(de Gruyter 2001
Unstable hyperplanes for Steiner bundles and multidimensional matrices
Vincenzo Ancona and Giorgio Ottaviani*
(Communicated by G. Gentili)
Abstract. We study some properties of the natural action of SL V0 SL Vpon non- degenerate multidimensional complex matricesAAP V0n nVpof boundary format (in the sense of Gelfand, Kapranov and Zelevinsky); in particular we characterize the non-stable ones as the matrices which are in the orbit of a ``triangular'' matrix, and the matrices with a stabilizer containingCas those which are in the orbit of a ``diagonal'' matrix. For p2 it turns out that a non-degenerate matrixAAP V0nV1nV2detects a Steiner bundleSA(in the sense of Dolgachev and Kapranov) on the projective space Pn, ndim V2 ÿ1. As a consequence we prove that the symmetry group of a Steiner bundle is contained in SL 2and that the SL 2-invariant Steiner bundles are exactly the bundles introduced by Schwarzen- berger [Schw], which correspond to ``identity'' matrices. We can characterize the points of the moduli space of Steiner bundles which are stable for the action of Aut Pn, answering in the
®rst nontrivial case a question posed by Simpson. In the opposite direction we obtain some results about Steiner bundles which imply properties of matrices. For example the number of unstable hyperplanes ofSA(counting multiplicities) produces an interesting discrete invariant ofA, which can take the values 0;1;2;. . .;dimV01 ory; they case occurs if and only ifSAis Schwarzenberger (andAis an identity). Finally, the Gale transform for Steiner bun- dles introduced by Dolgachev and Kapranov under the classical name of association can be understood in this setting as the transposition operator on multidimensional matrices.
1 Introduction
A multidimensional matrix of boundary format is an element AAV0n nVp
whereViis a complex vector space of dimensionki1 fori0;. . .;pand
k0 Xp
i1
ki:
* Both authors were supported by MURST and GNSAGA-INDAM
We denote by DetA the hyperdeterminant of A(see [GKZ]). Lete 0j;. . .;e kjj be a basis inVjso that everyAAV0n nVphas a coordinate form
AX
ai0;...;ipe 0i0 n ne pip :
Let x 0j;. . .;x kjj be the coordinates in Vj. Then A has the following di¨erent
descriptions:
1) A multilinear form
X
i0;...;ip
ai0;...;ipx 0i0 n nx ipp:
2) An ordinary matrix MA mi1i0 of size k11 k01 whose entries are multilinear forms
mi1i0 X
i2;...;ip
ai0;...;ipx 0i2 n nx pip : 1:1
3) A sheaf morphism fAon the productXPk2 Pkp:
OXk01!fA OX 1;. . .;1k11: 1:2
Theorem 3.1 of chapter 14 of [GKZ] easily translates into:
Theorem.The following properties are equivalent:
i) DetA00;
ii) the matrix MAhas constant rank k11on X Pk2 Pkp;
iii) the morphism fA is surjective so that SA kerfA is a vector bundle of rank k0ÿk1.
The above remarks set up a basic link between non-degenerate multidimensional matrices of boundary format and vector bundles on a product of projective spaces. In the particular casep2 the (dual) vector bundleSAlives on the projective spacePn, nk2, and is a Steiner bundle as de®ned by Dolgachev and Kapranov in [DK]. We can keep forSAthe name Steiner also in the case pX3.
The action of SL V0 SL Vp on V0n nVp translates to an action on the corresponding bundle in two steps: ®rst the action of SL V0 SL V1leaves the bundle in the same isomorphism class; then SL V2 SL Vp acts on the classes, i.e. on the moduli space of Steiner bundles. It follows that the invariants of matrices for the action of SL V0 SL Vpcoincide with the invariants of the action of SL V2 SL Vpon the moduli space of the corresponding bundles.
Moreover the stable points of both actions correspond to each other.
The aim of this paper is to investigate the properties and the invariants of both the above actions. When we look at the vector bundles, we restrict ourselves to the case p2, that is Steiner bundles on projective spaces. This is probably the ®rst case where Simpson's question ([Simp], p. 11) about the natural SL n1-action on the moduli spaces of bundles onPn has been investigated.
Section 2 is devoted to the study of multidimensional matrices. We denote by the same letter matrices inV0n nVp and their projections inP V0n nVp. In Theorem 2.4 we prove that a matrixAAP V0n nVpof boundary format with DetA00 is not stable for the action of SL V0 SL Vpif and only if there is a coordinate system such that ai0...ip 0 for i0>Pp
t1it. A matrix satisfying this condition is called triangulable. The other main results of this section are Theorems 2.5 and 2.6 which describe the behaviour of the stabilizer subgroup Stab A. In Remark 5.14 we introduce a discrete SL V0 SL V1 SL V2-invariant of non- degenerate matrices in P V0nV1nV2 and we show that it can assume only the values 0;. . .;k02;y.
The second part of the paper, consisting of Sections 3 to 6, can be read independently of Section 2, except that we will use Theorem 2.4 in two crucial points (Theorem 5.9 and Section 6). In this part we study the Steiner bundles on Pn P V. As we mentioned above, they are rank-n vector bundles S whose dual S appears in an exact sequence
0!S!WnO!fA InO 1 !0 1:3
whereWandIare complex vector spaces of dimensionnkandk, respectively. The map fA corresponds to AAWnVnI (which is of boundary format) and fA is surjective if and only if DetA00. We denote by Sn;k the family of Steiner bundles described by a sequence as (1.3).Sn;1 contains only the quotient bundle. Important examples of Steiner bundles are the Schwarzenberger bundles, whose construction goes back to the pioneering work of Schwarzenberger [Schw]. Other examples are the logarithmic bundles W logH of meromorphic forms on Pn having at most loga- rithmic poles on a ®nite union H of hyperplanes with normal crossing; Dolgachev and Kapranov showed in [DK] that they are Steiner. The Schwarzenberger bundles are a special case of logarithmic bundles, when all the hyperplanes osculate the same rational normal curve. Dolgachev and Kapranov proved a Torelli type theorem, namely that the logarithmic bundles are uniquely determined up to isomorphism by the above union of hyperplanes, with a weak additional assumption. This assumption was recently removed by ValleÁs [V2], who shares with us the idea of looking at the schemeW S fHAPn4jh0 SH00gHPn4of unstable hyperplanes of a Steiner bundle S. ValleÁs proves that any SASn;k with at least nk2 unstable hyper- planes with normal crossing is a Schwarzenberger bundle and W S is a rational normal curve. We strengthen this result by showing the following: for anySASn;k
any subset of closed points inW Shas always normal crossing (see Theorem 3.10).
Moreover SASn;k is logarithmic if and only if W S contains at least nk1 closed points (Corollaries 5.11 and 5.10). In particular ifW Scontains exactlyn k1 closed points thenSFW logW S. The Torelli Theorem follows.
It turns out that the length ofW Sde®nes an interesting ®ltration into irreducible subschemes of Sn;k which gives also the discrete invariant of multidimensional matrices of boundary format mentioned above. This ®ltration is well behaved with respect to the PGL n1-action onPn and also with respect to the classical notion of association reviewed in [DK]. Eisenbud and Popescu realized in [EP] that the association is exactly what nowadays is called Gale transform. For Steiner bundles corresponding toAAWnVnI this operation amounts to exchanging the role of V withI, so that it corresponds to the transposition operator on multidimensional matrices.
The Gale transform for Steiner bundles can be decribed by the natural isomorphism Sn;k=SL n1 !Skÿ1;n1=SL k:
Both quotients in the previous formula are isomorphic to the GIT-quotient P WnVnI=SL W SL V SL I
which is a basic object in linear algebra.
As an application of the tools developed in the ®rst section we show that all the points ofSn;k are semistable for the action of SL n1and we compute the stable points. Moreover we characterize the Steiner bundles SASn;k whose symmetry group (i.e. the group of linear projective transformations preserving S) contains SL 2or containsC.
Finally we mention that W S has a geometrical construction by means of the Segre variety. From this construction W S can be easily computed by means of current software systems.
We thank J. ValleÁs for the useful discussions we had on the subject of this paper.
2 Multidimensional matrices of boundary format and geometric invariant theory It is well known that all one dimensional subgroups of the complex Lie group SL 2
either are conjugated to the maximal torus consisting of diagonal matrices (which is isomorphic toC) or are conjugated to the subgroupCF 1 b
0 1
bAC
. De®nition 2.1.A p1-dimensional matrix of boundary formatAAV0n nVp
is called triangulable if one of the following equivalent conditions holds:
i) there exist bases inVj such thatai0;...;ip0 fori0>Pp
t1it;
ii) there exist a vector space U of dimension 2, a subgroup CHSL U and iso- morphismsVjFSkjUsuch that ifV0n nVp0nAZWnis the decomposi- tion into a direct sum of eigenspaces of the induced representation then we have AA0nX0Wn.
Proof of the equivalence between i)andii). Let x;y be a basis ofUsuch thattAC acts onxandyastxandtÿ1y. Sete kj:xkykjÿk kj
k ASkjU for j>0 ande 0k :
xk0ÿkyk k0
k ASk0Uso thate 0i0 n ne ippis a basis ofSk0Un nSkpUwhich diagonalizes the action ofC. The weight ofe 0i0 n ne pip is 2 Pp
t1itÿi0, hence ii) implies i). The converse is trivial.
The following de®nition agrees with the one in [WZ], p. 639.
De®nition 2.2.A p1-dimensional matrix of boundary formatAAV0n nVp
is called diagonalizable if one of the following equivalent conditions holds:
i) there exist bases inVj such thatai0;...;ip0 fori00Pp
t1it;
ii) there exist a vector space U of dimension 2, a subgroup CHSL U and iso- morphismsVjFSkjUsuch thatAis a ®xed point of the induced action ofC. The following de®nition agrees with the one in [WZ], p. 639.
De®nition 2.3.A p1-dimensional matrix of boundary formatAAV0n nVp
is an identity if one of the following equivalent conditions holds:
i) there exist bases inVj such that
ai0;...;ip 0 fori00Pp
t1it 1 fori0Pp
t1it;
ii) there exist a vector space Uof dimension 2 and isomorphisms VjFSkjU such thatAbelongs to the unique one-dimensional SL U-invariant subspace ofSk0U nSk1Un nSkpU.
The equivalence between i) and ii) follows easily from the following remark: the matrixAsatis®es the condition ii) if and only if it corresponds to the natural multi- plication map Sk1Un nSkpU!Sk0U (after a suitable isomorphism UFU has been ®xed).
From now on, we consider the natural action of SL V0 SL Vp on P V0n nVp. We may suppose pX2. The de®nitions of triangulable, diago- nalizable and identity apply to elements ofP V0n nVpas well. In particular all identity matrices ®ll a distinguished orbit in P V0n nVp. The hyper- determinant of elements of V0n nVp was introduced by Gelfand, Kapranov and Zelevinsky in [GKZ]. They proved that the dual variety of the Segre product P V0 P Vp is a hypersurface if and only if kjWP
i0jki for j0;. . .;p (which is obviously true for a matrix of boundary format). When the dual variety is a hypersurface, its equation is called the hyperdeterminant of format k01
kp1 and denoted by Det. The hyperdeterminant is a homogeneous polynomial function over V0n nVp so that the condition DetA00 is meaningful for AAP V0n nVp. The function Det is SL V0 SL Vp-invariant, in
particular if DetA00 thenAis semistable for the action of SL V0 SL Vp.
We denote by Stab AHSL V0 SL Vpthe stabilizer subgroup ofAand by Stab A0 its connected component containing the identity. The main results of this section are the following.
Theorem 2.4. Let AAP V0n nVp of boundary format such that DetA00.
Then
A is triangulable , A is not stable for the action of SL V0 SL Vp:
Theorem 2.5.Let AAP V0n nVpbe of boundary format such thatDetA00.
Then
A is diagonalizable ,Stab Acontains a subgroup isomorphic toC: We state the following theorem only in the case p2, although we believe that it is true for all pX2. We point out that in particular dim Stab AW3 which is a bound independent ofk0,k1,k2.
Theorem 2.6. Let AAP V0nV1nV2 of boundary format such that DetA00.
Then there exists a2-dimensional vector space U such thatSL Uacts over ViFSkiU and according to this action on V0nV1nV2 we have Stab A0HSL U.Moreover the following cases are possible:
Stab A0F
0 (trivial subgroup) C
C
SL 2 (this case occurs if and only if A is an identity).
8>
>>
<
>>
>:
Remark.WhenAis an identity then Stab AFSL 2.
Let Xj be the ®nite set f0;. . .;jg. We setB:Xk1 Xkp. Aslice (in theq- direction) is the subsetf a1;. . .;apAB:aqkgfor somekAXq. Two slices in the same direction are calledparallel. Anadmissible pathis a ®nite sequence of elements a1;. . .;apAB starting from 0;. . .;0, ending with k1;. . .;kp, such that at each step exactly one ai increases by 1 and all other remain unchanged. Note that each admissible path consists exactly ofk01 elements.
Tom Thumb's Lemma 2.7.Put a mark(or a piece of bread)on every element of every admissible path.Then two parallel slices contain the same number of marks.
Proof.Any admissible pathPcorresponds to a sequence ofk0integers between 1 and p such that the integer i occurs exactlyki times. We call this sequence the code of the path P. More precisely the j-th element of the code is the integerisuch that ai
increases by 1 from the j-th element of the path to the j1-th element. The occurrences of the integer i in the code divide all other integers di¨erent from i appearing in the code into ki1 strings (possibly empty); each string encodes the part of the path contained in one of theki1 parallel slices. The symmetric group Ski1acts on the setAof all the admissible paths by permuting the strings. LetPjithe number of elements (marks) of the pathPAAon the sliceai j. In particular for all sASki1 we have
X
PAA
Pji X
PAA
sPji X
PAA
Psiÿ1 j;
which proves our lemma.
We will often use the following well-known lemma.
Lemma 2.8. If OXk !f F is a morphism of vector bundles on a variety X with kW rankF f and cj F00for some jXf ÿk1,then the degeneracy locus Dk f fxAXjrank fxWkÿ1gis nonempty of codimensionWf ÿk1.
Proof.Suppose thatDk f q. Then consider the projectionXPkÿ1!p Xand let Hbe the pullback of the hyperplane divisor according to the second projection. The natural composition
O!pOknH !pFnH
gives a section of pFnH without zeroes, hence pFnH has a trivial line sub- bundle. It follows
0cf pFnH pcf F pcfÿk1 F Hkÿ1;
which is a contradiction because 1;. . .;Hkÿ1 are independent modulopH X;C.
We getDk f0 qand the result follows from the Theorem 14.4 (b) of [Fu].
A square matrix with a zero left-lower submatrix with the NE-corner on the diag- onal has zero determinant. The following lemma generalizes this remark to multi- dimensional matrices of boundary format.
Lemma 2.9. Let AAV0n nVp. Suppose that in a suitable coordinate system there is b1;. . .;bpABsuch that ai0...ip0for ikWbk kX1and i0Xb0:Pp
t1bt. ThenDetA0.
Proof.The submatrix ofAgiven by elementsai0...ipsatisfyingikWbk (kX1) gives on X Pb2 Pbp the sheaf morphism
OXb11!OX 1;. . .;1b0
whose rank by Lemma 2.8 drops on a subvariety of codimensionWb0ÿb1 Pp
t2btdimPb2 Pbp; hence there are nonzero vectorsviAVifor 1WiWp such thatA v1n nvp 0 and then DetA0 by Theorem 3.1 of Chapter 14 of [GKZ].
Lemma 2.10. Let pX2 and aji be integers with 0WiWp, 0WjWki satisfying the inequalities aj0Xa0j1 for 0WjWk0ÿ1,ajiWaj1i for i>0, 0WjWkiÿ1 and the linear equations
Xki
j0
aji0 for0WiWp a0Tp
t1biab11 abpp0 for all b1;. . .;bpAB:
Then there is N AQsuch that
ai0N k0ÿ2i; aijN ÿkj2i j>0:
Moreover NAZif at least one kjis not even,and2NAZif all the kj are even.
Proof.If 1WsWpandbsX1 we have the two equations a0Tp
t1btab11 abss abp
p 0;
a0Tp
t1btÿ1ab11 abssÿ1 abpp0:
Subtracting we obtain a0Tp
t1btÿa0Tp
t1btÿ1 ÿ abssÿabssÿ1;
so that the right-hand side does not depend ons.
Moreover for pX2 from the equations a0Tp
t1bt ab11 abqq1 abssÿ1 abpp0;
a0Tp
t1btÿ1ab11 abq
q abss abp
p0
we get
abqq1ÿabqqabssÿabssÿ1;
which implies that the right-hand side does not depend onbseither. Letabssÿabssÿ1 2NAZ. Thenatsa0s2Ntfort>0,s>0. By the assumptionPks
t0ats0 we get ks1a0s2NXks
t1
t0;
that is
a0s ÿksN:
The formulas for ais anda0i follow immediately. If some ks is odd we have 2NAZ andksNAZso thatNAZ.
Proof of Theorem2.4. IfAis triangulable it is not stable. Conversely suppose Anot stable and denote byAagain a representative ofAinV0n nVp. By the Hilbert±
Mumford criterion there exists a 1-parameter subgroup l:C!SL V0 SL Vpsuch that limt!0l tAexists. Let
a0sW Wakss; 0WsWp
be the weights of the 1-parameter subgroup of SL Vsinduced byl; with respect to a basis consisting of eigenvectors the coordinate ai0...ip describes the eigenspace of l whose weight isai00a1i1 aipp. Recall that
Xki
j0
asi0; 0WsWp:
We note that for all b1;. . .;bkABwe have a0Tp
t1bta1b1 abppX0; 2:1
otherwise the coe½cient ai0...ip is zero for ikWbk, 1WkWp and i0XPp
t1bt and Lemma 2.9 implies DetA0. The sum on all b1;. . .;bkAB for any admissible path of the left-hand side of (2.1) is nonnegative. The contribution ofat's in this sum is zero by Lemma 2.7. Also the contribution ofa0's is zero because it is zero on any admissible path. It follows that
a0Tp
t1btab11 abpp0 for all b1;. . .;bkAB;
and by Lemma 2.10 we get explicit expressions for the weights which imply thatAis triangulable.
Proof of Theorem2.5. Again we denote byAany representative ofAinV0n n Vp. IfAis diagonal in a suitable basise 0i0 n ne ipp, we construct a 1-parameter subgroup l:C!SL V0 SL Vp by the equation l te 0i0 n ne pip :
ti0ÿTt1p ite 0i0 n ne pip , so that CHStab A. Conversely let CHStab A. By Theorem 2.4, A is triangulable and by Lemma 2.9 all diagonal elements ai0...ip
with i0Pp
t1it are nonzero. We can arrange the action on the representative in order that the diagonal corresponds to the zero eigenspace. Then the assumption
CHStab Aand the explicit expressions of the weights as in the proof of Theorem 2.4 show thatAis diagonal.
We will prove Theorem 2.6 by geometric arguments at the end of Section 6.
3 Preliminaries about Steiner bundles
De®nition 3.1.A Steiner bundle overPnP Vis a vector bundleSwhose dualS appears in an exact sequence
0!S!WnO!fA InO 1 !0 3:1
whereWandIare complex vector spaces of dimensionnkandkrespectively.
A Steiner bundle is stable ([BS], Theorem 2.7 or [AO], Theorem 2.8) and is invariant by small deformations ([DK], Corollary 3.3). Hence the moduli spaceSn;k of Steiner bundles de®ned by (3.1) is isomorphic to an open subset of the Maruyama moduli scheme of stable bundles. On the other handSn;k is also isomorphic to the GIT-quotient of a suitable open subset of P Hom W;InV for the action of SL W SL I(see Section 6). It is interesting to remark that these two approaches give two di¨erent compacti®cations of Sn;k, but we do not pursue this direction in this paper. For other results aboutP Hom W;InV, see [EH] and [C].
De®nition 3.2. Let SASn;k be a Steiner bundle. A hyperplane H AP V is an unstable hyperplane ofSifh0 SjH 00. The setW Sof the unstable hyperplanes is the degeneracy locus overP Vof the natural mapH1 S ÿ1nO!H1 Sn O 1, hence it has a natural structure of scheme. W S is called the scheme of the unstable hyperplanes ofS. Note that sinceh0 SjH W1 ([V2]) the rank of the previous map drops at most by one.
3.3. Let us describe more explicitly the map H1 S ÿ1nO!H1 SnO 1.
From (3.1) it follows thatH1 S ÿ1FI andH1 SF VnI=W. The projec- tionVnI !B VnI=W can be interpreted as a mapVnH1 S ÿ1 !H1 S which induces onP Vthe required morphismH1 S ÿ1nO!H1 SnO 1.
For a generic S, W S q. Examples show that W Scan have a nonreduced structure.
We recall that ifDis a divisor with normal crossing then W logDis the bundle of meromorphic forms having at most logarithmic poles overD. IfHis the union of m hyperplanes Hi with normal crossing, it is shown in [DK] that for mWn1, W logH splits while for mXn2 we haveSW logHASn;k where kmÿ nÿ1.
The following is a simple consequence of [BS], Theorem 2.5.
Proposition 3.4.Let SASn;k,then
h0 S t 0,tWkÿ1:
Proof. S tF 5nÿ1S ÿkt. The5nÿ1-power of the sequence dual to (3.1) is 0!Snÿ1InO ÿn1ÿkt !Snÿ2InWnO ÿn2ÿkt !
!5nÿ1WnO ÿkt !5nÿ1S ÿkt !0;
and from this sequence the result follows.
Let us ®x a basis in each of the vector spaces W and I. Then the morphism fA in (3.1) can be represented by a k nk matrix A (it was called MA in the introduction, see (1.1)) with entries inV. In order to simplify the notations we will use the same letterAto denote also its class inP Hom W;InV.Ahas rankkat every point ofP V. Two such matrices represent isomorphic bundles if and only if they lie in the same orbit of the action of GL W GL I.
3.5.In particularH0 S tidenti®es with the space of nk 1-column vectorsv with entries inStVsuch that
Av0: 3:2
MoreoverHAW S(as closed point) if and only if there are nonzero vectorsw1of size nk 1 andi1 of sizek1 both with constant coe½cients such that
Aw1i1H: 3:3
This means thatw1is in the kernel of the mapWFH0 WnOH !H0 InOH 1:
3.6. According to the theorem stated in the introduction AAHom W;VnI has nonzero hyperdeterminant if and only if it corresponds to a vector bundle. The locus inP Hom W;VnIwhere the hyperdeterminant vanishes is an irreducible hyper- surface of degreek nk
k
([GKZ], Chapter 14, Corollary 2.6). It is interesting to remark that Proposition 3.4 can be proved also as a consequence of [GKZ], Chapter 14, Theorem 3.3.
3.7.The above description has a geometrical counterpart. HereP Vis the projective space of lines inV, dual to the usual projective spacePof hyperplanes inV. Consider inP VnIthe varietyXr corresponding to elements ofVnI of rankWr. In par- ticularX1 is the Segre varietyP V P I. Letmmin n;kÿ1so thatXm is the variety of non maximum rank elements. ThenAAHom W;VnIde®nes a vector bundle if and only if it induces an embeddingP WHP VnIsuch that at every smooth point of XmVP W, P W and Xm meet transversally. This follows from [GKZ], Chapter 14, Propostion 3.14 and Chapter 1, Proposition 4.11.
3.8.W Shas the following geometrical description. Let pV be the projection of the Segre varietyP V P Ion theP V. Then
W Sred pVP WV P V P Ired
(according to the natural isomorphismP V P V). In fact i1H in formula (3.3) is a decomposable tensor inVnI.
3.9.About the scheme structure we remark thatW Sis the degeneration locus of the morphismInOP V!VnI
W nOP V 1. The following construction is standard.
The projective bundle PP InOP V !p P V is isomorphic to the Segre varietyT P V P I P V P IandOP 1FOT 0;1. The morphism
C!VnI
W nVnI de®nes a section ofOT 1;1nVnI
W with zero locusZTVP W. Now assume that dimW S 0, hence dimT0. By applyingp to the exact sequence
OTn VnI W
!OT 1;1 !OZ!0
we get that the structure sheaf ofW Sis contained inpOZ. We do not know if the equality always holds. In particular if Z is reduced also W Sis reduced. We will show in Proposition 6.5 that a multiple point occurs inZi¨ it occurs inW S.
Theorem 3.10. Let SASn;k be a Steiner bundle. Then any set of distinct unstable hyperplanes of S has normal crossing.
Proof. We ®x a coordinate systemx0;. . .;xn on Pn and a basise1;. . .;enk ofW.
Let Abe a matrix representing S. If the assertion is not true, we may suppose that W S contains the hyperplanes x00;. . .;xj0, Pj
i0xi0 for some j such that 1WjWnÿ1. By (3.3) there are c0AW,b0 AI such that Ac0b0x0. We may suppose that the ®rst coordinate of c0 is nonzero, hence A c0;e2;. . .;enk
b0x0;. . . A0.
The matrix A0still representsS, hence by (3.3) there arec1AW,b1AI such that A0c1b1x1. At least one coordinate ofc1after the ®rst is nonzero, say the second. It follows thatA0 e1;c1;e3;. . .;enk b0x0;b1x1;. . . A00and againA00represents S. Proceeding in this way we get in the end that
b0x0;. . .;bjxj;. . .
is a matrix representingS, which we denote again byA.
By (3.3) there arec c1;. . .;cnkt AW,bAIsuch thatAc b0x0;. . .bjxj;. . .
cbPj
i0xi.
Now we distinguish two cases. If ci0 for iXj2 we get bc1b0c2b1
cj1bj, that is the submatrix of A given by the ®rst j1 columns has generically rank one. If we take the k nkÿj matrix which has bj as ®rst column and the lastnkÿjÿ1 columns ofAin the remaining places, we obtain a morphism
Ok!OlO 1nkÿjÿ1;
which by Lemma 2.8 has rankWkÿ1 on a nonempty subschemeZofPn. It follows that alsoAhas rankWkÿ1 onZ, contradicting the assumption thatSis a bundle.
So this case cannot occur.
In the second case there exists a nonzero ci for someiXj2, we may suppose cj200. Then the matrix
A0A e1;. . .;ej1;c;ej3;. . .;enk b0x0;. . .bjxj;bXj
i0
xi. . .
" #
representsS.
The lastnkÿjÿ2 columns ofA0 de®ne a sheaf morphismOk!O 1nkÿjÿ2 on the subspace Pnÿjÿ1 fx0 xj 0g and again by Lemma 2.8 we ®nd a point where the rank ofAisWkÿ1. So neither case can occur.
Proposition 3.11. Let SASn;k and let x1;. . .;xsAW S, sWnk. There exists a matrix representing S whose ®rst s columns are b1x1;. . .;bsxs, where the bi are vectors with constant coe½cients of size k1. Moreover any p columns among
b1;. . .;bs with pWk are independent. Conversely if the ®rst s columns of a matrix
representing S have the formb1x1;. . .;bsxsthenx1;. . .;xsAW S.
Proof. The last assertion is obvious. The proof of the existence of a matrixArepre- sentingShaving the required form is analogous to that of Theorem 3.10. Then it is su½cient to prove that b1;. . .;bp are independent. Suppose Pp
i1bili0. Letx Qp
i1xi. Letcbe the nk 1 vector (whith coe½cients inSpÿ1V) whosei-th entry
is lix=xi for i1;. . .;p and zero otherwise. It follows that AcxPp
i1bili0 and by (3.2) we get a nonzero section of S pÿ1, which contradicts Proposition 3.4.
3.12 Elementary transformations.ConsiderH fx0gAW S. The mapOH!SjH induces a surjective mapS!OHand an exact sequence
0!S0!S!OH!0 3:4
(see also [V2], Theorem 2.1); it is easy to check (e.g. by Beilinson's theorem) that S0ASn;kÿ1. According to [M] we say that S0 has been obtained from S by an
elementary transformation. By Proposition 3.11 there exists a matrixArepresenting Sof the following form
A
x
0
... A0
0 2 66 64
3 77
75 3:5
whereA0is a matrix representingS0. Sinceh0 SjH W1,S0is uniquely determined by SandH.
Theorem 3.13. With the above notations we have the inclusion of schemes W SH W S0UH.In particular we have:
i) lengthW S0XlengthW S ÿ1;
ii) ifdimW S0 0thenmultHW S0XmultHW S ÿ1,so that if H is a multiple point of W S,then H AW S0;
iii) ifdimW S0 0then for any hyperplane K0H, multKW S0XmultKW S.
Proof.The sequence dual to (3.4)
0!S!S0!OH 1 !0 gives the commutative diagram onP V:
0 ! O???y ! H1 S ÿ1nO ! H1 S0 ÿ1nO ! 0
??
?y
??
?y
0 ! H0 OH 1nO 1 ! H1 SnO 1 ! H1 S0nO 1 ! 0 It follows that the matrixB0of the map
H1 S0 ÿ1nO!H1 S0nO 1
can be seen as a submatrix of the matrixBof the map H1 S ÿ1nO!H1 SnO 1:
In a suitable system of coordinates:
B
y1
...
yn 0 B0 2 66 66 4
3 77 77
5 3:6
where y1;. . .;ynis the ideal ofH(in the dual space). It follows that I W S0 y1;. . .;ynHI W S;
which concludes the proof.
4 The Schwarzenberger bundles
Let U be a complex vector space of dimension 2. The natural multiplication map Skÿ1UnSnU!Snkÿ1U induces the SL U-equivariant injective map Snkÿ1U!Skÿ1UnSnUand de®nes a Steiner bundle onP SnUFPnas the dual of the kernel of the surjective morphism
OP SnUnSnkÿ1U!OP SnU 1nSkÿ1U:
It is called a Schwarzenberger bundle (see [ST], [Schw]). Let us remark that in the correspondence between Steiner bundles and multidimensional matrices mentioned in the introduction, the Schwarzenberger bundles correspond exactly to the identity matrices (see De®nition 2.3).
By interchanging the role of Skÿ1U andSnU we obtain also a Schwarzenberger bundle onP Skÿ1UFPkÿ1as the dual of the kernel of the surjective morphism
OP Skÿ1UnSnkÿ1U!OP Skÿ1U 1nSnU:
Both the above bundles are SL U-invariant. We sketch the original Schwarzen- berger construction for the ®rst one. The diagonal mapu7!unand the isomorphism P SnUFPn detect a rational normal curve P U CnHPn. In the same way a second rational normal curve P U Cnkÿ1 arises in P Snkÿ1U. We de®ne a morphism
P SnU SnP U !Gr Pnÿ1;P Snkÿ1U
npoints inP U 7!Span of npoints inCnkÿ1:
The pullback of the dual of the universal bundle on the Grassmannian is a Schwarzenberger bundle.
It is easy to check that if S is a Schwarzenberger bundle then W S CnH P SnU(the dual rational normal curve). See e.g. [ST], [V1].
This can be explicitly seen from the matrix form given by [Schw], Proposition 2
MA
x0 . . . xn
... ...
x0 . . . xn
2 64
3
75: 4:1
Let t1;. . .;tnk be any distinct complex numbers. Let wbe the nk nk
Vandermonde matrix whose i;jentry ist iÿ1j ; the i;j-entry of the productMAwis t iÿ1j Pn
k0xktjk; hencefPn
k0xktk 0gAW Sfor alltACby Proposition 3.11.
On the other handW Sis SL U-invariant; if it were strictly bigger thanCn then it would contain the hyperplane H fx0x10g, which lies in the next SL U- orbit; now equation (3.3) implies immediately thatw1i10.
In Theorem 5.13 we will need the following result.
Lemma 4.1. Let S be a Schwarzenberger bundle and let x0;. . .;xn be coordinates inP Vsuch that S is represented(with respect to suitable basis of I and W)by the matrix MAin 4:1.Let y0;. . .;ynbe dual coordinates inP V.Then the morphism H1 S ÿ1nO!H1 SnO 1(with respect to the obvious basis)is represented by the matrix
B
y1 ÿy0 y1 ÿy0
... ...
y1 ÿy0 y2 0 ÿy0
... ... ...
y2 0 ÿy0 y2 ÿy1
y3 0 0 ÿy0
... ... ... ...
2 66 66 66 66 66 66 66 66 66 66 4
3 77 77 77 77 77 77 77 77 77 77 5 :
Proof.By (3.3) it is enough to check that the composition W !A VnI!B VnI=W is zero, which is straightforward.
Theorem 4.2([Schw], Theorem 1, see also [DK], Proposition 6.6).The moduli space of Schwarzenberger bundles is PGL n1=SL 2, which is the open subscheme of the Hilbert scheme parametrizing rational normal curves.
In particularW Suniquely determinesSin the class of Schwarzenberger bundles.
5 A ®ltration ofSn;kand the Gale transform of Steiner bundles De®nition 5.1.
Sn;ki : fSASn;kjlengthW SXig:
In particular
Sn;kSn;0kISn;k1 I :
We will see (Corollary 5.5) thatSn;yk corresponds to Schwarzenberger bundles.
Each Sn;ki is invariant for the action of SL Von Sn;k. We will see in Section 6 that all the points of Sn;k are semistable (in the sense of Mumford's GIT) for the action of SL V.
LetSbe the open subset ofP Hom W;VnIrepresenting Steiner bundles. The quotientSn;k=SL Vis isomorphic toS=SL W SL I SL V.
By interchanging the role of V and I, also Skÿ1;n1=SL I turns out to be iso- morphic toS=SL W SL I SL V, so that we obtain an isomorphism
Sn;k=SL n1FSkÿ1;n1=SL k:
For anyEASn;k=SL n1we will call theGale transformofEthe corresponding class inSkÿ1;n1=SL kand we denote it byEG. In [DK] the above construction is called association. Here we follow [EP]. Our Gale transform is a generalization of the one in [EP]. In fact in the caseink1 Eisenbud and Popescu in [EP] review the classical association between PGL n1-classes of nk1 points of Pn in general position and PGL k-classes ofnk1 points ofPkÿ1 in general position and call it Gale transform. If we take the union H of nk1 hyperplanes with normal crossing in Pn (as points in the dual projective space) the Gale transform (as points in the dual projective space)HGconsists of a PGL k-class of nk1 hyperplanes with normal crossing in Pkÿ1. As remarked in [DK], W logHGF
W logHG. That is, the Gale transform in our sense reduces to that in [EP] when the Steiner bundles are logarithmic. It is also clear that the PGL-class of Schwar- zenberger bundles overP Vcorresponds under the Gale transform to the PGL-class of Schwarzenberger bundles overP I.
We point out that one can de®ne the Gale transform of a PGL-class of Steiner bundles but it is not possible to de®ne the Gale transform of a single Steiner bundle.
This was implicit (but not properly written) in [DK]. Nevertheless by a slight abuse we will also speak about the Gale transform of a Steiner bundleS, which will be any Steiner bundle in the class of the Gale transform ofSmod SL n1.
The following elegant theorem due to Dolgachev and Kapranov is a ®rst beautiful application of the Gale transform.
Theorem 5.2([DK], Theorem 6.8).Any SASn;2is a Schwarzenberger bundle.
Proof.
Sn;2=SL n1FS1;n1=SL 2;
and it is obvious that a Steiner bundle on the lineP1is Schwarzenberger.
Theorem 5.3. Two Steiner bundles having in common nk1 distinct unstable hyperplanes are isomorphic.
Proof. We prove that ifSis a Steiner bundle such that the hyperplanesfxi0g for
i1;. . .;nk1 belong toW S, thenSis uniquely determined. By Proposition
3.11 there exist column vectors aiACk such that S is represented by the matrix
a1x1;. . .;ankxnk. Moreover by (3.3) there arebACnk andcACk such that
a1x1;. . .;ankxnkbcxnk1:
We claim thatallthe components ofbare nonzero. The last formula can be written
a1b1;. . .;ankbnk;ÿc x1;. . .;xnk1t 0
where in the right matrix we identify xi with the n1 1 vector given by the coordinates of the corresponding hyperplane. We may suppose that there exists s with 1WsWnkÿ1 such that bi0 for 1WiWs and bi00 for s1WiW nk. IfsXk, it follows thatn1 hyperplanes among thexihave a nonzero syzygy, which contradicts Proposition 3.11. HencesWkÿ1 and we have
as1bs1;. . .;ankbnk;ÿc xs1;. . .;xnk1t0:
The rank of the right matrix isn1, hence the rank of the left matrix isWkÿs, in particular the ®rstkÿs1 columns are dependent and this contradicts Proposition 3.11. This proves the claim.
In particulara1;. . .;ank;ÿc B0 where
BDiag b1;. . .;bnk;1 x1;. . .;xnk1t
is a nk1 n1matrix with constant entries of rank n1. Therefore the matrixa1;. . .ank;ÿcis uniquely determined up to the (left) GL k-action, which implies thatSis uniquely determined up to isomorphism.
Corollary 5.4. A Steiner bundle is logarithmic if and only if it admits at least nk1unstable hyperplanes.
Proof.In factHHW W logHby formula (3.5) of [DK] and Proposition 3.11.
Corollary 5.5([V2], Theorem 3.1]).A Steiner bundle is Schwarzenberger if and only if it admits at least nk2unstable hyperplanes.In particularSn;ky coincides with the moduli space of Schwarzenberger bundles.
Proof. Let S be a Steiner bundle, and H AW S. Let us consider the elementary transformation (3.12)
0!S0!S!OH!0
where S0ASn;kÿ1; by Theorem 3.13, S0 has nk1 unstable hyperplanes. Pick- ing H0AW S0 and repeating the above procedure after kÿ2 steps we reach a S kÿ2ASn;2; by Theorem 5.2,S kÿ2is a Schwarzenberger bundle. In particular the remainingn4 unstable hyperplanes lie on a rational normal curve. It is then clear that any subset ofn4 hyperplanes inW Slies on a rational normal curve. Since there is a unique rational normal curve through n3 points in general position, it follows that W S is contained in a rational normal curve, so that S is a Schwar- zenberger bundle by Theorem 5.3.
Theorem 5.6.Let nX2,kX3.
i) Sn;ki for0WiWnk1is an irreducible unirational closed subvariety ofSn;k of dimension kÿ1 nÿ1 kn1 ÿi nÿ1 kÿ2 ÿ1.
ii) Sn;knk1contains as an open dense subset the variety of Steiner logarithmic bundles which coincides with the open subvariety of Symnk1Pn4 consisting of hyper- planes inPnwith normal crossing.
Proof.(ii) follows from Theorem 5.3.
The irreducibility in (i) follows from the geometric construction 3.8. The numerical computation in (i) is performed (for iWnk) by adding i nkÿ1(moduli of i points inP VnP I) ton kÿ1 nkÿi(dimension of Grassmannian of linear Pnkÿ1inP VnIcontaining the span of the aboveipoints) and subtractingk2ÿ1 (dim SL I).
Remark 5.7.In the case n;k 2;3the generic Steiner bundle is logarithmic (this was remarked in [DK], 3.18). In fact the genericP4 linearly embedded in P8 meets the Segre varietyP2P2 in degP2P26nk1 points.
Remark.The dimension ofSn;ki =SL n1is equal to nk1ÿi kÿ2 nÿ1
ÿ1 n kÿ1forkX3,nX2, 0WiWnk1 and it is 0 foriXnk2.
5.8. Corollary 5.5 implies the following property of the Segre variety: if a generic linearP WmeetsP V P Iinnk2 points, thenP Wmeets it in in®nitely many points.
Theorem 5.9. Consider a nontrivial (linear) action of SL 2 SL U over Pn. If a Steiner bundle isSL 2-invariant then it is a Schwarzenberger bundle andSL Uacts overPnP SnU.HenceSn;ky is the subset of the ®xed points of the action ofSL 2
onSn;k.
Proof. By Theorem 2.4 there exists a coordinate system such that all the entries (except the ®rst) of the ®rst column of the matrix representing the Steiner bundle S are zero. By Proposition 3.11, W S is nonempty. By the assumption W S is
SL 2-invariant and closed; it follows that W S is a union of rational curves and of simple points. If W S is in®nite we can apply Corollary 5.5. If W S is ®nite we argue by induction onk. We pick upHAW Sand we consider the elementary transformation 0!S0!S!OH!0. We get for allgASL Uthe diagram
S !f OH
??
?yi
gS !gf OH:
Since h0 SjH W1 we get that f and gfi coincide up to a scalar multiple. We obtain a commutative diagram
0 ! S???y0 ! S ! OH ! 0
??
?yF
??
?yF
0 ! gS0 ! gS ! OH ! 0:
It follows thatS0FgS0, hence SL UHSym S0and by the inductive assumption S0 is Schwarzenberger and SL U acts over Pn P SnU. Hence W S is in®nite and we apply again Corollary 5.5.
Corollary 5.10.IfHis the union of nk1hyperplanes with normal crossing then
W W logH
H whenHdoes not osculate a rational normal curve, Cn whenHosculates the rational normal curve Cn,
(this case occurs iff W logHis Schwarzenberger).
8<
:
Proof. HHW logHby Proposition 3.11. The result follows by Theorem 5.3 and Corollary 5.5.
Corollary 5.11. Let SASn;k be a Steiner bundle.If W Scontains at least nk1 hyperplanes then for every subset HHW S consisting of nk1 hyperplanes SFW logH,in particular S is logarithmic.
Corollary 5.12 (Torelli theorem, see [DK] for kXn2 or [V2] in general). Let H and H0 be two ®nite unions of nk1 hyperplanes with normal crossing in P V
with kX3not osculating any rational normal curve.Then HH0,W logHFW logH0: