15 march 2005

The research of the authors was performed through the years in the realm of the DFG SCHWERPUNKT ”Globale Methode in der komplexen Geometrie”, of the EAGER EEC Project, and of the VIGONI-DAAD exchange Program. The second author is a member of G.N.S.A.G.A. of I.N.d.A.M. and gratefully acknowledges the hospitality of the Institute of Mathematics Simion Stoilow of the Rumanian Academy of Sciences, in march 2002.
A.M.S. Subject classifications: 14D06, 14J29, 11G30.
<ph f="cmbx">Fibrations of low genus, I.</ph>

Fabrizio Catanese – Roberto Pignatelli

Contents

1 Introduction

The study of fibrations f : S → B   of an algebraic surface S   over a curve B   lies at the heart of surface classification (cf. e.g. [Enr] , [BPV] ).
Denoting by g   the genus of a general fibre, any surface birational to S   is called a a ruled surface if g = 0   , and an elliptic surface if g = 1   . In Enriques' classification of algebraic surfaces play a prominent role results establishing the existence of such fibrations with g = 0 , 1   .
For surfaces of general type, instead, do play a special role the fibrations with genus g = 2   , in that the presence of such a fibration constitutes the so called standard exception to the birationality of the bicanonical map (cf. [Bom1] , [CCML] , [Cil] ).
From a birational point of view, the datum of the fibration f : S → B   is equivalent to the datum of a curve C   of genus g   over the function field C ( B )   of B   .
Apart from complications coming from the fact that C ( B )   is not algebraically closed, curves of low genus g ≤ 5   are easily described (mostly as complete intersection curves in a projective space), so one has the hope of constructing, resp. describing such fibrations as complete intersections in some projective bundle over the base curve B   .
In order to do so, one needs a fixed biregular model for the birational equivalence class of f : S → B   . The classical approach is to consider the so called relative minimal model, which is unique except for the case g = 0   (cf. [Shaf] ), where a relative minimal model is a P 1   bundle, but one has the so called elementary transformations yielding non isomorphic surfaces out of the same birational genus 0   fibration.
While all the fibres of a relative minimal fibration are ∼ = P 1   in the case g = 0   , for g = 1   Kodaira gave, as a preliminary tool for his deep investigations of elliptic surfaces, a short complete list of the possible fibres (cf. [Kod] ). In the g = 2   case a similar but too long a list was provided by Ogg in [Ogg] . For instance Bombieri ([Bom2] ) was able to prove that the genus of the bicanonical pencil of a numerical Godeaux surface cannot be two by using Ogg's classification. In this respect, we show in this paper the power of our new methods by providing a one page proof of Bombieri's result.
In order to explain our new results, let us first remark that usually the existing literature on fibrations with low g   is divided into two lines of research :
1) Papers devoted to the necessary numerical restrictions which have to be satisfied by surfaces admitting such a fibration (cf. e.g. [Hor2] , [Xia2] , [Kon1] ) 2) Papers devoted to proving existence or devoted to the classification of such fibrations. For instance, the classification of surfaces with irregularity q = 1   leads naturally to the above issue 2), because the Albanese map of such surfaces is a genus g   fibration over a curve B   of genus 1   . Even when restricting to the case p g = q = 1   , where by general results such surfaces belong to a finite number of families, the classification has shown to be quite a hard task. The first result is quite old, and due independently to Bombieri and the first named author, respectively Horikawa. Theorem ([Cat1] , [Hor2] ) Minimal surfaces with K S 2 = 2 , p g = q = 1   are the double cover of the symmetric product B ( 2 )   of the elliptic curve B : = A l b ( S )   which is their Albanese image, with a branch divisor Δ   belonging to a fixed linear system. Their moduli space is unirational of dimension 7   .
Later, Ciliberto proposed to the first author to construct surfaces with p g = q = 1   and higher K S 2   as complete intersections in projective bundles over elliptic curves. Using Atiyah's classification of vector bundles on elliptic curves, and the representation theory of the Heisenberg groups it was possible to obtain the following result Theorem ([CC1] , [CC2] ) Minimal surfaces with K S 2 = 3 , p g = q = 1   are such that the genus g   of the Albanese fibre is either g = 2   or g = 3   .
For g = 3   one has an unirational connected component of the moduli space, of dimension 5   , given by surfaces birational to a divisor Σ   in the third symmetric product B ( 3 )   .
For g = 2   our surfaces are double covers of the symmetric product B ( 2 )   branched on a divisor Δ   belonging to a fixed algebraic system, and with two 4-uple singular points on the same fibre. These surfaces exist.
In [CC1] it was posed the problem whether the second family was irreducible (and hastily conjectured that the answer should be positive).
Our main application of our structure theorem for genus g = 2   is the complete classification of the moduli space of the above surfaces. This classification, while in particular giving a negative answer to the above question, shows however the existence of an irreducible family of the main stream of the surfaces with K S 2 = 3 , p g = q = 1 , g = 2 .   Theorem  7.2  The moduli space ℳ   of minimal surfaces of general type with K S 2 = 3 , p g = q = 1   has exactly four connected components, all (irreducible) unirational of dimension 5   .
Remark 1.1. It was observed in [Cat4] that the genus g   of the Albanese fibre is a differentiable invariant of the underlying complex surface (as a consequence of Seiberg Witten theory).
Restricting oneself thus to the case g = 2   , one may ask about existence of surfaces with p g = q = 1 , g = 2 , K S 2 > 3 .   Xiao ([Xia1] ) proved that one has then the inequality K S 2 ≤ 6   , and in [Cat3] existence was proven for the cases p g = q = 1 , g = 2 , K S 2 = 4 , 5 .  
We would then like to propose the following questions to the reader as a further application of our methods.
Problem I: Do there exist surfaces with p g = q = 1 , g = 2 , K S 2 = 6   ?
Problem II: How many are the connected and irreducible components of the respective moduli spaces which correspond to surfaces with p g = q = 1 , g = 2 , K S 2 = 4 , 5 , 6   ?
We would like now to introduce our structure theorems for fibrations with genus two, respectively three (the case g = 4   will hopefully be treated in a sequel to the present article, especially with a view towards the classification of numerical Godeaux surfaces, cf. [CP] ).
Referring to the theorems cited above, let us observe that the symmetric product B ( h )   of an elliptic curve parametrizes effective divisors of degree h   on B   , and using the group structure on B   we get the Abel-Jacobi map π : B ( h ) → B   , associating to the divisor ( P 1 , . . . , P h )   the point π ( P 1 , . . . , P h ) : = P 1 + ⋅ ⋅ ⋅ + P h .   Abel's theorem shows that π   makes B ( h )   a P h − 1   bundle over B   , and in fact we get the projectivization of an indecomposable vector bundle, denoted E ( h , 1 )   by Atiyah ([Ati] ).
In the study of surfaces with p g = q = 1   this bundle should also occur for higher values of K S 2   , the problems however which have to be faced are the following ones, where we concentrate to the case g = 4   for simplicity. It is true that a nonhyperelliptic curve of genus 4   has a canonical image in P 3   which is the complete intersection of a quadric and a cubic, but in general the most interesting (or most difficult) fibrations whith g = 4   will fail to be birational to the complete intersection of two divisors in a P 3   bundle. The reason for this is that, even if we assume that the general fibre is nonhyperelliptic, the general fibre F t   will be the complete intersection of a quadric Q t   and a cubic G t   ; but only the Q t   's will produce a global divisor Q   in a P 3   bundle over B   , so that our surface S   will be birational to a surface X   contained in the threefold Q   . The second trouble will be that X   exists only for very special such relative quadrics Q   , and the third trouble will be that we shall have to take care of singularities.
The problem of singularities occurs also in the a priori easiest case, where all the fibres are hyperelliptic curves (e.g., in the case g = 2   ).
Here S   is birational to the double cover of a P 1   bundle over B   , but the branch curve Δ   has several essential singularities.
For instance, in the case g = 2   Horikawa studied the genus 2   fibrations through a standard double cover of a P 1   bundle over B   , and taking into account the singularities of the branch curve, which are usually several pairs of 4-tuple points or of (3,3) points occurring on the same fibre.
Our approach is on one hand geometrical, but ultimately provides purely algebraic structure theorems, exactly in order to circumvent the problem of how to deal with singularities.
The key idea is to give a complete algebraic description of the relative canonical algebra ℛ ( f )   of the fibration, whose local structure was investigated for g ≤ 3   by Mendes Lopes in [M-L] and whose importance for general g   was stressed in [Rei2] .
We need some terminology in order to state our two main results:
for any fibre F t   of f : S → B   we consider its canonical ring ℛ ( F t ) : = ⊕ n = 0 ∞ V n ( t ) : = ⊕ n = 0 ∞ H 0 ( F t , O F t ( n K F t ) ) ,   (here K   denotes the canonical divisor). These vector spaces fit together yielding vector bundles V n   on B   ; these vector bundles and the relative canonical algebra ℛ ( f )   are defined as follows:
ℛ ( f ) : = ⊕ n = 0 ∞ V n : = ⊕ n = 0 ∞ f * ( O S ( n ( K S − f * ( K B ) ) ) ) .   If σ n : S y m n ( V 1 ) → V n   denotes the multiplication map, an important role play also the sheaf T : = C o k e r ( σ 2 )   introduced in [CC1] , which is here shown to be the sheaf O τ   of an effective divisor τ   on the curve B   , and the extension class ξ   of the extension 0 → S y m 2 ( V 1 ) → V 2 → T 2 → 0 .   There is then a last datum, denoted by w   , which is the one which has the most geometric meaning: in the case g = 2   it determines the branch divisor Δ   (intersecting the general fibre of the P 1   bundle P ( V 1 )   in 6   points), in the case g = 3   it determines a divisor Σ   in the P 2   bundle P ( V 1 )   intersecting a general fibre ∼ = P 2   in the quartic curve Σ t   which is the canonical image of the nonhyperelliptic fibre F t   .
We can give, for the experts, the exact statements of our two structure theorems Theorem  5.13  Let f   be a relatively minimal genus 2   fibration. Then its associated 5-tuple ( B , V 1 , τ , ξ , w )   is admissible.
Viceversa, every admissible genus two 5 −   tuple is the associated 5 −   tuple of a (relatively minimal) genus 2   fibration f : S → B   , and the surface S   has the following invariants:
χ ( O S ) = deg ( V 1 ) + ( b − 1 )
K S 2 = 2 deg V 1 + deg τ + 8 ( b − 1 ) .
Two relatively minimal genus 2   fibrations having isomorphic associated 5 −   tuples are isomorphic.
Here, b   is the genus of B   , and the condition of admissibility ( 5.12 ) is an open condition which limits the singularities of the so called relative canonical model X   of f   . In fact, the key idea for the genus 2   case consists in first ignoring the possible complication of the fibres of f   by replacing each of them by their canonical model. The second idea is to replace the P 1   bundle P ( V 1 )   (over B   ) by a conic bundle C   , the relative bicanonical image of S   , which is a birational image of P ( V 1 )   inside P ( V 2 )   . In this way, the ”bad” fibres correspond to double finite covers of reducible conics, and the gained result is that the new branch divisor has no essential singularities.
Similarly we have Theorem  8.13  Consider an admissible genus three 5-tuple ( B   , V 1   , τ   , ξ   , w   ).
Then, if its associated relative canonical model X   has at most rational double points as singularities, there is one and exactly one genus 3   nonhyperelliptic fibration f : S → B   with the property that every fibre is 2   -connected and whose associated 5 −   tuple is the given one. S   is indeed the minimal resolution of the singularities of X   .
The surface S   thus obtained has invariants χ ( O S ) = deg V 1 + 2 ( b − 1 )   , K S 2 = 3 deg V 1 + deg τ + 16 ( b − 1 )   .
Finally, every genus 3   nonhyperelliptic fibration with every fibre 2 −   connected (and S   smooth) yields an admissible genus three 5-tuple.
In the somewhat shorter last section we give an application of the structure theorem for g = 3   in a case where the base curve B   is P 1   .
This case is easier than the case where the genus b   is higher since by Grothendieck's theorem every vector bundle on P 1   is a direct sum of line bundles. Our theorem, establishing a new result, namely the existence of nonhyperelliptic genus 3   fibrations for some subvarieties of the moduli spaces of surfaces with p g = 3 , q = 0 , K S 2 = 2 , 3 , 4 , 5   , should be viewed as a guidebook to the use of our structure theorems for the case where B = P 1   .
As already mentioned, we plan in a sequel to this paper to describe the case of hyperelliptic fibrations of genus g = 3   and the case of nonhyperelliptic fibrations of genus g = 4   , giving applications to other questions of surface theory.
We also hope that our present results may be found useful for a development of the arithmetic theory of curves of genus g = 2   , resp. of genus g = 3   .

2 Generalities on fibrations of surfaces to curves

Throughout this paper S ′   will be a minimal projective complex surface, f ′ : S ′ ‒ ‒ ⃗ B   a rational map onto a smooth complex curve.
f ′   is necessarily a morphism if b   := genus ( B )   is strictly positive:
when b   is equal to zero, we let β : S → S ′   be a minimal sequence of blow ups which yields a morphism f : S → B   .
Without loss of generality, we may assume that f   is a fibration, i.e., that f   has connected fibres F   of arithmetic genus g   .
We shall also assume g ≥ 2   .
Since we are mostly interested in the case where S   has non negative Kodaira dimension, we can more generally assume that we have a relatively minimal fibration f : S → B   such that the following diagram commutes:
S
β ↓ ↘ f
S ′ ‒ ‒ ⃗ f ′ B .
For every p ∈ B   , we shall denote by F p   the fibre of f   over p   , i.e., the divisor f * ( p )   .
The canonical divisor K S   restricts on each fibre to the dualizing sheaf ω F p   ; recall that, f   being a fibration, ∀ p ∈ B   , h 0 ( O F p ) = 1   and h 0 ( ω F p ) =   genus ( F p ) = : g   .
One can put together the canonical maps of the fibres (cf., e.g. [Hor1] ) through the relative canonical map, given concretely as follows: let L   be a divisor on B   such that H 0 ( S , O S ( K S + f * L ) ) → H 0 ( F p , ω F p )   is surjective ∀ p ∈ B   , and consider the rational map h : S → P ( H 0 ( S , O S ( K S + f * L ) ) × B ,   induced by the linear system | O S ( K S + f * L ) |   and the projection f   .
h   is a birational map unless every fibre is hyperelliptic, in which case h   yields a double covering of a surface Y   ruled over B   .
Example 2.1. If B = P 1   and p g ( S ) = q ( S ) = 0   , then, setting L : = F   , we obtain h : S → P g − 1 × P 1   , and every fibre of f   is mapped to P g − 1   via its canonical map.
Proof. For every p ∈ P 1   , the exact sequence 0 → ω S → O S ( K S + F ) → ω F p → 0   yields an isomorphism H 0 ( O S ( K S + F ) ) → H 0 ( ω F p )   .
Q.E.D. The previous example is particularly relevant in the case of numerical Godeaux surfaces, to which we are interested (cf. Theorem 5.1).
Let us first collect some known results on canonical maps of Gorenstein curves ( cf. [CF] , [CFHR] , [CP] , [M-L] ).
These are based on Franchetta's definition (cf. [Fra1] , [Fra2] )
Definition 2.2. An effective divisor D   on a smooth algebraic surface S   is said to be k   -connected if, whenever we write D = A + B   as a sum of effective divisors A , B > 0   , then A ⋅ B ≥ k   .
Lemma 2.3. Let F   be a 2   -connected curve of genus g ≥ 2   : then the canonical map of F   is a morphism.
Lemma 2.4. Let C   be a curve of genus g ≥ 3   , let ω   be the dualizing sheaf of C   . Then φ ω   embeds C   ⇔   C   is 3   -connected and C   is not honestly hyperelliptic (i.e., a finite double cover of P 1   induced by the canonical morphism).
Remark 2.5. If g = 3   and φ ω   embeds C   , φ ω ( C )   is a plane quartic. If C   is contained in the canonical model X   of a surface of general type, then φ ω   embeds C   if C   is 2   -connected and C   is not honestly hyperelliptic.
Lemma 2.6. Let C   be a 3   -connected genus 4   Gorenstein curve, which is not honestly hyperelliptic: then φ ω   embeds C   as a complete intersection of type ( 2 , 3 )   .
Let us now recall an important definition, of the relative canonical algebra ℛ ( f )   of the fibration f   (cf. [CC1] , [Rei2] )
Definition 2.7. Consider the relative dualizing sheaf ω S | B : = O S ( K S − f * K B ) .   Then the relative canonical algebra ℛ ( f )   is the graded algebra ℛ ( f ) : = ⊕ 0 ∞ V n ,   where V n   is the vector bundle on B   given by V n : = f * ( ω S | B n )   .
Remark 2.8. V 0 = O B   because we have connected fibres, V 1   has rank g   , V n   has rank ( 2 n − 1 ) ( g − 1 )   .
By a theorem of Fujita ([Fuj1] , [Fuj2] ) V n   is semipositive, which means that every rank 1   locally free quotient of it has non negative degree.
While V 1   is a direct sum O B q ( S ) − b ⊕ V 1 a ⊕ V 1 0   where V 1 a   is ample, V 1 0   is a direct sum of stable degree 0 bundles ℰ i   with H 0 ( ℰ i ) = 0   .
Moreover, if rank ℰ i = 1   , ℰ i   is a line bundle associated to a torsion divisor in P i c 0 ( B )   ( cf. [Zuc] ).
Notice that, by relative duality, R 1 f * ω S | B = O B   , while R 1 f * ω S | B n = 0   for n ≥ 2   because the fibration is assumed to be relatively minimal.
It follows then by Riemann-Roch that for n ≥ 2  
χ ( V n ) = χ ( ω S | B n ) = = χ ( O S ) + 1 2 n ( n − 1 ) K S 2 − 2 n ( 2 n − 1 ) ( g − 1 ) ( b − 1 ) = = χ ( O S ) + 1 2 n ( n − 1 ) ( K S 2 − 8 ( g − 1 ) ( b − 1 ) ) − 2 n ( g − 1 ) ( b − 1 ) = = 1 2 n ( n − 1 ) [ K S 2 − 8 ( g − 1 ) ( b − 1 ) ] + r a n k ( V n ) ⋅ χ ( O B ) + [ χ ( O S ) − ( g − 1 ) ( b − 1 ) ] ,  
whereas
χ ( V 1 ) = χ ( O S ) − ( 2 g − 1 ) ( b − 1 ) = = r a n k ( V 1 ) ⋅ χ ( O B ) + [ χ ( O S ) − ( g − 1 ) ( b − 1 ) ] .  
Remark 2.9. Fujita's theorem shows that
d e g ( V n ) = 1 2 n ( n − 1 ) [ K S 2 − 8 ( g − 1 ) ( b − 1 ) ] + + { χ ( O S ) − ( g − 1 ) ( b − 1 ) } ≥ 0 .  
From this follow then as corollaries the Arakelov inequality [ K S 2 − 8 ( g − 1 ) ( b − 1 ) ] ≥ 0   and the inequality { χ ( O S ) − ( g − 1 ) ( b − 1 ) } ≥ 0 .   Indeed (cf. [EV] ) V n   is moreover ample for n ≥ 2   if the fibration is not a constant moduli fibration (i.e., unless all the smooth fibres are isomorphic), and in this case [ K S 2 − 8 ( g − 1 ) ( b − 1 ) ] > 0 .   Moreover, in the inequality χ ( O S ) − ( g − 1 ) ( b − 1 ) ≥ 0 ,   equality holds if and only if f   is a holomorphic bundle (by Noether's formula, cf. [Bea] ).
Finally , one has the inequality q ( S ) ≤ b + g   , equality holding if and only if f   is a product fibration F × B → B   ([Bea] ).

3 Invariants of the relative canonical algebra

Consider now the relative canonical algebra ℛ ( f ) = ⊕ 0 ∞ V n = ⊕ 0 ∞ f * ( ω S | B n ) .  
Definition 3.1. Denote by μ n , m : V n ⊗ V m → V n + m   , respectively by S n ( V 1 ) : = S y m n ( V 1 ) = S n ( f * ( ω S | B ) ) ⟶ σ n V n = f * ( ω S | B n ) ,   the homomorphisms induced by multiplication.
We define ℒ n : = ker σ n   and T n : = c o k e r σ n   .
Remark 3.2. By Noether's theorem on canonical curves, T n   is a torsion sheaf if the general fibre of f   is non-hyperelliptic, whereas more generally, if the general fibre is non-hyperelliptic, coker μ n , m   is a torsion sheaf as long as n , m ≥ 1 .   A more precise result was recently proved by K. Konno and M. Franciosi ([Kon2] , [Fran] ): if every fibre is 1   -connected (equivalently, there is no multiple fibre), coker μ n , m   = 0 if g ≥ 2 , n , m ≥ 2 , m a x { n , m } ≥ 3 .   In the case where there are no multiple fibres, the relative canonical algebra is generated by elements of degree ≤ 3   , while in the general case it is generated by elements of degree ≤ 4   .
The previous remark shows that the two cases
  • ∙   I) The general fibre is non hyperelliptic
  • ∙   II) All the fibres are hyperelliptic.
should be treated separately.
However, in the hyperelliptic case, one has the following useful method, of splitting the relative canonical algebra into the invariant, resp. anti-invariant part.
Assume in fact for the time being that a general fibre is hyperelliptic.
Then there is a birational involution σ   on S   , whence also on S ′   . Assuming that S   is not birationally ruled, σ   acts biregularly on S ′   . Moreover, since σ   preserves the rational map f ′   , it preserves its indeterminacy locus, therefore the minimal sequence of blow-ups turning f ′   into a morphism is σ   -equivariant, hence one concludes that σ   acts biregularly on the fibration f : S → B   (and trivially on the base B   ). Therefore, each open set U = f − 1 U ′   is σ   -invariant and σ   acts linearly on the space of sections O S ( U , ω S / B n )   , which splits as the direct sum of the ( + 1 )   -eigenspace and the ( − 1 )   -eigenspace.
Accordingly, we get direct sums:
V n = V n + ⊕ V n − = f * ( ω S | B n ) + ⊕ f * ( ω S | B n ) −   and we can split the relative canonical algebra as:
ℛ ( f ) = ℛ ( f ) + ⊕ ℛ ( f ) − ,   where we now observe that ℛ ( f ) +   is a subalgebra and ℛ ( f ) −   is an ℛ ( f ) +   -module.
Remark 3.3. In the hyperelliptic case, where obviously V 1 = V 1 −   , the cokernels T n   will be bigger than in the non hyperelliptic case.
We have the following table for the ranks of V n +   and V n −   (for n ≥ 2   ):
n rank V n +   rank V n −  
even n ( g − 1 ) + 1   ( n − 1 ) ( g − 1 ) − 1  
odd ( n − 1 ) ( g − 1 ) − 1   n ( g − 1 ) + 1  
Let the general fibre of the fibration f   be hyperelliptic: then the sheaf maps S n ( V 1 ) = S n ( f * ( ω S | B ) ) ⟶ σ n V n = f * ( ω S | B n ) ,   are injective iff g = 2   and their image lies in V n +   for n   even and in V n −   for n   odd.
So, if we define T n + =   coker S n ( f * ( ω S | B ) ) → V n +   for n   even, and T n − =   coker S n ( f * ( ω S | B ) ) → V n −   for n   odd, the decomposition in invariant and anti-invariant part of the cokernels T n   is T n = T n + ⊕ V n −   for n   even, and T n = T n − ⊕ V n +   for n   odd.
The sheaves T n ±   are torsion sheaves.
We end this section by observing that in the hyperelliptic case
deg ( S y m n ( V 1 ) ) = ( n + g − 1 g ) deg ( V 1 ) = = ( n + g − 1 g ) ( χ − ( g − 1 ) ( b − 1 ) ) ,  
where χ : = χ ( O S )   .

4 Genus 2 fibrations on non minimal surfaces

For genus 2   fibrations one can assume f   to be a morphism, by virtue of the following lemma, which is a minor improvement upon a lemma of Kodaira, cf. [Hor1] . Its proof shows that a similar result would hold for higher fibre genus, but with a longer (although finite) list of exceptions.1
Lemma 4.1. A relatively minimal genus 2   fibration on a non minimal surface S   of general type occurs only for the canonical pencil of a minimal surface S ′   with K S ′ 2 = 1 , p g ( S ′ ) = 2   .
Proof. We have K S = β * K S ′ + ∑ i = 1 , . . r E i   , where each E i   is an exceptional divisor of the first kind. We have 2 = K S F = F β * K S ′ + ∑ i = 1 , . . r F E i   . Let M   be the corresponding pencil on S ′   : thus F β * K S ′ = K S ′ M ≥ 1   . On the other hand, F E i ≥ 1   , else E i   is vertical and the fibration is not relatively minimal. Whence, either r = 0   and our claim is established, or r = 1   and we have blown up a single point.
In this latter case, F E = 1   , whence M 2 = 1 , K S ′ M = 1   . Applying the index theorem, we infer that K S ′ 2 = 1   and M   is numerically equivalent to K S ′   . Since K S ′ 2 = 1   , we get that q = 0   and M ≡ K S ′ + η   where η   is a torsion divisor class.
By Noether's inequality K S ′ 2 ≥ 2 p g − 4   , follows that p g ≤ 2   . If p g = 2   , the canonical model of S ′   is a hypersurface of degree 10   in P ( 1 , 1 , 2 , 5 )   and S ′   is simply connected, whence M = K S ′   , and this is the claimed exceptional case.
If instead p g = 1   , by [Cat2] the canonical model of S ′   is a complete intersection of type ( 6 , 6 )   in P ( 1 , 2 , 2 , 3 , 3 )   and again S ′   is simply connected, whence M = K S ′   and we have a contradiction.
Finally, if p g = 0   , h 0 ( S ′ , O S ′ ( K S ′ + η ) ) ≤ 1   (cf. [Rei1] ), again a contradiction.
Q.E.D.

1 We have later discovered that this result had already been obtained by Xiao Gang ([Xia1] , prop. 4.1).

5 Genus 2 fibrations: the structure theorem

Before proceeding to a description of the building data of the relative canonical algebra of a genus 2   fibration, it might be convenient to explain the underlying geometry (cf. also [CC1] ,[CC2] ).
Let f : S → B   be a genus 2   fibration. The rank 2   vector bundle V 1 : = f * ω S | B   induces a natural factorization of f   as π ∘ φ   , where φ : S ‒ ‒ ⃗ P : = P ( V 1 ) = P r o j ( S y m ( V 1 ) )   is a rational map of degree 2   , and π : P ( V 1 ) → B   is the natural projection.
The indeterminacy locus of φ   is contained in the fibres of f   which are not 2-connected, i.e., which split as F p = ℰ 1 + ℰ 2   with ℰ 1 ℰ 2 = 1   . Then ℰ i 2 = − 1   , ℰ i   has arithmetic genus 1   and is called an elliptic cycle. We shall see that these fibres are exactly the inverse images of the points in s u p p ( T 2 )   .
The typical example is given by a fibre consisting of two smooth elliptic curves ℰ 1 , ℰ 2   meeting transversally in a point P ′   . The blow-up of the point P ′   maps isomorphically to the fibre F ′ ′   of P   over the point P ∈ B   , while the elliptic curves ℰ 1 , ℰ 2   are contracted to two distinct points of the fibre F ′ ′   . The resolution φ ~   of φ   is the composition of the contraction of ℰ 1 , ℰ 2   to two simple − 2   -elliptic singularities, with a finite double cover where the branch curve Δ   in P   contains the fibre and has two distinct 4   -tuple points on it. More complicated fibres containing elliptic tails can produce different configurations of singularities of the branching divisor of φ   : a complete list is the one given by Ogg and by Horikawa in [Ogg] ,[Hor1] .
Lemma 5.1. Let f : S → B   be a genus 2   fibration. Then
  • (1) T 2   is the structure sheaf of an effective divisor τ ∈ D i v ≥ 0 ( B )   , whose support is given by the points whose corresponding fibres of f   are not 2-connected;
  • (2) τ   determines all the torsion sheaves T n   as follows:
    T 2 n + = O n τ ⊕ ( ⊕ i = 1 n − 1 O i τ ) ⊕ 2 T 2 n + 1 − = ( ⊕ i = 1 n O i τ ) ⊕ 2 ;   In particular deg T 2 n + = n 2 deg τ   and deg T 2 n + 1 − = n ( n + 1 ) deg τ   .
Proof. By a result of M.Mendes Lopes (thm. 3.7 of [M-L] , page 53), there are two possibilities for the canonical ring of a genus 2   fibre:
  • ∙   either the curve is honestly hyperelliptic, i.e., the graded ring is isomorphic to C [ x 0 , x 1 , z ] / ( z 2 − g 6 ( x 0 , x 1 ) ) ,   where deg x 0 = deg x 1 = 1   , deg z = 3   , deg g 6 = 6   ,
  • ∙   or the fibre is not 2   -connected and the ring is isomorphic to C [ x 0 , x 1 , y , z ] / ( Q 2 , Q 6 ) ,   where deg x 0 = deg x 1 = 1   , deg y = 2   , deg z = 3   and Q 2 : = x 0 2 − λ x 0 x 1   Q 6 : = z 2 − y 3 − x 1 2 ( α 0 y 2 + α 1 x 1 4 )   .
The first case is the one where the fibre is 2   -connected. In the second case, λ ≠ 0   corresponds to the case where there are on the canonical model two distinct honestly elliptic curves (i.e., finite double covers of P 1   of genus 1   ).
Clearly, x 0 , x 1   are anti-invariant sections, while y , z   are invariant ones; by the form of the equations we see that the sheaves T 2 n +   , T 2 n + 1 −   are supported on the points whose fibres are not 2   -connected.
V 2 +   is locally generated by x 0 2 , x 0 x 1 , x 1 2 , y   .
By flatness, if t   is a uniformizing parameter for B   such that t = 0   is the equation of a special point,we can lift the relation Q 2 : = x 0 2 − λ x 0 x 1   to x 0 2 − λ x 0 x 1 + μ ( t ) y + x 0 2 ψ 0 ( t ) + x 1 2 ψ 1 ( t ) + x 0 x 1 φ ( t ) .   If μ ( t )   were identically 0   , then the graded ring of the smooth fibre would not be generated in degree 2   by x 0 , x 1   .
Therefore, by a holomorphic change of coordinates in B   , we may assume μ ( t ) = t s   for a suitable positive integer s   : we will call this integer the ”multiplicity” of our special point. We get a relation Q 2 ( t ) : = t s y − Q ′ ( x 0 , x 1 , t ) .   which shows that the stalk of T 2   on a special point P   is the principal module O B , P / ( t s )   generated by the class of y   . We have then proved the first part of the statement, for the divisor τ   of the special points counted with the above defined multiplicity. We can also choose a lifting Q 6 ( t )   of Q 6   of the form Q 6 ( t ) : = z 2 − Q 6 ′ ( x 0 , x 1 , y , t ) .   Notice that, since Q 6   is invariant, also Q 6 ( t )   must be such (otherwise we would have also an anti-invariant relation, implying that z   is a zero divisor, a contradiction).
By flatness these are all the relation of the stalk of ℛ   at P   ; we leave to the reader the straightforward computation showing that ( T 2 n + 1 − ) P   equals ⊕ i = 1 n ( O B , P / ( t i s ) ⊕ O B , p / ( t i s ) )   with minimal ordered system of generators { x 0 x 1 2 n − 2 y , x 1 2 n − 1 y , … , x 0 y n , x 1 y n } ,   and ( T 2 n + ) P   equals ( ⊕ i = 1 n − 1 ( O B , p / ( t i s ) ⊕ O B , p / ( t i s ) ) ) ⊕ O B , p / ( t n s )   with minimal ordered system of generators { x 0 x 1 2 n − 3 y , x 1 2 n − 2 y , … , x 0 x 1 y n − 1 , x 1 2 y n − 1 , y n } .   Q.E.D. We introduce now a sheaf of graded algebras, whose P r o j   yields a conic bundle over B   admitting the relative canonical model as a finite double cover.
Definition 5.2. We denote by A   the graded subalgebra of ℛ   generated by V 1   and V 2   .
A n   will denote its graded part of degree n   , and we will accordingly write A = A e v e n ⊕ A o d d   with A e v e n = ⊕ n = 0 ∞ A 2 n   A o d d = ⊕ n = 0 ∞ A 2 n + 1 .   We decompose similarly ℛ = ℛ e v e n ⊕ ℛ o d d   .
Lemma 5.3. ℛ   is isomorphic to A ⊕ ( A [ − 3 ] ⊗ V 3 + )   as a graded A −   module; moreover A e v e n   is the invariant part of ℛ e v e n   and A o d d   is the anti-invariant part of ℛ o d d   .
Proof. In the proof of lemma  5.1 we wrote the stalk of ℛ   at a special point P   as an O B , P   -algebra generated by x 0 , x 1 , y   and z   , where the x i   's are anti-invariant of degree 1   , y   and z   are invariant of respective degrees 2   and 3   . In order to achieve a unified treatment of both cases, let's write the canonical ring of a honestly hyperelliptic fibre as C [ x 0 , x 1 , y , z ] / ( y , z 2 − g 6 ) ;   the stalk of ℛ   at the corresponding point will be then still generated by x 0 , x 1 , y   and z   with the same degrees as above, and again we have relations of respective degrees 2 , 6   ..
In both case the stalk of A   is the subalgebra generated by x 0 , x 1   and y   : therefore A e v e n   is invariant and A o d d   is anti-invariant.
The relations are generated by a relation in degree 2   (that is then a relation in A   ), and an invariant relation in degree 6   , that can locally be written as z 2 − Q 6 ′ ( x 0 , x 1 , y , t )   .
Therefore each element of the stalk of ℛ   can be uniquely written as a 1 + a 2 z   , with a 1   , a 2   elements of the stalk of A   . Since z   is a local generator of V 3 +   , this proves the first part of the statement.
Since we have seen that z   is invariant, that A e v e n   is invariant and that A o d d   is anti-invariant, the second part of the statement is clear.
Q.E.D. In the next lemma we describe the structure of A   .
Lemma 5.4. There are exact sequences
0 → ( det V 1 ) 2 ⊗ S n − 2 ( V 2 ) → i n S n ( V 2 ) → A 2 n → 0 ∀ n ≥ 2 ; (1)
V 1 ⊗ ( det V 1 ) ⊗ A 2 n − 2 → j n V 1 ⊗ A 2 n → A 2 n + 1 → 0 ∀ n ≥ 1 ; (2)
where i n ( ( x 0 ∧ x 1 ) ⊗ 2 ⊗ q ) : = ( σ 2 ( x 0 2 ) σ 2 ( x 1 2 ) − σ 2 ( x 0 x 1 ) 2 ) q ,   j n ( l ⊗ ( x 0 ∧ x 1 ) ⊗ q ) : = x 0 ⊗ ( σ 2 ( x 1 l ) q ) − x 1 ⊗ ( σ 2 ( x 0 l ) q ) .  
Proof. The above maps S n ( V 2 ) → A 2 n   and V 1 ⊗ A 2 n → A 2 n + 1   , induced by the ring structure of A   , are surjective because by definition A   is generated in degree 2   . Whence, the respective kernels are locally free sheaves on B   .
Both sequences are complexes by virtue of associativity and commutativity of multiplication in ℛ   .
To verify exactness, we first consider the stalk at a special point P ∈ S u p p ( T )   . The sheaf homomorphisms i n   , j n   and σ 2   induce linear maps i n , P   , j n , P   , σ 2 , P   on the fibres over P   of the corresponding vector bundles, and ring homomorphisms i n , ( P )   , j n , ( P )   , σ 2 , ( P )   on the stalks of the corresponding sheaves.
In the proof of lemma  5.1 we wrote the relation in degree 2   of A   as t s y − Q ′ ( x 0 , x 1 , t )   with Q ′ ( x 0 , x 1 , 0 ) = − x 0 2 + λ x 0 x 1   .
Therefore u 0 : = y , u 1 : = σ 2 , ( P ) ( x 0 x 1 )   and u 2 : = σ 2 , ( P ) ( x 1 2 )   are a basis of the fibre of the stalk (and by restriction of the fibre) of V 2   at P   , and σ 2 , ( P ) ( x 0 2 ) = − t s u 0 + λ u 1 + t ⋯   .
In this basis i n , P ( ( x 0 ∧ x 1 ) 2 ⊗ q ) = ( λ u 1 u 2 − u 1 2 ) q   , thus i n , P   is clearly injective.
At a general point we choose as basis u 0 : = σ 2 , ( P ) ( x 0 2 )   , u 1 : = σ 2 , ( P ) ( x 0 x 1 )   and u 2 : = σ 2 , ( P ) ( x 1 2 )   and since i n , P ( ( x 0 ∧ x 1 ) 2 ⊗ q ) = ( u 0 u 2 − u 1 2 ) q   we derive the same conclusion.
Since i n   injects ( det V 1 ) 2 ⊗ S n − 2 ( V 2 )   in S n ( V 2 )   as a saturated subbundle, the exactness of the sequence ( 1 ) follows then from the equality rk ( ( det V 1 ) 2 ⊗ S n − 2 ( V 2 )   ) + rk ( A 2 n   ) = ( n 2 ) + 2 n + 1 =   rk ( S n ( V 2 )   ).
To show that the image of j n   contains the kernel of the projection on A 2 n + 1   , which is locally free by our previous remark, it is enough to work on the fibres of the associated vector bundles. We identify the fibre of A k   to a subspace of the canonical ring of the fibre curve. We have then to show that, given p 0 , p 1   of even degree with x 0 p 1 = x 1 p 0   , there exists a q   with p i = x i q   ( i = 0 , 1   ).
This is straightforward for the fibre of a general point, because this subring is the ring C [ x 0 , x 1 ]   .
On a special fibre our subring is the ring C [ x 0 , x 1 , y ] / ( x 0 2 − λ x 0 x 1 ) = ( C [ x 0 , x 1 ] / ( x 0 2 − λ x 0 x 1 ) ) [ y ]   . We can thus replace p 0 , p 1   by their coefficients in C [ x 0 , x 1 ] / ( x 0 2 − λ x 0 x 1 )   .
In this ring, if k ≥ 2   is the degree of the polynomials p i   , we can uniquely find constants a i , b i   with p 0 = a 0 x 1 k − 1 x 0 + b 0 x 1 k   , p 1 = a 1 x 1 k − 1 ( x 0 − λ x 1 ) + b 1 x 1 k   . With this expression of the polynomials the condition x 0 p 1 − x 1 p 0   becomes b 0 = a 0 − b 1 = 0   and it suffices to choose q = a 0 x 1 k − 1 + a 1 x 1 k − 2 ( x 0 − λ x 1 )   . Q.E.D.
Remark 5.5. The two exact sequences above show that the structure of A   as an O B   -algebra is completely determined by the map σ 2 : S 2 ( V 1 ) → V 2   .
The structure of quotient algebra of S y m ( V 2 )   for A e v e n   gives a canonical embedding of P r o j ( A )   (canonically isomorphic to P r o j ( A e v e n )   ) into the P 2 −   bundle P ( V 2 )   .
Thus C : = P r o j ( A )   is a conic bundle and we define π A : C → B   as the restriction of the natural projection π 2 : P ( V 2 ) → B   .
Then the natural morphism φ A : S → C = P r o j ( A )   induced by the inclusion A ⊂ ℛ   yields a factorization f = π A ∘ φ A .   Since the multiplication map from V 3 + ⊗ V 3 +   to ℛ 6   has image contained in A 6   by lemma  5.3 , the ring structure on ℛ   induces a map δ : ( V 3 + ) 2 → A 6 .  
Definition 5.6. Let P   be a point in the support of τ   .
We have seen in lemma  5.1 that the map σ 2 , P   has rank 2   , therefore its image gives a pencil of lines in the plane which is the fibre of P ( V 2 )   over this point.
This pencil of lines has a base point. Taking all the points thus associated to the points of supp( τ   ) we get a subset of P ( V 2 )   that we will denote by P   . Note that the projection onto B   maps P   bijectively onto supp( τ   ).
Theorem 5.7. C = P r o j ( A )   is the divisor in the linear system | O P ( V 2 ) ( 2 ) ⊗ π 2 * ( det V 1 ) − 2 |   of P ( V 2 )   whose local equation is induced by the map i 2   defined in lemma  5.4 .
C   has at most rational double points as singularities.
φ A   is the minimal resolution of the singularities of the relative canonical model X : = P r o j ( ℛ )   , a double cover of C   whose branch locus has P   as set of isolated points, whereas its divisorial part Δ A   is the curve in the linear system | O C ( 3 ) ⊗ π A * ( V 3 + ) − 2 |   determined by δ   . Δ A   is disjoint from P   .
Proof. As noticed in remark  5.5 , by the exact sequence ( 1 ) A e v e n   is the quotient of the symmetric algebra S y m ( V 2 )   by the sheaf of principal ideals image of the map i 2 : ( det V 1 ) 2 → S 2 ( V 2 )   , therefore its P r o j   is a Cartier divisor C   in the corresponding linear system.
The map φ A   factors through X = P r o j ( ℛ )   . The natural map S → X   is the contraction of the ( − 2 ) −   curves contained in a fibre of f   ; the map X → C   is, by lemma  5.3 , finite of degree 2   ; in particular, having X   only rational double points as singularities, also the singularities of C   must be isolated and rational; since C   is a divisor of the smooth 3 −   fold P ( V 2 )   they are rational hypersurface singularities, i.e., Rational Double Points (cf. [Art] ).
There remains to compute the branch locus of the double cover ψ : X → C   .
The question being local, we may assume that X   is the subscheme of P ( 1 , 1 , 2 , 3 ) × B   defined by the equations q ( x 0 , x 1 , y , t ) = 0 , z 2 = g 6 ( x 0 , x 1 , y , t ) .   Observe that x 0 = x 1 = y = 0   implies then z = 0   .
At a point where x i ≠ 0   we simply localize the two equations dividing them by x i 2   , respectively by x i 6   . Whence, z = 0   is the ramification divisor and g 6 = 0   the branch locus. At the points where x 0 = x 1 = 0   , y = 1   we must have a point of P   . This point is fixed for the involution z → − z   and g 6 ≡ y 3 + x 0 φ 5 + x 1 ψ 5 ( m o d t )   does not vanish.
Q.E.D. Theorem  5.7 shows that a genus 2   fibration is determined by a sheaf of algebras A   constructed as in lemma  5.4 , a line bundle V 3 +   and a map from its square to A 6   . In fact these data are not completely independent, as the following easy computation shows.
Let us first observe how the birational invariants of S   determine d e g ( τ )   .
The exact sequence 0 → S y m 2 ( V 1 ) → V 2 → T 2 → 0   yields, by remark  2.9 
deg τ = deg ( T 2 ) = deg V 2 − 3 deg V 1 = = K S 2 − 9 ( b − 1 ) + χ − 3 χ + 3 ( b − 1 ) = K S 2 − 6 ( b − 1 ) − 2 χ ,  
which yields the Noether-type inequality ( N o e t h e r ) K S 2 ≥ 6 ( b − 1 ) + 2 χ .   Looking at the graded part in degree 3   we get the exact sequence 0 → S y m 3 ( V 1 ) → V 3 − → T 3 → 0 ,   which, together with lemma  5.1 yields
deg V 3 − = 6 deg V 1 + 2 deg τ = = 6 χ − 6 ( b − 1 ) + 2 K 2 − 12 ( b − 1 ) − 4 χ = 2 K S 2 − 18 ( b − 1 ) + 2 χ ,  
and thus follows
deg V 3 + = deg V 3 − deg V 3 − = K S 2 − 7 ( b − 1 ) − χ = deg V 1 + deg τ .  
In fact we have:
Proposition 5.8. V 3 + ∼ = det V 1 ⊗ O B ( τ ) .  
Proof. We have decomposed the fibration f   as S → r X → ψ C → π A B ,   where r   is a contraction of ( − 2 )   -curves (and φ A = ψ ∘ r )   . Therefore
V 1 : = ( f * ω S | B ) − = ( π A ) * ( ψ * ω X | B ) − . (3)
On the other hand, since standard adjunction for finite maps yields ψ * ω X | B = ψ * ψ ! ω C | B = ℋ o m O C ( ψ * O X , ω C | B ) ,   and since by Theorem  5.7 
ω C | B = π A * det V 2 ⊗ O C ( − 3 ) ⊗ π A * ( det V 1 ) − 2 ⊗ O C ( 2 ) = = O C ( − 1 ) ⊗ π A * ( det V 1 ⊗ O B ( τ ) ) ,  
multiplying both sides of ( 3 ) by ( det V 1 ⊗ O B ( τ ) ) − 1   we get the equation
(4) V 1 ⊗ ( det V 1 ⊗ O B ( τ ) ) − 1 = ( π A ) * ( ℋ o m O C ( ( ψ * O X ) − , O C ( − 1 ) ) ) .
We compute explicitly now the right-hand side of this equation.
Introducing local coordinates for X = P r o j ( ℛ )   and C = P r o j ( A )   as in the proofs of lemmas  5.1 and  5.3 , we see that a section of ( ψ * O X ) −   on a open set (small enough) can be always written as z ⋅ p q   , where p   and q   are weighted homogeneous polynomial in x 0 , x 1 , y   with coefficients in O U , π A ( P )   , and deg p q : = deg p − deg q = − 3   .
Target of our sheaf of homomorphisms is O C ( − 1 )   whose local sections are of the form p ′ q ′   , where p ′   and q ′   are again weighted homogeneous polynomial in x 0 , x 1 , y   with coefficients in O U , π A ( P )   , with deg p ′ q ′ = − 2   .
Therefore the space of the homomorphisms between these two sheaves restricted to the inverse image of a sufficiently small open set in B   is freely generated by the respective multiplications by x 0 z   and by x 1 z   .
Recalling that x 0   and x 1   are generators of V 1   and z   is a generator of V 3 +   , we conclude that the right-hand side of equation ( 4 ) is the rank 2   vector bundle V 1 ⊗ ( V 3 + ) − 1   . Substituting in equation ( 4 ) and taking determinants we conclude
( det V 1 ⊗ O B ( τ ) ) 2 = ( V 3 + ) 2 , (5)
which isn't enough to conclude yet.
Running a similar computation for V 3 + : = ( f * ω S | B 3 ) +   we get the isomorphism
V 3 + ⊗ ( det V 1 ⊗ O B ( τ ) ) − 3 ∼ = ( π A ) * ( ( ψ * ( ψ ! O C ) 3 ) + ⊗ O C ( − 3 ) ) . (6)
As in the previous case, we write explicitly the right-hand side of the equation ( 6 ).
We have seen that ψ ! O C   is the sheaf of O X −   modules associated to the dual of ψ * O X   : it is generated by the multiplication by p ( x 0 , x 1 , y ) / z   , with p   a homogeneous rational function of degree 3   .
Therefore every section of ( ψ ! O C ) 3   can be written as F / z 3   with F   a homogeneous rational function of degree 9   in x 0 , x 1 , y , z   . A section of this form is invariant if and only if F   is divisible by z   , therefore sections of ψ * ( ψ ! O C ) 3 ) +   are objects of the form F ( x 0 , x 1 , y ) / z 2   , with F   a rational homogeneous function of degree 6   . Twisting it with O C ( − 3 )   (whose sections are of the form 1 / F   with F   exactly as above), we conclude that ( π A ) * ( ψ * ( ψ ! O C ) 3 ) + )   is generated by 1 / z 2   , therefore it is isomorphic to ( V 3 + ) − 2   .
Substituting it in the right hand side of ( 6 ) we get the isomorphism ( V 3 + ) 3 ∼ = ( det V 1 ⊗ O B ( τ ) ) 3 ,   which, together with the isomorphism  5  ( )   concludes the proof.
Q.E.D. Before stating the main theorem of this section we need to give an appropriate definition.
Definition 5.9. Given a genus 2   fibration f : S → B   we define its associated 5-tuple ( B , V 1 , τ , ξ , w )   as follows:
  • ∙   B   is the base curve;
  • ∙   V 1 = f * ( ω S | B )   ;
  • ∙   τ   is the effective divisor of B   whose structure sheaf is isomorphic to T 2   ;
  • ∙   ξ ∈ E x t O B 1 ( O τ , S 2 ( V 1 ) ) / A u t O B ( O τ )   is the isomorphism class of the pair V 2 , σ 2   0 → S 2 ( V 1 ) → σ 2 V 2 → O τ → 0 ;  
  • ∙   setting A ~ 6 : = ℋ o m O B ( ( V 3 + ) 2 , A 6 ) .   w ∈ P ( H 0 ( B , A ~ 6 ) ) ∼ = | O C ( 3 ) ⊗ π A * ( V 3 + ) − 2 |   is the class of a section with associated divisor Δ A   .
Definition 5.10. Let B   be a smooth curve, V 1   , V 2   two vector bundles on B   of respective ranks 2   and 3   . Let σ 2 : S 2 ( V 1 ) → V 2   be an injective homomorphism whose cokernel is isomorphic to the structure sheaf of an effective divisor τ   on B   .
We define A ~ 6   to be the vector bundle ( c o k e r i 3 ) ⊗ ( det V 1 ⊗ O B ( τ ) ) − 2 ,   where the map i 3 : ( det V 1 ) 2 ⊗ V 2 → S 3 ( V 2 )   is the one induced by σ 2   as in lemma  5.4 (see exact sequence ( 1 )).
Remark 5.11. The fact that A ~ 6   is automatically a vector bundle follows by the assumption that the rank of σ 2   drops at most by 1   , which implies that the rank of the induced map i 3   never drops: this is a straightforward computation that we leave to the reader.
The previous results and definitions allow us to introduce the building package of a genus 2   fibration:
Definition 5.12. We shall say that a a 5 −   tuple ( B , V 1 , τ , ξ , w )   is an admissible genus two 5   -tuple if
  • ∙   B   a smooth curve;
  • ∙   V 1   a vector bundle on B   of rank 2   ;
  • ∙   τ ∈ D i v + ( B )   ;
  • ∙   ξ ∈ E x t O B 1 ( O τ , S 2 ( V 1 ) ) / A u t O B ( O τ )   yields a vector bundle V 2   ;
  • ∙   w ∈ P ( H 0 ( B , A ~ 6 ) )   , where A ~ 6   is the vector bundle determined by ξ   as in definition  5.10 ,
and moreover the following (open) conditions are satisfied:
  • i) Let A   be the sheaf of algebras determined by B , V 1 , τ , ξ   as in remark  5.5 ; then the conic bundle C : = P r o j ( A )   has at most rational double points as singularities.
  • ii) Observing that over each point b   of the support of τ   , the map σ 2 : S 2 ( V 1 ) → V 2   induced by ξ   has by construction always rank 2   , we denote by P ( b )   the base locus of the corresponding pencil of lines in the fibre of P ( V 2 )   over b   and denote by P   the set of points P ( b )   thus constructed. Let Δ A   be the divisor of w   on C   ; then Δ A   does not contain any point of P   .
  • iii) At each point of Δ A   the germ of the double cover X   of C   branched on Δ A   has at most rational double points as singularities.
Theorem 5.13. Let f   be a relatively minimal genus 2   fibration. Then its associated 5 −   tuple is admissible.
Viceversa, every admissible genus two 5 −   tuple is the associated 5 −   tuple of a (relatively minimal) genus 2   fibration f : S → B   , and the surface S   has the following invariants:
χ ( O S ) = deg ( V 1 ) + ( b − 1 )
K S 2 = 2 deg V 1 + deg τ + 8 ( b − 1 ) .
Two relatively minimal genus 2   fibrations having isomorphic associated 5 −   tuples are isomorphic.
Proof. That the associated 5 −   tuple of a genus 2   fibration is admissible was already shown.
Viceversa, let ( B , V 1 , τ , ξ , w )   be an admissible (genus two) 5 −   tuple, and let σ 2 : S 2 ( V 1 ) → V 2   be the map induced by ξ   . We can then construct a sheaf of algebras A   following the recipe in remark  5.5 :
P r o j ( A )   is then the divisor in P ( V 2 )   determined by the map i 3   in lemma  5.4 .
Finally, we denote by ℛ   the sheaf of graded A   -modules A ⊕ ( A [ − 3 ] ⊗ det V 1 ⊗ O B ( τ ) )   with the structure of sheaf of algebras induced by the map ( det V 1 ⊗ O B ( τ ) ) 2 → A 6   given by a representative of w   : it is clear that different choices of the representative induce isomorphic structures. We get then by construction a map ψ : X : = P r o j ( ℛ ) → C : = P r o j ( A )   finite of degree 2   whose branch locus is the union of the divisor Δ A   and of the finite set P   , which are disjoint by condition ii). Condition iii) ensures that X   has at most rational double points as singularities.
We then take S   to be a minimal resolution of the singularities of X   .
The induced map f : X → B   is clearly a genus 2   fibration whose associated 5 −   tuple coincides with the given admissible genus two 5 −   tuple.
The only non trivial verification is that f * ω S | B = V 1   : this can be done just repeating the first part of the proof of proposition  5.8 . Note that the relative canonical bundle is by construction f −   nef, i.e., f   is relatively minimal.
Let now f : S → B   and f ′ : S ′ → B   be two relatively minimal genus 2   fibration having the same associated 5 −   tuple, and let ℛ   be the sheaf of algebras induced by the quintuple. Then both f   and f ′   have a birational morphism onto X = P r o j ( ℛ )   : composing them we get a birational map between S   and S ′   (commuting with the fibrations) that will be biholomorphic because of the relative minimality of both fibrations.
There remains finally to compute the invariants of S   using the formulae in remark  2.9 for deg ( V 1 )   and deg ( V 2 )   , and the equality deg V 2 = deg V 1 + deg τ   .
Q.E.D.
Remark 5.14. We discuss here the open conditions in definition  5.12 .
The fibres of the conic bundle C   over a point not in supp ( τ )   are smooth conics, whereas the fibres over points of supp ( τ )   are singular: more precisely a singular reduced conic if the kernel of the map σ 2   on the corresponding fibre is not a square tensor, or a double line elsewhere.
Whence C   is smooth on the complement of the special fibres and on the nonsingular locus of C   the only condition to be fulfilled is that Δ A   have only simple singularities (cf. [BPV] ).
On the reduced singular fibres there is only one singular point, that is the corresponding point of P   , and a local computation shows that it is automatically a singularity of type A 2 s + 1   : in a neighbourhood of these fibres one has only to check that Δ A   does not contain this point.
The situation is slightly more complicated when C   has double fibres: in a neighbourhood of such a fibre one has first to check that the fibre is not contained in the singular locus.
If this is the case, writing the equation locally one finds two possible cases. In the first case C   has only a singular point (of type D 2 s   ) in the corresponding point of P   , which, as before, must not lie in Δ A   .
In the second case our point is a singularity A 1   , and C   has on this fibre a further singular point, still of type A 1   .
Here the branch curve is allowed to pass through this second singular point, and in this case one has to check that the double cover has a rational double point.
Remark 5.15. In [Hor1] Horikawa gave a classification of the special fibres of a genus 2   fibrations. Using the local equations we have used throughout all this section it is not difficult to recognize the classification of Horikawa, and more precisely how the parameters s   and λ   we have introduced determine the geometry of the fibre. We do not enter in the detail of the straightforward but long computation, but we think it can be of some interest for the reader to know that the special fibre is (in the notation of [Hor1] ) of type:
I s + 1 2 i f s i s o d d a n d λ ≠ 0 ;
I I s 2 i f s i s e v e n a n d λ ≠ 0 ;
I I I 1 o r V i f s = 1 a n d λ = 0 ; (7)
I I I s + 1 2 i f s ≥ 3 i s o d d a n d λ = 0 ;
I V s 2 i f s i s e v e n a n d λ = 0 .

6 The genus of the bicanonical pencil of a Godeaux surface

We give an application of theorem  5.13 to the classification of Godeaux surfaces. The following theorem was already proved by Bombieri ([Bom2] , cf. also footnote (1) on page 494 of [Bom1] ) using Ogg's list of genus 2   fibres (cf. [Ogg] ).
Theorem 6.1. Let S ′   be a minimal numerical Godeaux surface, i.e., the minimal model of a surface with K S ′ 2 = 1   , p g ( S ′ ) = 0   , and let f : S → P 1   be the fibration induced by the bicanonical pencil of S ′   . Then the genus of the fibre can only be 3   or 4   .
Proof. We already know (cf. [CP] and  4.1 ) that it suffices to exclude the case where S = S ′   and | 2 K S | = | M | + Φ   where | M |   is a base point free genus 2   pencil ( K S Φ = M 2 = 0 , K S M = 2   ).
Let's argue by contradiction: we have b = 0   , K S 2 = χ ( O S ) = 1   . By the formulae in theorem  5.13 we conclude deg V 1 = 2   , deg τ = 5   . By proposition  5.8 we get V 3 + = O P 1 ( 7 )   .
In particular
h 0 ( O S ( K S + Φ ) ) = h 0 ( O S ( 3 K S − M ) ) = h 0 ( f * O S ( 3 K S − M ) ) = = h 0 ( f * ( ω S | B 3 ) ⊗ O P 1 ( − 7 ) ) ≥ h 0 ( V 3 + ⊗ O P 1 ( − 7 ) ) = h 0 ( O P 1 ) = 1 .  
By the long cohomology exact sequence associated to 0 → O S ( K S ) → O S ( K S + Φ ) → O Φ ( K S + Φ ) → 0 ,   and p g ( S ) = q ( S ) = 0   , we conclude that h 0 ( O Φ ( K S + Φ ) ) ≥ 1   .
On the other hand the long cohomology exact sequence associated to 0 → O S → O S ( Φ ) → O Φ ( Φ ) → 0 ,   implies h 0 ( O Φ ( Φ ) ) = 0   since q ( S ) = 0   .
Since K S Φ = 0   we know (cf. [Art] ) that we have an isomorphism O Φ ( K S ) ∼ = O Φ   , and we derive a contradiction from h 0 ( O Φ ( K S + Φ ) ) ≥ 1   , h 0 ( O Φ ( Φ ) ) = 0   .
Q.E.D. Observe, by the way, that the above proof does not use the full strength of theorem  5.13 .

7 Surfaces with p g = q = 1   and K 2 = 3  

Let S   be a minimal surface of general type with p g = q = 1   ; recall that in this case by the Noether inequality and by the Bogomolov Miyaoka Yau inequality we have 2 ≤ K S 2 ≤ 9   . Then the Albanese map is a morphism f : S → B   where B   is a smooth elliptic curve.
Theorem  5.13 can be used to study the case in which the genus g   of the general fibre of f   is 2   . Xiao's inequality ([Xia1] , thm 2.2) in this case yields 2 ≤ K S 2 ≤ 6   . In fact, for K S 2 = 2   it was proved in [Cat1] that the genus of the Albanese fibre must be 2   , the surface is a double cover of B ( 2 )   , the second symmetric product of B   , and that the moduli space is irreducible, generically smooth, unirational of dimension 7   .
The class of surfaces of general type with K 2 = 3   , p g = q = 1   is investigated in [CC1] , [CC2] . In [CC1] it is proved that for this class of surfaces g = 2   or 3   . The second case is completely classified in [CC2] , where it is shown that the corresponding component of the moduli space is generically smooth, irreducible, unirational of dimension 5   .
In [CC1] it is shown that the surfaces with p g = q = 1   , K 2 = 3   and genus 2   of the Albanese fibre are double covers of B ( 2 )   . In order to show that this class of surfaces is not empty, it is shown in loc. cit.
(whose notation we follow here) the existence of curves C   algebraically equivalent to 6 D − E   and with exactly two ordinary triple points on a given fibre E   . By taking as branch locus the union C ∪ E   , and a corresponding double cover, one obtains surfaces with p g = q = 1   , K 2 = 3   and genus 2   of the Albanese fibre. It was conjectured there (see problem 5.5) that this family of surfaces should form an irreducible family of the moduli space: we will disprove this conjecture in this section using theorem  5.13 , and showing in particular that the corresponding locus of the moduli space is indeed disconnected.
More precisely:
Definition 7.1. We will denote by ℳ   the locus, in the moduli space of the minimal surfaces of general type, corresponding to the surfaces S   with p g ( S ) = q ( S ) = 1   , K S 2 = 3   .
Theorem 7.2. ℳ   has 4   connected components, all irreducible and unirational of dimension 5   .
Remark 7.3. The unirational family of dimension 5   studied in [CC2] is a connected component of ℳ   by virtue of the differentiable invariance of the genus g   (cf. [Cat4] ).
Therefore, if we denote by ℳ ′   the subset of ℳ   corresponding to the surfaces whose Albanese map has genus 2   fibres, to prove theorem  7.2 it suffices to show the following:
Proposition 7.4. ℳ ′   has 3   connected components, all irreducible and unirational of dimension 5   .
To prove the above we shall give a stratification of ℳ   ; we will now show the geometrical description for one of these strata, the one that we shall call of the main stream, because it is defined through some generality assumptions for the branch curve.
Indeed, the conjecture of [CC1] was related to existence and density of this main stream family, but the family constructed in [CC1] was not the one of the main stream!
The main stream family corresponds to double covers of B ( 2 )   branched on curves C + E   belonging to the 5   -dimensional family (vary also B   ) described in the following
Theorem 7.5. Let p : B ( 2 ) → B   be the Albanese (Abel-Jacobi) map, and let E   be a fixed fibre of p   . Let D t   be a natural section of p   given by the set of pairs (of points) on B   which contain a given point t ∈ B   .
Let P 1 , P 2   be general points of E   and let C   be a divisor algebraically equivalent to 6 D − E   having P 1 , P 2   as points of multiplicity at least 3   .
Then, for C   general in the algebraic equivalence class, the linear system | C |   has dimension 1   , the general curve C   inside the system is irreducible and its proper transform after the blow up of P 1   and P 2   has singularities at most double points.
Proof. Recall (cf. e.g. [Cat1] , [CC1] , [CC2] ) that on B ( 2 )   the bianticanonical system | − 2 K |   is a base point free pencil of elliptic curves Δ s = { { x , x + s } }   where s ∈ B   . Δ s   is an irreducible smooth elliptic curve isomorphic to B   , except when s   is a non trivial 2   -torsion point η   , and then Δ η = 2 T η   , where T η   is a smooth elliptic curve 2   -isogenous to B   . Observe moreover that − K   is algebraically equivalent to 2 D − E   .
CLAIM 1) Let C   be algebraically equivalent to 6 D − E   : then H 1 ( O B ( 2 ) ( C ) ) = 0 .   Proof of 1): by Serre duality it suffices to show the vanishing of H 1 ( O B ( 2 ) ( K − C ) )   .
Writing H 1 ( O B ( 2 ) ( K − C ) ) = H 1 ( O B ( 2 ) ( − ( C − K ) ) )   we infer the vanishing from the Ramanujam vanishing theorem ([Bom1]  [Ram] ). In fact, we have the following algebraic equivalences: C − K ∼ ( − 2 K ) + 4 D ∼ Δ s + D t 1 ⋅ ⋅ ⋅ + D t 4   which show that indeed C − K   is linearly equivalent to a reduced and connected effective divisor which is not composed of an irrational pencil. Q . E . D .   for 1) By Riemann Roch we obtain dim | C | = 1 2 C ( C − K ) − 1 = 13   , whence for any choice of distinct points P 1 , P 2 ∈ B ( 2 )   we have :
2) dim | C − 3 P 1 − 3 P 2 | ≥ 1 .   3) Letting P 3   be a third point of the fibre E   , distinct from P 1 , P 2   , we see that a curve in | C − 3 P 1 − 3 P 2 |   passing through P 3   must contain the whole fibre E   , whence the system | C − 3 P 1 − 3 P 2 |   contains the linear subsystem E + | C − E − 2 P 1 − 2 P 2 |   which is of codimension at most 1.
4) Observe that, by the same argumentation as in 1), we get dim | C − E | = 6   . Instead, we claim that C − 2 E ∼ − 3 K   is empty if the linear equivalence class of C   is general.
In fact, if M ∼ ( C − 2 E )   is an effective divisor, then M Δ s = 0   , thus M   consists of a fibre Δ s   of the bianticanonical pencil plus a curve T η   .
5) Assume now that the general curve in | C − 3 P 1 − 3 P 2 |   is always reducible, i.e., that generally there exists a fixed part Φ   .
There are two cases:
5.1) Φ ≠ E   .
5.2) Φ = E   .
CLAIM I): case 5.1) is not possible.
Proof of claim I): then the system | C − 3 P 1 − 3 P 2 |   contains, by 3), a reducible curve of the form E + Φ + M   , where C ′ : = Φ + M   is in | C ′ − 2 P 1 − 2 P 2 |   . Let the general pair specialize now to a pair (of distinct points) which is contained in the complete intersection Δ s ∩ E   .
The intersection number C ′ Δ s   is = Δ s ⋅ 2 D = 4   and therefore the system | C ′ − 2 P 1 − 2 P 2 |   contains the subsystem Δ s + | C ′ − Δ s − P 1 − P 2 |   as a subsystem of codimension ≤ 1   .
Since C ′ − Δ s ∼ 2 D   , and | 2 D − E | = ∅   for general choice of the linear equivalence classes of D , E   we easily see that dim | 2 D | = 2   always, and that the subsystem | C ′ − Δ s − P 1 − P 2 |   has dimension zero for general choice of the linear equivalence class of C ′   .
MAIN CLAIM: For special choice of { P 1 , P 2 } ⊂ Δ s ∩ E   , and for C   , Δ s   general, | C ′ − 2 P 1 − 2 P 2 |   equals the subsystem Δ s + | C ′ − Δ s − P 1 − P 2 |   .
Moreover, for a general choice of { P 1 , P 2 }   , and of C ′   , | C ′ − 2 P 1 − 2 P 2 |   has dimension 0   .
Proof of the Main Claim. Otherwise, the linear system | C ′ − 2 P 1 − 2 P 2 |   has always dimension ≥ 1   for a special choice of the points { P 1 , P 2 } ⊂ Δ s ∩ E   , and we see that this dimension is then generally equal to 1   by 4). Varying E   , Δ s   , and the linear equivalence class, we obtain an irreducible family of dimension 4   which dominates a subvariety of dimension 2   of the symmetric square of B ( 2 )   .
Thus, it suffices to show that at some curve Z   of the 4   -dimensional family given by curves in Δ s + G , G ∈ | C ′ − Δ s − P 1 − P 2 |   the equisingular deformations of Z   dominate the symmetric square of B ( 2 )   and the dimension of the fibre is at most 1   . In fact then we obtain a contradiction which proves the first assertion, and since the family has dimension at least 5   , also the second assertion follows.
Let us choose Z   of the form Δ s + D 1 + D 2   , where P i ∈ D i   .
Observe that D 1   and D 2   meet transversally at a point P   , while Δ s ∩ D i = { P i , Q i }   .
Let's make everything explicit, letting E   be the fibre of pairs with sum 0   , E = { ( y , − y ) | y ∈ B }   , letting P 1 = ( a , − a ) , P 2 = ( b , − b )   , D 1 = { ( a , x ) | x ∈ B }   , D 2 = { ( b , z ) | z ∈ B }   , s = 2 a   , thus Δ s = { ( u , u + 2 a ) }   .
Without loss of generality we may assume that 2 a = s = 2 b   . We can easily compute that Q 1 = ( a , 3 a )   , while Q 2 = ( b , 3 b )   .
We conclude that on the elliptic curve Δ s   the divisor P 1 + P 2   is not linearly equivalent to Q 1 + Q 2   for a general choice of s   , since − a − b ≠ a + b   .
Consider now the curve Z   and the normal sheaf N Z ′   of deformations which are equisingular at P 1 , P 2   .
Its space of global sections fits into the exact sequence:
0 → H 0 ( N Z ′ ) → → H 0 ( N D 1 ( Q 1 + P ) ) ⊕ H 0 ( N D 2 ( Q 2 + P ) ) ⊕ H 0 ( O Δ ( Q 1 + Q 2 ) ) → → C Q 1 ⊕ C Q 2 ⊕ C P →  
Since the normal sheaf N D i   has degree 1   , each summand H 0 ( N D i ( Q i + P ) )   in the middle surjects onto the direct sum C Q i ⊕ C P   of the two corresponding summands to the right.
It follows on one side that we have surjectivity onto the right term, that dim ( H 0 ( N Z ′ ) ) = 5   .
Since we saw that P 1 + P 2   is not linearly equivalent to Q 1 + Q 2   on Δ s   , we obtain that H 0 ( N Z ′ )   surjects onto the direct sum of the cotangent spaces of the surface at P 1 , P 2   , since H 0 ( N D 1 ( Q 1 + P − P 1 ) ) ⊕ H 0 ( N D 2 ( Q 2 + P − P 2 ) ) ⊕ H 0 ( O Δ ( Q 1 + Q 2 − P 1 − P 2 ) )   has dimension 4   and surjects onto C Q 1 ⊕ C Q 2 ⊕ C P   . We have therefore shown that the tangent dimension to the family is 5   at Z   , the differential of the map onto the symmetric square of B ( 2 )   is surjective, whence the family is smooth at Z   of dimension 5   and the fibre is smooth of dimension 1   .
Q . E . D .   for the Main Claim. We proceed now with the proof of claim 1): we let the pair P 1 , P 2   tend to a special pair: then the reducible curve Φ + M   has as limit a curve in the linear system | C ′ − 2 P 1 − 2 P 2 | = Δ s + G   . By a general choice of the class of C ′   we obtain that G ∼ 2 D   is irreducible. It follows that, modulo exchanging the roles of Φ   and M   , we may assume that Φ   tends to Δ s   , in particular Φ   is algebraically equivalent to Δ s   and thus belongs to the antibicanonical pencil. This however shows that the points P 1 , P 2   must always be special, a contradiction.
Q . E . D .   for claim 1) We finally observe that also case 5.2) is impossible, because otherwise we would have | C − 3 P 1 − 3 P 2 | = E + | C − E − 2 P 1 − 2 P 2 |   and we would have that dim | C − E − 2 P 1 − 2 P 2 |   is always ≥ 1   , contradicting the main claim.
Whence, we have proven that the general curve C   which has multiplicity at least 3   in each point P 1 , P 2   is generally irreducible, and for a fixed choice of P 1 , P 2   we have a linear pencil.
Observe however that the intersection number of C   with E   is 6   , whence it follows that each point P i   is of multiplicity exactly 3   , and that the fibre is not tangent. After blowing up the two points, the proper transform of C   , call it C ~   , does not intersect the proper transform of the fibre E   , and moreover C ~ 2 = 6   , whence we obtain a pencil of curves C ~   which generically have at most double points as singularities.
Q . E . D .   We have just shown how to construct the family of surfaces of the main stream, i.e., those double covers where all the choices made, elliptic curve, fibre E   , two points on the fibre, linear equivalence class, can be made general. However, for special choices, it can happen that the dimension of the corresponding linear system of curves jumps.
It is for instance clear that this happens when the points are very special, i.e., on a curve T η   , but in order to analyse all the possible cases, we will need to make full use of the theory developed in the previous section.
We need now to recall some standard notation and results about vector bundles on elliptic curves.
Definition 7.6. Given a point u   of a elliptic curve B   , and integers r , d   with r > 0   , ( r , d ) = 1   , we will denote by E u ( r , d )   (following [Ati] ) the only indecomposable vector bundle of rank r   on B   with det E u ( r , d ) = O B ( u ) ⊗ d   .
Lemma 7.7 ([CC1] , page 76). Let S   be a minimal surface of general type with p g ( S ) = q ( S ) = 1   whose Albanese map f : S → B   is a genus 2   fibration.
If K S 2 ≤ 3   , then there is a point u ∈ B   such that f * ω S | B = E u ( 2 , 1 )   .
If instead K S 2 ≥ 4   , either f * ω S | B = E u ( 2 , 1 )   or f * ω S | B = L ⊕ O B ( u )   with u ∈ B   and L   a non trivial torsion line bundle.
From this lemma and theorem  5.13 we immediately get the following
Remark 7.8. Let S   be a minimal surface of general type with p g ( S ) = q ( S ) = 1   and K S 2 = 3   whose Albanese map f : S → B   is a genus 2   fibration (onto an elliptic curve B   ).
Let ( B , V 1 , τ , ξ , w )   be the associated 5 −   tuple: then we may assume w.l.o.g. that V 1 = E [ 0 ] ( 2 , 1 )   .
Moreover deg τ = K S 2 − 2 = 1   , i.e., τ   is a point of B   .
To apply theorem  5.13 , we need to compute S 2 ( V 1 ) = S 2 ( E [ 0 ] ( 2 , 1 ) )   ; this is a computation based on the results in [Ati] : the interested reader can find the computation in [CC2] .
Remark 7.9. [cf. [CC2] ] Given an elliptic curve B   , let us denote by L i , i ∈ { 1 , 2 , 3 }   the three line bundles on B   with L i ≁ = O B   , L i 2 ∼ = O B   . Then S 2 ( E [ 0 ] ( 2 , 1 ) ) = 3 ⊕ i = 1 L i ( [ 0 ] ) .  
The 4 t h   element of the associated 5 −   tuple, ξ   , belongs to E x t O B 1 ( O τ , ⊕ i = 1 3 L i ( [ 0 ] ) ) / A u t O B ( O τ ) .   Note that ξ   can't be the class of 0   , since the extension must yield a vector bundle. Since τ   is a point, A u t O B ( O τ ) ∼ = C *   , acting on E x t O B 1 ( O τ , S 2 ( V 1 ) ) \ { 0 }   by scalar multiplication: therefore the space to which ξ   belongs is the projective space P ( E x t O B 1 ( O τ , S 2 ( V 1 ) ) )   . To compute this space we fix a section f 0 ∈ H 0 ( O B ( τ ) ) \ { 0 }   ; applying the functor H o m O B ( ⋅ , S 2 ( V 1 ) )   to the exact sequence 0 → O B ( [ 0 ] − τ ) ⟶ ⋅ ( − f 0 ) O B ( [ 0 ] ) → O τ → 0 ,   we get isomorphisms
(8) E x t O B 1 ( O τ , ⊕ i = 1 3 L i ( [ 0 ] ) ) ∼ = H o m O B ( O B ( [ 0 ] − τ ) , ⊕ i = 1 3 L i ( [ 0 ] ) ) ∼ = ∼ = H 0 ( ⊕ 1 3 L i ( τ ) ) ∼ = C 3 .
Therefore the associated projective space is a P 2   : we have parametrized the first 4   data of our 5 −   tuples via an unirational parameter space of dimension 4   (a P 2 −   bundle over a universal family of elliptic curves).
To complete the 5 −   tuple we need to compute H 0 ( A ~ 6 )   (in terms of B , τ , ξ   ). First we describe V 2   explicitly as follows
Remark 7.10. Let ( f 1 , f 2 , f 3 )   be an element of H 0 ( ⊕ 1 3 L i ( τ ) )   .
Then the corresponding extension class ξ   induced by f 0   as above described yields a sheaf V 2   that is the cokernel of the map from O B ( [ 0 ] − τ )   to O B ( [ 0 ] ) ⊕ ( ⊕ 1 3 L i ( [ 0 ] ) )   induced by the matrix t ( f 0 , f 1 , f 2 , f 3 )   .
In fact, by standard homological algebra we have a commutative diagram
0 → O B ( [ 0 ] − τ ) → ⋅ ( − f 0 ) O B ( [ 0 ] ) → O τ → 0
↓ ↓ | |
0 → ⊕ 1 3 L i ( [ 0 ] ) → σ 2 V 2 → O τ → 0 ,
where the vertical map on the left is the one given by t ( f 1 , f 2 , f 3 )   . This diagram clearly induces an exact sequence
0 → O B ( [ 0 ] − τ ) → O B ( [ 0 ] ) ⊕ ( ⊕ 1 3 L i ( [ 0 ] ) ) → V 2 → 0 . (9)
Lemma 7.11. V 2   is determined by the f i   's as follows:
I) f i ≠ 0 ∀ i ∈ { 1 , 2 , 3 }   ⇔   V 2 ( − [ 0 ] ) ∼ = E τ ( 3 , 1 )  
II) ∃ ! i ∈ { 1 , 2 , 3 }   with f i = 0   ⇔   V 2 ( − [ 0 ] ) ∼ = E τ i ( 2 , 1 ) ⊕ L i  
III) ∃ ! i ∈ { 1 , 2 , 3 }   with f i ≠ 0   ⇔   V 2 ( − [ 0 ] ) ∼ = L i ( τ ) ⊕ L j ⊕ L k  
where the point τ i ∈ B   is the divisor of a non trivial section of L i ( τ )   , and whenever j   and k   appear { i , j , k } = { 1 , 2 , 3 }   .
Proof. Let m   be the cardinality of the set { i | f i = 0 }   . Note that m ≤ 3   because f 0   is different from zero by assumption.
By the exact sequence ( 9 ), follows that V 2   is a vector bundle if and only if the f i   's have no common zeroes, i.e., if and only if m ≤ 2   (since the points τ i   are distinct).
Tensoring the exact sequence ( 9 ) by O B ( − [ 0 ] )   and since H 1 ( O B ( − τ ) ) → ⋅ f 0 H 1 ( O B )   is an isomorphism, we conclude that H 1 ( V 2 ( − [ 0 ] ) ) = 0   .
Further twisting the exact sequence ( 9 ) by any degree 0   line bundle L   and repeating the argument, we see that H 1 ( V 2 ( − [ 0 ] ) ⊗ L )   vanishes unless L = L i   for some i and f i = 0   : in this last case H 1 ( V 2 ( − [ 0 ] ) ⊗ L ) ∼ = C   .
Therefore V 2 ( − [ 0 ] )   is a vector bundle of rank 3   and determinant O B ( τ )   having the property that there are exactly m   line bundles L   of degree zero with H 1 ( V 2 ( − [ 0 ] ) ⊗ L ) ≠ 0   .
If V 2 ( − [ 0 ] )   is indecomposable, it is E τ ( 3 , 1 )   by Atiyah's classification.
This sheaf has trivial cohomology when twisted with any degree 0   line bundle, therefore m = 0   : we are in case I). If on the contrary V 2 ( − [ 0 ] )   is a sum of three line bundles, being a quotient of O B ⊕ ( ⊕ L i )   of degree 1   , we see that two summands have degree 0   : it is clear that in this case m = 2   and that V 2 ( − [ 0 ] )   is the one described in the statement (case II)). Else V 2 ( − [ 0 ] ) = W ⊕ L   sum of two indecomposable vector bundles of respective ranks 2   and 1   .
First we exclude the case deg W = 0   , deg L = 1   . Recall that, by theorem 5   of [Ati] , part ii), for each indecomposable vector bundle W   of degree zero there is exactly one line bundle L   with H 1 ( W ⊗ L ) ≠ 0   .
Therefore, if W   had degree 0   , we get m = 1   . But then, if f i   is the vanishing section, L i   is a direct summand of V 2 ( − [ 0 ] )   , a contradiction.
Then (by semipositivity) deg L = 0   and deg W = 1   ; every twist of W   by a degree 0   line bundle has trivial first cohomology group, therefore the corresponding twist of V 2 ( − [ 0 ] )   has nontrivial first cohomology group if and only if we twist by L − 1   : whence m = 1   and we are in case II). Q . E . D .   The preceding result suggests to consider a stratification of our moduli space of surfaces:
Definition 7.12. We stratify ℳ ′   as ℳ ′ = ℳ I ∪ ℳ I I ∪ ℳ I I I   according to the number of indecomposable summands for V 2 = f * ( ω S | B 2 )   , as in lemma  7.11 .
We decompose further ℳ I   as ℳ I ∘ ∪ ℳ I , 3   where ℳ I , 3   is given by the surfaces with 3 τ ≡ 3 [ 0 ]   ; ℳ I ∘   by the other surfaces in ℳ I   .
Remark 7.13. By a theorem of Clemens ([Cle] , see [CS] for the way it is applied) the dimension of an irreducible component of the moduli space of minimal surfaces of general type with p g = q = 1   is at least 10 χ − 2 K 2 + 1   , that is, 5   in our case. In particular each stratum of smaller dimension cannot contain an open set and can be disregarded for the determination of the irreducible components.
Studying the 4   strata introduced above we will find that ℳ I I I   consists of two not empty unirational families of dimension 5   , ℳ I ∘   is the unirational family corresponding to surfaces of the main stream, while the other strata have smaller dimension.
Finally we will study how the closures of these strata do intersect.
In order to simplify the exposition, we will first treat together the cases in which V 2   is decomposable: ℳ I I   and ℳ I I I   . We need the following
Remark 7.14. We view the natural map
( det V 1 ) 2 → S 2 ( S 2 ( V 1 ) )
( x 0 ∧ x 1 ) 2 → x 0 2 x 1 2 − ( x 0 x 1 ) 2
(since V 1 ∼ = E [ 0 ] ( 2 , 1 )   ) as a map O B ( 2 [ 0 ] ) → ( O B 3 ⊕ L 1 ⊕ L 2 ⊕ L 3 ) ( 2 [ 0 ] ) ,   given through a vector with 6   entries, whose last 3   are necessarily zero. Moreover the first three 3   cannot be zero, since the Segre image of P 1   in P 2   is a smooth conic, i.e., a conic of rank 3   . In particular (up to automorphisms) we can assume this vector to be the transpose of ( 1 , 1 , 1 , 0 , 0 , 0 )   .
Lemma 7.15. If V 2 ( − [ 0 ] )   splits as W ⊕ L   , with L ⊗ L ∼ = O B   and W   a vector bundle of rank 2   , then A ~ 6 ∼ = ( S 3 ( W ) ⊕ ( S 2 ( W ) ⊗ L ) ) ( [ 0 ] − 2 τ ) .  
Proof. Applying definition  5.10 in this case, we get that A ~ 6   is the cokernel of an injective map ( W ⊕ L ) ( [ 0 ] − 2 τ ) ↪ ( [ S 3 ( W ) ⊕ ( S 2 ( W ) ⊗ L ) ] ⊕ { W ⊕ L } ) ( [ 0 ] − 2 τ ) .   Applying remark  7.14 to the definition of the exact sequence  1 for n = 2   , one easily sees that the second component of this map is the identity, thus the above sequence splits.
Q . E . D .   We can now discuss the families ℳ I I   and ℳ I I I   separately.
Proposition 7.16. ℳ I I   is either empty or it has dimension 4   .
Proof. By lemma  7.11 , in case I I   , ξ   varies in a 1   -parameter family.
By lemma  7.15 and by the cited result of Atiyah ([Ati] ) A ~ 6 ∼ = E τ i ( 2 , 1 ) ( [ 0 ] − τ ) ⊕ E τ i ( 2 , 1 ) ( [ 0 ] − τ ) ⊕ ( ( ⊕ j = 1 3 L j ) ( [ 0 ] − τ ) ) ,   and we obtain h 0 ( A ~ 6 ) = 2   unless O B ( [ 0 ] − τ )   is a nontrivial 2-torsion bundle: in this last case h 0 ( A ~ 6 ) = 3   . Each of the two cases gives a (possibly empty) unirational family of dimension 4   .
Q . E . D .  
Proposition 7.17. ℳ I I I   has two connected components, both non empty, unirational of dimension 5   . For the first component τ = [ 0 ]   , in the other case τ   is a 2 −   torsion point.
Proof. By lemma  7.11 we can assume f j = f k = 0   ; applying lemma  7.15 we get
A ~ 6 ∼ = L i ( [ 0 ] + τ ) ⊕ L j ( [ 0 ] ) ⊕ L i ( [ 0 ] − τ ) ⊕ L j ( [ 0 ] − 2 τ ) ⊕ ⊕ L k ( [ 0 ] ) ⊕ O B ( [ 0 ] − τ ) ⊕ L k ( [ 0 ] − 2 τ ) .  
Therefore either h 0 ( A ~ 6 ) = 4   or h 0 ( A ~ 6 ) = 5   , the last case occurring when τ = [ 0 ]   or when O B ( [ 0 ] − τ ) ∼ = L i   .
The decomposition V 2 ( − [ 0 ] ) = L i ( τ ) ⊕ L j ⊕ L k   yields natural coordinates y i , y j , y k   on P ( V 2 )   . In these coordinates our conic bundle has equation f 0 2 y i 2 + y j 2 + y k 2   : let us note that it has only one singular point (of type A 1   ).
Each line bundle summand of S 3 ( V 2 )   corresponds to a monomial y i y j y k   . Therefore the relative cubic given by the corresponding section of H 0 ( A ~ 6 )   is cut by a relative cubic on P ( V 2 )   . If h 0 ( A ~ 6 ) = 4   we see that the only monomials allowed to have nonzero coefficient are the monomials y i 3 , y i 2 y j , y i 2 y k   : in particular the branch curve is not reduced (it contains { y i = 0 }   twice) and the generality conditions in definition  5.12 are never fulfilled: this case does not occur.
Therefore h 0 ( A ~ 6 ) = 5   and either τ = [ 0 ]   or O B ( τ − [ 0 ] ) = L i   .
If O B ( τ − [ 0 ] ) = L i   we have a relative cubic of the form a y i 3 + b y i 2 y j + c y i 2 y k + d y i y j 2   , where a   , b   , c   and d   are global sections of the respective line bundles.
This cubic is clearly reducible and we write the corresponding curve on the conic bundle as C 1 ∪ C 2   where C 1 = { y i = 0 }   and C 2 = { a y i 2 + b y i y j + c y i y k + d y j 2 = 0 }   ; of course d   is not allowed to be the zero section or we get the same contradiction as in the previous case); since d   is a section of the trivial bundle it never vanishes, which implies C 1 ∩ C 2 = ∅   .
C 1   is obviously smooth, while C 2   varies in a linear system that has no fixed points on the conic bundle, therefore by Bertini's theorem its general element is smooth. If we moreover assume that a   (section of O B ( 2 [ 0 ] )   ) does not vanish in [ 0 ]   we ensure that the curve does not pass through the singular point of the conic bundle: therefore there exists an element in H 0 ( A ~ 6 )   fulfilling the open conditions listed in definition  5.12 .
If [ 0 ] = τ   we have the relative cubic of equation a y i 3 + b y i 2 y j + c y i 2 y k + d y i y j y k   : the proof of the existence of a 'good' cubic is identical to the previous one.
The computation of the ”number of parameters” is easy: 1   parameter for B   , no parameters for τ   and ξ   , 4   from H 0 ( A ~ 6 )   .
Q . E . D .  
Remark 7.18. The example constructed in [CC1] , is in fact a minimal resolution of the singularities of a double cover of B ( 2 )   branched on T η + D t + D t + η + D t + η ′ + D t + η + η ′ + E 2 t + η ,   for some t ∈ B   and a pair of distinct non trivial 2 −   torsion elements η , η ′   . This curve has two quadruple points at the complete intersection T η ∩ E 2 t + η = { t , t + η } ∪ { t + η ′ , t + η ′ + η } .   We try to reconstruct the associate 5 −   tuple of this example.
Recall that V 1   is the bundle E [ 0 ] ( 2 , 1 )   ; τ   is the point 2 t + η   and we claim that ξ   the 'splitting' extension (case I I I   ), such that the kernel of σ 2   over the given point is the fibre of the bundle L i   corresponding to the chosen curve T η   . In fact, the two quadruple points for the branch curve, T η ∩ E 2 t + η   , are blown up, and the contraction of the proper transform of the fibre yields the singular point of the conic bundle C   : thus the kernel in the fibre of S 2 ( V 1 )   is as stated.
The above mentioned branch curve, after removing the fibre E 2 t + η   , stays in | 6 D 0 − E p |   for p = η − 4 t   , and it is therefore given by a map O B ( p ) → S 6 ( E [ 0 ] ( 2 , 1 ) )   ; so the branch curve in C   comes from a map O B ( p + 3 τ ) ∼ = L i ( 3 [ 0 ] + τ ) → A 6   . Since A 6   splits (following lemma  7.15 ) as ( S 3 ( L i ( τ ) ⊕ L j ) ⊕ ( S 2 ( L i ( τ ) ⊕ L j ) ⊗ L k ) ) ( 3 [ 0 ] )   , we find L i ( 3 [ 0 ] + τ )   as a direct summand in the left summand, and this characterises the second case of proposition  7.17 .
The first case of proposition  7.17 allows instead the following geometric example: take the minimal resolution of singularities of a double cover of B ( 2 )   branched on C + T η + E 2 t + η ,   where t ∈ B   , η   , is a non trivial 2 −   torsion element and C ∈ | 3 D t + D t + η |   has double points at T η ∩ E 2 t + η   ; the existence of such curve was proven in the proof of proposition  7.17 .
We note now that theorem  7.5 shows that ℳ I ∘ ≠ ∅   : in fact it shows that there is a component of the moduli space where the degree 0   line bundle det V 1 ( − τ )   can be chosen general, whereas for the stratum ℳ I , 3   we have that det V 1 ( − τ )   is a 3 −   torsion line bundle, and we have just shown that for each not empty irreducible component contained in ℳ I I ∪ ℳ I I I   , det V 1 ( − τ )   is a 2 −   torsion line bundle. In the next proposition we analyse this stratum.
Proposition 7.19. ℳ I ∘   forms an irreducible unirational family of dimension 5   .
Proof. By lemma  7.11 all f i   's are different from zero, and V 2 ( − [ 0 ] ) ∼ = E τ ( 3 , 1 )   .
By theorem 14' of [Ati] we have that S 2 ( E τ ( 3 , 1 ) ) ⊕ Λ 2 E τ ( 3 , 1 ) = E τ ( 3 , 1 ) ⊗ E τ ( 3 , 1 ) ∼ = E τ ( 3 , 2 ) ⊕ 3 ,   whence S 2 ( E τ ( 3 , 1 ) ) ∼ = E τ ( 3 , 2 ) ⊕ 2   , Λ 2 ( E τ ( 3 , 1 ) ) ∼ = E τ ( 3 , 2 )   , and that E τ ( 3 , 1 ) ⊗ E τ ( 3 , 2 ) ∼ = O B ( τ ) ⊕ ( ⊕ 1 8 M i ( τ ) ) ,   where the M i   's are the line bundles with M i 3 ∼ = O B   , M i ≁ = O B   .
Consider the Eagon Northcott exact sequence
0 → det E τ ( 3 , 1 ) → Λ 2 E τ ( 3 , 1 ) ⊗ E τ ( 3 , 1 ) → → E τ ( 3 , 1 ) ⊗ S 2 ( E τ ( 3 , 1 ) ) → S 3 ( E τ ( 3 , 1 ) ) → 0 .  
We have shown that the first 3   bundles are direct sums of line bundles of degree 1   : by cancellation we get S 3 ( E τ ( 3 , 1 ) ) = ( ⊕ 1 2 O B ( τ ) ) ⊕ ( ⊕ 1 8 M i ( τ ) ) .   By definition  5.10 we have an exact sequence
(10) 0 → E τ ( 3 , 1 ) ⊗ O B ( [ 0 ] − 2 τ ) → → ( ( ⊕ 1 2 O B ) ⊕ ( ⊕ 1 8 M i ) ) ( [ 0 ] − τ ) → A ~ 6 → 0 .
Note that E τ ( 3 , 1 ) ⊗ O B ( [ 0 ] − 2 τ )   is indecomposable of degree − 2   , therefore has no global sections and first cohomology group of dimension 2   (regardless of the choice of τ   ).
By assumption τ   is not a 3-torsion point: therefore the sheaf in the middle of the exact sequence ( 10 ) has no cohomology. It follows that H 0 ( A ~ 6 ) ∼ = C 2   , and w   varies in a P 1   . Summing up, we can parametrize ℳ I ∘   via a rational 5   -dimensional family ( 1   parameter for w   , 2   for ξ   , 1   for τ   and 1   for B   ).
Q . E . D .   We are left with the stratum ℳ I , 3   , computationally more complicated.
We will show that it has dimension ≤ 4   .
We will need the following algebraic lemma.
Lemma 7.20. Let B   be a smooth elliptic curve, τ ∈ B   a point, V 1 = E 0 ( 2 , 1 )   . Fix f 0 ∈ H 0 ( O B ( τ ) ) \ { 0 }   , and ∀ i ∈ { 1 , 2 , 3 }   f i ∈ H 0 ( ⊕ 1 3 L i ( τ ) )   .
Let then 0 → S 2 ( V 1 ) → σ 2 V 2 → O τ → 0 .   be the extension associated to ( f 0 , f 1 , f 2 , f 3 )   , and let A ~ 6   be the vector bundle associated to σ 2   (as in definition  5.10 ).
Then A ~ 6   is isomorphic to the cokernel of the map F   from ( O B ⊕ L 1 ⊕ L 2 ⊕ L 3 ⊕ O B ⊕ L 3 ⊕ L 2 ⊕ O B ⊕ L 1 ) ( [ 0 ] − 3 τ )   to
( O B ⊕ L 1 ⊕ L 2 ⊕ L 3 ⊕ O B ⊕ L 3 ⊕ L 2 ⊕ O B ⊕ L 1 ⊕ ⊕ L 2 ⊕ L 3 ⊕ L 1 ⊕ O B ⊕ L 1 ⊕ L 3 ⊕ L 2 ) ( [ 0 ] − 2 τ )  
given by the transpose of the matrix M = ( f 0 f 1 f 2 f 3 f 0 f 1 f 2 f 3 f 0 f 1 f 2 f 3 f 0 − f 3 f 1 − f 3 f 2 f 0 f 2 f 3 − f 1 − f 1 f 0 f 1 f 2 f 3 f 0 f 1 f 2 f 3 f 0 − f 2 f 1 f 3 − f 2 f 0 f 1 f 2 f 3 ) .   In particular H 0 ( A ~ 6 )   is isomorphic to the kernel of H 1 ( F )   .
Proof. In order to have all the diagrams fitting in the page, let us denote by ℬ   the vector bundle E n d ( V 1 ) = O B ⊕ L 1 ⊕ L 2 ⊕ L 3   .
Our by now familiar exact sequence 0 → O B ( [ 0 ] − τ ) → ℬ ( [ 0 ] ) → V 2 → 0   induces the exact sequence 0 → S 2 ( ℬ ) ( 3 [ 0 ] − τ ) → S 3 ( ℬ ) ( 3 [ 0 ] ) → S 3 ( V 2 ) → 0 .   Taking a twist A ~ 6   , given as a cokernel (cf. definition  5.10 ), we get the following diagram with exact rows and columns
0 ↓ 0 → O B ( 3 [ 0 ] − τ ) → ℬ ( 3 [ 0 ] ) → det ( V 1 ) 2 ⊗ V 2 → 0 ↓ i 3 0 → S 2 ( ℬ ) ( 3 [ 0 ] − τ ) → S 3 ( ℬ ) ( 3 [ 0 ] ) → S 3 ( V 2 ) → 0 ↓ A ~ 6 ( 2 [ 0 ] + 2 τ ) ↓ 0 . (11)
Note that all the vector bundles outside the last column are direct sum of line bundles. We look then for explicit 'vertical' maps α : O B ( 3 [ 0 ] − τ ) → S 2 ( ℬ ) ( 3 [ 0 ] − τ ) , β : ℬ ( 3 [ 0 ] ) → S 3 ( ℬ ) ( 3 [ 0 ] )   that can be inserted in the diagram ( 11 ) enlarging it to a bigger commutative diagram.
Now recall that i n   is defined by the formula i n ( ( x 0 ∧ x 1 ) ⊗ 2 ⊗ q ) = ( σ 2 ( x 0 2 ) σ 2 ( x 1 2 ) − σ 2 ( x 0 x 1 ) 2 ) q   , for q ∈ S n − 2 ( V 2 )   .
Consider then first i 2   : it factors as ( det V 1 ) 2 → S 2 ( S 2 ( V 1 ) ) → S 2 ( V 2 ) .   Remark  7.14 gave us an explicit form of the above (injective) map on the left between det V 1 2 ∼ = O B ( 2 [ 0 ] )   , and S 2 ( S 2 ( V 1 ) ) ∼ = S 2 ( ( L 1 ⊕ L 2 ⊕ L 3 ) ( [ 0 ] ) ) ∼ = ( O ⊕ L 3 ⊕ L 2 ⊕ O ⊕ L 1 ⊕ O ) ( 2 [ 0 ] ) .   it has the matrix t ( 1 , 0 , 0 , 1 , 0 , 1 )   (it is different from the previously given one just because we use now the lexicographic order).
We can now write the vertical maps we need: the map β   is the composition of the natural maps
( det V 1 ) 2 ⊗ ℬ ( [ 0 ] ) → S 2 ( S 2 ( V 1 ) ) ⊗ ℬ ( [ 0 ] ) → → S 2 ( ℬ ( [ 0 ] ) ) ⊗ ℬ ( [ 0 ] ) → S 3 ( ℬ ( [ 0 ] ) )  
where the first map is the one we have just described (tensored by ℬ ( [ 0 ] )   ), and the second map is induced by the inclusion S 2 ( V 1 ) ↪ O B ( [ 0 ] ) ⊕ S 2 ( V 1 ) = ℬ ( [ 0 ] )   .
A similar splitting can be done for α   : using again the lexicographic order we get S 2 ( ℬ ) ( 3 [ 0 ] − τ ) = ( O B ⊕ L 1 ⊕ L 2 ⊕ L 3 ⊕ O B ⊕ L 3 ⊕ L 2 ⊕ O B ⊕ L 1 ⊕ O B ) ( 3 [ 0 ] − τ )   and α = t ( 0 , 0 , 0 , 0 , 1 , 0 , 0 , 1 , 0 , 1 ) .   Taking the mapping cone we obtain a free resolution of A ~ 6   as follows:
0 → O B ( 3 [ 0 ] − τ ) → α ~ ℬ ( [ 3 [ 0 ] ) ⊕ S 2 ( ℬ ) ( 3 [ 0 ] − τ ) → → β ~ S 3 ( ℬ ) ( 3 [ 0 ] ) → A ~ 6 ( 2 [ 0 ] + 2 τ ) → 0  
with α ~ = ( f 0 f 1 f 2 f 3 0 0 0 0 − 1 0 0 − 1 0 − 1 ) β ~ = ( f 0 f 1 f 0 f 2 f 0 f 3 f 0 1 f 1 f 0 f 2 f 1 f 0 f 3 f 1 f 0 1 f 2 f 0 f 3 f 2 f 0 1 f 3 f 0 1 f 1 1 f 2 f 1 1 f 3 f 1 1 f 2 f 1 f 3 f 2 f 1 1 f 3 f 1 1 f 2 1 f 3 f 2 1 f 3 f 2 1 f 3 ) .   The reader can now easily simplify this resolution and get the minimal one that is given by the above matrix t M   .
Q . E . D .   We can now consider the missing case ℳ I , 3   : V 2   is indecomposable, and τ   is a 3-torsion point. Exact sequence ( 10 ) still holds but the vector bundle in the middle has nontrivial cohomology: both cohomology groups have the same dimension: 1   if τ ≠ [ 0 ]   , and 2   if τ = [ 0 ]   .
Proposition 7.21. ℳ I , 3   has dimension at most 4   .
Proof. Let us start with the case where τ   is a 3-torsion point, τ ≠ [ 0 ]   and assume by contradiction that we have then an irreducible family S   of dimension at least 5   . The data ( B , τ , ξ )   give a map from S   to an irreducible family Y   of dimension 3   (1-parameter for B   , no parameters for τ   , 2   parameters for ξ   ) whose fibre is the projective space P ( H 0 ( A ~ 6 ) )   which has dimension 2   .
We know in fact that h 0 ( A ~ 6 ) ≤ 3   , that the map dominates Y   , whence the fibre has dimension exactly two.
In other terms, for any triple ( B , τ , ξ )   with τ ≠ [ 0 ]   of 3 −   torsion we have h 0 ( A ~ 6 ) = 3   .
Let us then fix a general pair ( B   , τ   ) with τ   a nontrivial 3 −   torsion point. In lemma  7.20 we have shown that h 0 ( A ~ 6 )   equals the dimension of the kernel of H 1 ( F )   , where F   is determined by f 0 , f 1 , f 2 , f 3   : therefore we are assuming that for our choices of B , τ   and general choice of the f i   's H 1 ( F )   has a kernel of dimension 3   . Letting now f 1   tend to zero we obtain in the limit case ℳ I I   where, as shown in the proof of proposition  7.16 (since in this case τ   is not a 2 −   torsion), h 0 ( A ~ 6 ) = 2   :
this contradicts the semicontinuity theorem.
The case τ = [ 0 ]   , is more difficult, since in this case we only know 2 ≤ h 0 ( A ~ 6 ) ≤ 4   .
Also in this case we use lemma  7.20 . Dualizing, we see that H 0 ( A ~ 6 )   is isomorphic to the cokernel of the map F ′   from
H 0 ( ( O B ⊕ L 1 ⊕ L 2 ⊕ L 3 ⊕ O B ⊕ L 3 ⊕ L 2 ⊕ O B ⊕ L 1 ⊕ ⊕ L 2 ⊕ L 3 ⊕ L 1 ⊕ O B ⊕ L 1 ⊕ L 3 ⊕ L 2 ) ( [ 0 ] ) )  
to H 0 ( O B ⊕ L 1 ⊕ L 2 ⊕ L 3 ⊕ O B ⊕ L 3 ⊕ L 2 ⊕ O B ⊕ L 1 ) ( 2 [ 0 ] )   given by the matrix M   .
Since we are assuming V 2   to be indecomposable, all the f i   's are not zero and therefore each f i   generates the corresponding H 0   . We can then write explicitly generators of the image of F ′   as vectors whose entries have the form f i f j   .
The pairs { f 0 f i , f j f k }   give bases of H 0 ( L i ( 2 [ 0 ] ) )   (for { i , j , k } = { 1 , 2 , 3 }   ); the remaining 4   elements ( f 0 2 , f 1 2 , f 2 2 , f 3 2   ) generate H 0 ( O B ( 2 [ 0 ] ) )   : therefore there are two independent relations among them.
We identify ( f 0 , f 1 , f 2 , f 3 )   to scalar multiples of the functions ( θ 11 , θ 10 , θ 01 , θ 00 ) ,   where the θ i j   's are the half-integer theta functions θ i j ( z , μ )   (cf. e.g.
[Mum] , page 17: here μ   is a point of Poincaré 's upper half plane).
Therefore the relations among the f i   's come from the relations between the θ i j   's; using the relations ( E 1 )   and ( E 2 )   in [Mum] page 23, we can write them as f 2 2 = a f 0 2 + b f 1 2   , f 3 2 = c f 0 2 + d f 1 2   , with a , b , c , d   that vary freely in an open set of C 4   according to the choice of B   and of the f i   's.
We have now given a basis of each respective space of global sections, and we may write explicitly (depending on the parameters a , b , c , d   ) the corresponding matrix of F ′   . We wrote a Macaulay 2 script (available upon request) that told us
  • ∙   that F ′   is generically injective ( h O ( A ~ 6 ) = 2   ),
  • ∙   F ′   acquires a kernel of dimension 1   ( h O ( A ~ 6 ) = 3   ) on a subvariety of codimension 1   of the space of parameters
  • ∙   F ′   acquires a kernel of dimension 2   in codimension 4   (therefore never, the parameters a , b , c , d   giving only three moduli up to automorphisms).
We get then two (possibly empty) families:
i ) h O ( A ~ 6 ) = 2   gives a family of dimension at most 3 + 1 = 4   ; i i ) H O ( A ~ 6 ) = 3   gives a family of dimension at most 3 − 1 + 2 = 4   .
Q . E . D .   We can now conclude our classification theorem End of the proof of theorem  7.2 .
We have shown the existence of three irreducible components of dimension 5   : the main stream component whose general point is contained in ℳ I ∘   , and the two ones contained in ℳ I I I   , according to proposition  7.17 .
We have also seen that every other family in the moduli space has dimension at most 4   ; by remark  7.13 we conclude that ℳ ′   has three irreducible components of dimension 5   .
It remains to show that they do not intersect. The two components in ℳ I I I   do not intersect, since (proposition  7.17 ) in one case τ   is [ 0 ]   , in the other case it is a non trivial 2 −   torsion point.
Since a general point of the third component is contained in ℳ I ∘ ¯   it is then enough to show that ℳ I ∘ ¯ ∩ ℳ I I I = ∅   . It will be then enough to show that there is no flat family on a disc whose central fibre is a surface in ℳ I I I   and whose general fibre is a general surface in M I ∘   .
We borrow an argument used often in Horikawa's work. We note that by theorem  7.5 a general surface of type M I ∘   is a double cover of C   with irreducible branch curve. On the contrary we have shown that the same double cover, for a surface in ℳ I I I   , has a disconnected branch curve. Therefore we would have a family of double covers with a connected general branch curve and with a disconnected special branch curve, a contradiction.
Q.E.D.

8 Fibrations with non hyperelliptic general fibre: the case g = 3  

In the rest of the paper we will concentrate on the case when the general fibre of f : S → B   is nonhyperelliptic of genus g = 3   .
The natural morphism of graded O B   -algebras σ : S y m ( V 1 ) → ℛ ( f )   has as kernel the graded sheaf of ideals ℒ   and we denote as usual by T   its (graded) cokernel.
Remark 8.1. Letting X : = P r o j ℛ ( f )   the rational map ψ 1 : X → Σ ⊂ P r o j ( S y m V 1 ) = P ( V 1 )   allows to factor the relative canonical map φ : S → Σ   as ψ 1 ∘ r   , where r : S → Y   is a birational morphism.
Observe that, if the general fibre of f   is non hyperelliptic, all the above maps are birational and ℒ   is the ideal sheaf of the canonical image Σ : = φ ( S )   .
In the next lemma, we will investigate the sheaf T   : under the assumption that the general fibre is non hyperelliptic, we know that all the sheaves T n   are torsion sheaves on B   .
Main assumption: We will assume (often without explicit mention) in the rest of the section that every fibre of the genus 3   fibration f   is 2 −   connected.
The above assumption is used in the proof of lemma  8.2 . Weakening this assumption in lemma  8.2 should lead to a generalization of the main theorem of this section.
Lemma 8.2. Let f   be a genus 3 fibration whose general fibre is non hyperelliptic and such that every fibre is 2-connected.
Then: 1) the sheaf T 2   is the structure sheaf of a divisor τ   on B   ; 2) the sheaves T n   are (as O B   -modules) free O τ   -modules of rank 2 n − 3   ; 3) the multiplication map V 1 ⊗ V 2 → V 3   induces an isomorphism V 1 ⊗ T 2 ∼ = T 3   .
Proof. The argument is similar to the one given in Lemma  5.1 .
By the classification of genus 3   fibres due to M. Mendes Lopes (cf.
[M-L] ), and by the hypothesis of 2-connectedness, a fibre F   is either nonhyperelliptic, i.e., it has a canonical ring of the form C [ x 1 , x 2 , x 3 ] / < F 4 ( x i ) > ,   or it is honestly hyperelliptic, i.e. it has a canonical ring of the form R = C [ x 1 , x 2 , x 3 , y ] / < r 1 : = Q ( x i ) , r 2 : = y 2 − G ( x i ) > ,   where deg x i = 1   , deg y = 2   , deg Q = 2   , deg G = 4   .
If t   is a local parameter in B   such that the point p : = { t = 0 }   is the image of a hyperelliptic fibre, the relation r 1   lifts to a relation r ¯ 1 = Q ¯ ( x i , t ) + μ ( t ) y   where Q ¯ ( x i , 0 ) = Q ( x i )   , and μ ( 0 ) = 0   . The assumption that the generic fibre is nonhyperelliptic imposes μ ≢ 0   . If s > 0   is the multiplicity of F   in 0   , we can assume, up to a holomorphic change of the local parameter, that r ¯ 1 = Q ¯ ( x i , t ) + t s y   and then ( T 2 ) p = O B , P / t s   , which proves the first part of the statement.
We remark (using first the lift of r 2   to eliminate the multiples of y 2   ), that the set { t i q j y | i < s }   is a basis for the stalk of T k   at p   (as a complex vector space) if the set { q j }   yields a basis for the homogeneous part of degree k − 2   of the quotient ring C [ x 1 , x 2 , x 3 ] / ( Q )   . We see then that the stalk at p   of T k   is a free O B , p   -module with basis { q j y }   , which gives the second part of the statement.
For k = 3   the induced basis is { x 1 y , x 2 y , x 3 y }   , i.e., exactly the image of the natural basis of V 1 ⊗ T 2   , and the lemma is proven.
Q.E.D. In the next lemma we investigate the sheaf ℒ   . Denote by π : P ( V 1 ) → B   the canonical projection.
Lemma 8.3. If f   is a genus 3   fibration whose general fibre is nonhyperelliptic and such that every fibre is 2-connected, then 1) ℒ 2 = ℒ 3 = 0   .
2) deg V 1 = χ − 2 ( b − 1 )   , deg τ = K 2 − 3 χ ( O S ) − 10 ( b − 1 )   .
3) ℒ 4   is a line bundle of degree deg V 1 − deg τ   .
4) the maps ℒ 4 ⊗ S n − 4 ( V 1 ) ↪ ℒ n   are isomorphisms: in particular the “relative canonical image” Σ   is a divisor on P ( V 1 )   belonging to the linear system | O P ( V 1 ) ( 4 ) ⊗ π * ℒ 4 v |   .
Proof. We have already noticed that all the maps σ n   are generically surjective, so that their kernels are vector bundles of rank ( n + 2 2 ) − 4 n + 2   for n ≥ 2   . We get
  • ∙   ℒ 2 = ℒ 3 = 0   ;
  • ∙   ℒ 4   is a line bundle;
  • ∙   ∀ n ≥ 4   , rank ℒ n =   rank S n − 4 ( V 1 )   .
By remark  2.9  deg V 1 = χ ( O S ) − 2 ( b − 1 ) , deg V 2 = χ ( O S ) + K 2 − 18 ( b − 1 )   .
The exact sequence 0 → S 2 ( V 1 ) → V 2 → O τ → 0   yields then deg τ = deg V 2 − 4 deg V 1 = K 2 − 3 χ ( O S ) − 10 ( b − 1 ) ,   and we have proven the second part of the statement.
By remark  2.9  deg V 2 = 6 K 2 + χ ( O S ) − 98 ( b − 1 )   . Moreover deg S 4 ( V 1 ) = 20 deg V 1 = 20 χ ( O S ) − 40 ( b − 1 )   . By lemma  8.2  deg T 4 = 5 deg τ = 5 K 2 − 15 χ ( O S ) − 50 ( b − 1 )   , and we conclude
deg ℒ 4 = deg S 4 ( V 1 ) + deg T 4 − deg V 4 = = 5 ( χ ( O S ) + K 2 ) − 90 ( b − 1 ) − 6 K 2 − χ ( O S ) + 98 ( b − 1 ) = = 4 χ ( O S ) − K S 2 + 8 ( b − 1 ) = deg V 1 − deg τ .  
The fourth assertion is obvious since Σ   is a Cartier divisor in P ( V 1 )   .
Q.E.D. Now, with the informations on T   and ℒ   provided by lemmas  8.2 and lemma  8.3 , we can investigate the O B   -algebra structure of ℛ ( f )   , i.e the multiplication maps μ i , j : V i ⊗ V j → V i + j   .
For i = j = 1   we have the composition V 1 ⊗ V 1 → S 2 ( V 1 ) → σ 2 V 2   .
The next proposition shows that the triple ( V 1 , V 2 , σ 2   ) determines ℛ ( f )   in degree ≤ 3   .
Definition 8.4. Let A : V 1 ⊗ Λ 2 ( V 1 ) → S 2 ( V 1 ) ⊗ V 1   be defined by A ( c ⊗ ( a ∧ b ) ) : = b c ⊗ a − a c ⊗ b   and B : Λ 3 ( V 1 ) → V 1 ⊗ Λ 2 ( V 1 )   be defined by B ( a ∧ b ∧ c ) : = a ⊗ ( b ∧ c ) + b ⊗ ( c ∧ a ) + c ⊗ ( a ∧ b ) .  
Proposition 8.5. V 3 =   coker ( ( σ 2 ⊗ I d ) ∘ A ) : V 1 ⊗ Λ 2 ( V 1 ) → V 2 ⊗ V 1 )   and μ 2 , 1   is given by the projection onto the cokernel.
Proof. Consider the diagram
0 0 0
↑ ↑ ↑
0 → S 3 ( V 1 ) ⟶ σ 3 V 3 ⟶ T 3 → 0
↑ ↑ μ 2 , 1 ↑
0 → S 2 ( V 1 ) ⊗ V 1 ⟶ σ 2 ⊗ I d V 2 ⊗ V 1 ⟶ T 2 ⊗ V 1 → 0
↑ A ↑ ( σ 2 ⊗ I d ) ∘ A ↑
0 → V 1 ⊗ Λ 2 ( V 1 ) = V 1 ⊗ Λ 2 ( V 1 ) ⟶ 0
↑ B ↑
Λ 3 ( V 1 ) = Λ 3 ( V 1 )
↑ ↑
0 0 .
where the map S 2 ( V 1 ) ⊗ V 1 → S 3 ( V 1 )   is the natural one in S y m ( V 1 )   while the map T 2 ⊗ V 1 → T 3   is the push forward of μ 2 , 1   (it is an isomorphism by lemma  8.2 ).
It is then easy to check that the diagram commutes with exact rows and columns. This proves our statement.
Q.E.D. We will now analyse our algebra in degree 4   .
Definition 8.6. Let C : S 2 ( Λ 2 ( V 1 ) ) → S 2 ( S 2 ( V 1 ) )   be defined by C ( ( a ∧ b ) ( c ∧ d ) ) : = ( a c ) ( b d ) − ( a d ) ( b c ) .  
Lemma 8.7. The map S 2 ( σ 2 ) ∘ C : S 2 ( Λ 2 ( V 1 ) ) → S 2 ( V 2 )   is injective with locally free cokernel.
Proof. We have to show that the map is injective on every fibre. If φ   is map between vector bundles on B   , for every p ∈ B   we will denote with ( φ ) ( p )   the corresponding linear map between the fibres over p   .
The map C   is injective on every fibre because the rank of V 1   is three the injectivity fails in general, since, in terms of a basis x 1 , . . . x n   it sends ( x i ∧ x j ) ( x h ∧ x k )   to ( x i x h ) ( x j x k ) − ( x i x k ) ( x j x h )   ).
We saw in the proof of lemma  8.2 that, if F p   is not hyperelliptic, ( σ 2 ) ( p )   is an isomorphism, so ( S 2 ( σ 2 ) ∘ C ) ( p )   is injective. On the points p   of B   s.t. F p   is hyperelliptic the kernel of ( σ 2 ) ( p )   has rank 1   .
We can choose a basis x 0 , x 1 , x 2   for the fibre of V 1   over p   such that this kernel is generated either by x 0 2   , or by x 0 2 + x 1 2   , or by x 0 2 + x 1 2 + x 2 2   , according to the rank of the conic canonical image of the corresponding hyperelliptic fibre. We denote this conic by q   .
This choice induces a basis of S 2 ( Λ 2 ( V 1 ) )   given by three vectors of the form ( x i ∧ x j ) ( x i ∧ x j )   and three of the form ( x i ∧ x j ) ( x i ∧ x k )   .
By definition:
C ( ( x i ∧ x j ) ( x i ∧ x j ) ) = ( x i 2 ) ( x j 2 ) − ( x i x j ) ( x i x j )   C ( ( x i ∧ x j ) ( x i ∧ x k ) ) = ( x i 2 ) ( x j x k ) − ( x i x j ) ( x i x k )   The 6   image vectors of the base elements are linearly independent. It remains to show that they span a subspace which intersects transversally the subspace of S 2 ( S 2 ( V 1 ) )   given by the multiples of q   . This is straightforward:
it suffices to send to zero the subspace spanned by the elements ( x i 2 ) ( x j x k )   (any i , j , k   ).
Q.E.D.
Definition 8.8. 1) Given a curve B   , a rank 3   vector bundle V 1   on B   , an effective divisor τ   , and an extension 0 → S 2 ( V 1 ) → V 2 → O τ → 0   where V 2   is still a vector bundle, we will denote by V ~ 4   the sheaf S 2 ( V 2 ) / S 2 ( Λ 2 ( V 1 ) ) =   coker ( S 2 ( σ 2 ) ∘ C )   . Lemma  8.7 shows that it is a locally free vector bundle of rank 15   .
2) Note that coker C = S 4 ( V 1 )   , so the image of S 2 ( σ 2 ) ∘ C   is in the kernel of the map S 2 ( V 2 ) → V 4   , inducing a map V ~ 4 → V 4   :
we will denote by ℒ 4 ′   its kernel.
Proposition 8.9.
  • 1) The maps S 2 ( V 2 ) → V 4   and V ~ 4 → V 4   are surjective.
  • 2) ℒ 4 ∼ = det V 1 ⊗ O B ( − τ )   .
  • 3) ℒ 4 ′ ∼ = det V 1 ⊗ O B ( τ )   .
  • 4) Consider, for a points p   in the support of τ   , the maps induced on the fibres by the embedding ℒ 4 ′ ↪ V ~ 4   and by the composition map V 2 ⊗ S 2 ( V 1 ) → S 2 ( V 2 ) → V ~ 4   .
    Their images are vector subspaces intersecting only in zero.
Proof. The first part is an immediate consequence of the fact that ∀ p   the map S 2 ( H 0 ( F p , ω F p 2 ) ) → H 0 ( F p , ω F p 4 )   is surjective, as made clear by the explicit description of the canonical rings given in the proof of lemma  8.2 .
Recall that the map ℒ 4 ↪ S 4 ( V 1 )   defines Σ   as a divisor in | O P ( V 1 ) ( 4 ) ⊗ π * ( ℒ 4 ) v |   , as shown in lemma  8.3 .
The dualizing sheaf of P ( V 1 )   is ω P ( V 1 ) = O P ( V 1 ) ( − 3 ) ⊗ π * ( det V 1 ⊗ ω B )   (cf. [Har] , ex. III.8.4.(b)), therefore the dualizing sheaf of Σ   is the restriction of O P ( V 1 ) ( 1 ) ⊗ π * ( ℒ 4 v ⊗ det V 1 ⊗ ω B )   .
The morphism (cf. remark  8.1 ) ψ 1 : X → Σ   is an isomorphism when restricted to the preimage of B \ S u p p ( τ )   . We will now write local equations for X   in a neighborhood of a fibre F p   with p ∈ S u p p ( τ )   .
Let us now consider a representative G ′   in S 2 ( V 2 )   of the image of a local generator of ℒ 4 ′   in V ~ 4   . Recall that setting t = 0   in the relative canonical algebra we get the canonical ring of the fibre F p   , where the restrictions of the generators x i   of V 1   are the generators in degree 1   , y   is the generator in degree 2   . F p   is honestly hyperelliptic, so, completing the square, we may assume that the canonical ring of Y p   is C [ x 1 , x 2 , x 3 , y ] / ( Q ( x 1 , x 2 , x 3 ) , y 2 − G ( x 1 , x 2 , x 3 ) )   as seen in the proof of lemma  8.2 .
Recall that by the proof of lemma  8.2 , for a suitable local parameter t   in O B , p   , we can choose a local basis q 1 , … , q 6   of S 2 ( V 1 ) p   , respectively q 1 ′ , … , q 5 ′ , y   of V 2 , p   so that the map σ 2   on the stalks over p   can be written as σ 2 , p ( q i ) = q i ′   for i ≤ 5   , and σ 2 , p ( q 6 ) = t s y   . Note that ∀ i ≤ 5   we can assume q i = ∑ λ i j k x j x k   not to depend upon t   .
Clearly, q 6   specializes (for t = 0   ) to Q   . Moreover, adding multiples of Q   we can chose G   so that it belongs to the square of the ideal generated by the the q i   's, with i ≤ 5   , i.e., G = ∑ i , j = 1 5 a i , j q i q j   .
Then (up to a nonzero constant) G ′   specializes to y 2 − ∑ a i , j q i ′ q j ′   .
Moreover we can replace y   by y ⋅ ( 1 + t ⋯ )   so that G ′   has the form y 2 − ∑ i , j = 1 5 α i , j ( t ) q i ′ q j ′   where α i , j ( 0 ) = a i , j   .
Now we look for the image of a local generator of ℒ 4   in S 4 ( V 1 )   , i.e.
a generator of the kernel of the map S 4 ( V 1 ) → V 4   that we can factor as S 4 ( V 1 ) = S 2 ( S 2 ( V 1 ) ) S 2 ( Λ 2 V 1 ) → S 2 ( V 2 ) S 2 ( Λ 2 V 1 ) = : V ~ 4 → V ~ 4 ℒ 4 ′ = V 4 .   We are then looking for a generator of the subspace in S 4 ( V 1 )   of the elements that are mapped to a multiple of G ′   : the explicit equations we wrote shows that we can choose q 6 2 − t 2 s ∑ α i , j q i q j   .
Now we can locally represent our fibration as follows: chosen a suitably small open neighborhood U   of p   , Σ ∩ π − 1 ( U )   is the subvariety of π − 1 ( U ) = U × P 2   defined by the principal ideal generated by q 6 2 − t 2 s ∑ α i , j q i q j   , whence X   is the subvariety of U × P ( 1 , 1 , 1 , 2 )   defined by ideal generated by t s y − q 6   and y 2 − ∑ α i , j q i q j   . The map ψ 1 : X → Σ   is the one induced by the natural projection P ( 1 , 1 , 1 , 2 ) → P 2 = P ( 1 , 1 , 1 )   , so it is a finite map with conductor ideal ⊂ O Σ   containing the ideal ( t s , q 6 )   . If we view the conductor as an ideal ⊂ O X   , we know that, since both X   and Σ   are Cohen Macaulay, it is pure of codimension 1   , moreover its radical is ( t )   , therefore we easily conclude that the corresponding conductor divisor on X   is locally exactly D i v ( t s )   . We conclude that globally the conductor divisor on X   is r * f * ( τ )   .
Then the dualizing sheaf ω S   of S   is φ * ω Σ ⊗ f * O B ( τ ) v = φ * O P ( 1 ) ⊗ f * ( ( ℒ 4 ) v ⊗ O B ( − τ ) ⊗ det V 1 ⊗ ω B ) ,   and therefore the relative dualizing sheaf ω S | B   is ω S | B ∼ = φ * O P ( 1 ) ⊗ f * ( ( ℒ 4 ) v ⊗ O B ( − τ ) ⊗ det V 1 ) .   On the other hand by definition of the relative canonical map ω S | B = φ * O P ( 1 )   , and (in accordance with lemma  8.3 , part 3) part 2 of the statement follows.
We have the following commutative diagram with exact rows and columns:
0 0
↑ ↑
0 → ℒ 4 → S 4 ( V 1 ) ⟶ σ 4 V 4 → T 4 → 0
| | ↑ ↑
0 → S 2 ( S 2 ( V 1 ) ) S 2 ( Λ 2 V 1 ) ⟶ S 2 ( σ 2 ) / S 2 ( Λ 2 V 1 ) V ~ 4 → K → 0
↑
ℒ 4 ′
↑
0
By lemma  8.2  T 4 = O τ ⊕ 5 ,   and K   is isomorphic to coker S 2 ( σ 2 )   , whence K ∼ = O τ ⊕ 5 ⊕ O 2 τ .   Computing determinants in the three exact sequences of the diagram above, one finds:
det V 4 ⊗ ℒ 4 ′ = det V ~ 4 = det S 4 ( V 1 ) ⊗ O B ( 7 τ )
det V 4 ⊗ ℒ 4 = det S 4 ( V 1 ) ⊗ O B ( 5 τ )
therefore ℒ 4 = ℒ 4 ′ ( − 2 τ )   .
Finally we want to prove that the embedding ℒ 4 ′ ↪ V ~ 4   has the property stated in 4). So, consider a point p   in the support of τ   . The fibre over p   of ℛ ( f )   , i.e. the stalk modulo the maximal ideal of the point, is the canonical ring of the corresponding fibre of f   . In degree ≤ 4   we can now see that it has three generators x 1 , x 2 , x 3   in degree 1   given by V 1   , one generator in degree 2   , y   , generating T 2   , one relation, Q ( x i )   , generating the kernel of the map induced by σ 2   at the level of the fibres, and one further relation in degree 4   , given by a representative in S 2 ( V 2 )   of the image of a generator of ℒ 4 ′   . We know by the connectedness hypothesis that this relation has the form y 2 = G ( x i )   , whence the image of V 2 ⊗ S 2 ( V 1 )   is given by the polynomials without the term y 2   .
Q.E.D.
Definition 8.10. Summarizing the previous results, to every genus 3   fibration with nonhyperelliptic general fibre and such that every fibre is 2-connected, we associate its genus three 5-tuple
  • ∙   ( B   , V 1   , τ   , ξ   , w   ), where
  • ∙   ξ   is the element ξ ∈ E x t 1 ( O τ , S 2 ( V 1 ) ) / A u t ( O τ )   yielding the exact sequence 0 → S 2 ( V 1 ) → f * ( ω S | B 2 ) → O τ → 0   ,
  • ∙   if i   is the embedding ℒ 4 ′ ↪ V ~ 4 = S 2 ( V 2 ) / S 2 ( Λ 2 ( V 1 ) )   , w   is the corresponding point in P ( H 0 ( V ~ 4 ⊗ ℒ ′ 4 − 1 ) )  
We know moreover that i) ξ   yields a vector bundle ( V 2   ); ii) ℒ 4 ′ ∼ = O B ( τ ) ⊗ det V 1   ; iii) i   has locally free cokernel; iv) the image of the corresponding linear map on the fibre of every point p   in the support of τ   is not contained in the image of V 2 ⊗ S 2 ( V 1 )   .
The next theorem will show that these data determine the fibration.
Definition 8.11. 1) Define an admissible genus three 5-tuple ( B   , V 1   , τ   , ξ   , w   ) as follows:
  • ∙   B   is a curve,
  • ∙   V 1   is a rank 3   vector bundle V 1   ;
  • ∙   τ   is a 0 −   dimensional subscheme of B   ;
  • ∙   ξ   is an element of E x t 1 ( O τ , S 2 ( V 1 ) ) / A u t ( O τ )   yielding a vector bundle V 2   .
  • ∙   If V ~ 4   is the vector bundle constructed via σ 2   as in definition  8.8 , then w = P ( i )   , where i   is an embedding O B ( τ ) ⊗ det V 1 ↪ V ~ 4   with:
    i) locally free cokernel; ii) the image of the corresponding linear map on the fibre of every point p   in the support of τ   is not contained in the image of V 2 ⊗ S 2 ( V 1 )   .
2) Define its associated relative canonical model X   as follows:
letting W ⊂ P ( V 2 )   be the image of the rational Veronese map P ( V 1 ) ‒ ‒ ⃗ P ( V 2 )   induced by σ 2   , we let X   be the relative quadric divisor on W   corresponding to the principal ideal generated by the image of i   .
Remark 8.12. One can explicitly, cf. the proof of  8.13 , define a sheaf of graded O B   algebras ℛ   , generated by S y m ( V 1 ⊕ V 2 )   such that X = P r o j ( ℛ )   . The geometric procedure is more suitable to determine the singularities of X   .
Theorem 8.13. Consider an admissible genus three 5-tuple ( B   , V 1   , τ   , ξ   , w   ).
Then, if its associated relative canonical model X   has at most rational double points as singularities, there is one and exactly one genus 3   nonhyperelliptic fibration f : S → B   with the property that every fibre is 2 −   connected and whose associated 5 −   tuple is the given one. S   is indeed the minimal resolution of the singularities of X   .
The surface S   thus obtained has invariants χ ( O S ) = deg V 1 + 2 ( b − 1 )   , K S 2 = 3 deg V 1 + deg τ + 16 ( b − 1 )   .
Finally, every genus 3   nonhyperelliptic fibration with every fibre 2 −   connected (and S   smooth) yields an admissible genus three 5-tuple.
Proof. Step 1: we construct the truncation in degree ≤ 4   of the relative canonical algebra ℛ ≤ 4 = ⊕ 0 4 V i   .
The class of ξ   determines an exact sequence 0 → S 2 ( V 1 ) → σ 2 V 2 → O τ → 0 .   We can define as in  8.5  V 3 : =   coker ( σ 2 ⊗ I d ) ∘ A   , and μ 2 , 1   as the natural projection.
One can prove that V 3   is automatically locally free. We only sketch the proof, because we do not need it for the proof of theorem.
Considering the diagram in the proof of proposition  8.5 (defining T 3   as coker σ 3   ) one can prove that it has again exact rows and columns.
Writing explicitly the maps with the local expression of σ 2 , p   given above one can prove that V 3   is automatically locally free.
For the part in degree 4   we work as in proposition  8.9 .
Let us define ℒ 4 ′ : = O B ( τ ) ⊗ det V 1   , and V 4 : =   coker ( ℒ 4 ′ ↪ V ~ 4 )   which is locally free by our assumption.
The map μ 2 , 2   is the composition of the maps V 2 ⊗ V 2 → S 2 ( V 2 ) → V ~ 4 → V 4 .   that is automatically surjective. Let us notice that the whole structure in degree ≤ 4   is determined.
In fact to complete the structure we only need a map μ 1 , 3 : V 1 ⊗ V 3 → V 4   fulfilling the following (associativity) equality: μ 1 , 3 ∘ ( I d V 1 ⊗ μ 1 , 2 ) = μ 2 , 2 ∘ ( μ 1 , 1 ⊗ I d V 2 )   . This problem has always a uniquely determined solution because I d V 1 ⊗ μ 1 , 2   is surjective and has kernel contained in the kernel of μ 2 , 2 ∘ ( μ 1 , 1 ⊗ I d V 2 )   .
The map S 2 ( S 2 ( V 1 ) ) → S 2 ( σ 2 ) S 2 ( V 2 )   naturally induces a map S 4 ( V 1 ) → V ~ 4   , whose composition with the projection on V 4   will be denoted by σ 4   .
Define ℒ 4 : = ker σ 4   , T 4 : =   coker σ 4   , and K : =   coker S 2 ( σ 2 ) / S 2 ( Λ 2 V 1 ) : S 4 ( V 1 ) → V ~ 4   : we have a commutative diagram with exact rows and column as the one in the proof of proposition  8.9 .
Let us remark that K   is naturally isomorphic to the cokernel of S 2 ( σ 2 )   , that is to O τ ⊕ 5 ⊕ O 2 τ   ; the hypothesis ii) says that on each point p   in the support of τ   , the projection on the factor O 2 τ   of the composed map ℒ 4 ′ → K   does not belong to m p   , which implies T 4 = O τ ⊕ 5   .
Step 2: To finish the construction of the sheaf of algebras ℛ   , since outside S u p p ( τ )   ℛ   coincides with S y m ( V 1 ) / ℒ 4 S y m ( V 1 )   , it suffices to construct ℛ   as a quotient of S y m ( V 1 ⊕ V 2 )   locally at the points p ∈ S u p p ( τ )   .
Since the cokernel of σ 2   has rank 1   at p   , we have the usual local picture (cf. the proof of the proposition  8.9 ): on the stalks at p   we can find a local basis q 1 , … q 6   for S 2 ( V 1 ) p   , resp. q 1 ′ , … q 5 ′ , y   for V 2 , p   such that ∀ i ≤ 5   , σ 2 , p ( q i ) = q i ′   , whence σ 2 , p ( q 6 ) = t s y   , where s   is the multiplicity of p   in τ   . As before, given a local basis x 1 , x 2 , x 3   for V 1   and writing the q i   's accordingly , we can assume that for i ≤ 5   they do not depend upon t   .
The assumption that the image of ℒ 4 ′   in V ~ 4   is not contained in the image of V 2 ⊗ S 2 ( V 1 )   guarantees that, when we consider in the stalk at p   a representative in S 2 ( V 2 )   of the image of a generator of ℒ 4 ′   , the coefficient of y 2   does not vanish for t = 0   . Again multiplying y   by a unit, completing the square allows us to assume that the above representative can be chosen of the form y 2 − ∑ i , j = 1 5 α i , j ( t ) q i ′ q j ′   .
Denote by π : P ( V 1 ) → B   the natural projection map.
The map ℒ 4 ↪ S 4 ( V 1 )   induces a divisor (i.e. a surface) Σ ∈ | O P ( V 1 ) ( 4 ) ⊗ ℒ 4 v | .   If U × P 2   is an open set of P ( V 1 )   inverse image of a sufficiently small neighborhood U p   of p   , Σ U p : = Σ ∩ ( U p × P 2 )   in suitable coordinates has equation q 6 2 − t 2 s ∑ α i , j q i q j   .
Σ U p   is dominated by the surface X U p   in U × P ( 1 , 1 , 1 , 2 )   defined by the ideal ( q 6 − t s y , y 2 − ∑ i , j = 1 5 α i , j ( t ) q i q j )   , and the map X U p → Σ U p   is an isomorphism on the complement of the fibre over p   .
It suffices now to observe that the above local algebra O B [ x 1 , x 2 , x 3 , y ] / ( q 6 − t s y , y 2 − ∑ i , j = 1 5 α i , j ( t ) q i q j )   is isomorphic to the quotient S y m ( V 1 ⊕ V 2 )   given by O B [ x 1 , x 2 , x 3 , q 1 ′ , . . . q 5 ′ , y ] / ( q j ( x ) − q j ′ , q 6 ( x ) − t s y , y 2 − ∑ i , j = 1 5 α i , j ( t ) q i q j ) .   In other words, let Σ p   be the fibre over p   , and define Σ ′ : = Σ \ ⋃ p ∈ S u p p ( τ ) Σ p .   Our X   is then constructed glueing all the X U p   's with Σ ′   .
By our assumption, X   has at most rational double points as singularities and we get a fibration f : S → B   by taking the minimal resolution of these singularities.
If a fibre F   of f   is not 2 −   connected then two things can happen:
either
  • ∙   0) F   is not 1   -connected, i.e., by Zariski's lemma F = 2 C   , where C   has arithmetic genus 2   or
  • ∙   1) F = A + B   , A ⋅ B = 1   .
In case 0), since O F ′ ( 2 K S ) = O F ′ ( 2 K F ′ )   , the relative canonical algebra is not generated in degree ≤ 2   .
In case 1), the relative canonical map has base points, contradicting the form of our equations.
Step 3: The invariants of S   can be easily computed by the equalities deg ℒ 4 = 4 χ ( O S ) − K S 2 + 8 ( b − 1 ) = deg V 1 − deg τ   and deg V 1 = χ ( O S ) − 2 ( b − 1 ) .   The rest of the proof is straightforward.
Q.E.D.

9 Genus three fibrations on some surfaces with p g = 3 , q = 0  

The present section is meant to show, via a concrete example, how the structure theorem for genus three fibrations  8.13 can be applied.
Let us then assume that S   is a minimal surface of general type with p g = 3 , q = 0   , and with K S 2 = 2 + d   . We shall make several simplifying assumptions:
Assumption I : we have a genus 3   fibration f : S → P 1   ( B ∼ = P 1   since q = 0   ) such that, for a general fibre F   , the restriction map H 0 ( K S ) → H 0 ( ω F )   is surjective.
It follows then that V 1 ∼ = ⊕ 1 3 O P 1 ( 2 )   . Our standard formulae read out as K S 2 = 2 + d e g ( τ ) ⇒ d = d e g ( τ ) .   Assumption II : d ≤ 6   and σ 2   yields a general extension class
0 → ⊕ 1 6 O P 1 ( 4 ) → ( ⊕ 1 d O P 1 ( 5 ) ) ⊕ ( ⊕ 1 6 − d O P 1 ( 4 ) ) → O τ → 0 . (12)
It follows that S 2 Λ 2 V 1 ∼ = ⊕ 1 6 O P 1 ( 8 )   , ℒ 4 ′ ∼ = O P 1 ( d + 6 )   and
S 2 ( V 2 ) = ( 6 − d 2 ) O P 1 ( 8 ) ⊕ d ( 6 − d ) O P 1 ( 9 ) ⊕ ( d 2 ) O P 1 ( 10 ) (13)
Proposition 9.1. Assume d ≤ 3   . Then X   is the complete intersection of W   with a relative quadric Z ∈ | O P ( V 2 ) ( 2 ) ⊗ π * ( O P 1 ( − 6 − d ) ) |   , π : P ( V 2 ) → P 1   being the natural projection.
Proof. Tensoring the exact sequence 0 → S 2 ( Λ 2 ( V 1 ) ) → S 2 ( V 2 ) → V ~ 4 → 0 .   by ℒ ′ 4 − 1   we see that the restriction map H 0 ( S 2 ( V 2 ) ⊗ ℒ ′ 4 − 1 ) → H 0 ( V ~ 4 ⊗ ℒ ′ 4 − 1 )   is surjective as soon as H 1 ( S 2 ( Λ 2 ( V 1 ) ) ⊗ ℒ ′ 4 − 1 ) = 0   . The result follows then from S 2 ( Λ 2 ( V 1 ) ) ⊗ ℒ ′ 4 − 1 ∼ = ⊕ 1 6 O P 1 ( 2 − d )   . Q.E.D.
Theorem 9.2. Let d ≤ 3   . Let σ 2   be a general homomorphism as in the exact sequence ( 12 ) and let ζ   be a general section of the vector bundle S 2 ( V 2 ) ⊗ ℒ ′ 4 − 1   (cf. ( 13 )). Denote by Z   the divisor of ζ   on P ( V 2 )   and set X = W ∩ Z   . Then X   is a smooth minimal surface of general type with p g = 3   , q = 0   and K X 2 = 2 + d   . We obtain in this way a subvariety of the moduli space of dimension 33 − 2 d   .
Proof. Let U 1 = C x 1 ⊕ C x 2 ⊕ C x 3   and V 1 = U 1 ⊗ C O P 1 ( 2 )   . Let q 1 , … , q 6   be a general basis of S y m 2 ( U 1 )   . Let σ 2   be given, in the corresponding basis of S y m 2 ( V 1 )   , by the diagonal matrix whose entries are the first 6   elements of a sequence constituted by 1   repeated 6 − d   times followed by t 0   , t 1   , t 0 − t 1   .
If q 4 , q 5 , q 6   yield irreducible conics, W   is a P 2   bundle except possibly over the points ( 0 , 1 , ∞ )   where the fibre can be the cone over a rational normal quartic curve. The singular points of W   are then exactly the cited vertices.
For d ≤ 2   , S 2 ( V 2 ) ⊗ ℒ ′ 4 − 1   is globally generated and therefore the linear system cut by Z   on W   has no base points. So the smoothness assertion is proven.
If d = 3   we set U 2 : = C w 1 ⊕ C w 2 ⊕ C w 3   , U 2 ′ : = C z 1 ⊕ C z 2 ⊕ C z 3   and V 2 = ( U 2 ⊗ C O P 1 ( 4 ) ) ⊕ ( U 2 ′ ⊗ C O P 1 ( 5 ) )   . It follows easily that in this case the base locus of the linear system of Z   is the P 2   bundle of equations { z 1 = z 2 = z 3 = 0 }   .
Assume now that q 4   , q 5   , q 6   are smooth conics and that q 4 ∩ q 5 ∩ q 6 = ∅   .
Then we see that W   does not intersect the base locus of Z   , and that X   is smooth.
The minimality of X   follows since H 0 ( K X )   is given by x 1 , x 2 , x 3   and restricts to the canonical system of each fibre. By the local equation that we have one sees that this last is base point free.
We compute the dimension of the constructed subvariety of the moduli space. The choice of τ   gives d   parameters, the extension class ξ   (yielding σ 2   ) gives 6 d − d = 5 d   parameters since we have to quotient by A u t ( O τ )   , the linear system cut by Z   on W   has dimension 44 − 8 d   .
We have therefore 44 − 2 d   parameters and we have to subtract the automorphisms of P 1   (dimension 3   ) and of V 1   (dimension 8   ).
Q.E.D.
Remark 9.3. Minimal surfaces with K S 2 = 2 , p g = 3   are well known since the time of Noether to be double covers of P 2   branched on a curve D   of degree 8   . Whence, their moduli space is unirational of dimension 44 − 8 = 36   . However, the existence of such a genus three fibration implies that there is a one dimensional family of quartic curves C t   , images of the fibres, which are everywhere tangent to D   . It follows that D   is a square modulo C t   , thus the equation of D   can be written as a symmetrical determinant d e t ( A B B C ) = 0   . Each such matrix yields conversely such a family, but we obtain in this way only a 33   -dimensional family, since 3 × 15 − d i m G L ( 2 ) − d i m P G L ( 3 ) = 33   .
This shows how our above examples yield proper subvarieties of the moduli space.
References

  1. Artin, M., On isolated rational singularities of surfaces. Amer. J. Math. 88 (1966), 129–136.
  2. Atiyah, M. F., Vector bundles over an elliptic curve. Proc. Lond. Math. Soc. (3) 7 (1957), 414–452.
  3. Barth, W., Peters, C., Van de Ven, A., Compact complex surfaces. Ergebnisse der Mathematik und ihrer Grenzgebiete (3) [Results in Mathematics and Related Areas (3)], 4. Springer-Verlag, Berlin, 1984. x+304 pp.
  4. Beauville, A., L'inégalité p g ≥ 2 q − 4   pour les surfaces de type général. Bull. Soc. Math. France 110 (1982), no. 3, 344–346.
  5. Bombieri, E., Canonical models of surfaces of general type. Inst. Hautes Études Sci. Publ. Math. No. 42 (1973), 171–219.
  6. Bombieri, E., unpublished manuscript (1972), 151.
  7. Catanese, F., On a class of surfaces of general type. Proc. C.I.M.E. Conference 'Algebraic surfaces', 1977, Liguori Editore, Napoli (1981), pp. 269–284.
  8. Catanese, F., Pluricanonical Gorenstein curves. Enumerative geometry and classical algebraic geometry (Nice, 1981), pp. 51–95, Progr. Math., 24, Birkhäuser Boston, Boston, MA, 1982.
  9. Catanese, F., Singular bidouble covers and the construction of interesting algebraic surfaces., Proceedings of the Warsaw Conference in honour of F. Hirzebruch's 70th Birthday , A.M.S. Contemp. Math. 241 (1999), 97–120 .
  10. Catanese, F., Fibred surfaces, varieties isogenous to a product and related moduli spaces. Amer. J. Math. 122 (2000), no. 1, 1–44.
  11. Catanese, F,, Ciliberto, C., Surfaces with p g = q = 1   . Problems in the theory of surfaces and their classification (Cortona, 1988), 49–79, Sympos. Math., SSSII, Academic Press, London, 1991.
  12. Catanese, F., Ciliberto, C., Symmetric products of elliptic curves and surfaces of general type with p g = q = 1   . J. Algebraic Geom. 2 (1993), no. 3, 389–411.
  13. Catanese, F., Ciliberto, C., Mendes Lopes, M. On the classification of irregular surfaces of general type with nonbirational bicanonical map. Trans. Amer. Math. Soc. 350 (1998), no. 1, 275–308.
  14. Catanese F., Franciosi, M., Divisors of small genus on algebraic surfaces and projective embeddings. Proceedings of the Hirzebruch 65 Conference on Algebraic Geometry (Ramat Gan, 1993), 109–140, Israel Math. Conf. Proc., 9, Bar-Ilan Univ., Ramat Gan, 1996.
  15. Catanese F., Franciosi, M., Hulek, K., Reid, M., Embedding of curves and surfaces. Nagoya Math. J. 154 (1999), 185–220.
  16. Catanese, F., Pignatelli, R., On simply connected Godeaux surfaces. Complex analysis and algebraic geometry, 117–153, de Gruyter, Berlin, 2000.
  17. Catanese, F., Schreyer, F.-O., Canonical projection of irregular algebraic surfaces. Algebraic geometry, 79–116, de Gruyter, Berlin, 2002.
  18. Ciliberto, C. The bicanonical map for surfaces of general type. Algebraic geometry—Santa Cruz 1995, Proc. Sympos. Pure Math., 62, Part 1, Amer. Math. Soc., Providence, RI, (1997), 57–84.
  19. Clemens, H., Cohomology and obstruction I: on the geometry of formal Kuranishi theory. math.AG/9901084.
  20. Enriques, F., Le superficie algebriche. Zanichelli, Bologna, 1949.
  21. Esnault, H., Viehweg, E., Effective bounds for semipositive sheaves and for the height of points on curves over complex function fields. Algebraic geometry (Berlin, 1988). Compositio Math. 76 (1990), no. 1-2, 69–85.
  22. Franchetta, A., Sul sistema aggiunto ad una curva riducibile. Atti Accad. Naz. Lincei. Rend. Cl. Sci. Fis. Mat. Nat. (8) 6, (1949), 685–687.
  23. Franchetta, A., Sui sistemi pluricanonici di una superficie algebrica. Univ. Roma. Ist. Naz. Alta Mat. Rend. Mat. e Appl. (5) 8, (1949), 423–440.
  24. Franciosi, M. Adjoint divisors on algebraic curves. Adv. Math. 186 (2004), no. 2, 317–333.
  25. Fujita, T., On Kähler fiber spaces over curves. J. Math. Soc. Japan 30 (1978), no. 4, 779–794.
  26. Fujita, T., The sheaf of relative canonical forms of a Kähler fiber space over a curve. Proc. Japan Acad. Ser. A Math. Sci. 54 (1978), no. 7, 183–184.
  27. Hartshorne, R., Algebraic Geometry. Graduate Texts in Mathematics, No. 52. Springer-Verlag, New York-Heidelberg, 1977. xvi+496 pp. ISBN: 0-387-90244-9.
  28. Horikawa, E., On algebraic surfaces with pencils of curves of genus 2   . Complex analysis and algebraic geometry, pp. 79–90. Iwanami Shoten, Tokyo, 1977.
  29. Horikawa, E., Algebraic surfaces of general type with small c 1 2   . V. J. Fac. Sci. Univ. Tokyo Sect. IA Math. 28 (1981), no. 3, 745–755
  30. Kodaira, K. , On compact analytic surfaces. II, III. Ann. of Math. (2) 77 (1963), 563–626; ibid. 78 1963 1–40.
  31. Konno, K. Clifford index and the slope of fibered surfaces. J. Algebraic Geom. 8 (1999), no. 2, 207–220.
  32. Konno, K. 1-2-3 theorem for curves on algebraic surfaces. J. Reine Angew. Math. 533 (2001), 171–205.
  33. Mendes Lopes, M., The relative canonical algebra for genus three fibrations. Ph. D. Thesis, University of Warwick, 1989.
  34. Mumford, D., Tata lectures on theta I. Progress in Mathematics, 28. Birkhuser Boston, Inc., Boston, MA, 1983. xiii+235 pp. ISBN: 3-7643-3109-7.
  35. Ogg, A. P., On pencils of curves of genus two. Topology 5 (1966), 355–362.
  36. Ramanujam, C. P. Remarks on the Kodaira vanishing theorem. J. Indian Math. Soc. (N.S.) 36 (1972), 41–51, Supplement: J. Indian Math. Soc. (N.S.) 38 (1974), no. 1, 2, 3, 4, 121–124 (1975).
  37. Reid, M., Surfaces with p g = 0   , K 2 = 1   . J. Fac. Sci. Univ. Tokyo Sect. IA Math. 25 (1978), no. 1, 75–92.
  38. Reid, M., Problems on pencils of small genus. manuscript 1990.
  39. Shafarevich, I. R., Lectures on minimal models and birational transformations of two dimensional schemes. Notes by C. P. Ramanujam. Tata Institute of Fundamental Research Lectures on Mathematics and Physics, No. 37 Tata Institute of Fundamental Research, Bombay (1966),iv+175 pp.
  40. Xiao, G., Surfaces fibrées en courbes de genre deux. Lecture Notes in Mathematics, 1137. Springer-Verlag, Berlin, 1985. x+103 pp. ISBN: 3-540-15662-3
  41. Xiao, G., Fibered algebraic surfaces with low slope. Math. Ann. 276 (1987), no. 3, 449–466.
  42. Zucconi, F., A note on a theorem of Horikawa. Rev. Mat. Univ. Complut. Madrid 10 (1997), no. 2, 277–295.

Authors' addresses Prof. Dr. F. Catanese Lehrstuhl Mathematik VIII Universität Bayreuth Universitätsstr. 30, D-95447 Bayreuth Dr. R. Pignatelli Dipartimento di Matematica Università di Trento Via Sommarive, 14, I38050 Povo (TN)