Professional Documents
Culture Documents
1209-1225
by P. W. KASTELEYN
Synopsis
The number of ways in which a finite quadratic lattice (with edges or with periodic
boundary conditions) can be fully covered with given numbers of "horizontal" and
"vertical" dimers is rigorously calculated by a combinatorial method involving
Pfaffians. For lattices infinite in one or two dimensions asymptotic expressions for
this number of dimer configurations are derived, and as a n application the entropy
of a mixture of dimers of two different lengths on an infinite rectangular lattice is
calculated. The relation of this combinatorial problem to the Ising problem is brlefly
discussed.
a given volume can be divided into given numbers of single, double, triple
cells etc.; the equivalence of this combinatorial factor with the former one
is obvious.
A two-dimensional form of the arrangement problem is encountered in
the theory of adsorption of diatomic, triatomic etc. molecules on a regular
surface. The empty sites of the surface then play the role of "monomers".
As in many problems of this sort, it is easy to find the most general
solution for a one-dimensional lattice 2), it appears very difficult to find
a more or less general solution for two-dimensional lattices, whereas for
three dimensions any-exact solution seems extremely remote. Therefore one
generally uses approximation methods; we refer to the work of F o w l e r and
R u s h b r oo k e 9 (who also made some rigorous calculations on two- and
three-dimensional infinite strips of finite width), C h a n g , F l o r y , H u g g i n s ,
Miller, G u g g e n h e i m (for detailed references see ref. l ) , O r r 4), R u s h b r o o k e , S c o i n s and W a k e f i e l d s ) , a n d c o h e n , D e B o e r a n d S a l s b u r g 2 ) .
Recently, G r e e n and L eipnik6) claimed to have found a rigorous solution
for the case of monomer-dimer mixtures on a two-dimensional lattice, but
F i s h e r and T e m p e r l e y 7 ) , and K a t s u r a and I n a w a s h i r o 7) proved
that their results were not correct.
In this paper we present a rigorous solution to the above-mentioned
combinatorial problem for a very special case, viz. that of a two-dimensional
q~~adratic
lattice, completely covered with dirners (in terms of graph theory
we ask for the number of "perfect matchings" of the lattice 8)). Both the
absence of monomers and the dimension of the lattice form serious restrictions, but it is hoped that the,present investigation may be useful as a first
step. The situation has some resemblance t o that of the combinatorial
problem connected with the Ising model of cooperative phenomena 9), for
which an exact solution has been given for an equally special case 10-13).
I t will be shown that the two problems are to a certain extent analogous.
In $ 2 and $ 3 we shall develop the method of solution for the case of a finite
lattice imbedded in a plane (i.e. a rectangle with edges), and in 4 for that of
a lattice imbedded in a torus (i.e. with periodic boundary conditions). In
$ 5 an alternative method will be sketched. As an application, the entropy
of a certain mixture of dimers is calculated in 6.
I t is in.tended -to treat in a subsequent paper the statistics of dimers on
other two-dimensional lattices, to discuss boundary effects and to make some
remarks on three-dimensional lattices.
2. The planar quadratic lattice. Consider a planar quadratic m x n
lattice Q,, to which one can attach dimers (figures consisting of two linked
vertices) in such a way that every dimer occupies two lattice points connected by a bond. We indicate the lattice points by (i,j) or p (i = 1 , ...,rn;
j = 1, ..., n ;fi = 1 , ..., mn), the number of "horizontal" dimers (occupying
121 1
Zmn(z,z')
xkz,Nz, g(Nz,N;)
z~~~
z'-V2'
,
(li
where the sum runs over all combinations N2, N ; satisfying 2(Nz N;) =
= mn; if desired, the counting variables z and z' may be viewed as activities
and ,Z
,
as the configurational partition function. At least one of the two
numbers m and n has to be even; let m be even. We shall refer to the arrangement of dimers occupying the pairs of sites pl and $2, $3 and $4, p5 and
p6, etc. as to the config~~ation
C = \+I ; $21 p3; 9 4 I$:, ; Psi ... 1Pmn-1; pmnl.
A simple but important configuration is
which we shall call the standard configuration (fig. la). We could, however,
represent this arrangement of dimers equally well by 12,l; 1,113,l; 4,1/ ...
or by /4,1; 3 , l )1,l ; 2,11 ... etc. To make the representation unique we order
the points of the lattice row after row by choosing the $-numbering as
follows :
( i , j ) ++p = ( j -1) m
z,
(2)
Cb 6 ~ a ( k 1k2)
; a(k3; k4) ... ~(KN-1;AN),
(4)
where the sum runs over those permutations k1, kZ, ..., kN of the numbers
1,2, ... ,iV which obey
+ 1 according
1212
P. W. KASTELEYN
We begin by noting that if we define D($; 9') = 0 for all pairs of sites ( p ; p')
that are not connected by a bond, all terms in the Pfaffian that would not
correspond to a dimer configuration will vanish. Next we put the coefficients
D ( p ; p') corresponding to pairs of sites that are connected by a horizontal
or a vertical bond equal, in absolute magnitude, to z and z', respectively. In
this way we get all configurations represented by a term of the proper weight ;
the conditions (3) and (3') ensure that the correspondence is cne-to-one.
Finally, in order that all configurations are counted positively we have
to choose the signs of the non-zero elements such that the product
D(pl; $4D(P3; p4) ... has the same sign as the parity 6 ~ I .t isevident that
the product corresponding to the standard configuration Cohas to be positive.
Now, from Co one can obtain any arbitrary configuration C in the following
way. We draw a picture of the lattice in which every dimer of the configuration Co is represented by a dotted line and every dimer of the configu-
t - - - . - c - .
(1.2)
c-----.
(1.1) (21)
nc.i,i)
)
1
5
(
(a)
F1g 1
(b)
( a ) The standard configuration Co of the planar lattice Q64
(b) The construct~onof a new configuration C (full llnes) from Co (dotted hnes)
guration C by a full line (fig. l b ) . Since any lattice point is the endpoint
of just one line of each type the resulting figure consists of: a) pairs of sites
connected both by a dotted line and by a full line; b) closed polygons consisting of alternating dotted and full lines (to be called Co-bonds and nonCo-bonds). If we then take the configuration Co and we shift in each of the
"alternating" polygons all dimers clockwise or counter-clockwise by one
step, Co goes over into C.
B y analogy to the Ising problem 9) one might expect that each of the
polygons (representing a cyclic permutation of an even number of lattice
sites) would contribute a factor - 1 t o the parity 6p of the permutation P
corresponding to C. I n fact, this is true , but owing to the restrictions on
1213
1214
P.3V. KASTBLEYN
cyclic shift the opposite sides of a connected figure are shifted in opposite
directions, combined with the alternation of Co-bonds and non-Co-bonds
we conclude that both the numbers of unit squares in a strip and the number
of strips are odd (cf. fig. l b ) . Consequently, each alternating polygon
encircles an odd number of unit squares, and it will be sufficient to choose the
signs of the D ( p ; p') such that among the four bonds bounding a unit
square there is an odd number having a negative D ( p ; p'); we have further
seen that the standard configuration has to appear with a positive sign. This
can be realized e.g. by attributing minus signs to the coefficients of the
vertical bonds between lattice sites of odd i. We thus get the following set
of coefficients
for 1 g z
D(i,j;zf l,j)=+z
D(i, j ; z, j +- 1 ) = (-1)i.d for 1 5 z
( 2 ,j ; 2 , j )
=0
otherwise
s m -
< m,
I, 1 < j < n ,
l l j g n - 1 ,
(6)
The equations (5) and (6) are sufficient to derive the configuration
generating function Z,,(z, 2'). I t should be remarked that throughout this
paper we assume that the dimers are symmetric. If they were asymmetric
all elements D(p; p') would have to be multiplied by 2, since any pair of
sites might then be occupied in two distinguishable ways.
2')
[Pf 0 1 2 = det D,
(7)
where D is the matrix given by (6) together with the requirement of skew
symmetry :
Dti, j ; i',j')
-D(if, j r ; i , j),
(8)
12 15
1-1
1 0 0 .
0 1 0 .
0 - 1 0 1 . .
. . . . .
0 0 0 .
0
0 0 0 .
16
I
Q
v == /
. .
. .
001
001
.
. .
. .
0
0
.
0
0
'
. I J r=
0
o . . ..
O 01
1
0 . . .
001
0 - 1 . . .
0 0
. .
.
. . . .
0
o . . . - 1 0
0
O l ]
0 . . .
[ - -l 1o 0
(10)
166
. . - 101
and the indices indicate the order of the matrices. I t can be verified that
Q , can be diagonalized by a similarity transformation Q , = U;lQ,U,
with
the matrix U , given by
+ I)}$
U n ( l ;1') = { 2 / ( n
U;l(l; 1') = { 2 / ( n
il sin {LL' n / ( n I ) ) ,
1))f (-i)l' sin {11'z/(n 1 ) ) ;
(11)
i.e. the only non-zero elements are grouped in 2 x 2 blocks along the diagonal.
Thus the determinant is readily found:
2i2 cos
det D
det D
IT
k=ll=l
kn
L
n
---
m+ 1
7~
1~
-2iz'COS
--------
n+ 1
2i2' cos n+ 1
kn
-2iZCOSm+ 1
-
(12)
and, from (7) and ( 1 2), we find the following expression for the configuration
generating functioli of the lattice Q,,:
pm +n
nn
z+- n n
2'mn.
k = l 1=1
[ziwn-1)
kn
[2z C O S ~--
+m ~ n - 1 )
[22
k=l
z=l
m+l
kn
2'2
cos2 --- ,
n+l
C O S ~-
m+l
n+l
( n even)
(13)
( n odd)
I n writing down the expression valid for odd n use has been made of the
relation
1216
K-~STELLYA
+ ( 1 + ~2)']m-1 [ U - ( I + ~ ~ z ) ? ] ? n f l
2(1 + 2 . ~ 2 ) +
(14)
iz~
h= 1
i
i
!
t
this identity holds (for even 7%) because the two members represent two
polynomials in c~ of the same degree m, with the same zeros, and with equal
coefficients of the leading term. For numerical calculations it is sometimes
useful to perform, with the aid of (14), the product over k in (13). I n this
way we find
176
cos -n+ 1
+ (1 + [' cos")!]
1n
m+1
n+ 1
1
I
z,(z,
2') =lim
{z,,
( 2 . z1))llm=z,n
I =I
~ + O O
)'I.
(16)
m, n--too
n12
n,2
= exp { n P Zdw
/
0
0
nj2
= zi
exp{z-l/dw In [C cos o
+ ( 1 + CQosb)*j).
(17)
For 12'1 5 121 we may expand the latter integralid in terms of 8, integrate
term b y term, and sum the resulting series, which gives:
+ In z + n - l j dx x-1 arctan x
i
(18)
& In z'
1i c
+ n - l j dx x-I
arctan x.
From the equivalence of the last two expressions (which is easily proved)
and the analogous derivation in the case Iz'j 2 /z/ it follows that either
of them m a y be used for all values of z and 2'. Using the relation
arctan x = (2i)-l[ln(l
ix) - ln(1 - ix)] and introducing the function
1217
1
1
1
+ In z + n-lAa(z'/z)
$ In zJ+ n-lA3(z/~').
Ii
where
1G
= 1 - L 33-2
+ 5-2
- 7-2
'I
I
1
./
I
magnitude, to z and z', respectively; we can still choose the signs of these
elements. Further we have now to distinguish four classes of configurations.
The first class comprises those configurations that can be derived from the
standard configuration (which we take identical to that of 5 2) by cyclic
shifts along polygons not looping the torus either in horizontal or in vertical
direction, or, more generally, looping the torus an even number of times
in both directions; let us call them (e, e) configurations. In an analogous
.. . . . . . . . - . . . . ,
I------------------------
I
I
- 1
L-----------------------1
(o,e) CONFIGURATION
(e,e)CONFIGURATION
-----
I
I
I
I
I
I
. . . .
......
I
I
(e,o) CONFIGURATION
(op) CONFIGURATION
Fig. 2. Configurations from the four configuration classes of the toroidal lattice
QL~~
way we define (0, e), (e, o) and (0, o) configurations (cf. fig. 2 ; e = even,
m,
o = odd, first symbol refers to horizontal loops). If we define, for 1 5 i
1
Ij<%
D(m, j; 1 , j)
+z; D(i, n ; i, 1 )
(- l)iz',
(22.1)
D 3 with D(m, j ; 1 , j)
Dq with D(m, j ; 1 , j)
-2; D(i, n ; i, 1)
(-l)izf,
(22.3)
-2; D(i, n ; i, 1 )
(-l)i+lzl,
(22.4)
1219
(0,
TABLE I
Counting of configurations on a toroidal lattice
Class of configurations
Sign of corresponding
terms in Pf Dl,
DI
Dz
DB
D4
all configurations are counted with the right sign so that we have obtained
the analogue of eq. (5) for a toroidal quadratic lattice.
The evaluation of eq. (23) runs parallel to that of eq. (5),the only difference
being the occurrence of the matrices
.
.
.
, . . . . .
0 0 0 . . .
0 0 0 . . .
1 0 0 . .
-1
0 1 0 . .
0 - 1 0 1 . .
0
1
0-1
0
0
0
0
.
0
-1
-1
and
0
11
01
.
.
.
. . . . . .
0 0 0 . . .
1 0 0 . .
0 1 0 . .
0 - 1 0 1 . .
-1
0 0 0 . .
01'
0 0
0 0
. .
0 1
-101
instead of Q (cf. eq. (9) and (10)). They can be diagonalized successively by
a transformation with the matrices
V(1, 1') = ( 1 In), exp (211' niln),
V-(1, 1') = ( 1 In)+exp(l(21f - 1 )ni/n).
Proceeding as in
Zgi(z, z')
(24)
5 3 we find
{2ln/n)]f
k=l z=1
The first term of the right-hand member is easily seen to be equal to zero.
If desired, this equation can be put into a form analogous to (15) with the
aid of the following identities, valid for even m and non-negative values of u :
im
n 2 [ d +. sin2{2kn/m)]*
[ ~ ~ + +U~)fjim-[(l
LL+
( 1 +z,~2)i]im,
I;=1
)m
n 2[2$ + sinZ((2k
1 ) n/m)]+= [u
(26)
k= l
For m + m, i.e. for infinite cylindrical strips, the second and fourth term
of eq. (25) can be shown to be dominant and equal, and we find
!&%I
=2
12
[t
(21 - 1)n
sin
1- 1
+ ( I + (2
(21 - I ) n
sin2
)I
(27)
which is to be compared with eq. (16), valid for planar strips. In the limit
n -t co we finally obtain eq. (17) again. The values for the molecular
freedom ~2 in cylindrical strips of widths up to 8 calculated by F o w l e r and
R u s h b r o o k e 3) can be found as special cases from eq. (27). IE table I1
we list the various values of plz (exact and approximate) calculated for
strips and for the infinite lattice.
TABLE I1
The molecular freedom qz for planar and toroidal quadratic cc
n
1
2
3
4
5
6
7
8
planar lattice
/
I
1.000
1.618
1.551
toroidal lattice
2.414
1.685
1.754
1.716
1.701
x n
lattices
method
1.686
1.658
1.772
1.732
1.823
1.471
1.687
1.736
1.63
1.8
1.791 623
...
1
i
F l o r y 17)
C h a n g 18)
O r r 4,
M i l l e r 19)
F o w l e r and R u s h b r o o k e
this paper
approximate
3)
1221
We shall not go into details but instead mention another closely related
method for calculating Zmn.
I t follows from the considerations of $ 2 that there is a one-to-one correspondence between alternating polygons on Qmn and closed paths on a
corresponding oriented lattice Q&, sketched in fig. 3. In this lattice any bond
may be traversed in only one direction: the Co-bonds in the direction of
increasing (decreasing) i for odd (even) values of j, the non-Co-bonds in the
direction from the "head" of a Co-bond to the "tail" of another Co-bond.
a(;, j ; i, j)
d
+
+
j ;
1
. .
d(i, 7 ; z - 1, j)
d j i, j
1 )
( 2 , j ; i, j - 1
( i ,j ; i j )
= 1
= y
= y
{ -y'
= 0
for j odd,
for j even,
for i even, j odd,
for i odd, j even,
otherwise,
(28)
i.e. we attach weight factors y and iy' to horizontal and vertical oriented
bonds, respectively, and a factor 1 to each lattice site on its own. Thus in
detd each term will correspond to a configuration of closed paths on Qgn
(cf. ref. 9). A term representing a permutation consisting of v cycles (each
permuting an even number of points) occurs with a factor (- 1)v; since in a
determinant, as contrasted with a Pfaffian, there is no restriction on the
order of the indices of the elements, the difficulty mentioned in 5 2 does not
arise here. The argument of $ 2 shows again that the difference in sign
between permutations consisting of odd and even numbers of cycles is
compensated for by the negative signs attributed to the vertical bonds
between sites with odd values of i. So detd is just the path generating
1222
P. W . KASTELEYN
--
fuliction H,,(y,
det d =
CMC N Ih(i11,
,
&I1)y"ylM'= H m n ( ~Y'),
,
h(M, 14"') being the number of ways of combining &l horizontal and h1'
vertical steps to closed paths. According to 5 2 such a combination may be
considered as representing a configuration of M' vertical dimers, and
hence Smn - M' horizontal dimers. I t follows that
Zmn(z,2')
z'M'
z*mn det(d),
(29)
H ( y , y')
exp{~-2/dw / do' In [ ( l
0
y2)2
(30)
where the path of integration encircles the origin but excludes the singularities of Zmn(l,5). We shall introduce x = N2/amn and x' = Nilimn = 1 - x.
1223
For large m and n we can evaluate (31 ) by the saddle-point method. We find
lim (g(&mnx,$mnxf))l/mn = Y
(32)
711, n-oo
(33)
RANDOM MIXTURE
-INTERLOCKING MIXTURE
Fig. 4. The reduced entropy per dimer of a mixture of horizontal and vertical dimers
as a function of x (fradtion of horizontal dimers).
+ z'z + z2 cos w +
2'2
cos w']).
(34)
The same result was found when periodic boundary conditions were introduced. The effect of boundary conditions is, however, not entirely trivial
and will be discussed in more detail in a subsequent paper.
The right-hand member of (34) has a remarkable resemblance to Onsager's expression for the partition function per spin of a rectangular Ising
system 9 ) l O ) *). A more detailed examination reveals that Z ( z , z') as a
function of [ = z'/z has no singular points on the real positive axis; it
corresponds to Onsager's partition function at the critical point (or critical
line, if the strengths of the horizontal and vertical interactions vary with
respect t o each other).This fact might tempt one to conjecture that the more
general problem of monomer-dimer mixtures would be the analogue of the
Ising problem at arbitrary temperatures, and hence rigorously solvable.
However, this is not the case. I t is easy to see that the true analogue of the
partition function of an Ising system is the generating function H,,(y, y') =
= det d for closed paths on the lattice QgTz;
for an infinite lattice, the function
H ( y , y') given by (30) has a singularity a t y = 1 , which is just the value of
interest for the dimer problem.
t - - - - . - -
0---.
.......
.....
(9)
I. .'I I
..A
(b)
A. . . . . .
I ::I
Fig. 5. The construction of a monomer-dimer configuration from the standard configuration by (a)the omission of bonds and (b) the shift of dimers along chains of bonds.
shc
mi.
lat
abc
Rec
1225
6)
7)
8)
9)
10)
11)
12)
13)
14)
15)
16)
17)
18)
19)
20)