Contents

1 Phase Transitions 2
1.1 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Thermodynamics of Phase Transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Definitions in Ferromagnetic Case . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Landau Theory of Phase Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 Second-order Phase Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.2 Critical Exponents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.3 Application to Ferromagnetic Phase Transition . . . . . . . . . . . . . . . . . . . . 5
1.3.4 First-order Phase Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.5 Tricritical Phase Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Fluctuations and Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4.1 Correlation Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4.2 Additional Critical Exponents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4.3 Universality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5.1 Relations Between Critical Exponents . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5.2 Scaling Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5.3 Derivation of Exponent Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5.4 Critical Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6 Renormalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6.1 Scale Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2 Statistical Field Theory 17

2.1 Landau-Ginzburg Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Saddle-point Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

1
1 Phase Transitions
1.1 Classification
As a first step towards studying phase transitions, we introduce an order parameter, η, which is a variable
that quantifies the degree of order in a system. An example of an order parameter is the magnetization,
M , in the ferromagnetic state. η changes continuously on heating, up to a transition temperature, Tc ,
at which point the system undergoes a phase transition. The way that η changes at the phase transition
defines two types of transition (see Figure 1):
(1) η changes discontinuously to zero at Tc = first-order.
(2) η changes continuously to zero at Tc = second-order.
Second-order phase transitions are often accompanied by extreme behaviour of some physical and
thermodynamic properties (frequently with the divergence of some properties such as the heat capacity).

Figure 1: Variation of the order parameter for first-order and second-order phase transitions.

1.2 Thermodynamics of Phase Transition

The relevant quantity to consider is the Helmholtz free energy, which is defined via

F ≡ E − TS

The advantage of the free energy is that a global minimum in the the free energy function for all thermo-
dynamic variables defines a state of thermodynamic equilibrium. The phase transition determines how
the state of equilibrium changes as we scan through changes in external variables such as temperature or
pressure.
We assume that the free energy can be written as a function of the order parameter, i.e. F (η) (we
suppress the dependence on temperature for clarity), so that F (η) will have a minimum at the equilibrium
value of η for a particular temperature. For T > Tc , the equilibrium value of η is zero, so F (η) has a
minimum at η = 0, while for T < Tc , the equilibrium value of η is ±η0 , so F (η) has two minima at
η = ±η0 . At a second-order phase transition, the point F (η) = 0 will switch from a minimum to a
maximum at T = Tc (see Figure 2, left side). For a first-order phase transition, the discontinuous jump
in the equilibrium value of η at T = Tc means that we would expect the function F (η) to have minima
of equal depth at both η = 0 and η = ±η0 exactly at the transition temperature Tc (see Figure 2, right
side).

1.2.1 Definitions in Ferromagnetic Case

In the case of the ferromagnetic phase transition, the order parameter is the magnetization, M , which
we define here as simply the total magnetic moment (not the total magnetic moment per unit volume,
as is usually the case). The magnetization couples to the magnetic field, B (note that we are assuming

2
Figure 2: Variation of the free energy with order parameter at temperatures above, exactly at, and below the transition
temperature for second-order and first-order phase transitions.

the vectors M and B are parallel, so that we can simply use M and B), and we define the (isothermal)
susceptibility via  
∂M
χT ≡
∂B T
rather than the more usual definition, χT ≡ (∂M/∂H)T .
Noting that the Helmholtz free energy, F = E − T S, the corresponding exact differential is

dF = dE − T dS − SdT

If we only consider magnetic work, then the fundamental thermodynamic relation is

dE = T dS + BdM

which, after substituting in the above, yields

dF = −SdT + BdM

i.e. F ≡ F (T, M ), which allows us to define the field and susceptibility via
     2 
∂F ∂B ∂ F
B≡ and χ−1 T ≡ ≡
∂M T ∂M T ∂M 2 T

(Note that is also possible to start with dE = T dS −M dB, so that dF = −SdT −M dB and F ≡ F (T, B),
which corresponds to a different choice of thermodynamic variables - see Mandl or Kittel for details.)
Since at a second-order phase transition, F (M = 0) switches from a minimum to a maximum at
T = Tc , this means that ∂ 2 F/∂M 2 = 0 at T = Tc , and hence χT → 0 as T → Tc from below.

1.3 Landau Theory of Phase Transitions

1.3.1 Second-order Phase Transitions
Landau’s key idea was to expand the free energy as a Taylor series (i.e. power series) in terms of the two
variables, η and T , about η = 0 and T = Tc (see Boas, p.149; Stanley pp.168-9):

F (T, η) = F0 (T ) + F2 (T )η 2 + F4 (T )η 4 + . . .

where we have included only terms with even powers of η, since for second-order phase transitions we
always have F (η) = F (−η), and the ‘coefficients’ in each term are themselves power series in T − Tc :

Fi (T ) = fi0 + fi1 (T − Tc ) + fi2 (T − Tc )2 + . . .

3
where fij are constants.
We then make the following assumptions:
(1) Since we are interested in the region near the critical point, both η and T − Tc can be assumed
small, so that we may truncate the series for F (T, η) (and the expansions for the coefficients) after the
first few terms.
(2) In the coefficient of the term in η 2 , we choose f20 = 0 (see Stanley, P.169) and neglect terms above
first-order in T − Tc , letting f21 (T − Tc ) ≡ 21 a(T − Tc ), where a is a positive constant.
(3) In the coefficient of the term in η 4 , we discard all terms except the constant term, f40 ≡ 14 b.
If we truncate the series at the fourth-order term, then the constant coefficient b needs to be positive
for all T so that F (T, η) has the desired behaviour of being positive and increasing for all |η|. The
coefficient a(T − Tc ) can be of either sign. If a(T − Tc ) > 0, then
p F (T, η) has a single minimum at η = 0,
but if a(T − Tc ) < 0, then F (T, η) has two minima at η = ± −a(T − Tc )/b. We can therefore see how
a phase transition can arise from a(T − Tc ) changing sign at the transition temperature Tc .
To summarize, we now write the free energy function in the form

F (T, η) = F0 (T ) + 21 a(T − Tc )η 2 + 41 bη 4

and in what follows we shall drop the first term, F0 (T ), since it contains no information about the phase
transition.

1.3.2 Critical Exponents

The state of equilibrium is that corresponding to the minimum of the free energy, so that the equilibrium
value of η at any T is found by solving
 
∂F
= a(T − Tc )η + bη 3 = 0
∂η T

For a(T − Tc ) > 0 we know that there is a single minimum

p at η = 0, i.e. η(T > Tc ) = 0.pFor a(T − Tc ) < 0
on the other hand, there are two minima at η = ± a(Tc − T )/b, i.e. η(T < Tc ) = ± a/b(Tc − T )1/2 .
The relationship η ∝ (Tc − T )β , where β = 21 , from Landau theory (and any mean-field theory in
general) is characteristic of a type of equation that will often be encountered for physical quantities in
such a theory, and the exponent β is known as a critical exponent.
To determine the effect of an external field, Φ, on the order parameter, η, at the transition temperature,
Tc (N.B., for the magnetic phase transition, Φ = B and η = M ), we can add a field-dependent term to
the free energy as follows:

F (T, η) = F0 (T ) + 21 a(T − Tc )η 2 + 14 bη 4 − Φη

which defines the field as Φ = (∂F/∂η)T . At T = Tc the above expression gives

F (T = Tc , η) = F0 (Tc ) + 41 bη 4 − Φη

Minimizing w.r.t. η then yields

∂F 1
= bη 3 − Φ = 0 =⇒ η 3 = Φ =⇒ η ∝ Φ1/3
∂η b

We can express this as η ∝ Φ1/δ , where δ = 3 is another critical exponent.

To obtain the heat capacity, which is defined via
∂S
C=T
∂T
there are two approaches we can take. Firstly, we can start from

dF = −SdT + Φdη

from which
∂2F
   
∂F
S=− =⇒ C = −T
∂T η ∂T 2 η

4
or secondly, noting that F = E − T S, we see that S is given by the coefficient of the term in F that is
linear in T , i.e. S = − 21 aη 2 . We use this second approach, and since for T > Tc we have η = 0, this
implies that S = 0, and hence C = 0. For T < Tc , however

a2
r
a a
η= (Tc − T )1/2 =⇒ S = − 21 a · · (Tc − T ) = (T − Tc )
b b 2b

and so
a2 a2
 
∂
C=T (T − Tc ) = T
∂T 2b 2b
i.e. there is no dependence on T − Tc , only on T (no factor of T − Tc present). So, if we write the above
as (
(T − Tc )−α for T & Tc
C∼ 0
(Tc − T )−α for T . Tc
where α and α0 are the critical exponents for the heat capacity (and in general α = α0 ), then we have
α = α0 = 0.
The (isothermal) susceptibility is defined via
   2 
∂Φ ∂ F
χ−1 = =
∂η T ∂η 2 T

and hence
∂2 1
 
−1 2 1 4 ∂ 
a(T − Tc )η + bη 3 = a(T − Tc ) + 3bη 2

χ = a(T − Tc )η + bη =
∂η 2 2 4 ∂η

For T > Tc we have η = 0, and so χ−1 = a(T − Tc ). For T < Tc , however

r
a a
η= (Tc − T )1/2 =⇒ χ−1 = a(T − Tc ) + 3b · · (Tc − T ) = 2a(Tc − T )
b b

both of which imply that as T → Tc , χ−1 → 0 and so χ → ∞ (as in many second-order phase transitions).
If we express the above result as
(
(T − Tc )−γ for T & Tc
χ∼ 0
(Tc − T )−γ for T . Tc

where γ and γ 0 are the critical exponents for the susceptibility (and in general γ = γ 0 ), then we have
γ = γ 0 = 1.

1.3.3 Application to Ferromagnetic Phase Transition

We now determine the free energy expansion within Landau theory for a ferromagnet, by applying mean-
field theory to the Ising model for a system of classical spins in the presence of an external magnetic field.
The total Hamiltonian in this case can be expressed as
X X
Ĥ = −J si sj − µB si
hi,ji i

where hi, ji indicates that the sum is over nearest neighbours only, and si = ±1. We also assume that there
are N lattice sites (labelled by i = 1, . . . , N ), and that each one of these sites has q nearest neighbours.
(Note that some authors take B to be in units of the spin magnetic moment, µ, which removes the factor
of µ from the second term above.)
As our order parameter we choose the mean spin for a single site, or magnetization (note that this
differs from the standard definition of the term), which for the Ising model is given by
1 X
m̄ = hsi i
N i

5
The above implies that the total magnetization (with the term used in its conventional sense), M = N µm̄.
We can then express the interactions between neighbouring spins in terms of their deviation from the
average spin, m, i.e.

si sj = [m̄ + (si − m̄)] [m̄ + (sj − m̄)] = m̄2 + m̄(si − m̄) + m̄(sj − m̄) + (si − m̄)(sj − m̄)

In the mean-field approximation, we assume that the fluctuations of spins away from the average, e.g.
(si − m̄), are small, and hence we can neglect the last term above. The Hamiltonian then becomes
X X
m̄(sj + sj ) − m̄2 − µB

Ĥ = −J si
hi,ji i

Carrying out the summation over nearest neighbours for the N lattice sites yields
X X X X
[m̄(sj + sj )] = 21 m̄qsi + 21 m̄qsj = m̄qsi
hi,ji i j i

X
m̄2 = 12 N q m̄2
hi,ji

where the factors of 12 arise because we only want to count each pair of spins once, and simply summing
over lattice indices i and j leads to double counting of such pairs when i ↔ j. Hence we obtain
1 X
Ĥ = JN q m̄2 − (Jq m̄ + µB) si
2 i

This is just the familiar result from the Weiss theory of ferromagnetism, i.e. a particular spin feels the
average effect of its nearest neighbours’ spins via a magnetization term in an ‘effective’ magnetic field,
Beff = B + Jq m̄/µ.
The next step is to write down the partition function:
" #
X X X
−βH 1 2
Z= e = exp − 2 βJN q m̄ + β(Jq m̄ + µB) si
si =±1 si =±1 i

(N )
 X Y
= exp − 21 βJN q m̄2

exp [β(Jq m̄ + µB)si ]
si =±1 i=1

N
 1 2
Y
= exp − 2 βJN q m̄ {exp [β(Jq m̄ + µB)] + exp [−β(Jq m̄ + µB)]}
i=1

N
= exp − 21 βJN q m̄2 {2 cosh [β(Jq m̄ + µB)]}
 

from which the free energy follows via

F = −β −1 ln Z = 12 JN q m̄2 − N β −1 ln {2 cosh [β(Jq m̄ + µB)]}

In a result that will prove useful later, we note in passing that the thermal average of the spin at site i
can now be expressed as
1  −βE ↓ ↑

hsi i = e + e−βE = tanh(βµBeff )
Z
Since there is no phase transition when B 6= 0, we focus on the case where B = 0, in which case the
expression for the free energy above simplifies to

F = 12 JN q m̄2 − N β −1 ln(2 cosh βJq m̄)

Using the Taylor expansion

ln(2 cosh x) = ln 2 + 21 x2 − 1 4
12 x + O(x6 )

6
and assuming x is small, we can rewrite the above as

F ' 12 JN q m̄2 − N β −1 ln 2 − 12 N β −1 (βJq m̄)2 + 1

12 N β
−1
(βJq m̄)4

' −N β −1 ln 2 + 21 JN q(1 − βJq)m̄2 + 1 3 4

12 JN q(βJq) m̄

This is the same form as the Landau expansion from earlier, with a second-order coefficient that switches
from negative to positive at a transition temperature given by Tc = Jq, and a positive fourth-order
coefficient. Hence, applying the mean-field approximation to the Ising Hamiltonian has resulted in a
standard Landau second-order phase transition.

1.3.4 First-order Phase Transitions

In the Landau expansion for the free energy, if the coefficient of the fourth-order term is negative, we
have a first-order phase transition. In this case, we also need to include a sixth-order term with positive
coefficient to ensure that F (T, η) remains positive for large |η|, i.e. to ensure stability. Dropping the
‘constant’ term in the expansion, we therefore have

F (T, η) = 12 a(T − T0 )η 2 − 14 bη 4 + 16 cη 6

where we have replaced Tc with T0 for reasons that will become clear. As shown in Figure 2, at the
transition temperature Tc we have three minima in F (T, η) with the same value of F (in this case, F = 0,
since we dropped the ‘constant’ term). To calculate this temperature for the first-order transition we
note that we have F = 0 and ∂F/∂η = 0, so we simply equate these two expressions, solve for η and
substitute into either to obtain Tc . Starting from

F = 21 a(Tc − T0 )η 2 − 41 bη 4 + 61 cη 6 = 0

∂F
= a(Tc − T0 )η − bη 3 + cη 5 = 0
∂η

and equating the two yields η 2 = 3b/4c, which, when substituted back in to either of the above leads to

3b2
Tc = T0 +
16ac
For T < Tc , we can determine the equilibrium dependence of the order parameter on temperature,
η(T ), by minimizing the free energy, i.e. by solving
∂F
= a(T − T0 )η − bη 3 + cη 5 = 0
∂η
which yields (ignoring the unstable root η = 0):
p
2 b b2 − 4ac(T − T0 )
η = ±
c 2c
We can also derive other thermodynamic properties, such as the latent heat, which is given by the
change in enthalpy, H, at the phase transition. Since F = H − T S, the enthalpy is equal to the terms
independent of T in the expansion for F (T, η). Since we are neglecting the term F0 in this expansion, then
because η = 0 for T > Tc , the free energy, and hence also the enthalpy, will be zero in this temperature
range. The change in enthalpy at the phase transition, and hence the latent heat, will therefore simply
be equal to the enthalpy at T = Tc , which is given by the following terms in F (T, η):
3abT0
H = − 21 aT0 η 2 − 41 bη 4 + 16 cη 6 = −
8c
For some first-order phase transitions we no longer require the free energy to be symmetric under a
change in sign of the order parameter, i.e. F (−η) 6= F (η). In this case, the free energy is allowed to have
terms that are odd powers of η (except linear in η), for example

F (T, η) = 21 a(T − T0 )η 2 − 31 bη 3 + 14 bη 4

7
where the minus sign in front of the cubic term should be noted. At the transition temperature, this
function will now have two minima rather than three - one at η = 0 and another at nonzero η. As before,
we can solve the equations for F = 0 and ∂F/∂η = 0 simultaneously, but in this case we obtain the
transition temperature as
2b2
Tc = T0 +
9ac

1.3.5 Tricritical Phase Transitions

There are many materials for which external variables, such as pressure, and internal variables, such as
chemical composition, can change the fundamental parameters that determine the material’s properties,
including the coefficients in the Landau expansion of the free energy. One such example is where a change
in the variable leads to a change in the sign of the coefficient b. The case when b = 0 is then exactly on
the border between the phase transition being first-order or second-order, and such a phase transition is
known as a tricritical phase transition.
In such a case, the free energy can be written as

F (T, η) = 12 a(T − Tc )η 2 + 61 cη 6

and minimization of this w.r.t. η yields

 a 1/4
η= (Tc − T )1/4
c
so that the critical exponent, β = 41 , is a factor of two smaller than for a second-order transition.
To determine the heat capacity, we begin by writing the entropy:
1/2
a3

1
S = − 12 aη 2 = − (Tc − T )1/2
2 c

from which we obtain 1/2

a3

∂S 1
C=T = (Tc − T )−1/2
∂T 4 c
The critical exponent, α = − 21 , so that the heat capacity now diverges at a tricritical phase transition.
Finally, for the (isothermal) susceptibility we have
 2 
−1 ∂ F
χ = = a(T − Tc ) + 5cη 4 = 4a(Tc − T )
∂η 2 T

so that the critical exponent, γ = 1, which is the same as for a second-order phase transition.
Introducing an external field via

F (T, η) = 21 a(T − Tc )η 2 + 61 cη 6 − Φη

we find that minimizing F (T, η) w.r.t. η at T = Tc yields

 
∂F Φ
= cη 5 − Φ = 0 =⇒ η 5 = =⇒ η ∝ Φ1/5
∂η T c

so that the critical exponent, δ = 5.

Critical exponents for the Landau theory of second-order and tricritical phase transitions are compared
in Table 1.

1.4 Fluctuations and Correlations

1.4.1 Correlation Function
The mean-field theory considered so far has not taken into account fluctuations in the order parameter,
i.e. deviations from its mean value over a certain region - it has simply replaced the true value of η
at each point with its mean value. However, as a material is cooled towards a phase transition, regions

8
 Table 1: Comparison of critical exponents for Landau theory of second-order and tricritical phase transitions.

Exponent Second-order Tricritical

α 0 1/2
β 1/2 1/4
γ 1 1
δ 3 5

with the same value of order parameter (i.e. locally-ordered regions) begin to form, so that the local
environment seen by a particular site might be quite different from the global mean; this is not accounted
for in mean-field theory.
Consider a system of spins in a ferromagnet, such that the order parameter is the average spin (or
magnetization), which is zero above Tc . Below Tc , however, fluctuations start to occur as a result of
the interaction energy between individual spins (which tends to align them) winning out over thermal
energy (which tends to make them disordered) - i.e. regions are formed over which the spins point in the
same direction, or are correlated. The characteristic length scale of these regions (that is, of a typical
fluctuation) is known as the correlation length, ξ, and this is illustrated in Figure 3.

Figure 3: Ferromagnetic fluctuation of a size given by the correlation length, ξ.

A measure of the degree to which fluctuations in the order parameter at two particular sites (rather
than over a whole region) are correlated is given by a correlation function. For two spins, si and sj , we
define the two-point connected correlation function via
G(2)
c (ri − rj ) ≡ hsi sj i − hsi ihsj i

where ri and rj are the positions of si and sj , respectively (with |ri − rj | ≡ r) and h· · · i indicates a
(2)
thermal average. (Note that Gc (r) = hsi sj i − hsi2 for identical sites.) For T > Tc (and B = 0), hsi i = 0,
and the correlation function reduces to the usual form, hsi sj i. However, for T < Tc , subtracting off the
mean value hsi ihsj i allows us to see just the fluctuations away from the average behaviour.
We can gain some insight into the form of the correlation function for, as an example, a magnetic
system by extending the mean-field theory approach in the following ways:
(1) We now consider an external field, B, which is non-uniform, i.e. B = {Bi }, where Bi denotes the
field at site i.
(2) Rather than assuming that the q nearest neighbours of the site i have an average spin, m̄, calculated
from the spins of the whole sample, we now average only over the spins of the nearest neighbours, which
leads to the following new order parameter:
1X
m̄0 = hsj i
q
hi,ji

9
This order parameter will vary in space, and we make the further assumption that this spatial variation
is slow.
The thermal average of the variable si is given by the expression derived earlier:

hsi i = tanh(βµBeff )

where we now have

Jq m̄0 J X
Beff = Bi + = Bi + hsj i
µ µ
hi,ji

It can then be shown (see Pathria, pp.351-2) that the correlation function for the system has the form
(d−2)/2
a2
  
r
G(2)
c (r) ∼ K(d−2)/2
ξr ξ

where a is the inter-site distance (‘lattice constant’), d is the dimensionality of the system, Kµ (x) is a
modified Bessel function, and the correlation length ξ can be expressed in terms of the reduced temperature,
t ≡ (T − Tc )/Tc , as follows:
 c 1/2
ξ=a
t
where c is a number of order unity.
For x  1 (i.e. r  ξ), we have Kµ (x) ∼ x−1/2 e−x , so that the above yields

ad−2
G(2)
c (r) ∼ e−r/ξ (r  ξ)
ξ (d−3)/2 r(d−1)/2
for all d > 2.
On the other hand, for x  1 (i.e. r  ξ), we find that Kµ (x) ∼ x−µ for µ > 0, so that for d > 2:

ad−2
G(2)
c (r) ∼ (r  ξ; d > 2)
rd−2
while for the special case of d = 2 we obtain instead:
 
ξ
G(2)
c (r) ∼ ln (r  ξ; d = 2)
r

Finally, we note that the original expression simplifies considerably when d = 3, in which case
K1/2 (x) = (π/2x)1/2 e−x , which yields
 a
G(2)
c (r) ∼ e−r/ξ

d=3 r
which is true for all r, a result which was first derived by Ornstein and Zernike.
It is clear that the quantity ξ in the above expressions gives a measure of the distances over which
the spin-spin correlations in the system extend - hence the name ‘correlation length’. As long as T is
significantly above Tc , i.e. t ' 1, the expression ξ = a(c/t)1/2 tells us that ξ = O(a), as we would expect.
However, as T approaches Tc , t → 0 and hence ξ increases indefinitely, ultimately diverging at T = Tc .
This is a power-law singularity i.e.
ξ ∼ at−1/2 (T & Tc )
and the fact that ξ → ∞ at T = Tc indicates that correlations extend all over the system, paving the
way for the emergence of long-range order (even though the microscopic interactions are short-ranged).

1.4.2 Additional Critical Exponents

Although according to mean-field theory, the divergence of ξ at T = Tc is governed by a power law with
a critical exponent 21 , we anticipate that the experimental data on real systems may not obey this law.
We allow for this by introducing a new critical exponent, ν, such that

ξ ∼ t−ν (T & Tc , B = 0)

10
 (2)
As far as the correlation function is concerned, as t → 0, and hence ξ → ∞, the behaviour of Gc (r)
corresponds to
ad−2
 
(2) (2) ξ
Gc (r) ∼ d−2 (d > 2) and Gc (r) ∼ ln (d = 2)
r r
(2)
However, there is an exact result for Gc (r) at T = Tc for the 2D Ising model, according to which
−1/4
G(2)
c (r) ∼ r

which is quite different from the logarithmic mean-field prediction for d = 2. We therefore generalize our
result to cover all d as follows:
−(d−2+η)
G(2)
c (r) ∼ r (T = Tc , B = 0)

where we have introduced a final critical exponent, η.

1.4.3 Universality
Since correlations extend over macroscopic distances in a system, any structural details that differentiate
one system from another at the microscopic level lose significance. This has the remarkable result that
phase transitions in completely different physical systems can be described by the same set of critical
exponents, a phenomenon known as universality. The critical exponents do not depend on details of the
Hamiltonian or material parameters, but only on certain very general properties such as the fundamental
symmetries of the system. Hence we can assign each system to a universality class in such a way that any
two systems in the same universality class have the same dimensionality, d (usually 3), order parameters
of the same dimensionality, D, and the same critical exponents.

1.5 Scaling
1.5.1 Relations Between Critical Exponents
Critical exponents can be measured experimentally, obtained via computer simulations, or in some cases,
calculated analytically. The results for various models are shown in Table 2.

Table 2: Critical exponents obtained from various models.

Model α β γ δ ν η
2nd-order mean-field 0 1/2 1 3 1/2 0
Tricritical 1/2 1/4 1 5 1/2 0
2D Ising 0 1/8 7/4 15 1 1/4
2D 3-state Potts 1/3 1/9 13/9 14 5/6 4/15
2D XY 1 1/16 7/8 15 1/4 1/4
3D Ising ∼1/8 ∼5/16 ∼5/4 ∼5 ∼5/8 ∼0
3D XY ∼0 ∼0.34 ∼1.31 ∼4.92 ∼0.66 ∼0
3D Heisenberg ∼-0.11 ∼0.36 ∼1.39 ∼4.86 ∼0.70 ∼0.03

The six critical exponents turn out not to be independent of one another; in fact, it was noted by
various people in the 1960s, using basic thermodynamics and several assumptions, that the six exponents

11
satisfy a number of inequalities, e.g.

α0 + 2β + γ 0 > 2 (Rushbrooke)

α0 + β(δ + 1) > 2 (Griffiths)

γ 0 > β(δ − 1) (Griffiths)

(2 − α)(δ − 1)
γ> +1 (Griffiths)
δ
γ 6 (2 − η)ν (Fisher)

dν > 2 − α (Josephson)

d(δ − 1)
2−η 6 (Buckingham-Gunton)
δ+1

1.5.2 Scaling Hypothesis

In order to provide a theoretical basis for the above relations between the critical exponents, we use an
approach known as scaling, whereby the number of thermodynamic variables is reduced by defining new
variables that include a factor of some power of the reduced temperature, t. We start by returning to
the Landau expression for the free energy of a magnetic system (minus the ‘constant’ term), which we
rewrite in the form
F (T, M ) = ptM 2 + qM 4 − BM
where we have introduced new (constant) coefficients, p and q, in order to simplify what follows. At
equilibrium, the condition (∂F/∂M )T = 0 allows us to obtain an equation of state relating the three
variables, M , T and B:
B(T, M ) = 2ptM + 4qM 3
In the simple case of B = 0, differentiating the free energy, equating to zero and solving for M allows
us to obtain the expression
  1/2
p
M =± |t| (t . 0)
2q
We now try to obtain a similar expression for the case B 6= 0.
We begin by rewriting the equation of state in the form
" 3 #
p3/2 3/2
 1/2   1/2
q M q M
B(T, M ) = 1/2 |t| 2 sgn(t) +4
q p1/2 |t|1/2 p1/2 |t|1/2

where the function sgn(t) is needed because we have |t| everywhere, and we need to allow for t < 0.
Rearranging yields
3
q 1/2 B
 1/2   1/2
q M q M
= 2 sgn(t) + 4
p3/2 |t|3/2 p1/2 |t|1/2 p1/2 |t|1/2
The result is that we have replaced the relationship among three variables (M, B, T ) in the original form
of the equation of state with one between two new variables (M/|t|1/2 , B/|t|3/2 ) by scaling M with |t|1/2
and B with |t|3/2 .
If we define
q 1/2 B q 1/2 M
x ≡ 3/2 3/2 and y ≡ 1/2 1/2
p |t| p |t|
then the above can be written as
x = 2 sgn(t)y + 4y 3
which in turn implies that y must be a function of x, or equivalently

q 1/2 B
 1/2 
q B
= function of
p3/2 |t|3/2 p3/2 |t|3/2

12
which means that
p1/2 1/2 q 1/2 B
 
M= |t| × function of
q 1/2 p3/2 |t|3/2
We can rewrite the expression for the free energy in terms of the new scaled variables as follows:
" 2  1/2 4  1/2 #
p2 2
 1/2   1/2
2 4 q M q M q B q M
F (T, M ) = ptM +qM −BM = t sgn(t) + −
q p1/2 |t|1/2 p1/2 |t|1/2 p3/2 |t|3/2 p1/2 |t|1/2

and substituting in the above expression for M then yields

p2 q 1/2 B
 
F (T, B) = t2 × function of
q p3/2 |t|3/2

We have therefore replaced the expression for F in terms of the three variables (M, B, T ) with one in
terms of a single scaled variable, B/|t|3/2 (since M is dependent on B). Note that the functions appearing
on the RHS of the scaled expressions for M and F are universal, i.e. they have the same form for all
systems described by the Landau theory.
The next step is to generalize the mean-field expression for the free energy above to an arbitrary
system by introducing the so-called scaling hypothesis, developed by Widom and Kadanoff. This consists
of writing the free energy in the form
 
B
F (T, B) = P |t|2−α f Q ∆
|t|

where α and ∆ are universal numbers common to all systems in the given universality class, f (x) is
a universal function which is expected to have two different branches, f+ for t > 0 and f− for t < 0,
while P and Q (like p and q) are non-universal parameters characteristic of the particular system under
consideration. Note that F in this form (as well as the scaled forms of M and F in the mean-field theory
above) is an example of a homogeneous function (see Mathematics notes), since it has been expressed in
terms of a new function (i.e. f ) that depends on one less (scaled) variable than the original function.
We expect α and ∆ to determine all the critical exponents of the problem, while the amplitudes
appearing in the various power laws governing the behaviour near T = Tc will be determined by P , Q
and the limiting values of the function f (x) and its derivatives (as x → 0, i.e. as B → 0). We note
that the exponent of |t| outside the function f (x) has been chosen to be (2 − α), rather than 2 as in
the corresponding mean-field expression, to ensure that the singularity in the specific heat is correctly
reproduced (when taking the second derivative of the free energy); also, we make the assumption in what
follows that α = α0 and γ = γ 0 .
We can write down a similar expression for a scaled form of the correlation function:

g(rtν , B/t∆ )
G(2)
c (r; T, B) =
rd−2+η
where the function g(x) is again universal for a given universality class, and we have suppressed non-
universal parameters that vary from system to system within a class for clarity. Note that the natural
variable for r to scale with in g(x) is the correlation length, ξ ∼ t−ν . In the absence of the field (B = 0),
the above reduces to
g0 (rtν ) −(d−2+η)
G(2)
c (r; T ) = −→ G(2)
c (r) ∼ r (T = Tc , B = 0)
rd−2+η

1.5.3 Derivation of Exponent Relations

1.5.4 Critical Dimension

1.6 Renormalization
1.6.1 Scale Invariance
We consider a 2D ferromagnetic material in zero field, which can be described by the Ising model on a
square lattice. Such a model can be simulated on a computer, with black and white squares representing

13
 Figure 4: Computer simulation of system of spins described by Ising model on a square lattice at T  Tc .

opposite spins. At temperatures much greater than Tc , the thermal energy (which tends to randomize
the spins) dominates the exchange energy (which favours spin alignment), and although some nearest-
neighbour spins tend to lie parallel, this is a small perturbation on the random configuration (see Figure
4.
As the temperature is lowered, the effects of the exchange interaction become more apparent, with
nearest-neighbour spins more likely to point in the same direction, and clusters of aligned or correlated
spins appear. The size of the largest clusters is measured by the correlation length, which for say
T = 1.22Tc is of the order of a few lattice spacings - the system is then said to show short-range order
(see Figure 5).

Figure 5: Computer simulation of system of spins at T = 1.22Tc , showing short-range order.

As the temperature is lowered further, the correlation length increases, although fluctuations (i.e.
deviations of the spin from the average value over a certain regions) on a smaller scale remain important.
In fact, there are correlated regions of spins at all length scales up to that determined by the correlation
length. Each fluctuation is not actually a region of uniform spin alignment, but includes smaller fluctu-
ations, which in turn include still smaller ones, etc., down to the length scale set by the lattice spacing.

14
At T = Tc the correlation length becomes infinite, so that ordered structures (i.e. fluctuations) exist
at every length scale with no upper cutoff value (see Figure 4). This is the essence of a critical phase
transition: fluctuations at all length scales are important.

Figure 6: Computer simulation of system of spins at T = Tc , showing order at all length scales.

Below the critical temperature, say at 0.99Tc , there is a non-zero magnetization, with more spins
lying in one of the two spin states than the other. The model is then said to exhibit long-range order
(see Figure 7).

Figure 7: Computer simulation of system of spins at T = 0.99Tc , showing long-range order.

As we have seen, at a critical point all length scales are important, so that the physics is invariant
under scale transformations - a phenomenon known as scale invariance. Since the simplest mathematical
functions that are scale-invariant are power laws, this goes some way towards explaining why the critical
exponents of a phase transition (i.e. the power laws that predict the limiting behaviour of physical
properties at or near the transition) are so important in characterizing that transition.
In order to study scale invariance at a critical point, we can change the scale of system and see how
it behaves. This is done by taking a block of spins and replacing it by a single spin that takes the same
value as the majority of spins in the original cluster. This ‘zooming out’ procedure, which reduces the
scale of the system by a factor b, is known as renormalization. (If we start with a group of nine spins and
replace it with a single spin, then b = 3.) We then repeat the process to produce a series of snapshots
of the spin configuration, essentially seen under progressively decreasing magnification. If the correlation
length changes under successive iterations of this process, then that corresponds to an inverse change in
the temperature of the system.

15
 For a starting temperature above the critical temperature (see Figure 8), the scale change soon
obliterates any short-range order and the spins on the renormalized lattices become uncorrelated. This
corresponds to an infinite temperature: the system has been renormalized by the simple transformation
we have defined to T = ∞. This will be the case for all temperatures above Tc , and the nearer to the
critical temperature the starting point is, the more steps of the transformation it will taken to lose the
short-range order.
For temperatures below the critical temperature, there is an analogous change as renormalization is
carried out. However, now any fluctuations are relative to the ground state, and, as these are lost under
renormalization, the system moves to a completely ordered state characteristic of zero temperature (see
Figure 9).
Only at the critical temperature itself, where there are fluctuations on all length scales, does the
system remain invariant under renormalization (see Figure 10). This can be exploited to identify the
critical point and describe the behaviour of the thermodynamic functions in its vicinity.

Figure 8: Sequence of renormalized spin configurations at T > Tc , showing loss of short-range order.

Figure 9: Sequence of renormalized spin configurations at T < Tc , showing move to completely ordered state.

16
 Figure 10: Sequence of renormalized spin configurations at T = Tc , showing invariance under transformation.

2 Statistical Field Theory

2.1 Landau-Ginzburg Theory
In the Ising model for a system of classical spins on a lattice, we have previously defined the ‘magnetiza-
tion’ (average spin) via
1 X
m̄ = hsi i
N i
which is the equilibrium value calculated from the thermal averages of each spin. For what follows, we
shall find it convenient to define a slightly different expression for the average spin:
1 X
m= si
N i

which is now an average over a particular configuration of spins, and lies in the range −1 6 m 6 1. If we
introduce the notation {si }|m to mean the set of all configurations of spins such that their average value
m is given by the above expression, then we may write the partition function for the system as follows:
X X X
Z= e−βE[si ] ≡ e−βF (m)
m {si }|m m

where the double summation means that we first sum over all configurations of spins that correspond
to a magnetization m, and subsequently sum over all possible m, while E[si ] represents the energy of a
particular configuration. We have also defined an effective free energy, F (m), which depends on both
T and B, as well as the magnetization, m. Note that this is not the same as the thermodynamic free
energy, which is defined only in equilibrium.
Strictly speaking, m takes only discrete values (quantized in units of 1/N , but for the limit of large
N we can rewrite the above summation as an integral, i.e.
Z 1
Z= dm e−βF (m)
−1

In the Landau mean-field theory considered in the previous Section, the magnetization was assumed
to have the value, m̄ (or m in our new definition of the average spin) throughout the system. However,
what happens at or near the critical point depends on fluctuations away from the average spin, i.e on
spatial variations, and hence we need to generalize the Landau theory to allow the system to move away
from a homogeneous state - such a generalization is known as Landau-Ginzburg theory. The key idea is to

17
take the order parameter from Landau theory, which in the case of the Ising model is the magnetization
m, and promote it to a field which can vary in space, m(x) - this is referred to as a local order parameter.
As a first step, we return to the discrete lattice and divide it up into boxes with length of side b (with
b assumed smaller than the correlation length), where each of these boxes contains N 0 lattice sites (see
Figure 11). For each box, we define a magnetization (average spin) via

Figure 11: Division of lattice into boxes with side length b, each containing N 0 lattice sites.

1 X
m(x) = si
N0 i

where x is the location of the centre of the box, a procedure known as coarse-graining.
We can now write down the partition function of the system as
X X X
Z= e−βE[si ] ≡ e−βF [m(x)]
m {si }|m(x) m(x)

where {si }|m(x) is now the set of all configuration of spins that yield m(x) after coarse-graining.
We now make two assumptions in order to transform this into an integral: (1) Since m(x) takes
values only at certain values of x that label the positions of boxes, we assume that the number of boxes,
given by N/N 0 , is large enough that we may treat x as a continuous variable; (2) Since at each x, the
magnetization m(x) is quantized in units of 1/N 0 , we assume that N 0 is large enough for m(x) to be
treated as a continuous function.
In preparation for using our formalism to describe systems beyond the ferromagnetic one consid-
ered above, we also generalize the field to have more than one component, and write it as m(x) ≡
(m1 , m2 , . . . , mn ), where it is important to note that n is not in general the same as the dimensionality,
d, of the system. Specific physical phenomena/systems that are described by such a multi-component
order-parameter field are

n = 1 : liquid-gas transitions, binary mixtures, uniaxial magnets

n = 2 : superfluidity, superconductivity, planar magnets

n = 3 : classical magnets

In the above case of our uniaxial magnet, n = 1 and m(x) = m(x)ẑ, so that the field has just one
component.
Taking into account the above assumptions and generalizations, we may therefore now express the
partition function as a functional integral :
Z
Z = Dm(x) e−βF [m(x)]

where F [(x)] is now a functional, known as the Landau-Ginzburg free energy, and the notation Dm(x)
means that we should sum over all field configurations m(x). Since the partition function is now expressed

18
in terms of a field, rather than the more usual classical variables, the study of integrals such as the above
is known as statistical field theory.
To obtain an expression for the general form of the free energy, we first note a number of constraints
that arise from its microscopic origin:
(1) Locality and uniformity: In our discrete model, if the spins are assumed to be non-interacting, the
free energy will be given as a simple sum of local functionals, one for each location x. In a continuous
system, the sum becomes an integral and we can write
Z
F [m(x)] = dd x f [m(x)]

where f [m(x)] is the free-energy density. If we then introduce short-range interactions between spins, we
find that f becomes dependent on ∇m(x) and higher derivatives, which control how the field m(x) at
one point affects the field at neighbouring points. The result of assuming such local interactions is that
the free energy should take the form
Z
F [m(x)] = dd x f [m(x), ∇m(x), ∇2 m(x), . . .]

Note also that we are assuming that the f has no explicit dependence on x, i.e. that the material
is uniform in space. This will not be the case, however, when the system is in an external potential or
contains impurities.
(2) Symmetries One thing that survives the averaging (coarse-graining) process is the underlying
microscopic symmetries of the system, which therefore constrain the possible form of the effective free-
energy expansion. For example, in the absence of a magnetic field, all directions of magnetization are
equivalent, and so the free energy must be invariant under rotations in order-parameter space, Rn , i.e.
we must have F [Rn m(x)] = F [m(x)]. (In our uniaxial ferromagnet, for which n = 1 and m(x) = m(x)ẑ,
this corresponds to transforming spin-up, or +m(x), to spin-down, or −m(x), or vice versa.) A term
linear in m(x) in the free-energy expansion is not consistent with this symmetry (for example, R1 m(x) =
m(x) 6= m(x)), and the lowest-order term allowed is quadratic in m(x), i.e. proportional to:
n
X
m2 (x) ≡ m(x) · m(x) = mi (x)mi (x)
i=1

since, for example, R1 m2 (x) = m2 (x). By extension, terms in even powers of m(x) are allowed, while
those in even powers of m(x) are forbidden. However, for non-zero magnetic field, we can have terms
proportional to the field multiplied by odd powers of m(x), since the effects of transforming both the field
and the order parameter cancel out (e.g. the negative signs resulting from B → −B and m(x) → −m(x)
cancel).
When we consider possible terms involving gradients of the vector field, ∇m(x), we need to consider
the spatial symmetries of the system. In an isotropic system, all directions in space are equivalent, so we
require combinations of derivatives that are invariant under spatial rotations (as opposed to rotations in
order-parameter space, already considered above). It turns out that the simplest such term is quadratic:
n X
X d
(∇m)2 ≡ ∂α mi ∂α mi
i=1 α=1

where ∂α indicates the partial derivative along the αth direction in space. Higher-order allowed terms
are those that contain even powers of ∇m or a product of the latter with even powers of m, e.g.
(∇2 m)2 , m2 (∇m)2 , . . .
(3) Analyticity and stability Although at the microscopic scale there may be non-analyticities associ-
ated with the individual spins in a system (such as when they are quantized to specific values), we assume
that the averaging process associated with defining the field m(x) smoothes these out, so that the free
energy can be expressed as an analytic expansion in powers of m(x).
In order that the exponential, exp {−βF [m(x)]}, does not diverge for infinitely large values of m(x),
we require that the coefficient of the highest power of m in the expansion of the free energy be positive
(which will ensure that F → 0 as m → ∞, since the argument of the exponential is negative). There will
be similar constraints on the signs of terms involving gradients in the expansion.

19
 Taking into account the above considerations, we are now in a position to write down the general
form of the free energy expansion, the first few terms of which are:
Z  
t 2 K
F [m(x)] = F0 (T ) + dd x m (x) + um4 (x) + (∇m)2 + · · · − B · m(x)
2 2

where the constant term F0 (T ) is often ignored, since it does not depend on the order parameter.

2.2 Saddle-point Approximation

As a first attempt at calculating the Landau-Ginzburg partition function
Z Z R d
Z = Dm(x) e−βF [m(x)] = e−βF0 Dm(x) e− d x βf [m(x)]

we make the so-called saddle-point approximation, in which we assume that the integral over m(x)
can be approximated by the value of m that minimizes F [m(x)] (and hence maximizes the probability
exp[−βF ]), i.e.
Z ' Zsp = e−βF [m0 (x)]
where the particular field m0 is that for which the free energy takes its minimum value. To determine
the form that this field should take, we return to the expansion at the end of the previous Section and
note the following:
(1) The term in ∇m will be minimized by a field that is constant in x, so we assume such a property
for m0 , and hence this term will vanish.
(2) For a given value of B 6= 0, the term in B · m will be minimized when m is parallel to B. We
therefore make the assumption that m0 = m0 ẑ.
Given the above, we may now write the expression for the minimized free energy as
 
t 2
F [m0 ] ' F0 (T ) + V m0 + um40 − Bm0
2

where V is the system volume, which has exactly the same form as the Landau mean-field expression
for the free energy derived earlier. Hence, the saddle-point and mean-field approximations are equivalent
and lead to the same critical exponents.

20