You are on page 1of 7

Ramanujan J (2012) 29:423–429

DOI 10.1007/s11139-012-9387-5

Partitions and the grand canonical ensemble

Freeman Dyson

Received: 17 February 2012 / Accepted: 1 March 2012 / Published online: 31 August 2012
© Springer Science+Business Media, LLC 2012

Abstract Two disconnected remarks about partitions. First, a pedagogical remark


connecting pure mathematics with statistical physics. The grand canonical ensemble
of statistical mechanics is applied to the counting of partitions. This picture borrowed
from physics gives a simple approximation to the exact calculation of the partition
function by Hardy and Ramanujan. Second, an exact formula is guessed for the func-
tion NS (m, n) defined in a recent paper by Andrews, Garvan, and Liang. The formula
was subsequently proved by Garvan. We hope that it may lead to a better understand-
ing of the beautiful new congruence properties of partitions discovered by Andrews.

Keywords Grand canonical ensemble · Statistical physics · Partitions ·


SPT-function

Mathematics Subject Classification 11P83 · 11P84 · 82B10

1 A physical model for partitions

I am delighted to offer this small contribution to Ramanujan’s 125th birthday. I follow


his good example, jumping from one subject to another without any logical connec-
tion. First, I use a physical model to describe the partitions of a large integer n. The
model is the Grand Canonical Ensemble, used to describe physical systems with a
variable number of particles. Each part in the partition is regarded as a particle in
the model. For each integer k, there is a separate species of particle of energy k. The
particles are emitted and absorbed like the photons in a cavity filled with thermal ra-
diation. The states of the entire system are labeled by the numbers [m1 , m2 , m3 , . . . ],

F. Dyson ()
Institute for Advanced Study, Princeton, NJ, USA
e-mail: dyson@ias.edu
424 F. Dyson

where mk is the number of particles of energy k. So each state describes a partition


with sum

n= kmk . (1)
k
In the physical model, n is the energy of the state. The model is in thermal equi-
librium with temperature T , so the probability of the state [mk ] is

f (mk ) = Z −1 exp(−bkmk ), (2)
k

with
  −1 
Z= exp(−bkmk ) = 1 − xk = p(n)x n , (3)
m k k n

x = exp(−b), b = T −1 . (4)

The name “partition function” is given by physicists to the sum Z and by mathe-
maticians to the coefficients p(n). For mathematicians, p(n) is the number of parti-
tions of n. For physicists, the mean value of the energy of the ensemble is
  −1
E = −Z −1 dZ/db = k exp(bk) − 1 , (5)
k

the free energy is


  
F = −b−1 log Z = b−1 log 1 − exp(−bk) , (6)
k

and the entropy is


S = b(E − F ). (7)
In the grand canonical ensemble, the mean number of parts of size k is
 −1
a(k) = exp(bk) − 1 , (8)

and the mean number of parts greater than k is



 ∞
  −1
g(k) = a(j ) = exp(−bj k) exp(bj ) − 1 . (9)
k+1 1

If b is small and k is large, (9) reduces to


 
g(k) = −b−1 log 1 − exp(−bk) , (10)

which can be written in the symmetrical form


 
exp −bg(k) + exp(−bk) = 1. (11)
Partitions and the grand canonical ensemble 425

When b is small, the number of parts becomes large and the shape of the average
partition converges to the smooth curve

exp(−x) + exp(−y) = 1, (12)

where (x/b) is the number of parts larger than (y/b). The symmetry of (12) is the
well-known symmetry between a partition and its conjugate. According to (8), the
mean energy E is equal to the mean sum of all the parts,
    
E= ka(k) = k exp(−j kb) = (1/4) sinh−2 (j b/2) . (13)
k j j

The sum (13) is separated into two parts, the first exactly summable, the second
having a non-singular summand,
  
E= (j b)−2 + (1/4) sinh−2 (j b/2) − (j b)−2 . (14)
j j

The second sum can be approximated by an integral, so (14) becomes


 
E = π 2 /6 b−2 − (1/2)b−1 + (1/24), (15)

with an error of the order of b. The coefficient (π 2 /6) in (15) is equal to the area
under the curve (12). According to (5) and (6),

E = d(bF )/db, (16)

so that the free energy is obtained by integrating (15),


 
F = − π 2 /6 b−2 − (1/2)b−1 log(b/c), (17)

where c is a constant of integration and terms of order unity are neglected. A rather
tricky calculation of the asymptotic behavior of F for small b, using (6), gives the
result
c = 2π. (18)
The entropy is given by (7),
   
S = π 2 /3 b−1 + (1/2) log(b/2π) − 1 , (19)

neglecting terms of order b. The thermodynamics of the partitions of a large number


are described by (15), (17), (18), and (19), with temperature given by (4).
The statistical distribution of the energy n in the grand canonical ensemble can be
exactly described by the cumulant generating function
   
C(y) = ck y k /k! = log exp(ny) , (20)
426 F. Dyson

where y is a formal expansion parameter. The cumulants are the mean c1 , the variance
c2 , the skewness c3 , the kurtosis c4 , and so on. The definition (20), together with (1),
(2), and (3), gives
C(y) + log Z(b) = log Z(b − y), (21)
and then by (6) and (17),
     
C(y) = π 2 /6 y/ b(b − y) − (1/2) log b/(b − y) , (22)
  
ck = π 2 /6 b−k−1 − (2k)−1 b−k k!, (23)

neglecting terms of higher order in b. According to (20), the scale of the distribution-
function of the random variable n is set by the square-root of the variance c2 , so
the scale is proportional to b−3/2 . When scaled to the size of the distribution, the
skewness is of order b1/2 , the kurtosis is of order b, and all the higher cumulants ck
are of order b(k/2)−1 . So when b is small, the distribution of n becomes Gaussian
with deviations that vanish as b tends to zero.
It remains to use the thermodynamic model to estimate the size of the partition-
function p(n) appearing in (3). Let N be a fixed large integer. The ensemble of states
with the sum (1) exactly equal to N is the microcanonical ensemble M(N ), contain-
ing exactly p(N) states. The grand canonical ensemble with inverse temperature b is
obtained by adding together all the microcanonical ensembles M(n) with the statis-
tical weights x n according to (3) and (4). We choose the temperature b−1 so that the
quantity
   
log f (x)x −N = b(N − F ) (24)
is a minimum as b varies with N fixed. The condition for (24) to be a minimum is
according to (16)
N = F + b(dF /db) = E. (25)
So the grand and microcanonical ensembles have the same mean energy N , and
the value of (24) at the minimum is the entropy S of the grand ensemble according
to (7). Thus (6) and (17) give
 

f (x) = exp π 2 /6b (b/2π). (26)

The terms of the series



f (x)x −N = p(n)x n−N , (27)

have a maximum value p(N) at n = N , and decrease on each side with a distribution
that is Gaussian with variance given by (23),
 
V = π 2 /3 b−3 , (28)

to leading order in b. Therefore,


f (x) = p(N)exp(−bN ) (2πV ), (29)


Partitions and the grand canonical ensemble 427

with fractional error tending to zero with b. Putting together (29) with (15), (17), and
(26), we find the asymptotic formula
√ 

p(N) = (4N 3)−1 exp π (2N/3) , (30)

in agreement with Hardy and Ramanujan [4, Eq. (1.41), p. 280]. The asymptotic
formula (30) was the starting-point for the historic analysis by which Hardy and Ra-
manujan obtained an exact formula for p(N). Physical arguments can only take us
to the starting-point. To go further, Hardy and Ramanujan studied f (x) as a function
of the complex variable x analytic in the unit circle, and introduced the magic of
modular forms to determine its behavior precisely all around the circle.

2 An exercise in experimental mathematics

Recently, George Andrews, Frank Garvan, and Jie Liang wrote a paper [1], with
the title, “Combinatorial Interpretations of Congruences for the SPT-function”. The
SPT-function spt(n) is the total number of smallest parts in all the partitions of an
integer n. They proved that

spt(n) = NS (m, n), (31)
m

where the integer m is called the SPT-crank of a partition of n and takes values from
n − 1 to 1 − n. The NS (m, n) are defined as sums over partitions of contributions
equal to plus one or minus one, and it came as a surprise to Andrews et al. when
they were able to prove that all the NS (m, n) are positive. In their paper, they display
a table of numerical values of the NS (m, n) for n up to 20 and m up to 10. Since
NS (m, n) is an even function of m, only non-negative values of m are displayed.
As an exercise in experimental mathematics, I looked for a simple pattern in the
numerical values of the NS (m, n), and guessed a simple formula which makes it
easy to calculate them. The formula also explains why they are all positive. The
formula was proved correct by Garvan [3], by showing that it follows easily from
[1, Corollary 2.5]. The clue to guessing the formula was the behavior of the num-
bers in the right-hand side of the table where m ≥ n/2. In this region the values
are all equal to the standard partition function p(n − m − 1). When I subtracted
p(n − m − 1) from all the numbers in the table, the result was again simple in the
region where m ≥ n/3. In that region, the residuals were all negative and equal to mi-
nus (p(n − 2m − 3) + p(n − 2m − 5)). Adding these terms to the residuals, I found
the result to be again simple in the region m ≥ n/4. In that region, the second residu-
als were positive and equal to (p(n − 3m − 6) + p(n − 3m − 9) + p(n − 3m − 12)).
The pattern was now clear. My guess for the general formula was

 
k−1
  
NS (m, n) = (−1)k−1 p n − k(m + j ) − k(k + 1)/2 , (32)
k=1 j =0

which fitted all the data that I had available. For each n, the sum terminates after a
small number of terms.
428 F. Dyson

The formula (32) explains in a simple way why the NS (m, n) are all positive. It is
convenient to use the generating function



Am (x) = NS (m, n)x n . (33)
n=1

Then (32) becomes


 
k−1
Am (x) = Z(x) (−1)k−1
x (k(k+1)/2)+k(j +m) , (34)
k=1 j =0

where Z(x) is the generating function (3) of ordinary partitions. Now we group the
terms of the sum (34) together in groups of four,

(k, j + 1), (k + 1, j ), (k + 1, j + 1), (k + 2, j ), (35)

with k and j both odd. These groups of four cover the double sum (34) exactly, except
for the terms (k, k) which appear instead of the correct (k, k − 1). So the sum (34)
becomes
k
∞ 
   
Am (x) = Z(x) x (k(k+1)/2)+k(j +m) 1 − x j +m 1 − x j +m+k+1
k=1 j =1

 
+x (k(3k−1)/2)+mk
1−x k
, (36)

with the sums taken over k and j both odd. The minus signs of (32) have now disap-
peared into factors like (1 − x j +m ) which merely cancel the corresponding factors in
Z(x). From (36) it follows immediately that NS (m, n) is positive. In fact, NS (m, n)
is a sum of terms counting the number of partitions of (n − (k(k + 1)/2) − k(m + j ))
into parts with (j + m) and (j + m + k + 1) missing, plus partitions of (n − (k(3k −
1)/2) − mk) with k missing.
Another interesting consequence of (34) is the identity


  
Am (x) − Am+1 (x) = Z(x) (−1)k−1 x (k(k+1)/2)+mk − x (k(3k+1)/2)+mk . (37)
k=1

This shows the difference (Aj − Aj +1 ) to be closely related to the generating


functions of ranks and cranks according to [2, Eq. (7.7) and (7.17), pp. 68 and 70]. It
looks as if there are more treasures to be found here if we dig a little deeper.

Acknowledgements I am grateful to Andrews, Garvan, and Liang for sharing their results with me
before publication and for encouraging me to publish this little addendum.
Partitions and the grand canonical ensemble 429

References

1. Andrews, G.E., Garvan, F.G., Liang, J.L.: Combinatorial interpretations of congruences for the SPT-
function. Ramanujan J. (this issue)
2. Garvan, F.G.: New combinatorial interpretations of Ramanujan’s partition congruences mod 5, 7 and
11. Trans. Am. Math. Soc. 305, 47–77 (1988)
3. Garvan, F.G.: private communication
4. Hardy, G.H., Ramanujan, S.: Asymptotic formulae in combinatory analysis. Proc. Lond. Math. Soc.
(2) 17, 75–115 (1918)

You might also like