What Can Be Computed
in Algebraic Geometry?

Dave Bayer
David Mumford

May 7, 1992

Partially supported by NSF grant DMS-90-06116.

This paper evolved from a long series of discussions between the two authors, going back to around 1980, on the problems of making effective computations in algebraic geometry, and it took more definite shape in a survey talk given by the second author at a conference on Computer Algebra in 1984. The goal at that time was to bring together the perspectives of theoretical computer scientists and of working algebraic geometers, while laying out what we considered to be the main computational problems and bounds on their complexity. Only part of the talk was written down and since that time there has been a good deal of progress. However, the material that was written up may still serve as a useful introduction to some of the ideas and estimates used in this field (at least the editors of this volume think so), even though most of the results included here are either published elsewhere, or exist as “folk-theorems” by now.

The article has four sections. The first two parts are concerned with the theory of Gröbner bases; their construction provides the foundation for most computations, and their complexity dominates the complexity of most techniques in this area. The first part introduces Gröbner bases from a geometric point of view, relating them to a number of ideas which we take up in more detail in subsequent sections. The second part develops the theory of Gröbner bases more carefully, from an algebraic point of view. It could be read independently, and requires less background. The third part is an investigation into bounds in algebraic geometry of relevance to these computations. We focus on the regularity of an algebraic variety (see Definition 3.2), which, beyond its intrinsic interest to algebraic geometers, has emerged as a measure of the complexity of computing Gröbner bases (see [BS87a], [BS87b], [BS88]). A principal result in this part is a bound on the regularity of any smooth variety by the second author: Theorem 3.12(b). This bound has stimulated subsequent work, and has now been generalized by [BEL91]. Another result of interest is Proposition 3.13, which elucidates the scheme structure of the ideal membership problem. The fourth part is a short discussion of work on algorithms for performing some other key operations on varieties, some open problems about these operations and some general ideas about what works and what doesn’t, reflecting the prejudices of the authors.

One of the difficulties in surveying this area of research is that mathematicians from so many specialties have gotten involved, and they tend both to publish in their own specialized journals and to have specific agendas corresponding to their area. Thus one group of researchers, the working algebraic geometers, are much more interested in actually computing examples than in worst-case complexity bounds. This group, including the first author, has put a great deal of work into building a functioning system, Macaulay, based on Gröbner bases, which has solved many problems and provided many examples to the algebraic geometry community [BS92a]. Another group comes from theoretical computer science and is much more interested in theoretical bounds than practical systems (cf. the provocative comments in Lenstra’s survey [Len92]). It seems to us that more communication would be very helpful: On the one hand, the working algebraic geometer knows lots of facts about varieties that can be very relevant to finding fast algorithms. Conversely asymptotic and/or worst-case performance bounds are sometimes, at least, important indicators of real-time performance. These theoretical bounds may also reveal important distinctions between classes of procedures, and may pose new and deep problems in algebraic geometry. Thus we will see in Section 3 how regularity estimates flesh out a picture explaining why Gröbner basis computations can have such explosive worst case behavior, yet be so useful for the kinds of problems typically posed by mathematicians. Finally, to make this article more useful in bridging this gap, we have tried to include a substantial number of references in our discussions below.

1 A Geometric Introduction

Let X be a subvariety or a subscheme of projective n-space Pn, over a field k. Let be a vector bundle or a coherent sheaf supported on X. We would like to be able to manipulate such objects by computer. From algebra we get finite descriptions, amenable to such manipulations: Let S = k[x0,,xn] be the homogeneous coordinate ring of Pn. Then X can be taken to be the subscheme defined by a homogenous ideal I S, and can be taken to be the sheaf associated to a finitely generated S-module M. We can represent I by a list of generators (f1,,fr), and M by a presentation matrix F, where

M1FM0M0

presents M as a quotient of finitely generated free S-modules M0, M1. We concentrate on the case of an ideal I; by working with the submodule J = Im(F) M0, the module case follows similarly.

The heart of most computations in this setting is a deformation of the input data to simpler data, combinatorial in nature: We want to move through a family of linear transformations of Pn so that in the limit our objects are described by monomials. Via this family, we hope to pull back as much information as possible to the original objects of study.

Choose a one-parameter subgroup λ(t) GL(n + 1) of the diagonal form

λ(t) = tw0 tw1 tw n ,

where W = (w0,,wn) is a vector of integer weights. For each t0, λ(t) acts on X via a linear change of coordinates of Pn, to yield the subscheme Xt = λ(t)XX. The limit

X0 = lim t→0 Xt

is usually a simpler object, preferable to X for many computational purposes.

Even if we start out by restricting X to be a subvariety rather than a subscheme of Pn, it does not suffice to take the limit X0 set-theoretically; often all we will get pointwise in the limit is a linear subspace L Pn, reflecting little besides the dimension of the original variety X. By instead allowing this limit to acquire embedded components and a nonreduced structure, we can obtain an X0 which reflects much more closely the character of X itself.

We compute explicitly with the generators f1,,fr of I: Let λ act on S by mapping each xi to twix i; λ maps each monomial xA = x 0a0x nan to tWAxA = tw0a0++wnanx 0a0x nan. If f = axA + bxB + , then λf = atWAxA + btWBxB + . We take the projective limit in(f) = lim t→0 λf by collecting the terms of λf involving the least power of t; in(f) is then the sum of the terms axA of f so W A is minimal. For a given f and most choices of λ, in(f) consists of a single term.

The limit X0 we want is defined with all its scheme structure by the ideal in(I) = lim t→0 λI, generated by the set {in(f)|f I}. For a given I and most choices of λ, in(I) is generated by monomials. Unfortunately, this definition is computationally unworkable because I is an infinite set, and in(I) need not equal (in(f1),,in(fr)) for a given set of generators f1,,fr of I. To understand how to compute in(I), we need to look more closely at the family of schemes Xt defined by λ.

Let S[t] be the polynomial ring k[x0,,xn,t]; we view S[t] as the coordinate ring of a one-parameter family of projective spaces Ptn over the affine line with parameter t. For each generator fj of I, rescale λfj so the lowest power of t has exponent zero: Let gj = tλf j, where = W A is the least exponent of t in λfj. Then fj = gjt=1 and in(fj) = gjt=0. Now, let J S[t] be the ideal generated by (g1,,gr); J defines a family Y over A1 whose central fiber is cut out by (in(f1),,in(fr)).

What is wrong with the family Y ? Y can have extra components over t = 0, which bear no relation to its limiting behavior as t 0. Just as the set-theoretic limit lim t→0 Xt can be too small (we need the nonreduced structure), this algebraically defined limit can be too big; the natural limit lies somewhere in between.

The notion of a flat family captures exactly what we are looking for here. For example, if Y is flat, then there are no extra components over t = 0. While the various technical definitions of flatness can look daunting to the newcomer, intuitively flatness captures exactly the idea that every fiber of a family is the natural scheme-theoretic continuation of its neighboring fibers.

In our setting, all the Xt are isomorphic for t0, so we only need to consider flatness in a neighborhood of t = 0. Artin [Art76] gives a criterion for flatness applicable here: The syzygies of g1,,gr are the relations h1g1 + + hrgr = 0 for h1,,hr S[t]. Syzygies correspond to elements (h1,,hr) of the S[t]-module S[t]r; the set of all syzygies is a submodule of S[t]r. Y is a flat family at t = 0 if and only if the restrictions (h1t=0,,hrt=0) of these syzygies to the central fiber generate the S-module of syzygies of g1t=0,,grt=0.

When g1t=0,,grt=0 are single terms, their syzygies take on a very simple form: The module of syzygies of two terms axA, bxB is generated by the syzygy bxC(axA) axD(bxB) = 0, where xE = xCxA = xDxB is the least common multiple of xA and xB. The module of syzygies of r such terms is generated (usually not minimally) by the syzygies on all such pairs.

We want to lift these syzygies to syzygies of g1,,gr, working modulo increasing powers of t until each syzygy lifts completely. Whenever we get stuck, we will find ourselves staring at a new polynomial gr+1 so tg r+1 J for some > 0. Including gr+1 in the definition of a new J J has no effect on the family defined away from t = 0, but will cut away unwanted portions of the central fiber; what we are doing is removing t-torsion. By iterating this process until every syzygy lifts, we obtain explicit generators g1,,gr,gr+1,,gs for a flat family describing the degeneration of X = X1 to a good central fiber X0. The corresponding generators g1t=1,,gst=1 of I are known as a Gröbner basis for I.

This process is best illustrated by an example. Let S = k[w,x,y,z] be the coordinate ring of P3, and let I = (f1,f2,f3) S for

f1 = w2 xy, f 2 = wy xz, f3 = wz y2.

I defines a twisted cubic curve X P3; X is the image of the map (r,s)(r2s,r3,rs2,s3). Let

λ(t) = t16 t4 t1 t0 .

If waxbyczd is a monomial of degree < 4, then λ waxbyczd = twaxbyczd where = 16a + 4b + c. Thus, sorting the monomials of S of each degree < 4 by increasing powers of t with respect to the action of λ is equivalent to sorting the monomials of each degree in lexicographic order.

We have

g1 = t32λf 1 = w2 t27xy, g2 = t17λf 2 = wy t13xz, g3 = t16λf 3 = wz t14y2.

The module of syzygies on w2, wy, wz is generated by the three possible pairwise syzygies; we start with the syzygy y(w2) w(wy) = 0. Substituting g1, g2 for the lead terms w2, wy we get

y(w2 t27xy) w(wy t13xz) = t13wxz t27xy2

which is a multiple t13x of g3. Thus, the syzygy

yg1 wg2 t13xg 3 = 0

of g1, g2, g3 restricts to the monomial syzygy y(w2) w(wy) = 0 when we substitute t = 0, as desired.

Similarly, the syzygy

zg1 t14yg 2 wg3 = 0

restricts to the monomial syzygy z(w2) w(wz) = 0. When we attempt to lift z(wy) y(wz) = 0, however, we find that

z(wy t13xz) y(wz t14y2) = t13xz2 + t14y3.

xz2 is not a multiple of w2, wy, or wz, so we cannot continue; J = (g1,g2,g3) does not define a flat family. Setting t = 1, the troublesome remainder is xz2 + y3. Making this monic, let f4 = xz2 y3; f4 I and

g4 = t4λf 4 = xz2 ty3.

Adjoin g4 to the ideal J, redefining the family Y . Now,

zg2 yg3 + t13g 4 = 0

restricts to z(wy) y(wz) = 0 as desired.

The module of syzygies of w2, wy, wz, and xz2 is generated by the pairwise syzygies we have already considered, and by the syzygy xz(wz) w(xz2) = 0, which is the restriction of

ty2g 2 + xzg3 wg4 = 0.

Thus, J = (g1,g2,g3,g4) defines a flat family Y , and

w2 xy, wy xz, wz y2, xz2 y3

is a Gröbner basis for I. The limit X0 is cut out by the monomial ideal in(I) = (w2,wy,wz,xz2), which we shall see shares many properties with the original ideal I. Note that xz2 y3 = 0 defines the projection of X to the plane P2 in x, y, and z.

The scheme structure of X0 is closely related to the combinatorial structure of the monomial k-basis for Sin(I): For each degree d in our example, the monomials not belonging to in(I) consist of three sets {xd,xd1y,,yd}, {xd1z,xd2yz,,yd1z}, {yd,yd1z,,zd}, and a lone extra monomial xd1w. The first two sets correspond to a double line supported on w = z = 0, the third set to the line w = x = 0, and the extra monomial to an embedded point supported at w = y = z = 0. Together, this describes the scheme structure of X0. The first two sets consist of d + 1 and d monomials, respectively; the third set adds d 1 new monomials, and overlaps two monomials we have already seen. With the extra monomial, we count 3d + 1 monomials in each degree, which agrees with the dimensions of the graded pieces of SI. The embedded point is crucial; it makes this count come out right, and it alone keeps X0 nonplanar like X.

The new monomial generator xz2 of in(I) excludes the line w = y = 0 from X0; combinatorially, it excludes all but three monomials of the set {xd,xd1z,,zd} from the monomial k-basis for each degree of the quotient Sin(I). We can see that this line is unwanted as follows: Away from t = 0, Y is parametrized by (r,s,t)(t16r2s,t4r3,trs2,s3,t). Thus, fixing r and s, the curve (r,ts,t)(t17r2s,t4r3,t3rs2,t3s3,t), with projective limit (0, 0,r,s, 0) as t 0. Similarly, the curve (r,t3s,t2) has as its limit (0,r2,s2, 0, 0). These calculations show that the lines w = z = 0 and w = x = 0 indeed belong set-theoretically to the limit X0. We can find no such curve whose limit is a general point on the line w = y = 0, for (r,t4s,t3) doesn’t work. Thus, the line w = y = 0 sticks out of the good total space Y .

One usually computes Gröbner bases by working directly in the ring S, dispensing with the parameter t. The one-parameter subgroup λ is replaced by a total order on the monomials of each degree, satisfying the multiplicative property xA > xB xCxA > xCxB for all xC. In fact, for our purposes these are equivalent concepts: The weight vector W associated with λ induces the order xA > xBW A < W B, which is a total multiplicative order in low degrees as long as no two monomials have the same weight. Conversely, given any multiplicative order and a degree bound d, one can find many λ which induce this order on all monomials of degree < d. See [Bay82], [Rob85] for characterizations of such orders.

We shall be particularly interested in two multiplicative orders, the lexicographic order used in our example, and the reverse lexicographic order. The lexicographic order simply expands out the monomials of each degree into words, and sorts them alphabetically, i.e. xA > xB iff the first nonzero entry in A B is positive. The reverse lexicographic order pushes highest powers of xn in any expression back to the end, then within these groups pushes highest powers of xn1 to the end, etc., i.e. xA > xB iff the last nonzero entry of A B is negative.

What do these orders mean geometrically? The dominant effect of the lexicographic order is a projection from Pn to Pn1, eliminating x0. A second order effect is a projection to Pn2, and so forth. We could compute the deformation from X to X0 with respect to the lexicographic order in stages carrying out these projections, first applying a λ with W = (1, 0,, 0), then with W = (1,1, 0,, 0), etc. Alternatively, for monomials of each degree < d, we can apply the single λ with W = (dn1,,d,1, 0), generalizing the λ used in our example. Use of the lexicographic order tends to muck up the family Y more than necessary in most applications, because projections tend to complicate varieties.

For the reverse lexicographic order, the dominant effect is a projection of Pn down to the last coordinate point (0,, 0, 1). As a secondary effect, this order projects down to the last coordinate line, and so forth. In other words, this order first tries to make X into a cone over the last coordinate point, and only then tries to squash the result down to or cone it over the last coordinate line, etc. For monomials of each degree < d, this can be realized by applying λ with W = (0, 1,d,,dn1). Like such cones, the reverse lexicographic order enjoys special properties with respect to taking linear sections of X or X0 by intersection with the spaces defined by the last variable(s) (see [BS87a]). The preferred status of the reverse lexicographic order can be attributed to this relationship, because generic linear sections do not complicate varieties.

For example, if we take X to be three general points in P2, then using the lexicographic order X0 becomes a triple point on a line, because the first order effect is the projection of the three points to a line, and the second order limiting process keeps the points within this line. By contrast, if we use the reverse lexicographic order then X0 becomes the complete first order neighborhood of a point (a point doubled in all directions). This is because the first order limiting process brings the three points together from distinct directions, tracing out a cone over the three points. The first order neighborhood of the vertex in this cone has multiplicity 3, and is the same as the complete first order neighborhood in the plane of this vertex.

For those familiar with the theory of valuations in birational geometry [ZS76, Vol. II, Ch. VI], the lexicographic and reverse lexicographic orders have simple interpretations. Recall that if X is a variety of dimension n, and

F : X = Z0 Z1 Z2 Zn

is a flag of subvarieties, codimX(Zi) = i, with Zi smooth at the generic point of Zi+1, then we can define a rank n valuation vF on X as follows: For each i = 1,,n 1, fix fi to be a function on Zi1 with a 1st order zero on Zi. Then for any function f, we can define e1 = ordZ1(f), e2 = ordZ2((ff1e1) Z1), etc., and vF (f) = (e1,,en) Zn, where the value group Zn is ordered lexicographically. The arbitrarily chosen fi are not needed to compare two functions f, g: We have vF (f) vF (g) if and only if ordZ1(fg) > 0, or if this order is zero and ordZ2((fg)Z1) > 0, and so forth. Such a valuation also defines an order on each graded piece Sd of the homogeneous coordinate ring: take any f0 Sd and say f > g if and only if vF (ff0) vF (gf0). More generally, one may take the Zi to be subvarieties of a variety X dominating X and pull back functions to X before computing vF .

The lexicographic order on monomials of each degree of Pn is now induced by the flag

Pn V (x 0) V (x0,x1) V (x0,,xn1).

For example, the first step in the comparison defining vF (xAf 0) vF (xBf 0) has the effect of asking if a0 b0 > 0.

The reverse lexicographic order is induced by a flag on a blowup X of Pn: First blow up V (x0,,xn1) and let E1 be the exceptional divisor. Next blow up the proper transform of V (x0,,xn2), and let E2 be this exceptional divisor. Iterating, we can define a flag

X E1 E1 E2 E1 En

which induces the reverse lexicographic order on monomials in each degree. For example, looking at the affine piece of the first blow up obtained by substituting x0 = x0x n1, , xn2 = xn2x n1, the power of xn1 in the transform of xA is a0 + + an1, which is the order of vanishing of this monomial on E1. Thus, the first step in the comparison defining vF (xAf 0) vF (xBf 0) has the effect of asking if a0 + + an1 b0 bn1 > 0, which is what we want.

Taking into account the equivalence between multiplicative orders and one-parameter subgroups, the process we have described in S[t] is exactly the usual algorithm for computing Gröbner bases. It is computationally advantageous to set t = 1 and dismiss our extra structure as unnecessary scaffolding, but it is conceptually advantageous to treat our viewpoint as what is “really” going on; many techniques of algebraic geometry become applicable to the family Y , and assist in analyzing the complexity of Gröbner bases. Moreover, this picture may help guide improvements to the basic algorithm. For example, for very large problems, it could be computationally more efficient to degenerate to X0 in several stages; this has not been tried in practice.

The coarsest measure of the complexity of a Gröbner basis is its maximum degree, which is the highest degree of a generator of the ideal in(I) defining X0. This quantity is bounded by the better-behaved regularity of in(I): The regularity of an ideal I is the maximum over all i of the degree minus i of any minimal ith syzygy of I, treating generators as 0th syzygies. When I is the largest (the saturated) ideal defining a scheme X, we call this the regularity of X. We take up regularity in detail in Section 3; here it suffices to know that regularity is upper semi-continuous on flat families, i.e. the regularity can only stay the same or go up at special fibers.

Let reg(I) denote the regularity of I, and reg0(I) denote the highest degree of a generator of I. In our case, t = 0 is the only special fiber, and the above says that

reg0(I) reg(I) reg(in(I)) reg0(in(I)),

where reg0(I) can be immediately determined from the input data, and reg0(in(I)) is the degree-complexity of the Gröbner basis computation. In practice, each of these inequalities are often strict.

However when k is infinite, then for any set of coordinates for Pn chosen from a dense open set U GL(n + 1) of possibilities, Galligo ([Gal74]; see also [BS87b]) has shown that the limiting ideal in(I) takes on a very special form: in(I) is invariant under the action of the Borel subgroup of upper triangular matrices in GL(n + 1). This imposes strong geometric conditions on X0. In particular, the associated primes of in(I) are also Borel-fixed, so they are all of the form (x0,,xi) for various i. This means that the components of X0 are supported on members of a flag.

In characteristic zero, it is shown in [BS87a] that the regularity of a Borel-fixed ideal is exactly the maximum of the degrees of its generators, or in our notation, that reg(in(I)) = reg0(in(I)) when in(I) is Borel-fixed. Thus, for generic coordinates in characteristic zero, the degree-complexity of computing Gröbner bases breaks down into two effects: the gap reg0(I) reg(I) between the input degrees and the regularity of X, and the gap reg(I) reg(in(I)) allowed by upper-semicontinuity.

A combination of theoretical results, hunches and experience guides the practitioner in assessing the first gap; what about the second? Does the regularity have to jump at all? One can easily find examples of ideals and total orders exhibiting such a jump, but in [BS87a], it is shown that for the reverse lexicographic order, in generic coordinates and any characteristic, there is no jump: reg(I) = reg(in(I)), so in characteristic zero we have

reg0(in(I)) = reg(I).

In this sense, this order is an optimal choice: For the reverse lexicographic order, the degree-complexity of a Gröbner basis computation is exactly the regularity of the input data. This agrees with experience; computations made on the same inputs using the lexicographic order can climb to much higher degrees than the reverse lexicographic order, in practice.

For many applications, one is free to choose any order, but some problems restrict us to using orders satisfying combinatorial properties which the reverse lexicographic order fails to satisfy. An example, developed further in Section 2, is that of eliminating variables, or equivalently, of computing projections. To compute the intersection of I with a subring R = k[xi,,xn], it is necessary to use an order which in each degree sorts all monomials not in R ahead of any monomial in R. The lexicographic order is an example of such an order, for each i simultaneously. This strength comes at a cost; we are paying in regularity gaps for properties we may not need in a particular problem. An optimal order if you need one specific projection (in the same sense as above) is constructed by sorting monomials by total degree in the variables to be eliminated, and then breaking ties using the reverse lexicographic order. See [BS87b] for this result, and a generalization to the problem of optimally refining any nonstrict order.

Using this elimination order, one finds that the inherent degree-complexity of a computation is given not by the regularity of X itself, but rather by the regularity of the flat projection X of X, which is the central fiber of a flat family which animates the desired projection of X as t 0. The jump in regularity between X and X is unavoidable; by choosing an optimal order, we avoid the penalty of a further jump in regularity between X and X0.

The regularity of algebraic varieties or schemes X is far from being well understood, but there is considerable interest in its study; this computational interpretation of regularity as the inherent degree-complexity of an ideal is but one more log on the fire.

From a theoretical computer science perspective, the full complexity of computing Gröbner bases is determined not merely by the highest degree reg0(I) in the basis, but by the total number of arithmetic operations in the field k required to compute this basis. This has not been analyzed in general, but for 0-dimensional ideals I, Lakshman and Lazard ([Lak91], [LL91]) have shown that the complexity of computing reduced Gröbner bases is bounded by a polynomial in dn, where d is the maximum degree of the generators, and n is the number of variables.

2 Gröbner Bases

Let S = k[x0,,xn] be a graded polynomial ring over the field k, and let I S be a homogeneous ideal.

Let Sd denote the finite vector space of all homogeneous, degree d polynomials in S, so S = S0 S1 Sd . Writing I in the same manner as I = I0 I1 Id , we have Id Sd for each d. Recall that the Hilbert function of I is defined to be the function p(d) = dim(Id), for d 0.

A total order > on the monomials of S is said to be multiplicative if whenever xA > xB for two monomials xA, xB, then xCxA > xCxB for all monomials xC. This condition insures that if the terms of a polynomial are in order with respect to >, then they remain in order after multiplication by a monomial.

Definition 2.1 Let > be a multiplicative order. For a homogeneous polynomial f = c1xA1 + + c mxAm with xA1 > > xAm, define the initial term in(f) to be the lead (that is, the largest) term c1xA1 of f. For a homogeneous ideal I S, define the initial ideal in(I) to be the monomial ideal generated by the lead terms of all elements of I.

Note that the definitions of in(f) and in(I) depend on the choice of multiplicative order >. See [BM88] and [MR88] for characterizations of the finite set of in(I) realized as the order > varies.

Fix a multiplicative order > on S.

Proposition 2.2 (Macaulay) I and in(I) have the same Hilbert function.

Proof. ([Mac27]) The lead terms of Id span in(I)d, because every monomial xA in(I) is itself the lead term in(f) of some polynomial f I: Since xA = xCxB for some xB = in(g) with g I, we have xA = in(f) for f = xCg.

Choose a k-basis Bd Id with distinct lead terms, and let in(Bd) be the set of lead terms of Bd; in(Bd) has cardinality p(d) = dim(Id). Since any element of Id is a linear combination of elements of Bd, any lead term of Id is a scalar multiple of an element of in(Bd). Thus, in(Bd) is a basis for in(I)d, so p(d) = dim(in(I)d).     _

One can compute the Hilbert function of I by finding in(I) and applying this result; see [MM83], [BCR91], and [BS92b].

Corollary 2.3 The monomials of S which don’t belong to in(I) form a k-basis for SI.

Proof. These monomials are linearly independent in SI, because any linear relation among them is a polynomial belonging to I, and all such polynomials have lead terms belonging to in(I). These monomials can be seen to span SI by a dimension count, applying Proposition 2.2.     _

Two examples of multiplicative orders are the lexicographic order and the reverse lexicographic order. xA > xB in the lexicographic order if the first nonzero coordinate of A B is positive. For example, if S = k[w,x,y,z], then w > x > y > z in S1, and

w2 > wx > wy > wz > x2 > xy > xz > y2 > yz > z2

in S2.

xA > xB in the reverse lexicographic order if the last nonzero coordinate of A B is negative. For example, if S = k[w,x,y,z], then w > x > y > z in S1, and

w2 > wx > x2 > wy > xy > y2 > wz > xz > yz > z2

in S2. These two orders agree on S1, but differ on the monomials of S of degree > 1 when n 2.

The lexicographic order has the property that for each subring k[xi,,xn] S and each polynomial f S, f k[xi,,xn] if and only if in(f) k[xi,,xn]. The reverse lexicographic order has the property that for each f k[x0,,xi], xi divides f if and only if xi divides in(f).

One can anticipate the applications of these properties by considering a k-basis Bd Id with distinct lead terms, as in the proof of Proposition 2.2. With respect to the lexicographic order, Bd k[xi,,xn] is then a k-basis for Id k[xi,,xn] for each i. With respect to the reverse lexicographic order, Bd (xn) is then a k-basis for Id (xn). Thus, these orders enable us to find polynomials in an ideal which do not involve certain variables, or which are divisible by a certain variable. For a given degree d, one could construct such a basis Bd by applying Gaussian elimination to an arbitrary k-basis for Id. However, this cannot be done for all d at once; such a computation would be infinite. We will finesse this difficulty by instead constructing a finite set of elements of I whose monomial multiples yield polynomials in I with every possible lead term.

Such sets can be described as follows:

Definition 2.4 A list F = [f1,,fr] I is a (minimal) Gröbner basis for I if in(f1),,in(fr) (minimally) generate in(I).

in(I) is finitely generated because S is Noetherian, so Gröbner bases exist for any ideal I.

The order of the elements of F is immaterial to this definition, so F can be thought of as a set. We are using list notation for F because we are going to consider algorithms for which the order of the elements is significant. For convenience, we shall extend the notation of set intersections and containments to the lists F.

A minimal set of generators for an ideal I need not form a Gröbner basis for I. For example, if S = k[x,y] and I = (x2 + y2,xy), then with respect to the lexicographic order, in(x2 + y2) = x2 and in(xy) = xy. Yet y(x2 + y2) x(xy) = y3 I, so y3 in(I). Thus, any Gröbner basis for I must include y3; it can be shown that in(I) = (x2,xy,y3) and [x2 + y2,xy,y3] is a Gröbner basis for I.

On the other hand,

Lemma 2.5 If F = [f1,,fr] is a Gröbner basis for I, then f1,,fr generate I.

Proof. For each degree d, we can construct a k-basis Bd Id with distinct lead terms, whose elements are monomial multiples of f1,,fr: For each xA in(I) d, xA is a scalar multiple of xCin(f i) for some xC and some i; include xCf i in the set Bd. Thus, the monomial multiples of f1,,fr span I.     _

Proposition 2.6 (Spear, Trinks) Let R S be the subring R = k[xi,,xn]. If F = [f1,,fr] is a Gröbner basis for the ideal I with respect to the lexicographic order, then F R is a Gröbner basis for the ideal I R. In particular, F R generates I R.

Proof. ([Spe77], [Zac78], [Tri78]) Let f I R; in(f) is a multiple of in(fi) for some i. Since in(f) R, in(fi) R, so fi R. Thus, F R is a Gröbner basis for I R. By Lemma 2.5, F R generates I R.     _

Proposition 2.6 has the following geometric application: If I defines the subscheme X Pn, then I k[xi,,xn] defines the projection of X to Pni = Proj(k[x i,,xn]).

Recall that the saturation Isat of I is defined to be the largest ideal defining the same subscheme X Pn as I. Isat can be obtained by taking an irredundant primary decomposition for I, and removing the primary ideal whose associated prime is the irrelevant ideal (x0,,xn). I is saturated if I = Isat.

If the ideal I is saturated, and defines a finite set of points X Pn, then I k[xn1,xn] is a principal ideal (f), where {f = 0} is the image of the projection of X to P1 = Proj(k[x n1,xn]). Given a linear factor of f of the form (bxn1 axn), we can make the substitution xn1 = az, xn = bz for a new variable z, to obtain from I an ideal J k[x0,,xn2,z] defining a finite set of points in Pn1. For each point (c0,,cn2,d) in the zero locus of J, (c0,,cn2,ad,bd) is a point in the zero locus of I.

If X Pn1 is of dimension 1 or greater, then in general I k[xn1,xn] = (0), because a generic projection of X to P1 is surjective. In this case, an arbitrary substitution xn1 = az, xn = bz can be made, and the process of projecting to P1 iterated. Thus, the lexicographic order can be used to find solutions to systems of polynomial equations.

Recall that the ideal quotient (I : f) is defined to be the ideal {g S|fg I}. Since S is Noetherian, the ascending chain of ideals (I : f) (I : f2) (I : f3) is stationary; call this stationary limit (I : f) = {g S|fmg I for some m}.

Proposition 2.7 If [xna1f 1,,xnarf r] is a Gröbner basis for the ideal I with respect to the reverse lexicographic order, and if none of f1,,fr are divisible by xn, then F = [f1,,fr] is a Gröbner basis for the ideal (I : xn). In particular, f1,,fr generate (I : xn).

Proof. ([Bay82], [BS87a]) We have F (I : xn). Let f (I : xn); xnmf I for some m, so in(xnmf) is a multiple of in(xnaif i) for some i. Since fi is not divisible by xn, in(fi) is not divisible by xn, so in(f) is a multiple of in(fi). Thus, F is a Gröbner basis for (I : xn). By Lemma 2.5, f1,,fr generate (I : xn).     _

If I = q0 q1 qt is a primary decomposition of I, then (I : xn) = (q i : xn) = (q i : xn). We have (qi : xn) = (1) if the associated prime pi of qi contains xn, and (qi : xn) = q i otherwise. Thus, if I defines the subscheme X Pn, then (I : xn) defines the subscheme consisting of those primary components of X not supported on the hyperplane {xn = 0}.

(I : xn) is saturated, because it cannot have (x0,,xn) as an associated prime. If xn belongs to none of the associated primes of I except (x0,,xn), or equivalently if {xn = 0} is a generic hyperplane section of X Pn, then (I : xn) = Isat. Thus, the reverse lexicographic order can be used to find the saturation of I.

One of the most important uses of Gröbner bases is that they lead to canonical representations of polynomials modulo an ideal I, i.e. a division algorithm in which every f S is written canonically as f = gifi + h, where [f1,,fr] is a Gröbner basis for I, and h is the remainder after division.

Recall the division algorithm for inhomogeneous, univariate polynomials f(x), g(x) k[x]: Let in(f) denote the highest degree term of f. The remainder of g under division by f can be recursively defined by

Rf(g) = Rf(g cxaf)

if in(f) divides in(g), where cxa = in(g)in(f), and by

Rf(g) = g

otherwise.

Division can be generalized to homogeneous polynomials f1,,fr,g S, given a multiplicative order on S ([Hir64], [Bri73], [Gal74], [Sch80]): The remainder RF (g) of g under division by the list of polynomials F = [f1,,fr] can be recursively defined by

RF (g) = RF (g cxAf i)

for the least i so in(g) is a multiple cxA of in(fi), and by

RF (g) = in(g) + RF (g in(g))

if in(g) is not a multiple of any in(fi). RF (g) is an element of S.

Thus, the fate of in(g) depends on whether or not in(g) (in(f1),,in(fr)). Let I be the ideal generated by f1,,fr. If F = [f1,,fr] fails to be a Gröbner basis for I, then the remainder is poorly behaved. For example, with respect to the lexicographic order on k[x,y],

R[xy,x2+y2](x2y) = x2y x(xy) = 0,

but

R[x2+y2,xy](x2y) = x2y y(x2 + y2) = y3,

so the remainder RF (g) is dependent on the order of the list F. Note that x2y (x2 + y2,xy).

If on the other hand, F is a Gröbner basis for the ideal I, then RF (g) is a k-linear combination of monomials not belonging to in(I). By Corollary 2.3, these monomials form a k-basis for SI, so each polynomial in S has a unique representation in terms of this k-basis, modulo the ideal I. The remainder gives this unique representation, and is independent of the order of F (but dependent on the multiplicative order chosen for the monomials of S). In particular, RF (g) = 0 if and only if g I.

An algorithm for computing a Gröbner basis for I from a set of generators for I was first given by Buchberger ([Buc65], [Buc76]). This algorithm was discovered independently by Spear ([Spe77], [Zac78]), Bergman [Ber78], and Schreyer [Sch80]. It was termed the division algorithm by Schreyer, after the division theorem of Hironaka ([Hir64], [Bri73], [Gal74]).

Define S(fi,fj) for i < j by

S(fi,fj) = bxBf i cxCf j,

where xA = bxBin(f i) = cxCin(f j) is the least common multiple of in(fi) and in(fj). bxBf i and cxCf j each have xA as lead term, so xA cancels out in S(fi,fj), and xA > in(S(f i,fj)).

If F is a Gröbner basis for the ideal I, then RF (S(fi,fj)) = 0 for each i < j, since S(fi,fj) I. Conversely,

Proposition 2.8 (Buchberger) If RF (S(fi,fj)) = 0 for each i < j, then F = [f1,,fr] is a Gröbner basis for the ideal I = (f1,,fr).

See [Buc65], [Buc76]. We postpone a proof until the theory has been extended to S-modules. This result can also be thought of as an explicit converse to the assertion that if F is a Gröbner basis, then division is independent of the order of F: Whenever we have a choice in division between subtracting off a multiple of fi and a multiple of fj, the difference is a multiple of S(fi,fj). If division is independent of the order of F, then these differences must have remainder zero, so by Proposition 2.8, F is a Gröbner basis.

As sketched in Section 1, Proposition 2.8 can be used to compute a Gröbner basis from a set of generators f1,,fr for the ideal I: For each i < j so fr+1 = RF (S(fi,fj))0, adjoin fr+1 to the list F = [f1,,fr]. Note that fr+1 I. By iterating until no new polynomials are found, a Gröbner basis F is obtained for I. This process terminates because S is Noetherian, and each new basis element corresponds to a monomial not in the ideal generated by the preceding lead terms.

We now extend this theory to S-modules. Let M be a graded, finitely generated S-module, given by the exact sequence of graded S-modules

M1FM0M0,

where M0 = Se01 Se0q and M1 = Se11 Se1r are free S-modules with deg(eij) = dij for each i, j. We now think of F both as a list [f1,,fr] of module elements, and as a map between free modules: Let fi = F(e1i)0 for i = 1,,r, and let I M0 be the homogeneous submodule generated by f1,,fr. Thus, M = M0I.

A monomial of M0 is an element of the form xAe 0i; such an element has degree deg(xA) + d 0i. An order on the monomials of M0 is multiplicative if whenever xAe 0i > xBe 0j, then xCxAe 0i > xCxBe 0j for all xC S. For some applications, such as developing a theory of Gröbner bases over quotients of S, one wants this order to be compatible with an order on S: If xA > xB, then one wants xAe 0i > xBe 0i for i = 1,,r. The orders encountered in practice invariably satisfy this second condition, but it does not follow from the first, and we do not require it here.

One way to extend a multiplicative order on S to a compatible multiplicative order on M0 is to declare xAe 0i > xBe 0j if i < j, or if i = j and xA > xB. Another way is to assign monomials xC1,, xCq in S to the basis elements e01,,e0q of M0, and to declare xAe 0i > xBe 0j if xA+Ci > xB+Cj, or if A + Ci = B + Cj and i < j.

Fix a choice of a multiplicative order > on M0. The constructions developed for S carry over intact to M0, with the same proofs ([Gal79], [Sch80], [Bay82]): Given an element f M0, define in(f) to be the lead term of f. Define in(I) to be the submodule generated by the lead terms of all elements of I M0; in(I) is a monomial submodule of M0 with the same Hilbert function as I. Define F = [f1,,fr] I to be a Gröbner basis for I if in(f1),,in(fr) generate in(I); a set of generators for I need not be a Gröbner basis for I, but a Gröbner basis for I generates I. Given an element g M0, define RF (g) M0 exactly as was done for the free module S. If F is a Gröbner basis for I, then RF (g) = 0 if and only if g I.

The quotient of g under division by f1,,fr can be recursively defined by

QF (g) = cxAe 1i + QF (g cxAf i)

for the least i so in(g) is a multiple cxA of in(fi), and by

QF (g) = QF (g in(g))

if in(g) is not a multiple of any in(fi). The quotient is an element of M1.

Following the recursive definitions of the remainder and quotient, it can be inductively verified that

g = F(QF (g)) + RF (g).

If F is a Gröbner basis for I, and g I, then RF (g) = 0, so the quotient lifts g to M1. In this case, the quotient can be thought of as expressing g in terms of f1,,fr.

Define S(fi,fj) for i < j by

S(fi,fj) = bxBf i cxCf j,

if in(fi) and in(fj) have a least common multiple xAe 0k = bxBin(f i) = cxCin(f j). Leave S(fi,fj) undefined if in(fi) and in(fj) lie in different summands of M0, and so don’t have common multiples.

Recall that the module of syzygies of f1,,fr is defined to be the kernel of the map F, which is the submodule of M1 consisting of all h M1 so F(h) = 0. Thus, if h = h1e11 + + hre1r is a syzygy, then h1f1 + + hrfr = 0. Let J M1 denote the module of syzygies of f1,,fr, and let K M1 denote the module of syzygies of in(f1),,in(fr).

Define the map in(F) : M1 M0 by in(F)(e1i) = in(fi); K is the kernel of in(F). For each i < j so S(fi,fj) is defined, define tij to be the element

tij = bxBe 1i cxCe 1j M1,

where xAe 0k = bxBin(f i) = cxCin(f j) is the least common multiple of in(fi) and in(fj), as before. in(F)(tij) = 0, so each tij belongs to the syzygy module K. Observe that F(tij) = S(fi,fj).

Assign the following multiplicative order on M1, starting from the order on M0 ([Sch80]; see also [MM86]): Let xAe 1i > xBe 1j if xAin(f i) > xBin(f j), or if these terms are k-multiples of each other and i < j. If the order on M0 is compatible with an order on S, then this order on M1 is compatible with the same order on S.

With respect to this order on M1, we have

Lemma 2.9 The list [tij] is a Gröbner basis for the module K of syzygies of in(f1),,in(fr).

Proof. Let h M1, so in(F)(h) = 0. Then in(F)(in(h)) is canceled by in(F)(h in(h)) in M0. Therefore, if in(h) = xAe 1i, then h has another term xBe 1j so xAin(f i) and xBin(f j) are k-multiples of each other and i < j. Thus, tij is defined and in(tij) divides in(h), so [tij] is a Gröbner basis for K.     _

Thus, the set {tij} generates K. In general, the [tij] are far from being a minimal Gröbner basis for K; we consider the effects of trimming this list in Proposition 2.10 below.

Define

sij = tij QF (S(fi,fj))

whenever RF (S(fi,fj)) = 0. Note that in(sij) = in(tij). Each sij is the difference of two distinct elements of M1, each of which is mapped by F to S(fi,fj), so F(sij) = 0. In other words, sij belongs to the syzygy module J. Conversely,

Proposition 2.10 (Richman, Spear, Schreyer) Choose a set of pairs T = {(i,j)} such that the set {tij}(i,j)T generates the module K of syzygies of in(f1),,in(fr). If RF (S(fi,fj)) = 0 for each (i,j) T, then

(a) F = [f1,,fr] is a Gröbner basis for I;

(b) the set {sij}(i,j)T generates the module J of syzygies of f1,,fr.

Moreover,

(c) if [tij](i,j)T is a Gröbner basis for K, then [sij](i,j)T is a Gröbner basis for J.

Proof. ([Ric74], [Spe77], [Zac78], [Sch80]) First, suppose that [tij](i,j)T is a Gröbner basis for K. Let h J, so F(h) = 0. By the same reasoning as in the proof of Lemma 2.9, we can find (i,j) T so in(tij) divides in(h). Since in(sij) = in(tij), in(sij) also divides in(h), so [sij](i,j)T is a Gröbner basis for J, proving (c).

Now, suppose that {tij}(i,j)T merely generates K. Let T be a set of pairs so [tm](,m)T is a Gröbner basis for K. It is enough to construct a list [um](,m)T of elements of J, generated by {sij}(i,j)T , so in(um) = in(tm) for all (,m) T. Then by the preceding argument, [um](,m)T is a Gröbner basis for J, so {sij}(i,j)T generates J.

Write each tm = gmijtij, for (,m) T and (i,j) T, in such a way that the terms of tm and each term of each product gmijtij map via in(F) to multiples of the same monomial in M0. In other words, find a minimal expression for each tm, which avoids unnecessary cancellation. Then define

um = gmijsij.

We have in(um) = in(tm), proving (b).

Let f I, and choose g M1 so f = F(g). Let h M1 be the remainder of g under division by [um](,m)T; f = F(h). Since in(h) is not a multiple of any in(um) = in(tm), the lead term of F(in(h)) is not canceled by any term of F(h in(h)). Therefore, if in(h) = axAe 1i, then in(fi) divides in(F). Thus, F = [f1,,fr] is a Gröbner basis for I, proving (a).     _

Proposition 2.8 follows as a special case of this result.

The above proof can be understood in terms of an intermediate initial form in0(h) for h M1: Apply the map in(F) separately to each term of h, and let xA M 0 be the greatest monomial that occurs in the set of image terms. Define in0(h) to be the sum of all terms of h which map via in(F) to multiples of xA. Then in refines in0, for according to the order we have defined on M1, in(h) is the term of in0(h) lying in the summand of M1 whose basis element ei has the smallest index i.

In this language, tij = in0(tij) = in0(sij). Our expressions for the tm have the property that each gmijtij = in0(gmijtij), with each term of each product for a given tm mapping via in(F) to multiples of the same monomial xA. Thus, each in0(gmijsij) = gmijtij; the tails gmij(sij in0(sij)) stay out of our way, mapping termwise via in(F) to monomials which are less than xA with respect to the order on M0.

Observe that QF (g) is a linear combination of monomials not belonging to in(J), for any g M0.

In [Buc79], Buchberger gives a criterion for selecting a set T of pairs (i,j) in the case where I is an ideal: If (i0,i1), (i1,i2),, (is1,is) T, and the least common multiple of in(fi0),in(fi1),,in(fis) is equal to the least common multiple of in(fi0) and in(fis), then (i0,is) need not belong to T. In other words, if ti0is (ti0i1,,tis1is), then the pair (i0,is) is unnecessary; this condition is equivalent to the condition of Proposition 2.10, for the case of an ideal.

Suppose that we wish to compute the syzygies of a given set of elements g1,,gs of M0. To do this, compute a Gröbner basis f1,,fr for the submodule I M0 generated by g1,,gs. Keep track of how to write each fi in terms of g1,,gs. Using these expressions, each syzygy of f1,,fr can be mapped to a syzygy of g1,,gs. These images generate the module of syzygies of g1,,gs; the set of syzygies obtained in this way is not in general minimal.

Syzygies can be used to find a minimal set of generators for a submodule I M0 from a given set of generators g1,,gs: If h1g1 + + hrgr = 0 is a syzygy of g1,,gs with h1 k, then g1 = (h2g2 + + hrgr)h1, so g1 is not needed to generate I. All unnecessary generators can be removed in this way.

Alternatively, a careful implementation of Gröbner bases can directly find minimal sets of generators for submodules: Starting from an arbitrary set of generators, we can eliminate unnecessary generators degree by degree, by removing those which reduce to zero under division by a Gröbner basis for the ideal generated by the preceding generators.

Either way, we can trim the set of syzygies computed via Gröbner bases for a given set of generators g1,,gs of I, to obtain a minimal set of generators for the syzygy module J. By starting with a minimal generating set for I, and iterating this method, a minimal free resolution can be found for I.

A beautiful application of these ideas yields a proof of the Hilbert syzygy theorem, that minimal free resolutions terminate (Schreyer [Sch80], [Sch91], for an exposition see also Eisenbud [Eis92]). At each stage of a resolution, order the Gröbner basis F for I in such a way that for each i < j, letting in(fi) = axAe 0k and in(fj) = bxBe 0, we have xA > xB in the lexicographic order. If the variables x1,,xm are missing from the initial terms of the fi, then the variables x1,,xm+1 will be missing from the initial terms of the syzygies sij. Iterating, we run out of variables, so the resolution terminates.

3 Bounds

How hard are the algorithms in algebraic geometry? We describe some key bounds. The best known example is the bound established by G. Hermann [Her26] for ideal membership:

Theorem 3.1 (G. Hermann) Let k be any field, let (f1,...,fk) k[x1,...,xn] and let d = max(deg(fi)). If g (f1,...,fk), then there is an expression

g = i=1ka ifi

where deg(ai) deg(g) + 2(kd)2n1.

This type of bound is called “doubly exponential”. However, with the advent of the concept of coherent sheaf cohomology [Ser55] and the systematic study of vanishing theorems, it has become apparent that the vanishing of these groups in high degrees is almost always the most fundamental bound. The concept of an ideal being “m-regular” or “regular in degrees m” was introduced by one of us [Mum66] by generalizing ideas of Castelnuovo:

Definition 3.2 1 Let k be any field, let I k[x0,,xn] be an ideal generated by homogeneous polyomials, let Id be the homogeneous elements in I of degree d, let I be the corresponding sheaf of ideals in 𝒪Pn, and let I(d) be the dth twist of I. Then the following properties are equivalent and define the term “m-regular”:

(a) the natural map Im H0((m)) is an isomorphism and Hi((m i)) = (0), 1 i n

(b) the natural maps Id H0((d)) are isomorphisms for all d m and Hi((d)) = (0) if d + i m, i 1.

(c) Take a minimal resolution of I by free graded k[X]-modules:

0 α=1 rn k[x] eα,nϕnϕ1 α=1 r0 k[x] eα,0ϕ0k[x]k[x]I 0.

Then deg(eα,i) m + i for all α, i. (In particular, if fα = ϕ0(eα,0), then f1,,fr0 are minimal generators of I, and deg(eα,0) = deg(fα) m.)

The intuitive idea is that past degree m, nothing tricky happens in the ideal I. Unfortunately, neither (a), (b) nor (c) can be verified by any obvious finite algorithm. This lack of a finitely verifiable criterion for m-regularity has been remedied by a joint result of the first author and M. Stillman [BS87a]:

Theorem 3.3 (Bayer-Stillman) I is m-regular if and only if the degrees of the minimal set of generators of I are at most m, and there exists a set y0,,y of linear combinations of x0,...,xn such that for all homogeneous f of degree m,

y0f I f I y1f I f I + k[x] y0 yf I f I + i=01k[x] y i

and

f I + i=0k[x] y i.

Moreover, if this holds at all, it holds for y0,,y taken arbitrarily from a Zariski-open set in the space of + 1 linear forms.

To see why m-regularity is a key bound, we want to show that it controls some of the geometric features of the ideal I. Let’s introduce several refined notions of the “degree” of I:

Definition 3.4 If I = q0 q1 qt is a primary decomposition of I, q i = pi is prime and V (pi) is the subvariety Zi of Pn for i 1, while p0 = (x0,,xn) (so that V (p0) = ), then first let q1,, qs be the isolated components, (i.e., ZiZj if 1 i s, 1 j t, ij, or equivalently, V (I) = Z1 Zs is set-theoretically the minimal decomposition of V (I) into varieties). Then let

mult(qi) = length  of a maximal chain of pi-primary ideals: qi = J J1 J1 = pi

(Equivalently, this is the length of the local ring k[x]piIk[x]pi, or, in the language of schemes, if η is the generic point of Zi, then this is the length of 𝒪η,Pn.)

deg(Zi) = usual geometric degree of Zi: the cardinality of Zi L for almost all linear spaces L of complementary dimension. geom-degr(I) = 1 i si such that  dim Zi = rmult(qi) deg(Zi)

If qi is one of the non-isolated, or embedded components, then we extend the concept of multiplicity more carefully: Let

Ii = qjj such that pj pi or equivalently Zj Zi pi

and

multI(qi) = length  of a maximal chain of ideals: qi Ii = J J1 J0 = Ii where each Jk satisfies: ab Jk,api b Jk.

(Equivalently, Jk equals qk Ii for some pi-primary ideal qk.) In particular:

I0 = j=1 t qj is known as Isat, and multI(q0) = length  of a maximal chain of ideals I = J J1 J0 = Isat = dim k(IsatI).

For s + 1 i t, an equivalent way to define multI(qi) is as the length of the module

Iik[x]piIk[x]pi

or, in the language of schemes, the length of

Ii𝒪η,PnI𝒪η,Pn

where η is the generic point of Zi.

Then write

arith-degr(I) = 1 i si such that  dim Zi = rmultI(qi) deg(Zi)

and

arith-deg1(I) = multI(q0).

The idea here is best illustrated by an example: let

I = (x12,x 1x2) k[x0,x1,x2].

Then

I = q1 q2 q1 = (x1), p1 = (x1), Z1 = {line x1 = 0} q2 = (x12,x 1x2,x22), p 2 = (x1,x2), Z2 = {point (1, 0, 0)}.

Then

deg(Z1) = 1, mult(q1) = 1

so

geom-deg1(I) = arith-deg1(I) = 1.

One might be tempted to simply define

multI(q2) =  length of chain of p2-primary ideals between q2, p2

and since

k[x]q2q2k[x]p2K 1 + K x1 + K x2,K = k(x0)

this is 3. But embedded components are not unique! In fact,

I = q1 q2 q2 = (x 12x 2) also,

which leads to

k[x]p2q2k[x] p2K 1 + K x2

which has length 2. The canonical object is not the local ring k[x]p2q2k[x]p2 but the ideal

Ker k[x]p2Ik[x]p2 k[x]p2p2k[x]p2 k x1

which has length 1. Thus, the correct numbers are

multI(q2) = 1

and

geom-deg0(I) = 0 arith-deg0(I) = 1.

Now the question arises: find bounds on these degrees in terms of generators of I. For geometric degrees, a straightforward extension of Bezout’s theorem gives:

Proposition 3.5 Let d(I) be the maximum of the degrees of a minimal set of generators of I. Then

geom-degr(I) d(I)nr.

A proof can be found in [MW83]. The idea is clear from a simple case: Suppose f,g,h K[x,y,z] and f = g = h = 0 consists of a curve C and points Pi off C. We can bound like this: Choose 2 generic combinations f, g of f, g, h so that f = g = 0 does not contain a surface. It must be of the form C C, C one-dimensional, containing all the Pi but not the generic point of C. Then by the usual Bezout theorem

deg C deg f deg g = d(I)2.

Let h be a 3rd generic combination of f,g,h. Then C{h = 0} consists of a finite set of points including the Pi’s. Thus

= #Pi #(C{h = 0}) deg C d(I) by Bezout’s theorem d(I)3.

Can arith-deg(I) be bounded in the same way? In fact, it cannot, as we will show below. Instead, we have

Proposition 3.6 If m(I) is the regularity of I, then for 1 r n,

arith-degr(I) m(I) + n r 1 n r m(I)nr

which replaces d(I) by the regularity of I. A proof is given in the technical appendix.

We have introduced two measures of the complexity of a homogeneous ideal I. The first is d(I), the maximum degree of a polynomial in a minimum set of generators of I. The second is m(I), which bounds the degrees of generators and of all higher order syzygies in the resolution of I (Definition 3.2 (c)). Obviously,

d(I) m(I).

A very important question is how much bigger can m(I) be than d(I)? The nature of the answer was conjectured by one of us in his thesis [Bay82] and this conjecture is being borne out by subsequent investigations. This conjecture is that in the worst case m(I) is roughly the (2n)th power of d(I) – a bound like G. Hermann’s. But that if I = I(Z) where Z is geometrically nice, e.g. is a smooth irreducible variety, then m(I) is much smaller, like the nth power of d(I) or better. This conjecture then has three aspects:

(1)a doubly exponential bound for m(I) in terms of d(I),
which is always valid,
(2) examples of I where the bound in (1) is best possible, or nearly so,
(3) much better bounds for m(I)
valid if V (I) satisfies various conditions.

All three aspects are partially proven, but none are completely clarified yet. We will take them up one at a time.

A doubly exponential bound for m(I) in terms of d(I) may be deduced easily in characteristic zero from the work of M. Giusti [Giu84] and A. Galligo [Gal79]:

Theorem 3.7 If char(k) = 0 and I k[x0,xn] is any homogeneous ideal, then

m(I) (2d(I))2n1 .

It seems likely that Theorem 3.7 holds in characteristic p, too. A weaker result can be derived quickly in any characteristic by straightforward cohomological methods:

Proposition 3.8 If I k[x0,xn] is any homogeneous ideal, then

m(I) (2d(I))n!.

The proof is given in the technical appendix.

Next, we ask whether Theorem 3.7 is the best possible, or nearly so. The answer is yes, because of a very remarkable example due to E. Mayr and A. Meyer [MM82].

Example 3.9 Let InA be the ideal in 10n variables S(m),F(m),C i(m),B i(m), 1 i 4, 1 m n defined by the 10n 6 generators

2 m n S(m) S(m1)C 1(m1) F(m) S(m1)C 4(m1) Ci(m)F(m1)B 2(m1) C i(m)B i(m)F(m1)B 3(m1), 1 i 4 1 m n 1 F(m)C 1(m)B 1(m) S(m)C 2(m) F(m)C 2(m) F(m)C 3(m) S(m)C 3(m)B 1(m) S(m)C 2(m)B 4(m) S(m)C 3(m) F(m)C 4(m)B 4(m)        Ci(1)S(1) C i(1)F(1)(B i(1))2, 1 i 4

Let InH be the ideal gotten from InA by homogenizing with an extra variable u. Then Mayr and Meyer [MM82, lemma 8, p. 318] prove:

Lemma 3.10 Let en = 22n. If M is any monomial in these variables, S(n)C i(n) F(n)M I nA if and only if

M = Ci(n)(B i(n))en ,

and S(n)C i(n) S(n)M I nA if and only if

M = Ci(n).

Now note that the generators of InA and InH are all of the very simple type given by a difference of two monomials. Quite generally, if

J k[x1,,xn] J = (, xαi xβi ,)1ik

then the quotient ring k[x]J has a very simple form. In fact, we get an equivalence relation between monomials generated by

xαi+γ xβi+γ, any i,γ

and

k[x]J δ k xδ

where δ runs over a set of representatives of each equivalence class.

Bearing this in mind, let’s look at the 1st order syzygies for the homogeneous ideal:

JnH = (S(n),F(n),I nH).

S(n) and F(n) are part of a minimal set of generators, and let fα InH complete them. Then syzygies are equations:

pS(n) + qF(n) + r αfα = 0.

One such is given by:

uen+eC i(n) S(n) + ue(B i(n))en Ci(n) F(n) + R αfα = 0

for some Rα, and some e 0 (the extra power ue is necessary because some terms Rαfα have degree greater than en + 2) whose degree is 2 + en + e. Now express this syzygy as a combination of a minimal set of syzygies. This gives us in particular:

uen+eC i(n) = a λpλ ue(B i(n))en Ci(n) = a λqλ pλS(n) + q λF(n) + R αλfα = 0.

Then for some λ, pλ must have a term of the form u or uC i(n), hence the monomial uS(n) or uCi(n)S(n) occurs in pλS(n). But by the general remark on quotient rings by such simple ideals, this means that this term must equal some second term MS(n) (M a monomial in pλ) or MF(n) (M a monomial in qλ) mod InH. By the lemma, the first doesn’t happen and the second only happens if the term uC i(n)(B i(n))en occurs in qλ, in which case en + 1 deg qλ = deg(syzygy(pλ,qλ,Rαλ)) 1. This proves:

Proposition 3.11 JnH has for its bounds:

d(J) = 4 m(J) 22n + 1.

Going on to the 3rd aspect of the conjecture, consider results giving better bounds for m(I) under restrictive hypotheses on V (I).

Theorem 3.12 If Z Pn is a reduced subscheme purely of dimension r, and I = I(Z) is the full ideal of functions vanishing on Z, then

(a) if r 1, or Z is smooth, char(k) = 0 and r 3, then:

m(I) deg Z n + r + 1

(b) if char(k) = 0 and Z is smooth,

m(I) (r + 1)(deg(Z) 2) + 2.

Since deg(Z) d(I)nr (Proposition 3.5), these bound m(I) in terms of d(I).

Part (a) of this are due to Gruson-Lazarsfeld-Peskine [GLP83] for r 1, and to Pinkham [Pin86], Lazarsfeld [Laz87], and Ran [Ran90] for r 3. It is conjectured by Eisenbud and Goto [EG84], and others, that the bound in (a) holds for all reduced irreducible Z, and it might well hold even for reduced equidimensional Z which are connected in codimension 1. As this problem is now understood, the needed cohomological arguments follow formally, once one can control the singularities of a projection of the variety. These singularities become progressively harder to subdue as the dimension of the variety increases, and are what impedes definitive progress beyond dimension 3.

Part (b) is due to the second author and is proven in the technical appendix. It has been generalized by Bertram, Ein, and Lazarsfeld [BEL91] to show that any smooth characteristic 0 variety of codimension e defined as a subscheme of Pn by hypersurfaces of degrees d1 dm is (d1 + de e + 1)-regular. Since we cannot decide the previous conjecture, this is a result of considerable practical importance, for it strongly bounds the complexity of computing Gröbner bases of smooth characterisitic 0 varieties in terms of the degrees of the input equations.

The biggest missing link in this story is a decent bound on m(I) for any reduced equidimensional ideal I. We would conjecture that if a linear bound as in part (a) doesn’t hold, at the least a so-called “single exponential” bound, i.e. m(I) d0(n) ought to hold. This is an essential ingredient in analyzing the worst-case behavior of all algorithms based on Gröbner bases, and would complete the story about what causes the bad examples discussed above. At least in some cases Ravi [Rav90] has proven that the regularity of the radical of a scheme is no greater than the regularity of the scheme itself.

There is a direct link between the bounds that we have given so far and the G. Hermann bound with which we started the section. This results from the following:

Proposition 3.13 Let IA k[x 1,,xn] have generators f1,,fk and let IH k[x 0,x1,,xn] be the ideal generated by homogenizations f1h,,f kh of the fi. Let IH = q 0 qt be the primary decomposition of IH, let Zi = V (qi) and let

mult(IH) = max mult I(qi1) + + multI(qik) + multI(q0)

where the max is taken over chains V ((x0)) Zi1 Zik. If g IA, then we can write:

g = i=1ka ifi

where

deg ai deg g + mult(IH).

The proof goes like this: Let gh be the homogenization of g. Consider the least integer m such that x0mgh IH. Since g I, this m is finite. Moreover, if

x0mgh =x 0mi aihf ih

then

g = aifi

and

deg ai = deg(ah) deg(x 0mgh) deg f j m + deg(g).

Now in the primary decomposition of IH, suppose that for some k,

x0kgh iS qi, and x0kghq j if jS.

Choose S such that V (q) is maximal. Since g IA, we know V (q) V ((x0)), hence x0 p. Let

IS = iS qi.

Then multI(q) is easily seen to be the length of a maximal chain of ideals between:

I k[x]p and IS k[x]p.

But look at the ideals Jp, for p 0, defined by

Ik[x]p (I,x0k+pgh)k[x] p Jp ISk[x]p.

If Jp = Jp+1, then

x0k+pgh (I,x 0k+p+1gh) i.e., x0k+pgh = ax 0k+p+1gh + b, b I.

But 1 ax0 is a unit in k[x]p, so

Jp = x0k+pgh = (1 ax 0)1b Ik[x] p.

This means that in any case

x0k+multI(q)gh I k[x] p

hence, because q is p-primary:

x0k+multI(q)gh q

Induction now shows that

x0mult (IH)gh IH

Corollary 3.14 Let IA, IH be as above. If g IA, then

g = aifi

where deg(ai) deg(g) + m(I)+n+1 n+1 .

Proof. Combine Propositions 3.6 and 3.13.

If we further estimate m(I) by Theorem 3.7 in characteristic 0 or by Proposition 3.8, we get somewhat weaker versions of Hermann’s Theorem 3.1. But if I = V (Z), Z a good variety, we may expect the Corollary to give much better bounds than Theorem 3.1.

Corollary 3.14 shows that any example which demonstrates the necessity of double exponential growth in Hermann’s ideal membership bound (Theorem 3.1) also demonstrates the necessity of double exponential growth in the bounds on m(I) given in Theorem 3.7 and Proposition 3.8. Thus we can make use of the general arguments for the existence of such examples given in [MM82], rather than depending on the single example of Proposition 3.11, to show that the bounds on m(I) inevitably grow double exponentially: Since in Corollary 3.14, the degrees of the ai are bounded by a single exponential function of m(I), in all examples where the degrees of the ai grow double exponentially, m(I) also grows double exponentially.

This line of argument gives a geometric link between the ideal membership problem and m(I): In Corollary 3.14, if IA exhibits ai of high degree, then IH has primary components of high multiplicity. These components force m(I) to be large, and distinguish IH from good ideals considered in Theorem 3.12 and related conjectures.

A major step in understanding the gap between the double exponential examples and the strong linear bounds on the regularity of many smooth varieties was taken by Brownawell [Bro87] and Kollár [Kol88]. They discovered the beautiful and satisfying fact that if we replace membership in I by membership in I, then there are single exponential bounds on the degrees of ai:

Theorem 3.15 (Brownawell, Kollár) Let k be any field, let I = (f1,...,fk) k[x1,...,xn] and let d = max(deg(fi),i = 1,,k; 3). If n = 1, replace d by 2d 1. If g I, then there is an expression

gs = i=1ka ifi

where s dn and deg(ai) (1 + deg(g))dn. In particular:

Idn I.

What this shows is that although the bad examples have to have primary components at infinity of high degree, nonetheless these primary ideals contain relatively small powers of IH. The picture you should have is that these embedded components at infinity are like strands of ivy that creep a long way out from the hyperplane at infinity, but only by clinging rather closely to the affine components.

Technical Appendix to Section 3

1. Proof of the equivalence of the conditions in Definition 3.2:

In [Mum66, pp. 99-101], it is proven that for any coherent sheaf on Pn, Hi((i)) = (0), i 1 implies that the same holds for (d), all d 0, and that H0((d)) is generated by H0() H0(𝒪(d)). In particular, if you apply this to = (m), the equivalence of (a) and (b) follows. (Note the diagram:

Id H0((d)) k[x]dH0(𝒪Pn (d))

which shows that Im H0((m)) is injective for every d). To show that (b) (c), first note that we may rephrase the reults in [Mum66] to say that if Hi((i)) = (0), i 1, then the degrees of the minimal generators of the k[x]-module

d Z H0((d))

are all zero or less. So we may construct the resolution in (c) inductively: at the kth stage, say

α=1 rk + 1 k[x] eα,k1ϕk1k[x]k[x]I0

has been constructed, let Mk =  ker(ϕk) and let k be the corresponding sheaf of ideals. The induction hypothesis will say that Hi( k(m + k 1)) = (0), i 1. Therefore Mk is generated by elements of degree m + k, i.e., dα = deg eα,k m + k, all α. We get an exact sequence

0 Mk+1 α=1 rk k[x] eα,k Mk 0

hence

0 k+1 α=1 rk 𝒪Pn(dα) k 0 (1)

Therefore

α=1 rk Hi(𝒪 Pn(m+kidα))Hi( k(m+ki)) (2) Hi+1( k+1(m+(k+1)(i+1))) α=1 rk Hi+1(𝒪 Pn(m+kidα))

is exact. But m + k i dα i so Hi+1(𝒪Pn (m + k i dα)) = (0). This shows that k+1 satisfies the induction hypothesis and we can continue. Thus (c) holds. To see that (c) (a), we just use the same exact sequences (1) and prove now by descending induction on k that Hi( k(m + k i)) = (0), i 1. Since I = 0, this does it. The inductive step again uses (2), since Hi(𝒪Pn (m + k i dα)) = (0) too.

2. Proof of Proposition 3.6:

Look first at the case r = 0. Let be the sheaf of ideals defined by I and let be the sheaf defined by omitting all 0-dimensional primary components of I. Consider the exact sequence:

0 (m 1) (m 1) ()(m 1) 0

This gives us:

H0((m 1)) H0(()(m 1)) H1((m 1))

Now H1((m 1)) = (0) by m-regularity, and h0(()(m 1)) = h0() =  length() = arith-deg 0(I) since has 0-dimensional support. But H0((m 1)) H0(𝒪Pn (m 1)), so

arith-deg0(I) h0((m 1)) h0(𝒪 Pn(m 1)) = m + n 1 n

If r > 0, we can prove the Proposition by induction on r. Let H be a generic hyperplance in Pn, given by h = 0. Let IH = (I,h)(h) k[x0,,xn](h)k[x0,,x n1] for suitable linear combinations xi of xi. Then it is easy to check that:

arith-degr(I) = arith-degr1(IH)

and that IH is also m-regular, so by induction

arith-degr1(IH) m + (n 1) (r 1) 1 (n 1) (r 1) = m + n r 1 n r

If r = 1, we use the fact that

0IdH0((d))(Isat) d

if d m, hence

dim(IsatI) dim k[x](x0,,xn)m = m + n n + 1 .

3. Proof of Proposition 3.8:

Let I k[x0,xn] and assume, after a linear change of coordinates, that xn is not contained in any associated prime ideals of I. Let I¯ k[x0,xn1] be the image of I. Then d(I¯) = d(I) and by induction we may assume

m(I¯) (2d(I))(n1)!.

We will prove, in fact, that

m(I) m(I¯) + m(I¯) 1 + n n (3)

and then we will be done by virtue of the elementary estimate:

if m = (2d(I))(n1)!, and d 2, then m + m 1 + n n (2d(I))n!

To prove (3), we use the long exact sequence

0 (I : (x0))k1 x0 Ik I¯k 0 0 H0((k 1)) H0((k 1))H0(¯(k 1))δ δH1((k 1)) H1((k)) H1(¯(k))

where (I : (x0)) = {f|x0f I}. Let m¯ = m(I¯). Note that Hi(¯(k 1)) = (0), i 1, k m¯, hence

Hi((k 1)) Hi((k))

is an isomorphism if k m¯ 1 + 1 and i 2. Since Hi((k)) = (0), k 0, this shows that Hi((k)) = (0), i 2, k m¯ i. Moreover I¯k H0(¯(k)) is an isomorphism if k m¯, hence δ = 0 if k m¯, hence H1((k)) = (0), k m¯ 1. But now look at the surjectivity of Ik H0((k)). For all k, let Mk be the cokernel. Then Mk is a k[x]-module of finite dimension. Multiplication by x0 induces a sequence:

0 (I : (x0))k1 Ik1 Mk1 x0 Mk 0

which is exact if k m¯. But if, for one value of k m¯,

(I : (x0))k = Ik (4)

then by Theorem 3.3, I is k-regular and (4) continues to hold for larger k, and Mk must be (0). In other words,

dim Mk, k m¯ 1

is non increasing and monotone decreasing to zero when k m¯. Therefore

m(I) m¯ + dim Mm¯1 m¯ + dim k[x]m¯1 m¯ + m¯ 1 + n n

which proves (3).

4. Proof of Theorem 3.12(b):

Let Z be a smooth r-dimensional subvariety of Pn and d = degree of Z. We first consider linear projections of Z to Pr and to Pr1. To get there, let L1 Pn be a linear subspace of dimension n r 1 disjoint from Z and L2 L1 a linear subspace of dimension n r 2. Take these as centers of projection:

Pn L 1 Z p2 p1 Pr+1 {P} Z 1 p2 Pr

Let x0,xr+1 be coordinates on Pr+1 so that p = (0,, 0, 1), hence x0,xr are coordinates on Pr. Let f(x0,xr+1) = 0 be the equation of the hypersurface Z1.

Now there are two ways of getting r-forms on Z: by pullback of r-forms on Pr and by residues of (r + 1)-forms on Pr1 with simple poles along Z1. The first gives us a sheaf map

p2 Ω PrrΩ Zr

whose image is ΩZr(B 1), B1 the branch locus of p2. Corresponding to this on divisor classes:

KZ p2(K Pr) + B1 (5) (r + 1)H + B1,

where H = hyperplane divisor class on Z. The second is defined by

a(x) dx1 dxr+1 f p1a(x) dx1 dxr fxr+1 (6)

and it gives us an isomorphism

p1(Ω Pr+1r+1(Z 1)Z1) ΩZr(B 2)

B2 is a divisor which can be interpreted as the conductor of the affine rings of Z over those of Z1: i.e.,

f 𝒪Z(B2) f (p1,𝒪Z) 𝒪Z1.

In particular,

p1,(𝒪Z(B2)) sheaf of 𝒪Z1  ideals C in 𝒪Z1. (7)

A classical reference for these basic facts is Zariski [Zar69], Prop. 12.13 and Theorem 15.3. A modern reference is Lipman [Lip84] (apply Def. (2.1)b to p1 and apply Cor. (13.6) to Z1 Pr+1). (4) gives us the divisor class identity:

KZ + B2 p1(K Pr+1 + Z1) (8) (d r 2)H.

(5) and (8) together tell us that

B1 + B2 (d 1)H.

In fact, the explicit description (6) of the residue tells us more: namely that if y1,,yr are local coordinates on Z, then

(x1,,xr) (y1,,yr) 1 fxr+1 dy1 dyr

generates ΩZr(B 2) locally. But (x1,,xr) (y1,,yr) = 0 is a local equation for B1, so this means that fxr+1 = 0 is a local equation for B1 + B2. But fxr+1 = 0 is a global hypersurface of degree d 1 in Pr+1, hence globally:

B1 + B2 = p1(V ( f xr1))

(equality of divisors, not merely divisor classes). All this is standard classical material.

(7) has an important cohomological consequence: let C 𝒪 Pr+1 be the sheaf of ideals consisting of functions whose restriction to Z1 lies in C. Then we get an exact sequence:

0 𝒪Pr+1(Z1) C𝒪 Pr+1 C𝒪Z1 0

hence an exact sequence

0 𝒪Pr+1( d) C𝒪 Pr+1() p1,(𝒪Z(H B2)) 0

for all integers . But H1(𝒪 Pr+1( d)) = (0), hence

H0(C𝒪 Pr+1()) H0(𝒪 Z(H B2))

is surjective, hence

H0(𝒪 Z(H B2))  Im H0(𝒪 Pr+1()) H0(𝒪 Z(H)) . (9)

Now let us vary the projections p1 and p2. For each choice of L1, we get a different B1: call it B1(L1), and for each choice of L2, as different B2: call it B2(L2). By (5) and (8), all divisors B1(L1) are linearly equivalent as are all divisors B2(L2). Moreover:

L1 B1(L1) = L2 B2(L2) =

This is because, if x Z, then there is a choice of L1 such that p1 : Z Pr is unramified at y; and a choice of L2 such that p2(x) Z1 is smooth, hence p2 is an isomorphism near x. Thus

B1(L1) = KZ + (r + 1)H

and

B2(L2) = KZ + (d r 2)H

are base point free linear systems.

Next choose (r + 1) L2’s, called L2α, 1 α r + 1, so that if B2(α) = B 2(L2(α)), then α B2(α) = . Look at the Koszul complex:

0 𝒪Z(H B2(α)) α,β𝒪Z(H B2(α) B 2(β)) α𝒪Z(H B2(α)) 𝒪 Z(H) 0.

This is exact and diagram chasing gives the conclusion:

Hi(𝒪 Z(H (i + 1)B2)) = (0), all i 1

αH0(𝒪 Z(H B2(α))) H0(𝒪 Z(H)) surjective

hence by (9)

H0(𝒪 Pn()) H0(𝒪 Z(H)) surjective

and

Hi+j(𝒪 Z(H (i + 1)B2)) = (0), all i 0

Hj(𝒪 Z(H)) = (0).

Now I(Z) is m-regular if and only if Hi( Z(m i)) = (0), i 1, hence if and only if

H0(𝒪 Pn(m 1)) H0(𝒪 Z(m 1)) surjective

Hi(𝒪 Z(m i 1)) = (0), i 1.

By the previous remark, this follows provided that

Hi+j(𝒪 Z((m i 1)H (j + 1)B2)) = (0), if i,j 0, i + j 1.

But let us rewrite:

(m i 1)H (j + 1)B2 KZ + jB1 + (m i (j + 1)(d 1) + r)H

using (5) and (8). Note that jB1 + H is an ample divisor if 1,j 0, because B1 is base point free. Therefore by the Kodaira Vanishing Theorem,

Hi(𝒪 Z(KZ + jB1 + H)) = (0), i,j 1, j 0

and provided m = (r + 1)(d 2) + 2, this gives the required vanishing.

4 Applications

From some points of view, the first main problem of algebraic geometry is to reduce the study of a general ideal I to that of prime ideals, or the study of arbitrary schemes to that of varieties. One way of doing this is to find a decomposition of the ideal into primary ideals: i.e. write it as an intersection of primary ideals. But even when non-redundancy is added, this is not unique, and usually one actually wants something less: to find its radical and perhaps write the radical as an intersection of prime ideals, or to find its top dimensional part, or to find its associated prime ideals and their multiplicities. There are really four computational problems involved here which should be treated separately: (i) eliminating the multiplicities in the ideal I, (ii) separating the pieces of different dimension, (iii) “factoring” the pieces of each dimension into irreducible components, and finally (iv) describing the original multiplicities, either numerically or by a primary ideal. Three of these four problems are the direct generalizations of the basic problems for factoring a single polynomial: we can eliminate multiple factors, getting a square-free polynomial, we can factor this into irreducible pieces and we can ask for the multiplicities with which each factor appeared in the original polynomial. There is a fifth question which arises when we work, as we always must do on a computer, over a non-algebarically closed field k: we can ask (v) for an extension field k of k over which the irreducible components break up into absolutely irreducible components.

Classical algorithms for all of these of these rely heavily on making explicit projections of V (I) to lower dimensional projective spaces. This can be done either by multi-variable resultants if you want only the set-theoretic projection, or by Gröbner bases with respect to the lexicographic order or an elimination order, to get the full ideal I k[X0,,Xm]. Recent treatments of multi-variable resultants can be found in [Can89], [Cha91], and a recent treatment of the basis method can be found in [GTZ88]. There is no evidence that either of these is an efficient method, however, and taking Gröbner bases in the lexicographical order or an elimination order is often quite slow, certainly slow in the worst case. The general experience is that taking projections can be very time consuming. One reason is that the degree of the generators may go up substantially and that sparse defining polynomials may be replaced by more or less generic polynomials. A specific example is given by principally polarized abelian varieties of dimension r: they are defined by quadratic polynomials in (4r 1)-space, but their degree here (hence the degree of their generic projection to Pr+1) is 4rr! [Mum70a]. In fact, any variety is defined purely by quadratic relations in a suitable embedding [Mum70b].

Instead of using real computational experience, the fundamental method in theoretical computer science for analyzing complexity of algorithms is to count operations. For algebraic algorithms, the natural measure of complexity is not the number of bit operations, but the number of field operations, addition, subtraction, multiplication and (possibly) division that are used. In this sense, any methods that involve taking Gröbner bases for any order on monomials will have a worst-case behavior whose complexity goes up with the regularity of the ideal hence will take “double exponential time”. However, it appears that this worst-case behavior may in fact only concern problem (iv) – finding the primary ideals – and that problems (i), (ii) and (iii) may be solvable in “single exponential time”. The idea that such algorithms should exist for finding V (I) set-theoretically was proposed in the 1984 lecture on which this article is based, but turned out, in fact, to have been already proven by Chistov and Grigoriev, cf. their unpublished 1983 note [CG83]. Their line of research led, in some sense, to the work of Brownawell and Kollár, showing the single exponential bound Im I for m = dn, where d = max( degrees of generators of I).

Based on this work, Giusti and Heintz [GH91] give a singly exponentially bounded algorithm for computing ideals qi such that V (qi) are the irreducible components of V (I) (over the ground field k). The method depends on computing what is essentially the Chow form of each component, and leads to an ideal defining this variety but not its full ideal. In fact, their qi may be guaranteed to be prime except for possible embedded components.

A direct approach to constructing both I and the intersection of the top-dimensional primary components of I, denoted Top(I), is given in a recent paper by Eisenbud, Huneke and Vasconcelos [EHV92]. Their construction of the radical uses the Jacobian ideals, i.e. the ideals of minors of various sizes of the Jacobian matrix of generators of I. This is certainly the most direct approach, but, again they have trouble with possible embedded components, and must resort to ideal quotients, hence they need a Gröbner basis of I in the reverse lexicographic order. They compute Top(I) as the annihilator of Extcodim(I)(k[X 0,,Xn]I,k[X0,,Xn]), which is readily found from a full resolution using Gröbner bases. Their algorithm appears to be practical in some cases of interest, but still has double exponential time worst-case behavior.

It may turn out to be most effective in practice to combine these ideas. Often an ideal under study has regularity far smaller than the geometric degree of its top dimensional components; projecting these components to a hypersurface requires computing in degrees up to the geometric degree, which is wasteful. On the other hand, methods such as those in [EHV92] work better in low codimensions, if only because there are fewer minors to consider in the Jacobian matrix. Thus, projecting an arbitary scheme down to low codimension and then switching to direct methods may work best of all.

This still does not settle the issue of the complexity of calculating I, or, for that matter, calculating the full prime ideal of any subvariety of codimension greater than one. Chow form type methods give you an effective method of defining the set V (I) but only of generating I up to possible embedded components. For this reason, the two schools of research, one based on the algebra of I, the other based on subsets of Pn have diverged. If we knew, as discussed in the previous section, that the regularity of a reduced ideal could be bounded singly exponentially, then we could bound the degrees of the generators of I, and, using Brownawell-Kollár, we could determine I up to these degrees and get the whole ideal. But without such a bound, it is still not clear whether only V (I) and not I can be found in worst-case single exponential time.

Let’s look at problem (iii). Assume you have found a reduced equidimensional I. To study splitting it into irreducible or absolutely irreducible pieces, we shall assume initially it is a hypersurface, i.e. I = (f). Computationally, there may often be advantages to not projecting a general I to a hypersurface, and we will discuss one such approach below. Geometrically, there is nothing very natural about irreducible but not absolutely irreducible varieties: from the standpoint of their properties, they behave like reducible varieties, except that, being conjugate over k, their components have very similar properties. If the ground field k gets bigger or smaller, the set of absolutely irreducible components gets partitioned in finer or coarser ways into the k-components. If one has never done any calculations, one would therefore be inclined to say – let’s extend k as far as needed to split our algebraic set up into absolutely irreducible components. This is a very bad idea! Unless this extension k happens to be something simple like a quadratic or cyclotomic extension of k, the splitting field k is usually gigantic. This is what happens if one component of V (I) is defined over an extension field k1 of k of degree e, and the Galois group of k1k is the full symmetric group, a very common occurence. Then V (I) only splits completely over the Galois closure of k1k and this has degree e!. The moral is: never factor unless you have to.

In fact, unless you need to deal simultaneously with more than one of its irreducible components, you can proceed as follows: the function field K = k[X0,,Xn](f) contains as a subfield an isomorphic copy of k1: you find that field as an extension k1 = k[y](p(y)) of k, and solve for the equation of one irreducible component f1 k1[X0,,Xn] by the formula Normk1k(f1) = f.

Pursuing this point, why should one even factor the defining equation f over k? Factoring, although it takes polynomial time [LLL82], is often very slow in real time, and, unless the geometry dictates that the components be treated separately, why not leave them alone. In some situations, for instance, [DD84] one may have an ideal, module or other algebraic structure defined by polynomials or matrices of polynomials over a ground ring D = k[y](p(y)), where p is a square-free polynomial. Thus D is a direct sum of extension fields, but there is no need to factor p or split up D until the calculations take different turns with the structures over different pieces of Spec(D).

The standard methods of factoring in computer algebra all depend on (i) writing the polynomial over a ring, finitely generated over Z, and reducing modulo a maximal ideal m in that ring, obtaining a polynomial over a finite field; and (ii) restricting to a line L, i.e. substituting Xi = aiX0 + bi,i 1 for all but one variable, obtaining a polynomial in one variable over a finite field. This is then factored and then, using Hensel’s lemma, one lifts this factorization modulo higher powers of m and of the linear space L. One then checks whether a coarsened version of this factorization works for f. This is all really the arithmetic of various small fields. Geometrically, every polynomial in one variable factors over a suitable extension field and the question of counting the absolutely irreducible components of a variety is really more elementary: it is fundamentally topological and not arithmetic. One should, therefore, expect there to be direct geometric ways of counting these components and separating them. Assuming I is a reduced, equi-r-dimensional ideal, the direct way should be to use Serre duality, computing the cohomology Hr(Ω V (I)r), where ΩV (I)r ω V (I) is the subsheaf of the top-dimensional dualizing sheaf of V (I) of absolutely regular r-forms. Its dimension will be the number of absolutely irreducible components into which V (I) splits. Calculating this cohomology involves two things: algebraically resolving the ideal I and geometrically resolving the singularities of V (I) far enough to work out ΩV (I)r. Classically, when I = (f) was principal, ΩV (I)r was called its ideal of “subadjoint” polynomials.

There is one case where this is quite elementary and has been carried out: this is for plane curves. One can see immediately what is happening by remarking that a non-singular plane curve is automatically absolutely irreducible, hence one should expect that its singularities control its decomposition into absolutely irreducible pieces. Indeed, if C k[X0,X1,X2](f) is the conductor ideal, then ΩV (f)1 is given by the homogeneous ideal C, but with degree 0 being shifted to be polynomials of degree d 3, d the degree of f. To calculate H1, assume X0 is not zero at any singularity of V (f) and look at the finite-dimensional vector space of all functions k[X1X0,X2X0](C + (f)) modulo the restrictions g(X0d3) for all homogeneous polynomials g of degree d 3. This will be canonically the space of functions on the set of components of V (f) with sum 0. In particular, it is (0) if and only if V (f) is absolutely irreducible. This follows from standard exact sequences and duality theory. It was known classically as the Cayley-Bacharach theorem, for the special case where V (f) was smooth except for a finite number of ordinary double points. It states that V (f) is absolutely irreducible if and only if for every double point P, there is a curve of degree d 3 passing through all the double points except P.

This example gives one instance where a deeper computational analysis of varieties requires a computation of its resolution of singularities. We believe that there will be many instances where practical problems will require such an analysis. In many ways, resolution theorems look quite algorithmic, and, for instance, Abhyankar and his school have been approaching the problem in this way [Abh82], as have Bierstone and Milman [BM91]. However, the only case of resolution of singularities to be fully analyzed in the sense of computational complexity is that of plane curves. This has been done by Teitelbaum [Tei89], [Tei90]. His analysis is notable in various ways: he is extremely careful about not making unnecessary factorizations, let alone taking unnecessary field extensions, and uses the “D” formalism discussed above. He describes his algorithm so precisely that it would be trivial to convert it to code and, as a result, he gives excellent bounds on its complexity.

References

[Abh82]    S. S. Abhyankar, Weighted expansions for canonical desingularization, Lecture Notes in Math., vol. 910, Springer-Verlag, 1982.

[Art76]    M. Artin, Lectures on deformations of singularities, Tata Institute on Fundamental Research, Bombay, 1976.

[Bay82]    Dave Bayer, The division algorithm and the Hilbert scheme, Ph.D. thesis, Harvard University, Department of Mathematics, June 1982, order number 82-22588, University Microfilms International, 300 N. Zeeb Rd., Ann Arbor, MI 48106.

[BEL91]    Aaron Bertram, Lawrence Ein, and Robert Lazarsfeld, Vanishing theorems, a theorem of Severi, and the equations defining projective varieties, J. Amer. Math. Soc. 4 (1991), 587–602.

[Ber78]    G. M. Bergman, The diamond lemma for ring theory, Adv. in Math. 29 (1978), 178–218.

[BM88]    Dave Bayer and Ian Morrison, Standard bases and geometric invariant theory I. Initial ideals and state polytopes, J. Symb. Comput. 6 (1988), no. 2–3, 209–217, reprinted in [Rob89].

[BM91]    E. Bierstone and P. Milman, A simple constructive proof of canonical resolution of singularities, Effective methods in algebraic geometry (Castiglioncello, 1990), Progr. Math., vol. 94, Birkhauser Boston, 1991, pp. 11–30.

[BCR91]    A. M. Bigatti, M. Caboara, and L. Robbiano, On the computation of Hilbert-Poincare series, Applicable Algebra in Engineering, Communications, and Computing 2 (1991), 21–33.

[Bri73]    J. Briancon, Weierstrass prepare a la Hironaka, Astérisque 7,8 (1973), 67–73.

[Bro87]    W. D. Brownawell, Bounds for the degrees in the Nullstellensatz, Ann. of Math. (2) 126 (1987), 577–591.

[BS87a]    Dave Bayer and Mike Stillman, A criterion for detecting m-regularity, Invent. Math. 87 (1987), 1–11.

[BS87b]    Dave Bayer and Mike Stillman, A theorem on refining division orders by the reverse lexicographic order, Duke Math. J. 55 (1987), no. 2, 321–328.

[BS88]    Dave Bayer and Mike Stillman, On the complexity of computing syzygies, J. Symb. Comput. 6 (1988), 135–147.

[BS92a]    Dave Bayer and Mike Stillman, Macaulay: A system for computation in algebraic geometry and commutative algebra, 1982–1992, computer software available via anonymous ftp from zariski.harvard.edu.

[BS92b]    Dave Bayer and Mike Stillman, Computation of Hilbert functions, J. Symb. Comput. 6 (1992), 31-50.

[Buc65]    B. Buchberger, Ph.D. thesis, Univ. Innsbrück, 1965.

[Buc76]    B. Buchberger, A theoretical basis for the reduction of polynomials to canonical forms, ACM SIGSAM Bull. 39 (1976), 19–29.

[Buc79]     B. Buchberger, A criterion for detecting unnecessary reductions in the construction of Gröbner bases, Symbolic and Algebraic Computation (Proceedings of EUROSAM 79), Lecture Notes in Computer Science, vol. 72, Springer-Verlag, 1979, pp. 3–21.

[Can89]    J. Canny, Generalized characteristic polynomials, Symbolic and Algebraic Computation (Proceedings of ISSAC 88), Lecture Notes in Computer Science, vol. 358, Springer-Verlag, 1989, pp. 293–299.

[CG83]    A. L. Chistov and D. Yu. Grigoriev, Subexponential-time solving systems of algebraic equations I, II, Steklov Mathematical Institute, Leningrad department, LOMI Preprints E-9-93, 0E-10-c83, 1983.

[Cha91]    Marc Chardin, Un algorithme pour le calcu des résultants, Effective methods in algebraic geometry (Castiglioncello, 1990), Progr. Math., vol. 94, Birkhauser Boston, 1991, pp. 47–62.

[DD84]    C. Dicrescenzo and D. Duval, Computations on curves, Lecture Notes in Computer Science, vol. 174, Springer-Verlag, 1984.

[EG84]    David Eisenbud and Shiro Goto, Linear free resolutions and minimal multiplicity, J. Algebra 88 (1984), no. 1, 89–133.

[EHV92]    David Eisenbud, Craig Huneke, and Wolmer Vasconcelos, Direct methods for primary decomposition, Invent. Math. (1992), to appear.

[Eis92]    David Eisenbud, Commutative algebra with a view toward algebraic geometry, 1992, in preparation.

[Gal74]    A. Galligo, A propos du theoreme de preparation de Weierstrass, Fonctions de Plusieurs Variables Complexes, Lecture Notes in Math., vol. 409, Springer-Verlag, 1974, pp. 543–579.

[Gal79]    A. Galligo, Theoreme de division et stabilite en geometrie analytique locale, Ann. Inst. Fourier (Grenoble) 29 (1979), 107–184.

[GH91]    Marc Giusti and Joos Heintz, Algorithmes—disons rapides—pour la decomposition d’une variete algebrique en composantes irreductibles et equidimensionnelles [“Fast” algorithms for the decomposition of an algebraic variety into irreducible and equidimensional components], Effective methods in algebraic geometry (Castiglioncello, 1990), Progr. Math., vol. 94, Birkhauser Boston, 1991, pp. 169–194.

[Giu84]    Marc Giusti, Some effectivity problems in polynomial ideal theory, EUROSAM 84), Lecture Notes in Computer Science, vol. 204, Springer-Verlag, 1984, pp. 159–171.

[GLP83]    L. Gruson, R. Lazarsfeld, and C. Peskine, On a theorem of Castelnuovo, and the equations defining space curves, Invent. Math. 72 (1983), 491–506.

[GTZ88]    P. Gianni, B. Trager, and G. Zacharias, Gröbner bases and primary decomposition of polynomial ideals, J. Symb. Comput. 6 (1988), no. 2–3, 149–167, reprinted in [Rob89].

[Her26]    Grete Hermann, Die Frage der endlich vielen Schritte in der Theorie der Polynomideale, Math. Ann. 95 (1926), 736–788.

[Hir64]    H. Hironaka, Resolution of singularities of an algebraic variety over a field of characteristic zero: I, II, Ann. of Math. (2) 79 (1964), 109–326.

[Kol88]    János Kollár, Sharp effective Nullstellensatz, J. Amer. Math. Soc. 1 (1988), no. 4, 963–975.

[Lak91]    Y. N. Lakshman, A simple exponential bound on the complexity of computing gröbner bases of zero-dimensional ideals, Effective methods in algebraic geometry (Castiglioncello, 1990), Progr. Math., vol. 94, Birkhauser Boston, 1991, pp. 227–234.

[Laz87]    Robert Lazarsfeld, A sharp Castelnuovo bound for smooth surfaces, Duke Math. J. 55 (1987), 423–429.

[Len92]    H. W. Lenstra, Jr., Algorithms in algebraic number theory, Bull. Amer. Math. Soc. (N.S.) 26 (1992), no. 2, 211–244.

[Lip84]    Joseph Lipman, Dualizing sheaves, differentials and residues on algebraic varieties, Astérisque, vol. 117, 1984.

[LL91]    Y. N. Lakshman and D. Lazard, On the complexity of zero-dimensional algebraic systems, Effective methods in algebraic geometry (Castiglioncello, 1990), Progr. Math., vol. 94, Birkhauser Boston, 1991, pp. 217–225.

[LLL82]    A. K. Lenstra, H. W. Lenstra, Jr., and L. Lovász, Factoring polynomials with rational coefficients, Math. Ann. 261 (1982), 515–534.

[Mac27]    F. S. Macaulay, Some properties of enumeration in the theory of modular systems, Proc. London Math. Soc. 26 (1927), 531–555.

[MM82]    Ernst W. Mayr and Albert R. Meyer, The complexity of the word problem for commutative semigroups and polynomial ideals, Adv. in Math. 46 (1982), 305–329.

[MM83]    H. Michael Möller and Ferdinando Mora, Upper and lower bounds for the degree of Gröbner bases, Computer Algebra (EUROCAL 83), Lecture Notes in Computer Science, vol. 162, Springer-Verlag, 1983, pp. 157–167.

[MM86]    H. Michael Möller and Ferdinando Mora, New constructive methods in classical ideal theory, J. Algebra 100 (1986), no. 1, 138–178.

[MR88]    T. Mora and L. Robbiano, The Gröbner fan of an ideal, J. Symb. Comput. 6 (1988), no. 2–3, 183–208, reprinted in [Rob89].

[Mum66]    David Mumford, Lectures on curves on an algebraic surface, Princeton University Press, Princeton, New Jersey, 1966.

[Mum70a]   David Mumford, Abelian varieties, Oxford University Press, Oxford, 1970.

[Mum70b]   David Mumford, Varieties defined by quadratic equations, Questions on Algebraic Varieties, Centro Internationale Matematica Estivo, Cremonese, Rome, 1970, pp. 29–100.

[MW83]    D. W. Masser and G. Wüstholz, Fields of large transcendence degree generated by values of elliptic functions, Invent. Math. 72 (1983), 407–464.

[Pin86]    Henry C. Pinkham, A Castelnuovo bound for smooth surfaces, Invent. Math. 83 (1986), 491–506.

[Ran90]    Ziv Ran, Local differential geometry and generic projections of threefolds, J. Differential Geom. 32 (1990), 131–137.

[Rav90]    M. S. Ravi, Regularity of ideals and their radicals, Manuscripta Math. 68 (1990), 77–87.

[Ric74]    F. Richman, Constructive aspects of Noetherian rings, Proc. Amer. Math. Soc. 44 (1974), 436–441.

[Rob85]    L. Robbiano, Term orderings on the polynomial ring, Proceedings of EUROCAL ’85 (Linz), Lecture Notes in Computer Science, vol. 204, Springer-Verlag, 1985, pp. 513–517.

[Rob89]    Lorenzo Robbiano (ed.), Computational aspects of commutative algebra, Academic Press, 1989, ISBN 0-12-589590-9.

[Sch80]    Frank-Olaf Schreyer, Die Berechnung von Syzygien mit dem verallgemeinerten Weierstrass’schen Divisionssatz, Diplomarbeit am Fachbereich Mathematik der Universität Hamburg, 1980.

[Sch91]    Frank-Olaf Schreyer, A standard basis approach to syzygies of canonical curves, J. Reine Angew. Math. 421 (1991), 83–123.

[Ser55]    J.-P. Serre, Faisceaux algebrique coherents, Ann. of Math. (2) 61 (1955), 197–278.

[Spe77]    D. Spear, A constructive approach to commutative ring theory, Proceedings of the 1977 MACSYMA Users’ Conference, NASA CP-2012, 1977, pp. 369–376.

[Tei89]    Jeremy Teitelbaum, On the computational complexity of the resolution of plane curve singularities, Symbolic and algebraic computation (Rome, 1988), Lecture Notes in Computer Science, vol. 358, Springer, 1989, pp. 285–292.

[Tei90]    Jeremy Teitelbaum, The computational complexity of the resolution of plane curve singularities, Math. Comp. 54 (1990), no. 190, 797–837.

[Tri78]    W. Trinks, Über B. Buchberger’s Verfahren, Systeme algebraischer Gleichungen zu lösen, J. Number Theory 10 (1978), 475–488.

[Zac78]    G. Zacharias, Bachelor’s thesis, Mass. Inst. of Technology, 1978.

[Zar69]    Oscar Zariski, An introduction to the theory of algebraic surfaces, Lecture Notes in Math., vol. 83, Springer-Verlag, 1969.

[ZS76]    Oscar Zariski and Pierre Samuel, Commutative algebra, Volumes I, II, Graduate texts in mathematics, vol. 28–29, Springer-Verlag, 1975–1976.