Statistical Thermodynamics: Molecules to Machines

Venkat Viswanathan
May 5, 2015

Module 5: Ising model

Learning Objectives:
Introduce the Ising model for magnetic systems.
Exactly solve the Ising model in 1 dimension.
Give the mean-field approximation for the Ising model in D 2.
Discuss critical behavior in magnetic systems and in liquid-vapor
systems
Show how the mean-field approximation does not accurately ac- count
for correlations that dominate the behavior near the critical point.

Key Concepts:
Ising model, dimensionality, ferromagnetic and antiferromagnetic, magnetization, transfer function, transfer matrix, meanfield approximation,
critical behavior, fluctuations, critical fluctuations, Rayleigh scattering

statistical thermodynamics: molecules to machines

Non-interacting molecules
Thus far, our analyses have focused on non-interacting systems:
Ideal gas Indistinguishable, non-interacting bosons
Electron gas Indistinguishable, non-interacting fermions
Crystal fluctuations Distinguishable, non-interacting phonons (quasiparticles).
Non-interacting systems require the evaluation of a single-molecule
partition function, which generally is tractable (either exactly or approximately). However, very few practical problems involve either noninteracting molecules or approximately non-interacting molecules. Furthermore, a variety of important physical phenomena are not exhibited
in non-interacting systems. For example, phase transitions are a hallmark issue in interacting systems. Our goal for the remainder of this
class is to discuss the effect of interactions on thermodynamic behavior.

Ising model
Consider a lattice of N sites. The definition of the lattice can include
many different geometries (square lattice, triangular lattice, etc..) and
dimensionality (1 dimension, 2 dimensions, etc...). Each site contains
exactly 1 spin that is defined by the spin state si that is either spin up
(si = +1) or spin down (si = 1). The lattice spins are subjected to an
external field h, and the spins only interact with there nearest neighbors
with a coupling strength J. The system energy is defined to be:
E = h

N
X
i=1

si J

si sj

(1)

hiji

where the sum over hiji implies a summation over all nearest neighbor
pairs in the lattice.
This general model can be applied to a variety of problems (magnetic
systems, binary alloys, liquid-gas phase transition, neuron network Hopfield model).
The sign of the coupling constant J dictates whether the spins prefer
to align with their neighbors or anti-align (Fig. 1):
Ferromagnetic the spins prefer to align with their neighbors, occurring when J > 0.
Antiferromagnetic the spins prefer to anti-align with their neighbors,
occurring when J < 0.

Figure 1: Ground states (minimum energy) for the 2-dimensional Ising model
(square lattice) for a ferromagnetic system (with coupling constant J > 0)
and an antiferromagnetic system (with
coupling constant J < 0).

statistical thermodynamics: molecules to machines

Non-interacting case (J = 0)
The partition function for the Ising model with J = 0 is given by

Q=

...

s1 =1,1 s2 =1,1

exp h

sN =1,1

ehs1

s1 =1,1

N
X

!
si

i=1

ehs2 ...

s2 =1,1

ehsN

sN =1,1

ehs

N

= eh + eh
= 2N coshN (h)

s=1,1

which is valid for any lattice type and dimensionality.

Define the total magnetization M to be:
X

N
log Q
= N tanh (h)
M=
si =
h

(2)

i=1

The average energy hEi and entropy S are given by:

X

N
hEi = h
si = hM = hN tanh (h)
i=1

EF
hN tanh (h) + N kB T log [2 cosh (h)]
S=
=
T
T
Note, hEi hN and S 0 as T 0, and hEi 0 and S
N kB log 2 as T .
The non-interacting Ising model (J = 0) exhibits a gradual change
in the magnetization M with the external field h (Fig. 2). Thus, no
phase transition is exhibited in the non-interacting case; there is never
a phase transition for non-interacting molecules. We now turn to cases
where the interactions are turned on, such that J 6= 0.

Ising model in 1 dimension

Consider a line of spins in 1 dimension with periodic boundary conditions
such that sN interacts with s1 . For the Ising model in 1 dimension, the
energy is written as:
E = h

N
X
i=1

si J

N
X

si si+1

(3)

i=1

where h = h and J = J.
Define a transfer function T (si , si+1 ) to be:
h s
i
s 
i
i si+1
T (si , si+1 ) = exp h
+ i+1 + Js
2
2

(4)

Figure 2: The total magnetization

M = N versus the external field h for
the non-interacting Ising model J = 0).

statistical thermodynamics: molecules to machines

With this definition, we write the partition function as:

Q=

s1 =1,1 s2 =1,1

...

T (s1 , s2 )T (s2 , s3 )...T (sN , s1 )

sN =1,1

Now construct the transfer matrix to be:

"
# "

T (1, 1)
T (1, 1)
eh + J
T =
=

T (1, 1) T (1, 1)
eJ

eJ

eh+J

The partition function is then written as:

X
Q=
T N (s1 , s1 ) = T r{T N }

(5)

(6)

s1 =1,1

where T r{T N } indicates a trace of the matrix M .

The trace of a matrix is the sum of its eigenvalues, and the eigenvalues
of T N are N
, where are the eigenvalues of T . In our case:

1/2

= eJ cosh h e2J + e2J sinh2 h

(7)

Therefore, the exact solution for the partition function is:



1/2 N
J
2J
2J
2

+
Q = e cosh h + e
+ e sinh h


1/2 N

eJ cosh h e2J + e2J sinh2 h

which agrees with
For J = 0, the partition function Q = 2N coshN h,

our previous result. For h = 0, the partition function Q = 2N coshN J +

2N sinhN J 2N coshN J for large N .
Since + > , the partition function is Q N
+ for N  1. The
Q
sinh h
average magnetization is given by M = log
=
N
. In
h

e4J +sinh h

the limit h 0+ , the magnetization approaches zero. Thus, this exact

solution for the 1D Ising model does not exhibit a phase transition from
a disordered state to an ordered state. The physical justification for
this lies in the fact that forming a disordered phase from an ordered
phase occurs through an energy change E = 4J in 1D (independent of
N ). In higher dimensions (D 2), the energy cost scales with N (e.g.
E N 1/2 J in 2D), thus leading to a finite-T phase transition (Fig. 3).

Mean-field approximation
The solution for the Ising model for D 2 is not as easily found. In
1944, Lars Onsager found an exact solution for the 2D Ising model for
h = 0 that exhibits a spontaneous magnetization for sufficiently large
J. However, there is no exact solution for the Ising model for D 3.

Figure 3: Energy cost of a phase transition from an ordered phase to a disordered phase in 1D and 2D.

statistical thermodynamics: molecules to machines

In this module we introduce the mean-field approximation, and the

limits of this approximation. The mean-field approximation involves
approximating the detailed interactions with the average of these interactions (Fig. 4).
1 PN
Define the average magnetization per site m = M
i=1 hsi i =
m = N
hsi, noting that spatial invariance implies hsi i = hsi. Define the coordination number z as the number of nearest neighbors to any given site
divided by 2 (e.g. z = 2 for 2D square lattice and z = 3 for 3D square
lattice). The mean-field approximation replaces the near-neighbor interactions by the average interaction, thus sj hsj i, giving the energy
approximation:
E = h

N
X
i=1

si J

si sj h

N
X
i=1

hiji

si Jzm

N
X

si

Figure 4: Schematic of the mean-field

approximation.

(8)

i=1

which reduces the problem to an effective non-interacting model.

Therefore, we write the average magnetization per site as:
P
s=1,1 s exp [(h + Jzm)s]
= tanh(h + Jzm)
m= P
s=1,1 exp [(h + Jzm)s]

(9)

which performs the average of the site magnetization using the meanfield site energy in the Boltzmann weight for the spin probability.
For the case h = 0 (no external field), the magnetization satisfies
m = tanh(Km), where K = Jz. This is a self-consistent mean-field
equation. Local order is dictated by the surrounding order that is itself
defined by the local order, leading to a self-consistent condition that
must be met. Two scenarios exist (Fig. 5):

Figure 5: Plots of tanh(Km) versus m

for various values of K. Circles mark
points that satisfy the self-consistent
equation m = tanh(Km).

For K < 1, there is only one solution for m that corresponds to

m = 0.
For K > 1, there are 3 solutions for m corresponding to m = 0 and
m = m . The solution m = 0 is unphysical as it is a free energy
maximum.
Therefore, K = Kc = 1 defines a critical temperature Tc = Jz/kB ,
and we can write K = Tc /T (Fig. 6).
Near T = Tc , we can expand the self-consistent equation near m = 0
to get:

m = tanh

Tc
m
T

Tc
1
m
T
3

Tc
T

3

r
3
m3 m
(Tc T )1/2
Tc

Generally, the external field h acts as the intensive variable that is

conjugate to the magnetization M (extensive variable). In Module 2 we
learned that thermodynamic stability dictates that a derivative of an

Figure 6: Phase diagram for the meanfield solution for the Ising model, showing m versus T /Tc .

statistical thermodynamics: molecules to machines

intensive variable with respect to its conjugate extensive variable is a

positive quantity, e.g.:


T
S


,
V ,N

p
V


,

T ,N

i
Ni


>0

(10)

T ,V ,Nj6=i

Similarly, thermodynamic stability dictates that:



h
>0
M Jz,N

(11)

The phase diagram for the mean-field solution is dictated by the

self-consistent mean field equation (Fig. 7):
m = tanh(h + Jzm)

(12)

This equation of state can be likened to the vapor-liquid equation of

state p = p(v ) (e.g. the van der Waals equation of state from Module
2, Fig. 9).
The phase diagram for the mean-field approximation for the Ising
model exhibits similar characteristics as the liquid-vapor phase diagram.
Correspondence between M h in the magnetic system and p V in
the liquid-vapor system reflects the universal physical issues that are
addressed with the Ising model and why it can be used to understand a
variety of phenomena (Fig. 9).

Critical point in vapor-liquid and magnetic systems

Previously, we discussed the phase behavior of a magnetic system by
analyzing the exact 1D and mean-field approximation of the Ising model.
We saw that no phase transition occurs for the 1D Ising model and the
model exhibits a true phase transition for D 2, which was argued by
physical arguments and through the mean-field approximation.
The h M phase diagram in a magnetic system was found to behave
akin to the p V phase diagram in vapor-liquid systems. In both systems, a critical point exists that marks the onset of phase coexistence
in the system, i.e. for T > Tc , the system does not exhibit a phase
transition, and for T < Tc , phase transitions exist. Now, we discuss the
physical behavior near the critical point for both a magnetic system and
a liquid-vapor system.
The fluctuations of the energy, density, and magnetization for a system are found using various ensemble manipulations. From the canonical
ensemble, the variance of the energy is:

Figure 7: Plot of m versus h for several values of K = Jz = Tc = T .

The dashed part of each curve indicates
where the system is unstable.

Figure 8: Comparison of the phase diagrams for a magnetic system modeled

using the mean-field approximation of
the Ising model (left panel) and the
vapor-liquid coexistence for a van der
Waals fluid (right panel).

Figure 9: Plot of m versus h for K =

Jz = Tc /T = 1.5. The arrows indicate the hysteresis in magnetization.

statistical thermodynamics: molecules to machines




 
1
1 2 Q 
Q 2 

hE i hEi =

Q 2 V ,N Q2
V ,N
 2
 2




F 
ln Q 
=
= kT 2 CV
=

2
2 V ,N
V ,N
2

(13)

Thus connecting the energy fluctuations to CV . The standard devip

ation E = hE 2 i hEi2 for energy fluctuations results in the characteristic magnitude of the energy fluctuation per molecule in terms of the
heat capcity per molecule cV = CNV as:
E
1p
N kT 2 cV =
=
N
N

kT 2 cV
N

(14)

This calculation verifies that the average energy per molecule is a

well defined quantity with negligible fluctuations, provided that the heat
capacity does not diverge. Similarly, the grand canonical ensemble for a
one-component system yields the molecule number variance to be:

hN i hN i =

2 ln
()2

=
,V

2 (pV )
()2

= kT
,V

hN i

T ,V

(15)
And noting the Gibbs-Duhem equation for a 1 component system
V
(d = sdT + vdp) and v = hN
i , the variance in N can be written as:

hN 2 i hN i2 = kT

V
v3

Using the definition of molar density =

h2 i hi2 =

kT
V v3

v
p
1
v

v
p

(16)
T

hN i
V

(17)
T

 
 >

T
0, and if V , the density fluctuations
in a stable system are neg
 
p 
ligible. However, the condition v
 0 dictates that the density
T
fluctuations in the system diverge. Thus,
the limit of stability for a
  
p 
1-component system, marked by v  = 0, exhibits wild density
T
fluctuations.
As a system approaches the limit of stability for T < Tc , density
fluctuations become substantial, leading to the eventual formation of the
The condition for thermodynamic stability dictates that

v
p

Figure 10: Plot of p versus v for a =

3.375 = 27/8 (critical point, blue), a
= 3.527, a = 3.679, a = 3.831, and a
= 3.982 (red). The solid black curves
are the binodal curves (coexistence),
and the dashed black curves are the
spinodal curves (limit of metastability).
The critical point is identified as the
point where the spinodal curves converge.

statistical thermodynamics: molecules to machines

2 coexisting phases (spinodal decomposition) as can be seen in Fig. 10.

As the system approaches the critical point, the density fluctuations
dominate the physical behavior.
Rayleigh Scattering occurs when the length scale of density fluctuations is comparable to the wavelength of light, where the local mismatch
in the index of refraction leads to light scattering, as can be seen in
Fig. 11.
The fluctuation in the magnetization for the Ising model is governed
by:

hM i hM i =

2Q
(h)2

=
J,N

hM i
(h)

=N
J,N

m
(h)

(18)
As in the vapor-liquid system, the Ising model exhibits wild fluctua

(h) 
tions at the limit of metastability, marked by the condition M 
=
J,N

0. In this case, the fluctuations are in the magnetization rather than the
density. If we are far away from the critical point, the standard deviation
in the magnetization per site is:

M
=
N

1
N

m
(h)

(19)
J

Thus, suggesting that the magnetization is well-defined in the thermodynamic limit N . As we approach the critical point, the role of
fluctuations dominates the system thermodynamics, as we will proceed
to discuss.

Critical behavior for the Ising model

We will discuss the critical behavior for the Ising model for the Meanfield approximation using our treatment developed earlier.
To illustrate these concepts, focus on the 2-dimensional Ising model
on a square lattice. The analytical solution for the 2D Ising model shows
that the critical temperature is Tc = 2.269J /k and the magnetization
1
scales as M N (T Tc ) 8 .

Mean-field approximation
We showed that the equation of state for the mean-field approximation
to the Ising model is:

m = tanh(h + Jzm) = tanh(h + Km)

(20)

Figure 11: The sky is blue due to a

phenomenon called Rayleigh scat- tering, where light is scattered due to local
density fluctuations in the fluid. Such
density fluctuations occur for a fluid
that is near its critical point, which is
the case for the oxygen and nitrogen
in the upper atmosphere. The shortest wavelength light (violet end of spectrum) is scattered the most, leading to
an overall perceived color that is light
blue.

statistical thermodynamics: molecules to machines

The mean-field critical point is at h = 0 and Kc = 1, thus we can

write K = Tc /T , where Tc = 2J /k for a 2D square lattice. Just below
T = Tc and h = 0, on expanding m near m = 0, we get:

m = tanh

Tc
m
T

Tc
1

m
T
3

Tc
T

3

r
3
m m
(Tc T )1/2
Tc
(21)
3

To find the fluctuations in the magnetization near T = Tc , we find:

m
m
= [1 tanh2 (h + Km)][1 + K ]

h
h
1 m2
=
1 K + Km2

(22)

For T > Tc , we have m = 0. Therefore:

m
1
1
=

1K
T Tc
h
And for T < Tc , we have:
1 3 + 3/K
1
1
m
=

1 K + K (3 3/K )
K 1
Tc T
h
Thus, the mean-field approximation to the Ising model predicts the
fluctuations in the magnetization to diverge near the critical temperature
as:

hM 2 i hM i2

N
|T Tc |

(23)

For T = Tc (K = 1) and h 6= 0, expand m about m = 0 to get:

m ' (h + m) 1/3(h + m)3 m ' 31/3 h 1/3 h h1/3

(24)

Deviation of the mean-field approximation from the exact result lies

in the neglect of correlations near the critical point, thus the mean-field
approximation is best far from the critical point.
For a vapor-liquid system governed by a mean-field theory (e.g. van
der Waals equation of state), the critical behavior is:

c (Tc T )1/2 , h2 i hi2

c (p pc )1/3

1
T
|T Tc |
(25)

statistical thermodynamics: molecules to machines

Which shows a direct correspondence between m and h p. The

critical behavior of the Ising model and of other related thermodynamic
systems reflects the underlying impact of correlations on the thermodynamic behavior. The mean-field approximation does not properly
account for these correlations, a technique called renormalization group
theory can account for spatial correlations which is beyond the scope of
this course.

10