KNOT THEORY AND CLUSTER ALGEBRAS

VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

Abstract. We establish a connection between knot theory and cluster alge-

bras via representation theory. To every knot diagram (or link diagram), we
associate a cluster algebra by constructing a quiver with potential. The rank
of the cluster algebra is 2n, where n is the number of crossing points in the
knot diagram. We then construct 2n indecomposable modules T (i) over the
Jacobian algebra of the quiver with potential. For each T (i), we show that
the submodule lattice is isomorphic to the corresponding lattice of Kauffman
states. We then give a realization of the Alexander polynomial of the knot
as a specialization of the F -polynomial of T (i), for every i. Furthermore, we
conjecture that the collection of the T (i) forms a cluster in the cluster algebra
whose quiver is isomorphic to the opposite of the initial quiver, and that the
resulting cluster automorphism is of order two.

1. Introduction
We establish a connection from cluster algebras and representation theory to
knot theory. Let K be a knot diagram (or link diagram) with n crossings. A
segment of K is a segment of the strand from one crossing point to the next. We
associate to K a quiver Q with 2n vertices, one for each segment of K, as well as a
potential W . We then construct 2n indecomposable representations T (i), each of
which encodes the Alexander polynomial of the link.
To be more precise, each crossing point p of the diagram K gives rise to an
oriented cycle wp of length four in the quiver Q and each region R in K gives rise
to an oriented cycle wR in Q whose length is the number of segments at the region
R. The potential W is the difference of the sum of the crossing point cycles and
the sum of the region cycles.
Denote by B the Jacobian algebra of the quiver with potential over an alge-
braically closed field. Then B is a non-commutative algebra which may be infinite
dimensional. The representations T (i) are finite-dimensional B-modules. We con-
struct them explicitly as representations of the quiver Q by specifying a finite
dimensional vector space at every vertex and a linear map for every arrow in Q.
This construction is a representation theoretic analogue of the construction of the
Kauffman states in [14]. The direct sum T = ⊕T (i) is called the link diagram
module of K.
Let A be the cluster algebra with principal coefficients of the quiver Q as defined
in [12]. A is a commutative algebra with a special combinatorial structure. It
is defined as a subring of a field of rational functions by constructing a set of
The second author was supported by the NSF grants DMS-1800860 and DMS-2054561. This
work was partially supported by a grant from the Simons Foundation. The authors would like to
thank the Isaac Newton Institute for Mathematical Sciences for support and hospitality during the
programme Cluster Algebras and Representation Theory when work on this paper was undertaken.
This work was supported by: EPSRC Grant Number EP/R014604/1.
1
2 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

generators, the cluster variables, via a recursive method called mutation that is
determined by the quiver Q. Each mutation step connects two sets of 2n cluster
variables and these sets are called the clusters of A.
The cluster variables are Laurent polynomials in two sets of indeterminates xi
and yi , for i = 1, 2, . . . , 2n, with positive integer coefficients [11, 16]. Moreover,
their specialization, obtained by setting all xi equal to 1 is a polynomial, called the
F -polynomial [12].
It was shown in [9] that F -polynomials can also be computed from modules over
our Jacobian algebra B. If M is a B-module then its F -polynomial is
X
FM = χ(Gre (M )) ye ,
e
where the sum runs over all dimension vectors e = (ei )i=1,2,...,2n of submodules of
M and ye = y1e1 y2e2 . . . y2n
e2n
. Moreover, Gre (M ) is the quiver Grassmannian of M ,
meaning the variety of all submodules of M whose dimension vector is e, and χ
denotes the Euler characteristic. In general, this Euler characteristic is very hard
to compute, because it is known that every projective variety can be realized as a
quiver Grassmannian.
We show that the F -polynomials of our B-modules T (i) have a much simpler
formula, since every submodule is uniquely determined by its dimension vector.
Therefore X
FT (i) = ydim L ,
L⊂T (i)

where the sum runs over all submodules L of T (i). We write FT (i) |t for the spe-
cialization of the F -polynomial at

 −t if segment j runs from an undercrossing to an overcrossing;
yj = −t−1 if segment j runs from an overcrossing to an undercrossing;
−1 if segment j connects two overcrossings or two undercrossings.

(1)
The Alexander polynomial ∆K of an oriented link diagram K is an important
polynomial invariant of the link. It is a Laurent polynomial in one variable t
with integer coefficients. Introduced by Alexander in [1], it has several equivalent
definitions, see for example [18]. In this paper, we follow Kauffman’s approach that
realizes the Alexander polynomial as a state sum [14]. More recently, the Alexander
polynomial has been generalized in the work of Osváth and Szabó [23], as well as
Rasmussen [25], on knot Floer homology.

We are now ready to state our main result. Recall that a link is prime if it
cannot be decomposed as the connected sum of two non-trivial links.
Theorem 1.1. Let K be a diagram of a prime link. Then, for every segment i of
K, the Alexander polynomial of K is equal to the specialized F -polynomial of the
B-module T (i). That is
∆K = FT (i) |t .
We point our that the quiver Q, and hence the algebra B and the cluster algebra
A, is not an invariant of the link, because Q depends on the choice of the diagram
K. For one, the number of vertices in Q is equal to the number of segments in
K, which is not invariant under Reidemeister moves. Moreover, the definition of Q
 KNOT THEORY AND CLUSTER ALGEBRAS 3

does not take into account the difference between an overpass and an underpass in
K. This difference is only recovered in the specialization (1).
The key step in the proof is the following result which is of interest in its own
right.
Theorem 1.2. The lattice of Kauffman states of K relative to a segment i is
isomorphic to the lattice of submodules of the B-module T (i).
An interesting question is how the different T (i) are related to each other. We
conjecture the following. Recall that a B-module T is called a tilting module if the
projective dimension of T is at most one, Ext1B (T, T ) = 0 and there is an exact
sequence 0 → B → T 0 → T 1 with T 0 , T 1 in the additive closure of T . Also recall
that, given a quiver Q, its opposite quiver Qop is obtained by reversing the direction
of all arrows.
Conjecture 1.3. (a) The module T = ⊕i T (i) is a B-tilting module.
(b) The Gabriel quiver of the endomorphism algebra of T is isomorphic to the
quiver Qop .
(c) The 2n cluster variables in A corresponding to T form a cluster.
(d) There is a permutation σ of order two such that the mapping that sends the
initial cluster variable xi to the cluster variable corresponding to T (σ(i)) is
a cluster automorphism of order two in the sense of [2].
Evidence for the conjecture has been obtained in previous work by David Whiting
and the second author in [27]. They considered a very special family of links, namely
2-bridge links whose continued fraction has at most two parameters. For a slightly
simpler quiver than our quiver Q, they constructed some of the modules T (i) and
showed that their direct sum ⊕T (i) can be completed to a tilting module that
satisfies the conditions in the conjecture.
As an application, we use a well-known property of the Alexander polynomial to
show the following result that is related to the rank-unimodality conjecture of [20].
Theorem 1.4. Let M be a module of Dynkin type An and L the submodule lattice
of M . Then 
X
h(L) ±1 if |L| is odd;
(−1) =
0 if |L| is even,
L∈L
P
where h(L) = dim L = j∈Q0 dim Lj is the total dimension of the submodule L.
Relation to other work. A first connection between cluster algebras and knot theory
was given by Kyungyong Lee and the second author in [17] in the special case of
2-bridge links. The authors realized another invariant, the Jones polynomial, as
a specialization of a cluster variable in a cluster algebra of Dynkin type A. This
result was based on an ad hoc construction using the fact that both the 2-bridge
links and the cluster variables of type A are parametrized by continued fractions.
We now can see this correspondence as a special case of our general construction,
as explained in section 8. This provides a more conceptual explanation for the
results in [17]. However, we do not know how to generalize the Jones polynomial
specialization to arbitrary links.
Nagai and Terashima used ancestral triangles constructed from continued frac-
tions to give a formula for the cluster variables of type A and then defined a spe-
cialization that produces the Alexander polynomial of the corresponding 2-bridge
link, see [22]. Our specialization is a generalization of theirs.
4 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

In [6], Cohen, Dasbach and Russel gave a realization of the Alexander polynomial
for arbitrary knots as a sum over perfect matchings of the bipartite graph whose
vertices are given by the crossing points and the regions of the diagram. Their
graph can be recovered from our quiver by the methods used for plabic graphs,
see for example [10]. In the case of 2-bridge knots, their graph is equivalent to
the snake graph associated to the continued fraction in [5] and in that case their
formula seems to be a special case of the cluster variable expansion formula of [21]
and therefore may be related to ours as well. However, in their approach, the weight
of a perfect matching is given by edge weights, which in the cluster algebra setup
corresponds to x-variables, whereas we use the y-variables instead. For arbitrary
knots, it is unclear if their formula corresponds to a cluster variable.
All of the articles above consider a single segment of the link to produce a formula
for the invariant. In our approach, we rather aim at a conceptual understanding of
the collection of the 2n objects given by all of the segments of the link inside the
cluster algebra and in the module category of the Jacobian algebra.

The paper is organized as follows. After fixing the notation and recalling certain
facts and terminology in section 2, we review Kauffman’s construction of the state
poset and the state polynomial in section 3. In section 4, we define our quiver with
potential and its Jacobian algebra B. The link module T = ⊕i T (i) is constructed
in section 5. Section 6 is devoted to the proof of the lattice isomorphism in The-
orem 1.2. Then Theorem 1.1 is proved in section 7. We end the paper with the
special case of 2-bridge links and the proof of Theorem 1.4 in section 8.

Acknowledgments. We thank Dylan Rupel for pointing out a typo in an earlier

version of this article.

2. Preliminaries
We recall basic notions and results from knot theory and cluster algebras.

2.1. Knots and links. A knot is a subset of R3 that is homeomorphic to a circle.

A link with r components is a subset of R3 that is homeomorphic to r disjoint
circles. Thus a knot is a link with one component. Links are considered up to
ambient isotopy. A link is said to be prime if it is not the connected sum of two
nontrivial links.
A link diagram K is a projection of the link into the plane, that is injective except
for a finite number of double points that are called crossing points. In addition, the
diagram carries the information at each crossing point which of the two strands is on
top and which is below. A diagram is called alternating if traveling along a strand
alternates between overcrossings and undercrossings. A link is called alternating if
it has an alternating diagram. A link is said to be oriented if for each component
a direction of traveling along the strand is fixed.
A curl is a monogon in the diagram. We usually assume without loss of generality
that our link diagrams are without curls, because one can always remove them (by
a Reidemeister I move) without changing the link.
Throughout this paper, we assume that all links are prime and all link diagrams
have a finite number of crossing points.
 KNOT THEORY AND CLUSTER ALGEBRAS 5

K− K+ K0

Figure 1. Skein relations for the Alexander polynomial

2.1.1. The Alexander polynomial. The Alexander polynomial ∆ of an oriented link

1
is a polynomial invariant of the link ∆ ∈ Z[t± 2 ] that can be defined in terms of
homology, see [18, Chapter 6]. For the original definition of Alexander, see [1]. The
Alexander polynomial is defined up to multiplication by a signed power of t.
In [7], Conway showed that the Alexander polynomial ∆(K) of an oriented link
K can be defined recursively as follows. The Alexander polynomial of the unknot
is 1, and whenever three oriented links K− , K+ and K0 are the same except in the
neighborhood of a point, where they are as shown in Figure 1, then
∆K+ − ∆K− = (t1/2 − t−1/2 )∆K0 .
This property also provides a normalization of the Alexander polynomial, but we
will not use it here.
The Alexander polynomial ∆ has the following properties, see for example [18,
Chapter 6].
. .
(i) For any link, ∆(t) = ∆(t−1 ), where the symbol = means “equal up to a
signed power of t”.
(ii) ∆(1) = ±1 for any knot, and ∆(1) = 0 for any link with at least 2 compo-
nents.
(iii) For any knot
.
∆ = a0 + a1 (t−1 + t) + a2 (t−2 + t2 ) + . . .
with a0 odd.
(iv) If a knot has genus g then 2g ≥ breadth(∆), where the breadth is the
difference between the maximal and the minimal degree of the polynomial.
Kauffman gave a description of the Alexander polynomial as a state sum. This
approach is crucial for us and we review it in section 3.
Let us close this subsection by mentioning a recent breakthrough in a closely
related question. In 1982, Freedman showed in [13] that a knot in the 3-sphere is
topologically slice if its Alexander polynomial is trivial. A famous pair of knots with
11 crossings, the Kinoshita-Terasaka knot and the Conway knot are the smallest
non-trivial knots for which the Alexander polynomial is trivial. In particular, both
knots are topologically slice. The Kinoshita-Terasaka knot is also known to satisfy
the stronger property of being smoothly slice. Recently Lisa Piccirillo solved a long-
standing open problem in [24] by proving that the Conway knot is not a smoothly
slice knot. For an illustration of the quiver of the Conway knot see Example 9.3.

2.2. Cluster algebras. In this section, we recall the definition of a skew-symmetric

cluster algebra with principal coefficients following [12, 26]
Let P be the free abelian group on generators y1 , . . . , yn written multiplicatively.
Let ZP be the ring of Laurent polynomials in the variables y1 , . . . , yn and let QP
denote its field of fractions. Denote by F = QP(x1 , . . . , xn ) the field of rational
6 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

functions in n variables and coefficients in QP. We also define an auxiliary addition

⊕ by
Y a Y b Y min(a ,b )
yj j ⊕ yj j = yj j j
. (2)
j j j

The cluster algebra is determined by the choice of an initial seed (x, y, Q), which
consists of the following data.
• Q is a finite connected quiver without loops ◦ d and 2-cycles ◦ o /◦,
and with n vertices;
• y = (y1 , . . . , yn ) is the n-tuple of generators of P, called initial coefficient
tuple;
• x = (x1 , . . . , xn ) is the n-tuple of variables of F, called initial cluster.
The cluster algebra A = A(x, y, Q) is the ZP-subalgebra of F generated by
so-called cluster variables, and these cluster variables are constructed from the
initial seed by a recursive method called mutation. A mutation transforms a seed
(x, y, Q) into a new seed (x0 , y0 , Q0 ). Given any seed there are n different mutations
µ1 , . . . , µn , one for each vertex of the quiver, or equivalently, one for each cluster
variable in the cluster.
The seed mutation µk in direction k transforms (x, y, Q) into the seed µk (x, y, Q) =
(x0 , y0 , Q0 ) defined as follows:
• x0 is obtained from x by replacing one cluster variable by a new one, x0 =
x \ {xk } ∪ {x0k }, and x0k is defined by the following exchange relation
!
1 Y Y
xk x0k = yk xi + xi (3)
yk ⊕ 1
i→k i←k
where the first product runs over all arrows in Q that end in k and the
second product runs over all arrows that start in k.
• y0 = (y10 , . . . , yn0 ) is a new coefficient n-uple, where

−1
yk
 if j = k;
0
yj = y Y y (y ⊕ 1)−1 Y (y ⊕ 1) if j 6= k.
 j
 k k k
k→j k←j

Note that one of the two products is always trivial, hence equal to 1, since
Q has no oriented 2-cycles. Also note that y0 depends only on y and Q.
• The quiver Q0 is obtained from Q in three steps:
(1) for every path i → k → j add one arrow i → j,
(2) reverse all arrows at k,
(3) delete 2-cycles.
Mutations are involutions, that is, µk µk (x, y, Q) = (x, y, Q). Note that Q0 only
depends on Q, that y0 depends on y and Q, and that x0 depends on the whole seed
(x, y, Q).
Let X be the set of all cluster variables obtained by finite sequences of mutations
from (x, y, Q). Then the cluster algebra A = A(x, y, Q) is the ZP-subalgebra of F
generated by X .
By definition, the elements of A are polynomials in X with coefficients in ZP,
so A ⊂ ZP[X ]. On the other hand, A ⊂ F, so the elements of A are also rational
functions in x1 , . . . , xn with coefficients in QP.
 KNOT THEORY AND CLUSTER ALGEBRAS 7

2.2.1. Laurent phenomenon and positivity. We have the following important results.
Theorem 2.1 (Laurent Phenomenon). [11] Let u ∈ X be any cluster variable.
Then
f (x1 , . . . , xn )
u=
xd11 · · · xdnn
where f ∈ ZP[x1 , . . . , xn ], di ∈ Z.
Theorem 2.2 (Positivity). [16] The coefficients of the Laurent polynomials in The-
orem 2.1 are positive in the sense that f ∈ Z≥0 P[x1 , . . . , xn ].
2.2.2. F -polynomials. Let u be any cluster variable in the cluster algebra A =
A(x, y, Q). By the two theorems above, we can write u as a positive Laurent
polynomial in the initial cluster as u = Lu ∈ Z≥0 [x1 , . . . , xn , y1 , . . . , yn ]. Then the
F -polynomial of x is defined as the evaluation of Lu at x1 = · · · = xn = 1. Thus
Fx = Lu (1, . . . , 1, y1 , . . . , yn ).

2.3. Quivers with potential. In this subsection, we recall an alternative ap-

proach to F -polynomials using quivers with potential. Let Q be a finite quiver.
Following [8] we let the vertex span of Q be the commutative algebra R over C with
basis the constant paths ei , i ∈ Q0 and multiplication ei ej = δi,j ei . Furthermore,
the arrow span of Q is the R-bimodule A with C-basis the set of arrows Q1 and
R-bimodule structure ei Aej = ⊕α:j→i Cα. Q∞
The complete path algebra of Q then isQRhhAii = d=0 A⊗R d , with m-adic topol-
∞
ogy given by the two-sided ideal m = d=1 A⊗R d . The elements of RhhAii are
(possibly infinite) C-linear
P combinations of paths in Q.
A potential W = c∈mcyc λc c on Q is a (possibly infinite) linear combination
of cyclic paths in RhhAii. The cyclic derivative ∂α , for α ∈ Q1 is defined on a
non-constant cyclic path α1 . . . αd by
X
∂α (α1 . . . αd ) = αp+1 . . . αd α1 . . . αp−1 ,
p : αp =α

and extended linearly to the whole potential.

The Jacobian algebra Jac(Q, W ) of the quiver with potential is defined as the
quotient RhhAii/J(W ), where J(W ) is the closure of the two-sided ideal generated
by all cyclic derivatives ∂α W , with α ∈ Q1 .
For every finitely generated module M over the Jacobian algebra, Derksen, Wey-
man and Zelevinsky introduced its F -polynomial in [9] as
X Y
FM = χ(Gre (M )) yiei , (4)
e i∈Q0

where the sum is over all dimension vectors e = (ei )i∈Q0 and χ(Gre (M )) ∈ Z is
the Euler characteristic of the quiver Grassmannian of all submodules N ⊂ M of
dimension vector e.
Furthermore, they introduced the notion of mutations of (decorated) representa-
tions and showed that if µ is a mutation sequence that transforms the zero module
into M then the F -polynomial of M is equal to the F -polynomial of the cluster
variable obtained by the same mutation sequence from the initial seed in the cluster
algebra A(x, y, Q).
8 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

3. Kauffman states
In this section, we recall Kauffman’s realization of the Alexander polynomial as
a state sum.
3.1. Poset of Kauffman states. Consider an oriented link and fix a diagram K
without curls. Denote by n the number of crossings. Then, there are n + 2 regions
and 2n segments. We chose a segment and label it 1 and then, label other segments
following the orientation of the string by 2, 3, . . . , 2n. In this paper, a pair (x, R)
of a crossing point x and a region R such that x is incident to R is called an arrow.
To define a Kauffman state, we chose a segment i = 1, 2, . . . , n and label the
adjacent regions Ri and Ri0 . A Kauffman state is a set of arrows (x, R), called
markers, such that:
• each crossing point is used in exactly one marker;
• each region except for Ri , Ri0 is used in exactly one marker.
The regions Ri , Ri0 are used in no marker.
A state S 0 is obtained from a state S by a counterclockwise transposition at a
segment j if S 0 is obtained from S by switching two markers at the segment j as
in Figure 2.

R1 R1
• •
x • y x • y
R2 R2

(a) State S (b) State S 0

Figure 2. Kauffman counterclockwise transposition from S to S 0 .

More precisely, let x, y be the endpoints of the segment j and let R1 , R2 be the
adjacent regions at j such that, going clockwise around x, we go from R1 to R2
crossing j. Then, S contains the markers (x, R2 ), (y, R1 ), S 0 contains the markers
(x, R1 ), (y, R2 ) and the other markers in S and S 0 are the same.
We define a partial order on the set of all Kauffman states by S1 < S2 if
there is a sequence of counterclockwise transpositions that transforms S1 into S2 .
Kauffman proved that the resulting poset is a lattice whose maximal element is
a state that admits no counterclockwise transposition and is therefore called the
clocked state in [14]. Similarly, the minimal element is called the counterclocked
state in [14]. We will refer to these states as the maximal and the minimal state.
Example 3.1. Let’s use the following labeling for the segments of the figure-eight
knot.
6
3
5 2
8
7
4
1
 KNOT THEORY AND CLUSTER ALGEBRAS 9

•
•
•
•

• •
• •
• •
• •

•
•
•
•

• •

• •

Figure 3. Lattice of Kauffman states of the figure-eight knot

Figure 3 shows the lattice of Kauffman states for the figure-eight knot with
regards to segment 1.

3.2. The State polynomial. Following Kauffman, we define the weight w(x, R)
of an arrow (x, R) as shown in the following two cases.

R1

R2 x R4

R3

In this case, w(x, R1 ) = B, w(x, R2 ) = 1, w(x, R3 ) = W and w(x, R4 ) = 1.

R1

R2 x R4

R3

In this case, w(x, R1 ) = W , w(x, R2 ) = 1, w(x, R3 ) = B and w(x, R4 ) = 1.

10 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

B j 1 W j 1 W j 1 B j 1
1 B 1 W 1 B 1 W
w(j) = B −2 w(j) = W −2 w(j) = W −1 B −1 w(j) = W −1 B −1
w(j) 7→ t w(j) 7→ t−1 w(j) 7→ 1 w(j) 7→ 1

1 j 1 1 j 1 j
1 1 1 j 1
W B B W
B B W W
w(j) = W B −1 w(j) = W −1 B
w(j) = 1 w(j) = 1
w(j) 7→ t w(j) 7→ t−1

B j W W j B
W j W B j B
1 1 1 1
1 1 1 1
w(j) = W B −1 w(j) = W −1 B
w(j) = 1 w(j) = 1
w(j) 7→ t w(j) 7→ t−1

1 j W 1 j B 1 j W 1 j B
W 1 B 1 B 1 W 1
w(j) = W 2 w(j) = B 2 w(j) = W B w(j) = W B
w(j) 7→ t w(j) 7→ t−1 w(j) 7→ 1 w(j) 7→ 1
1
Figure 4. Possible values and specializations at W = t 2 and
1
B = t− 2 for the weight of a segment j

The weight w(S ) of a state S is defined as

Y
w(S ) = w(x, R).
(x,R)∈S

The state polynomial is the sum of the weights of all states S

X
σ(S )w(S ),
S

where σ(S ) = (−1)b with b is the exponent of B in w(S ).

Theorem 3.2 ([14]). The Alexander-Conway polynomial of a diagram L is equal
1 1
to the specialization of the state polynomial at W = t 2 , B = t− 2 .
If a state S 0 is obtained from a state S by a counterclockwise transposition at
a segment at a segment j, then we denote the weight ratio between S 0 and S by
w(j). Thus,
w(S 0 )
w(j) = .
w(S )
Note that w(j) depends only on the segment j and not on the state S and S 0 .
1 1
The possible values for w(j) and its specialization at W = t 2 , B = t− 2 are shown
in Figure 4.
 KNOT THEORY AND CLUSTER ALGEBRAS 11

/6
B3 e γ3 O
γ2
α3 γ4 δ4
 
8O /5o γ1
2O
β1
α4 β4 δ3
 β2
%
4o
β3
@7

α2 δ1
α1 δ2


1p

Figure 5. The quiver of the figure-eight knot of Example 3.1

4. The Jacobian algebra of a link diagram

Let K be a reduced diagram of an oriented prime link without curls. Denote by
n the number of crossings and label the segments 1, 2, . . . , 2n as in section 3.1. We
shall use the notation K0 for the set of crossing points, K1 for the set of segments,
and K2 for the set of regions (including the unbounded region) of R2 \ K.
We are going to construct a quiver with potential and consider its Jacobian
algebra.
We define the quiver Q as follows. The set of vertices Q0 is the set of segments
of K. Thus Q0 = K1 . The set of arrows Q1 is the set of arrows of K introduced in
section 3.1, more precisely, there is an arrow i → j in Q if and only if
• the segments i and j of Q meet at a crossing point p;
• the segments i and j bound the same region R;
• going clockwise around p, we cross i then R then j.
For example the quiver of the figure-eight knot in Example 3.1 is shown in
Figure 5.
The planar link diagram K induces a planar embedding of Q, and since K has
no curls, Q has no loops. On the other hand, Q may have 2-cycles, see however
section 4.1.
The quiver Q has the following two types of chordless cycles. For each crossing
point p ∈ K0 , we obtain a 4-cycle ωP and for each region R bounded by r segments,
we obtain an r-cycle ωR . Each arrow (p, R) lies in exactly two chordless cycles ωp
and ωR .
We define a potential W as
X X
W = ωp − ωR .
p∈K0 R∈K2

In the example in Figure 5, the potential is

W = α1 α2 α3 α4 + β1 β2 β3 β4 + γ1 γ2 γ3 γ4 + δ1 δ2 δ3 δ4
−α1 δ2 β3 − α2 γ3 δ1 − α3 β1 γ2 − β2 δ3 γ1 − α4 β4 − γ4 δ4
12 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

where the first row consists of the four 4-cycles of the four crossing points in K and
the second row consists of the four 3-cycles and two 2-cycles of the six regions in
K.

Definition 4.1. The algebra B = Jac(Q, W ) is called the (completed) Jacobian

algebra of the link diagram K.

Remark 4.2. (a) The quiver Q and the algebra B are not invariants of the link. For
example, the second Reidemeister move will change the number of vertices in Q.
(b) The quiver does not see the difference between an overpass and an underpass
in K.

4.1. Removal of 2-cycles. Each bigon in the link diagram gives rise to a 2-cycle
in the quiver. We can replace the quiver with potential (Q, W ) by a quiver with
potential (Q0 , W 0 ) without 2-cycles as follows. The quiver Q0 is obtained from Q by
removing all 2-cycles. The potential W 0 is obtained from W as follows. For every
bigon R, given by two segments i, j that cross each other in two crossing points
p1 , p2 , we replace ωp1 + ωp2 − ωR by the 6-cycle (∂(p1 ,R) ωp1 )(∂(p2 ,R) ωp2 ) obtained
by joining the two 4-cycles ωp1 and ωp2 . This identification on all 2-cycles induces
an isomorphism of algebras

B = Jac(Q, W ) ∼
= Jac(Q0 , W 0 ).

This realization of the algebra B by a quiver without loops and 2-cycles will be
important when we describe the connection to cluster algebras.
In our running example, there are two bigons formed by the pairs of segments
(4, 8) and (2, 6) in Example 3.1 and these give rise to two 2-cycles α4 β4 and γ4 δ4
in the quiver in Figure 5. The above reduction produces the potential

W0 = α1 α2 α3 β1 β2 β3 + γ1 γ2 γ3 δ1 δ2 δ3
−α1 δ2 β3 − α2 γ3 δ1 − α3 β1 γ2 − β2 δ3 γ1 .

Note that ∂α4 W = −β4 + α1 α2 α3 and thus, in the Jacobian algebra Jac(Q, W ),
the arrow β4 is equal to the path of length three α1 α2 α3 . Similarly α4 is equal to
β1 β2 β3 .

5. The link diagram module T

Let K be a curl free diagram of a prime link and let B = Jac(Q, W ) be its
Jacobian algebra. In this section, we associate a B-module T = ⊕i∈K1 T (i) to
K, where each T (i) is an indecomposable summand. The T (i) are constructed
explicitly as representations of the quiver Q.

5.1. A partition of K1 . Let i be a fixed segment of K. We shall define a partition

of the set of all segments K1 = td≥0 K(d) and use it later to define the representa-
tion T (i). The sets K(d) depend on the choice of the segment i, but, in the interest
of simplicity, our notation does not reflect this dependency. This should not create
confusion, since i is fixed here. The construction is recursive and the case d = 0 is
slightly different from the cases d > 0. But first let us run through the construction
in the following example.
 KNOT THEORY AND CLUSTER ALGEBRAS 13

5.1.1. An example. Consider the knot diagram K illustrated in the top picture in
Figure 6. This is the knot 1066 in the Rolfsen table. We choose the segment i = 1.
The set K(0) is the set of all edges that share a region with the segment i = 1,
including i itself. This set is shown in red in the second picture in the figure. Thus
we have K(0) = {1, 15, 11, 5, 13, 14}. We think of this set as a union of two paths
both starting and ending with the segment 1. The first path wL,0 starts on segment
1 in the direction given by the orientation of the knot and turns left at each crossing
point until it comes back to 1. Thus wL,0 = 1, 15, 11, 5, 13, 1. The second path wR,0
also starts on segment 1 in the same direction, but it turns right at each crossing
point. Thus wR,0 = 1, 14, 1.
The set K(1) is constructed in two steps. First, we remove the set K(0) from
K, and then we define the set K 0 (1) as the set of all segments that are incident to
the unbounded region of K \ K(0). This set is shown in red in the third picture
in Figure 6. Note that there are precisely two crossing points p1 and q1 that are
incident to exactly one segment of K 0 (1). Again, we can think of this set as the
union of two paths, but this time they start at p1 and end at q1 . The first path wL,1
makes a left turn at every crossing point. Thus wL,1 = 2, 19, 10, 16, 7, 4, 12, 6, 3, 20.
The second path wR,1 makes a right turn at every crossing point. Thus wR,1 = 2, 20.
In our example there are two crossing points x1 and x2 that are of degree 4 in
K 0 (1). In this situation, the set K(1) is strictly larger than K 0 (1). It is shown in
the last picture in the figure and is defined as follows. The path wL,1 goes through
each of the points x1 , x2 exactly twice. Let D(xi ) denote the domain in the plane
bounded by the subpath of wL,1 from xi to xi . Thus D(x1 ) is bounded by the
path 19, 10, 16, 7, 4, 12, 6, 3 and D(x2 ) is bounded by the path 4, 12, 6. The domain
D(x2 ) actually consist of a single region of the diagram. On the other hand, the
domain D(x1 ) contains 5 regions of K. We let R(xi ) be the unique region of K
inside D(xi ) that is incident to xi . Then R(x2 ) = D(x2 ) and R(x1 ) is the region
bounded by the segments 19, 9, 17, 7, 3.
Then K(1) is defined as the union of K 0 (1) with the segments of the regions
R(xi ). Thus we need to add the segments 9 and 17 to our set. We are now done
with the case d = 1.
The set K(2) is again defined in two steps, but the second step will be trivial.
First let K 0 (2) be the set of all segments that are incident to the unbounded region
of K \ (K(0) ∪ K(1)). Thus K 0 (2) = {18, 8}. There are no crossing points of degree
4 in this set, and therefore we have K(2) = K 0 (2).

5.1.2. The general case for d = 0. For a general link, define

K 0 (0) = K(0) = {j ∈ K1 | j and i bound the same region of K} ∪ {i},
and let K 0 (0) be the closure of K 0 (0); here closure means that the set also contains
the endpoints of the segments.
We can describe K(0) as the union of two paths given by the boundaries of the
two regions incident to i. We describe these paths below in a way that will generalize
to an iteration of this procedure to d > 0. The set K(0) can be described as the
union of the segments along two paths
wL,0 : p0 wp,0
p00 0
q00 wq,0
q0
wL,0

wR,d : p0 wp,0
p00 0
q00 wq,0
q0
wR,0
14 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

15 11
10 16
18 8
K
19 9 17 7
2 4
3
1 14 12 5
20 6

13

15 11
10 16
18 8
K(0)
19 9 17 7
2 4
3
1 14 12 5
20 6

13

10 16
18 8
0
K (1)
19 9 17 7
2 R(x1 ) 4
p1
x1 3 x2 R(x2 )
12
q1 20 6

10 16
18 8
K(1)
19 9 17 7
p1 2 4
x1 3 x2
12
q1 20 6

Figure 6. An example of the construction of the partition of the

segments of the knot into disjoint subsets K(d). The quiver of this
diagram is shown in Figure 11.
 KNOT THEORY AND CLUSTER ALGEBRAS 15

γ γ γ

p p
p j

i i i

Figure 7. Proof of Lemma 5.1. At the point p, there is an even

number of edges on either side of the curve γ in the two pictures
on the left, and an odd number in the picture on the right.

as follows. Let p0 and q0 be the endpoints of the segment i. Since K has no curls,
we have p0 6= q0 . Define p00 = q0 and q00 = p0 and let wp,0 = wq,0 be the segment
i. At p00 and at every subsequent crossing point, the path wL,0 turns left, and
therefore wL,0 is the boundary of the region to the left of the segment i from p0
to q0 . Similarly wR,0 is the boundary of the region to the right of the segment i.
These are exactly the segments in K 0 (0).
Note that the two points p0 , q0 are of degree 3 in K(0) and all other crossing
points in K have degree 0 or 2 in K(0).
0 0
Lemma 5.1. The two subpaths wL,0 and wR,0 do not share a crossing point besides
0 0
p0 and q0 .
Proof. Let R1 , R2 ∈ K2 denote the two regions at i. Suppose there exists p ∈ K0
such that p is not an endpoint of i and p ∈ R1 ∩ R2 . Then we can draw a closed
curve γ from p to p that runs through R1 , R2 , crosses i once, and does not cross any
other segment of K. We consider two cases, depending on the local configuration
of the four segments at p relative to γ, see Figure 7.
Suppose first that there is an even number of these four segments on either side
of γ. This case is illustrated in the two pictures on the left of the figure. Note that
at either endpoint of the segment i, there are three loose segments, so that in total
there is an odd number of loose segments on either side of γ, and (using the Jordan
curve theorem) it is thus impossible to pair them up without crossing γ in order to
form a link.
Therefore, out of the four segments at p, there must be an odd number on either
side of γ. This case is illustrated in the right picture of the figure. Then, on one
side, there is only one segment; call it j. Moving γ slightly away from p toward
the segment j, we obtain a closed curve γ 0 that crosses only two segments i and
j. This shows that K is the connected sum of two links and thus not prime, a
contradiction. 

5.1.3. The general case for d ≥ 1. We define K(d) recursively.

Definition 5.2. Assume K(e) is defined for all e < d. Let K 0 (d) be the set of all
segments j in K1 \ (∪e<d K(e)) for which there exists a segment k ∈ K(d − 1) such
that j and k bound the same region of K. Let K 0 (d) be the closure of K 0 (d).
Definition 5.3. (a) A crossing point p ∈ K0 is called an external point in K 0 (d)
if exactly one of its incident segments lies in K 0 (d). A segment j ∈ K1 is called
16 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

x4

x3
D(x1 )

p0d (x2 ) qd0 (x2 )

R(x2 )
x2

j1
p0d (x1 ) qd0 (x1 )

R(x1 )

x1
pd qd

Figure 8. The domain D(x1 ) and the region R(x1 ) associated to

an interior point x1 ∈ K(d). The domain D(x1 ) is bounded by the
red curve and the region R(x1 ) is shaded. It is bounded above by
the black segments, one of which is labeled j1 . The domain D(x2 )
for the second interior point x2 is only the part of D(x1 ) that lies
above the point x2 .

external in K 0 (d) if j ∈ K 0 (d) and exactly one the endpoints of j is an external

point in K 0 (d).
(b) A crossing point p is called an internal point in K 0 (d) if all four segments at
p lie in K 0 (d). Note that for d = 0 there are no internal crossing points since our
diagram K has no curls. If there exists a non-constant path w starting and ending
at x that uses only segments of K 0 (d), we let D(x) be the bounded domain enclosed
by w in the plane. Then D(x) is a union of regions of K 0 (d). Let R(x) ∈ K2 be
the unique region in D(x) that contains x.

An example is shown in Figure 8. In that figure, there are four interior points
x1 , . . . , x4 . The domain D(x1 ) is bounded by the red subcurve w and the region
R(x1 ) is shaded.
Now let j be a segment of K. We define d (j) ∈ {0, 1} for d ≥ 1 as follows.

if there exists an internal point x in K 0 (d) such that j lies



 1
in the interior of D(x) and the region R(x) ∈ K2 contains the

d (j) =

 segment j and the point x;
0 otherwise.

(5)
 KNOT THEORY AND CLUSTER ALGEBRAS 17

Definition 5.4. For d ≥ 1, let

K(d) = K 0 (d) ∪ {j ∈ K1 | d (j) = 1}.
For example, the segment j1 in Figure 8 satisfies the first condition for x = x1 .
Thus d (j1 ) = 1 in this case. The set K(d) contains every segment of the red curve
and every segment of the black curves bounding R(x1 ) and R(x2 ).
Lemma 5.5. Let d ≥ 1.
(a) Each connected component C of K 0 (d) is either a single path w from pd to
qd or the union of two paths

wL,d : pd wp,d
p0d 0
qd0 wq,d
qd
wL,d

wR,d : pd wp,d
p0d 0
qd0 wq,d
qd
wR,d

from pd to qd , where pd and qd are external points in K 0 (d), the initial and terminal
subpaths wp,d and wq,d are the same in both paths, deg p0d = deg qd0 = 3 in C, p0d 6= qd0
0 0
and wL,d (respectively wR,d ) is obtained by turning left (respectively right) at every
0
crossing point in K (d).
If no such points p0d , qd0 exist then wL,d = wR,d and K 0 (d) is a single path from
pd to qd .
(b) All other crossing points, besides pd , p0d , qd , qd0 , have degree 0, 2 or 4 in C.
In particular, C has exactly two external points pd and qd , and moreover, pd , qd ∈
K 0 (d) ∩ K 0 (d − 1) and pd , qd ∈ / K 0 (e), with e 6= d, d − 1.
0 0
(c) The paths wL,d and wR,d do not share a crossing point besides p0d and qd0 .

Proof. (a) Suppose first d = 1. The external points are p1 = p00 and q1 = q00 , and
there are no other external points in K 0 (1). If there are points of degree 3 in K 0 (1),
we let p01 be the point of degree 3 closest to p1 , and let q10 be the point of degree 3
closest to q1 . Note that every connected component has an even number of points
of degree 3, because of parity, and that there are at most two because there are
only two external points.
Let wp,1 be the unique path from p1 to p01 in K 0 (1) that
• does not use an edge twice;
• is of maximal length; (6)
• turns left or right at every point of degree 4 in K 0 (d).
Such a path exists and is unique by the following argument. The starting point p1
is of degree 1 in C, so the first step is uniquely determined. At every point of degree
2, the incoming edge leaves only one choice for the outgoing edge. At a point of
degree 4, there are a priori two possibilities, turn left or turn right, but only one
of these will produce a path of maximal length. Similarly, Let wq,1 be the unique
path from q10 to q1 in K 0 (1) that respects conditions 6.
Then the paths wL,1 , wR,1 form the boundary of the regions on the left and right
of the path (wq,1 i wp,1 ) in (K \ K(0)) ∪ {i}. These are exactly the segments in
K 0 (1). This completes the proof of (a) for d = 1.
(b) The degree formulas follow directly from (a). In particular p1 and q1 are the
only external points in C. Furthermore, three of the segments at p1 lie in K(0),
and the remaining segment, which is the first segment of wp,1 , lies in K 0 (1). Thus
18 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

0 0
wL,d wL,d

` D
pd p0d x qd0 qd
0
wR,d y `0 0
wR,d

Figure 9. Proof of Lemma 5.5 part (c).

p0 ∈ K 0 (1) ∩ K 0 (0) and p0 ∈/ K 0 (e), with e 6= 0, 1. The proof for q1 is similar. This
also implies that all other points have degree 0, 2 or 4 in K 0 (1).
0 0
(c) Suppose wL,1 and wR,1 share a point x 6= p01 , q10 , see Figure 9. Let D denote
the domain in the plane bounded by the segments of the paths wL,1 and wR,1 from
p01 to x. Let ` denote the segment at p01 that does not belong to wL,1 or wR,1 . Since
wL,1 turns left at p01 and wR,1 turns right, the segment ` lies in D. Following the
link, starting at p01 in direction of `, we must reach a point y where we leave D. Let
`0 denote the segment outside D right after y. In K, this segment `0 bounds the
0
same region as a segment on the path wR,1 , and therefore `0 must lie in K(0)∪K(1).
0
However, ` cannot lie in K(0), because K(0) has no external segments. On the
other hand, `0 cannot be in K 0 (1), because (a) implies that every segment of K 0 (1)
lies on one of the two paths wL,1 , wR,1 . This is a contradiction, and thus the two
0 0
paths wL,1 , wR,1 cannot have the point x in common. This completes the proof for
d = 1.
For d > 1, the proof is similar, with the additional feature that now we may have
components that arise from internal points in K 0 (d − 1). Indeed, every internal
point x of K 0 (d − 1), such that there is a segment j bounding the region R(x) with
d−1 (j) = 1, determines two crossing points p0d (x) and qd0 (x) as the unique points
in R(x) \ {x} that lie on the boundary of D(x), and such that the path w = wL,d−1
or wR,d−1 is of the form
w : pd x p0d (x) qd0 (x) x qd ,
see Figure 8. Therefore every internal point x of K 0 (d − 1) gives rise to a connected
component of K(d) in which the points pd = p0d−1 (x) and qd = qd−1 0
(x) are the
0 0
unique external points. For all other components, we have pd = pd−1 and qd = qd−1 .
Note that pd and qd have degree 3 in K(d − 1) and thus degree 1 in K 0 (d). Hence
they are external points.
The rest of the proof of (a) is analogous to the case d = 1. The point p0d is the
point of degree 3 closest to pd , and qd0 is the point of degree 3 closest to qd . The
paths wp,d and wq,d are the unique paths from pd to p0d and from qd to qd0 that
satisfy the conditions (6). Moreover the paths wL,d , wR,d form the boundaries of
the regions to the left and right of the path
wq,d . . . wq,2 wq,1 i wp,1 wp,2 . . . wp,d
in (K \ ∪d−1
e=0 K(e)) ∪ {wq,d−1 , . . . wq,1 , i, wp,1 , . . . , wp,d−1 }.
Thus wL,d and wR,d con-
0
sist of the edges of K (d).
The proof of (b) is analogous to the case d = 1 except that the point pd may be
equal to a point p0d−1 (x) for some interior point x in K 0 (d).
 KNOT THEORY AND CLUSTER ALGEBRAS 19

In the proof of (c), the only difference to the proof in case d = 1 is that now
the segment `0 in Figure 9 cannot lie in K(d − 1), because otherwise y would be an
external point of K(d − 1) that lies in K(d), a contradiction to (b). 
Remark 5.6. Each interior point x of K 0 (d − 1) whose region R(x) contains a
segment j with d−1 (j) = 1 gives rise to a connected component of K 0 (d).
Proposition 5.7. For every segment i ∈ K1 , we have constructed a partition
K1 = td≥0 K(d).
Proof. This follows directly from the construction. 
We are now ready to define the dimension vector of the representation T (i).
Definition 5.8. Let K1 = td≥0 K(d) be the partition with respect to a segment
i ∈ K1 . For every segment j ∈ K1 , we define d(i)j = d if j ∈ K(d).
In the example of Figure 6, we have d(1)j = 1 if j = 2, 3, 4, 6, 7, 9, 10, 12, 16, 17, 19, 20;
d(1)j = 2 for j = 8, 18, and d(i)j = 0 for all other j.
Our next result says that the dimension difference at adjacent vertices is at most
one.
Proposition 5.9. Let i, j, k ∈ K1 . If there is an arrow j → k ∈ Q then
|d(i)j − d(i)k | ≤ 1.
Proof. Let d = d(i)j . Thus j ∈ K(d). The existence of the arrow j → k implies that
j and k bound the same region in K. If k ∈ / ∪e≤d K(e) then Definition 5.2 implies
that k ∈ K 0 (d+1), thus k ∈ K(d+1) and d(i)j −d(i)k = −1. If k ∈ K(d)∪K(d−1)
then d(i)k ∈ {d, d − 1} and there is nothing to show. Finally suppose k ∈ K(e),
with e ≤ d − 2. Then Definition 5.2 implies that j lies in K 0 (e + 1) unless it already
lies in ∪e0 <e K(e0 ). In both cases, we have d(i)j ≤ e + 1 ≤ d − 1, a contradiction. 
5.2. Properties of K(d). It will be convenient to use the following terminology.
Given two segments i, j ∈ K1 , a curve in R2 is called a dimension curve from j to
i if it starts at a point on segment j, ends at a point on segment i and does not go
through a crossing point of K.
Let dim◦ (i, j) be the minimal number of crossings between the segments of K
and a dimension curve from segment j to segment i. We call a segment j ∈ K1 an
interior segment of K(d) if K(d) contains all the segments on the boundary of the
two regions incident to j in K.
Note that the segment i is an interior segment of K(0). The following lemma
says that there are no other interior segments.
Lemma 5.10. If d ≥ 1 then K(d) has no interior segments.
Proof. Suppose a segment j belongs to two regions R1 , R2 in K and each segment
in R1 ∪ R2 lies in K(d), see Figure 10. Since d ≥ 1, the dimension curve of j must
cross a segment k of R1 ∪ R2 , so it has one more crossing than the dimension curve
of the segment k. Since j, k ∈ K(d) this means that d (j) = 1. Thus there exists an
internal point x in K(d) satisfying the condition (5). In particular, one of the two
regions at j, say R1 , contains x and j. Thus R1 is the region R(x) of condition (5).
The other region of K at j is the region R2 and both lie entirely in K(d). Since
j∈/ K 0 (d), the segments of R1 ∪ R2 that do lie in K 0 (d) all lie in the same region in
20 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

R2
` y j
`00
00
y
`0 y0 R1
h
x

Figure 10. Proof of Lemma 5.10

K 0 (d). In particular, no segment k of R1 ∪ R2 shares a region with another internal

point x0 6= x such that k lies in D(x0 ). Therefore each segment k of R2 \{j} satisfies
d (k) = 0. Thus k ∈ K 0 (d), which implies k ∈ D(x).
Now consider an endpoint y of j. Three of the segments incident to y lie in
R1 ∪ R2 and the fourth segment ` doesn’t, see Figure 10. Therefore ` either lies on
the boundary of D(x) and hence dim◦ (i, `) = d, or ` lies outside D(x) and hence
dim◦ (i, `) = d − 1. The latter case is impossible by Lemma 5.5(b), because ` would
be an external segment of K 0 (d − 1), but ` has an endpoint in K 0 (d) (and not in
K 0 (d − 2).
Thus the segment ` lies on the boundary of D(x). Since ` is not in R1 , there
exists a point y 0 ∈ K0 such that two of its incident segments lie in R1 and one, call
it h, lies on the boundary of D(x) between x and y as in the figure. Denote by `0
the fourth segment at y 0 . Note that it must lie inside D(x), because otherwise y 0
would be an external point of K 0 (d − 1) that does not lie in K 0 (d − 2), contradicting
Lemma 5.5(b). Since R1 is a region in K, the segment ` must lie in the connected
component C of D(x) \ R1 that contains y 0 . Following the link starting at y 0 in
direction `0 , we must reach a point y 00 , where we leave the component C. Let `00 be
the segment outside C right after y 00 . Then `00 is an external segment of K 0 (d − 1)
with external point y 00 not in K(d − 2), again a contradiction to Lemma 5.5(b). 
It will be convenient to consider the following dual graph. For an illustration,
see Example 5.16
Definition 5.11. (a) Let Q(d) be the full subquiver of Q on the vertices j such
that j ∈ K(d).
(b) Let G(d) be the graph with vertex set the set of chordless cycles in Q that
also lie in Q(d), and two chordless cycles are connected in G(d) by an edge if they
share an arrow in Q(d).
Notice that the graph G(d) has two types of vertices, the crossing point vertices
and the region vertices. The first type corresponds to the chordless 4-cycles ωp ,
with p ∈ K0 , and the second type corresponds to the chordless cycles ωR , with
R ∈ K2 .
Corollary 5.12. In G(d), the degree of a crossing point vertex is at most 2.
Proof. If a crossing point vertex x has three adjacent regions in G(d), hence in
K(d), then K(d) has an interior segment, contradicting Lemma 5.10. 
 KNOT THEORY AND CLUSTER ALGEBRAS 21

Definition 5.13. Let R be a region of K such that each segment of R lies in K(d).
By Lemma 5.5, one of the two paths w = wL,d or wR,d encloses the region R. Let
x(R) be the first crossing point of the region R on the path w. We call x(R) the
root of the region R.
Recall that a leaf in a graph is a vertex of degree one.
Lemma 5.14. (a) The mapping R 7→ x(R) is a bijection between the sets of region
vertices of G(d) and crossing point vertices of G(d).
(b) Every connected component of G(d) has a unique vertex that is a leaf and a
crossing point vertex.
Proof. (a) Since the path w starts and ends at vertices outside R it must go through
the point x(R) twice, in the sense that it contains all four segments at x(R). Thus
x(R) is an internal point of K 0 (d) and therefore a vertex of G(d). This shows that
the mapping is well-defined.
The mapping is injective by definition. Now let x be any crossing point vertex
in G(d). Then all four segments at x lie in K(d). This implies that x is an internal
point of K 0 (d), because the endpoints of the segments j with d (j) have at most
degree 3 in K(d). Lemma 5.5 then implies that the four segments at x all lie on
one path w = wL,d or w = wR,d , and thus w goes through x twice. Then w is of
the form
w : pd x x qd
w(x)

and the subpath w(x) forms the boundary of the domain D(x) in condition (5).
There is a unique region R(x) of K that lies within D(x) and contains x. By
definition of K(d), all segments of R lie in K(d). Thus R is a region vertex of G(d)
and x(R) = x. This shows that the mapping is surjective.
(b) Let C be a connected component of G(d). By Lemma 5.5, there is a path
w = wL,d or w = wR,d that encloses all regions of the region vertices of C. Let x
be the first point on w that is a crossing point vertex of C. Then x is a leaf of C.
To show that there is no other crossing point that is a leaf, note first that every
crossing point vertex of G(d) is of degree at most 2 in G(d), by Corollary 5.12.
Now we proceed by induction on the number of region vertices. If there is only one
region vertex R in C then C = R x(R) and we are done. Suppose there
is more than one region vertex. Take a leaf `. If ` = R is a region vertex, then
C \ {`} has x(R) as a leaf and thus C \ {`, x(R)} is connected, and by induction it
has no other crossing point vertex that is a leaf than x. On the other hand, if C
contains no leaf that is a region vertex, then there are more crossing point vertices
than region vertices in C, which is impossible by part (a). 

Remark 5.15. We don’t know if G(d) is a forest.

5.3. Definition of the link diagram module T . Let k be an algebraically closed
field. Let K be an oriented diagram without curls of a prime link with n crossings.
Let (Q, W ) be the associated quiver with potential and B = Jac(Q, W ) its Jacobian
algebra. Let Id denote the identity matrix of rank d. We define the link diagram
module
T = ⊕i∈K1 T (i)
of K as follows.
22 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

For each segment i of K, the direct summand T (i) = (T (i)j , T (i)α )j∈Q0 ,α∈Q1 is
the representation of Q given by
T (i)j = kd(i)j ,
for each vertex j, where d(i)j is the dimension defined in Definition 5.8; and for
each arrow α : j → `, we define the corresponding linear map
  
 0
 .. Id−1 


if d(i)j = d(i)` + 1 = d



  . 

0




T (i)α =  



  I

d−1 
 

 
   if d(i)j + 1 = d(i)` = d


0 ··· 0


and if d(i)j = d(i)` = d then

  
 0
 .. Id−1




  .

 if α is of the form (x, R(x)) with
x an internal point of K 0 (d)

  
 0
T (i)α =



 0 0 ··· 0



Id


otherwise.
Because of Proposition 5.9, there are no other possibilities for the dimensions and
thus T (i)α is well-defined.
Example 5.16. The quiver Q of the knot 1066 in Figure 6 is shown in the top
picture and the representation T (1) in the bottom picture of Figure 11. The quiver
Q(1) is the full subquiver on the vertices 2,3,4,6,7,9,10,12,16,17,19,20. It contains
four chordless cycles, the two crossing point cycles wx1 , wx2 the two region cycles
wR(x1 ) , wR(x2 ) , where we use the notation of Figure 6. Therefore its dual graph is
wx1 wR(x1 ) wx2 wR(x2 ) .

6. Kauffman states and submodules of the link diagram module

We keep the notation of the previous sections. Again we choose a segment i ∈ K1
and consider the Kauffman states and the B-module T (i). Our goal is now to prove
that the lattice of Kauffman states of a link K relative to a segment i is isomorphic
to the lattice of submodules of the direct summand T (i) of the corresponding link
diagram module T .

6.1. The state module M (S ). Let S be a Kauffman state. We will define a

B-module M (S ) = (Mj , Mα )j∈Q0 ,α∈Q1 . Consider a sequence s of counterclockwise
transpositions that transforms the minimal Kauffman state into the state S . Then
we define
Mj = kej ,
 KNOT THEORY AND CLUSTER ALGEBRAS 23

15 / 11
H a O
}
10 o
O = 16
!
18O /8
O
  
o 9o 17 o 67
= 19
 (3 !
2O e 4a

} 
1 oe / 14 = 12
o /5

! y 
20 ;6

# y
13

kO o
1
[ 10 ] ?k
[0 1]

k2O
1 / k2
1 O 1

[0 1] [ 10 ] [0 1] [ 10 ]
  
o 1
ko
1 o 1 7k
@k k
1 1
0
'
1 
kO kd k^
1

1
1 1
0 @k

z  1
k k

Figure 11. The quiver Q and the representation T (1) for the knot
diagram of Figure 6.
24 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

where ej is the number of occurrences of j in s. The order in which the transpo-

sitions at j occur determines a basis for Mj , which we call the basis induced by
s.
Next we define the linear maps of the representation M . In the remainder of
this section, we use the following matrices
 
0    
0
 .. I`  
 . I`  I` 
J` =  .

 V` =  ..  H` = 
 
 0   
0 0 ··· 0
0 0 ··· 0
where I` denotes the identity matrix of size `. We point out that J` is a Jordan
block of size ` + 1 with eigenvalue 0, and that H` V` = J` and V` H` = J`−1 .
Every crossing point p ∈ K0 indues a subsequence s(p) of s consisting of all
occurrences of the transpositions at the four segments incident to p. Let a, b, c, d
denote these four segments in counterclockwise order around p such that a is the
first entry in s(p). Then s(p) is of one of the following forms
(abcd)` , (abcd)` a, (abcd)` ab, (abcd)` abc, (7)
for some ` ≥ 0. Let
aO
δ /d
α γ

bo c
β

be the corresponding 4-cycle in the quiver Q, and let wp = δγβα ∈ B.

Since every arrow of Q lies in a unique 4-cycle induced by a crossing point, it
suffices to define the linear maps of the representation M on these four arrows
α, β, γ, δ. There are four cases depending on the sequence s(p).
(i) If s(p) = (abcd)` then ea = eb = ec = ed = ` and
Mδ = J`−1 Mγ = Mβ = Mα = I` .
(ii) If s(p) = (abcd)` a then ea = ` + 1, eb = ec = ed = ` and
Mδ = V` Mγ = Mβ = I` , Mα = H`
`
(iii) If s(p) = (abcd) ab then ea = eb = ` + 1, ec = ed = ` and
Mδ = V ` Mγ = I` , Mβ = H` Mα = I`+1 .
`
(iv) If s(p) = (abcd) abc then ea = eb = ec = ` + 1, ed = ` and
Mδ = V` M γ = H` Mβ = Mα = I`+1 .
Definition 6.1. The B-module M (S ) is called the state module associated to the
Kauffman state S .
Remark 6.2. In all four cases (i)-(iv) above the composition of the four matrices
along the cycle wp is equal to J`−1 . Thus the action of wp on M (S ) is given by
this matrix.
From the construction of the state module, we have the following results.
Lemma 6.3. Let M (S ) = (Mx , Mα )x∈Q0 ,α∈Q1 . Then for every arrow α : j → k,
we have | dim Mj − dim Mk | ≤ 1.
 KNOT THEORY AND CLUSTER ALGEBRAS 25

Lemma 6.4. If the state S 0 is obtained from the state S by applying the trans-
position at a segment a then the module M (S 0 ) is obtained from M (S ) by
(i) increasing the dimension at vertex a by one;
(ii) increasing the rank of the map on each arrow α : a → • starting at a by
one.
The dimension at the other vertices and the rank on the other arrows do not change.
We also note the following for future reference.
Lemma 6.5. Let S be a state and s a sequence of transpositions that transforms
the minimal state into S . If
αt−1
w = a0
α1
/ a1 α2
/ ... / at αt
/ a0
is a chordless cycle in Q then the subsequence of s of all occurrences of transpositions
at vertices of w is of the form
aj . . . a2 a1 a0 (at at−1 . . . a1 a0 )` at at−1 . . . ak
for some j, k and `. In particular, the order in s is opposite to the order in w.
Proof. We have already proved this result in equation (7) in the case where w = wp
is the chordless 4-cycle given by a crossing point p ∈ K0 . It suffices to show the
result in the case where w = wR is the chordless cycle of a region R ∈ K2 . The
transposition at ai is defined by moving the markers counterclockwise around the
endpoints of ai , but it can also be seen as moving the markers in the clockwise
direction along the segment ai , see Figure 2. Thus the proof for the region cycle
wR is dual to the proof for the crossing point cycle wp . 
As an immediate consequence we have the following.
Corollary 6.6. If α : a → d is an arrow in Q such that the transpositions at a and
d occur consecutively in s then d comes before a. 
6.2. Lattice isomorphism. We start with the following result.
Proposition 6.7. Let K be a link diagram without curls and let i ∈ K1 be a
segment. Let S , S 0 be two Kauffman states relative to i. Then
(a) M (S ) is a B-module.
(b) M (S ) ∼6 M (S 0 ) if S 6= S 0 .
=
(c) If S is the minimal Kauffman state then M (S ) = 0.
(d) If S is the maximal Kauffman state than M (S ) = T (i).
(e) If S < S 0 then M (S ) is a submodule of M (S 0 ).
(f) For every submodule M of T (i) there is a unique Kauffman state S such
that M ∼= M (S ).
Proof. (a) By definition, M (S ) is a representation of Q, so we only need to check
that M (S ) satisfies the relations given by the cyclic derivatives of the potential
W.
Let (p, R) be an arrow in Q, thus p ∈ K0 and R ∈ K2 such that the region R is
incident to the crossing point p. By definition of the potential, we have
∂(p,R) W = wp − wR ,
where wp = (p, R)wp0
is the 4-cycle in Q given by the four arrows around the crossing
0
point p and wR = (p, R)wR is the cycle in Q given by the arrows around the region
26 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

·O / ak+1 o ·
p ak+1
R (p,R)
ak 
·o ak /·

Figure 12. A crossing point p with an adjacent region R in the

link diagram on the left and the corresponding paths in the quiver
on the right. The cycle wp is the 4-cycle

R, see Figure 12. We must show that ∂(p,R) W acts trivially on M (S ), and for that
it suffices to show that the composition of the linear maps in the representation
M (S ) along the paths wp0 and wR 0
are equal.
We will write Mw for the composition of the arrows in M (S ) along a path w.
According to Remark 6.2, we have Mwp = J` for some ` ≥ 0, and thus it suffices to
show MwR = J` .
As before, we let s be the sequence of counterclockwise transpositions that trans-
forms the minimal Kauffman state into the state S . Let s(R) be the subsequence
of s consisting of all occurrences of the transpositions at the segments that bound
the region R. Let a0 , a1 , . . . , at denote these segments in clockwise order around R
such that a0 is the first entry in the sequence s(R). By Lemma 6.5, the subsequence
s(R) is of the form
s(R) = (a0 a1 . . . at )` a0 a1 . . . au , with u < t or
s(R) = (a0 a1 . . . at )`
for some ` ≥ 0. In the first case, the dimension of M (S ) is ` at vertices au+1 , . . . , at
and it is ` + 1 at vertices a0 , a1 , . . . , au , while in the second case, the dimension is
` at all vertices a0 , a1 . . . , at .
Denote the crossing point of the segments aj−1 , aj by pj . So the arrow (pj , R)
is aj → aj−1 . Then by definition of M (S ) at the crossing point pj , we have
M(pj ,R) = J`−1 if j = 0 and s(R) = (a0 , . . . , at )`
and otherwise

 Id if dim Maj = dim Maj−1 = d;
M(pj ,R) = Vd if dim Maj − 1 = dim Maj−1 = d;
Hd if dim Maj = dim Maj−1 − 1 = d.


In particular MwR = J` . This shows that M (S ) satisfies all relations of the

form ∂α W , α ∈ Q1 .
We also have to consider the closure I of the ideal generated by the relations
∂α W . For this we must show that arbitrary long paths act as zero; more precisely,
if w is a path such that for all N there exists a path un of length n > N such that
w = un then Mw = 0. Suppose w is such a path. Then, for all m there exists n
such that there is an x ∈ Q0 through which un passes at least m times. Thus un
decomposes as
un = uw1 w2 . . . wm v,
where each wi is an oriented cycle that starts and ends at x and that does not pass
through x another time. Take m > dim Mx = d. We shall show below that, on
 KNOT THEORY AND CLUSTER ALGEBRAS 27

each cycle wi , the matrix product Mwi is some power of the matrix Jd−1 . Therefore
Mw1 w2 ...wn = (Jd−1 )m+k , which is zero. Thus Mun = 0. Since un = w, we have
Mw = 0, as desired.
It remains to show that, for every oriented cycle
αt−1
w = a0
α1
/ a1 α2
/ ... / at αt
/ a0

in Q such that ai 6= aj if i 6= j, the matrix Mw is a power of the matrix J` , for some

`. Let w be such a cycle. By definition of M (S ), for every arrow αj , the matrix
Mαj is one of the four matrices I, V, H, J, which satisfy the relations H` V` = J`
and V` H` = J`−1 . Hence if Mw is not of the claimed form then Mw = I` and
all Mαj = I` . Then dim Maj = ` at each vertex aj in w, which means that the
transposition at aj appears exactly ` times in the sequence s. Let ak be the last
transposition in the sequence s at a vertex in w. Consider the crossing point p
where ak−1 and ak meet in the link diagram. By Lemma 6.4, the last transposition
at ak does not increase the rank of the matrix Mαk , since the arrow αk ends in ak .
Thus Mαk = J`−1 and hence Mw 6= I` , and we are done. This completes the proof
of part (a) of the proposition.
(b) Let S , S 0 be two Kauffman states and suppose that M (S ) = M (S 0 ). Let s
and s0 be the sequences of transpositions that transform the minimal state into the
the state S and S 0 , respectively. Let p be any crossing point and denote by s(p)
and s0 (p) the subsequences of s and s0 consisting of all occurrences of transpositions
at p. Since M (S ) ∼= M (S 0 ), both representations have the same dimension vector
and thus s(p) and s0 (p) are equal up to a permutation. In fact, since the minimal
state has exactly one marker at the point p, it follows that s(p) = s(p0 ). At
every crossing point p, the states S and S 0 are determined by the last entry in
s(p) = s0 (p), and thus S = S 0 .
(c) If S is the minimal state then its sequence of transpositions s is empty. Thus
M (S ) is the zero module.
(d) Let S be the maximal state. We have described Kauffman’s construction of
S in [14] as a partition of K1 in section 5.1. The fact that M (S ) and T (i) have
the same dimension vector follows directly from that. We now show that M (S )
and T (i) also have the same linear maps.
Let α : a → d be an arrow in Q. It is clear from the definition of M (S ) and
T (i) and by Lemma 6.3 that the linear maps on α are the same if the dimension
at vertex a is different from the dimension at vertex d. Suppose therefore that
dim Ma = dim Md = `. Recall that the arrow α corresponds to a pair (p, R), where
p is a crossing point and R is an adjacent region in K. In the quiver Q, we have
two corresponding chordless cycles wp and wR that share the arrow α as follows.
bO /ao aO 1
α

co d / at−1

The crossing point cycle wp is the cycle of length 4 on the left. The length of
the region cycle wR is the number of segments that bound R in the link diagram.
We denote this length by t + 1. Note that in these two cycles the arrow α is
the only arrow that ends at d. Consider the sequence of transpositions s that
transforms the minimal state into the maximal state S and let s(p) and s(R) be
28 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

the subsequences of all occurrences of transpositions at segments incident to p,

respectively at segments bounding R. Suppose first that the first occurrence of a
is before the first occurrence of d in s(p). Then a must be the first entry in s(p),
because the direct predecessor of a would have to be d, by Corollary 6.6. Similarly,
d must be the last entry in s(p), because Ma and Md have the same dimension `
and the direct successor of d in s(p) would have to be a. Thus s(p) = (abcd)` and
this shows that
(i) Mα = J`−1 , by definition of M (S );
(ii) The dimension of M (S ) is ` at each vertex of wp and thus p is an internal
point of K 0 (`) as in Definition 5.3.
Now consider the sequence s(R). Since a occurs before d in s(p), it also does so
in s(R), and by the same argument as above, we see that a must be the first entry
of s(R) and d must be the last. Thus s(R) = (aa1 a2 . . . at−1 d)` and therefore the
dimension of M (S ) is equal to ` at every vertex of wR . The transposition at a
moves two state markers at the endpoints of a counterclockwise. By our convention
on the orientation of the quiver, the marker at the endpoint p must lie in the region
R. Moreover, the fact that a is the first entry in both sequences s(p) and s(R)
implies that the position (p, R) carries a state marker already in the minimal state.
Similarly, since d is the last entry in s(p) and s(R), the position (p, R) also has a
state marker in the maximal state. It follows from the construction of the minimal
and maximal states in [14] that the region R is the region R(p) of the internal point
p as in Definition 5.3. Now the definition of the maps in the diagram module T (i)
implies that T (i)α = J`−1 . Hence M (S )α = T (i)α .
It remains the case where the first occurrence of d is before the first occurrence
of a in s(p). Then Lemma 6.4 implies that each occurrence of a in s(p) augments
the rank of M (S )α by one. Thus our assumption dim Ma = dim Md = ` implies
that Mα = I` . On the other hand, we also have T (i)α = I` , because the position
(p, R) does not carry the state marker of maximal state, and thus R is not the
region R(p) of the internal point p. This completes the proof of pard (d).
(e) It suffices to show that if the state S 0 is obtained from the state S by
a single transposition at some segment a then M (S ) is a submodule of M (S 0 ).
We use the notation M (S ) = (Mx , Mγ ), M (S 0 ) = (Mx0 , Mγ0 ) and dx = dim Mx ,
d0x = dim Mx0 . Define a morphism f : M (S ) → M (S 0 ) by

Idj if j 6= a;
fj =
Hdj if j = a.
Clearly f is injective. To show that f is a morphism of B-modules, we need
to consider arrows b
α /a β
/ c in Q and show that the following diagram
commutes.
Mβ
Mb
Mα
/ Ma / Mc

fb fa fc
 0
Mα  Mβ0 
Mb0 / Ma0 / Mc0

Since S 0 is obtained from S by the transposition at a, Lemmata 6.3 and 6.4 imply
that da ∈ {db , db −1} = {dc , dc −1}. Moreover d0a = da +1, d0b = db and d0c = dc , and
the maps Mα , Mβ , Mα0 , Mβ0 are uniquely determined by the fact given in Lemma 6.4
 KNOT THEORY AND CLUSTER ALGEBRAS 29

0
b d0 Ab ?d
β α δ0 γ0
p a p 0
 
c c0 c] a_ c0

d b0 γ δ α0  β0
d b0

Figure 13. A local configuration in the link diagram on the left

and the corresponding configuration of the quiver on the right.

that rank(Mα0 ) = rank(Mα ) and rank(Mβ0 ) = rank(Mβ ) + 1. Thus

M α = I` and Mα0 = H` if da = db = `;
M α = V` and Mα0 = J` if da = db − 1 = `;
Mβ = J`−1 and Mβ0 = V` if da = dc = `;
Mβ = H` and Mβ0 = I`+1 if da = dc − 1 = `.

This shows that the diagram commutes and the proof of (e) is complete.
(f) Let Smax denote the maximal Kauffman state. Thus T (i) = M (Smax ) by
part (d). Let da = dim T (i)a . We fix a sequence of transpositions s that transforms
the minimal state into the maximal state, and we use the basis of T (i) induced by
s. In particular, we have a basis {e1 , e2 , . . . , eda } for every vector space T (i)a with
a ∈ Q0 .
Let M = (Ma , Mα ) be a submodule of T (i). Each vector space Ma is a sub-
space of T (i)a , and thus the points in Ma can be expressed as coordinate vec-
tors (x1 , . . . , xda ) with respect to our basis of T (i)a . For every vertex a ∈ Q0 ,
let πa denote the canonical projection from the vector space M = ⊕j∈Q0 Mj to
the vector space Ma . For any point x ∈ M , we define an integer m(a, x) as fol-
lows. If πa (x) = (x1 , . . . , xda ) 6= 0, we let m(a, x) be the unique integer such that
xm(a,x) 6= 0 and xk = 0, for all k = m(a, x) + 1, . . . , da . If πa (x) = 0, we let
m(a, x) = 0. We then define a function m : Q0 → Z by m(a) = maxx∈M m(a, x).
We will show that m(a) = dim Ma .
If m(a) = 0 then πa (x) = 0, for all x ∈ M , and thus M is not supported at a,
whence dim Ma = 0. If m(a) = 1 then πa (x) ∈ span{e1 }, for all x ∈ M , and hence
dim Ma = 1.
Suppose now that m(a) ≥ 2. Then dim T (i)a ≥ 2 and thus the sequence s
contains the transposition at a at least twice. Denote the crossing points at the
ends of the segment a in K by p and p0 , and denote the adjacent segments by b, c, d
and b0 , c0 , d0 in counterclockwise order as shown in the left picture of Figure 13.
The sequence s must also contain the transpositions at b, c, d, b0 , c0 and d0 . The
corresponding subquiver of Q is shown on the right in Figure 13. We have chordless
cycles wp = δγβα and wp0 = δ 0 γ 0 β 0 α0 . Because of Remark 6.2, the action of wp on
Mα is as follows. If πa (x) = (x1 , x2 , . . . , xm(a) , 0, . . . , 0) then

πa (x · wp ) = (x2 , x3 , . . . , xm(a) , 0, . . . , 0)
πa (x · wpk ) = (xk+1 , xk+2 , . . . , xm(a) , 0, . . . , 0) (8)
m(a)−1
πa (x · wp ) = (xm(a) , 0, . . . , 0)
30 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

m(a)−1
Hence, since xm(a) 6= 0, the vectors πa (x), πa (x · wp ), . . . , πa (x · wp ) are lin-
early independent vectors in Ma . Thus Ma = span{e1 , e2 , . . . , em(a) } and hence
dim Ma = m(a) as claimed.
In fact, each subspace Ma ⊂ T (i)a , and hence the submodule M ⊂ T (i), is
completely determined by m(a) = dim Ma . In particular, the submodules of T (i)
are determined by their dimension vector.
We will now show thatPM corresponds to a Kauffman state using induction on
the total dimension ` = a∈Q0 dim Ma of M . If ` = 0 then M = M (Smin ) is the
zero module. Suppose now that ` ≥ 1. Let a ∈ Q0 be such that S(a) is a direct
summand of top M . Recall that top M = M/radM = M/(M · radB) is the largest
semisimple quotient of M . In particular, we have a short exact sequence

0 /L f
/M g
/ S(a) / 0, (9)

where L is the kernel of g. By induction, we can assume that there exists a state
S such that L = M (S ). We will show that M = M (S 0 ), where S 0 is the state
obtained from S by the transposition at the segment a.
Since M is determined by its dimension vector, we only need to show that the
state S admits the transposition at a. In other words, the markers at segment a
must be in the positions indicated on the left of Figure 13 and corresponding to
the arrows δ, δ 0 ∈ Q1 on the right of the same figure. It suffices to show that the
subsequence s(p) of s ends in the transposition at d and the subsequence s(p0 ) ends
in the transposition at d0 . We will show this for s(p) only, since the other case is
symmetric.
By Lemma 6.3, we know that | dim La −dim Ld | ≤ 1 and | dim Ma −dim Md | ≤ 1.
Moreover, the short exact sequence (9) implies dim Ma = dim La + 1, and thus
dim La = dim Ld − 1 or dim La = dim Ld . In the former case, the sequence s(p)
must end in d, and we are done.
Suppose now dim La = dim Ld . The morphism g of (9) gives rise to the commu-
tative diagram

Mb
Mα
/ Ma

ga
 
0 / S(a)a

with ga 6= 0. Thus the commutativity implies that ga Mα = 0, and hence Mα is

not surjective. Because of the description of the action of wp in equation (8), the
cokernel of Mα ◦ Mβ ◦ Mγ ◦ Mδ is of dimension one, and thus the cokernel of Mα
is of dimension one. Consequently the exactness in (9) implies that the map Lα in
the representation L is surjective, and therefore the sequence s(p) does not end in
a, by Lemma 6.4(ii).
Next we show that s(p) does not end in b or c. Suppose first that dim Mb =
dim Ma = ` + 1 then dim Lb = dim La + 1 and the sequence s(p) starts with b and
Lα = V` . The morphism f of (9) induces a commutative diagram
 KNOT THEORY AND CLUSTER ALGEBRAS 31

Lb
Lα =V`
/ La

fb =I`+1 fa =H`
 
Mb
Mα
/ Ma
whence Mα = J` .
Then wp = δγβα acts like Mα , and we obtain Mβ = Mγ = Mδ = I. Thus
Lβ = Lγ = I. Again Lemma 6.4(ii) implies that the sequence s(p) does not end in
b or c. Therefore s(p) must end in d, and we are done.
On the other hand, if dim Mb 6= dim Ma then dim Mb = dim Ma − 1 = ` and
dim Lb = dim La = `. Then the definition of L = M (S ) implies that the map
Lα : Lb → La is either I` or J`−1 . In the latter case, the sequence s(p) would end in
a, a contradiction. Thus Lα = I` . In view of the sequence (9), we obtain Mα = H` ,
and from equation (8), we have Mα ◦ Mβ ◦ Mγ ◦ Mδ = J` . In particular, Mβ , Mγ
and Mδ are surjective, and using (9) again, we see that Lβ and Lγ are surjective
as well, since β and γ are not incident to the vertex a. Now Lemma 6.4(ii) implies
that s(p) does not end in b or c. This completes the proof of part (f). 
Corollary 6.8. (a) For every dimension vector e the quiver Grassmannian Gre (T (i))
is either empty or a point. In particular, the Euler characteristic
χ(Gre (T (i))) = 0 or 1.
(b) The F -polynomial of T (i) is
X
FT (i) = ydim L ,
L⊂T (i)
Q2n
where the sum is over all submodules of T (i) and ydim L = i=1 yidim Li .
Proof. In the proof of part (f) of the proposition, we have seen that every submodule
of T (i) is determined by its dimension vector. Thus if there is a submodule of
dimension vector e then it is unique, and Gre (T (i)) is a point. Otherwise it is
empty. This shows (a), and (b) follows directly. 

We are ready for the main result of this section.

Theorem 6.9. The map S 7→ M (S ) is a lattice isomorphism from the lattice of
Kauffman states of K relative to the segment i to the lattice of submodules of the
module T (i).
Proof. The map is well-defined by parts (a) and (d) of Proposition 6.7, injective
by part (b) and surjective by part (f). Part (e) implies that it is order preserving.
Moreover the maximum and minimum elements correspond by parts (c) and (d).

6.3. Indecomposability.
Proposition 6.10. The B-module T (i) is indecomposable.
Proof. First we show that the support of T (i) induces a connected subquiver of
Q. By definition of T (i), the support consists of all segments in K1 \ K(0), which
is obtained from K by removing the two regions R1 , R2 that are incident to the
32 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

segment i. If K \ {R1 , R2 } is disconnected then we can draw a closed curve γ in

the plane that separates one connected component from the rest. Putting back
R1 and R2 , we see that γ crosses exactly two regions in K. This means that K
is a connected sum which contradicts our assumption that K is prime. Thus the
support of T (i) is connected.
Suppose now T (i) = M ⊕ N is the direct sum of two nonzero B-modules. Let
α : j → k be an arrow in Q such that T (i)j and T (i)k are nonzero. By definition,
the linear map T (i)α are nonzero, except possibly if there is a crossing point cycle
j / k / l / m / j such that T (i) has dimension one at vertices j, k, l and m, in
α

which case one of the four maps is zero and the others are the identity. Therefore,
since the support of T (i) induces a connected subquiver of Q, the supports of the
two submodules M and N cannot be disjoint.
Then there exists j ∈ Q0 such that Mj and Nj are both nonzero. Since j is a
segment in K, there exists a crossing point p ∈ K0 that is incident to j. Let wp
denote the corresponding 4-cycle in Q. Because of Remark ??, the vector space
T (i)j = Mj ⊕ Nj has a basis {e1 , e2 , . . . , ed } with respect to which the action of wp
is given by the matrix Jd−1 .
Now let m ∈ Mj , n ∈ Nj be nonzero elements and denote their expansions in
the basis as
Xn X n
m= µk ek n= νk ek ,
k=1 k=1
with µk , νk ∈ k. Let km and kn be the largest indices such that µkm 6= 0 and
νkn 6= 0. Then m · wpkm −1 = µkm e1 ∈ Mj and n · wpkn −1 = νkn e1 ∈ Nj , because M
and N are right B-modules. Dividing by the scalars shows that e1 ∈ Mj ∩ Nj , a
contradiction to the assumption that the sum T (i) = M ⊕ N is direct. 

7. The main result

In this section, we prove that the Alexander polynomial of the link is a special-
ization of the F -polynomial of any indecomposable summand of the link diagram
module.
Let K be an oriented link diagram of a prime link and assume that K contains
no curls. Let n be the number of crossing points in K, and let i be a segment in
K. Let T (i) be the P corresponding indecomposable summand of the link diagram
module and FT (i) = L⊂T (i) ydim L its F -polynomial as in Corollary 6.8(b).
For f ∈ Z[y1 , . . . , y2n ] we write f |t for the specialization of f at

 −t if segment j runs from an undercrossing to an overcrossing;
yj = −t−1 if segment j runs from an overcrossing to an undercrossing;
−1 if segment j connects two overcrossings or two undercrossings.

(10)
Theorem 7.1. The Alexander polynomial of K is equal to the specialization (10)
of the F -polynomial of every indecomposable summand T (i) of the link diagram
module T . That is
∆ = FT (i) |t .
Proof. Let ∆ denote the Alexander polynomial of K. Kauffman’s theorem says
that
. X
∆= σ(S ) w(S ), (11)
S
 KNOT THEORY AND CLUSTER ALGEBRAS 33

where the sum is over all states, σ(S ) = ±1, and w(S ) is a power of t. The symbol
.
= means that the expressions on either side are equal up to sign and up to a power
of t. We denote by Smin the minimal state. Normalizing the above identity, we
find
. X σ(S ) w(S )
∆= . (12)
σ(Smin ) w(Smin )
S
Let s = j1 , j2 , . . . , jt be a sequence of transpositions that transforms Smin into
S , and let M (S ) denote the state module introduced in Definition 6.1. Then
t
X
dim M (S ) = ejk , (13)
k=1

where ejk ∈ Z2n is the vector that is 1 at position jk and 0 elsewhere.

Recall that, if a state S 0 is obtained from a state S by a s single transposition at
a segment j ∈ K1 then w(j) = w(S 0 )/w(S ) is independent of the particular states
S , S 0 and only depends on the segment j. Moreover, in this situation, Lemma 2.7
of [14] implies that σ(S 0 ) = −σ(S ). Thus
σ(S ) w(S )
= (−w(j1 ))(−w(j2 )) . . . (−w(jt )). (14)
σ(Smin ) w(Smin )
From our table in Figure 4 we know that −w(j) is equal to the specialization of yj
at (11). Thus
−w(j) = yj |t = yej |t .
Therefore the right hand side of equation (14) is equal to ydim M (S ) |t . Applying
this result to the formula in (12), we obtain
. X dim M (S )
∆= y |t .
S
Now Theorem 6.9 implies
. X dim L
∆= y |t = FT (i) |t ,
L⊂T (i)

where the sum is over all submodules of T (i), and thus it is the specialized F -
polynomial. 

8. A special case: 2-bridge links

A special class of links is the family of 2-bridge links K[a1 ,a2 ,...,an ] which were first
studied by Schubert in [28]. These links are parametrized by continued fractions
1
[a1 , a2 , . . . , an ] = a1 + (15)
1
a2 +
1..
.+
an
with ai ∈ Z≥1 . The link K[a1 ,...,an ] consists of n braids on two strands whose number
of crossings is given by the ai and that are joined together in a linear fashion as
shown in Figure 14 and such that the resulting link is alternating. K[a1 ,...,an ] is a
knot if the numerator of the continued fraction (15) is odd, and it is a link with
exactly two components otherwise. For example, K[2, 1, 2, 3] in Figure 14 is a knot,
because [2, 1, 2, 3] = 27/10.
34 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

Figure 14. The 2-bridge knot K[2, 1, 2, 3]

Let i be the long segment at the bottom of the link diagram that connects the
a1 -braid to the an -braid. Then our construction of the link module T (i) produces
a Dynkin type A module whose support is given by one of the quivers
1o 2o ... o `1 / ... / `2 o ... o `n−1 / ... / (`n − 1) ,

1o 2o ... o `1 / ... / `2 / ... / `n−1 o ... o (`n − 1) ,

where `j = a1 + a2 + · · · + aj and direction of the arrows at the right end depend on
the parity of n. The module T (i) is of dimension one at each vertex 1, 2, . . . (`n − 1)
and its linear maps on the arrows shown above are the identity maps.
In this situation, the F -polynomial of T (i) can be computed as a sum over all
perfect matchings of the snake graph G[a1 ,...,an ] associated to the continued fraction
in [5]. We have X
FT (i) = y(P )
P ∈Match G[a1 ,...,an ]

where y(P ) is the height function of the poset of perfect matchings [21].
Remark 8.1. This module T (i) was implicitly used in [17], where the Jones poly-
nomial was realized as the specialization of the F -polynomial of T (i) at y1 = t−2
and yj = −t−1 for all j = 2, 3, . . . , n. We do not know how to generalize this
specialization to other segments i of this link, or to other type of links.
We obtain two consequences from the above discussion.
8.1. Type A cluster variables correspond to links. We have the following
result.
Theorem 8.2. Let Q be a quiver of Dynkin type A and let A(Q) be its cluster
algebra. For every non-initial cluster variable x ∈ A(Q) there exists a link diagram
K and a segment i ∈ K1 such that the indecomposable summand T (i) of the link
module is mapped to x under the Caldero-Chapoton map.
Proof. Let CC denote the Caldero-Chapoton map. Since x is non-initial, there
exists an indecomposable kQ-module M such that CC(M ) = x, [4, 3]. The support
of M defines a connected subquiver of Q which in turn determines a continued
fraction [5] and hence a 2-bridge link K. From the discussion above and our main
theorem, we have a segment i ∈ K1 such that T (i) = M . 
8.2. An application to q-deformed rationals. In [20], Morier-Genoud and
Ovsienko introduced q-deformed rationals and q-deformed continued fractions. They
propose a unimodality conjecture that can be rephrased in terms of the specialized
height function as follows.
Let M be an indecomposable type A module and let h be the linearization
of the lattice of submodules of M that maps submodules L of M to their total
 KNOT THEORY AND CLUSTER ALGEBRAS 35

P
dimension. Thus h(L) = dim L = j∈Q0 dim Lj . Equivalently, we can think of h
as a linearization of the lattice of perfect matchings of the associated snake graph G
that maps a perfect matching P to the specialization of the height function setting
all y-variables equal to t. Thus h(P ) = y(P )|yj =t . In other words, h associates
to each lattice element the length of the shortest chain from the element to the
minimal element in the lattice.
Conjecture 8.3 (Morier-Genoud–Ovsienko). The function h is unimodal.
Progress towards this conjecture has been made in [19].
Using our main theorem and properties of the Alexander polynomial, we have
the following result, which says that the alternating sum of the number of objects
on each level of the poset is −1, 0, or 1.
Theorem 8.4. Let M be a module of Dynkin type An and L the submodule lattice
of M . Then

X
h(L) ±1 if |L| is odd;
(−1) =
0 if |L| is even.
L∈L

Proof. Combining Theorems 8.2 and 7.1, we see that there exists a 2-bridge link K
whose Alexander polynomial is the specialized F -polynomial of M . More precisely,
. X
∆ K = FM | t = ydim L |t .
L∈L

From the definition of the specialization (10), we see that evaluating the above
equation at t = 1 gives
X
∆K (1) = ydim L |yi =−1 . (16)
L∈L
Furthermore
yidim Li |yi =−1 = (−1)
P
ydim L |yi =−1 = dim Li
= (−1)h(L) .
Q
i
i

Thus equation (16) becomes

X
∆K (1) = (−1)h(L) .
L∈L

Now the result follows from property (ii) of subsection 2.1.1 and the fact that K is
a knot if and only if the number of submodules of M is odd. 

9. Examples
Example 9.1. Consider the figure-eight knot. We use the same labelling of seg-
ments as in Example 3.1. The lattice of the sumbodules of the module T (1) is
shown at Figure 15a.
The lattice isomporphism with the lattice of Kaufman states with regards to the
segment 1 is obvious, see Figure 3. The F -polynomial of T (1) is
FT (1) = 1 + y2 + y8 + y2 y8 + y2 y5 y8
and its specialization at y2 = −t, y5 = −t−1 and y8 = −t, as given by Equation 10,
is
FT (1) |t = 1 − 3 + t2 .
36 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

4
1
3
8

28 1
5 3
8
2 8
5 5
3
8
5

0 8
(a) Submodules of T (1)

0
(b) Submodules of T (2)

Figure 15. Lattices of submodules of indecomposable modules

on the Jacobian algebra given by the figure-eight knot.

Remark that the lattice of sumbodules of T (2) is completely different for T (1),
see Figures 15a and 15b. The F -polynomial of T (2) is
FT (2) = 1 + y8 + y3 y8 + y1 y3 y8 + y1 y3 y4 y8
and its specialization at y1 = −t−1 , y3 = −t−1 , y4 = −t and y8 = −t is
FT (2) |t = −t−1 + 3 − t.
Example 9.2. For the representation T (1) in Figure 11, the F -polynomial is the
following polynomial with 75 terms. This was computed using [15].
1 + y4 + y12 y4 + y8 + y18 y8 + y10 y18 y8 + y10 y18 y19 y8 + y10 y18 y19 y2 y8 + y10 y18 y19 y2 y20 y8 + y4 y8
+y12 y4 y8 + y18 y4 y8 + y10 y18 y4 y8 + y12 y18 y4 y8 + y10 y12 y18 y4 y8 + y10 y18 y19 y4 y8 + y10 y12 y18 y19 y4 y8
+y10 y18 y19 y2 y4 y8 + y10 y12 y18 y19 y2 y4 y8 + y10 y18 y19 y2 y20 y4 y8 + y10 y12 y18 y19 y2 y20 y4 y8 + y10 y18 y19 y8 y9
+y10 y17 y18 y19 y8 y9 + y10 y18 y19 y2 y8 y9 + y10 y17 y18 y19 y2 y8 y9 + y10 y18 y19 y2 y20 y8 y9
+y10 y17 y18 y19 y2 y20 y8 y9 + y10 y18 y19 y4 y8 y9 + y10 y12 y18 y19 y4 y8 y9 + y10 y17 y18 y19 y4 y8 y9
+y10 y12 y17 y18 y19 y4 y8 y9 + y10 y18 y19 y2 y4 y8 y9 + y10 y12 y18 y19 y2 y4 y8 y9 + y10 y17 y18 y19 y2 y4 y8 y9
+y10 y12 y17 y18 y19 y2 y4 y8 y9 + y10 y18 y19 y2 y20 y4 y8 y9 + y10 y12 y18 y19 y2 y20 y4 y8 y9
+y10 y17 y18 y19 y2 y20 y4 y8 y9 + y10 y12 y17 y18 y19 y2 y20 y4 y8 y9 + y10 y17 y18 y19 y4 y7 y8 y9
+y10 y12 y17 y18 y19 y4 y7 y8 y9 + y10 y16 y17 y18 y19 y4 y7 y8 y9 + y10 y12 y16 y17 y18 y19 y4 y7 y8 y9
+y10 y17 y18 y19 y2 y4 y7 y8 y9 + y10 y12 y17 y18 y19 y2 y4 y7 y8 y9 + y10 y16 y17 y18 y19 y2 y4 y7 y8 y9
+y10 y12 y16 y17 y18 y19 y2 y4 y7 y8 y9 + y10 y17 y18 y19 y2 y20 y4 y7 y8 y9 + y10 y12 y17 y18 y19 y2 y20 y4 y7 y8 y9
+y10 y16 y17 y18 y19 y2 y20 y4 y7 y8 y9 + y10 y12 y16 y17 y18 y19 y2 y20 y4 y7 y8 y9 + y10 y17 y18 y19 y2 y20 y3 y4 y7 y8 y9
+y10 y12 y17 y18 y19 y2 y20 y3 y4 y7 y8 y9 + y10 y16 y17 y18 y19 y2 y20 y3 y4 y7 y8 y9 + y10 y12 y16 y17 y18 y19 y2 y20 y3 y4 y7 y8 y9
+y10 y12 y17 y18 y19 y2 y20 y3 y4 y6 y7 y8 y9 + y10 y12 y16 y17 y18 y19 y2 y20 y3 y4 y6 y7 y8 y9
2 2 2 2
+y10 y16 y17 y18 y19 y4 y7 y8 y9 + y10 y12 y16 y17 y18 y19 y4 y7 y8 y9 + y10 y16 y17 y18 y19 y4 y7 y8 y9
2 2 2 2
+y10 y12 y16 y17 y18 y19 y4 y7 y8 y9 + y10 y16 y17 y18 y19 y2 y4 y7 y8 y9 + y10 y12 y16 y17 y18 y19 y2 y4 y7 y8 y9
2 2 2 2 2
+y10 y16 y17 y18 y19 y2 y4 y7 y8 y9 + y10 y12 y16 y17 y18 y19 y2 y4 y7 y8 y9 + y10 y16 y17 y18 y19 y2 y20 y4 y7 y8 y9
2 2 2 2 2
+y10 y12 y16 y17 y18 y19 y2 y20 y4 y7 y8 y9 + y10 y16 y17 y18 y19 y2 y20 y4 y7 y8 y9 + y10 y12 y16 y17 y18 y19 y2 y20 y4 y7 y8 y9
2 2 2 2
+y10 y16 y17 y18 y19 y2 y20 y3 y4 y7 y8 y9 + y10 y12 y16 y17 y18 y19 y2 y20 y3 y4 y7 y8 y9 + y10 y16 y17 y18 y19 y2 y20 y3 y4 y7 y8 y9
2 2 2
+y10 y12 y16 y17 y18 y19 y2 y20 y3 y4 y7 y8 y9 + y10 y12 y16 y17 y18 y19 y2 y20 y3 y4 y6 y7 y8 y9
2 2
+y10 y12 y16 y17 y18 y19 y2 y20 y3 y4 y6 y7 y8 y9

The specialization is
FT (18) |t = 3 − 9t + 16t2 − 19t3 + 16t4 − 9t5 + 3t6 ,
which is equal to the Alexander polynomial of the corresponding knot 1066 .
 KNOT THEORY AND CLUSTER ALGEBRAS 37

Example 9.3. For the Conway knot, the quiver is

/7
C1 _ I

o r
D6 D2

  
17\ 16\ 15W /9 / 13
O O

   
;5 ;3 14 o 8U

 # {
22 g 21O ? 10

  
4o 20 g 19O

  ' 
18 i / 11 12

The F -polynomial of T (18) has 131 terms. The highest degree term is

y2 y3 y5 y6 y8 y9 y10 y13 y14 y15 y16 y17 y19 y21

The specialization gives FT (18) |t = t, confirming that the Alexander polynomial is
trivial.

References
[1] J. W. Alexander, Topological invariants of knots and links, Trans. Amer. Math. Soc. 30 (2),
(1928) 275–306.
[2] I. Assem, R. Schiffler, V. Shramchenko, Cluster automorphisms, Proc. London Math. Soc. 3
no. 104, 1271-1302 (2012).
[3] P. Caldero and F. Chapoton, Cluster algebras as Hall algebras of quiver representations.
Comment. Math. Helv. 81 (2006), no. 3, 595–616.
[4] P. Caldero, F. Chapoton and R. Schiffler, Quivers with relations arising from clusters (An
case). Trans. Amer. Math. Soc. 358 (2006), no. 3, 1347–1364 (electronic).
[5] İ. Çanakçı and R. Schiffler, Cluster algebras and continued fractions, Compos. Math., 54 (3)
(2018) 565–593.
[6] M. Cohen, O. Dasbach and H. Russell, A twisted dimer model for knots. Fund. Math. 225
(2014), no. 1, 57–74.
38 VÉRONIQUE BAZIER-MATTE AND RALF SCHIFFLER

[7] J. H. Conway, An enumeration of knots and links, and some of their algebraic properties.
1970 Computational Problems in Abstract Algebra (Proc. Conf., Oxford, 1967) pp. 329–358
Pergamon, Oxford.
[8] H. Derksen, J. Weyman and A. Zelevinsky, Quivers with potentials and their representations.
I. Mutations. Selecta Math. (N.S.) 14 (2008), no. 1, 59–119.
[9] H. Derksen, J. Weyman and A. Zelevinsky, Quivers with potentials and their representations
II: Applications to cluster algebras. J. Amer. Math. Soc. 23 (2010), no. 3, 749–790.
[10] S. Fomin, L. Williams and A. Zelevinsky, Introduction to Cluster Algebras, Chapter 7,
arXiv:2106.02160.
[11] S. Fomin and A. Zelevinsky, Cluster algebras I: Foundations, J. Amer. Math. Soc. 15 (2002),
497–529.
[12] S. Fomin and A. Zelevinsky, Cluster algebras IV: Coefficients, Compos. Math. 143 (2007),
112–164.
[13] M. H. Freedman, A surgery sequence in dimension four; the relations with knot concordance.
Invent. Math. 68 (1982), no. 2, 195-226.
[14] L. Kauffman, Formal Knot Theory, Mathematical Notes, 30. Princeton University Press,
Princeton, NJ, 1983.
[15] B. Keller, Mutation applet.
[16] K. Lee and R. Schiffler, Positivity for cluster algebras, Annals of Math. 182 (1), (2015)
73–125.
[17] K. Lee and R. Schiffler, Cluster algebras and Jones polynomials, Sel. Math. New Ser. (2019)
25: 58.
[18] W. B. R. Lickorish, An introduction to knot theory. Graduate Texts in Mathematics, 175.
Springer-Verlag, New York, 1997.
[19] T. McConville, B. Sagan and C. Smyth, On a rank-unimodality conjecture of Morier-Genoud
and Ovsienko. Discrete Math. 344 (2021), no. 8, Paper No. 112483, 13 pp.
[20] S. Morier-Genoud and V. Ovsienko, q-deformed rationals and q-continued fractions, preprint,
arXiv:1812.00170. Forum Math. Sigma 8 (2020), Paper No. e13, 55 pp.
[21] G. Musiker and R. Schiffler, Cluster expansion formulas and perfect matchings, J. Algebraic
Combin. 32 (2010), no. 2, 187–209.
[22] W. Nagai and Y. Terashima, Cluster variables, ancestral triangles and Alexander polynomials,
Adv. Math. 363 (2020), 106965, 37 pp.
[23] P. Ozsváth, and Z. Szabó, Holomorphic disks and knot invariants. Adv. Math. 186 (2004),
no. 1, 58–116.
[24] L. Piccirillo, The Conway knot is not slice, Ann. of Math. (2) 191 (2020), no. 2, 581–591.
[25] J. A. Rasmussen, Floer homology and knot complements. Thesis (Ph.D.)–Harvard University.
2003. 126 pp.
[26] R. Schiffler, Cluster algebras from surfaces: lecture notes for the CIMPA School Mar del
Plata, March 2016. Homological methods, representation theory, and cluster algebras, 65–99,
CRM Short Courses, Springer, Cham, 2018.
[27] R. Schiffler and D. Whiting, Tilting modules arising from knot invariants, preprint, 18 pages,
arXiv:2001.04004.
[28] H. Schubert, Knoten mit zwei Brücken. (German) Math. Z. 65 (1956), 133–170.

Department of Mathematics, University of Connecticut, Storrs, CT 06269-1009, USA

Email address: [email protected]

Department of Mathematics, University of Connecticut, Storrs, CT 06269-1009, USA

Email address: [email protected]