Mean Field Theory Solution of the Ising Model

Franz Utermohlen

September 12, 2018

Contents
1 Ising model 1

2 Mean field theory solution of the Ising model 2

2.1 Decoupling the Hamiltonian . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Self-consistency equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Critical behavior at zero field . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.1 Magnetization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.2 Susceptibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Free energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 Mean field theory solution of the Ising model with single-ion anisotropy 15
3.1 Decoupling the Hamiltonian . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Self-consistency equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1 Ising model
The (ferromagnetic) Ising model is a simple model of ferromagnetism that provides some
insight into how phase transitions and the non-analytic behavior of thermodynamic quantities
across phase transitions occur in physics. Consider a lattice containing a spin at each site that
can point either up (+1) or down (−1). The (classical) nearest-neighbor Ising Hamiltonian
for this system is
X N
X
H = −J si sj − h si , (1)
hiji i=1

where J is a positive (ferromagnetic) coupling constant, hiji denotes summation over nearest
neighbors,1 si (= +1 or −1) is the value of the spin on the ith site, h is an external magnetic
field pointing along the z direction,2 and N is the number of sites on the lattice.
1
Each pair of sites is only included once in the sum.
2
h > 0 corresponds to the field pointing in the +z direction.

1
As we will see, in dimensions higher than 1, the Ising model has two distinct phases, namely
a paramagnetic phase in which its spins are disordered due to thermal fluctuations, and a
ferromagnetic phase in which its spins start preferentially aligning in one direction. These
two phases are separated by a phase transition at some critical temperature T = Tc below
which the system becomes ferromagnetic. We can quantitatively distinguish these two phases
by defining the magnetization 3
N
1 X
m≡ hsi i , (2)
N i=1

where
Tr(si e−βH )
hsi i = (3)
Z
is the thermal expectation value (or mean value) of si , β = 1/(kB T ) is the inverse tempera-
ture,
Z = Tr(e−βH ) (4)
is the canonical partition function (from which we can extract thermodynamic information
about the system), and Tr (trace) denotes a sum over all of the system’s possible configura-
tions, which in the case of the Ising model can be written explicitly as
N
!
Y X X X X
Tr → = ··· . (5)
i=1 si =±1 s1 =±1 s2 =±1 sN =±1

The magnetization serves as this system’s order parameter, meaning we can use it to describe
how ordered the system is; for example, in this model, m = 0 corresponds to all of the spins
being disordered (paramagnetic state) and m 6= 0 corresponds to the spins having a preferred
direction (ferromagnetic state).

2 Mean field theory solution of the Ising model

The Ising model can only be exactly solved in one and two dimensions, although these
solutions are not very simple. It can, however, be approximately solved in any number of
dimensions in a relatively simple manner using the mean field theory (MFT) approximation
(usually just referred to as mean field theory). This approximation consists of assuming
that the system’s thermal fluctuations are relatively small and can therefore be neglected to
a certain extent. As we will see, from this simple assumption we can then treat a system
of interacting particles as a system of non-interacting4 particles in which each particle only
3
Some authors define the magnetization without the thermal expectation value, so what we are calling
m in these notes, they call hmi.
4
That is, not interacting with each other.

2
interacts with a “mean field” that captures the average behavior of the particles around
it. In other words, MFT effectively decouples the Hamiltonian into a simpler Hamiltonian
describing a non-interacting system. This makes it a very powerful method that is used often
in physics to explore the behavior of complicated many-body systems that cannot be solved
exactly.
Remarkably, even though the predictions MFT makes are quantitatively incorrect, it cor-
rectly predicts the Ising model’s qualitative behavior for two dimensions and higher. This is
because fluctuations are more important in lower dimensions, so the MFT approximation is
less accurate in lower dimensions.

2.1 Decoupling the Hamiltonian

In order to decouple the Ising Hamiltonian using MFT, we start by writing each spin in the
spin interaction terms si sj in the form

si = hsi i + δsi , (6)

where
δsi ≡ si − hsi i (7)
denotes fluctuations about the mean value of si . The spin interaction terms si sj thus be-
come

si sj = (hsi i + δsi )(hsj i + δsj )

= hsi i hsj i + hsj i δsi + hsi i δsj + δsi δsj . (8)

We now make the assumption that the fluctuations are very small, so we can ignore the term
quadratic in fluctuations:
δsi δsj = 0 . (9)
This is the approximation that MFT relies on: assuming fluctuations are small. The quantity
si sj is then approximately

si sj ≈ hsi i hsj i + hsj i δsi + hsi i δsj

= hsi i hsj i + hsj i (si − hsi i) + hsi i (sj − hsj i)
= hsj i si + hsi i sj − hsi i hsj i . (10)

Since this system is translationally invariant, the expectation value hsi i of any given site i
is independent of the site, so we have

hsi i = m . (11)

We can then further simplify si sj to

si sj = m(si + sj ) − m2 = m[(si + sj ) − m] (12)

3
and approximate the Ising Hamiltonian as
X N
X
HMF = −Jm (si + sj − m) − h si , (13)
hiji i=1

where the subscript on the Hamiltonian reminds us that this is only an approximation of
the Ising Hamiltonian. Since X X
si = sj (14)
hiji hiji

due to the symmetry of i and j in the sum over nearest neighbors, we can write
X X
(si + sj ) = 2si , (15)
hiji hiji

so the Hamiltonian reads

X N
X
HMF = −Jm (2si − m) − h si . (16)
hiji i=1

We can write the sum over nearest neighbors as

N
X 1X X
→ , (17)
2 i=1
hiji j∈nn(i)

where the factor of 1/2 is to avoid double counting pairs of sites and nn(i) denotes nearest
neighbors of i. Since there is no explicit j dependence inside the summation, this inner sum
is simply X
= q, (18)
j∈nn(i)

where the coordination number q is equal to the number of neighbors of any given site.5 We
then have
N
X qX
→ , (19)
2 i=1
hiji

so the Ising Hamiltonian further simplifies to

N N
qJm X X
HMF =− (2si − m) − h si
2 i=1 i=1
N N
N qJm2 X X
= − qJm si − h si
2 i=1 i=1
N
N qJm2 X
= − (h + qJm) si , (20)
2 i=1
5
For example, for a 1D lattice, q = 2; for a 2D triangular lattice, q = 3; for a 2D square lattice, q = 4;
etc.

4
or simply
N
N qJm2 X
HMF = − heff si , (21)
2 i=1

where
heff ≡ h + qJm (22)
is the effective magnetic field felt by the spins. We have now effectively decoupled the
Hamiltonian into a sum of one-body terms. Again, this conceptually means that particles
no longer interact with each other in this approximation, but rather interact only with an
effective magnetic field heff that is comprised of the external field h and the mean field qJm
induced by neighboring particles.

2.2 Self-consistency equation

Let’s calculate the partition function using the mean field Ising Hamiltonian:

ZMF = Tr(e−βHMF )
N
!
Y X
= e−βHMF
i=1 si =±1
N
!
Y X 2 /2
PN
= e−βN qJm eβheff j=1 sj

i=1 si =±1
N
!
−βN qJm2 /2
Y X
=e eβheff si
i=1 si =±1
N
2 /2
Y
= e−βN qJm (eβheff + e−βheff )
i=1
| {z }
=2 cosh(βheff )
−βN qJm2 /2
=e [2 cosh(βheff )]N , (23)

so we find
2 /2
ZMF = e−βN qJm [2 cosh(βheff )]N . (24)
Now, recall from Eq. 2 that the magnetization is given by
N
1 X
m≡ hsi i .
N i=1

5
Using Eq. 3, we can rewrite this as
N
1 X Tr(si e−βHMF )
m=
N i=1 ZMF
N
1 1 X −βHMF
= si e
N ZMF i=1
1 1 ∂ZMF
=
N β ZMF ∂heff
1 ∂(ln ZMF )
= , (25)
N β ∂heff
where on the third line we have used the fact that
∂ZMF ∂
= Tr(e−βHMF )
∂heff ∂heff
 
∂ −βHMF
= Tr e
∂heff
 
∂HMF −βHMF
= −β Tr e
∂heff
" N
! #
X
= −β Tr − si e−βHMF
i=1
" N
#
X
= β Tr si e−βHMF . (26)
i=1

Now, from Eq. 24 we can calculate ln ZMF :

βN qJm2
ln ZMF =− + N ln 2 + N ln[cosh(βheff )] , (27)
2
so
∂(ln ZMF )
= N β tanh(βheff ) . (28)
∂heff
Inserting this back into Eq. 25 gives

m = tanh(βheff ) . (29)

Inserting the definition of heff (Eq. 22) back into this equation gives us the self-consistency
equation
m = tanh[β(h + qJm)] . (30)
This is a transcendental equation, so we cannot solve for m analytically. However, we can
solve for m graphically by plotting m and tanh[β(h + qJm)] (for fixed values of β, h, q, and
J).

6
Let’s consider the special case of h = 0. In Figure 1 we plot m and tanh(βqJm) on the same
plot for various values of βqJ. We can graphically see that the solutions are qualitatively
different when βqJ ≤ 1 and βqJ > 1.6 These two cases can be understood as follows:
• βqJ ≤ 1 (i.e., kB T ≥ qJ): There is only one solution: m = 0. This corresponds to the
system being in a paramagnetic state.
• βqJ > 1 (i.e., kB T < qJ): There are three solutions: m = 0 and m = ±m0 , where
m0 ≤ 1. The m = ±m0 solutions correspond to the system being in a ferromagnetic
state (the system is magnetized). As we will discuss later, the m = 0 solution turns
out to be unstable, so we physically only observe either m = m0 or m = −m0 at these
temperatures.
The critical temperature Tc below which the system becomes spontaneously magnetized
without any external magnetic fields is therefore given by

kB Tc = qJ . (31)

It should be noted that for the 1D case, MFT thus predicts a magnetic phase transition
at kB Tc = 2J; however, solving this system exactly we find that there is in fact no phase
transition in 1D. Thermal fluctuations turn out to be strong enough to destroy the system’s
magnetic ordering in 1D, so MFT provides a qualitatively incorrect result in this specific
case.
As mentioned earlier, MFT is more accurate in higher dimensions. For example, for the 2D
square lattice, MFT √ predicts kB Tc = 4J, whereas solving the system exactly we find that
kB Tc = 2J/ ln(1 + 2) ≈ 2.27J. Also note that since MFT assumes fluctuations are small,
it generally overestimates the system’s tendency to order and thus overestimates the value
of Tc .

2.3 Critical behavior at zero field

In this section we discuss the behavior of the system near the critical temperature Tc . In
order to more easily look at the behavior for T → Tc , let’s define the reduced temperature t
6
We can also reach this conclusion a bit more rigorously by noting that there are three solutions when
 
d 
tanh(βqJm)  > 1,
dm m=0

and one solution otherwise. Evaluating the derivative gives

   
d  1 
tanh(βqJm)  = βqJ 2
 = βqJ ,
dm m=0 cosh (βqJm) m=0

so we find that we have three solutions solution when βqJ > 1, and one solution otherwise.

7
 1.0

βqJ = 0.5
0.5

m
m
-1.0 -0.5 0.5 1.0 tanh(0.5 m)

-0.5

-1.0

1.0

βqJ = 1
0.5

m
m
-1.0 -0.5 0.5 1.0 tanh(m)

-0.5

-1.0

1.0

βqJ = 1.5
0.5

m
m
-1.0 -0.5 0.5 1.0 tanh(1.5 m)

-0.5

-1.0

Figure 1: Solving the self-consistent equation graphically for h = 0 and various values of βqJ.
The top plot is at T > Tc , the middle plot is at T = Tc , and the bottom plot is at T < Tc .

8
by
T − Tc
t≡ . (32)
Tc
We can then look at the critical behavior by expanding around small t.

2.3.1 Magnetization

We already know that when there is no external magnetic field, we have m = 0 when T ≥ Tc .
Let’s explore further what happens in the critical regime just below Tc (i.e., T → Tc− ) at
zero field (h = 0). Using the expansion
x3
tanh(x) = x − + O(x5 ) (33)
3
around x = 0, we can rewrite the self-consistency equation (Eq. 30) for h = 0 around m = 0
(near Tc ) as
1
m = βqJm − (βqJm)3
3
   3
Tc 1 Tc
= m− m3 , T → Tc− , (34)
T 3 T
or moving everything to the same side,
"   3 #
Tc 1 Tc
m −1 − m2 = 0 , T → Tc− . (35)
T 3 T

This gives the solutions m = 0 and

s  2  
T Tc − T
m=± 3 , T → Tc− . (36)
Tc Tc

Note that this expression is only sensible for T ≤ Tc , otherwise m would be imaginary (which
is unphysical). However, we have to keep in mind that this expression only applies whenever
m is very close to 0, meaning that we are looking only at the region below Tc but very close
to Tc . In terms of the reduced temperature t, Eq. 36 reads
p
m = ± −3(1 + t)2 t , T → Tc− . (37)

Using the binomial approximation

(1 + x)α ≈ 1 + αx , |x|  1 , (38)

we can simplify this for t → 0− (i.e., T → Tc− ) as

p √
m ≈ ± −3(1 + 2t)t ≈ ± −3t , T → Tc− . (39)

9
 m
1

0
T
Tc

-1

Figure 2: Behavior of the magnetization m(T, 0) obtained by solving the Ising model using MFT.

We have thus found that at zero field, the magnetization m(T, h) behaves as
(
±(3|t|)1/2 , T → Tc−
m(T, 0) = . (40)
0, T → Tc+

In the zero-temperature limit (T → 0 , β → ∞), the self-consistency equation gives m0 = 1,

so
m(0, 0) = ±1 . (41)
We can use all of this information to sketch the magnetization as a function of temperature in
Figure 2. The physical interpretation of this is that above Tc , thermal fluctuations are strong
enough to prevent the system from becoming magnetized; this is the paramagnetic phase.
However, at Tc a phase transition occurs and the system becomes spontaneously magnetized
as you decrease the temperature past Tc ; this is the ferromagnetic phase. Even though this
behavior is quantitatively wrong (and notably qualitatively wrong in 1D!), it is nevertheless
qualitatively correct in 2D and higher dimensions, which is quite remarkable.

2.3.2 Susceptibility

The isothermal (magnetic) susceptibility χT is given by

 
∂m
χT (T, h) ≡ . (42)
∂h T
Taking the derivative of both sides of the self-consistency equation (Eq. 30) with respect to
h while keeping T constant gives
1
χT = 2 β(1 + qJχT ) . (43)
cosh [β(h + qJm)]

10
Solving for χT , we find

β
χT (T, h) = 2 . (44)
cosh [β(h + qJm)] − βqJ

Let’s now look at the behavior of the at zero field (h = 0):

β
χT (T, 0) = 2 . (45)
cosh (βqJm) − βqJ

For T → Tc+ , we have m = 0, so

β 1 1 1
χT (T, 0) = = = , T → Tc+ . (46)
1 − βqJ kB (T − Tc ) kB Tc |t|

For T ≤ Tc but very close to Tc , we have m = ±(3|t|)1/2 (as found in the previous section),
which is very small. We can thus use the expansion

x2
cosh(x) = 1 + + O(x4 ) (47)
2
around x = 0 to write
β
χT (T, 0) = 1
[1 + 2
(βqJ(3|t|)1/2 )2 ]2 − βqJ
β
≈
1 + 3(βqJ)2 |t| − βqJ
1 1 1
≈ , T → Tc− . (48)
2 kB Tc |t|

We have thus found that at zero field, the isothermal susceptibility goes as
(
A|t|−1 , T → Tc− 1 1
χT (T, 0) = , where A = , (49)
2A|t|−1 , T → Tc+ 2 kB Tc

and diverges at T = Tc . We can use this information to sketch the susceptibility as a function
of temperature near Tc in Figure 3.

2.4 Free energy

We can also gain some useful insight into the physics of this system by looking at its
(Helmholtz) free energy
F ≡ U − T S = −kB T ln Z , (50)

11
 χT (T, 0)

T
Tc

Figure 3: Behavior of the isothermal susceptibility χT (T, 0) near Tc obtained by solving the Ising
model using MFT.

where S is the system’s entropy. Understanding the behavior of the system’s free energy
is helpful because the equilibrium state of a system at constant temperature and volume
minimizes its free energy. Specifically, we are going to be looking at how the system’s free
energy depends on its magnetization. At zero field, the free energy of this system is

FMF |h=0 = −kB T ln ZMF

N qJm2
= − N kB T ln 2 − N kB T ln[cosh(βheff )]
2
N qJm2 (βheff )2 (βheff )4
 
6


= − N kB T ln 2 − N kB T − + O (βheff )
2 2 12
N qJm2
 2
β (qJm)2 β 4 (qJm)4

≈ − N kB T ln 2 − N kB T −
2 2 12
2 2 4
N qJm N (qJm) N (qJm)
= − N kB T ln 2 − +
2 2kB T 12(kB T )3
N kB Tc N kB Tc4 4
= −N kB T ln 2 + (T − Tc ) m2 + m , (51)
2T 12T 3
which, as a function of m, is the form

FMF (m)|h=0 = F0 + a(T − Tc )m2 + bm4 (52)

up to order m4 . Adding an external magnetic field adds a term linear in h and m:7

FMF (m) = F0 − hm + a(T − Tc )m2 + bm4 . (53)

Note that the term constant in m is not relevant to the system’s equilibrium state, since it
is just an overall energy shift that does not affect the free energy’s minima with respect to
m. In Figures 4 and 5 we will therefore therefore implicitly ignore the F0 term.
7
As well as some higher-order terms that we are going to ignore.

12
 FMF (m)
T > Tc
h=0

m
FMF (m)
T = Tc
h=0

m
FMF (m)
T < Tc
h=0

m
-m0 m0

Figure 4: Mean field solution of the free energy as a function of magnetization at zero field at
various temperatures. As we cool the system below Tc , the free energy changes smoothly to a
“Mexican hat” shape, thereby making the m = 0 solution unstable and two new stable solutions
at m = ±m0 appear.

Plots of the free energy at zero field are shown in Figure 4. We can see that as we cool the
system below Tc , the m = 0 solution becomes metastable and any small perturbation causes
the system to spontaneously “roll down” either to m = m0 or m = −m0 . When we apply a
weak external magnetic field below Tc , the solution magnetized in the direction of the field
becomes more energetically favorable and more stable (see Figure 5). If the field is strong
enough, the solution magnetized in the direction of the field becomes the only solution.

13
 FMF (m)
T > Tc
h>0

m
FMF (m)
T < Tc
h > 0 (small)

m
-m0 m0

FMF (m)
T < Tc
h > 0 (large)

Figure 5: Mean field solution of the free energy as a function of magnetization above and below
Tc with an external magnetic field h in the +z direction. Above Tc , the system “rolls” down into a
state with positive magnetization, which is what we would expect. Below Tc , the solution m = m0
is now more stable than the m = −m0 solution. For large enough h, m = −m0 stops being a
solution and m = m0 becomes the only solution.

14
3 Mean field theory solution of the Ising model with
single-ion anisotropy
Suppose we add a single-ion anisotropy term to the Ising model, so our Hamiltonian is
now
X N
X N
X
2
H = −J si sj − D (si ) − h si , (54)
hiji i=1 i=1

where J > 0 is a ferromagnetic coupling constant and D is the single-ion anisotropy constant
that favors easy-axis magnetization for D > 0 and easy-plane magnetization for D < 0.

3.1 Decoupling the Hamiltonian

For the interaction term, we will write si sj in the form

si = hsi i + δsi , (55)

where
δsi ≡ si − hsi i (56)
denotes fluctuations about the mean value of si . The spin interaction terms si sj thus be-
come

si sj = (hsi i + δsi )(hsj i + δsj )

= hsi i hsj i + hsj i δsi + hsi i δsj + δsi δsj . (57)

We now make the assumption that the fluctuations are very small, so we can ignore the term
quadratic in fluctuations:
δsi δsj = 0 . (58)
The quantity si sj is then approximately

si sj ≈ hsi i hsj i + hsj i δsi + hsi i δsj

= hsi i hsj i + hsj i (si − hsi i) + hsi i (sj − hsj i)
= hsj i si + hsi i sj − hsi i hsj i . (59)

Since this system is translationally invariant, the expectation value hsi i of any given site i
is independent of the site, so we have

hsi i = m . (60)

We can then further simplify si sj to

si sj = m(si + sj ) − m2 = m[(si + sj ) − m] . (61)

15
Similarly, for the single-ion anisotropy term, we will write (si )2 as
(si )2 = 2msi − m2 = m(2si − m) . (62)
The mean field Hamiltonian is then
X N
X N
X
HMF = −Jm (si + sj − m) − Dm (2si − m) − h si
hiji i=1 i=1

X N
X N
X
= −Jm (2si − m) − Dm (2si − m) − h si
hiji i=1 i=1
N N N
qJm X X
X
=− (2si − m) − Dm (2si − m) − h si
2 i=1 i=1 i=1
  N
qJ X
= + D N m2 − [h + (qJ + 2D)m] si ,
2 i=1

where q is the coordination number. We thus find

  N
qJ X
HMF = + D N m2 − heff si , (63)
2 i=1

where
heff ≡ h + (qJ + 2D)m (64)
is the effective magnetic field felt by the spins.

3.2 Self-consistency equation

Let’s calculate the partition function using the mean field Hamiltonian:
ZMF = Tr(e−βHMF )
N
!
Y X
= e−βHMF
i=1 si =±1
N
!
Y X qJ 2
PN
= e−β( 2 +D)N m eβheff j=1 sj

i=1 si =±1
N
!
−β( qJ +D)N m2
Y X
=e 2 eβheff si
i=1 si =±1
N
qJ 2
Y
= e−β( 2 +D)N m (eβheff + e−βheff )
i=1
| {z }
=2 cosh(βheff )
−β( qJ +D)N m2
=e 2 [2 cosh(βheff )]N , (65)

16
so we find
qJ 2
ZMF = e−β( 2 +D)N m [2 cosh(βheff )]N . (66)
Now, recall from Eq. 2 that the magnetization is given by
N
1 X
m≡ hsi i .
N i=1

Using Eq. 3, we can rewrite this as

N
1 X Tr(si e−βHMF )
m=
N i=1 ZMF
N
1 1 X −βHMF
= si e
N ZMF i=1
1 1 ∂ZMF
=
N β ZMF ∂heff
1 ∂(ln ZMF )
= , (67)
N β ∂heff
where on the third line we have used the fact that
∂ZMF ∂
= Tr(e−βHMF )
∂heff ∂heff
 
∂ −βHMF
= Tr e
∂heff
 
∂HMF −βHMF
= −β Tr e
∂heff
" N
! #
X
= −β Tr − si e−βHMF
i=1
" N
#
X
= β Tr si e−βHMF . (68)
i=1

Now, from Eq. 66 we can calculate ln ZMF :

 
qJ
ln ZMF = −β + D N m2 + N ln 2 + N ln[cosh(βheff )] , (69)
2
so
∂(ln ZMF )
= N β tanh(βheff ) . (70)
∂heff
Inserting this back into Eq. 67 gives

m = tanh(βheff ) . (71)

17
Inserting the definition of heff (Eq. 64) back into this equation gives us the self-consistency
equation


m = tanh β h + (qJ + 2D)m . (72)
This corresponds to a system with a phase transition at kB Tc = qJ + 2D. Above this
temperature, it is paramagnetic, and below this temperature, it is ferromagnetic. Note that
the system will be paramagnetic for all (nonzero) temperatures if D < −qJ/2 (remember
that D < 0 favors easy-plane magnetization).

References
[1] Mean-Field Theory. https://www.physik.uni-muenchen.de/lehre/vorlesungen/
sose_14/asp/lectures/ASP_Chapter5.pdf.

[2] Martin Evans. Mean-Field Theory of the Ising Model. https://www2.ph.ed.ac.uk/

~mevans/sp/sp10.pdf.

18