Notes on intersection theory on Mg

Tabes Bridges
Fall 2014

Basic Definitions
A stack can, for our purposes, be thought of as a hideous categorical black box designed to be
a juiced-up moduli space. In particular, although the coarse moduli scheme does not admit
a universal curve, and so does not satisfy the universal property of fine moduli spaces, the
universal curve is hardwired into the definition of the stack via the category of flat familes
X → B of genus g stable curves. Anyway, the coarse moduli space has finite quotient
singularities, so its associated groups of divisor classes and line bundle classes are equal after
tensoring with Q: A1 (Mg ) ⊗ Q = P ic(Mg ) ⊗ Q. Now instead of taking the time to define a
stack, at which point we would hope that the corresponding stacky definition would naturally
fall out of the discussion, we give a series of ad hoc definitions which contain the necessary
shadows of the general theory to allow us to accurately carry out calculations on stacks.
Definition. A rational divisor class on the moduli stack, that is to say an element of
P icf un (M g ) ⊗ Q, is a choice of class in P ic(B) ⊗ Q for every family of genus g stable curves
X → B such that if we carry out base change f : B 0 → B, the class of the base-changed
family is the pullback under f of the original class.
We now completely side-step the issue of properly defining bundles on stacks because it
apparently does not matter. The following isomorphism (and the recipe in the proof) will
allow us to somewhat freely pass between the worlds of stacks and mere mortals:
Proposition. P icf un (Mg ) ⊗ Q ∼
= P ic(Mg ) ⊗ Q.
Proof. Given a rational divisor class D on the moduli space, some integer multiple mD can
be represented by a line bundle. For every family X → B of stable curves, we get a map
ϕ : B → Mg so that we get a pullback bundle ϕ∗ (mD) on B. The rational divisor class
1 ∗
m
ϕ (mD) is the desired association.
The construction of the inverse map depends on a technical result, namely the existence
of a tautological family of stable curves whose base naturally surjects finitely onto Mg ;
we also need a variation, the existence of a family such that the natural map is surjective,
generically finite, and finite over a given point [C] in the moduli space. Now we can use
this to “compress” the information of a rational divisor class on the stack into a class on the
moduli space. Given such a class γ on the stack, choose a tautological family ρ : X → Ω,
take its associated divisor class γ(ρ) ∈ P ic(Ω) ⊗ Q, and push γ(ρ) forward to Mg via the

1
natural map Ω → Mg . Finally, divide by the degree of ϕ. The resulting divisor class is
independent of the choice of tautological family because any two tautological families are
covered by their fiber product, so the universal property of rational divisor classes on stacks
guarantees equality.

Exercise. We collect here a few facts, each finer than the last, which give an impression of
the flavor of these gadgets:

1. Any rational divisor class on the moduli stack is determined by its values on families
X → B with smooth, one-dimensional base B.

2. Let Σ ⊂ Mg be any proper subvariety. Any rational divisor class on the stack is
determined by its values on families with smooth, one-dimensional base such that the
image of B under the natural map B → Mg is not contained in Σ.

3. In view of the fact that P ic(Mg ) is discrete, we have that any rational divisor class
γ ∈ P icf un (Mg ) ⊗ Q is determined by the degrees deg(γ(ρ)) ∈ Q of its values on
families ρ : X → B with smooth, one-dimensional base B whose image in the moduli
space is not contained in Σ.

Solution. The first item should follow from the existence of suitably nice maps from an
arbitrary scheme to a smooth curve. .

Generalities on Calculating δ, κ, λ on 1-parameter Fami-

lies.
Let f : X → B be a flat family of stable curves with B a smooth curve. To compute the
number of nodal fibers in a family, it suffices to compute the number of nodes and divide by
the number of nodes in a general singular fiber:

# {nodes in X }
# {singular fibers in X → B} = .
# {nodes in a general Xb }

To compute the numerator, we may use a juiced up version of Riemann-Hurwitz.

Fk First
−1
we excise the singular fibers Xb1 , ..., Xbk to obtain X = f (B \ Γ) t i=1 Xbi , where Γ =
{b1 , ..., bk } ⊂ B. Now we compute the Euler characteristic on the locus where f is a fiber
bundle, then reintroduce the singular fibers by hand. With F a general fiber,
X
χ(X ) = χ(B \ Γ)χ(F ) + χ(Xbi )
i
X
= χ(B)χ(F ) − kχ(F ) + χ(Xbi )
i
X
= χ(B)χ(F ) + (χ(Xbi ) − χ(F )).
i

2
From topology, we can see that each degeneration to a node increases the Euler characteristic
by 1 (since each 1-cell in a CW decomposition of a Riemann surface contributes −1), so the
differences in the summation are the number of nodes in each fiber. Thus
X
# {nodes} = (χ(Xbi ) − χ(F )) = χ(X ) − χ(B)χ(F ).
i

While χ(B) and χ(F ) are generally clear (e.g. for a family of genus g curves over P1 ,
χ(B)χ(F ) = 2(2 − 2g) = 4 − 4g), χ(X ) is not always immediate. Again using CW decom-
positions in the analytic topology, we note that blowing up a smooth point has the effect
of excising a point (subtracting 1) and introducing a 2-sphere (adding 2), so blowing up a
surface in n smooth points increases its Euler characteristic by n.

General Pencil of Plane Quartics

Example. For a pair of general quartics V (F ), V (G) on P2 , consider the pencil X = V (sF +
tG) ⊂ P1 × P2 . Since a general line in the parameter space of plane quartics does not meet
the locus of curves with worse-than-nodal singularities (which has codimension 2), we obtain
a family of stable curves whose general fiber is a smooth curve of genus 3. Since a general
singular quartic has a single node (isn’t this pretty much the case for anything?), we see
that the number of nodes is exactly the number of singular fibers. By Bezout’s theorem, our
pencil has 16 base points, so the total space X is P2 blown up in these 16 points. Since the
usual stratification expresses P2 as the disjoint union of P1 and the contractible space A2 , we
conclude χ(P2 ) = 3, so we compute

δ = # {nodes} = χ(P2 ) + 16 − (4 − 4g) = 3 + 12 + 4(3) = 27.

Example. We now compute κ on the same pencil of plane quartics. Note that OP1 ×P2 (X ) =
OP1 ×P2 (1, 4) due to the bidegree of the defining equation. Pulling back from the factors, we
have KP1 ×P2 = OP1 ×P2 (−2, −3), and by adjunction,

KX = KP1 ×P2 ⊗ OP1 ×P2 (X )|X

= OP1 ×P2 (−2, −3) ⊗ OP1 ×P2 (1, 4)|X
= OP1 ×P2 (−2 + 1, −3 + 4)|X
= OX (−1, 1).

Now to obtain the relative dualizing sheaf, we “divide out” by differentials on the base:

ωX /P1 = KX ⊗ π1∗ KP∨1

= OX (−1, 1) ⊗ OX (2, 0)
= OX (1, 1).
R
Computing κ on this family amounts (I think) to evaluating X c1 (ωX /P1 )2 . Since the total
space of our pencil has class (1, 4), we are looking at computing (1, 4) · (1, 1)2 . Since the

3
product of projective spaces has Chow ring A∗ (P1 × P2 ) = Z [α, β] /(α2 , β 3 ), we calculate
directly

(α + 4β)(α + β)2 = (α + 4β)(|{z}

α2 +2αβ + β 2 )
=0
= 2α β +αβ + 8αβ 2 + 4β 3
2 2
| {z } |{z}
=0 =0
2
= 9αβ
= 9 [pt] ,

so κ = 9.
Example. Now to compute λ, we return to our calculation of the relative dualizing sheaf,
but push it down to P1 before taking the Chern class:

E = (π1 )∗ OX (1, 1)
= (π1 )∗ (OX (1, 0) ⊗ OX (0, 1))
= (π1 )∗ π1∗ (OP1 (1)) ⊗ (π1 )∗ OX (0, 1)
= OP1 (1) ⊗ (π1 )∗ OX (0, 1)
= OP1 (1) ⊗ OP⊕3 1
⊕3
= OP1 (1),

where the isomorphism (π1 )∗ OX (0, 1) = OP⊕31 (1) follows because a general divisor in the class

upstairs dominates P1 under π1 ? Roughly, the sections of the pushforward over an open set
should be the sections of OX (0, 1) over the preimage of that open set...not entirely clear on
the numerics of pulling a line bundle on a surface back to a line bundle on another surface,
then pushing this down to a rank 3 vector bundle on a curve. And I suppose the pushforward
is trivial because the bundle upstairs was only twisted over the P2 factor? Anyway, since the
first Chern class of a vector bundle is the sum of the first Chern classes of any splitting into
line bundles (formal or otherwise, as is the case here), we conclude λ = deg c1 (E) = 3.

Hyperplane Sections on a Smooth Quartic Surface.

Example. Let S ⊂ P3 be a smooth quartic surface cut out by F ∈ H 0 (P3 , OP3 (4)). Now
choose a pair of independent linear forms G, H ∈ H 0 (S, OS (1)), i.e. cut S by a pair of
independent hyperplanes. Since deg S = 4, the common intersection of these three is 4
points. Thus to obtain a family of stable curves over X → P1 , it suffices to blow up S in the
four basepoints of the linear system V (sG + tH) ⊂ P1 × S ⊂ P1 × P3 . Apparently S is a K3
surface and hence χ(S) = 24, or we can apparently use the Hopf index theorem for this as
well. I’m not sure how to check either of these. Anyway, since we blew up 4 points, we get
χ(X ) = 24 + 4 = 28. Since our curves are essentially smooth plane quartics, the general fiber
has genus 3, so δ = χ(X ) − (4 − 4g) = 28 − (4 − 12) = 36.
Example. Continuing with the pencil of hyperplane sections, we introduce pullback hyper-
plane classes on X , specifically, η = π ∗ OP1 (1) and ζ = α∗ OS (1), where π : χ → P1 and

4
α : X → S. Recall that the total space of this family is cut out in P1 × S by a bihomogeneous
polynomial of bidegree (1, 1), so we can write [X ] = η + ζ and thereby compute the classes
of the canonical bundle and the relative dualizing sheaf. By adjunction,

KP1 ×S = (−2η − 4ζ) + 4ζ = −2η.

To compute KX , we again use adjunction KX = (KP1 ×S + [X ])|C = −2η + (η +ζ) = ζ − η.

Again the class of the relative dualizing sheaf is just this guy plus 2η, so ωX /P1 = η + ζ.
Since OP1 (1) is the class of a point, η is the class of a fiber, so η 2 = 0. Since ζ is the class
of a line in S, η · ζ = 4. For some reason, ζ 2 = 4. Thus we have

κ = deg c1 (ωX /P1 ) = (η + ζ)2 = 2η · ζ + ζ 2 = 8 + 4 = 12.

Example. Now we finish with the pencil of hyperplane sections. Note that ωX /P1 = π ∗ OP1 (1)⊗
α∗ OS (1)). Pushing down to P1 , we obtain OP1 (1) on the first factor, while the second factor
is apparently a bundle of rank 3 whose fiber over p ∈ P1 is H 0 (OXp (1)), so apparently we
have a short exact sequence

0 → OP1 (1) → OP1 (1)⊕4 → π∗ α∗ OS (1) → 0.

Supposedly this should allow us to compute c1 and then λ, but I think there are typos that
I cannot get around.

A Less General Pencil

Example. Consider a general pair of plane quartics with exactly one intersection point of
contact order 2. After choosing the first quartic, there remains a codimension 2 locus Γ ⊂ P14
of quartics with a simple tangency to the first. The locus of unstable curves has codimension
2 in Γ, so the family parametrized by a general line in Γ consists entirely of stable curves.
Again, we can compute δ by counting nodes in the total space using Riemann-Hurwitz. This
pencil has 15 base points due to multiplicity, but we have to blow up the point of tangency
a second time, so as in the general case χ(X ) = 3 + 16 = 19. Again, our general fiber is a
smooth curve of genus 3, so we conclude δ = 19 − (4 − 4g) = 27.

The Hyperelliptic Locus in M 3

Write the class of the hyperelliptic locus Dhyp = aλ + bδ0 + cδ1 ∈ P ic(M 3 ⊗ Q). For a
general pencil of plane quartics C1 , we have intersection numbers C1 · λ = 3, C1 · δ0 = 27,
C1 · δ1 = 0. This is done in the third section above. Since the curves in this pencil are
canonically embedded, none are hyperelliptic, so we conclude 0 = 3a + 27b.
For our next test family C2 , take a general pencil of plane cubics with p one of its base
points. Next fix a curve of genus 2 and fix q a non-Weierstrass point; attach this curve to the
pencil by gluing p to q. We have intersection numbers C2 · λ = 1, C2 · δ0 = 12, C2 · δ1 = −1.
To calculate C2 · λ, we pullback to M 1,1 × M 2,1 . Since the second factor is constant, we need
only calculate the class on the pencil of cubics. The total space X is a divisor of type (1, 3)

5
on P1 × P2 , hence has canonical class KX = OX (−1, 0). The base is P1 , so ωX/P1 = OX (1, 0).
Pushing down to P1 , we get λ = 1. To calculate C2 · δ0 , we need the number of singualar
(rational nodal) fibers in the pencil. We computed this previously as χ(X) − χ(B)χ(F ),
the latter of which is 0 since F is a torus. To compute χ(X), note that X is P2 blown up
in 9 points, so χ(X) = 3 − 9(1) + 9(2) = 12. Finally, to calculate C2 · δ1 , note that this
family is contained in ∆1 since every fiber has an elliptic component, so we need to compute
the class of the normal bundle for the sections we glued to get the family. One component
is a constant family, so the relevant section has self-intersection 0. On the other hand, we
consider a constant section, which has self-intersection 0, but we have to blow up once, so
the overall self-intersection is −1.
For our third test family C3 , fix a curve C of genus 2 and fix a non-Weierstrass point
p ∈ C. Let q ∈ C vary, and let C3 be the family obtained by gluing p and q as q varies;
denote the total space by X. We have intersection numbers C3 · Dhyp = 1 (since the single
hyperelliptic fiber in C3 arises when q coincides with the image of p under the hyperelliptic
involution of C), C3 · λ = 0, C3 · δ0 = −4, C3 · δ1 = 1. To compute C3 · λ, observe that
we have an exact sequence on Xq = C/p ∼ q relating its differentials to those on C. Every
holomorphic differential on C descends to Xq , so we have H 0 (KC ) ⊂ H 0 (ωXq ). Since a section
of ωXq is strictly meromorphic if and only if it has a pole hence nonzero residue at p, we
obtain for each q ∈ C an exact sequence of vector spaces
resp
0 → H 0 (KC ) → H 0 (ωXq ) −−→ C → 0.

This globalizes and pushes forward to a sequence on C

0 → H 0 (KC ) ⊗ OC → π∗ (ωX/C ) → OC → 0.

We have C3 · λ = c1 (π∗ (ωX/C )) = c1 (H 0 (KC ) ⊗ OC ) + c1 (OC ) = 0.

The first Chern class of the middle guy is C3 · λ, and is equal to the sum of the first
Chern classes of the outer guys, hence trivial. To compute C3 · δ1 = 1, observe that all fibers
of C3 are irreducible except when p = q and the fiber sprouts a rational tail. On the other
hand, C3 lies in ∆0 , so we need to compute normal bundles again: the total space is C × C
blown up at (p, p). The relevant sections are a constant and the diagonal, which have self
intersections 0 and χ(C) = 2 − 2(2) = −2; after blowing up, these drop to −1 and −3, so
C3 · δ0 = −4.

Admissible Covers
The basic question is this: suppose we have a generically smooth family of stable rational
curves, and consider a family of branched covers Ct → P1 . When the base specializes (i.e.
P1 degenerates to a rational nodal curve B), can we extend the family to include a branched
cover of B? The answer turns out to be yes. If we treat the degeneration of P1 as the gradual
contraction of a loop γ ⊂ P1 , the inverse images of γ in the Ct consist of some disjoint loops
which degenerate as γ does. The inverse image π −1 (B \ {q}) of the smooth locus is smooth
with simple branching over the limits of the branch points on P1 . It remains to see what
happens over the node. By some monodromy analysis (is he really just saying that if when

6
winding around a singularity it takes k times to get back to the base point, then the cover
is locally of degree k?), we find that locally, the node of B is covered by a node ai -to-1 over
both branches, where ai is the length of the cycle corresponding to the connected component
of π −1 (γ) that degenerates to this particular node. Applying Riemann-Hurwitz to the two
irreducible components of the map and accounting for the number of points at which they
meet upstairs (the degree d minus non-simple branching), we get that the genus upstairs is
−d + 1 + 2b where b is the branching of the general fibers. Now applying Riemann-Hurwitz to
the general fiber (where the irreducible curve upstairs has genus g) confirms that the special
fiber is a stable curve of genus g. Generalizing,
Definition. Let (B, p1 , ..., pb ) be a stable b-pointed rational curve with nodes q1 , ..., qk . An
admissible cover of B is a nodal curve C together with a map π : C → B such that
1. The preimage of the smooth locus in B is precisely the smooth locus in C. Restricted
to the smooth locus, branching is permitted only over the pi and it must be simple.

2. The preimage of the singular locus in B is precisely the singular locus in C, for each qi
and each r ∈ π −1 (qi ), the branches meeting at r cover the branches meeting at qi with
the same ramification.
This naturally extends to the idea of a family of admissible covers over family of stable
b-pointed rational curves.
Since a cover of a rational curve is determined up to monodromy by the branch points,
a point of the Hurwitz scheme Hd,g is locally a finite quotient of a (2d + 2g − 5)-dimensional
polydisc. The compactification by admissible covers has boundary ∆ = Hd,g \ Hd,g . Each
node in the base increases the codimension by 1; for example, ∆δ ⊂ ∆δ−1 .
Exercise. How many irreducible components in the boundary?
Solution. My first thought is that this should depend on b, but this number seems to be
irrelevant to the definition so probably
not. Actually scratch that, b is a function of d
g+d−1 b
and g (b = 2 ). My thought is 2 , one component for each pair of branch points on
P1 that can be the “first” to collide, and then these intersect as in Keel’s theorem (the
subset relations), yielding relations if one were to calculate the cohomology? Or maybe
b


2
− 3 since an automorphism of P1 should induce an automorphism of π, but this might not
happen because we are worrying about monodromy...actually I don’t think that is an issue.
Monodromy just came up in determining which isomorphism classes of covers are admissible
in the first place. But my proposed −3 might actually arise as Keel-type relations as well
(the type pulled back from M 0,4 )? .
Example. Suppose we have b branch points, and p1 , p2 collide. A twig branches off with
these two points. Since the branching of the cover is simple, the monodromy around each
point is a transposition. If σ1 = σ2 , then σ1 σ2 = σ12 = 1, so the cover is unramified over the
node.
Since any component over the twig with p1 and p2 can branch only over those two points,
Riemann-Hurwitz allows for components mapping isomorphically, or degree two branched
covers by components with Euler characteristic 2χ(P1 ) − {1, 2} = 3, 2, so also rational. Any

7
higher degree cover with only 2 branch points exceeds the maximal Euler characteristic of 2
for a curve.
Apparently in the stable model (so the gadget in the boundary of Hd,g ) we blow down all
of these rational curves, leaving behind a node in place of the branched component. I’m not
sure what happens to the twig downstairs.

Definition. A nodal curve C 0 is stably equivalent to a stable curve C if we have C 0 → C

by blowing down all smooth rational components which meet the other components in ≤ 2
points.

Boundary of the Hyperelliptic Locus

Consider the family X → D obtained by gluing a fixed point p of a general elliptic curve E
to a general genus 2 curve D at a varying points q ∈ D. Let Cq = D t E/p ∼ q be the fiber
over q ∈ D. The fibers in X ∩ H3 are those which are stably equivalent to an admissible
double cover of a rational curve B. Clearly such a cover C has components isomorphic to D
and E which are each 2 : 1 over B. For some monodromatic reason the map must ramify
over the node, so q ∈ D must be a Weierstrass point, of which there are 6.
Now we must deal with the question of transversality of D → M 3 and H3 . We employ
another family, a pencil of cubics Et with a general point q on a fixed genus 2 curve D
attached to a base point p ∈ Et . If Et is smooth, Et ∪ D is never hyperelliptic since q is
general.