the di are at most 2 and n is arbitrary.

Since a base change of the field K to a larger field does not affect the projec-
tive dimension, in the remainder of this paper we often pass to the case where
K is infinite.

2 The Main Results

We use recursion on h to define a function B(m, n, h) of three nonnegative
integers, with h ≤ n, as follows. Let B(m, n, 0) = m + mn = m(n + 1). If h ≥ 1,
let

B(m, n, h) = (m + h)(n3 + n2 + n + 1) + h(n + 1) + B (m + h)n2 , n, h − 1 .

We also define C(s) to be the largest value of B(m, n, h) + h for m + n = s

and 0 ≤ h ≤ n − 1. Let C0 (s) be the largest value of B(0, n, h) + m + h for
m + n = s and 0 ≤ h ≤ n − 1. It is shown in §6 that both C0 (s) and C(s) are
asymptotic to 2s2s .

Our goal is to prove the following result.

Main Theorem 1. Let R = K[x1 , . . . , xN ] be a polynomial ring over a field K.
Let F1 , . . . , Fm+n be forms of degree at most 2 generating ideal I, and suppose
that the linear forms among them span a K-vector space of dimension m. Let
h denote the height of the ideal generated by the images of the quadratic Fj in
the ring obtained from R by killing the ideal generated by all of the Fi that are
linear.
If K is infinite, after a linear change of variables, there are at most b ≤
B(m, n, h) variables y1 , . . . , yb and at most c ≤ h quadratic forms G1 , . . . , Gc
in I such that
(1) The elements y1 , . . . , yb , G1 , . . . , Gc form a regular sequence in R.

2
(2) The Fi are in the polynomial ring K[y1 , . . . , yb , G1 , . . . , Gc ].
(3) The quadratic forms G1 , . . . , Gc are also a regular sequence modulo those
generators of I that involve only the variables y1 , . . . , yb .
(Condition (3) is automatic: see Lemma 2 below).
Hence, if K is infinite, the elements F1 , . . . , Fm+n are contained in the
K-subalgebra of R generated by a regular sequence of at most B(m, n, h) + h
linear and quadratic forms. It also follows that, if K is infinite, the elements
F1 , . . . , Fm+n are contained in the K-subalgebra of R generated by a regular
sequence of at most C(m + n) forms of degree at most 2.
Consequently, for every field K, the projective dimension of R/I is at most
B(0, n, h) + m + h, and is also at most C0 (m + n).
Note that the last statement follows from the first because the projective
dimension drops by m when we kill m variables in I and pass to the polynomial
ring in the remaining variables, replacing the original Fj by the images of the
quadratic Fj in the smaller polynomial ring.
Note also that if h = n then there is a much better result, in that F1 , . . . , Fm+n
is already a regular sequence.
Observe also that, as mentioned earlier, the projective dimension does not
change when we enlarge the base field.

Given polynomials of degree at most 2 that need not be forms, each is the
sum of at most one quadratic form, one linear form, and a scalar that is already
in K. Hence:
Main Theorem 2. Let R = K[x1 , . . . , xN ] be a polynomial ring over a field
K. Let F1 , . . . , Fs be polynomials of degree at most 2, and let I = (F1 , . . . , Fs ).
If K is infinite, the elements F1 , . . . , Fs are contained in the K-subalgebra of
R generated by a regular sequence of at most C(2s) forms of degree at most 2.
Consequently, for any field K, the projective dimension of R/I is at most
C(2s). 
Finally, observe that these theorems hold even if we allow the ambient poly-
nomial ring to have a set of variables of arbitrary infinite cardinality, because
any given finite set of polynomials will only involve finitely many of the variables.

3 Preliminary results
To prove these theorems, we will need the following four lemmas.
Lemma 1. Let I be a homogeneous ideal in the polynomial ring R. If we take
the image of I after killing some of the variables, its height does not increase.
Much more generally, let R be a locally equidimensional catenary Noetherian
ring, let x be any element not in any minimal prime of R, and let I be any ideal
such that for every minimal prime P of I, P + xR 6= R. Then the height of
I(R/xR) is at most the height of I.

3
 Proof. The first statement follows from the second by induction on the
number of variables killed. Note that when I is proper homogeneous, so are all
of its minimal primes, and this is preserved both by adding xR for a variable
x and by killing xR. Without loss of generality, we may replace the ideal I by
a minimal prime P of the same height. Since R is locally equidimensional and
catenary, the proper ideal P + xR has height equal to either ht(P ) or ht(P ) + 1,
depending on whether x ∈ P or not: this can be checked after localizing at any
minimal prime of P + xR. After xR is killed, the height drops by 1, because x
is not in any minimal prime of R. 
Lemma 2. If Fr+1 , . . . , Fr+s form a regular sequence modulo a set of vari-
ables containing the variables occurring in F1 , . . . , Fr , then they form a regular
sequence modulo the ideal (F1 , . . . , Fr ).
Proof. The variables killed generate a ring A, and we may form the quotient
ring B = A/(F1 , . . . , Fr )A. The polynomial ring C in the rest of the variables
over B is flat over B. It suffices to prove the statement after replacing B, C
by their localizations at their homogeneous maximal ideals. But then the result
follows from [8], Ch. 8, (20.F): if the image of a sequence in the closed fiber of
a flat local extension is regular, so is the original sequence. 

A regular sequence of forms in a polynomial ring over a field K can be

extended to a homogeneous system of parameters. The original polynomial
ring is module-finite and free over the polynomial ring generated over K by the
homogeneous system of parameters. Hence:
Lemma 3. Let F1 , . . . , Ft be a regular sequence of forms in a polynomial ring
R over a field K. Then R is free, hence, faithfully flat over A = K[F1 , . . . , Ft ].
Hence, for any ideal J of R whose generators lie in A, pd(R/J) ≤ t. 
Lemma 4. Let K be any field, let y1 , . . . , yr , z1 , . . . , zs be r +s indeterminates
over K, let α1 , . . . , αn ∈ K[y1 , . . . , yr ], and let β1 , . . . , βn ∈ K[z1 , . . . , zs ].
Then we can map the polynomial ring K[T1 , . . . , Tn ] onto K[α1 , . . . , αn ] as a
K-algebra so that each Ti 7→ αi . Call the kernel P . Likewise, we can map the
polynomial ring K[T1 , . . . , Tn ] onto K[β1 , . . . , βn ] as a K-algebra so that each
Ti 7→ βi . Suppose that the kernel of this map is also P . Finally, we can map
the polynomial ring K[T1 , . . . , Tn ] onto

K[α1 + β1 , . . . , αn + βn ] ⊆ K[y1 , . . . , yr , z1 , . . . , zs ]

as a K-algebra so that each Ti 7→ αi + βi . Suppose that the kernel of this map

contains P . Assume also that P is homogeneous, which is automatic if the αi
(or the βi ) are homogeneous of the same degree over K. Then P is generated
by linear forms over K.
Proof. Let L denote an algebraic closure of K. Since

0 → P → K[T1 , . . . , Tn ] → K[α1 , . . . , αn ] → 0

4
is exact and K[α1 , . . . , αn ] → K[y1 , . . . , yr ] is injective, the same holds after we
apply L ⊗K . Similar remarks apply when we consider the K-algebra spanned
by the βi . It follows that L ⊗K P , which may be identified with expansion of P
to L[y1 , . . . , yr ], is the kernel of the map of L[T1 , . . . , Tn ] onto L[α1 , . . . , αn ],
and taken together with the corresponding facts for the βi and the αi + βi ,
this shows that our hypotheses are preserved after base change to L. If we
know the result for L, we know that L ⊗K P is generated as an ideal by its
degree 1 part, [L ⊗K P ]1 = L ⊗K [P ]1 . Then [P ]1 K[T1 , . . . , Tn ] expands to
P L[T1 , . . . , Tn ]. Since L[T1 , . . . , Tn ] is faithfully flat over K[T1 , . . . , Tn ], the
contractions [P ]1 K[T1 , . . . , Tn ] and P , respectively, are equal. That is, P is
generated by [P ]1 .
Hence, we may assume without loss of generality that K is algebraically
closed. Consider the algebraic set V defined by P in affine n-space over K. V
is closed under scalar multiplication because P is homogeneous. We claim that
V is closed under addition. For suppose that (c1 , . . . , cn ), (d1 , . . . , dn ) ∈ V .
Then there is a K-homomorphism K[α1 , . . . , αn ] → K sending each αi → ci
and, similarly, a K-homomorphism K[β1 , . . . , βn ] → K sending every βi → di .
This yields a K-homomorphism θ : K[α1 , . . . , αn ] ⊗K K[β1 , . . . , βn ] → K
such that for every i, αi 7→ ci and βi 7→ di . We may tensor the injections
K[α1 , . . . , αn ] ⊆ K[y1 , . . . , yr ] and K[β1 , . . . , βn ] ⊆ K[z1 , . . . , zs ] over K to
obtain an injection

K[α1 , . . . , αn ] ⊗K K[β1 , . . . , βn ] ⊆ K[y1 , . . . , yr , , z1 , . . . , zs ]

that enables us to identify K[α1 , . . . , αn ] ⊗K K[β1 , . . . , βn ] with its image

K[α1 , . . . , αn , β1 , . . . , βn ] ⊆ K[y1 , . . . , yr , z1 , . . . , zs ]. Hence, we may think
of θ as a K-homomorphism K[α1 , . . . , αn , β1 , . . . , βn ] → K such that for every
i, 1 ≤ i ≤ n, αi + βi 7→ ci + di . Since every polynomial H ∈ P vanishes on
(α1 + β1 , . . . , αn + βn ), we have that H(c1 + d1 , . . . , cn + dn ) = 0 as well. This
completes the proof that V is closed under addition, and so V is a vector sub-
space of K n . But then its defining prime ideal P is generated by linear forms.


4 The basic step: putting the Fi and variables

in standard form.
We are going to make iterative of use the construction of putting the Fi and the
variables in standard form described just below.

We shall say that the Fi and the variables xj are in standard form if the
following conditions hold:

(1) Fn+i = xi for 1 ≤ i ≤ m and the remaining Fi are quadratic or 0.

(2) No Fi for 1 ≤ i ≤ n has a monomial term in K[x1 , . . . , xm ].

5
(3) F1 , . . . Fh , x1 , . . . , xm , xm+h+1 , . . . , xN is a regular sequence.
(4) There is a nonnegative integer r ≤ (m + h)n such that when the Fi are
written as polynomials in the variables x1 , . . . , xh+m with coefficients in
the ring K[xh+m+1 , . . . , xN ], the coefficients occurring for the variables
x1 , . . . , xh+m are in the K-span of xh+m+1 , . . . , xh+m+r .
(5) There is a nonnegative integer s ≤ (m + h)n2 such that when the Fi are
written as polynomials in the variables x1 , . . . , xh+m+r with coefficients in
the ring K[xh+m+r+1 , . . . , xN ], the coefficients occurring for the variables
xh+m+1 , . . . , xh+m+r are in the K-span of xh+m+r+1 , . . . , xh+m+r+s .
(6) We call xm+1 , . . . , xm+h front variables. We shall also denote them u1 , . . . , uh .
Let fi denote the image of Fi under the K-homomorphism π : R →
K[u1 , . . . , uh ] that kills all the variables except the front variables while
fixing the front variables (note that π kills Fi for i > n). Then there is an
integer d, h ≤ d ≤ n, such that f1 , . . . , fd are linearly independent over K
while fi = 0 for i > d. We call f1 , . . . , fn the front polynomials.

Under these conditions we call x1 , . . . , xm leading variables, and, as al-

ready mentioned, we call xm+1 , . . . , xm+h the front variables, and denote them
u1 , . . . , uh . We call xh+m+1 , . . . , xh+m+r the primary coefficient variables and
use the alternative notation v = v1 , . . . , vr for them. We call

xh+m+r+1 , . . . , xh+m+r+s

the secondary coefficient variables and we write w = w1 , . . . , ws as an alterna-

tive notation for them. We refer to xh+m+r+1 , . . . , xN as the tail variables. We
use the alternative notation z = z1 , . . . , zN −(h+m+r) for the tail variables. It is
very important to note that the tail variables include the secondary coefficient
variables w.
Let τ denote the K-homomorphism from R to the polynomial ring in the tail
variables that fixes the tail variables and kills the others. We write gi = τ (Fi ).
Again, τ kills Fi for i > n. We call g1 , . . . , gn the tail polynomials.
We can map the polynomial ring K[T1 , . . . , Tn ] onto K[f1 , . . . , fn ] by Ti 7→
fi . The kernel is a prime ideal P in K[T1 , . . . , Tn ], which we call the ideal of
front relations.

We want to see that these conditions can be achieved. Note, however, that
standard form is far from unique.

Achieving standard form.

We want to show that the sequence can be put in standard form without
changing the ring it generates.
If the Fi are linearly dependent over K, one or more Fi may be replaced by
0, and we assume that any 0 elements occur in the final spots among the first n

6
terms of the sequence. Hence, we may assume that the linear forms that occur
are linearly independent. We may assume that the Fi are numbered so that any
linear forms occur as a final segment, and we may assume these forms are also
an initial segment of the variables. Thus, we have xi = Fn+i , 1 ≤ i ≤ m.
We may now subtract from each Fi , 1 ≤ i ≤ n, the sum of all terms that
occur that involve only x1 , . . . , xm .
Since K is infinite, after replacing F1 , . . . , Fh by suitably general linear com-
binations of the quadratic forms F1 , . . . , Fn , we may assume that the images of
F1 , . . . , Fh form a maximal regular sequence in (F1 , . . . , Fn )R/(x1 , . . . , xm )R.
Note that for i ≤ n, every Fi will be quadratic or else 0. Since regular sequences
of forms are permutable, we have that F1 , . . . , Fh , x1 , . . . , xm is a regular se-
quence, and since K is infinite we may extend F1 , . . . , Fh , x1 , . . . , xm to a homo-
geneous system of parameters using linear forms, which we may assume are the
variables xm+h+1 , . . . , xN . That is, F1 , . . . , Fh , x1 , . . . , xm , xh+m+1 , . . . , xN
form a homogeneous system of parameters for K[x1 , . . . , xN ]. The images of
F1 , . . . , Fh when we kill the variables in this regular sequence are precisely the
elements f1 , . . . , fh ∈ K[u1 , . . . , uh ]. Hence, the quadratic forms f1 , . . . , fh
are a homogeneous system of parameters for K[u1 , . . . , uh ]. In particular, they
are linearly independent over K. The remaining Fi may be permuted so that
f1 , . . . , fd is a K-vector space basis for the K-span of the fi . For i > d we may
subtract a K-linear combination of F1 , . . . , Fd from Fi to arrange that fi be
zero.
When F1 , . . . , Fn are written as polynomials in the leading and front vari-
ables x1 , . . . , xm , u1 , . . . , uh (with coefficients in the polynomial ring in the
remaining variables) there are most m + h linear terms: the coefficients are
linear forms in xm+h+1 , . . . , xN . All these coefficients from all of the Fi span
a K-vector space of dimension r ≤ n(m + h). We make a linear change of
the variables xm+h+1 , . . . , xN so that the variables xm+h+1 , . . . , xm+h+r span
this space. These are the primary coefficient variables, which are also denoted
v1 , . . . , vr .
Next, when F1 , . . . , Fn are written as polynomials in the leading, front,
and primary coefficient variables x1 , . . . , xm , u1 , . . . , uh , v1 , . . . , vr (with co-
efficients in the polynomial ring in the remaining variables) there are most
n(m + h) terms that are linear involving one of the vj : the coefficients are
linear forms in xm+h+r+1 , . . . , xN . All these coefficients from all of the Fi span
a K-vector space of dimension s ≤ n2 (m + h). We make a linear change of the
variables xm+h+1 , . . . , xN so that the variables xm+h+r+1 , . . . , xm+h+r+s span
this space. These are the secondary coefficient variables, which we also denote
w1 , . . . , ws .
Finally, we refer to xm+h+r+1 , . . . , xN as the tail variables. This set is
the complement in {x1 , . . . , xN } of the union of leading, front, and primary
coefficient variables. We emphasize again that the tail variables include the
secondary coefficient variables.
All of the conditions for standard form are now satisfied. We remark that
once we have standard form, it is unaffected by permuting Fh+1 , . . . , Fd , or by
multiplying by an invertible matrix over K, thereby replacing them with linear

7
combinations that have the same K-span.
We show in the next section that if this procedure is then carried through
a second time using the forms consisting of all leading, front, primary and
secondary coefficient variables (these are the new leading variables) and the
tails, thus producing a second set of tails, then one of two things happens:
(1) The second set of tails generates an ideal of height at most h − 1 modulo the
ideal generated by the new leading variables. (It may also happen that the
set of original tails generates an ideal of height at most h − 1 modulo the
ideal generated by the original leading variables, which is an easier case.)
(2) The leading, front, primary and secondary coefficient variables and the
nonzero elements in first set of tails form a regular sequence.
Either condition yields an estimate of what is needed for B(m, n, h). The
first condition requires a larger value and leads to the recursive definition given
earlier. Details are given in the next section.

5 Key behavior of standard form and the proof

of the main theorems
Placing F1 , . . . , Fm+n and the variables in standard form puts surprising con-
straints on the forms. In particular, parts (c), and (d) of the Key Lemma below
play a central role in the proof of the main results.
Key Lemma. Let F1 , . . . , Fm+n consist of quadratic and linear forms in the
polynomial ring R = K[x1 , . . . , xN ] which, together with the variables x1 , . . . , xN
are assumed to be in standard form. Let all notation and terminology, including
m, n, h, r, s, d, fi , gi , ui , vi , wi , P, π, and τ , be as in the preceding section. Let
I = (F1 , . . . , Fm+n )R. Then:

(a) The elements f1 , . . . , fh form a homogenous system of parameters for the

ring K[u1 , . . . , uh ], and since

K[f1 , . . . , fh ] ⊆ K[f1 , . . . , fd ] = K[f1 , . . . , fn ] ⊆ K[u1 , . . . , uh ]

the domain K[f1 , . . . , fn ] has Krull dimension h, so that P has height n−h.
(b) For 1 ≤ i ≤ n, Fi uniquely has the form fi + ei + gi where

ei ∈ K[x1 , . . . , xm , u, v, w]

and is also in the ideal of this ring generated by (x1 , . . . , xm , v). Hence,

K[F1 , . . . , Fm+n ] ⊆ K[g1 , . . . , gn , x1 , . . . , xm , u, v, w].

(c) If H ∈ P , then H(g1 , . . . , gn ) = 0. That is, the tail polynomials satisfy the
front relations.

8
(d) If i > d, the gi = 0, i.e., Fi = ei is in the ideal generated by the leading and
primary coefficient variables, and in the polynomial ring generated by the
leading variables, front variables and the primary and secondary coefficient
variables.

Proof. The statement that f1 , . . . , fh is a homogeneous system of parameters

was proved in the fourth paragraph of the subsection on Achieving standard
form, and the other assertions in (a) are immediate.

For (b), ei is clearly the sum of all terms in Fi not involving only front
variables nor only tail variables. From the condition (2) of the definition of
standard form there are no terms involving only leading variables. The terms
in ei that involve some xi or ui in degree 1 have coefficients in the span of the
vi by the definition of primary coefficient variables. That is,
m
X h
X
ei = Lj xj + L′k uk + e′i
j=1 k=1

where the Lj , L′j are in the K-span of v1 , . . . , vr and the terms in e′i are
quadratic in the vt or linear in the vt , and those that are linear in the vt have
coefficients in the K-span of w1 , . . . , ws by the definition of secondary coeffi-
cient variables.

To prove (c), note first that H is a sum of homogeneous elements of P and

so may be assumed homogeneous of degree k. Since −fi ≡ ei + gi modulo I, we
have

0 = (−1)k H(f1 , . . . , fn ) = H(−f1 , . . . , −fn ) ≡ H(e1 +g1 , . . . , en +gn ) modulo I,

so that
H(e1 + g1 , . . . , en + gn ) ∈ I.
If H(g1 , . . . , gn ) 6= 0, we obtain a contradiction by showing that

F1 , . . . , Fh , H(e1 + g1 , . . . , en + gn )

is a regular sequence of length h+1 in R/(x1 , . . . , xm ). To this end, it suffices to

show that we have a regular sequence modulo (x1 , . . . , xm , v1 , . . . , vr ). Killing
the vj kills the ej by (b) above, and so it suffices to show that

f1 + g1 , . . . , fh + gh , H(g1 , . . . , gn )

is a regular sequence in the polynomial ring in the front and tail variables. This
is immediate from Lemma 2: since regular sequences of forms are permutable,
it suffices to show that

H(g1 , . . . , gn ), f1 + g1 , . . . , fh + gh

9
is a regular sequence. Since the first element is nonzero, it is a nonzerodivisor
involving only the tail variables, and modulo the tail variables the remaining
terms become the regular sequence f1 , . . . , fh .

Finally, (d) follows from (c) and the fact that fi = π(Fi ) = 0 for i > d: we
may take H = Ti , and this gives that gi = 0. The remaining statements now
follow from (b) above. 

Proof of Main Theorem 1. If h = 0, then F1 , . . . , Fn are expressible as sums

of multiples by linear forms of x1 , . . . , xm . All the multipliers together span a
vector space of dimension at most mn, and so all of F1 , . . . , Fn can be expressed
in terms of m + mn variables, which include the variables x1 , . . . , xm .

We now assume h ≥ 1 and use induction on h. We may assume standard

form. We consider two cases, (1) and (2). Case (2) has subcases (2a) and (2b).
Note that Case (1) is, in a sense, subsumed in Case (2a). Let h′ be the height
of the ideal generated by the tail polynomials modulo the secondary coefficient
variables. Of course, h′ ≤ h.

Case 1: h′ ≤ h − 1. By the induction hypothesis, we may work in the poly-

nomial ring in the tail variables z with the sequence of tail polynomials and

secondary coefficient variables w to select at most k ≤ B(n2 (m + h), n, h − 1
tail variables zj1 , . . . , zjk , such that K[g1 , . . . , gn , w1 , . . . , ws ] is contained in
the K-algebra generated by these zjν and a sequence of at most h − 1 distinct
quadratic forms that, together with the zjν form a regular sequence. When we
include the leading, front, and primary coefficient variables (which are disjoint
from the tail variables), we still have a regular sequence, and by part (b) of the
Key Lemma, the algebra these generate contains K[F1 , . . . , Fm+n ]. In this case
the constraint placed on B(m, n, h) is that it be at least

m + h + n(m + h) + B n2 (m + h), n, h − 1).

Case 2: h′ = h. Let S be the polynomial ring over K generated by the tail

variables. We work with sequence g1 , . . . , gn , w1 , . . . , ws in S. We may now
put these in standard form. We can do so without changing g1 , . . . , gh : we
claim these already form a regular sequence of length h modulo the secondary
coefficient variables. (To meet the condition that an initial segment of the new
front polynomials be independent over K while the rest are zero, we may adjust
Fh+1 , . . . , Fd by multiplying by an invertible matrix over K: note the com-
ment on p. 7.) The reason is that there is a homomorphism of K-algebras
K[f1 , . . . , fn ] → K[g1 , . . . , gn ] sending fi 7→ gi for all i. This is well-defined
precisely because of part (d) of the Key Lemma: the gi satisfy all the relations
in P on the fi . It is clearly surjective. it follows that K[g1 , . . . , gn ] is module-
finite over K[g1 , . . . , gh ]. But K[g1 , . . . , gn ] maps onto the new ring of front

10
polynomials, which will have Krull dimension h. Thus, K[g1 , . . . , gh ] must have
dimension h, and no less. This implies that g1 , . . . , gh is a regular sequence,
even modulo the secondary coefficient variables. Thus, the prime ideal of re-
lations on g1 , . . . , gn must be exactly P . The same holds for the new ideal of
front relations on the front polynomials of the gi . We let α1 , . . . , αn denote the
new front polynomials let β1 , . . . , βn denote the new tail polynomials. Let h′′
denote the height of the ideal generated by the new tail polynomials in the ring
generated over K by the new tail variables modulo the ideal generated by the
new secondary coefficient variables.

Subcase (2a): h′′ ≤ h − 1. This is very much like Case (1). We have that
K[F1 , . . . , Fm+n ] is contained in the polynomial ring generated over K by the
original leading, front, primary coefficient variables, secondary coefficient vari-
ables, and the original tail polynomials. But K[w1 , . . . , ws , g1 , . . . , gn ] is con-
tained in turn in the K algebra generated by the secondary coefficient variables
(which are also the new leading variables), the new front, new primary coef-
ficient , and new secondary coefficient variables, and the new tails. We count
variables as follows:

Type of variable or term Cardinality

leading m
front h
primary coefficient (m + h)n
secondary coefficient = new leading (m + h)n2
new front h
(m + h)n2 + h n


new primary coefficient
covered by B (m + h)n2 , n, h − 1


new secondary coefficient and new tails

In this case the constraint placed on B(m, n, h) is that it be at least

m + h + (m + h)n + (m + h)n2 + h + (m + h)n2 + h n + B (m + h)n2 , n, h − 1

which may be rewritten as

(m + h)(n3 + n2 + n + 1) + h(n + 1) + B (m + h)n2 , n, h − 1 .

Since this is larger than what was needed in Case (1), we no longer need to con-
sider Case (1). This is the formula used in the recursive definition of B(m, n, h).

Subcase (2b): h′′ = h. In this case we show that d = h. To see this, note
that the the relations on g1 , . . . , gn are exactly given by P , and because h′ = h
we know that the relations on α1 , . . . , αn are exactly given by P again. But
the elements β1 , . . . , βn are homomorphic images of the elements g1 , . . . , gn , so
that these satisfy P , and so are the elements α1 + β1 , . . . , αn + βn , which must
also satisfy P . Moreover, since h′′ = h, the algebra generated by the elements

11
β1 , . . . , βn has dimension h, and P gives the ideal of all relations on the ele-
ments β1 , . . . , βn . We may apply Lemma 4 to conclude that the P is generated
by linear forms. Since standard form was set up so that f1 , . . . , fd are linearly
independent but the fi for i > h are integral over K[f1 , . . . , fh ], we must have
that fi = 0 for i > h, i.e., d = h as claimed. Hence, the only nonzero tail polyno-
mial are g1 , . . . , gh , which form a regular sequence modulo x1 , . . . , xm , u, v, w.
Then K[F1 , . . . , Fm+n ] ⊆ K[g1 , . . . , gh , x1 , . . . , xm , u, v, w], and the only con-
straint on B(m, n, h) is that it be at least (m + h)(1 + n + n2 ). Therefore, using
the formula obtained in Subcase (2a) will cover all cases. 

6 Size estimates
We study briefly the size of the functions B(m, n, h), C0 (s), and C(s). We shall
use the alternative notation B(m, n, h) = Bh (m, n) = Bh . Let

g = g(n) = n3 + n2 + n + 1,

and let
g1 = g1 (n) = g(n) + n + 1.
Ph−1
Thus, g and g1 are both monic in n of degree 3. Let θh (t) = i=0 ti , so that
θ0 = 0 and θh = tθh−1 + 1 for h ≥ 1. We claim that for all h,

(m + h) (n + 1)(n2 )h + gθh (n2 )

is a lower bound for Bh , while

(m + h) (n + 1)(n2 + 1)h + g1 θh (n2 + 1)

is an upper bound. Note that the two agree with B0 when h = 0. The proof
that the given expressions are bounds is by induction on h. If we assume the
estimates for a given h − 1 ≥ 0, we have

Bh = (m + h)g + h(n + 1) + Bh−1 (m + h)n2 , n

≤ (m+h)g+(m+h)(n+1)+ (m+h)n2 +h−1 (n+1)(n2 +1)h−1 +g1 θh−1 (n2 +1) .

We may replace the summand h−1 in the first factor of the third term by m+h.
The expression then becomes
 
2 2 h−1 2
(m + h) g1 + (n + 1) (n + 1)(n + 1) + g1 θh−1 (n + 1) ,

which simplifies to the required form. We also have

Bh ≥ (m + h)g + h(n + 1) + (m + h)n2 + (h − 1) (n + 1)(n2 )h−1 + gθh−1 (n2 ) .

12
We may simply drop the terms h(n + 1) and h − 1 to get the required estimate.

A lower bound for both C0 (s) and for C(s) may be obtained from the case
m = 0, s = n, h = s − 1, which yields

C(s) ≥ (s − 1) (s + 1)(s2 )s−1 + g(s)θs−1 (s2 ) + (s − 1).

Since whenever h = s we must have m = 0 and that the given forms are already
a regular sequence, we may assume h ≤ s − 1. This yields the upper bound

C(s) ≤ (s − 1) (s + 1)(s2 + 1)s−1 + g1 (s)θs−1 (s2 + 1) + (s − 1).

Both bounds are asymptotic to 2s2s , since θs−1 (t) = (ts−1 − 1)/(t − 1), and
(s2 +1)s−1 is asymptotic to (s2 )s−1 . Hence, both C0 (s) and C(s) are asymptotic
to 2s2s .

7 The question in higher degree

We raise here the following question: is there an integer C(n, d) such that given
n forms F1 , . . . , Fn of degree at most d in a polynomial ring R over an infinite
field K, there exists a regular sequence of forms G1 , . . . , Gk ∈ R of degree at
most d with k ≤ C(n, d) such that F1 , . . . , Fn ∈ K[G1 , . . . , Gk ]? Our main
result here is, of course, the case d = 2. An affirmative answer would yield
results analogous to our main theorems for arbitrary d: C(n, d) would bound
the projective dimension of R/(F1 , . . . , Fn ) over any field, and C(dn, d) would
give corresponding results when the Fi are polynomials of degree at most d, not
required to be homogeneous.

References
[1] J. Beder, J. McCullough, L. Nunez-Betancourt, A. Seceleanu, B. Snapp,
B. Stone Ideals with larger projective dimension and regularity, preprint,
10pp., arXiv:1101.3368v1 [math.AC].
[2] L. Burch A note on the homology of ideals generated by three elements in
local rings, Proc. Cambridge Philos. Soc. 64 (1968), pp. 949–952.
[3] G. Caviglia, M. Kummini Some ideals with large projective dimension, Proc.
Amer. Math. Soc. 136 (2008), no. 2, pp. 505–509.
[4] B. Engheta Bounds on projective dimension, Ph.D. thesis, University of
Kansas, 2005.
[5] B. Engheta A bound on the projective dimension of three cubics, J. Symbolic
Comput. 45 (2010), no. 1, pp. 60–73.

13
 [6] B. Engheta On the projective dimension and the unmixed part of three
cubics, J. Algebra 316 (2007), no.2, pp. 715-734.
[7] P. Kohn Ideals generated by three elements, Proc. Amer. Math. Soc. 35
(1972), pp. 55–58.
[8] H. Matsumura Commutative Algebra, Benjamin, New York, 1970.
[9] J. McCullough A family of ideals with few generators in low degree and large
projective dimension, Proc. Amer. Math. Soc. 139 (2011), pp. 2017–2023.
[10] I. Peeva, M. Stillman Open problems on syzygies and Hilbert functions, J.
Commut. Algebra Vol. 1, no. 1 (2009) pp. 159–195.

Department of Mathematics Department of Mathematics

University of Michigan Adrian College
Ann Arbor, MI 48109–1109 Adrian, MI 49221
USA USA
E-mail: E-mail:
[email protected] [email protected]

14