STABILITY OF PROJECTIVE POINCARE´ AND PICARD BUNDLES I. BISWAS, L. BRAMBILA-PAZ, AND P. E. NEWSTEAD Abstract. LetX beanirreduciblesmoothprojectivecurveofgenusg ≥3definedover thecomplexnumbersandletM denotethemodulispaceofstablevectorbundlesonX 9 ξ 0 of rank n and determinant ξ, where ξ is a fixed line bundle of degree d. If n and d have 0 a common divisor, there is no universalvector bundle on X×Mξ. We prove that there 2 is a projective bundle on X ×M with the property that its restriction to X ×{E} is ξ n isomorphic to P(E) for all E ∈Mξ and that this bundle (called the projective Poincar´e a bundle) is stable with respect to any polarization; moreover its restriction to {x}×M ξ J is also stable for any x ∈ X. We prove also stability results for bundles induced from 1 the projective Poincar´e bundle by homomorphisms PGL(n) → H for any reductive H. 2 ′ We show further that there is a projective Picard bundle on a certain open subset M of M for any d > n(g−1) and that this bundle is also stable. We obtain new results ] ξ G on the stability of the Picard bundle even when n and d are coprime. A . h t a m 1. Introduction [ Let X be an irreducible smooth projective curve over C of genus g ≥ 3. For any integer 3 v n ≥ 2 and any algebraic line bundle ξ of degree d on X, let M denote the moduli space ξ 1 of stable vector bundles E of rank n and degree d on X with detE := nE ∼= ξ. 3 1 V If d is coprime to n, there is a universal vector bundle U on X×M . The direct image 4 ξ . of U on M , which we will denote by W, is called a Picard bundle; we need to assume 5 ξ 0 d ≥ 2n(g −1) in order to ensure that the Picard sheaf is locally free. The stability of U 8 0 andW was proved in[3]and[5]respectively. Moreover, for any point x ∈ X, semistability : v of the restriction of U to {x}×Mξ was proved in [3], while its stability was established i X in [9]. r a In this paper we will consider the situation where n and d have a common divisor. In this case there is no universal vector bundle on X × M (see [18]; also [15]). However, ξ there is a projective Poincar´e bundle PU −→ X ×M ξ such that, for any point E ∈ M , the restriction of PU to X × {E} ⊂ X × M is ξ ξ isomorphic to the projective bundle P(E) over X that parametrizes all lines in the fibres of the stable vector bundle E on X (see section 2 for more details). For any x ∈ X, we denotebyPU therestriction ofPU to{x}×M . AlthoughU doesnotexist, onecanalso x ξ Date: January 21, 2009. 2000 Mathematics Subject Classification. 14H60, 14J60. All authors are members of the international researchgroup VBAC. The second author acknowledges the support of CONACYT grant 48263-F. 1 2 I.BISWAS, L.BRAMBILA-PAZ,ANDP. E. NEWSTEAD define an adjoint Poincar´e bundle adU. In fact, since there is a bijective correspondence between projective bundles of fibre dimension n−1 and principal PGL(n)-bundles, one can associate with any homomorphism of algebraic groups ρ : PGL(n) → H with H reductive an induced Poincar´e principal H-bundle Uρ. When H = GL(m), one can regard Uρ as a vector bundle on X ×M ; the adjoint Poincar´e bundle is a special case of this ξ construction. We construct also the projective Picard bundle PW on the Zariski-open subset M′ of M consisting of those E ∈ M for which H1(X,E) = 0. The fibre of PW over any point ξ ξ E ∈ M′ is identified with the projective space P(H0(X,E)). The construction requires no restriction on d, but PW = ∅ for d ≤ n(g −1). The moduli space M is an irreducible smooth quasiprojective variety defined over C of ξ dimension (n2−1)(g−1). It is the smooth locus of the moduli space M of (S-equivalence ξ classes of)semistable vector bundles onX ofranknanddeterminant ξ (recallthatg ≥ 3). The variety M is locally factorial (see [8]) and ξ Pic(M ) ∼= Pic(M ) ∼= Z. ξ ξ In particular, M has a unique polarization represented by a divisor Θ. It follows that ξ there is a unique notion of (slope-)stable vector bundle on M . The notion of semistable ξ and stable vector bundles extends to principal bundles (see [20], [19], [21], [1] for the definitions) and in particular to projective bundles; we shall give a direct definition for projective bundles in section 2. In section 3 we prove the following results on the stability of the projective Poincar´e bundle. Theorem 3.6. Let X be a smooth projective algebraic curve of genus g ≥ 3, n ≥ 2 an integer and ξ a line bundle on X of degree d. Let M denote the moduli space of stable ξ vector bundles on X of rank n and determinant ξ and let PU be the projective Poincar´e bundle on X ×M . Then PU is stable for all x ∈ X. ξ x Theorem 3.9. Under the hypotheses of Theorem 3.6, PU is stable with respect to any polarization on X ×M . ξ The proofs involve Hecke transformations (see [12], [14]) and use the same construc- tions as in [5]. Since the concept of stability for projective bundles agrees with that for principal PGL(n)-bundles (see Remark 2.2), these theorems can be restated in terms of principal bundles. Using a theorem of [19] concerning principal bundles and recalling that a homomorphism G → H with H reductive is irreducible if its image is not contained in any proper parabolic subgroup, we obtain Theorem 3.10. Under the hypotheses of Theorem 3.6, let PGL(n) → H be a homomor- phism with H reductive and let Uρ be the induced Poincar´e H-bundle. Then (i) Uρ is semistable for all x ∈ X; x (ii) Uρ is semistable with respect to any polarization on X ×M ; ξ STABILITY OF PROJECTIVE POINCARE´ AND PICARD BUNDLES 3 (iii) if ρ is irreducible, then Uρ is stable with respect to any polarization on X ×M . ξ For the last part of this theorem, we need to show that P(E)ρ is stable for general E ∈ M . We offer two proofs of this. The first (see Lemma 2.10) is analgebraic argument ξ based on the concept of monodromy as introduced in [6]. The second (see Remark 2.11) involves an argument of Subramanian [23] using unitary representations. When n = 2 and ρ is the adjoint representation, we give a third proof of the theorem (Theorem 3.11) using the methods of the present paper. In section 4, we define the Picard bundle PW and prove Theorem 4.4. Under the hypotheses of Theorem 3.6, suppose further that d > n(g−1). Then the projective Picard bundle PW on M′ is stable. When gcd(n,d) = 1, we can define a Picard sheaf W on M whose restriction to M′ ξ ξ is a vector bundle W′ such that P(W′) ∼= PW. As a corollary of Theorem 4.4 (Corollary 4.5), we obtain the stability of W and W′, thus extending the result of [5]. ξ Notation. We shall consistently write E (respectively P ) for the restriction of a vector x x bundle E (respectively projective or principal bundle P) on X × Z to {x} × Z. We suppose throughout that X is a smooth irreducible projective algebraic curve of genus g ≥ 3 defined over C. 2. The projective Poincar´e bundle We begin with a definition of stability for a projective bundle. Let Y be a polarized irreducible locally factorial projective variety and let Z be a Zariski-open subset of Y whose complement has codimension ≥ 2 in Y. Fix a divisor D on Y defining the polarization. For any vector bundle E on Z, we can define c (E) as a 1 divisor class on Y and write degE = [c (E).Dm−1](Y), 1 where m = dimY. Now let P be a projective bundle on Z and let P′ be a projective subbundle of the restriction of P to a Zariski-open subset Z′ of Z whose complement has codimension ≥ 2. Write q, q′ for the projections of P, P′ to Z, Z′ respectively. We have an exact sequence of vector bundles (2.1) 0 −→ TrelP′ −→ TrelP| −→ N −→ 0 q′ q Z′ 1 on P′, where TrelP′ and TrelP are the relative tangent bundles and N is the normal q′ q Z 1 bundle. The direct image N := q′(N ) is a vector bundle on Z′. (The higher direct ∗ 1 images are all 0 by (2.1) and the fact that Hi(P,TP) = 0 for i ≥ 1 and any projective space P.) Definition 2.1. The projective bundle P is stable (semistable) if, for every such P′, the condition degN > 0 (degN ≥ 0) 4 I.BISWAS, L.BRAMBILA-PAZ,ANDP. E. NEWSTEAD holds. Remark 2.2. We have adopted this form of the definition because it is the most con- venient for our purposes. Moreover, if P is the projectivization of a vector bundle V on Z, then a projective subbundle P′ defines a subbundle V′ of V over Z′. In that case, the bundle N is identified with Hom(V′,V/V′)) = (V′)∗ (V/V′). Using this it follows immediately that P is stable (semistable) if and only ifNthe vector bundle V is stable (semistable). Definition 2.1 is also equivalent to the standard definition of stability for a principal PGL(n)-bundle (see [19, Definition 4.7] or [21]). To see this, note that, if we denote also by P the PGL(n)-bundle corresponding to P, the subbundle P′ of our defini- tion corresponds to a reduction of structure group σ : Z′ → P/Q, where Q is a maximal parabolic subgroup of PGL(n). If T denotes the tangent bundle along the fibres of P/Q P/Q, then σ∗(T ) is isomorphic to the bundle N of Definition 2.1; now compare [20, P/Q Definition 1.1 and Lemma 2.1], which are stated for curves but generalize immediately to higher dimension by requiring σ to be defined on the complement of a subvariety of codimension ≥ 2. Recallthat the standard definition ofM is as a quotient π : R → M of a Zariski-open ξ ξ subset R of a Quot-scheme Q by a free action of PGL(M) for some M (see [22] or [16, Chapter 5]). In fact Q is a closed subset of Quot(OM;P), the Grothendieck scheme of quotients of OM with Hilbert polynomial P, for some positive integer M and polynomial P. There is a natural action of GL(M) on Q which descends to an action of PGL(M) and RisaZariski-openPGL(M)-invariantsubsetofQ; therestrictionoftheactionofPGL(M) to R is free and defines the quotient π. There also exists a universal quotient on X ×Q to which the action of GL(M) lifts naturally, but λI acts by multiplication by λ, so the action does not descend to PGL(M). The universal quotient restricts to a vector bundle on X×R, which, after tensoring by the pullback of some bundle O(−m) on X, becomes a vector bundle E such that E | is the stable bundle π(r) for all r ∈ R. As indicated R R X×{r} above, GL(M) acts on this bundle with λI acting by multiplication by λ, so PGL(M) acts on the associated projective bundle P(E ). The quotient PU := P(E )/PGL(M) R R is then a projective bundle whose restriction to X × {E} is isomorphic to P(E) for all E ∈ M . The uniqueness of PU as constructed in this way is a corollary of the following ξ result which we shall need later. Proposition 2.3. Let E be a vector bundle on X × Z such that the restriction of E to X × {z} is stable of rank n and determinant ξ for all z ∈ Z and let φ : Z → M be E ξ the corresponding morphism. Then the projective bundles P(E) and (id ×φ )∗(PU) are X E isomorphic. STABILITY OF PROJECTIVE POINCARE´ AND PICARD BUNDLES 5 Proof. We have a pullback diagram Y −φ→Y R (2.2) π′ π   Z −φ→E M . y yξ The vector bundles (id × φ )∗E and (id × π′)∗E on X × Y have the property that X Y R X their restrictions to X × {y} are stable and isomorphic for all y ∈ Y. If we denote by p : X ×Y → Y the natural projection, it follows that Y p (Hom((id ×φ )∗E ),(id ×π′)∗E) Y∗ X Y R X is a line bundle L on Y, and there is a natural isomorphism (2.3) (id ×φ )∗E ⊗p∗ L −→ (id ×π′)∗E. X Y R Y X Moreover GL(M) acts naturally on both (id × φ )∗E and (id × π′)∗E; in the first X Y R X case, λI acts by multiplication by λ, in the second by the identity. There is also a natural action of GL(M) on L and an induced action on the left-hand side of (2.3) which descends to PGL(M). In particular (2.3) is PGL(M)-equivariant and the same holds for (cid:3) the corresponding isomorphism of projective bundles. Now take quotients. Corollary 2.4. Suppose that π′ : R′ → M defines M as a quotient of R′ by a free ξ ξ action of PGL(M′) and • E is a vector bundle on X×R′ such that E | is the stable bundle π′(r′) for R′ R′ X×{r′} all r′ ∈ R′; • the action of PGL(M′) lifts to P(E ). R′ Then P(ER′)/PGL(M′) ∼= PU. Proof. Apply Proposition 2.3 to E to get an isomorphism R′ ∼ ∗ P(ER′) = idX ×φER′ (PU). (cid:0) (cid:1) It follows from the proof of the proposition that the isomorphism can be chosen to be PGL(M′)-equivariant. Now take quotients. (cid:3) In view of this, we shall call PU the projective Poincar´e bundle on X ×M . It should ξ be noted that it is not the same as the universal projective bundle constructed in [2], which exists on a certain open set in the moduli space of projective bundles with the appropriate topological invariants. This open set is a quotient by a finite group of the Zariski-open set in M constructed in the following lemma. ξ Lemma 2.5. There is a non-empty Zariski-open subset Z of M such that, for each ξ stable vector bundle E ∈ Z, the corresponding projective bundle P(E) does not admit any nontrivial automorphism. 6 I.BISWAS, L.BRAMBILA-PAZ,ANDP. E. NEWSTEAD Proof. LetE ∈ M besuchthattheassociatedprojectivebundleP(E)admitsanontrivial ξ automorphism τ′ : P(E) −→ P(E). The automorphism τ′ gives an isomorphism of vector bundles (2.4) τ : E −→ E ⊗L, where L is some line bundle of degree 0. From the given condition that τ′ is nontrivial it ∼ follows that L 6= O . Taking the top exterior product of both sides of (2.4) we conclude X that L⊗n ∼= O . X Suppose now that τ exists with L of order r ≥ 2 as an element of the Jacobian J(X). Choose an isomorphism L⊗r −∼=→ O and let s be the section of L⊗r corresponding to X r the constant section 1 of O . Via this isomorphism τr defines an automorphism of E, X which has the form λid since E is stable. Let σ : Y → X be the cyclic covering defined E as the subvariety of the total space of L given by the equation tr−λs = 0. Then E is the r direct image of a vector bundle V on Y of rank n and degree d; moreover V is necessarily r stable (see [13] for details of the construction; also [4, Example 3.4 and Proposition 3.6] for the case r = n). Note that σ is determined by L and that there are only finitely many choices for L (up to isomorphism). Since V depends on n 2(g(Y)−1)+1 parameters, r it follows that the stable vector bundles E of determina(cid:0)nt(cid:1)ξ arising in this way depend on at most ν parameters, where n2 ν = (g(Y)−1)+1−g r2 n2 = r(g −1)+1−g r2 n2 = −1 (g −1) < (n2 −1)(g −1) = dimM . ξ (cid:18) r (cid:19) (cid:3) This completes the proof. As an immediate consequence we have Corollary 2.6. The projective Poincar´e bundle PU on X × M does not admit any ξ nontrivial automorphism. (cid:3) Remark 2.7. Aprecise descriptionofthevarietyofstablebundles E forwhich E ∼= E⊗L for a fixed L is given in [13, Proposition 3.3]. For an analytic proof of Lemma 2.5, see [10, Theorem 1.2 and Proposition 1.6]. An algebraic proof in the coprime case is given in [3, Proposition 3.8]. In the topologically trivial case (i.e. d is divisible by n), the lemma is also a special case of [2, Proposition 2.6]. Remark 2.8. If E ∼= E ⊗ L, then we have a non-zero homomorphism L → EndE. Now suppose that E is stable. Since we are in characteristic 0, EndE is semistable of degree 0 (see [19]), so this homomorphism embeds L as a subbundle of EndE. Now ∼ ∼ EndE = O⊕adE. Hence, if L 6= O, L embeds as a subbundle of adE. We deduce from Lemma 2.5 and its proof that, if E ∈ M , then the following conditions are equivalent: ξ STABILITY OF PROJECTIVE POINCARE´ AND PICARD BUNDLES 7 • E ∈ Z, where Z is the Zariski-open subset of M consisting of those E for which ξ ∼ ∼ E = E ⊗L ⇒ L = O; • P(E) admits no non-trivial automorphism; • adE possesses no line subbundle of degree 0 (see also [10, Proposition 1.6], [3, Proposition 3.10]). Remark 2.9. It follows from Remark 2.8 (see [3, Corollary 3.11]) that, if E ∈ Z and n = 2, then adE is stable. In fact, more generally, we have Lemma 2.10. For any n and any irreducible homomorphism ρ : PGL(n) → H with H reductive, the principal H-bundle P(E)ρ is stable for general E ∈ M . ξ Proof. We recall the concept of monodromy introduced in [6]. For a stable G-bundle E G on X, its monodromy is a reductive subgroup of G (see [6, Lemma 4.13]; any irreducible subgroup is automatically reductive). Hence all stable G-bundles whose monodromy is a proper subgroup of G admit reduction of structure group to some proper reductive subgroup of G. There are countably many proper reductive subgroups G′ of G up to conjugation, for each of which dimM (G) = dimG(g −1)+dim(centre of G) X > dimG′(g −1)+dim(centre of G′) = dimM (G′), X where M (G) (resp. M (G′)) denotes the moduli space of stable G-bundles (resp. G′- X X bundles). Therefore, all stable G-bundles with monodromy a proper subgroup of G are contained in a countable union of subvarieties. In our case, we take G = PGL(n) and deduce that there exists a bundle E ∈ M such ξ that the monodromy of P(E) is PGL(n). It follows from [6, Lemma 4.13] that P(E)ρ is stable for this E and hence for general E since stability is an open property. (cid:3) Remark 2.11. For an analytic proof of this lemma, recall that the principal H-bundle given by any irreducible unitary representation of the fundamental group is stable [20, Theorem 7.1] (here unitary means that the image of the representation is contained in a maximal compact subgroup). The result follows by an argument of Subramanian [23, §3] (see also [3, proof of Theorem 2.7]). Remark 2.12. Foranyhomomorphismρ : PGL(n) → H,theinducedPoincar´eH-bundle Uρ on X×M has the property that Uρ| is isomorphic to P(E)ρ for all E ∈ M . In ξ X×{E} ξ ∼ particular adU| = adE. The bundle adU can also be constructed by noting that X×{E} the action of PGL(M) on R lifts to an action on adE , which therefore descends to a R bundle on X ×M . This bundle coincides with adU. ξ 3. Stability of Poincar´e bundles In this section, we shall prove our results on the stability of Poincar´e bundles. We begin by recalling two constructions from [5]. 8 I.BISWAS, L.BRAMBILA-PAZ,ANDP. E. NEWSTEAD ∼ Let x ∈ X and let F be a vector bundle over X of rank n such that detF = ξ(x). Let P := P(F∗) be the projective space parametrizing the hyperplanes in the fibre F . Let x x p : X ×P −→ X be the projection and ι : P ֒→ X × P the inclusion map defined by z 7−→ (x,z). We have the following diagram of homomorphisms of sheaves on X ×P 0 0   p∗(F(−x)) = p∗(F(−x)) y y   (3.1) 0 −→ E −→ p∗F −→ ι∗OP(1) −→ 0 y y k   0 −→ ι∗(Ω1P(1)) −→ Fx ⊗C ι∗OP −→ ι∗OP(1) −→ 0 y y   0 0 y y (see [5, p. 565, (5)]). Lemma 3.1. There exists an exact sequence of vector bundles (3.2) 0 −→ OP(1) −→ Ex −→ Ω1P(1) −→ 0 on P. Proof. Pulling back the left hand column of (3.1) by ι, we get 0 −→ Ω1P(1) −→ Fx ⊗C OP −→ Ex −→ Ω1P(1) −→ 0. This splits into two short exact sequences (3.3) 0 −→ Ω1P(1) −→ Fx ⊗C OP −→ K −→ 0 and (3.4) 0 −→ K −→ E −→ Ω1(1) −→ 0, x P whereKisalinebundleonP. Sincedeg(Ω1(1)) = −1,itfollowsfrom(3.3)thatdegK = 1. P (cid:3) Now (3.4) gives the result. Lemma 3.2. Let W ⊂ E be a nonzero coherent subsheaf of the vector bundle E in (3.2) x x such that • the quotient E /W is torsion-free, and x • deg(W) ≥ deg(Ex). rk(W) rk(Ex) Then W contains the line subbundle OP(1) of Ex in (3.2). STABILITY OF PROJECTIVE POINCARE´ AND PICARD BUNDLES 9 Proof. Note that, by (3.2), deg(E ) = 0. Let W ⊆ W be the first term of the Harder– x 1 Narasimhan filtration of the subsheaf W. So W is semistable, and 1 deg(W ) deg(W) 1 ≥ . rk(W ) rk(W) 1 Since deg(W)/rk(W) ≥ deg(E )/rk(E ), we have x x deg(W ) deg(E ) 1 deg(Ω1(1)) 1 x P (3.5) ≥ = 0 > = . rk(W ) rk(E ) 1−n rk(Ω1(1)) 1 x P The vector bundle Ω1(1) is stable (see [17, Chapter II Theorem 1.3.2]). Therefore, from P (3.5) it follows that there is no nonzero homomorphism from W to Ω1(1). Consequently, 1 P W1 is contained in the line subbundle OP(1) of Ex in (3.2). Since Ex/W1 is torsion-free, (cid:3) this completes the proof of the lemma. We shall apply this construction in the case where F is (0,1)-stable (see [12, Definition 8.1] or [14, Definition 5.1]). For convenience, we recall this definition: a vector bundle F on X is (0,1)-stable if, for any proper subbundle F′ of F, degF′ degF −1 < . rkF′ rkF We have ([5, Lemma 1] or [14, Lemma 5.5]) Lemma 3.3. Let F be a (0,1)-stable vector bundle of rank n and determinant ξ(x). Then the vector bundle E in (3.1) is a family of stable vector bundles of rank n and determinant ξ on X. (cid:3) Let M denote the moduli space of stable vector bundles on X of rank n and deter- ξ(x) minant ξ(x). If F ∈ M is (0,1)-stable, Lemma 3.3 gives us a morphism ξ(x) φ : P(F∗) → M . F x ξ We have Lemma 3.4. For any (0,1)-stable F ∈ M , the morphism φ is an embedding. More- ξ(x) F over the locus of (0,1)-stable vector bundles F ∈ M is a non-empty Zariski-open ξ(x) subset. Proof. For the first part, see [5, Lemma 3] or [14, Lemma 5.9]. For the second part, see [5, Lemma 2] and note that, in the last line of the proof, we do not require n and d to be coprime since g ≥ 3. (cid:3) The second construction is the reverse of the one just considered. For any vector bundle E on X of rank n and determinant ξ, take a line ℓ ⊂ E in the x fibre of E over x. Let F be the vector bundle over X that fits in the following short exact sequence of sheaves (3.6) 0 −→ F(−x) −→ E −→ E /ℓ −→ 0. x 10 I.BISWAS, L.BRAMBILA-PAZ,ANDP. E. NEWSTEAD Consider the subset H of PU consisting of pairs (E,ℓ) for which the bundle F defined x x by (3.6) is (0,1)-stable. We have a diagram p H −→ M x ξ (3.7) q  M ξ(x) y given by p(E,ℓ) = E and q(E,ℓ) = F. The map p is the restriction of the projection PU → M and its image is the set of bundles E ∈ M for which there exists a line x ξ ξ ℓ ∈ E such that the vector bundle F in (3.6) is (0,1)-stable. x Lemma 3.5. (i) H is non-empty and Zariski-open in PU and q is a morphism; x x (ii) p(H ) is non-empty and Zariski-open in M . x ξ Proof. (This is proved in the coprime case in [5, p. 566] but the proof uses a Poincar´e vector bundle, so we give full details here.) (i) We return to the construction of M as a quotient ξ π : R → M . The bundle P(E ) on R parametrizes a family of pairs (E,l), where ξ R x E ∈ M and ℓ is a line in E , and hence parametrizes sequences (3.6); the subset H′ of ξ x x P(E ) for which F is (0,1)-stable is therefore Zariski-open and is non-empty by Lemmas R x 3.3 and 3.4. The group PGL(M) acts on P(E ) with H′ as an invariant subset. We can R x x now identify PU and H with P(E ) /PGL(M) and H′/PGL(M). Moreover the map x x R x x H′ → M given by sending a point of H′ to the corresponding F is a morphism by the x ξ(x) x universal property of M ; hence q is a morphism. ξ(x) (ii) By (i), p(H ) is non-empty. It is also Zariski-open since p is the restriction of the x projection morphism PU → M . (cid:3) x ξ We are now ready to state and prove the first of our main theorems. Theorem 3.6. Let X be a smooth projective algebraic curve of genus g ≥ 3, n ≥ 2 an integer and ξ a line bundle on X of degree d. Let M denote the moduli space of stable ξ vector bundles on X of rank n and determinant ξ and let PU be the projective Poincar´e bundle on X ×M . Then PU is stable for all x ∈ X. ξ x Proof. Let P′ be a projective subbundle of the restriction of PU to a Zariski-open subset x Z′ of M with complement of codimension ≥ 2. By (3.7) and Lemma 3.5, p−1(Z′) is a ξ Zariski-open subset of H whose complement S has codimension ≥ 2, in other words x dimS ≤ dimM +n−3. ξ Since the image of q has dimension dimM , the intersection of S with the general fibre of ξ q is a closed subset of dimension ≤ n−3. It follows from this, taking account of Lemma 3.4, that there exists a (0,1)-stable bundle F ∈ M such that φ−1(Z′) has complement ξ(x) F