ON CAPORASOS CONJECTURE ON

BRILL-NOETHER LOCI FOR TRIVALENT GRAPHS

TAKEO NISHINOU
Abstract. We report on our result on Brill-Noether theory on
graphs. In particular, we prove Caporasos conjecture which states
that the existence of graphs on which there are no special divisors
if the Brill-Noether number is negative. In fact, we can prove
full Brill-Noether theorem for a class of trivalent graphs. This,
combined with Bakers specialization lemma, gives a purely combinatorial proof of classical Brill-Noether theorem.

1. Introduction
This is a report on the authors recent result concerning the divisor theory of graphs. The divisor theory of graphs has undergone
big progress after the seminal work of Baker and Norine [2] about
Riemann-Roch theory of graphs appeared. Since then, this theory has
been developed in several directions. One of remarkable results is the
Brill-Noether theory of graphs developed by Cools, Draisma, Payne and
Robeva [4], which is a deepening of the Riemann-Roch theory. They
showed (among other results) the existence of particular four-valent
graphs which are Brill-Noether general, that is, the set of divisors of
given Brill-Noether number (see Section 2) has the expected dimension. However, graph theoretically four-valent graphs are not general:
general graphs are trivalent. So the natural question is whether there
exist trivalent graphs which are Brill-Noether general.
In this context, L. Caporaso conjectured [3] that there should be
trivalent graphs which has the property that the set of eective divisors
whose Brill-Noether number is negative is empty. In this report, we
give an example of a class of trivalent graphs which conrms Caporasos
conjecture. In fact, these graphs are Brill-Noether general (that is, the
Brill-Noether theorem holds even when the Brill-Noether number is
nonnegative).
2. Review on divisor theory on graphs
In this section we recall the divisor theory on metric graphs very
briey. See for example [2] for more information.
2.1. Divisors on graphs. Let be a compact connected metric graph.
A divisor on is a nite Z-linear combination of points of , and we
1

write Div() for the additive group of all divisors. A general element
of Div() is written as
D = a1 v1 + + ar vr ,
where ai Z and vi . We dene the degree of D by
deg(D) = a1 + + ar .
A rational function on is a piecewise linear function on with
integral slopes. If f is a rational function of and v , dene ordv (f )
by the sum of incoming slopes of f at v. Then dene div(f ) Div()
by

div(f ) =
ordv (f ) v
v

and call it the divisor of f . The divisors of rational functions on

compose a subgroup P rin() of Div(), the subgroup of principal divisors.
The Picard group of is dened by
P ic() = Div()/P rin().
One sees every principal divisor has degree 0, so there is a well-dened
map
deg : P ic() Z.
An element of P ic() is called a divisor class. Denote by P icd () the
subset deg 1 (d) of P ic(). In particular, P ic0 () is a subgroup of
P ic().
2.2. Linear system of divisors. A divisor D = a1 v1 + + ar vr
Div() is eective if each coecient ai is nonnegative. A divisor D0 is
linearly equivalent to D if D D0 P rin().
The rank r(D) of an eective divisor D is the largest integer r such
that, for every eective divisor E of degree r, D E is linearly equivalent to an eective divisor. If D is not linearly equivalent to an eective
divisor, then set r(D) = 1.
For nonnegetive integers r and d, the Brill-Noether locus
Wdr () P icd ()
is the set of divisor classes of degree d and rank at least r.
In classical theory of divisors on Riemann surfaces, for an Riemann
surface of genus g, the Brill-Noether locus Wdr (C) is similarly dened.
Let
(g, r, d) = g (r + 1)(g d + r)
be the Brill-Noether number. Then it is known that for a general curve,
if (g, r, d) is nonnegetive, the dimension of Wdr (C) is equal to (g, r, d),
and if (g, r, d) is negative, then Wdr (C) is an empty set.
2

For graphs however, there are open sets in the moduli space of metric
graphs of genus g, where dim Wdr () is strictly larger than (g, r, d) (see
[3]). One of the conjectures of Caporaso is the following.
Conjecture 2.1. Assume g 2 and (g, r, d) < 0. Then there exists
a 3-regular graph of genus g for which Wdr () = .
Theorem 2.2. Caporasos conjecture holds for the following graphs.
l2

l0

l7

l4

l1

l3

l3g-4

l5

l3g-5

l3g-2

l3g-1

l3g-3

l6

Figure 1. The graph g

Here the edge lengths of the edges li , 1 i 3g 3 are given by
length(li ) = i ,
with small positive constant.
Remark 2.3. The edge length need not strictly take these values. If we
perturb the lengths of edges slightly (compared to its original length, for
example, we can change the length of the edge li to i + O(i+1 )), then
the resulting graph still satises the conclusion of Theorem 2.2. Thus,
we actually have the open subset of the moduli space of the metric graphs
where the conclusion of Theorem 2.2 holds.
Remark 2.4. The same line of argument proves that in fact the full
Brill-Noether theorem is true for these graphs. That is, when the BrillNoether number is nonnegative, then the corresponding Brill-Noether
locus has dimension . In this note, we concentrate on the case < 0.
3. Idea of the proof
Let us write the graph in the statement of Theorem 2.2 by g . Recall
that Brill-Noether number is given by
= g (r + 1)(g d + r).
One sees that = 1 if and only if the genus g of g , the degree d of
the divisor and rank r of the linear system satisfy the relation
(g, d) = (r(p 1) + p 2, pr 1)
for some positive integer p. The cases when is smaller than 1 are
easier than the case = 1, so we consider the cases where = 1.
3

An important tool for the proof is the chip-ring deformation which

we introduce below. It gives a convenient way to construct linearly
equivalent divisors of a given divisor.
Recall the chip-ring move on a graph [1, 2]. Let be a graph and
D a divisor on it. Let v be a k-valent vertex of and v1 , , vk be the
neighboring vertices of v. Then applying a chip-ring move at v to D
gives the new divisor D0
{
D0 = D + kv v1 v2 vk (borrowing move), or
D0 = D kv + v1 + v2 + + vk (giving move).
Any divisor linearly equivalent to D can be obtained as a result of a
sequence of chip-ring moves.
In the case of a metric graph, a straightforward generalization of
chip-ring move can be dened, which we call chip-ring deformation.
To dene it, we prepare some notation. Let be a compact connected
metric graph. Let v be a k-valent point of (a point on the interior of
an edge is a divalent point. The valency at the vertices of is dened
as usual). Consider the open subset \ {v} of . This is a graph with
k open ends. Add one valent vertices to each of these ends. This gives
a closed graph 0 which is not necessarily connected.
Definition 3.1. We call the graph 0 the graph obtained from by
cutting at v.
Let D be a divisor on whose summand at the point v is lv with
l > 0.
Definition 3.2. A divisor D0 on 0 is obtained from D by cutting the
pair (, D) at v if
k

D0 = D lv +
c i vi
with ci 0 and

i=1 ci

i=1

= l.

Note that D lv does not have a summand at the point v, so we

can think of it as a divisor on the graph 0 . Thus, the expression of D0
in Denition 3.2 makes sense.
Now we dene the chip-ring deformation. For simplicity, we only
dene it for the case of positive ends. Let be a metric graph and D
an eective divisor on . Let
v1 , , v a
be points of and 0 be the graph obtained from be cutting it at
each of the points v1 , , va . Let
01 , , 0b
be the connected components of 0 . Assume that by cutting the pair
(, D) suitably at each of the points v1 , , va , we obtain a component
4

of 0 (let it be 01 for simplicity) and an eective divisor D10 on it which

have the following property:
Each one-valent vertex is contained in the support of the divisor
D10 .
Let
w1 , , w c
be the one-valent vertices of 01 and write D10 as
c

00
0
ai wi
D1 = D1 +
i=1

with ai > 0 for all i. Here D100 is the subdivisor of D10 whose support is
disjoint from the one-valent vertices of 01 .
Take a positive number so that it is not greater than the length of
any leaf. Here a leaf is an edge of 01 which has one of w1 , , wc as
one of its ends. Let 01, be the graph obtained from 01 by shortening
each leaf fi , i = 1, , c by the length ai . Here fi has wi as one of its
ends and ai is the coecient of wi in D10 (see the above expression for
D10 ).
Let wi, be the point of 01, corresponding to wi in the obvious way.
If the length of the leaf fi is greater than , then wi, is again a onevalent vertex, and if the length of fi is equal to , then wi, coincides
with another end of fi . We point out that in the latter case, some wi,
and wi0 , may give the same point even if i 6= i0 . So we have a divisor
c

ai wi,

i=1

01, .

on
Note that the graph 01, is naturally
ca sub metric graph of although
0
1 is not in general. So the divisor i=1 ai wi, can be seen as a divisor
on . Also note that
c

ai wi
i=1

naturally gives an eective divisor on by mapping wi to one of the

points from v1 , , va where it is cut. Let
c

ai w
i
i=1

be this divisor on .
c
i of D to a divisor
Summarizing,
we
replaced
the
subdivisor
i=1 ai w
c
i=1 ai wi, , and obtained a new eective divisor
=D
D

ai w
i +

i=1

i=1

ai wi, .

 from D a chipDefinition 3.3. We call the above process to obtain D

ring deformation.
The following results hold for chip-ring deformation, analogously to
the case of chip-ring move.
is obtained from D by a chip-ring deformation,
Lemma 3.4. If D

then D is linearly equivalent to D.


Proposition 3.5. If D0 is an eective divisor which is linearly equivalent to D, then D0 is obtained from D by a sequence of chip-ring
deformations.

As an example of the use of the notion of chip-ring deformation, we
prove the following result.
Lemma 3.6. Let li , i {0, 1, 4, , 3g 2} be a vertical edge of g .
Let D be an eective divisor on g whose degree is bounded by 3g.
Assume that the intersection
Supp(D) int(li ),
where int(li ) = li \ li , is one point whose coecient is 1 and which is
contained in a medium neighborhood of the middle point of li . Then, if
D0 is any eective divisor linearly equivalent to D, the intersection
Supp(D0 ) int(li )
is not empty.
Remark 3.7. Note that the right most vertical edge l3g1 is excluded
from the statement of the lemma.
Proof. Let p be the point Supp(D) int(li ). Let a, b be the two end
points of li . For any divisor

E=
b j qj
jJ

on g , consider the following real valued function fi . Namely, let

E0 =
b k qk
kK

be the subdivisor of E where

Supp(E) li = {qk }kK .
Then
fi (E) =

bk (d(a, qk ) d(b, qk )),

kK

where d is the distance function on g . By assumption,

1
fi (D) i + O(i+1 ) mod i .
2
6

If the statement of the lemma is not true, there is an eective divisor

D0 linearly equivalent to D such that
fi (D0 ) 0 mod i+1 .
Consider a chip-ring deformation which changes the value of fi .
Obviously the relevant graph 01 must contain a part of the edge li .
Then there are two cases:
(1) There are two vertices of 01 contained in li .
(2) There is only one vertex of 01 contained in li .
In the case (1), the graph 01 is a subsegment of li , or contains the
complement of a subsegment of li . In each case, it is clear that the
corresponding chip-ring deformation does not change fi . In the case
(2), the graph 01 contains a part of some edge lj with j > i. Then,
the corresponding chip-ring deformation can change the value of fi at
most O(i+1 ). Since the degree of the divisor D is bounded by 3g, it
follows that we can change the value of fi at most O(i+1 ), proving the
lemma.

We name the vertices of the graph g as follows:
v1

l0

l7

l4

l1

v2

v4

l3

v2g-5 l3g-7 v2g-3

l5 v5

v3

l2

l3g-8

l3g-4

v2g-4 l3g-3 v2g-2

v6

l6

l3g-5

Figure 2
Consider the following open graph k+1 of the graph g :
v1

l0

v3

l2

l7

l4

l1

v2

l3

v2k-1 l3k-1 v2k+1

l5 v5

v4

l6

l3k-2

v2k

v6

l3k

v2k+2

Figure 3
It has k + 1 vertical edges l0 , l1 , l4 , , l3k2 . Let a0 , a1 , a4 , , a3k2
be the middle points of these edges. Let Sk+1 be the set of ordered
sequences of k + 1 nonnegetive integers (z0 , z1 , z4 , , z3k2 ) satisfying
z0 + z1 + z4 + + z3k2 = k + 1.
7

Let Dk+1 be the following set of eective divisors

Dk+1 = {z0 a0 +z1 a1 +z4 a4 + +z3k2 a3k2 | (z0 , z1 , z4 , , z3k2 ) Sk+1 }.
Using chip-ring deformation, we can also prove the following result.
Proposition 3.8. Let D be an eective divisor of degree d on g ,
where d is a positive integer not larger than 3g. Let E = D k+1 be
the subdivisor of D whose support is contained in k+1 , where k is a
xed nonnegative integer with k g 1. Assume the divisor D satises
the following two conditions:
(a) For any element of Dk+1 , there exists a divisor linearly equivalent to D which contains as a subdivisor.
(b) D is linearly equivalent to an eective divisor D0 such that the
subdivisor E 0 = D0 k+1 satises
deg E 0 k,
such that the
Then D is linearly equivalent to an eective divisor D
k+1 satises
subdivisor E = D
deg E 2k + 2.

Using this proposition, we can prove the following result.
Proposition 3.9. Let D be an eective divisor on g with deg(D) 3g
such that for any member of Dk+1 , k g 1, there is a divisor D0
linearly equivalent to D such that D0 contains as a subdivisor. Then
linearly equivalent to D, the degree of the
for any eective divisor D

intersection E = D k+1 is at least k.


Now we outline the proof of Theorem 2.2. In this note we give a
detailed argument only for the case p = 2. Namely,
(g, d) = (r, 2r 1).
Assume that there is an eective divisor D on g with deg(D) = 2g 1
which has rank at least r. Let a, b be points on the edge l3g1 such that
d(v2g3 , a) = 3g1 103g2 and d(v2g3 , b) = 3g1 53g2 . Let ab
be the interval on l3g1 of length 53g2 whose end points are a, b. Let
0g be the open subgraph of g dened by
0g = g \ ab.
Apply Proposition 3.9 to the graph 0g . Note that the graph 0g is
not the same as the graph g in Proposition 3.9, the proposition is
applicable to 0g . Then any eective divisor D0 linearly equivalent to D
contains a subdivisor of degree at least g whose support is contained
in 0g . On the other hand, there must be an eective divisor linearly
equivalent to D which has ga as a subdivisor, where a is a point in ab.
However, such a divisor has degree at least 2g, a contradiction. This
8

proves Theorem 2.2 for the case p = 1 ( = 1). The other cases with
= 1 can be proved by similar argument.

References
[1] Biggs, N., Chip-firing and the critical group of a graph. J. Algebraic Combinatorics., 9 (1999), 25-45.
[2] Baker, M. and Norine, S., Riemann-Roch and Abel-Jacobi Theory on a
Finite Graph., Advances in Mathematics 215 (2007), 766788.
[3] Caporaso,L.,
Algebraic and combinatorial Brill-Noether theory.
ArXiv:1106.1140.
[4] Cools, F., Draisma, J, Payne, S. and Robeva, E. A tropical proof of the
Brill-Noether Theorem. Advances in Mathematics. 230 (2012), 759-776.
Department of Mathematics, Rikkyo University, Tokyo, Japan
E-mail address: [email protected]