PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A

TRIANGULAR LATTICE

PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

Abstract. We prove the Pfaffian Sign Theorem for the dimer model on a triangular lat-
arXiv:1711.00032v1 [math-ph] 31 Oct 2017

tice embedded in the torus. More specifically, we prove that the Pfaffian of the Kaste-
leyn periodic-periodic matrix is negative, while the Pfaffians of the Kasteleyn periodic-
antiperiodic, antiperiodic-periodic, and antiperiodic-antiperiodic matrices are all positive.
The proof is based on the Kasteleyn identities and on small weight expansions. As an appli-
cation, we obtain an asymptotics of the dimer model partition function with an exponentially
small error term.

1. Introduction
1.1. Dimer model on a triangular lattice. We consider the dimer model on a triangular
lattice m,n = (Vm,n , Em,n ) on the torus Zm Zn = Z2 /(mZ nZ) (periodic boundary
conditions), where Vm,n and Em,n are the sets of vertices and edges of m,n , respectively. It
is convenient to consider m,n as a square lattice with diagonals. A dimer on m,n is a set
of two neighboring vertices hx, yi connected by an edge. A dimer configuration on m,n
is a set of dimers = {hxi , yi i, i = 1, . . . , mn
2
} which cover Vm,n without overlapping. An
example of a dimer configuration is shown in Fig. 1. An obvious necessary condition for a
configuration to exist is that at least one of m, n is even, and so we assume that m is even,
m = 2m0 .

Figure 1. Example of a dimer configuration on a triangular 6 6 lattice on the torus.

To define a weight of a dimer configuration, we split the full set of dimers in a configuration
into three classes: horizontal, vertical, and diagonal with respective weights zh , zv , zd > 0.
If we denote the total number of horizontal, vertical and diagonal dimers in by Nh (),

Date: November 2, 2017.

The first author is supported in part by the National Science Foundation (NSF) Grants DMS-1265172
and DMS-1565602.
1
2 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

Nv (), and Nd (), respectively, then the dimer configuration weight is

mn
2
Y N () Nv () Nd ()
w() = w(xi , yi ) = zh h zv zd , (1.1)
i=1

where w(xi , yi ) denotes the weight of the dimer hxi , yi i . We denote by m,n the set of
all dimer configurations on m,n . The partition function of the dimer model is given by
X
Z= w(). (1.2)
m,n

Notice that if all the weights are set equal to one, then Z simply counts the number of dimer
configurations, or perfect matchings, on m,n .
1.2. Main result. As shown by Kasteleyn [7][9], the partition function Z of the dimer
model on the torus can be expressed in terms of the four Kasteleyn Pfaffians as
1
Z = (Pf A1 + Pf A2 + Pf A3 + Pf A4 ) , (1.3)
2
with periodic-periodic, periodic-antiperiodic, antiperiodic-periodic, and antiperiodic-antiperiodic
boundary conditions in the x- and y-axis, respectively. For the reader convenience we include
a proof of formula (1.3) in Appendix B below. For an extension of formula (1.3) to graphs on
Riemannian surfaces of higher genera see the works of Galluccio and Loebl [3], Tesler [12],
Cimasoni and Reshetikhin [1], and references therein. Formula (1.3) is very powerful in the
asymptotic analysis of the partition function as m, n , because the absolute value of
the Pfaffian of a square antisymmetric matrix A is determined by its determinant through
the classical identity
(Pf A)2 = det A. (1.4)
The asymptotic behavior of det Ai as m, n can be analyzed by a diagonalization of
the matrices Ai (see, e.g., [7], [2], and Appendix C below), and an obvious problem arises to
determine the sign of the Pfaffians Pf Ai in formula (1.3).
In [7] Kasteleyn considered the dimer model on the square lattice, which corresponds to
the weight zd = 0. He showed that in this case Pf A1 = 0 and he assumed that Pf Ai 0 for
i = 2, 3, 4. Kenyon, Sun and Wilson [6] established the sign of the Pfaffians Pf Ai for any
critical dimer model on a lattice on the torus, including the square lattice. The dimer model
on the triangular lattice is not critical and the results of [6] are not applicable in this case.
Different conjectures about the Pfaffian signs for the dimer model on a triangular lattice are
stated, without proof, in the works of McCoy [10], Fendley, Moessner, and Sondhi [2], and
Izmailian and Kenna [5].
Our main result in this paper is the following theorem:
Theorem 1.1 (Pfaffian Sign Theorem). Let zh , zv , zd > 0. Then
Pf A1 < 0, Pf A2 > 0, Pf A3 > 0, Pf A4 > 0. (1.5)
The proof of this theorem is given below and it is based on the following two important
ingredients:
(1) The Kasteleyn formulae for the Pfaffians Pf Ai in terms of algebraic sums of the
partition functions of the dimer model restricted to different Z2 homology classes.
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 3

(2) An asymptotic analysis of Pf Ai as one of the weights tends to zero. It is worth

noting that due to various cancellations in the Pfaffians the leading terms in the
small weight asymptotics of the Pfaffians Pf Ai depend on arithmetic properties of the
lattice dimensions m and n, and it requires a geometric description of configurations
giving the leading contribution to the Pfaffians Pf Ai for different m, n.
As an application of Theorem 1.1, we obtain an asymptotic behavior of the partition
function Z as m, n with an exponentially small error term.
The set-up for the remainder of the paper is the following. In Section 2 we review Pfaffians
of the Kasteleyn matrices and their properties. In Section 3 we prove various preliminary
results about the Pfaffians Pf Ai . In Section 4 we prove that Pf A3 > 0, Pf A4 > 0, in Section
5 that Pf A2 > 0, and in Section 6 that Pf A1 < 0, which is the most difficult part of our
study. In Section 7 we prove the identities Pf A1 = Pf A2 and Pf A3 = Pf A4 for odd n. In
Section 8 we obtain the asymptotics of the partition function as m, n . In Appendix A
we prove a sign formula for the superposition of two dimer configurations. In Appendix B
we prove Kasteleyn identities for the triangular lattice on the torus, and in Appendix C we
review the block diagonalization of the Kasteleyn matrices and the double product formulae
for their determinants. Finally, in Appendix D we present numerical data for the Pfaffians
Pf Ai for different dimensions m and n.

2. Dimer model and Kasteleyn matrices

We consider different orientations on the set of the edges Em,n : O1 (periodic-periodic)
O2 (periodic-antiperiodic), O3 (antiperiodic-periodic), and O4 (antiperiodic-antiperiodic),
depicted in Fig. 2 for m = 4, n = 3. All these orientations are Kasteleyn orientations, so

Figure 2. The Kasteleyn orientations for m = 4, n = 3: O1 (periodic-

periodic), O2 (periodic-antiperiodic), O3 (antiperiodic-periodic), and O4
(antiperiodic-antiperiodic).

that for any face the number of arrows on the boundary oriented clockwise is odd.
With every orientation Oi we associate a sign function i (x, y), x, y Vm,n , defined as
follows: if x and y are connected by an edge then
1, if the arrow in the Kasteleyn orientation Oi points from x to y,

i (x, y) = (2.1)
1, if the arrow in the Kasteleyn orientation Oi points from y to x,
and
i (x, y) = 0, if x and y are not connected by an edge. (2.2)
More specifically, the sign functions are given by the following formulae.
4 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

Let e1 = (1, 0), e2 = (0, 1), and x = (j, k) Zm Zn . Then the sign function 1 takes the
values
1 (x, x + e1 ) = 1,
1 (x, x + e2 ) = (1)j , (2.3)
1 (x, x + e1 + e2 ) = (1)j+1 .
The sign function 2 is obtained from 1 by the reversal of the vertical and diagonal arrows
in the upper row, so that
2 (x, x + e1 ) = 1,
2 (x, x + e2 ) = (1)j+k,n1 , (2.4)
j+1+k,n1
2 (x, x + e1 + e2 ) = (1) .
Similarly, the sign function 3 is obtained from 1 by the reversal of the horizontal and
diagonal arrows in the last column, so that
3 (x, x + e1 ) = (1)j,m1 ,
3 (x, x + e2 ) = (1)j , (2.5)
3 (x, x + e1 + e2 ) = (1)j+1+j,m1 .
Finally, the sign function 4 is obtained from 1 by the reversal of both the vertical and
diagonal arrows in the upper row and the horizontal and diagonal arrows in the last column,
so that
4 (x, x + e1 ) = (1)j,m1 ,
4 (x, x + e2 ) = (1)j+k,n1 , (2.6)
4 (x, x + e1 + e2 ) = (1)j+1+j,m1 +k,n1 .
In addition, (2.2) holds and i (y, x) = i (x, y).
With every orientation Oi we associate a Kasteleyn matrix Ai . To define the Kasteleyn
matrices, consider any enumeration of the vertices, Vm,n = {x1 , . . . , xmn }. Then the Kaste-
leyn matrices Ai are defined as


Ai = ai (xj , xk ) 1j,kmn , i = 1, 2, 3, 4, (2.7)
with
i (x, y)w(x, y), if x and y are connected by an edge,

ai (x, y) = (2.8)
0 otherwise,
where w(x, y) = zh , zv , zd is the weight of the dimer hx, yi and i is the sign function. Consider
now the Pfaffians Pf Ai .
The Pfaffian, Pf Ai , of the mn mn antisymmetric matrix Ai , i = 1, 2, 3, 4, is a number
given by X
Pf Ai = (1) ai (xp1 , xp2 )ai (xp3 , xp4 ) ai (xpmn1 , xpmn ), (2.9)

where the sum is taken over all permutations,
 
1 2 3 mn 1 mn
= , (2.10)
p1 p2 p3 pmn1 pmn
which satisfy the following restrictions:
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 5

(1) p2`1 < p2` , 1 ` mn

2
,
mn
(2) p2`1 < p2`+1 , 1 ` 2 1.
Such permutations are in a one-to-one correspondence with the dimer configurations, and
(2.9) can be rewritten as
X Y
Pf Ai = (1)() w() i (x, y), i = 1, 2, 3, 4. (2.11)
m,n hx,yi

An important property of the Kasteleyn Pfaffians Pf Ai is that they do not depend on the
enumeration of the vertices, Vm,n = {x1 , . . . , xmn }.
The sign of a configuration , sgn () = sgn (; Ai ), is the following expression:
Y
sgn () = (1)() i (x, y), (2.12)
hx,yi

where i (x, y) is given by (2.1). Having (2.12), the Pfaffian formula for a Kasteleyn matrix
Ai can be rewritten as
X
Pf Ai = sgn ()w(), sgn () = sgn (; Ai ). (2.13)
m,n

Given two configurations and 0 , we consider the double configuration 0 , and we call
it the superposition of and 0 . In 0 , we define a contour to be a cycle consisting of
alternating edges from and 0 . Each contour consists of an even number of edges. The
superposition 0 is partitioned into disjoint contours {k : k = 1, 2, . . . , r}. We will call
a contour consisting of only two edges a trivial contour.
Let us introduce a standard configuration st as follows. Consider the lexicographic or-
dering of the vertices (i, j) Zm Zn . Namely,
(i, j) = xk , k = jm + i + 1, 1 k mn. (2.14)
Then n mn o
st = hx2l1 , x2l i, l = 1, . . . , . (2.15)
2
Observe that
sgn (st ; Ai ) = +1, i = 1, 2, 3, 4. (2.16)
because (st ) = Id and i (x2l1 , x2l ) = +1.
We will use the following lemma:
Lemma 2.1. (see [6], [12]). Let , 0 be any two configurations and {k : k = 1, 2, . . . , r}
all contours of 0 . Then
r
Y
0
sgn (; Ai ) sgn ( ; Ai ) = sgn (k ; Oi ), i = 1, 2, 3, 4, (2.17)
k=1

with
sgn (k ; Oi ) = (1)k (Oi )+1 , (2.18)
where k (Oi ) is the number of edges in k oriented clockwise with respect to the orientation
Oi .
For the convenience of the reader we give a proof of this lemma in Appendix A.
6 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

3. Preliminary results
As shown by Kasteleyn [7, 9] (for recent expositions see the works [6], [10], [11] and
references therein), the partition function Z can be decomposed as
Z = Z 00 + Z 10 + Z 01 + Z 11 , (3.1)
rs
the four partition functions Z corresponding to dimer configurations of the homology classes
(r, s) Z2 Z2 , and the Pfaffians Pf Ai are expressed as
Pf A1 = Z 00 Z 10 Z 01 Z 11 , Pf A2 = Z 00 Z 10 + Z 01 + Z 11 ,
(3.2)
Pf A3 = Z 00 + Z 10 Z 01 + Z 11 , Pf A4 = Z 00 + Z 10 + Z 01 Z 11 .
These equations are called the Kasteleyn identities. Observe that equations (3.1), (3.2) imply
(1.3).
The proof of the Kasteleyn identities (3.2) can be found in the works of Galluccio and
Loebl [3], Tesler [12], Cimasoni and Reshetikhin [1], and in Appendix B below. It follows
from formula (2.11) that Pf Ai are multivariate polynomials with respect to the weights
zh , zv , zd . By diagonalizing the matrices Ai , one can obtain the double product formulas,
m
2
1 n1 "
Y Y 2(j + i ) 2(k + i )
det Ai = 4 zh2 sin2 + zv2 sin2
j=0 k=0
m n
 # (3.3)
2(j + i ) 2(k + i )
+ zd2 cos2 + ,
m n
with
1 1 1
1 = 1 = 0 ; 2 = 0 , 2 = ; 3 = , 3 = 0 ; 4 = 4 = (3.4)
2 2 2
(see, e.g., [2] and [5]). For the convenience of the reader, we give a proof of formula (3.3) in
Appendix C.
The function
S(x, y) = 4 zh2 sin2 2x + zv2 sin2 2y + zd2 cos2 (2x + 2y)
 
(3.5)
is the spectral function of the dimer model on the triangular lattice. In its terms equation
(3.3) is conveniently written as
m
2
1 n1  
Y Y j + i k + i
det Ai = S , . (3.6)
j=0 k=0
m n
The function S(x, y) is periodic in x and y,
   
1 1
S x + , y = S x, y + = S(x, y), (3.7)
2 2
and if zh , zv , zd > 0, then
S(x, y) > 0, (x, y) R2 . (3.8)
Indeed, obviously, S(x, y) 0. Suppose S(x, y) = 0. Then from (3.5) we obtain that
1
2x Z, 2y Z, 2x + 2y + Z, (3.9)
2
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 7

which are inconsistent. From (3.6), (3.8) we obtain that

det Ai > 0, if zh , zv , zd > 0, (3.10)
because all the factors in (3.6) are positive.
As a consequence of (3.10), we have that Pf Ai does not change the sign in the region
zh , zv , zd > 0; hence, it is sufficient to establish the sign of Pf Ai at any point of the region
zh , zv , zd > 0. As a first step in this direction, let us prove the following proposition:
Proposition 3.1. We have that
(1) If zh > 0, zv > 0, and zd = 0, then
det A1 = 0, det A3 > 0, det A4 > 0,
= 0, if n 1 (mod 2),

(3.11)
det A2
> 0, if n 0 (mod 2).
(2) If zh > 0, zv = 0, and zd > 0, then
= 0, if n 0 (mod 4),

det A1
> 0, if n 6 0 (mod 4).
= 0, if n 2 (mod 4), (3.12)

det A2
> 0, if n 6 2 (mod 4),
det A3 > 0, det A4 > 0.
(3) If zh = 0, zv > 0, and zd > 0, then
= 0, if m 0 (mod 4),

det A1
> 0, if m 2 (mod 4).
= 0, if m 0 (mod 4) and n 1 (mod 2),

det A2
> 0, otherwise.
(3.13)
= 0, if m 2 (mod 4),

det A3
> 0, if m 0 (mod 4).
= 0, if m 2 (mod 4) and n 1 (mod 2),

det A4
> 0, otherwise.
Proof. (1) Assume that zh > 0, zv > 0, and zd = 0. By (3.3), det Ai = 0 if and only if for
some 0 j m2 1 and 0 k n 1,
2(j + i ) 2(k + i )
Z and Z. (3.14)
m n
In particular, this gives that det A1 = 0, due to the factor j = k = 0. On the other hand,
det A3 > 0, because 2j+1
m
6 Z (since m is even). The same argument works for det A4 > 0.
Consider det A2 . If n 1 (mod 2), then det A2 = 0, due to the factor j = 0, k = n1
2
. On
2k+1
the other hand, if n 0 (mod 2), then n 6 Z and det A2 6= 0.
(2) Assume that zh > 0, zv = 0, and zd > 0. To have det Ai = 0, we need that
2(j + i ) 2(k + i ) 1
Z and + Z. (3.15)
m n 2
8 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

We have that det A1 = 0, provided n 0 (mod 4), due to the factor j = 0, k = n4 . On the
other hand, if n 6 0 (mod 4), then 2k
n
6 12 + Z, hence det A1 > 0.
Consider det A2 . If n 2 (mod 4), then det A2 = 0, due to the factor j = 0, k = n2 4
.
On the other hand, if n 6 2 (mod 4), then 2k+1 n

6 1
2
+ Z, and therefore det A2 > 0. Also,
det A3 > 0, because 2j+1
m

6 Z. The same argument is applied to det A 4 .
(3) Assume that zh = 0, zv > 0, and zd > 0. We need that
2(j + i ) 1 2(k + i )
+ Z and Z. (3.16)
m 2 n
This equation is similar to (3.15), but since we assume that m is even and n can be odd, the
analysis is slightly different. If m 0 (mod 4), then det A1 = 0, due to the factor j = m4 ,
k = 0. On the other hand, if m 2 (mod 4), then 2j m
6 12 + Z, hence det A1 > 0.
If m 0 (mod 4) and n 1 (mod 2), then det A2 = 0, due to the factor j = m4 , k = n1 2
.
On the other hand, if n 0 (mod 2), then 2k+1 n

6 Z, hence det A 2 > 0. If m
6 0 (mod 4),
then 2jm

6 1
2
+ Z, hence det A 2 > 0.
If m 2 (mod 4), then det A3 = 0, due to the factor j = m2 4
, k = 0. On the other hand,
2j+1 1
if m 0 (mod 4), then m 6 2 + Z, hence det A3 > 0.
If m 2 (mod 4) and n 1 (mod 2), then det A4 = 0, due to the factor j = m2 4
,
n1 2k+1
k = 2 . On the other hand, if n 0 (mod 2), then n 6 Z, hence det A4 > 0. If m 0
(mod 4), then 2j+1
m
6 12 + Z, hence det A4 > 0. 

4. Positivity of Pf A3 and Pf A4
Let us turn to the proof of Theorem 1.1. We first prove that Pf A3 > 0 and Pf A4 > 0.
Lemma 4.1. Let zh , zv , zd > 0. Then
Pf A3 > 0, Pf A4 > 0. (4.1)
Proof. If zh > 0, zv > 0, and zd = 0, then by Proposition 3.1 (1), det A1 = 0, hence from
(3.2) we deduce that
Z 00 Z 01 = Z 10 + Z 11 . (4.2)
On the other hand, if zh > 0, zv > 0, and zd = 0, then det A3 > 0. This implies that
Pf A3 6= 0, hence from (3.2), (3.11), and nonnegativity of Z rs we obtain that
Pf A3 = 2Z 10 + 2Z 11 > 0, if zh > 0, zv > 0, zd = 0. (4.3)
By continuity, Pf A3 > 0 for the chosen zh > 0, zv > 0, and small zd > 0. This proves that
Pf A3 > 0 in the whole region zh , zv , zd > 0. The same argument works for Pf A4 > 0. 
We will finish the proof of Theorem 1.1 in the subsequent two sections by showing that
Pf A2 > 0 and Pf A1 < 0, respectively.

5. Positivity of Pf A2
We begin with the following case:
Lemma 5.1. Let zh , zv , zd > 0. Then if either n 0 (mod 2) or m 2 (mod 4), then
Pf A2 > 0. (5.1)
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 9

Proof. First assume that n 0 (mod 2) and consider the case zh > 0, zv > 0, and zd = 0.
Since n is even, then by Proposition 3.1 (1), det A2 > 0 and det A1 = 0. By continuity,
Pf A2 > 0 for the chosen zh , zv , and small zd . This proves Pf A2 > 0 in the whole region
zh , zv , zd > 0. Now assume that m 2 (mod 4) and consider the case zh = 0, zv > 0, and
zd > 0. By Proposition 3.1 (3), in this case det A2 > 0, det A3 = 0, hence
Z 01 Z 10 = Z 00 + Z 11 = Pf A2 = 2Z 00 + 2Z 11 > 0. (5.2)
By continuity, Pf A2 > 0 for the chosen zv > 0, zd > 0, and small zh > 0. This proves that
Pf A2 > 0 in the whole region zh , zv , zd > 0. 
The positivity of Pf A2 in the case m 0 (mod 4), n 1 (mod 2) is more difficult. To
deal with this case, we consider the asymptotic behavior of Pf A2 for zh = 1, zv = 0, as
zd 0, and prove the following lemma:
Lemma 5.2. Let m 0 (mod 4), n 1 (mod 2), and zh = 1, zv = 0. Then as zd 0,
 m n
Pf A2 = 2 zdn (1 + O(zd )). (5.3)
2
Remark. This will imply that Pf A2 > 0 for zh = 1, zv = 0 and sufficiently small zd , and
hence Pf A2 > 0 for all zh , zv , zd > 0.
Proof. We have that
X
Pf A2 = sgn ()w(). (5.4)
m,n

By our assumption, zv = 0, hence there are no vertical dimers. Consider first the limiting
case, zd = 0. In this case there are only horizontal dimers. Let be any configuration
of horizontal dimers and Tk a configuration obtained from by the shift x x + e1 ,
e1 = (1, 0), on the horizontal line y = k. Then the superposition Tk consists of
trivial contours and one horizontal loop around the torus, which has a negative sign as
there are m 0 (mod 2) arrows in the direction of movement from left to right. Hence,
sgn (Tk ) = sgn (), w(Tk ) = w(). From here it follows that
sgn ()w() + sgn (Tk )w(Tk ) = 0,
and therefore, and Tk () cancel each other in Pf A2 . This implies that

Pf A2 z =0 = 0. (5.5)
d

Now let zd > 0, so that there are both horizontal and diagonal dimers. We will consider
zd 0, and we will call configurations consisting of only horizontal dimers the ground
state configurations, because they have the biggest weight w(). As we saw, the ground
state configurations cancel each other in Pf A2 . We will call not ground state configurations
excited state configurations.
Consider the set of configurations (k) in which the line y = k is occupied completely by
horizontal dimers. Let (k) and Tk obtained from by the shift x x + e1 on the
line y = k. Then again sgn (Tk ) = sgn () and hence
X
sgn ()w() = 0.
(k)
10 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

By adding over k Zn , we obtain that

X
sgn ()w() = 0, (5.6)
S
kZn (k)

hence to get excited states without cancellations, we have to consider the set of configurations
such that on each line y = k there is at least one vertex covered by a diagonal dimer. In
fact, since m is even, there are at least two vertices covered by diagonal dimers, hence the
total number of diagonal dimers Ndiag () is at least n. We denote
0 = { m,n | Ndiag (, k) = 2, k Zn }, (5.7)
where Ndiag (, k) is the number of vertices on the line y = k in covered by diagonal dimers.
Denote by Ndiag (, k, k +1) the number of diagonal dimers connecting horizontal line y = k
to horizontal line y = k + 1. Then
Ndiag (, k) = Ndiag (, k 1, k) + Ndiag (, k, k + 1),
hence
Ndiag (, k, k + 1) = 0, 1 or 2, 0 .
Suppose that for some 0 and some k Zn ,
Ndiag (, k, k + 1) = 0,
then
Ndiag (, k + 1, k + 2) = 2, Ndiag (, k + 2, k + 3) = 0, Ndiag (, k + 3, k + 4) = 2, . . . ,
hence the whole lattice is stratified into horizontal strips of width 2 with 2 diagonal dimers
in each strip. But n is odd, hence such stratification is not possible. Similarly, Ndiag (, k, k +
1) = 2 is not possible as well. This implies that if 0 , then
Ndiag (, k, k + 1) = 1, k Zn .
This means that for every k there is a unique diagonal dimer connecting horizontal lines
y = k and y = k + 1. Let (jk , k) be the vertex covered by this diagonal dimer on the
horizontal line y = k. Then, since all vertices between (jk + 1, k + 1) and (jk+1 , k + 1) must
be covered by horizontal dimers we have that
jk+1 (jk + 1) 1 (mod 2), k Zn , (5.8)
or, equivalently,
jk+1 jk 0 (mod 2), k Zn . (5.9)
This implies that
jk j` 0 (mod 2), k, ` Zn . (5.10)
Hence,
1
G
0 = i0 (5.11)
i=0
and i0 , i = 0, 1, if
j` i (mod 2) {(j` , k` ), (j` + 1, k` + 1)}. (5.12)
The proof of Lemma 5.2 is based on the following lemma:
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 11

Lemma 5.3. For any 0 ,

sgn () = +1. (5.13)
Let us finish the proof of Lemma 5.2, assuming Lemma 5.3. Since w() = zdn (recall that
zh = 1) for any 0 , we obtain that
X
sgn ()w() = |0 |zdn . (5.14)
0
m
At y = 0 we have m choices for j0 and then, because of condition (5.9), we have 2
choices
for j1 , j2 , . . . , jn1 . Therefore,
 m n1
|0 | = m , (5.15)
2
hence X  m n1
sgn ()w() = m zdn . (5.16)

2
0

The higher excited states have at least (n + 1) diagonal dimers and therefore their weight is
at most zdn+1 , hence Lemma 5.2 follows. 
It remains to prove Lemma 5.3. We define a stack configuration, stack , to be a configuration
in which all diagonal dimers form a stack between the vertical lines x = 0 and x = 1. The
remaining vertices are occupied by horizontal dimers. See Fig. 4 (a).
Proof of Lemma 5.3. The proof consists of two steps. At Step 1 we show that if 00 (see
equations (5.11), (5.12)), then
sgn () = sgn (stack ), (5.17)
1
and if 0 , then
sgn () = sgn (stack + e1 ). (5.18)
At Step 2 we show that
sgn (stack ) = sgn (stack + e1 ) = +1, (5.19)
hence sgn () = +1, 0 .

Step 1. Observe that any configuration 0 is determined by the position of its

diagonal dimers. Let us call an elementary move the change of to 0 , where a diagonal
dimer {(j, k), (j + 1, k + 1)} is shifted to {(j + 2, k), (j + 3, k + 1)}. Assume that in the
intermediate vertices (j + 1, k) and (j + 2, k + 1) are not covered by diagonal dimers. Then
the superposition 0 consists of trivial contours and one nontrivial contour of the length
6, which is positive as shown in Fig. 3 (a) (note that there are always three arrows opposite
to any direction of movement along that contour), hence sgn () = sgn ( 0 ). If exactly one
of the vertices (j + 1, k), (j + 2, k + 1) is covered by a diagonal dimer, then the superposition
0 consists of trivial contours and one nontrivial contour of the length m + 2, m + 1 1
(mod 2) of whose arrows are oriented from left to right, as shown in Fig. 3 (b), hence again
sgn () = sgn ( 0 ). Finally, if both vertices (j + 1, k), (j + 2, k + 1) are covered by diagonal
dimers, then sgn () = sgn ( 0 ), because we can consider a clockwise sequence of elementary
moves from to 0 , without intermediate vertices covered by diagonal dimers. See Fig. 3
(c). Thus, for any elementary move 0 we have that
sgn ( 0 ) = sgn (). (5.20)
12 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

If 00 , then by elementary moves we can first move the diagonal dimer at the horizon-
tal line y = 0 to the position {(0, 0), (1, 1)}, and then inductively at horizontal line y = k,
k = 1, 2, . . . , n 1, to the position {(0, k), (1, k + 1)}, forming a stack of diagonal dimers
above the dimer at the horizontal line x = 0. In other words, we have moved to stack . If
10 , then using the above argument we move it to stack + e1 . Hence, formula (5.17) holds.

Step 2. As shown in Fig. 4 (b), the superposition stack st consists of trivial contours
and exactly one nontrivial contour which is a zigzag path between the vertical lines x = 0
and x = 1. The latter one is positive, since there are 2n 1 1 (mod 2) arrows in the
direction of movement from top to bottom and hence sgn (stack ) = +1.
Similarly, the superposition (stack + e1 ) (st + e1 ) consists of trivial contours and a zigzag
contour with n + 1 0 (mod 2) arrow in the direction of movement from top to bottom, and
so sgn ((stack +e1 )(st +e1 )) = 1. In addition, the sign of the superposition st (st +e1 )
is (1) as well. Indeed, the superposition st (st + e1 ) consists of n horizontal loops of
the length m. Since m is even, the sign of each loop is (1), and since n is odd, the sign of
the superposition st (st + e1 ) is (1) as well. Now, since the sign of the superpositions
(stack +e1 )(st +e1 ) and st (st +e1 ) is (1), we obtain that sgn ((stack +e1 )st ) = +1,
hence sgn (stack + e1 ) = +1.
Lemma 5.3 is proven. 

(a) 0 (b) 0 (c) 0

Figure 3. Examples of different types of the superpositions 0 on a 4 5 lattice.

6. Negativity of Pf A1
We begin with the following case:
Lemma 6.1. Let zh , zv , zd > 0. Then if either n 2 (mod 4) or m 2 (mod 4), then
Pf A1 < 0.
Proof. First assume n 2 (mod 4) and consider the case zh > 0, zv = 0, and zd > 0. By
Proposition 3.1 (2), in this case det A1 > 0, det A2 = 0, hence
Z 00 Z 10 = Z 01 Z 11 = Pf A1 = 2Z 01 2Z 11 < 0. (6.1)
By continuity, Pf A1 < 0 for the chosen zh > 0, zd > 0, and small zv > 0. This proves that
Pf A1 < 0 in the whole region zh , zv , zd > 0.
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 13

(a) stack 0 (b) stack st

Figure 4. The stack configuration and its superposition with the standard
configuration on a 4 5 lattice.

Now assume m 2 (mod 4) and consider the case zh = 0, zv > 0, and zd > 0. By
Proposition 3.1 (3), in this case det A1 > 0, det A3 = 0, hence
Z 00 Z 01 = Z 10 Z 11 = Pf A1 = 2Z 10 2Z 11 < 0. (6.2)
By continuity, Pf A1 < 0 for the chosen zv > 0, zd > 0, and small zh > 0. This proves that
Pf A1 < 0 in the whole region zh , zv , zd > 0. 
The case m 0 (mod 4), n 1 (mod 2), is more difficult. To deal with it, we consider
the asymptotic behavior of Pf A1 for zh = 1, zv = 0, as zd 0, and prove the following
lemma:
Lemma 6.2. Let m 0 (mod 4), n 1 (mod 2), and zh = 1, zv = 0. Then as zd 0,
 m n
Pf A1 = 2 zdn (1 + O(zd )). (6.3)
2
Remark. This will imply that Pf A1 < 0 for zh = 1, zv = 0 and sufficiently small zd , and
hence Pf A1 < 0 for all zh , zv , zd > 0. Observe that comparing (6.3) and (5.3), we have that
Pf A1 = Pf A2 (1 + O(zd )). (6.4)
Proof. We have that X
Pf A1 = sgn ()w(). (6.5)
m,n

We will assume in this proof that zv = 0.

As in Lemma 5.2, we have that
X
sgn ()w() = 0, (6.6)
S
kZn (k)

hence to get the lowest excited states without cancellations, we have to consider 0 .
The proof of Lemma 6.2 is based on the following lemma:
Lemma 6.3. For any 0 ,
sgn () = 1. (6.7)
We now conclude as in the end of Lemma 5.2. 
14 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

Proof of Lemma 6.3. The proof is identical to the one of Lemma 5.3. Namely, we have that
sgn () = sgn (stack ), 00 , (6.8)
and
sgn () = sgn (stack + e1 ), 10 . (6.9)
The only important change is that
sgn (stack ) = sgn (stack + e1 ) = 1. (6.10)
Indeed, the superposition stack st consists of trivial contours and one nontrivial contour
which is a zigzag path between the vertical lines x = 0 and x = 1, similar to Fig. 4 (b) except
that the diagonal dimer {(0, n 1), (1, 0)} has the opposite orientation, hence sgn (stack ) =
1.
The superposition (stack + e1 ) (st + e1 ) consists of trivial contours and a zigzag con-
tour with n 1 (mod 2) arrow in the direction of movement from top to bottom, and so
sgn ((stack + e1 ) (st + e1 )) = +1. In addition, the sign of the superposition st (st + e1 )
is (1). Indeed, the superposition st (st + e1 ) consists of n horizontal loops of the
length m. Since m is even, the sign of each loop is (1), and since n is odd, the sign
of the superposition st (st + e1 ) is (1). Now, since the sign of the superposition
(stack + e1 ) (st + e1 ) is (+1), and the sign of the superposition st (st + e1 ) is (1), we
have that sgn ((stack + e1 ) st ) = 1. Hence, sgn (stack + e1 ) = 1. 
The final case m 0 (mod 4), n 0 (mod 4) is the most difficult yet. We consider the
asymptotic behavior of Pf A1 for zh = 1, 0 < zv zd2 , zd 0, and we prove the following
theorem upon establishing four lemmas:
Theorem 6.4. Let m 0 (mod 4), n 0 (mod 4), zh = 1, and 0 < zv zd2 . Then as
zd 0,  m n
Pf A1 = n2 zv2 zdn2 (1 + O (zd )) .
2
Remark. This will imply that Pf A1 < 0 for zh = 1, 0 < zv zd2 , and sufficiently small zd ,
and hence Pf A1 < 0 for all zh , zv , zd > 0.
First note that for the same reasons as in the beginning discussion of the proof of Lemma
5.2, configurations with only horizontal dimers on some line y = k cancel each other in
the Pfaffian, so we next look at 0 defined as in (5.7). However, since n 0 (mod 4),
configurations in 0 now cancel each other completely. Indeed, we can decompose 0 as
0 = 10 t 20 , (6.11)
where 10 and 20 can be further decomposed as
1
G
10 = 1,i
0 , 1,i
0 = { 0 | Ndiag (; k, k + 1) = 1, jk i (mod 2), k Zn },
i=0
1
(6.12)
G
20 2,i 2,i

= 0 , 0 = 0 | Ndiag (; 2k + i, 2k + i + 1) = 2, k Z n2 .
i=0

Here Ndiag (; k, k + 1) is the number of diagonal dimers in connecting horizontal line y = k

to horizontal line y = k + 1, and the set of diagonal dimers for every configuration 10
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 15

is {(jk , k), (jk + 1, k + 1)}, k Zn . As before, equation (5.9) holds since all other vertices
must be covered by horizontal dimers.
On the other hand, configurations in 2,i 0 have the following structure: for every k Z 2
n

and for i = 0 or i = 1 there is one diagonal dimer which connects a vertex x1 = (j1 , 2k + i) to
y1 = (j1 + 1, 2k + i + 1) and another diagonal dimer which connects a vertex x2 = (j2 , 2k + i)
to y2 = (j2 + 1, 2k + i + 1). Note that j2 j1 1 (mod 2), since all other vertices are covered
by horizontal dimers. We have the following lemma:

n
Lemma 6.5. |10 | = |20 | = 2 m2 and for every 0 ,
(
1, if 10 ,
sgn () =
+1, if 20 .
X
Hence, sgn ()w() = 0.
0

n
Proof. The counting argument for |10 | = 2 m2 is the same as in the end of the proof of
Lemma 5.2, and the proof that for every 10 , sgn () = 1 is identical to the proof of
Lemma 6.3. Namely, we have that
sgn () = sgn (stack ), 1,0
0 , (6.13)

sgn () = sgn (stack + e1 ), 01,1 , (6.14)

and
sgn (stack ) = sgn (stack + e1 ) = 1. (6.15)
The only important change is that now the signs of the superpositions (stack + e1 ) (st + e1 )
and st (st + e1 ) are (1) and (+1), respectively. Indeed, the superposition (stack + e1 )
(st + e1 ) consists of trivial contours and a zigzag contour with n 0 (mod 2) arrow in the
direction of movement from top to bottom, and so sgn ((stack + e1 ) (st + e1 )) = +1. In
addition, the superposition st (st +e1 ) consists of n horizontal loops of the length m. Since
m is even, the sign of each loop is (1), and since n is even, the sign of the superposition
st (st + e1 ) is (+1). Since the sign of the superposition (stack + e1 ) (st + e1 ) is (1), and
the sign of the superposition st (st +e1 ) is (+1), we have that sgn ((stack +e1 )st ) = 1.
Hence, sgn (stack + e1 ) = 1.

n
To show that |20 | = 2 m2 , we count as follows: choose either i = 0 or i = 1. Now for
each k Z n2 , there are m choices for the first diagonal dimer and m2 choices for the second,
but the diagonal dimers are equivalent, so we divide this resulting number by 2.
Fix i = 0 and let 2,0 0 be arbitrary. Consider, again, the elementary move described
in Step 1 of the proof of Lemma 5.3, that is, the change of to 0 where a diagonal dimer
{(j, 2k), (j + 1, 2k + 1)} is shifted to {(j + 2, 2k), (j + 3, 2k + 1)} for some k Z n2 . Assume
that in the intermediate vertices (j + 1, 2k) and (j + 2, 2k + 1) are not occupied by a
diagonal dimer. Then the superposition 0 consists of trivial contours and one nontrivial
contour of the length 6 comparable to that in Fig. 3 (a), which is positive, hence sgn () =
sgn ( 0 ). If the intermediate vertices are occupied by a diagonal dimer, simply apply the
clockwise sequence of elementary moves m2 1 times as before. Hence, for every , 0
2,0 0
0 , sgn () = sgn ( ). Now fix a configuration 0,0 0
2,0
with diagonal dimer pairs
16 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

Figure 5. 0,0 st

{(0, 2k), (1, 2k + 1), (1, 2k), (2, 2k + 1)}kZ n and all other dimers horizontal. Then 0,0 st
2
consists of trivial contours and n2 0 (mod 2) identical nontrivial contours of length m + 2
as shown in Fig. 5, hence sgn () = sgn (st ). This shows that for any 02,0 , sgn () = +1.
The proof in the case of i = 1 is entirely analogous. Since 10 and 20 have the same cardinality
but have configurations with opposite signs, they cancel each other and the claim follows.

Let us consider then the set of configurations
1 = { | Nd,v (, k) = 2, k Zn ; Ndiag () = n 1, Nvert () = 1},
where Nd,v (, k) is the total number of vertices covered by diagonal and vertical dimers in
on the line y = k, and Nvert () is the number of vertical dimers in . However, by the
following lemma this set is empty:
Lemma 6.6. 1 = .
Proof. Assume without loss of generality that the vertical dimer is {(0, n 1), (0, 0)} and let
{(jk , k), (jk + 1, k + 1)}kZn1 denote the set of diagonal dimers. Since we require all other
vertices to be covered by horizontal dimers, we have that
(1) jk + 1 jk+1 1 (mod 2), k Zn2 jk+1 jk 0 (mod 2), k Zn2 ,
(2) j0 0 1 (mod 2) j0 1 (mod 2),
(3) jn2 + 1 0 1 (mod 2) jn2 0 (mod 2),
where (1) is the requirement that diagonal vertices within a given horizontal line be odd
spacing apart, (2) is the requirement that (0, 0), the top vertex of the vertical dimer, be odd
spacing from (j0 , 0), the bottom vertex of the diagonal dimer {(j0 , 0), (j0 + 1, 1)}, and (3) is
the requirement that (0, n1), the bottom vertex of the vertical dimer, be odd spacing apart
from (jn2 + 1, n 1), the top vertex of the diagonal dimer {(jn2 , n 2), (jn2 + 1, n 1)}.
From (1) it follows that
jn2 jn3 + jn3 jn4 + jn4 . . . + j3 j2 + j2 j1 + j1 j0 0 (mod 2).
Hence, jn2 j0 (mod 2). Combining this and (2) gives us that jn2 1 (mod 2), which
contradicts (3). 
Since 1 is empty, let us consider
2 = { | Nd,v (, k) = 2, k Zn ; Ndiag () = n 2, Nvert () = 2}.
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 17

Analogous to the decomposition of 0 above, let us write

2 = 12 t 22 , (6.16)
where
12 = { 2 | Nd,v (; k, k + 1) = 1, k Zn },
1
G (6.17)
22 = 2,i
2 , 2,i
2 = { 2 | Nd,v (; 2k + i, 2k + i + 1) = 2, k Z 2 },
n

i=0

i.e. 12 and 22 consist respectively of configurations whose vertical and diagonal dimers
connect all horizontal lines to one another and of configurations whose vertical and diagonal
dimers are pairwise placed on horizontal lines y = 2k + i and y = 2k + 1 + i, k Z n2 for i = 0
or 1. We again require that every pair of vertices of diagonal and/or vertical dimers in a
given horizontal line be an odd spacing apart since the remaining dimers must be horizontal.
See Fig. 6 (a), (b), and (c) for examples of configurations 12 , 2,0
2 , and 2 ,
2,1

respectively.

(a) 12 (b) 22,0 (c) 22,1

Figure 6. Examples of configurations from 2 .

We have the following lemma:

n
Lemma 6.7. |12 | = n(n 1) m2 , and sgn () = 1 for every 12 .
Proof. There are n2 choices of horizontal lines for the vertical dimers, m choices to place


the first vertical dimer within its horizontal line and m2 choices to place

n each of the second
vertical dimer and all n 2 diagonal dimers, hence |12 | = n(n 1) m2 .
Let us prove that sgn () = 1 for every 12 . We have seen in the proof of Lemma
5.3 that the sign of a configuration is invariant with respect to elementary moves, horizontal
shifts of diagonal dimers by two units to the right. The same reasoning applies to vertical
dimers as well.
In 12 define an elementary swap that interchanges a diagonal dimer {(j, k), (j +
1, k + 1)} with a vertical dimer {(j + 1, k 1), (j + 1, k)} above it as follows: increase the
j-coordinate of every vertex of every dimer along the horizontal line y = k, k Zn , to
produce a new configuration 0 in which the vertical dimer is now above the diagonal dimer.
The superposition 0 consists of trivial contours and one nontrivial contour of length
m + 2, m 1 1 (mod 2) of whose arrows are in the direction of movement from left to
right if j 1 (mod 2) and m + 1 1 (mod 2) of whose arrows are in the direction of
18 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

movement from left to right if j 0 (mod 2), the former case exemplified in Fig. 7 with
j = 1, k = 2, m = 4. Hence, for every , 0 12 , sgn () = sgn ( 0 ). In other words, for
any elementary move (horizontal shifts to the right by two or swaps) 0 we have that
formula (5.20) holds.
We now show that any configuration 12 can be moved to a configuration stack in
which vertical dimers are placed at positions {(0, 0), (0, 1)} and {(1, 1), (1, 2)}, respectively,
whereas diagonal dimers form a stack above these two vertical dimers and between lines
x = 0 and x = 1. The remaining vertices are occupied by horizontal dimers. See Fig. 8 (a).
Fix 12 and assume that vertical dimers are at positions {(j1 , k1 ), (j1 , k1 + 1)} and
{(j2 , k2 ), (j2 , k2 + 1)}, respectively. Note that

j2 j1 1 (mod 2). (6.18)

For the sake of contradiction, let us assume that j2 j1 0 (mod 2). If these two vertical
dimers are not on the same horizontal line y = k, apply sequence of swaps on one of them
until k1 = k2 . Note that in order to swap a diagonal and vertical dimer {(j, k), (j, k + 1)}
one vertex of a diagonal dimer has to be on the vertical line x = j and the other vertex on
a horizontal line y = k. If this were not the case, horizontally shift diagonal dimer until this
property is satisfied. Therefore, let us assume that after swaps vertical dimers are positioned
at {(j1 , k), (j1 , k + 1)} and {(j2 , k), (j2 , k + 1)}, respectively. Now, j2 j1 0 (mod 2)
implies that along horizontal line y = k there is an odd number of vertices between (j1 , k)
and (j2 , k), but since all other vertices on y = k have to be covered by horizontal dimers we
have a contradiction. Therefore, equation (6.18) holds. Further, note that this means that
two vertical dimers cannot be placed along the same vertical line x = j.
In other words, we may assume that j1 0 (mod 2) and j2 1 (mod 2). If for a vertical
dimer {(j1 , k1 ), (j1 , k1 + 1)}, k1 6= 0, then use swaps and horizontal shifts of diagonal dimers
where necessary until this dimer is at line y = 0. In a similar fashion, apply swaps with
necessary shifts of diagonal dimers until vertical dimer {(j2 , k2 ), (j2 , k2 + 1)} is at line y = 1.
Now, shift these two vertical dimers horizontally until they are positioned at {(0, 0), (0, 1)}
and {(1, 1), (1, 2)}, respectively. If it so happens that after the above elementary moves they
are positioned at {(0, 1), (0, 2)} and {(1, 0), (1, 1)}, respectively, apply swaps until they are in
the desired position. Similarly to the proof of Lemma 5.3, by horizontal shifts by two to the
right, we can first move the diagonal dimer at horizontal line y = 2 to position {(0, 2), (1, 3)}
and then inductively each diagonal dimer on each of the lines y = k, k = 3, 4, . . . , n 1,
to position {(0, k), (1, k + 1)}, forming a stack of diagonal dimers above the vertical dimers.
Hence, we have obtained a configuration stack .
In the superposition stack st , there are trivial contours and one nontrivial contour of
length 2n, 2n 2 0 (mod 2) of whose arrows are in the direction of movement from top
to bottom as shown in Fig. 7 (b). Hence, sgn (stack ) = 1 and the claim follows.


n
Lemma 6.8. |22 | = n m2 , and sgn () = 1 for every 22 .

Proof. There are n choices of horizontal lines within which to place the lower vertices of the
two vertical dimers, and within this horizontal line there are m choices to place the first
vertical dimer and m2 choices to place the second, but the dimers are equivalent so we divide
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 19

Figure 7. Example of a type of superposition on a 4 4 lattice.

(a) stack 12 (b) stack st

Figure 8. The stack configuration stack 12 and its superposition with the
standard configuration st .

m 2
choices in placing each of the n2 1


this number by 2. Subsequently, there are likewise 2
n
diagonal dimer pairs, hence |22 | = n m2 .

For any given configuration 22 , by applying a sequence of elementary moves, i.e.

horizontal shifts of diagonal and vertical dimers by two units to the right, we can position
all vertical and diagonal dimers between the vertical lines x = 0 and x = 1 to obtain a
configuration stack . See Fig. 9 (a). Since the elementary moves do not change the sign of
configuration, we obtain that sgn () = sgn (stack ).
The superposition stack st contains trivial contours, one nontrivial positive square-
shaped contour, and n2 1 1 (mod 2) negative contours as shown in Fig. 9 (b). Hence,
sgn (stack ) = 1.


Proof of Theorem 6.4 We have that

X
Pf A1 = sgn ()w(). (6.19)

20 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

(a) 22 (b) stack st

Figure 9. Examples of types of superpositions on a 4 4 lattice.

We assume here that zh = 1, zv zd2 , and zd is small. For the same reasons as those given
in Lemma 5.2, we have that
X
sgn ()w() = 0, (6.20)
kZn (k)

hence to get lowest excited states without cancellations we next consider 0 . However, by
Lemma 6.5 X
sgn ()w() = 0, (6.21)
0
so we next turn to 1 , but this is empty by Lemma 6.6. Therefore, we finally consider 2 ,
which, by Lemmas 6.7 and 6.8 yields
X X X
Pf A1 = sgn ()w() = sgn ()w() + sgn ()w()
2 \
 m n X (6.22)
= n2 zv2 zdn2 + sgn ()w(),
2
\

where = 2i=0 i kZn (k). However, any configuration in \ will either have a
greater total number of vertical and diagonal dimers or it will have a weight
zv` 2 n2
z z zd` zv2 zdn2 , 2 ` n 2, (6.23)
zd` v d
from which the claim follows. 

7. Identities Pf A1 = Pf A2 , Pf A3 = Pf A4 for odd n

We prove the following theorem:
Theorem 7.1. Let n 1 (mod 2). Then for all zh , zv , zd 0,
Pf A1 = Pf A2 , Pf A3 = Pf A4 . (7.1)
Proof. Let us prove that
det A1 = det A2 . (7.2)
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 21

Denote  
2s 2t 2s 2t
g(s, t) = zh2 sin2
+ zv2 sin2 + zd2 cos2 + . (7.3)
m n m n
Then by (3.3),
m m
1 1
Y n1
2 Y Y n1
2 Y  1

mn mn
det A1 = 2 g(j, k), det A2 = 2 g j, k + . (7.4)
j=0 k=0 j=0 k=0
2
Let n = 2n0 + 1.
Lemma 7.2. We have the identity,
 
1
g s, t + n0 + = g(s, t). (7.5)
2
Proof. Note that
1 1



2 t + n0 + 2
2 t + n0 + 2 2t 2t
= = + = (mod ). (7.6)
n 2n0 + 1 2n0 + 1 n
Since sin2 (x + ) = sin2 x, cos2 (x + ) = cos2 x, (7.5) follows. 
From (7.5) we obtain that
n0 n0   2n0  
Y Y 1 Y 1
g(j, k) = g j, k + n0 + = g j, k + (7.7)
k=0 k=0
2 k=n
2
0

and
0 1
nY   0 1
nY 2n0
1 Y
g j, k + = g(j, k + n0 + 1) = g(j, k). (7.8)
k=0
2 k=0 k=n0 +1
Hence,
2n0 2n0  
Y Y 1
g(j, k) = g j, k + (7.9)
k=0 k=0
2
and m m
2
1
2n0 2
12n0  
mn
YY
mn
YY 1
det A1 = 2 g(j, k) = 2 g j, k + = det A2 . (7.10)
j=0 k=0 j=0 k=0
2
Thus, (7.2) is proved.
From (7.2) it follows that
(Pf A1 )2 = (Pf A2 )2 . (7.11)
Since we have proved that Pf A1 < 0 and Pf A2 > 0, we conclude that Pf A1 = Pf A2 as
claimed.
Similar to (7.10), we obtain that
m m
2
1 2n0   2
1 2n0  
mn
Y Y 1 mn
Y Y 1 1
det A3 = 2 g j + ,k = 2 g j + ,k + = det A4 . (7.12)
j=0 k=0
2 j=0 k=0
2 2
Since Pf A3 > 0, Pf A4 > 0, we conclude that Pf A3 = Pf A4 . Theorem 7.1 is proved. 
22 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

Numeric data for Pfaffians Pf Ai in Appendix D illustrate the identities Pf A2 = Pf A2

and Pf A4 = Pf A3 on the 4 3 lattice, while disprove such identities on the 4 4 lattice.

8. Poisson Summation Formula and Asymptotics of The Partition Function

As an application of the Pfaffian Sign Theorem, we prove the following theorem:
Theorem 8.1. Suppose that m, n in such a way that
m
C1 C2 (8.1)
n
for some positive constants C2 > C1 . Then for some c > 0,
1
Z = 2e 2 mnF 1 + O ec(m+n) ,



(8.2)
where
Z1 Z1
F = ln 2 + f (x, y) dx dy, (8.3)
0 0
and
1  2 2
ln zh sin (2x) + zv2 sin2 (2y) + zd2 cos2 (2x + 2y) .

f (x, y) = (8.4)
2
Proof. By the double product formula,
m1 n1
1 1 XX j + i k + i
ln det Ai = ln 2 + f (xj , yk ), xj = , yk = , (8.5)
mn mn j=0 k=0 m n
with
1  2 2
ln zh sin (2x) + zv2 sin2 (2y) + zd2 cos2 (2x + 2y) .

f (x, y) = (8.6)
2
Note that in (8.5) we were able to change the upper limit in the first sum from m2 1 to
m 1 due to the symmetries of sin2 x and cos2 x functions. The sum in (8.5) is a Riemann
sum and we evaluate its asymptotics by the Poisson summation formula. To that end, since
f (x + 12 , y) = f (x, y) and f (x, y + 12 ) = f (x, y), we expand f (x, y) into Fourier series
X
f (x, y) = a(p, q)e2i(px+qy) , (8.7)
(p,q)Z2

where
Z1 Z1
a(p, q) = f (x, y)e2i(px+qy) dx dy. (8.8)
0 0
Then
m1 n1 X a(p, q) m1 n1
1 XX XX
f (xj , yk ) = e2i(pxj +qyk ) . (8.9)
mn j=0 k=0 2
mn j=0 k=0
(p,q)Z
Note that
m1 m1
(
1 X 2ipxj 2ip mi 1
X j i 0, p 6 0 (mod m)
e =e e2ip m = e2ip m , (8.10)
m j=0 m j=0 1, p 0 (mod m)
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 23

and similarly,
n1 n1
(
1 X 2iqyk i 1 X k i 0, q 6 0 (mod n)
e = e2iq n e2iq n = e2iq n . (8.11)
n k=0 n k=0 1, q 0 (mod n)
By substituting equations (8.10) and (8.11) into equation (8.9) we obtain the Poisson sum-
mation formula
m1 n1
1 XX X
f (xj , yk ) = a(ms, nt)e2i(si +ti )
mn j=0 k=0
(s,t)Z2
X0 (8.12)
2i(si +ti )
= a(0, 0) + a(ms, nt)e .
(s,t)Z2

Figure 10. Contour of integration

We want to estimate equation (8.8). To that end assume that p > 0 and q 0 and because
f (x, y) is real analytic in x, we can integrate over the contour in Fig. 10 for some > 0. By
the Cauchy integral formula and the periodicity of f (x, y), it follows that
Z1 Z1
a(p, q) = f (x, y)e2i(px+qy) dx dy
0 0
Z1 Z1 (8.13)
2p 2i(px+qy)
=e f (x i, y)e dx dy
0 0
2p


=O e .
If p < 0, then use the contour above reflected across the real axis to get
a(p, q) = O e2|p| .


(8.14)
We can likewise assume p 0 and |q| > 0 and perform the same reasoning with respect to
y to obtain
a(p, q) = O e2(|p|+|q|) .


(8.15)
24 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

From this equation and equation (8.12), we conclude that

m1 n1
1 XX
f (xj , yk ) = a(0, 0) + r(m, n), (8.16)
mn j=0 k=0
where X0
|r(m, n)| C e2(m|s|+n|t|) C0 e2(m+n) , C0 > 0. (8.17)
(s,t)Z2

Returning to equation (8.5) we have that

Z1 Z1
1
f (x, y) dx dy + O e2(m+n) .


ln det Ai = ln 2 + (8.18)
mn
0 0

From (8.18) it follows that for each i = 1, 2, 3, 4, we can write

det Ai = emnF 1 + O ec(m+n) .



(8.19)
Finally, from equations (1.3) and (1.5) the claim follows. 

9. Conclusion
In this paper we establish the sign of the Pfaffians Pf Ai in the dimer model on the
triangular lattice on the torus. We prove that Pf A1 < 0, while Pf Ai > 0 for i = 2, 3, 4. Our
proof is based on the Kasteleyns expressions for Pf Ai in terms of restricted partition sums
Z rs and on low weight expansions. As an application, we obtain an asymptotic expansion
of the partition function in the limit as m, n .
As shown by Galluccio and Loebl [3], Tesler [12], and Cimasoni and Reshetikhin [1], the
dimer models on an orientable Riemann surface of genus g is expressed as a linear algebraic
combination of 22g Pfaffians. It is extended to non-orientable surfaces in the work of Tesler
[12]. The Pfaffian Sign Theorem in this general setting is an interesting open problem.

Appendix A. Proof of Lemma 2.1

We have that
r
G
0
= i . (A.1)
i=1
Since each vertex in i is occupied by a dimer either from or 0 , and the dimers in 0
alternate, we conclude that each i is of even length.
By (2.13),
X
Pf A = sgn ()w(). (A.2)

If we enumerate the vertices on m,n , i.e. permute the set of vertices Vm,n , then by the
well-known fact (see e.g. [4]) that for an arbitrary matrix P of order mn mn,
Pf (P AP T ) = det(P )Pf (A), (A.3)
we get
Pf (A) = (1) Pf (A), (A.4)
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 25

where is some permutation on Vm,n . Here (A) denotes a matrix A whose rows and columns
have been permuted by . In other words,
X X
Pf A = sgn ()w() = (1) [sgn ()] w(), (A.5)

where [sgn ()] indicates the sign of with respect to some new enumeration of vertices.
From (A.5), we have that
sgn () = (1) [sgn ()] . (A.6)
If we take any two configurations and 0 , then (A.6) implies that
sgn () sgn ( 0 ) = [sgn ()] [sgn ( 0 )] , (A.7)
i.e. the sign of 0 is invariant under any renumeration of vertices.
Let
 
1 2 n1 n1 + 1 n1 + n2 n1 + . . . + nr1 + nr
= , (A.8)
v1,1 v1,2 v1,n1 v2,1 v2,n2 vr,nr
where vj,k {1, 2, . . . , mn} denotes the k-th vertex of the j-th contour. Note that is a
renumeration of vertices so that along each i they are rearranged in a cyclical order, starting
from one contour and continuing to the next one. Now, the underlying permutations ()
and ( 0 ) with respect to are then:
[()] = Id, (A.9)
Yr
0
[( )] = C(i ), (A.10)
i=1

where  
v vi,ni
C(i ) = i,1 . (A.11)
vi,2 vi,1
From this equation, equation (A.7), and the fact that each i corresponds to a cycle of even
length, Lemma 2.1 follows.

Appendix B. Proof of Kasteleyn identities (3.2)

From the general results of Galluccio and Loebl [3], Tesler [12], and Cimasoni and Reshetikhin
[1], it follows that
Pf Ai = Z 00 Z 10 Z 01 Z 11 , i = 1, 2, 3, 4. (B.1)
The sign is determined by the sign of the superposition of the standard configuration st
with any configuration from the given Z2 homology class. To evaluate the sign, we consider
four configurations, we call them model configurations, one from each homology class, and
evaluate their signs in Pf Ai .
(1) As a model configuration for Z 00 , we take st . Then the superposition st st
consists of trivial contours, and it has the (+1) sign for each Kasteleyn orientation
Oi , i = 1, 2, 3, 4. This implies that the sign at Z 00 in (B.1) is (+1) for all i = 1, 2, 3, 4.
26 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

(2) As a model configuration for Z 10 , we take the configuration 10 , which coincides with
the standard configuration st everywhere except for the line y = 0 where it is of the
form 10 = st +e1 . Note that the superposition 10 st has one nontrivial horizontal
contour of the even length m. See Fig. 11(a). For the orientations O1 and O2 , the
sign of 10 st is (1), because there are m 0 (mod 2) arrows oriented from left to
right. For the orientations O3 and O4 , the sign is (+1), because there are m 1 1
(mod 2) arrow oriented from left to right.
(3) As a model configuration for Z 01 , we take the configuration 01 , which coincides with
the standard configuration st everywhere except for the diagonal dimers h(0, k), (1, k+
1)i, k = 0, 1, . . . , n1. Note that the superposition 01 st has one nontrivial zigzag
contour of the length 2n. See Fig. 11(b). For the orientations O1 and O3 , the sign
of 10 st is (1), since there are 2n 0 (mod 2) arrows oriented down along the
contour. For the orientations O2 and O4 , the sign is (+1) since there are 2n 1 1
(mod 2) arrow oriented down.
(4) As a model configuration for Z 11 , we take the configuration 11 , which coincides with
the standard configuration st except for
(a) horizontal dimers h(2l 1, 1), (2l, n 1)i, 2 l m/2 ;
(b) vertical dimers h(1, 1), (1, 2)i and h(n 1, 0), (0, 0)i;
(c) diagonal dimers h(1, 0), (2, 1)i and h(0, j), (1, j + 1)i, 2 j n 1.
See Fig. 11(c) and Fig. 12. Observe that for the orientations O1 and O4 , the sign of
10 st is (1), while for O2 and O3 the sign is (+1).
This proves Kasteleyn identities (3.2).

(a) 10 st (b) 01 st (c) 11 st

Figure 11. Oriented contours of the superposition of the model configura-

tions and the standard one, with respect to the periodic-periodic orientation
O1 .

Appendix C. Block Diagonalization of the Kasteleyn Matrices and the

Double Product Formulae for det Ai , i = 1, 2, 3, 4
In this Appendix we diagonalize the Kasteleyn matrices Ai and derive the double product
formulae for det Ai , i = 1, 2, 3, 4. We write
m
j = 2j0 + r, 0 j0 m0 1, m0 = , r = 0, 1, (C.1)
2
and we represent a vertex x as
x = (j0 , k, r) Zm0 Zn Z2 . (C.2)
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 27

Figure 12. Contours of 11 st for m = 4 and n = 5.

Then the matrix elements of the Kasteleyn matrix AK 1 are written in terms of the Kronecker
-symbol jk as
A1 (x, x0 ) = zh j0 j00 kk0 (r0 r0 1 r1 r0 0 ) + j0 +1,j00 kk0 r1 r0 0 j0 ,j00 +1 kk0 r0 r0 1
 
 
+ zv j0 j00 (k+1,k0 k,k0 +1 )(r0 r0 0 r1 r0 1 )
 (C.3)
+ zd j0 j00 k+1,k0 r0 r0 1 + j0 j00 k,k0 +1 r1 r0 0 + j0 +1,j00 k+1,k0 r1 r0 0

j0 ,j00 +1 k,k0 +1 r0 r0 1 .
It is convenient to write the matrix A1 in terms of tensor products as follows. Denote by In
the identity n n matrix and put
0 1 0 ... 0 0

0 0 1 ... 0 0
0 0 0 ... 0 0

+
Jn = (k+1,k0 )k,k0 Zn =
... .. .. ... .. .. ,
. . . .
0 0 0 ... 0 1
1 0 0 ... 0 0
(C.4)
0 0 0 ...
1 0

1 0 0 ...
0 0
0 1 0 ...
0 0

Jn = (k,k0 +1 )k,k0 Zn =
... .. .. ..
.. ...
. . .
. .
0 0 0 . . . 0 0
0 0 0 ... 1 0
Then (C.3) can be rewritten as follows:
      
0 1 + 0 0 0 1
A1 = zh Im0 In + Jm0 In Jm0 In
1 0 1 0 0 0
    
1 0 0 1
+ zv Im0 (Jn+ Jn ) + zd Im0 Jn+ (C.5)
0 1 0 0
     
0 0 + + 0 0 0 1
+ Im0 Jn + Jm0 Jn Jm0 Jn .
1 0 1 0 0 0
28 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

We want to block-diagonalize the Kasteleyn matrix A1 with 2 2 blocks along the diagonal.
To that end introduce the vectors

1
1

2i 2i
e m0 e n

4i 4i
g = e m0 Rm0 , h =
e n
Rn , (C.6)
.. .
.

.

.
2(m0 1)i 2(n1)i
e m0 e n
and
f,,r = g h er Rmn , Zm0 , Zn , r Z2 , (C.7)
with    
1 2 0
e0 = R , e1 = R2 . (C.8)
0 1

Observe that for Zm0 and Zn , g is an eigenfunction of the matrices Jm 0
and h of
the ones Jn . Namely,
2i 2i
Jm g =e
0
m0 g , Jn h = e n h , Zm0 , Zn . (C.9)
By (C.5), this implies that
A1 f,,0 = 00 f,,0 + 10 f,,1 , A1 f,,1 = 01 f,,0 + 11 f,,1 , (C.10)
with
2 2i 2i 2i 2i

00 = 2izv sin , 01 = zh (1 e m 0 ) + zd e n e m0 n
,
n (C.11)
2i
2i
2i
+ 2i
2
10 = zh (1 + e m0
) + zd e n +e m0 n
, 11 = 2izv sin .
n
From (C.10),
     
f,,0 f,,0 00 01
A1 = , , , = , Zm0 , Zn . (C.12)
f,,1 f,,1 10 11
Observe that , is an anti-Hermitian matrix so that
, = , , (C.13)
and also
, = , , Tr , = 0. (C.14)
The matrix
= (, ,0 ,0 ),0 Zm 0 Z (C.15)
0 ; , n

is a block-diagonal matrix, which is a block-diagonalization of A1 , so that

 2ij m0 1  2ik n1
1
A1 = U U , U = U1 U2 I2 , U1 = e m0
, U2 = e n . (C.16)
j,=0 k,=0

This implies that

m
1
Y n1
2 Y
det A1 = det , . (C.17)
=0 =0
 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 29

It remains to calculate det , . From (C.11) we have that

det , = 00 11 01 10
2  2i 2i 2i 2i

= 4zv2 sin zh (1 e m0 ) + zd e n e m0 n
n
2i 2i 2i
+ 2i
 (C.18)
zh (1 + e m0 ) + zd e n + e m0 n

 
2 2 2 2 2 2 2 2 2 2
= 4zh sin + 4zv sin + 4zd cos + ,
m n m n

hence
m
1
Y n1
2 Y 2 2

2 2

det A1 = 4zh2 sin2
+ 4zv2 sin2 + 4zd2 cos2 + . (C.19)
=0 =0
m n m n

This proves the double product formula for det A1 .

Consider now det A2 . The matrix elements of the matrix A2 are

A2 (x, x0 ) = zh j0 j00 kk0 (r0 r0 1 r1 r0 0 ) + j0 +1,j00 kk0 r1 r0 0 j0 ,j00 +1 kk0 r0 r0 1

 

+ zv j0 j00 (1)k,n1 k+1,k0 (1)k0 ,n1 k,k0 +1 (r0 r0 0 r1 r0 1 )



(C.20)
+ zd j0 j00 (1)k,n1 k+1,k0 r0 r0 1 + j0 j00 (1)k0 ,n1 k,k0 +1 r1 r0 0


+ j0 +1,j00 (1)k,n1 k+1,k0 r1 r0 0 j0 ,j00 +1 (1)k0 ,n1 k,k0 +1 r0 r0 1 ,



and the matrix A2 can be written as

     
0 1 + 0 0 0 1
A2 = zh Im0 In + Jm0 In Jm0 In
1 0 1 0 0 0
    
1 0 0 1
+ zv Im0 (Kn+ Kn ) + zd Im0 Kn+ (C.21)
0 1 0 0
     
0 0 + + 0 0 0 1
+ Im0 Kn + Jm0 Kn Jm0 Kn ,
1 0 1 0 0 0

with

0 1 0 0 ... 0 0 0 0 1
0 ...

0 0 0 1 ... 0 1 0 0
0 0 ...
0 0 0 0 ... 0 0 1 0
0 0 ...

Kn
+

Kn = .. .... ,
.. .. .. =
... .. ..
.. ..
.. . (C.22)

. . . . . . . .
. . .
0 0 0 . . . 0 1 0 0 0 ... 0 0
1 0 0 . . . 0 0 0 0 0 ... 1 0
30 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

Observe that the vectors

1
2i(+ 2
1
)

e n

4i(+ 2 )
1
h+ 1 = e n Rn , Zn , (C.23)
2
..

.

2(n1)i(+ 2 1
)
e n

are eigenvectors of the matrices Kn . Namely,

( 1
2i + 2 )
Kn h+ 1 = e n h+ 1 , Zm0 , Zn . (C.24)
2 2

Consider the vectors

f,+ 1 ,r = g h+ 1 er , Zm0 , Zn , r Z2 . (C.25)
2 2

Then
A2 f,+ 1 ,0 = 00 f,+ 1 ,0 + 10 f,+ 1 ,1 , A2 f,+ 1 ,1 = 01 f,+ 1 ,0 + 11 f,+ 1 ,1 , (C.26)
2 2 2 2 2 2

with
1
2 + 12



2 + 2
00 = 2izv sin , 11 = 2izv sin ,
n n
2i(+ 2
1
) 2i(+ 1
2)

01
2i
= zh (1 e m0 ) + zd e n
2i
e m0
n
, (C.27)

2i ( 1
2i + 2 ) 2i
+
( 1
2i + 2 )


10 = zh (1 + e m0
) + zd e n +e m0 n
.
The matrix
= (, ,0 ,0 ),0 Zm 0 Z (C.28)
0 ; , n

is now a block-diagonalization of A2 with

!n1
 2ij
m0 1 (
2ik + 1
2 )
A2 = U U 1 , U = U1 U2 I2 , U1 = e m0
, U2 = e n . (C.29)
j,=0
k,=0

From here it follows that

m
1 "
Y n1 1
2


Y 2 2 +
det A2 = 4zh2 sin2 + 4zv2 sin2 2

=0 =0
m n
! # (C.30)
2 2 + 12


+ 4zd2 cos2 + .
m n

Let us turn to det A3 .

 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 31

The matrix elements of the matrix A3 are

A3 (x, x0 ) = zh j0 j00 kk0 (r0 r0 1 r1 r0 0 ) + (1)j0 ,m0 1 j0 +1,j00 kk0 r1 r0 0


0 
(1) j0 ,m0 1 j0 ,j00 +1 kk0 r0 r0 1
 
+ zv j0 j00 (k+1,k0 k,k0 +1 )(r0 r0 0 r1 r0 1 ) (C.31)

+ zd j0 j00 k+1,k0 r0 r0 1 + j0 j00 k,k0 +1 r1 r0 0
j 0 ,m
+ (1)j0 ,m0 1 j0 +1,j00 k+1,k0 r1 r0 0 (1)

0 0 1 j0 ,j00 +1 k,k0 +1 r0 r0 1 ,
hence
     i
0 1 0 0 0 1
h
+
A3 = zh Im0 In + Km0 In Km0 In
1 0 1 0 0 0
    
1 0 0 1
+ zv Im0 (Jn+ Jn ) + zd Im0 Jn+ (C.32)
0 1 0 0
     
0 0 + + 0 0 0 1
+ Im0 Jn + Km0 Jn Km0 Jn .
1 0 1 0 0 0
Now, the matrix
= (, ,0 ,0 ),0 Zm 0 Z (C.33)
0 ; , n

is a block-diagonalization of A3 with
!m0 1
( 1
2ij + 2 )  2ik n1
A3 = U U 1 , U = U1 U2 I2 , U1 = e m0
, U2 = e n . (C.34)
k,=0
j,=0

Hence,
m
1 "
Y n1 1
2


Y 2 + 2
det A3 = 4zh2 sin2 2
+ 4zv2 sin2
=0 =0
m n
! # (C.35)
2 + 21


2
+ 4zd2 cos2 + .
m n
The double product formula for det A4 is proved similarly:
m
2
1 n1 " 1

1


Y Y 2 + 2 +
det A4 = 4zh2 sin2 2
+ 4zv2 sin2 2

=0 =0
m n

! # (C.36)
1 1


2 + 2 +
+ 4zd2 cos2 2
+ 2
.
m n

Appendix D. Numerical Data for the Pfaffians Ai

In this Appendix we present numerical data for the Pfaffians Ai on the m n lattices
on the torus for different values m and n. It is interesting to compare these data with
the asymptotics of the Pfaffians Ai , obtained in Sections 5 and 6 above, and also with the
identities Pf A1 = Pf A2 , Pf A3 = Pf A4 for odd n, proven in Section 7.
32 PAVEL BLEHER, BRAD ELWOOD, AND DRAZEN PETROVIC

The Pfaffians Ai for m = 4, n = 3.

Pf A1 = 4zh zd (3zv2 + zd2 )(4zh2 + 3zv2 + 3zd2 ),

Pf A2 = 4zh zd (3zv2 + zd2 )(4zh2 + 3zv2 + 3zd2 ),
Pf A3 = 2(zh2 + zd2 )[(2zh2 + 3zv2 )(2zh2 + 3zv2 + 4zd2 ) + zd4 ],
Pf A4 = 2(zh2 + zd2 )[(2zh2 + 3zv2 )(2zh2 + 3zv2 + 4zd2 ) + zd4 ].

The Pfaffians Ai for m = 4, n = 4.

Pf A1 = 256zh2 zv2 zd2 (zh2 + zv2 + zd2 ),

Pf A2 = 16(zv2 + zd2 )2 (2zh2 + zv2 + zd2 )2 ,
Pf A3 = 16(zv2 + zd2 )2 (zh2 + 2zv2 + zd2 )2 ,
Pf A3 = 16(zv2 + zd2 )2 (zh2 + zv2 + 2zd2 )2 .

The Pfaffians Ai for m = 4, n = 6.

Pf A1 = 16zh2 zd2 (4zh2 + 3zv2 + 3zd2 )2 (3zv2 + zd2 )2 ,

Pf A2 = 16zv2 (4zh2 + zv2 + zd2 )2 (zh2 + zv2 + zd2 )(zv2 + 3zd2 )2 ,
Pf A3 = 4(zd2 + zh2 )2 (zd4 + 4zd2 (2zh2 + 3zv2 ) + (2zh2 + 3zv2 )2 )2 ,
Pf A4 = 4(zd2 + zh2 + 2zv2 )2 (zd4 + 8zd2 zh2 + 4zh4 + 4zd2 zv2 + 4zh2 zv2 + zv4 )2 .

The Pfaffians Ai for m = 4, n = 8.

Pf A1 = 4096zd2 zh2 zv2 (zd2 + zv2 )2 (zd2 + zh2 + zv2 )(zd2 + 2zh2 + zv2 )2 ,
Pf A2 = 16(zd4 + 6zd2 zv2 + zv4 )2 (zd4 + 8zh4 + 8zh2 zv2 + zv4 + 2zd2 (4zh2 + zv2 ))2 ,
Pf A3 = 256(zd2 + zh2 )2 (zh2 + zv2 )2 (2zd2 + zh2 + zv2 )2 (zd2 + zh2 + 2zv2 )2 ,
Pf A4 = 16(zd4 + 2zh4 + 4zh2 zv2 + zv4 + 2zd2 (2zh2 + zv2 ))2 (zd4 + 2zh4 + 4zh2 zv2 + zv4 + zd2 (4zh2 + 6zv2 ))2 .

The Pfaffians Ai for m = 6, n = 6.

Pf A1 = 4zd2 (zd2 + 3zh2 )2 (zd2 + 3zv2 )2 (zd2 + 3(zh2 + zv2 ))2 (4zd2 + 3(zh2 + zv2 ))2 ,
Pf A2 = 4zv2 (3zd2 + zv2 )2 (3zh2 + zv2 )2 (3(zd2 + zh2 ) + zv2 )2 (3(zd2 + zh2 ) + 4zv2 )2 ,
Pf A3 = 4zh2 (3zd2 + zh2 )2 (zh2 + 3zv2 )2 (3zd2 + zh2 + 3zv2 )2 (3zd2 + 4zh2 + 3zv2 )2 ,
Pf A4 = 4(zd2 + zh2 + zv2 )3 (4zd2 + zh2 + zv2 )2 (zd2 + 4zh2 + zv2 )2 (zd2 + zh2 + 4zv2 )2 .

The Pfaffians Ai for m = 8, n = 8.

 PFAFFIAN SIGN THEOREM FOR THE DIMER MODEL ON A TRIANGULAR LATTICE 33

Pf A1 = 1048576zh2 zv2 zd2 (zd2 + zh2 )2 (zd2 + zv2 )2 (zh2 + zv2 )2 (zd2 + zh2 + zv2 )(2zd2 + zh2 + zv2 )2
(zd2 + 2zh2 + zv2 )2 (zd2 + zh2 + 2zv2 )2 ,
Pf A2 = 256(zd4 + 6zd2 zv2 + zv4 )2 (zd4 + 4zd2 zh2 + 2zh4 + 2zd2 zv2 + 4zh2 zv2 + zv4 )2
(zd4 + 4zd2 zh2 + 2zh4 + 6zd2 zv2 + 4zh2 zv2 + zv4 )2 (zd4 + 8zd2 zh2 + 8zh4 + 2zd2 zv2 + 8zh2 zv2 + zv4 )2 ,
Pf A3 = 256(zd4 + 6zd2 zh2 + zh4 )2 (zd4 + 2zd2 zh2 + zh4 + 4zd2 zv2 + 4zh2 zv2 + 2zv4 )2
(zd4 + 6zd2 zh2 + zh4 + 4zd2 zv2 + 4zh2 zv2 + 2zv4 )2 (zd4 + 2zd2 zh2 + zh4 + 8zd2 zv2 + 8zh2 zv2 + 8zv4 )2 ,
Pf A4 = 256(2zd4 + 4zd2 zh2 + zh4 + 4zd2 zv2 + 2zh2 zv2 + zv4 )2 (8zd4 + 8zd2 zh2 + zh4 + 8zd2 zv2 + 2zh2 zv2 + zv4 )2
(zh4 + 6zh2 zv2 + zv4 )2 (2zd4 + 4zd2 zh2 + zh4 + 4zd2 zv2 + 6zh2 zv2 + zv4 )2 .
References
[1] D. Cimasoni, N. Reshetikhin, Dimers on Surface Graphs and Spin Structures. I, Commun. Math. Phys.
275, (2007), 187208
[2] P. Fendley, R. Moessner, S. L. Sondhi, Classical dimers on the triangular lattice, Phys. Rev. B 66
(2002), 214513
[3] A. Galluccio, M. Loebl, On the theory of Pfaffian orientations. I. Perfect matchings and permanents,
Electron. J. Combin. 6, Research Paper 6 (1999), 18 pp. (electronic)
[4] C. D. Godsil, Algebraic Combinatorics, Chapman and Hall, 1993
[5] N. Sh. Izmailian, R Kenna, Dimer model on a triangular lattice, Phys. Rev. E 84 (2011), 021107
[6] R.W. Kenyon, N. Sun, and D.B. Wilson, On the asymptotics of dimers on tori, Probability Theory and
Related Fields 166(3), (2016), 9711023
[7] P. W. Kasteleyn, The statistics of dimers on a lattice. I. The number of dimer arrangements on a
quadratic lattice, Physica 27 (1961), 12091225
[8] P. W. Kasteleyn, Dimer statistics and phase transitions, J. Math. Phys. 4 (1963), 287293
[9] P. W. Kasteleyn, Graph theory and crystal physics, in Graph Theory and Theoretical Physics, Academic
Press, London, 1967
[10] B. M. McCoy, Advanced Statistical Mechanics, Oxford: Oxford University Press, 2010
[11] B. M. McCoy, T. T. Wu, The TwoDimensional Ising Model, Second Edition, Dover Publications Inc.,
Mineola, New York, 2014
[12] G. Tesler, Matchings in graphs on non-orientable surfaces, J. Combin. Theory Ser. B 78 (2000), 198231