arX

iv:1

411.

5724

v1 [

mat

h.A

G]

20 N

ov 2

014

Tate Resolutionsfor Products of Projective Spaces

for Ngo Viet Trung on the occasion of his sixtieth birthday

David Eisenbud, Daniel Erman, and Frank-Olaf Schreyer∗

Abstract

We describe the Tate resolution of a coherent sheaf or complex of co-herent sheaves on a product of projective spaces. Such a resolution makesexplicit all the cohomology of all twists of the sheaf, including, for example,the multigraded module of twisted global sections, and alsothe Beilinsonmonads of all twists. Although the Tate resolution is highlyinfinite, any fi-nite number of components can be computed efficiently, starting either froma Beilinson monad or from a multigraded module.

Introduction

A complex of coherent sheavesF on projective space may be specified in finiteterms by giving a complex of graded modulesM, or by giving a Beilinson monadB, that is, a finite complex written in terms of a strong exceptional sequence of thevector bundles∧iU , whereU is the universal rankn sub-bundle. The complexMis a convenient way of simultaneously specifying the complexes of global sectionsH0(F(d)) of twistsF(d) for all sufficiently larged, while the Beilinson monad is

∗This paper reports on work started during the Commutative Algebra Program, 2012-13, atthe Mathematical Sciences Research Institute in Berkeley,and continued at the MathematischesForschungsinstitut Oberwolfach. We are grateful to these institutes for providing a beautiful andexciting environment for this work. The first and second authors are grateful to the NationalScience Foundation, and the second and third author are grateful to the Simons Foundation forpartial support during this period. This material is based upon work supported by the NationalScience Foundation under Grant No. 0932078 000, while the authors were in residence at theMathematical Science Research Institute in Berkeley, California in 2012–2013.

1

a convenient way of specifying the hypercohomology ofF itself; but the Beilinsonmonads of twistsF(d) generally look quite different. TheTate resolutionof Fis a way of packaging all the cohomology spaces and the Beilinson monads ofall twists ofF simultaneously. It is a doubly infinite exact complex of finitelygenerated free modules over the exterior algebraE that is the Koszul dual of thehomogeneous coordinate ring of projective space. Any finitenumber of termscan be computed efficiently, in terms ofE-free resolutions of finitely generatedE-modules, from either a module of twisted global sections ora Beilinson monadfor F . Tate resolutions in this case were treated in [EFS03] and [ESW03]. Inmany cases they yield the fastest algorithms for computing cohomology.

In this paper we will provide an analogous, efficiently computable, construc-tion of a Tate resolution for finite complexes of coherent sheaves on products ofprojective spaces. A new feature, which makes this case muchmore difficult thanthe case of a single projective space, is that there are no finitely generated mod-ules among the terms of the Tate resolution. Despite this, wecan use the Tateresolution to effectively compute the monads and (hyper)cohomology of any fi-nite number of twists ofF in terms of free resolutions of certain finitely generatedmodules over an appropriate exterior algebra.

To state the main results we introduce some notation:Let P = Pn1 × · · · × Pnt = P(W1) × · · · × P(Wt) be a product oft pro-

jective spaces over an arbitrary fieldK. SetVi = W ∗i andV = ⊕iVi. Let E

be theZt-graded exterior algebra onV , where elements ofVi ⊂ E have degree(0, . . . , 0,−1, 0, . . . , 0) with −1 in thei-th place.

Let ωE be the freeE-module of rank 1 with generator in multidegree(n1 +1, . . . , nt + 1); this module has socle in multidegree0 and is the injective hullof the residue fieldK. We will generally write free complexes ofE-modules assums of twists ofωE. We haveωE = HomK(E,K), whereK is regarded as a1-dimensional vector space concentrated in degree 0.

A Zt-graded complex ofE-modules is calledlocally finite if the sum of thegraded components of each multidegree is finite dimensional.

LetUk = ker(H0(Pnk ,O(1))⊗O → O(1)) denote the tautological subbundleonPnk of ranknk. Fora ∈ Zt we set

Ua := ⊠tk=1Λ

akUk = π∗1Λa1U1 ⊗ · · · ⊗ π∗tΛ

atUt

the tensor product of the pullbacks toP of exterior powers of theUk. Of courseUa is nonzero if only if0 ≤ a ≤ n = (n1, . . . , nt), where the partial order onmulti-indices is termwise. TheUa form a full strong exceptional collectionfor

2

the derived categoryDb(P) [Huy06, Def. 8.31]. In particular every sheaf canbe expressed as the homology of a complex whose terms are direct sums of thesheavesUa, called aBeilinson monadfor F .

Consider the additive functor on the category of direct sumsof finitely gener-ated free graded rightE-modules defined on objects by

U : ωE(a) 7→ Ua.

We defineU on morphisms by using the identifications

HomE(ω(a), ω(b)) = Eb−a = HomP(Ua, U b).

This identifies the morphism induced by multiplication by anelement

e ∈ Eb−a = ⊗tk=1Λ

ak−bkVk

with a morphism of sheaves

U

(ωE(a)

e∧−✲ ωE(b))=

(Ua e¬✲ U b

),

wheree¬ is the map induced by the contraction operator, which we write in thesame way,

⊗tk=1Λ

akWk ⊗K Oe¬✲ ⊗t

k=1 ΛbkWk ⊗K O.

See for example [EFS03] for the caset = 1.If T is a multigraded complex of freeE-modules,I ( {1, . . . , t} is a proper

subset, andc = (ci)i∈I are integers, then theI-th strand ofT throughc is thesubquotient complex obtained fromT by taking all the free summands of termsof T of the formωE(a) whereai = ci for all i ∈ I. When all theci are zero, wespeak simply of theI-th strand ofT . Thus for exampleT itself is the∅-strand.

We will say that a locally finiteZt-graded complex of freeE-modules is aTate Resolutionif, for every multi-indexc, all strands ofT throughc are exact. AcomplexT of freeE-modules is called minimal ifT ⊗E K has zero differential.Every complex of freeE-modules is isomorphic to the direct sum of a minimalcomplex and a split exact complex.

Theorem 0.1. For any sheafF on P there is a unique minimal Tate resolutionT(F) such thatU

(T(F)(c)[|c|]

)is a Beilinson monad forF(c) for everyc ∈ Zt.

Moreover,T(F)d = ⊕a∈Zt HomK(E,Hd−|a|(P,F(a))),

where the cohomologyHd−|a|(P,F(a)) is regarded as a vector space concen-trated in degreed− |a|.

3

The second statement follows from the first using the well-known result ofTheorem2.1. The construction of the Tate resolution is given in Section1, andthe proofs of its properties are given in Sections2 and3.

Example 0.2.Consider the case wheret = 1 andP = P1. The Tate resolution ofOP1 onP1 is a complex of the form

T := T(OP1) = · · · ✲ ω2E(3)

(

e0 e1)

✲ ωE(2)

(

e0e1)

✲ ωE

e0e1

✲ ω2E(−1) ✲ · · · .

Every term in the resolutionT(F) is finitely generated in this case; indeed, this istrue whenevert = 1.

By contrast, if we taket > 1 then each term in the Tate resolution ofanynon-zero sheaf is infinite. For instance, in the caset = 2 with P = P1 × P1 andF = OP1×P1 = OP1 ⊠OP1 , thenT(F) = T ⊠ T , a complex in whicheverytermis infinite. By the formula in Theorem0.1we have, for example,

T(F)0 = ⊕(p,q)∈Z2 HomK

(E,H0−p−q (P,F(p, q))

)

= HomK

((E,H0(P,F)

)⊕

⊕

p∈Z

HomK

((E,H1(P,F(p,−p− 1))

)

= H0(P,F)⊗ ωE ⊕⊕

p∈Z

H1 (P,F(p,−p− 1))⊗ ωE(−p, p+ 1)

= ωE ⊕⊕

p∈Z

ωE(−p, p + 1)p(p+1)

Other examples are given in Section4.

SinceT(F) is locally finite we can form, for any finite intervalA in Zt, thefinite subquotient complexT ′ of T(F) with terms

T ′d ∼=⊕

a∈A

HomK(E,Hd−|a|(F(a)).

We give an algorithm, implemented in Macaulay2, for computing any such finitesubquotient complex [M2-Tate]. The algorithm makes use of free resolutions overE, and can be executed starting either from a multigraded complex of modulesrepresenting the global sections of high twists ofF or from a Beilinson monad forF . In particular, our method computes any finite number of cohomology groups

4

of a coherent sheaf onPn1 × · · · × Pnt without passing to a Segre embedding, aprocess that would introduce a much larger number of variables.

We can use the Tate resolution ofF to compute the direct image complexalong the projectionπJ : P → PJ :=

∏j∈J P

nj for any subsetJ ⊂ {1, . . . , t}.Let I be the complement ofJ . The differential on theI-th strand ofT(F) is givenby matrices with entries inEJ := ∧ ⊕j∈J Vj. Thus theI-th strand ofT(F) hasthe formTJ ⊗ ωEI

, whereTJ is a complex overEJ . Because a strand of a strandis a strand,TJ is a Tate resolution onPJ .

Corollary 0.3. With notation as above,U(TJ) is a Beilinson monad forRπJ∗F .

SinceU(T(F)) is quasi-isomorphic toF , we see thatU(T(F(c))) must bequasi-isomorphic toU(F)(c). The formula forT(F(c)) above suggests the fol-lowing:

Theorem 0.4(Twist and shift formula). LetT be a Tate resolution. ThenU(T (c))andU(T )(c)[−c] are quasi-isomorphic. In particular, for any non-empty subsetJ ⊂ {1, . . . t} with complementI, theI-th strand ofT throughc computes thedirect imageRπJ∗

(U(T )(c)

)alongπJ : P→ PJ .

Example 0.5. The quasi-isomorphismU(T(F(c))) ≃ U(F)(c) above is gener-ally not an isomorphism. For exampleU(T(OP1×P1)) is the complex

0→ OP1×P1 → 0.

By contrast,U(T(OP1×P1(3,−2))) is

0→ (U1 ⊠ U2)⊕6 → (U1 ⊠O)

⊕3 ⊕ (O ⊠ U2)⊕8 → (O ⊠O)⊕4 → 0

which we may rewrite as

0→ O6(−1,−1)→ O3(−1, 0)⊕O8(0,−1)→ O4 → 0.

ThusU(T(OP1×P1))(3,−2) andU(T(OP1×P1)(3,−2)[−1]

)are only quasi-isomorphic.

Our methods can also be used to treat the hypercohomology of complexesof sheaves. We say that a complexU is aBeilinson representativefor a boundedcomplexF of sheaves ifF is quasi-isomorphic toU and each term ofU is a directsum of copies of the sheavesUa. Our construction ofT(F) whenF is a sheaf

5

generalizes immediately to the case whenF is a bounded complex of sheaves,and thenU(T(F)) is a Beilinson representative ofF . As above, it follows that

(1) T(F)d ∼=⊕

a∈Zt

HomK

(E,Hd−|a|(P,F(a))

).

whereHd−|a| denotes hypercohomology. For simplicity we will generallyfocus onthe case of sheaves, leaving some details of the case of complexes to the interestedreader.

The construction of the Tate resolution is given in Section1, and the con-nection with Beilinson monads as well as the proof of Theorem0.1 appears inSection2.

Associated to a Tate resolution are many exact complexes built from inducedsubquotient complexes—see Theorem3.3; one type, the “corner complexes” de-fined in Section3, are the key to our finitistic construction of Tate resolutions andalso to our proof of Corollary0.3and the general twist and shift formula stated inTheorem0.4.

Given a finite complexBW of freeE-modules we can apply the functorUto get a finite complex of sheaves, and then apply the functorT to get a doublyinfinite exact complexT overE. In Section6 we show how to go directly fromBW toE by a computation in terms of free and injective resolutions overE.

In Section7 we consider a broad generalization, to products of projectivespaces, of Horrocks criterion for the splitting of a vector bundle. We prove thatthe criterion holds under an additional hypothesis, which may not be necessary.

We are grateful to Mike Stillman and Dan Grayson, the authorsof the com-puter algebra system Macaulay2 [M2], without which we would not have dis-covered the results in this paper! We are also grateful to Christine Berkesch andFlorian Geiss for useful conversations.

1 Construction of the Tate resolution

As above,P = Pn1 × · · · × Pnt = P(W1)×· · ·×P(Wt) is a product oft projectivespaces over an arbitrary fieldK. We letVi = W ∗

i , W = ⊕iWi andV = ⊕iVi.We letS = Sym(W ) be the Cox ring ofP, with theZt grading where elementsin Wi ⊂ S have degree(0, . . . , 0, 1, 0, . . . , 0) with the1 in theith place. We alsolet E be the exterior algebra onV with the dual grading, i.e. elements ofVi ⊂ Ehave degree(0, . . . , 0,−1, 0, . . . , 0). We write{xi,j} and{ei,j} for dual bases ofWi andVi.

6

By [Cox95], there is a correspondence between finitely generated, multigradedS-modules and coherent sheaves onPn. More precisely, ifM is any finitely gen-erated, multigradedS-module, then there is a corresponding coherent sheafM onP, and every coherent sheaf arises in this way.

If N is aK-vector space then we regardHomK(E,N) as a rightE-module bythe formula(φe)(f) = φ(ef).

We will use the BGG correspondence, which we briefly recall.

Definition 1.1. Let M =∑

a∈Zt Ma be a multigradedS-module. Ford ∈ Z setMd =

∑a,|a|=dMa, the sum of the components in total degreed. Let R(M) be

the complex

R(M) : . . .→ HomK(E,Md)→ HomK(E,Md+1)→ . . .

whereHomK(E,Md) = ωE ⊗K Md = R(M)d is in cohomological degreed andthe differential is given by

φ 7→ {f 7→∑

xi,jφ(fei,j)}.

This is the same formula as in the singly graded case, but nowR(M) is a multi-graded complex of free rightE-modules. Similarly, ifP =

∑a∈Zt Pa is a multi-

graded rightE-module, we define a complex of freeS-modules,

L(P ) : . . .→ S ⊗K Pd → S ⊗K Pd−1 → . . .

with S ⊗K Pd = L(P )d in homological degreed andPd =∑

a,|a|=d Pa. Thedifferential is given by

s⊗ p 7→∑

sxi,j ⊗ pei,j.

Note that our convention isdeg ei,j = −1 = − deg xi,j .The functorL defines an equivalence between the categorygrmod(E) of

finitely generated multigradedE-modules and the categorylincplx(S) of finitelinear complexes overS. Indeed, suppose that

F : . . .∂✲ Fd

∂✲ Fd−1∂✲ . . .

is a linear complex withFd =∑

a:|a|=dBa ⊗K S(−a), where we think ofBa

as a vector space in degree0. Let P =∑

d∈Z

∑a:|a|=d Pa with Pa = Ba(−a).

We giveP the structure of a graded rightE-module by settingpei,j = pi,j if

7

∂(p ⊗ 1) =∑

pi,j ⊗ xi,j. Because∂2 = 0, this action extends to an action ofthe exterior algebraE, andF = L(P ). Similarly,R : grmod(S) → lincplx(E)is an equivalence between the category of multigradedS-module and the cate-gory of linear freeE-complexes. Sometimes it is more convenient to indexL(P )cohomologically. For that purpose we defineP d = P−d.

The functorsR andL extend naturally to functors

cplx(S)R

// cplx(E)L

oo

between the categories of complexes. We first defineL(P [k]) = L(P )[k] andR(M [k]) = R(M)[k], where, for any complexF andk ∈ Z we define the shiftedcomplexF [k] by F [k]d = F k+d. Thus the shift operator commutes with thefunctorsR andL. With this convention we defineR on a complex ofS-modulesby applyingR to each term, and taking the total complex of the resulting doublecomplex; we defineL on a complex ofE-modules similarly.

The shift operators also commute with the twist operators:M [k](a) = M(a)[k]for any graded module overS or P . However, the twist operator only commuteswith R andL up to a shift, as in the following Lemma. We can only shift a com-plex by an integer, not a multi-index so, to simplify notation, we will usually writeM [b] in place ofM [|b|] (recall that|b| denotes

∑i bi, not the absolute value).

Lemma 1.2. For any multigradedS-moduleM , any multigradedE-moduleP ,and anyb ∈ Zt, we have:

R(M(b)) = R(M)(b)[−b]

L(P (b)) = L(P )(b)[−b].

Proof. We regardMa as a vector space in degreea, so thatM(b)a = Mb+a(b).Thus

R(M(b))d =∑

a:|a|=d

HomK(E,Mb+a(b))

=∑

c:|c|=d+|b|

HomK(E,Mc)(b)

= (R(M)(b)[−b])d

The computation forL is similar.

A fundamental reciprocity result proven for the caset = 1 in [EFS03], Theo-rem 3.7, also holds in general:

8

Theorem 1.3(Reciprocity). LetM be a finitely generated gradedS-module, andP a finitely generated gradedE-module. ThenR(M) is an injective resolution ofP if and only ifL(P ) is a projective resolution ofM .

Proof. We may regardM andP also as singly graded modules for which theReciprocity Theorem [EFS03, Theorem 3.7] applies. The constructions respectthe finer grading.

Corollary 1.4. LetM be a finitely generated multigradedS-module.M≥c(c) hasa linear resolutionF if and only if R(M≥c) is acyclic. Moreover, if we writeF = L(P ) , thenH|c|(R(M≥c)) ∼= P (−c).

Proof. H0(R(M≥c(c))) ∼= P is equivalent toH|c|(R(M≥c)) ∼= P (−c).

We next recall the construction from [EFS03] of the Tate resolution of a sheafF onPn. Suppose thatF is the sheafification of theZ-gradedS-moduleM .

1. Choose an integerc sufficiently large that the moduleM≥c(c) has a linearresolution (this is equivalent toc ≥ regM), and so thatM contains nosubmodule of finite length (this is satisfied ifc > regM). It follows that

R(M≥c) : 0→ T c → T c+1 → . . .

is acyclic by Theorem1.3and is minimal because no element ofM is anni-hilated by every linear form ofS.

2. Attach toR(M≥c) a minimal free resolution. . . → T c−2 → T c−1 →Hc(R(M≥c))→ 0 to obtain a doubly infinite complex

T(F) : . . .→ T d−1 → T d → T d+1 → . . . .

We have

Theorem 1.5([EFS03] Thm 4.1). The complexT(F) depends only on the sheafF . Moreover,

T d =

n∑

i=0

HomK(E,Hi(F(d− i))).

9

Steps 1 and 2 work also in multigraded setting: Given aZt-gradedS-moduleM = ⊕a∈ZtMa and a multidegreec ∈ Zt we denote by

M≥c = ⊕a≥cMa

its truncation. Herea ≥ c stands for the componentwise partial order on multide-grees. We note that, for anyc, the sheavesM andM≥c are isomorphic [CLS11,Proposition 5.3.10].

We will show that, for a sufficiently large multidegreec, the complexR(M≥c)is acyclic and minimal. Thus we may attach a free resolution of H|c|(R(M≥c))to obtain a doubly infinite complex of (finitely generated) freeE-modules. How-ever it is no longer true that the resulting complex encodes all of the cohomologygroups of twists ofF = M . Instead, as we shall see, it encodes only the coho-mology groups of the twists that are≥ c and those that are≤ c− (1, . . . , 1); theseare only2 of the2t orthants ofZt.

To get a complex that encodes all cohomology groups we must pass to a re-stricted inverse limit. Fix a large multidegreeb and consider allc ≥ b. LetTailc(M) denote the projective resolution ofH|c|(R(M≥c)). It is easy to see thatthe complexesR(M≥c) form a directed system of subcomplexesR(M≥c) →R(M≥b). We will show that the tails form an inverse system of complexes viaa sequence of epimorphismsTailc(M) ։ Tailb(M) defined whenc > b≫ 0.

Definition 1.6. The Tate resolution ofF = M on the productP of t projectivespaces is the restricted inverse limit

T(F) =(lim∞←c

′Tailc(M))[−t + 1].

where “restricted” means that, thinking oflim∞←cTailc(M) as a subcomplex of∏c Tailc(M), we take only those sequences of elements of bounded degree.

To make sense of this definition we must define the mapsTailc(M) ։ Tailb(M).The first step is to prove some properties of high truncationsof multigradedmodules that are standard for singly graded modules. Recallthat for anyJ ⊂{1, . . . t}, we denote the projectionP→ PJ :=

∏j∈J P

nj by πJ .

Proposition 1.7. Let M be a finitely generated multigradedS-module and letF = M be the coherent sheaf onP = Pn1 × · · · × Pnt represented byM . Thereexists a multidegreeb such that, for any multidegreec ≥ b:

10

1. The truncated twisted moduleM≥c(c) is generated in degree0 and has alinear resolution, i.e. the free modulesFk = ⊕aS

βk,a(−a) in the minimalfree resolution

(0←M≥c(c)←)F0 ← F1 ← · · ·

ofM≥c(c) satisfiesβk,a 6= 0 only if k = |a| anda ≥ 0,

2. Mc = H0(P,F(c)) andHp(P,F(c)) = 0 for p > 0. More generally, foranyJ ⊂ {1, . . . t} andp > 0, we haveRpπJ ∗F(c) = 0.

Definition 1.8. We call a multidegreeb sufficiently positive forM if it satisfies theconditions of Proposition1.7.

Proof. Consider first the caseM = S. WriteO = OP for the structure sheaf. Weclaim that0 is sufficiently positive forS.

Denote bymi = 〈Wi〉 ⊂ S the minimal primes of the irrelevant ideal in theCox ring. For everyc ≥ 0 the ideal

S≥c =⋂

i

mcii =

∏

i

mcii

has a resolution which is the tensor product of the resolutions of the idealsmcii .

Since (up to twist) these resolution are linear, the tensor product resolution ofS≥c(c) is linear, i.e. the first assertion holds forS and everyc ≥ 0. Furthermore,we haveS =

∑aH

0(O(a1, . . . , at)) and∑

a≥−n Hp(O(a1, . . . , at)) = 0 for p ≥

1 by the Kunneth formula. Finally for a pair of complementaryindex setsJ ∪ I ={1, . . . , t} and multidegreec = (cJ , cI) ≥ 0, the truncated module of globalsections(πJ)∗O(c) = H0(PI ,O(cI))⊗O(cJ) has a linear resolution sincecJ ≥ 0.

We now consider an arbitrary finitely generated multigradedS-moduleM . Bythe Hilbert syzygy theorem,M has a finite free multi-homogeneous resolution

0←M ← G0 ← G1 ← · · · ← GN ← 0

with N ≤∑

(ni+1). WriteGk = ⊕S(−a)βk,a and setbi = max{ai | ∃βk,a 6= 0}.

We claim thatb = (b1, . . . bt) is sufficiently positive forM . Let c ≥ b. Each(Gk)≥c(c) has a linear resolution because, for every summandS(−a), the trunca-tion

S(−a)≥c(c) = mc1−a11 ∩ . . . ∩m

ct−att (c− a)

has a linear resolution. An iterated mapping cone (see Section5) over the complex(G•)≥c(c) yields a non-minimal resolutionF ′ with graded Betti numbersβk,a(F

′)

11

non-zero only if|a| ≤ k. SinceM≥c(c) is generated in degree0 its minimalresolutionF satisfiesβk,a(F ) = 0 for |a| < k. Thus the first assertion holds forM≥c(c), because we can obtainF fromF ′ by canceling trivial subcomplexes.

For the second assertion, we note that in the sheafified complex

0← F(c)← G0(c)← G1(c)← . . .

we haveHp(Pn1 × · · · × Pnt , Gk(c)) = 0 for p ≥ 1. Thus the complex is exacton global sections,

H0(F(c)) = coker (H0(G1(c))→ H0(G0(c))) = coker ((G1)c → (G0)c) = Mc,

and the higher cohomology ofF(c) vanishes. A similar argument shows thatRpπJF(c) = 0 for p > 0.

Proposition 1.9. If c is sufficiently positive forM then0 is sufficiently positive forΓ≥0(πJ ∗(F(c))).

Proof. If c is sufficiently positive forM andJ ⊂ {1, . . . , t} then

Γ≥0(πJ∗(F(c))) =∑

a≥0, aj=0∀j /∈J

Mc+a,

where we considerMc+a to be in degreea.For property 1 of the definition, we note that the subquotientcomplex of the

free resolution ofM , consisting of elements that have degrees of the formc + awith a ≥ 0 andaj = 0 for all j /∈ J , is a linear free resolution of the module∑

a≥0, aj=0∀j /∈J Mc+a over the homogeneous coordinate ring ofPJ .For property 2, note that fora ≥ 0 with aj = 0 ∀j /∈ J we have

H0(PJ , πJ∗(F(c))(a)) = H0(P,F(c+ a)) = Mc+a

andHp(PJ , (πJ)∗(F(c))(a)) = Hp(P,F(c+ a)) = 0,

since the direct imagesRp(πJ)∗(F(c)) vanish forp > 0. Also, for I ⊂ J and thefurther projectionπI⊂J : PJ → PI we have

(πI⊂J)∗(πJ)∗(F(c))(a)) = (πI)∗(F(c+ a))

andRp(πI⊂J)∗((πJ)∗(F(c))(a)) = Rp(πI)∗(F(c+ a)) = 0

for all a ≥ 0 havingaj = 0 for j /∈ J .

12

Notation 1.10(Vectors of Ones). Set

1t := (1, . . . , 1) ∈ Zt

0i1t−i := (0 . . . 0, 1, . . . 1) ∈ Zi × Zt−i.

We denote the standard basis vectors by1i = 0i−110t−i = (0 . . . , 0, 1, 0 . . .0) ∈Zt.

Notation 1.11 (Degree restrictions). Let M =∑

a∈Zt Ma be a multigradedS-module,J∪I = {1, . . . t} complementary subsets of the index set,c = (cI , cJ) ∈Zt a multidegree broken up into two parts, andd ∈ Z. We set:

M≥c =∑

a: a≥c

Ma , McI ,≥cJ =∑

a=(aI,aJ )aI=cI, aJ≥cJ

Ma and M>cI ,≥cJ =∑

a=(aI ,aJ )a≥c, aI 6=cI

Ma

M≤c =∑

a: a≤c

Ma andM≥d =∑

a: |a|≥d

Ma.

DecomposingS = SI⊗SJ accordingly we may regardMcI ,≥cJ∼= M≥c/M>cI ,≥cJ

either as anS-module or anSJ -module. LetE = EI ⊗ EJ be the correspondingdecomposition of the exterior algebra and

RJ : grmod(SJ)→ lincplx(EJ)

the correspondingR-functor. Note that

R(McI ,≥cJ )∼= HomK(EI , K(cI))⊗RJ(McI ,≥cJ ).

is a flat extension of scalars, soR(McI ,≥cJ ) is exact iffRJ(McI ,≥cJ ) is exact.

Proposition 1.12. Let {j} ∪ I = {1, . . . , t} be a disjoint decomposition with asingleton. Ifb = (bj , bI) andc = (bj + 1, bI) are multidegrees such thatM≥b(b)andM≥c(c) have linear resolutions, thenR(M≥b), R(M≥c) andR(Mbj ,≥bI ) areacyclic andMbj ,≥bI (b) has a linear resolution.

Proof. SinceM≥c ⊂ M≥b is a submodule with quotientMbj ,≥bI we have a shortexact sequences of complexes

0→ R(M≥c)→ R(M≥b)→ R(Mbj ,≥bI )→ 0

By the Reciprocity Theorem the only non-zero cohomology groups in the longexact sequence are

0→ H|b|(R(M≥b))→ H|b|(R(Mbj ,≥bI ))→ H|b|+1(R(M≥c))→ 0.

SoMbj ,≥bI (b) has a linear resolution as well.

13

Theorem 1.13.Let c be a multidegree that is sufficiently positive forM . There isa naturalsplice mapϕc : R(Mc) → R(Mc+1t) that is a surjection to the moduleof cyclesP of homological degree|c|+ t of the complexR(M≥c+1t). Moreover, ifc − k1t is sufficiently positive forM for somek ≥ 0, thenR((M≤c)≥|c|−k) is thebeginning of a projective resolution ofP .

Proof. We will prove the first statement by induction on the numbert of factorsin the productP. If t = 1 then the statement is equivalent to the acyclicity ofR(M≥c), which follows from the Reciprocity Theorem. In this case the splicemapϕc is simply the differential.

Now consider the caset > 1. The moduleP is part of the module of cyclesin the larger complexR(M≥c+011t−1). By the exactness of this complex,P iscontained in the image ofR(Mc+011t−1); and thus is contained in the image of themoduleP ′ of cycles of homological degree|c|+ t− 1 in the subquotient complexR(Mc1,≥(c2+1,...,ct+1)), sinceP is congruent to 0 in this subquotient.

By our hypothesisMc1,≥(c2+1,...,ct+1) is a truncated section module on a productof t− 1 factors. By induction, there is a splice mapϕ′c : R(Mc)→ R(Mc+011t−1)that is a surjection ontoP ′. Letϕc be the composition of this with the componentof the differential that mapsR(Mc+011t−1) to R(Mc+1t).

By construction, the image of the mapϕc containsP . Thus to prove the firststatement of the Theorem it suffices to show that the image ofϕc composes to0 with the first differential ofR(M≥c+1t). Becauseϕc is defined as the compo-sition of ofϕ′c with the component of the differential that mapsR(Mc+011t−1) toR(Mc+1t), it is at least clear that theϕc composes to 0 with the component of thedifferential that goes fromR(Mc+1t) to R(Mc+211t−1) is 0.

Unraveling the induction used to defineϕc, we see thatϕc is the composition

R(Mc)→ R(Mc+0t−111)→ . . .→ R(Mc+021t−2)→ R(Mc+011t−1)→ R(Mc+1t)

of t maps that are components of differentials ofR(M≥c). If we permute thetfactors ofP, then we would define the mapR(Mc) → R(Mc+1t) as a differentcomposition. However, since all squares inR(M≥c) anti-commute, the result-ing map would only differ by the signature of the permutation. In particular wesee that the composition ofϕc with every component of the first differential ofR(M≥c+1t) is zero as required.

We will not give a proof of the second statement of the Theoremhere, sincethe result follows from Theorem3.3 proven below; in any case, it can be provenby a straightforward diagram chase.

14

↑ ↑ ↑ ↑

RM(1,3) → RM(2,3) → RM(3,3) → RM(4,3) →

↑ ↑ ↑ ↑

RM(1,2) → RM(2,2) → RM(3,2) → RM(4,2) →

↑ ↑ ↑ ↑

RM(1,1) → RM(2,1) → RM(3,1) → RM(4,1) →

ր

RM(−2,0) → RM(−1,0) → RM(0,0)

↑ ↑

RM(−1,−1) → RM(0,−1)

↑

RM(0,−2)

Figure 1: Nonzero terms of the complexR((M≤0)≥−2)→ R(M≥1t) in the case wheret = 2 and whereM is a module for whichb = (−2,−2) is sufficiently positive.

Corollary 1.14. Suppose the multidegreeb is sufficiently positive for the multi-gradedS-moduleM and letc ≥ b + 1t. If the graded betti numberβ|a|,a(M) isnonzero, then0 ≤ ai ≤ ni. Moreover,β|n|,n 6= 0. In particulardepthM≥c = t.

Proof. LetP be theE-module of cycles of homological degree|c| in R(M≥c). Bythe Reciprocity Theorem, the minimal free resolution ofM≥c(c) is the complexL(P (c)) = L(P )(c)[−c], so it suffices to examine the Hilbert function ofP .

The socle ofP coincides with the socle ofR(Mc), which is nonzero in degreec, while the generators ofP have the same degrees as the generators ofR(Mc−1t),that is,c− n, and the first statement follows. The last statement followsfrom thefirst by the Auslander-Buchsbaum formula.

LetM be gradedS-module and letb a sufficiently positive multidegree forM .For eachc ≥ b + 1t we consider the module of cyclesP = H|c|(R(M≥c) as anE-module of cohomological degree|c|. Let

Tailc(M)→ P

be a minimal free resolution ofP as anE-module. By Theorem1.13the first termof Tailc(M) is R(Mc−1t). By definition, this module has socle in degreec− 1t.

Now suppose thatI is the complement of the singleton{j} ⊂ {1, . . . , n}and writec + 1j, recalling that1j was defined in Notation1.10. The short exact

15

sequence of modules

0→M≥c+1j → M≥c →Mcj ,≥cI → 0

gives rise to a short exact sequence of minimal injective resolutions, from whichwe deduce a short exact sequence of modules of cycles:

0→ H|c|(R(M≥c))→ H|c|(R(Mcj ,≥cI))→ H|c|+1(R(M≥(c+1j)))→ 0.

Corresponding to the left hand map, we get a mapαc,j from the minimal free reso-lutionTailc(M) ofH|c|(R(M≥c)) to the minimal free resolution ofH|c|(R(Mcj ,≥cI))

whose mapping cone is a free resolution ofH|c|+1(R(M≥(c+1j))).

Proposition 1.15.The minimal free resolutionTailc+1j ofH|c|+1(R(M≥(c+1j))) isisomorphic to the result of canceling the first map in the mapping cone ofαc,j.Thus we may regardTailc(M) as a quotient complex ofTailc+1j(M).

Proof. It is immediate from the long exact sequence in homology thatthe homol-ogy of the mapping cone isH|c|+1(R(M≥(c+1j))), in homological degree 1. Sinceαc,j is an isomorphism on the terms of homological degree 0, we maycancel theseto obtain a free resolution ofH|c|+1(R(M≥(c+1j))).

It remains to prove minimality. The moduleMcj ,≥cI is annihilated byWj , sothe differential of the complexR(R(Mcj ,≥cI) does not involve any of the exteriorvariables inVj. Thus the moduleH|c|(R(Mcj ,≥cI )) is free over the tensor factorEj = ∧Vj of E, and its socle has degreecj in thej-th component.

It follows that the free modules in the minimal free resolution ofH|c|(R(Mcj ,≥cI ))all have socle with degreecj in thej-th component. On the other hand it followsfrom Theorem1.13that the free modules in the minimal free resolution resolutionTailc(M) of H|c|(R(M≥c)) have socle degree in thej-th component all≤ cj − 1.Thus the mapping cone is minimal as claimed.

If we choose a sequence of sufficiently positive multi-indices c(i) going toinfinity in each component, then Proposition1.15allows us to give the collection

{Tailc(i)(M) | c ≥ b}

the structure of a directed systemTailc(i+1)(M) ։ Tailc(i)(M). We define theTate resolution ofF = M as the restricted inverse limit

T(F) :=(lim∞←c

′Tailc(M))[1− t],

16

that is, as the complex generated by limit elements that are represented by se-quences of elements all of which are homogeneous of the same degree. The shift[1 − t] is necessary to adjust for the shift in the complex of Theorem1.13. ThecomplexT(F) depends, up to isomorphism, only onF = M .

In Corollary 3.5 we will show thatT(F) is in fact a Tate resolution in thesense defined in the introduction—that is, all its strands are exact. For now weprove a weaker property:

Proposition 1.16.LetF be a coherent sheaf onP.

1. For each multidegreea the space of homogeneous elementsT(F)a of mul-tidegreea is finite-dimensional.

2. For any multi-indexa we have

T(F(a)) ∼= T(F)(a)[−a].

3. T(F) exact and uniquely determined byF up to isomorphism.

Proof. 1. Since eachTailc(M) is a minimal resolution of a finitely generatedmodule, and the dimensiondimK E is finite as aK-vector space, the space ofhomogeneous elements(Tailc(M))a of degreea is finite dimensional. In the in-ductive construction of theT(F) the kernel of the mapTailc+1j M → TailcM ispart of the resolution of a submodule ofR(Mcj ,≥cI ), so the only multidegrees inwhich modules in the resolution are nonzero havej-th coordinate betweencj andcj + nj. Thus a given degree can appear in only finitely many kernels,and thus,for everyc, the dimensiondimK(Tailc(M))a stabilizes asc → ∞. ThusT(F)ais a finite dimensional for each multidegreea.

3. Since each(Tailc(M)) is acyclic, this stabilization shows that each(Tailc(M))ais exact, and it follows thatT(F)a is exact. If the gradedS-modulesN andMrepresent the same sheaf thenN≥c ∼= M≥c for c ≫ 0. SoR(N≥c) ∼= R(M≥c)andTailc(N) ∼= Tailc(M) as minimal free resolution over a graded ring. Finally(lim∞←c

′Tailc(M)) [1− t] ∼= (lim∞←c′Tailc(M)) [1− t].

2. From Lemma1.2 it follows thatT(F(a)) ∼= T(F)(a)[−a] for any multi-indexa.

2 Beilinson Monads

To prove the more precise statement about the graded pieces of the Tate resolutiongiven in Theorem0.1 we will use Beilinson monads. Recall that aBeilinson

17

monadfor a coherent sheafF onP is a finite complexB whose only homology isF in homological degree 0, and whose terms that are direct sumsof the sheavesUa. It is minimal if, under the identificationsHomP(U

a, U b) = Eb−a, all thenonzero components of the morphisms ofB have nonzero degree. The followingresult is well-known:

Theorem 2.1. If B is a minimal Beilinson monad for a coherent sheafF on P,then thed-th term is

Bd =⊕

a∈Zt

Hd−|a|(F(a))⊗ U−a

where (in this formula) we regardHd−|a|(F(a)) as a vector space in degree0. Inparticular,Bd can be nonzero only in the range−|n| ≤ d ≤ |n|.

Proof. Write Bd = ⊕0≤a≤nBda ⊗ U−a for vector spacesBd

a in degree zero. Thecollection{U−a| − n ≤ a ≤ 0} forms a strong full exceptional collection for thederived categoryDb(P) of coherent sheaves onP = Pn1 × · · · × Pnt (see [Huy06,Def. 8.31]) which is right orthogonal to the strong full exceptional collection{O(c)|0 ≤ c ≤ n} in the sense that, forc in this range,

HpRHom(O(c), U−a) = Hp(U−a(−c)) =

{K if c = −a andp = −|a|,

0 otherwise.

Hence the complexRHom(O(a),U(F)) is the sequence

. . .→ Bda → Bd+1

a → . . .

of vector spaces, whereBda sits in cohomological positiond−|a|, and all maps are

zero sinceB is minimal. SinceB is quasi isomorphic toF this complex coincideswith RHom(O(a),F) which is a sequence of vector spaces with terms

0→ H0(F(a))→ H1(F(a))→ . . .→ H |n|(F(a))→ 0

in cohomological position0, . . . , |n| and zero differential. Thus

Bda = Hd−|a|F(b)

as desired. Note that this group is possibly nonzero only if−|n| ≤ d ≤ |n|,because0 ≤ d− |a| ≤ |n| and0 ≤ a ≤ n.

18

We can now show thatU(F) := U(T(F)) is a Beilinson monad forF . Be-cause eachT(F)a is finite dimensional, the result of applyingU to the Tate reso-lution ofF is at least a bounded complex.

Theorem 2.2.U(F) is a minimal Beilinson monad for the sheafF .

Proof. As in the case of a single projective space, the result follows by applyingthe functorL to a bounded part of the Tate resolution.

We first consider the complexesL(ωE(a)). For anE-moduleQ =∑

aQa

we setQ≥1t :=∑

a≥1t Qa. If we writeQa = Ba(−a) with Ba a vector space in

degree0, then the sheafificationL(Q≥1t) of the complexL(Q≥1t) has the form

· · · → ⊕jB1t+1j ⊗O((−1)t − 1j)→ B1t ⊗O((−1)

t)→ 0

It is more convenient to index the complex cohomologically,so that this termbecomesL−t(Q≥1t). For example

L((ωE)≥1t) : . . .→ (⊗tk=1Wk)⊗O(−1, . . . ,−1)→ 0

is the tensor product oft truncated Koszul complexes

(πk)∗(0→ Λnk+1Wk ⊗O(−nk − 1)→ . . .→Wk ⊗O(−1)→ 0).

Since each of these complexes is a resolution ofO we see that

H−t(L((ωE)≥1t)) = O

and all other cohomology groups are zero. Similarly we find that the rightmostterm ofL((ωE(a))≥1t) is (⊗t

k=1Λak+1Wk)⊗O(−1, . . . ,−1) and the whole com-

plex is a tensor product of possibly truncated Koszul complexes. Thus the com-plexesL((ωE(a))≥1t) are acyclic with only cohomologyH−t(L((ωE(a))≥1t) =Ua, which is nonzero if and only if0 ≤ a ≤ n.

For any sheafF we can form the double complex

L((T(F))≥1t)[t].

What we have proven shows that the homology of this double complex with re-spect to the differential coming from the functorL, which we think of as thehorizontal homology, is the complexU(F).

Let M be a gradedS-module whose sheafification isF . The complexU(F)depends only on the summandsωE(a) of T(F) with 0 ≤ a ≤ n. Choose a

19

multidegreec that is sufficiently positive forM and large enough such that everysuch summand is contained inTailc(M). With this choice,L((Tailc(M))≤1t) is abounded double complex whose horizontal cohomology isU(F)[−t].

Recall thatTailc(M) is, by definition, a resolution of the module of cyclesP = H|c|(R(M≥c)), which has cohomological degree|c| − t. Thus the verticalhomology ofL((Tailc(M))≤1t) is the complex

L(P )[−t + |c|]

whose only homology isH−t(L(P )[−t+ |c|]) ∼= M≥c ∼= F .It follows that the only homology ofU(F) isH0(U(F)) ∼= F . More precisely,

the total complex ofL((Tailc(M))≥1t)[t] is quasi isomorphic to bothU(F) andL(P )[−|c|] via the natural maps

L((Tailc(M))≥1t)[t]

L(P )[−|c|]✛

U(F).

✲

Proof of Theorem0.1. The complexT(F) is minimal by construction, and wewill see that it is a Tate resolution in Corollary3.5below. Applying Theorem2.2and Theorem2.1 to the sheafF(b) we see that, if we writeT(F(b))d = ⊕aB

da ⊗

ω(−a) then for−n ≤ a ≤ 0 we must haveBda = Hd−|a|F(a + b). Moreover, by

Proposition1.16,T(F) = T(F(b))(−b)[−b]

for every multi-indexb. It follows thatU(T(F)(b)[b]) is a Beilinson monad forF(b). T(F) is uniquely determined byF up to isomorphism by Proposition1.16.It is also uniquely determined up to isomorphism by any of itsBeilinson monadsU(T(F)(b)[b]) by Corollary6.7below.

3 Exactness properties of the Tate resolution

We will next establish the exactness of the strands and othersubquotient com-plexes ofT(F), showing that it is indeed a Tate resolution as defined in the

20

Introduction. We begin by establishing notation. First, wecan restrict the no-tions already defined to any nonempty subsetJ of the indices{1, . . . , t}: We setWJ := ⊕j∈JWj, andSJ := SymWj We use similar notation forV andE. As inthe Introduction we writeπJ : P→ PJ =

∏j∈J P

nj .

We denote byωJ = EJ(− dimVJ) theEJ -injective hull ofK, and we writeUJ for the functor whose value onωJ(a) is⊗j∈J ∧

aj Uj , analogous toU.

Notation 3.1 (Strands, quadrant complexes, and region complexes). Let T be alocally finite complex of graded freeE-modules with termsT d =

∑a∈Zt Bd

a ⊗ωE(−a). For c ∈ Zt and three disjoint subsetsI, J,K ⊂ {1, . . . , t} we denote byTc(I, J,K) the subquotient complex with terms

Tc(I, J,K)d =∑

a∈Z

ai<ci for i∈I

ai=ci for i∈J

ai≥ci for i∈K

Bda ⊗ ωE(−a),

and we call this aregion complexof T . A strandof T , which was defined in theIntroduction, may be viewed as a region complex of the special form Tc(∅, J, ∅)whereJ ( {1, . . . , t} is a proper subset. Note thatT itself is the strand corre-sponding toJ = ∅.

If I ∪ J ∪K = {1, . . . , t} we callTc(I, J,K) a quadrant complex. A regioncomplex which is not a quadrant complex is called aproper region complex. IfT = T(F) is the Tate resolution of a sheaf we will see, that any proper regioncomplexTc(I, J,K) is exact.

To simplify the notation we sometimes write the quadrant complexes as

qTc,I := Tc(I, ∅, J)

whereJ is the complement ofI, andT≥c = Tc(∅, ∅, {1, . . . , t}) and T<c =Tc({1, . . . , t}, ∅, ∅) for the first and last quadrant complex.

Inclusions of regions give various short exact sequences ofcomplexes. Fori /∈ I ∪ J ∪K we have exact sequences

0→ Tc(I, J,K ∪ {i})→ Tc(I, J,K)→ Tc(I ∪ {i}, J,K)→ 0,

0→ Tc+1i(I, J,K ∪ {i})→ Tc(I, J,K ∪ {i})→ Tc(I, J ∪ {i}, K)→ 0,

0→ Tc(I, J ∪ {i}, K)→ Tc+1i(I ∪ {i}, J,K)→ Tc(I ∪ {i}, J,K)→ 0.

21

Notation 3.2 (Corner complexes). We define the map

ϕc : T<c[−t]→ T≥c

to be the composition

T<c[−t] = qTc,{1,...,t}[−t]→ . . .→ qTc,{1,...k}[−k]→ . . .→ qTc,∅ = aT≥c

of t morphisms deduced from the exact sequences

0→ qTc,{1,...k−1} → Tc({1, . . . , k − 1}, ∅, {k + 1, . . . , t})→ qTc,{1,...k} → 0.

This is a morphism of complexes since all mapsTc,k[−k] → Tc,k−1[−k + 1] aremorphisms of complexes. We define thecorner complexT�c as the mapping coneof ϕc.

Theorem 3.3.LetT be a locally finite complex of freeE-modules. The followingstatements are equivalent

1. Every strand ofT is exact.

2. Every proper region complex ofT is exact.

3. Every corner complexT�c is exact.

4. The corner complexesT�c are exact for every sufficiently largec.

5. The proper region complexesTc(I, ∅, ∅) are exact for every sufficiently largec.

Proof. By the definition of local finiteness, the complex of vectorspacesTb con-sisting of homogeneous elements of degreeb is a finite complex of finite dimen-sional vector spaces. Note thatT is exact if and only ifTb is exact for everyinternal degreeb.

1. ⇒ 2. For a fixedc, we must prove the exactness ofTc(I, J,K)b for all b.We note that, by hypothesis,Tc(∅, J, ∅) is exact for every proper subsetJ . Letus consider the caseI = {k} andK = ∅. If bk ≪ ck sufficiently large, thenwe haveTc({k}, J, ∅)b = Tc(∅, J, ∅)b, so the exactness ofTc({k}, J, ∅)b followsfrom the exactness ofTc(∅, J, ∅). For smallerck we use descending induction onck together with the exactness of the strandsTc(∅, {k}, ∅) and exact sequences

0→ Tc(∅, J ∪ {k}, ∅)b → Tc+1k({k}, J, ∅)b → Tc({k}, J, ∅)b → 0.

22

Note that the term on the left is also exact by hypothesis so long asJ ∪{k} is stilla proper subset, yielding the exactness of the term on the right.

The exactness of complexesTc(∅, J, {k}) can be argued similarly, using theexact sequence

0→ Tc(∅, J ∪ {k}, ∅)→ Tc(∅, J, {k})→ Tc+1k(∅, J, {k})→ 0

and an ascending induction onck.The general case follows by an induction on the size ofI ∪ J ∪K using the

exact sequences above Notation3.2. Note that this induction stops with#(I ∪J ∪K) = t − 1 because the complexTc(∅, {1, . . . , n}, ∅) is a bounded complexof E-modules, and thus is never exact unlessT is a split exact complex.

2. ⇒ 3. By Notation3.2, we have thatT�c is the mapping cone of the mapφc : T<c[−t] → T≥c. Note that, by definitionφc is the composition of morphismsφc,k arising via the exact sequences

0→ qTc,{1,...k−1} → Tc({1, . . . , k − 1}, ∅, {k + 1, . . . , t})→ qTc,{1,...k} → 0.

In fact, the middle term is the mapping cone ofφc,k : qTc,{1,...k}[−1]→ qTc,{1,...k−1}.By hypothesis, the middle term is exact and hence, by the induced long exactsequence, we see thatφc,k is quasi-isomorphism. Since a composition of quasi-isomorphisms is a quasi-isomorphism, it follows thatφc is a quasi-isomorphism.By the induced long exact sequence for the mapping cone ofφc, we then concludethat the mapping cone ofφc, which equalsT�c, is exact.

3.⇒ 4. is trivial.4. ⇒ 5. The complex of vector spaces(Tc(I, ∅, ∅))a obtained by fixing an

internal degreea in the region complexTc(I, ∅, ∅) is the same as the degreea part(T�b)a in the corner complexT�b for bI = cI andbj > −aj + n for somej ∈ I ′.This is because, for suchb, the free modules in the the complexT≥b do not containelements of degreea.

5. ⇒ 1.: For anyI ( {1, . . . , t} and c, the complexTc(I, ∅, ∅) is exactfor sufficiently largec by hypothesis. By Lemma3.4 below, the subcomplexTc−1i(I \ {i}, {i}, ∅) is exact for eachi ∈ I. From the exact sequence

0→ Tc−1i(I \ {i}, {i}, ∅)→ Tc(I, ∅, ∅)→ Tc−1i(I, ∅, ∅)→ 0

we see thatTc−1i(I ∪ {i}, ∅, ∅) is exact too. Descending induction now shows thatTc(I, ∅, ∅) is exact for allc.

By Lemma3.4, Tc(I, J, ∅) is exact for allc, I, J , so long asI∪J ( {1, . . . , t}.In particular the strands ofT are all exact.

23

Lemma 3.4. If Tc(I ∪ J,K, ∅) is exact, then so is its subcomplexTc−1J (I, J ∪K, ∅).

Proof. To prove the exactness atTc−1J (I, J ∪K, ∅)d we decomposeTc(I ∪ J,K, ∅)d−1

as a direct sum of graded freeE-modulesTc(I ∪ J,K, ∅)′d−1⊕Tc−1J (I, J ∪K, ∅)d−1.

SinceTc(I ∪ J,K, ∅) is assumed exact, the module of cyclesQ ⊂ Tc−1J (I, J ∪K, ∅)d

is equal to the sum of the boundaries inTc−1J (I, J ∪K, ∅)d and the intersectionP

of the image ofTc−1J (I, J ∪K, ∅)′d−1

with Tc−1J (I, J ∪K, ∅)d. It thus sufficesto show thatP is contained in the maximal ideal ofE timesQ.

The differential inTc−1J (I, J ∪K, ∅) involves only the variables ofE(J∪K)′,and thus the module of cycles ofTc−1J (I, J ∪K, ∅) in Tc−1J (I, J ∪K, ∅)d is gen-erated by linear combinations of the free generators ofTc−1J (I, J ∪K, ∅)d withcoefficients inEL, whereL is the complement ofJ ∪K.

On the other hand, it is clear from the form of the differential of Tc(I ∪ J,K, ∅)thatP ⊂ mEJ

Tc−1J (I, J ∪K, ∅)d, wheremEJdenotes the maximal ideal ofEJ .

Thus no element ofP can be a minimal generator ofQ.

Corollary 3.5. For any coherent sheafF on P, the complexT(F) satisfies theequivalent conditions of Theorem3.3. In particular all strandsT(F) are exact,and henceT(F) is a Tate resolution.

Proof. By constructionT is a subcomplex of an inverse limit of the acyclic com-plexesTailc(M), which are defined for sufficiently positivec. FurthermoreT�c

coincides with the exact complex obtained as the mapping cone of

Tailc(M)[−t]→ R(Γ≥c(F)),

soT satisfies condition4. of Theorem3.3.

The following proposition implies Corollary0.3.

Proposition 3.6. Let T be a locally finite minimal complex of graded freeE-modules. LetI ∪ J = {1, . . . , t} be a decomposition into disjoint sets, and letTI

denote the complex ofEJ -modules such that

TI ⊗ ωEI

is theI-th strandT0(∅, I, ∅) of T . The complexUJ(TI) is a Beilinson monad forRπJ ∗(U(T )) in Db(PJ).

24

Proof. We have

RpπJ ∗Ua =

{UaJJ if aI = 0 andp = 0,

0 otherwise.

Since, in particular,U(T ) is πJ∗-acyclic,RπJ ∗(U(T )) is represented by the com-plexπJ∗U(T ) = UJ (TI).

Proposition3.6 gives an interpretation of the 0-strand of a Tate resolution.Theorem0.4, which we now prove, provides a similar interpretation for everystrand.

Proof of Theorem0.4. We may assumec ≥ 0. Let b = c+1t and letP • the imageof the lower quadrant complexT≤c in the upper quadrant complexT≥b along thecorner map. Then by the argument of Theorem2.2, U(T ) andL(P •) are quasi-isomorphic, and similarlyU

(T [c](c)

)andL(P •[c](c)) are quasi-isomorphic. By

Lemma1.2we haveL(P •[c](c)) = L(P •)(c), and the first result follows.The second statement follows from Proposition3.6.

Example 3.7. ForF = Ua andT = T(Ua) the corner complexT�−a of T at−ahas terms

. . .→ ⊗tj=1Λ

nj−aj+1Vj ⊗ ωE(a)→ ωE(a)→ ⊕tj=1Λ

aj+1Wj ⊗ ωE(a)→ . . .

with cohomological indexing such that(T�−a)0 = ωE(a). Notice that the socle of

⊗tj=1Λ

nj−aj+1Wj ⊗ ωE(a) sits in a single degreen + 1t, while the socle of theright hand side sits in several degreesa− aj − 1j. The functorU takes all but themiddle term to zero. ThusU(T�−a) = Ua.

The corner complexT�1t has terms

· · · → ⊕tj=1Λ

nj−aj+1Vj ⊗ ωE(a)→ ωE(a)→ ⊗tj=1Λ

aj+1Wj ⊗ ωE(a)→ · · ·

and satisfiesU(T�1t) = Ua[1] as well.

Definition 3.8. Let ℓ1 ≤ ℓ2 be integers. A locally finite free complexT of freeE-modules hasfinite amplitude[ℓ1, ℓ2], if for all a ∈ Zt andd ∈ Z, when we write

T d =∑

a

Bda ⊗ ωE(−a)

we have thatBda 6= 0 only if ℓ1 ≤ d− |a| ≤ ℓ2.

25

Note that ifT has finite amplitude[ℓ1, ℓ2], thenT (c)[c] has finite amplitude[ℓ1, ℓ2] as well, whileT [c] or T (c) have a different amplitude. The Tate resolutionof a coherent sheaf has amplitude[0, |n|] by Theorem0.1.

Corollary 3.9. Every Tate resolution has a finite amplitude.

Proof. U(T (c)[c]) andU(T )(c) represent the same object inDb(P) for everyc. In particular the homology sheavesHk(U(T (c)[c])) occur only in finitelymany places, say fromk1 to k2 independent ofc. Thus the hypercohomologyHi(U(T (c)[c]) is nonzero only ifi ∈ [ℓ1, ℓ2] = [k1, k2 + |n|], and the terms ofThave the formT d = ⊕aH

i−|a|(U(T )(c))⊗ ωE(−a) by Theorem2.1.

Remark 3.10. If T is a Tate resolution with amplitude[ℓ1, ℓ2], then its BeilinsonrepresentativeU(T ) can only have non-zero terms in cohomological degrees[ℓ1−|n|, ℓ2]

Maclagan and Smith have defined a notion of regularity for a multigradedmodule over the Cox ring of a smooth toric variety [MS04, Definition 1.1]. Usingthe corner complex and reciprocity, we compare this to the notion of sufficientlypositive from Defintion1.8. Although these notions can diverge (see Remark4.1),they are closely related. We make use of Notation1.10in the following.

Proposition 3.11.LetM be a finitely generated, multigradedS-module such thatno associated prime ofM contains the irrelevant ideal ofS. If b is sufficientlypositive forM , thenb lies in the multigraded regularity ofM . Conversely, ifblies in the multigraded regularity ofM , then for anyi, b + 1t − 1i is sufficientlypositive forM .

Proof. Let b be the irrelevant ideal ofS. ThenH0bM = 0 by assumption. Ifb is

sufficiently positive forM , sinceMc = H0(P, M(c)) for all c ≥ b, it follows that(H1

bM)c = 0 for all c ≥ b as well. In addition,M≥b admits a linear resolution

M≥b(b)← F0 ← F1 ← · · · ← Fp ← 0,

whereβk,a(M≥b(b)) only if a ≥ 0 and|a| = k. We fix somej = (j1, . . . , jt) ∈ Nt

with |j| = i. To show thatb lies in the multigraded regularity, we need to showthat(H i+1

bM)b−j = 0 which amounts to showing thatH i(P, M(b− j)) = 0. For

this, it suffices to check thatH i+k(P, Fk(−j)) = 0 for all k. Since the resolution islinear, this amounts to checking thatH i+k(P,O(−a− j)) = 0 whenevera, j ≥ 0and|a| = k and|j| = i, and this holds for line bundles onP.

26

Conversely, letb lie in the multigraded regularity ofM . The first half of thesecond condition of Proposition1.7 is immediate from the definition of multi-graded regularity [MS04, Definition 1.1]. It then suffices to show thatM≥b+1t−1i

has a linear resolution, as the second half of the second condition will follow.By Corollary1.4, it suffices to show thatR(M≥b+1t−1i) is acyclic. LetT be theTate resolution ofM and consider the corner complexT≤b−1i [−t] → T≥b+1t−1i .Since the higher cohomology ofM vanishes for all multidegreesc ≥ b − 1i, itfollows that the quadrant complexT≥b+1t−1i equalsR(M≥b+1t−1i). So we wantto show thatT≤b−1i [−t], has no terms in cohomological degrees> |b− 1i| − t, orequivalently thatT≤b−1i has no terms in cohomological degrees≥ |b|.

By Theorem 1.12, any such term would correspond to a nonzero cohomologygroupHp(P, M(b−1i−j)) wherej ∈ Nt andp ≥ |1i+j|. However these groupsare all zero sinceb is in the multigraded regularity.

4 An example onP1 × P1

We recall that, by [Cox95], we can present a coherent sheaf onP1 × P1 as a bi-graded module over the Cox ringS = K[x0, x1, y0, y1]. Any sheaf onP3 with aGm action will define such a module. For example, consider the universal sub-bundleU onP3 and the corresponding sheafF onP1 × P1. As in [EFS03, Theo-rem 4.1], the cohomology table ofU is given by the Betti table of its Tate resolu-tion:

-5 -4 -3 -2 -1 0 1 2 3 4 53: 120 70 36 15 4 . . . . . .2: . . . . . . . . . . .1: . . . . . 1 . . . . .0: . . . . . . 6 20 45 84 140

Consider the mapT 0(U)→ T 1(U) indicated by the numbers 1 and 6 in the tableabove. This corresponds, in theZ2 grading, to the mapωE → ωE(−2, 0) ⊕ω4E(−1,−1)⊕ ωE(0,−2) defined by the matrix

m = (e0e1, e0f0, e1f0, e0f1, e1f1, f0f1)t

whereV = V1 ⊕ V2 = 〈e0, e1, f0, f1〉. By reciprocity,

L(imagem)→M → 0

is the minimal free resolution of the module of global sectionsM =∑

(a,b)∈Z2 H0(F(a, b))of a rank3 vector bundleF .

27

The cohomology table ofF , written as a matrix overZ[h], is:

( 2∑

i=0

hi(F(a, b)) · hi)−3≤a,b≤3

=

28h 18h 8h 2 12 22 3220h 13h 6h 1 8 15 2212h 8h 4h 0 4 8 124h 3h 2h h 0 1 24h2 2h2 0 2h 4h 6h 8h12h2 7h2 2h2 3h 8h 13h 18h20h2 12h2 4h2 4h 12h 20h 28h

Note that the corresponding Tate resolution

T = T(F) : · · · → T−4 → T−3 → T−2 → T−1 → T 0 → T 1 → T 2 → T 3 → T 4 →

has terms from the diagonal colored bands of the cohomology table withT 0 corre-sponding to the main diagonal above. The maps have components correspondingto arrows pointing to terms in north, north-west and west direction in the nextcolored band.

For example, the matrixm above corresponds to the three arrows encoded inthe submatrix

1

4h 1

.

As another example, there is a(3 + 2 + 2 + 3) × (8 + 2 + 2 + 8) submatrix ofT−2 → T−1 encoded by

8h3h 2h2h2 2h

2h2 3h 8h

.

Here6 of the16 blocks are zero because of the north/west condition.In this example, the cohomology ofF is “natural”, corresponding to the fact

that each entry of the cohomology table is a monomial. In the more general casethe terms of an entry that is not a monomial would contribute summands to differ-entT d.

The Tate resolutionT (U) on P3 can also be thought of the as the complexobtained by considering theT�0(F) with respect to the natural “coarse”Z grading.

28

Remark 4.1. In the above example, by considering the cohomology table ofF ,we can see that(1, 1) lies in the multigraded regularity ofM . However, sinceH1(F) = k1, this would yield a term of homological degree1 in the cornercomplex forM at (1, 1). The total betti numbers of the corner complexT�12F inthe coarse grading are

-3 -2 -1 0 1 2 3 43: 36 15 4 . . . . .2: 10 8 6 4 1 . . .1: . . . . . . . .0: . . . . 4 16 39 76

as one can read from the cohomology table ofF given above (using some valuesin addition to those shown in the table). In particular the complexR(M≥(1,1)),with betti numbers4, 16, 39, 76, . . . , is not exact, as one sees from the “1” in thesecond row of the Betti table, which represents an element ofthe kernel of thedifferential

ω8E(−2,−1)⊕ ω8

E(−1,−2)→ ω12E (−3,−1)⊕ ω15

E (−2,−2)⊕ ω12E (−1,−3)

which is not in the image ofω4E(−1,−1) → ω8

E(−2,−1) ⊕ ω8E(−1,−2). Thus

R(M≥(1,1)) is not acyclic andM≥(1,1) cannot have a linear resolution by Theo-rem 1.3. Hence(1, 1) is not sufficiently positive forM in our sense. However,both(1, 2) and(2, 1) are sufficiently positive.

5 Injective and projective resolutions

To prepare for the proof of Theorem6.1, we remind the reader of some generalresults about resolutions of complexes.

LetM• : · · · →M i →M i+1 → · · · be a bounded-above complex of modulesover a ringR. A projective resolutionof M• is a complex of projective modulesF • and a quasi-isomorphismF • → M•, and similarly for injective resolutions.Such resolutions were constructed in the famous book of Cartan-Eilenberg [CE99]by putting together resolutions of the various kernels and cokernels of maps inM•, but the same goal can be accomplished using iterated mapping cones. Forthe reader’s convenience we give a proof of this elementary result.

Proposition 5.1(Resolution of complexes by iterated mapping cones). LetR bea ring. Let

M• : · · · →Mk−1 → Mk → · · ·

29

be a bounded above complex ofR-modules, and let

M•p : 0→Mp →Mp+1 → · · ·

be the subcomplex ofM• obtained by truncation. For eachp, let

G•p : . . .→ Gp−1p → Gp

p →Mp → 0

be a projective resolution ofMp. There is a sequence of projective complexes

· · · ⊂ F •p+1 ⊂ F •p ⊂ · · ·

and surjective quasi-isomorphismsF •p → M•p such thatF •p is the mapping cone

of a mapG•p[−1]→ F •p+1, andF • := ∪pF •p is a projective resolution ofM•. Thusthek-th term ofF • isF k =

∑p≤k F

kp .

A similar result holds for bounded below complexes and injective resolutions.

If φ : A → B is a map of complexes we write the mapping cone ofφ as[φ]

or [Aφ✲ B] or even[A → B] whenφ is clear from context, so that there is an

exact sequence of complexes

0→ B → [φ]→ A[1]→ 0.

Proof. Note thatM = ∪pM•p . SinceM• is bounded above there is ak such that

Mp = 0 for all p > k, and we may takeF •p = M•p , the complex whose terms are

all 0, for p > k. We now use descending induction onp.Suppose thatF •p+1 → M•

p+1 is a projective resolution. We will show that themap of complexes

φ :(0→Mp → 0

)[1] −→

(0→Mp+1 → Mp+2 → · · ·

)

induced by the differential ofM• lifts to a mapφ′ : G•p[−1] → F •p+1, so thatthe mapsφ andφ′ are quasi-isomorphic, and we defineF •p to be the mappingconeF •p := [φ′]. From the long exact sequence of the mapping cone, we obtainH iF •p

∼= H iF •p+1∼= H iM• for i > p+ 1, as well as a four-term exact sequence

0→ HpF •p → HpG•p → Hp+1F •p+1 → Hp+1F •p → 0.

SinceHpG•p = Mp andHp+1F •p+1 = ker(Mp+1 →Mp+2)we immediately obtainthatH iF •p = H iM•

p for i = p, p + 1. HenceF •p is a projective resolution ofM•p .

30

By construction,F •p+1 ⊂ F •p and they agree in degree≥ p + 1. It follows thatF •

is a projective resolution ofM•, as claimed.It remains to produce the mapφ′. The moduleZ of (p + 1)-cycles inM•

p+1 isthe same as that inM•

p , andM•p+1 has no(p+ 1)-boundaries, so the moduleZ of

(p+ 1)-cycles inF •p+1 maps surjectively toZ. SinceM qp+1 = 0 for q < p+ 1, the

quotient complex ofF •p+1 obtained factoring outF rp+1 for r ≥ p+1 is a resolution

of the kernel ofZ → Z. SinceG•p is a projective resolution ofMp, the mapGp

p → Mp lifts to a mapGpp → Z, and we continue lifting to obtain the desired

map of complexesφ′ : G•p[−1]→ F •p+1, as in the diagram:

F p+2p+1

✲ Mp+2

F p+1p+1

✻

✲ Mp+1

✻

∪ ∪

F p−1p+1

✲ F pp+1

✲

✲

Z ✲ Z ✲ 0

φ′ :

Gp−2p

✻

✲ Gp−1p

✻

✲ Gpp

✻

✲ Mp

✻

✲ 0.

The reciprocity theorem for resolutions of modules overE andS is a specialcase of a reciprocity theorem for complexes, proved in the same way.

Theorem 5.2. Let M• be a bounded complex of finitely generatedS-modulesand P • a bounded complex of finitely generatedE-modules. ThenR(M•) isan injective resolution of the complexP • if and only if L(P •) is a projectiveresolution ofM•.

Proof. The key point is thatL andR are adjoint functors. See [EFS03, Theorem2.6], [BGG78, Theorem 3] for details.

6 Tate Resolutions from Beilinson Monads

LetF be a finite complex of sheaves. Generalizing the case of a single sheaf, wemay construct aTate resolutionT(F) as follows. We may representF by a finite

31

complex of gradedS-modulesM•. SinceR is a functor, we may apply it term byterm to the complexM• to get a double complex of modules overE. We then canapply the procedure of Section1. We writeM•

≥c for the complex

· · · →Mp≥c →Mp+1

≥c → · · · ,

whereMp≥c is the truncation in degrees≥ c as in Section1.

Theorem 6.1 (Tate resolution and Beilinson representative of a complex). LetF be a bounded complex of coherent sheaves. Then there exists aminimal TateresolutionT = T(F) such thatU(T (c)[c]) is quasi-isomorphic toF(c) for everyc ∈ Zt. In particularT(F) satisfies Equation1.

Proof. Let F be represented by a finite complexM• of finitely generated gradedS-modules, and letc ∈ Zt be sufficiently positive for each of the finitely manynonzero modulesMp in M•. Each moduleMp

≥c has a linear minimal free res-olution Lp and the differentials ofM• lift to homogeneous mapsLk → Lk+1.Since adding a homotopy would make the maps inhomogeneous, the lifted mapsare unique and compose to 0. We may writeLk = L(P k) for suitable finitelygeneratedE-modulesP k, and it follows that the lifted maps on theLk make theP k into a complex, which we denoteP •.

SinceR is a functor, we may regardR(M•) as a double complex. By The-orem1.3 the complexR(Mk

≥c) is the injective resolution ofP k. We regard thedifferentials inR(M•) coming from the individualR(Mk

≥c) as the “horizontaldifferentials” in this double complex. With this convention the “horizontal ho-mology” ofR(M•

≥k) is the complexP •.Replacingc by c + 1t if necessary, we may ensure thatc − 1t is sufficiently

positive for eachMk. Then the minimal projective resolution ofP k starts with thetermR(Mk

c−1t) by Theorem1.13.Let Tailc(M•)→ P • be a minimal projective resolution. By Theorem5.1we

can obtain such resolution by minimizing an iterated mapping cone of projectiveresolutions of theP k. As in the case of a single module, we set

T(F) :=(lim∞←c

′Tailc(M•))[1− t].

SinceL(P •)→ M•≥c is a resolution, the result follows as in Section2.

Conversely given any locally finite complexT of freeE-modules, we get afinite complex of sheavesU = U(T ), which we may regard as a Beilinson repre-sentative of an objectF in Db(P). We will now show how to construct the Tate

32

resolutionT(U(T )) directly, in the context of resolutions over the exterior alge-bra, without passing through the category of sheaves. We will achieve this goal inseveral steps.

There is a unique smallest free subquotient complexBW(T ) of T , called theBeilinson windowof T , such thatU(T ) = U(BW(T )). Indeed, if

T d =∑

a∈Zt

Ba ⊗ ωE(−a)

BW(T ) is the subquotient complex with terms

BW d =∑

a∈Zt

0≥a≥−n

Ba ⊗ ωE(−a).

Now suppose thatBW is any complex of finitely generated freeE-modulesthat are direct sums of modules of the formωE(b) with 0 ≤ b ≤ n. We callU(BW ) minimal if BW is minimal. In general,U(BW ) is the direct sum ofa minimal complexUmin and trivial complexes of type0 → U b → U b → 0for various degreesb with 0 ≤ b ≤ n and various shifts; this follows from thecorresponding fact forBW .

Theorem 6.2.LetT be a minimal Tate resolution andBW = BW(T ) its Beilin-son window. LetP (BW ) ⊂ BW be the subcomplex

P (BW ) = 〈f ∈ BW homogeneous| deg f 6≥ 1t〉 ⊂ BW.

If BW/P (BW ) → I is a minimal injective resolution andF → P (BW ) is aminimal projective resolution, then the corner complexT�1t is isomorphic to

[ [F → BW ]→ I ].

and the shifted corner complexcT := (T�1t)[−1] hasBW as a subquotient com-plex.

Example 6.3. If BW consist of a single termωE(b) with 0 ≤ b ≤ n thenU(BW ) = U b and the corner complex(T�1t)[−1] as described in Example3.7for T = T(U b) coincides withcT . Indeed, in this case

P (BW ) := 〈f ∈ ωE(b) homogeneous| deg f 6≥ 1t〉

= 〈f ∈ ωE(b) homogeneous| ∃j : (deg f)j ≤ 0〉

= ker(ωE(b)→ ⊗tj=1Λ

bj+1Wj ⊗ ωE(b)).

Here we consider the vector spaceWj to be concentrated in degree−1j , soΛbj+1Wj ⊗ ωE(b)) is isomorphic to a direct sum of copies ofωE(−1

t).

33

Proof of Theorem6.2. The lower quadrantqT = T≤0 is a quotient complex ofcTand the shifted upper quadrantuT = T≥1t [t− 1] is a subcomplex ofcT , which inturn is a cone over the corner map between these two quadrant complexes.BWis a subcomplex ofqT , and we let

qT− := qT/BW ∼=∑

a∈Zt

0≤−a,−a 6≤n

Ba ⊗ ωE(−a)

denote the part ofqT outside the Beilinson window. Since

Λni+2Vi = 0,

the corner map induces the zero map fromqT− to uT .Thus

cT [1] = [ [qT−[−1]→ BW ]→ uT ] = [qT−[−1]→ [BW → uT ] ]

where[A→ B] denotes the cone over a map of complexes as in Section5.By the exactness ofcT we have

P (BW ) = 〈f ∈ BW homogeneous| deg f 6≥ 1t〉 = ker(BW → uT ).

The last equality holds since, by Example6.3, it holds for every termωE(b) ofBW , and because the injective resolution can be obtained by minimizing an iter-ated mapping cone as in Proposition5.1. SinceT and hencecT are minimal, werecoveruT from BW as the minimal injective resolution ofBW/P (BW ). AlsoF = qT−[−1] is the minimal projective resolution ofP (BW ) ⊂ BW . Hence werecoverqT = [F → BW ] andcT [1] = [qT → uT ] as cones.

Algorithm 6.4. [lower quadrant of the Tate extension]Input: BW : 0 → BW r → . . . → BW s → 0, a finite minimal complex withtermsBW d =

∑0≥a≥−n B

da ⊗ ωE(−a) and an integerd with d < r.

Output: The lower quadrant complex

qT = T1t({1, . . . t}, ∅, ∅) = [F → BW ]

of Theorem6.2in homological degreesd, . . . , s.

1. Setk := s, qT k := BW s andqT k+1 := 0.

2. Whilek > d do:

34

(a) ComputeLk = ker(qT k → qT k+1)/image(BW k−1 → qT k).

(b) Compute minimal generators of

P (Lk) = 〈f ∈ Lk homogeneous) | deg f 6≥ 1t〉

.

(c) Choose a mapF k → ker(qT k → qT k+1) from a freeE-module ofthe formF k =

∑0≥a6≥−n B

ka ⊗ ωE(−a) such that the image of the

generators inF k generatesP (Lk) minimally.

(d) SetqT k−1 := BW k−1 ⊕ F k, and extend the differential ofqT to in-clude the mapqT k−1 → qT k.

(e) Replacek by k − 1.

3. Return the complex

qT d → qT d+1 . . .→ qT s → 0.

Proof. We keep the notation of the proof of Theorem6.2. Representatives inqT k

of elements ofP (Lk) map to zero inuT [1] for degree reasons. By the exactnessof cT they must be covered by elements inqT−. The result follows.

Algorithm 6.5. [Corner Complex]Input: BW : 0 → BW r → . . . → BW s → 0, a finite minimal complex

with termsBW d =∑

0≥a≥−n Bda ⊗ ωE(−a) and a degreea ∈ Zt with a < −n.

Output: A finite piece of the corner complexT�a of the Tate resolutionT =T(U(BW )).

1. Setd = r + |a| − |n| − t ande = r + |a|

2. Apply Algorithm6.4to computeqT : qT d → qT d+1 . . .→ qT s → 0.

3. CollectℓT , uT the terms and maps ofqT contributing to the lower partTa({1, . . . , t}, ∅, ∅) and the upper partTa(∅, ∅, {1, . . . , t}) respectively.

4. Compute the corner mapsℓT [t − 1] → uT and form the subquotient com-plex cT = [ℓT [t− 1]→ uT ] of T�a.

5. Then the differentialcT e−1 → cT e and(T�a)e−1 → (T�a)

e coincide.

35

6. Extend the part of the corner complex to the complex

(T�a)e−1 → . . .→ (T�a)

s

by injective (and possibly projective) resolutions of the single correct dif-ferential from Step 5, and return the result.

Proof. Since the nonzero summandsBrc ⊗ ωE(−c) of BW r satisfyc ≤ 0 taking

syzygy we see that the nonzero summandsBe−1c ⊗ωE(−c) of qT e−1 satisfyc 6≥ a.

ThusqT e is the last term which possibly could contribute touT andqT d is the lastterm which possibly maps touT e via a corner map. Moreover sincea < −n thelower quadrantTa({1, . . . , t}, ∅, ∅) maps to no termBℓ

c ⊗ωE(c) of T ℓ with c 6≤ 0.ThuscT e−1 → cT e and(T�a)

e−1 → (T�a)e coincide.

Algorithm 6.6. [Cohomology Table]Input: BW : 0 → BW r → . . . → BW s → 0, a finite minimal complex withtermsBW d =

∑0≥a≥−n B

da ⊗ ωE(−a); and two degreesa, b ∈ Zt with a < −n

and0 < b.Output: The cohomology table ofF = U(BW )

{c ∈ Zt | a ≤ c ≤ b} → Z[h, h−1]

c 7→∑

k∈Z

dimHk(F(c))hk

1. Sete = r + |a|

2. Use Algorithm6.5to compute the differential

(T�a)e−1 → (T�a)

e

of the corner complexT�a of T = T(F).

3. Setk = e− 1 and∂k : uT k → uT k+1 = (T�a)k → (T�a)

k+1.

4. repeat

(a) replacek by k + 1

(b) Compute a minimal free injective hullI =∑

c∈Zt Bk+1c ⊗ ωE(−c) of

coker ∂k−1.

36

(c) Set

uT k+1 :=∑

c∈Zt

c≤b

Bk+1c ⊗ ωE(−c)

and define∂k : uT k → uT k+1 as the induced map.

until uT k+1 = 0.

5. Read off the cohomology table from the complex

uT e → . . .→ uT k

by collecting the dimensions

k∑

ℓ=e

dimBℓc h

ℓ−|c| ∈ Z[h, h−1]

for all c ∈ Zt with a ≤ c ≤ b.

Proof. Let T = T(U(BW )). The upper quadrantuT = Ta(∅, ∅, {1, . . . , t})is a subcomplex of the corner complex, which by exactnessss can be computedfrom any of its differentials. Since the nonzero summandsBr

c ⊗ ωE(−c) of BW r

satisfyc ≤ 0 taking syzygy we see that the nonzero summandsBe−1c ⊗ωE(−c) of

uT e−1 satisfyc 6≥ a. So these do not contribute to the desired cohomology table.For the computation in the injective direction we may drop terms Bℓ

c ⊗ ωE(c)with c 6≤ b, because the differential of theT and the corner complex goes in thecohomology table in northeast direction, so that these terms cannot contribute toterms or computations of terms in the desired cohomologicalrange. Hence thecorrectness follows from Equation1 applied to the complexU(BW ).

Corollary 6.7. Starting from a Beilinson windowBW = BW(T ) of a Tate reso-lution, one can recover arbitrarily large finite parts of theTate resolutionT withcomputations over the exterior algebra.

Proof. Apply the three algorithms above.

Using this result we can answer the question: “When is a complex of sheavesonP quasi-isomorphic to a vector bundle in homological degree 0?”

37

Corollary 6.8. LetT be a Tate resolution. ThenU(T ) is a monad of a sheaf if andonly if the cohomology table of a positive quadrantTc(∅, ∅, {1, . . . , t}) for somec ≫ 0 has onlyh0 entries. It is a monad of a vector bundle if only if in additionalso the cohomology table of every very negative quadrantTc({1, . . . , t}, ∅, ∅) forc≪ 0 has onlyh|n| entries.

The corollary above is not an effective criterion, because we have no algorithmto compute the whole Tate resolution. However as in the case of a single projectivespace, see for example [EFS03, Example 7.3], there is an effective method forproving that a bounded complex is quasi-isomorphic to a vector bundle.

Proposition 6.9. Considera, b ∈ Zt with b > 0 and let t0 < t1 ∈ Z. LetF ∈ Db(P). Suppose

H∗F(a+ bt) = H|n|F(a+ bt) for t ∈ [t0, t0 + |n|]

andH∗F(a+ bt) = H0F(a+ bt) for t ∈ [t1 − |n|, t1].

ThenF is quasi isomorphic to a vector bundle onP.

Proof. ConsiderL = O(b1, . . . , bt). The line bundleL is very ample and definesan embeddingι : P → PN with N + 1 =

∏tj=1

(bj+nj

nj

). The cohomology table

of G = ι∗F(a) with respect toOPN (1) can be read of from the values of thecohomology table ofF along the integral lineZ → Zt, t 7→ a + bt. If π :P → P|n| denotes a linear Noether normalization ofι(P) ⊂ PN , thenG andπ∗Ghave the same cohomology table, andG is a vector bundle, i.e. Cohen-Macaulaysheaf of dimension|n|, iff and only if π∗G is a vector bundle. The assumptionH∗F(a + bt) = H0F(a + bt) for t ∈ [t1 − |n|, t1] implies that forT(π∗G) theassumption of [EFS03, Lemma 7.4] is satisfied at positiont1. So in particularHk∗(π∗G) = 0 for k > |n|. Similarly using the dual complex,H∗F(a + bt) =

H|n|F(a + bt) for t ∈ [t0, t0 + |n|] implies using the Lemma,Hk∗(π∗G) = 0 for

k < 0. Moreover, the two applications of the Lemma imply that the intermediatecohomology groupsHk

∗(G) = ⊕d∈ZHk(π∗G(d)) for 0 < k < |n| have finite

length. Thusπ∗G on P|n|, and henceF on P, is quasi-isomorphic to a vectorbundle by the Auslander-Buchsbaum formula [Eis95].

Remark 6.10. We have implemented these algorithms in our Macaluay2 packageTateOnProducts.m2[M2-Tate].

38

7 Application to split vector bundles

Horrocks’ splitting criterion ([Hor64,OSS11]) says that if a vector bundle onPn

has no intermediate cohomology, then that vector bundle splits as a sum of linebundles. That is:if the cohomology table of a vector bundle looks like a sum ofline bundles, then the vector bundle itself splits as a sum ofline bundles.

Question 7.1.Suppose that the cohomology table of a vector bundleE onP :=Pn1×Pn2 ×· · ·×Pnt can be written as a positive integral sum of the cohomologytables of line bundles onP. Is E a direct sum of line bundles?

We prove that the answer is “yes” under an additional hypothesis that is auto-matically satisfied in the caset = 1. For c ∈ Zs we writeγi,c(E) := hi(P, E(c))and we writeγ(E) = (γi,c(E))i,c for the cohomology table ofE . As throughoutthis paper, we use the termwise partial order for comparing integer vectors inZt.

Theorem 7.2.LetE be a vector bundle onP := Pn1 × Pn2 × · · · × Pnt such thatthe cohomology table ofE decomposes as a positive sum of line bundles:

γ(E) =s∑

i=1

γ(O(c(i))mi) wherec(i) ∈ Zt.

If c(1) ≥ c(2) ≥ · · · ≥ c(s) thenE splits as

E ∼=

s⊕

i=1

O(c(i))mi .

The proof uses our construction of the Beilinson monad. We begin with amore general lemma.

Lemma 7.3. Suppose that the cohomology table ofE splits asγ(E) = γ(Om1) +γ(E ′), whereγ(E ′) is the cohomology table of some vector bundleE ′, and whereU−1(E ′) = 0. ThenE splits as

E = Om1 ⊕ E ′′,

for some vector bundleE ′′ satisfyingγ(E ′′) = γ(E ′).

Proof. By assumption,U0(E) hasU0(Om1) = Om1 as a summand, so we maywriteU

0(E) = Om1 ⊕U0(E ′). Note thatU−1(E) = U

−1(Om1)⊕U−1(E ′) = 0.

ThusU(E) has the form

= · · · → 0→ Om1 ⊕U0(E ′)→ U

1(E)→ . . .

39

Further, the summandOm1 must map to 0 inU1(E): minimality ensures that anymap between sums ofO are zero, and the fact that the terms in the image of thefunctorU form a strong exceptional collection implies thatO = U0 admits nononzero maps to anyUa with a 6= 0. Thus,Om1 is a direct summand of the zerothcohomologyE = H0(U(E)) as claimed.

Proof of Theorem7.2. By induction, it suffices to show thatO(c1)m1 is a sum-mand ofE . Without loss of generality, we may assume thatc(1) > c(2). Replac-ing E by E(−c(1)), we can further assume thatc(1) = 0 and thus thatc(i) < 0for all i = 2, . . . , s. We will complete the proof by verifying the hypotheses ofLemma7.3.

Let E ′ :=⊕s

i=2O(c(i))mi . Then we haveγ(E) = γ(Om1)+ γ(E ′) by assump-

tion, and we need to show thatU−1(E ′) = 0. Since

U−1(E ′) =

s⊕

i=2

U−1(O(c(i))mi),

it suffices to show thatU−1(O(b)) = 0 for anyb < 0. By Theorem2.1we have

U−1(O(b)) =

⊕

p+|a|=−1

U−a ⊗Hp(P,O(b+ a))

Assume thatHp(P,O(b + a)) 6= 0 for some−n ≤ a ≤ 0. Then there existj1, . . . , jt such that

∑i ji = p and

Hp(P,O(b+ a)) =

t⊗

i=1

Hji(Pni,O(bi + ai)).

There are two possibilities for eachji: eitherji = 0 or ji = ni. If ji = 0, thensincebi andai are both nonpositive, we must havebi = ai = 0 andji + ai = 0.If ji = ni, then sinceai ≥ −ni, we must haveji + ai ≥ 0. This implies thatp + |a| =

∑i ji + ai ≥ 0 wheneverU−a andHp(P,O(b+ a)) are both nonzero,

and it follows thatU−1(O(b)) = 0 as claimed. We may now apply Lemma7.3tosplit off a copy ofOm1 , completing the proof.

8 More Open Questions

Question 8.1.Consider the caset = 2. If F ∼= F1 ⊠ F2 then its Tate resolutionhas the structure of a double complex:

T(F) ∼= T(F1)⊗K T(F2) as complex ofE = E1 ⊗K E2 modules.

40

Similarly a direct sum of box-products has a Tate resolutionwhich is naturallya double complex. Does the converse hold? In other words, assume thatT(F)admits the structure of a double complex, so that the differential ∂ decomposes as∂ = ∂h + ∂v, and assume moreover that the entries of∂h come fromE1 while theentries of∂v come fromE2. Does it follow thatF is a direct sum of box products?One can ask a similar question for anyt ≥ 2.

Question 8.2.What is the geometric meaning of the other exact subquotientcom-plexes, say the “half plane” complexesTc({k}, ∅, ∅) or Tc(∅, ∅, {k}), defined inSection3?

Question 8.3.We showed in [ES10] that every complex onAm is the direct imageof a vector bundle onAm× Pn for somen. Does every complex onPm occur as apushforward of a vector bundle onPm × Pn1 × · · · × Pnt for some(n1, . . . , nt)?

References[Beı78] A. A. Beılinson, Coherent sheaves onPn and problems in linear algebra,

Funktsional. Anal. i Prilozhen.12 (1978), no. 3, 68–69 (Russian). MR509388(80c:14010b)

[BGG78] I. N. Bernsteın, I. M. Gel˜fand, and S. I. Gel˜fand, Algebraic vector bundles onPn

and problems of linear algebra, Funktsional. Anal. i Prilozhen.12 (1978), no. 3,66–67 (Russian). MR509387 (80c:14010a)

[CE99] Henri Cartan and Samuel Eilenberg,Homological algebra, Princeton Landmarks inMathematics, Princeton University Press, Princeton, NJ, 1999. With an appendix byDavid A. Buchsbaum; Reprint of the 1956 original. MR1731415(2000h:18022)

[CDRMR11] L. Costa, S. Di Rocco, and R. M. Miro-Roig,Derived category of fibrations,Math. Res. Lett.18 (2011), no. 3, 425–432, DOI 10.4310/MRL.2011.v18.n3.a3.MR2802576 (2012f:14030)

[CMR04] L. Costa and R. M. Miro-Roig,Tilting sheaves on toric varieties, Math. Z.248 (2004), no. 4, 849–865, DOI 10.1007/s00209-004-0684-6. MR2103545(2005m:14023)

[Cox95] David A. Cox,The homogeneous coordinate ring of a toric variety, J. AlgebraicGeom.4 (1995), no. 1, 17–50. MR1299003 (95i:14046)

[CLS11] David A. Cox, John B. Little, and Henry K. Schenck,Toric varieties, GraduateStudies in Mathematics, vol. 124, American Mathematical Society, Providence, RI,2011. MR2810322 (2012g:14094)

[Eis95] David Eisenbud,Commutative algebra with a view toward algebraic geometry,Graduate Texts in Mathematics, vol. 150, Springer-Verlag,New York, 1995.

41

[ES09] David Eisenbud and Frank-Olaf Schreyer,Betti numbers of graded modules andcohomology of vector bundles, J. Amer. Math. Soc.22 (2009), no. 3, 859–888, DOI10.1090/S0894-0347-08-00620-6. MR2505303 (2011a:13024)

[ES10] , Cohomology of coherent sheaves and series of supernatural bundles,J. Eur. Math. Soc. (JEMS)12 (2010), no. 3, 703–722, DOI 10.4171/JEMS/212.MR2639316 (2011e:14036)

[ESW03] David Eisenbud, Frank-Olaf Schreyer, and Jerzy Weyman,Resultants and Chowforms via exterior syzygies, J. Amer. Math. Soc.16 (2003), no. 3, 537–579, DOI10.1090/S0894-0347-03-00423-5. MR1969204 (2004j:14067)

[EFS03] David Eisenbud, Gunnar Floystad, and Frank-Olaf Schreyer,Sheaf cohomology andfree resolutions over exterior algebras, Trans. Amer. Math. Soc.355(2003), no. 11,4397–4426 (electronic).

[ES13] David Eisenbud and Frank-Olaf Schreyer,The banks of the cohomology river,Kyoto J. Math. 53 (2013), no. 1, 131–144, DOI 10.1215/21562261-1966089.MR3049309

[GM03] Sergei I. Gelfand and Yuri I. Manin,Methods of homological algebra, 2nd ed.,Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2003. MR1950475(2003m:18001)

[Kap88] M. M. Kapranov,On the derived categories of coherent sheaves on some homoge-neous spaces, Invent. Math.92 (1988), no. 3, 479–508, DOI 10.1007/BF01393744.MR939472 (89g:18018)

[Hor64] G. Horrocks,Vector bundles on the punctured spectrum of a local ring, Proc. Lon-don Math. Soc. (3)14 (1964), 689–713. MR0169877 (30 #120)

[Huy06] D. Huybrechts,Fourier-Mukai transforms in algebraic geometry, Oxford Mathe-matical Monographs, The Clarendon Press, Oxford University Press, Oxford, 2006.MR2244106 (2007f:14013)

[MS04] Diane Maclagan and Gregory G. Smith,Multigraded Castelnuovo-Mumford reg-ularity, J. Reine Angew. Math.571 (2004), 179–212, DOI 10.1515/crll.2004.040.MR2070149 (2005g:13027)

[M2] Daniel R. Grayson and Michael E. Stillman,Macaulay2, a soft-ware system for research in algebraic geometry. Available athttp://www.math.uiuc.edu/Macaulay2/.

[M2-Tate] David Eisenbud, Daniel Erman, and Frank-Olaf Schreyer,TateOnProducts, package for Macaulay2. Available athttp://www.math.uni-sb.de/ag-schreyer/home/computeralgebra.

[OSS11] Christian Okonek, Michael Schneider, and Heinz Spindler,Vector bundles on com-plex projective spaces, Modern Birkhauser Classics, Birkhauser/Springer Basel AG,Basel, 2011. Corrected reprint of the 1988 edition; With an appendix by S. I.Gelfand. MR2815674 (2012d:14073)

42

[Ser55] Jean-Pierre Serre,Faisceaux algebriques coherents, Ann. of Math. (2)61 (1955),197–278 (French). MR0068874 (16,953c)

Author Addresses:

David EisenbudDepartment of Mathematics, University of California, Berkeley, Berkeley CA94720eisenbud@math.berkeley.edu

Daniel ErmanDepartment of Mathematics, University of Wisconsin, Madison WI 53706derman@math.wisc.edu

Frank-Olaf SchreyerMathematik und Informatik, Universitat des Saarlandes, Campus E2 4, D-66123Saarbrucken, Germanyschreyer@math.uni-sb.de

43