5 Phase Transitions and Critical Phenomena

5.1 The Ising Model

Ferromagnetism is an important phenomenon in the study of solids. Certain
metals, including iron, spontaneously develop a nite magnetization at ordinary
temperatures. Above the so-called Curie temperature, however, these systems
exhibit randomly oriented spins. Only below this temperature is a permanent,
spontaneous magnet formed. It is also found that the heat capacity of such
magnetic systems diverges near the Curie temperature.
The Ising model is a seemingly very simple model that was developed to
understand such behavior. Although developed specically in connection with
ferromagnetism, the Ising model has proven to be a very model that can also
be applied to such diverse systems as interacting gases, simple binary liquid
mixtures, and alloys. It also turns out to be far from simple to solve! Well
discuss one approach to solve the Ising model, as well as to understand phase
transitions in general.
The following Hamiltonian represents a simple model for a paramagnet, a
material that does not exhibit a spontaneous magnetization in the absence of
a magnetic eld, but which responds by developing a magnetization in the
presence of a eld:
H = h
N

i
s
i
, (5.1)
where s
i
= 1 for up and down magnetic spins. The number of microstates
with a given magnetization M = N

is then
=
N!
N

!N

!
. (5.2)
The entropy is then
S = Nk

log 2
1
2
(1 +s) log (1 +s)
1
2
(1 s) log (1 s)

, (5.3)
where we have used a version of Sterlings approximation
log N! N log N N, (5.4)
and the fact that N
,
= N (1 s) /2. Here, s = M/N is the average spin.
For a given magnetic eld h, the energy of a single up/down spin is
,
= h.
Since these are the only two possibilities, the probability that a single spin is
up/down is given by
e
h
/Q, (5.5)
where Q = e
h
+e
h
, and = 1/(kT). Thus, the average spin
s =
e
h
e
h
Q
= tanh(h). (5.6)
119
The important physics that this simple model leaves out, of course, is the fact
that the spins do interact. One must take interactions into account in order to
understand transitions between phases. The interactions in a ferromagnet tend
to align the a single spin with its neighbors. In the Ising model, we simply add
to the Hamiltonian above a term that favors alignment of neighboring spins:
H = h

i
s
i
J

i,j
s
i
s
j
. (5.7)
Here, we have assigned an energy J for each pair of neighboring spins, de-
pending on alignment. For J > 0, the energy is lower for spins that are aligned.
This is what we expect for ferromagnets. The sum in this expression is over all
pairs of neighboring spins (bonds). Note that each pair of spins that interact
should only be counted once. For a regular square lattice in two dimensions,
there are four nearest neighbor spins, corresponding to spins to the left and
right, as well as above and below a single spin. In d dimensions, there will be
z = 2d neighbors.
One approach to solving the Ising model is to consider the average or mean
eect of the z spins neighboring a single spin s
i
. This approach is known as
mean eld theory. We can write the Hamiltonian in the following way,
H = h

i
s
i

1
2
J

i
s
i

nn
s
nn

, (5.8)
where nn refers to the z nearest neighbors of spin s
i
. The factor of 1/2 here is
introduced to correct for double counting pairs of spins, since each spin that is
included in the rst sum will also appear in the second, although each pair of
spins should only be counted once. So far, no approximation has been made.
The idea behind mean eld theory is to treat the neighboring spins only in an
average way. The part of the energy above that involves an individual spin s
i
can be written
H
i
= hs
i
Js
i

nn
s
nn

. (5.9)
120
Note that H =

H
i
, due to double counting of spin pairs. We approximate
the term

nn
s
nn
by zs, since there are z nearest neighbors. This is a serious
approximation, but we might expect that it is not a bad approximation if the
number of neighbors is large, so the number of terms in

nn
s
nn
is large. The
result looks just like the model for a paramagnet, except with an eective mag-
netic eld h
e
= h +zJs. In other words, the eect of the surrounding spins
is similar to a molecular magnetic eld acting on spin s
i
due its neighbors.
With this approach, we obtain the average spin self-consistently using Eq.
(5.6).
s = tanh ( [h +zJs]) . (5.10)
This can be solved graphically, with the result that the only solution for h = 0
and zJ < 1 is s = 0. In other words, there is no spontaneous magnetization.
For kT < zJ, however, nonzero solutions exist. Physically, these solutions
correspond to states for which there is a spontaneous magnetization.
We can solve this model in another way, that is more instructive. Within
the mean eld approximation, the energy per spin is approximately
u = hs
1
2
zJs
2
, (5.11)
where we have dropped the brackets denoting the average spin. What about
the entropy? This we know from our solution of the paramagnet above. The
free energy per spin in this approximation is therefore approximately given by
f = hs +
kT zJ
2
s
2
+
kT
12
s
4
, (5.12)
for small s.
In the absence of a magnetic eld h, the minimum of this free energy with
respect to the average spin s occurs for s = 0 when kT > zJ. For kT <
zJ = kT
c
, there are two degenerate minima for s > 0 and s < 0. These states
correspond to spontaneous magnetization up and down. These minima
occur for
f
s
= 0. Furthermore, near the transition,
s
2
3(T
c
T)/T. (5.13)
121
Thus, near the transition the magnetization is given approximately by
M N

3(1 T/T
c
), (5.14)
provided that T < T
c
.
Near the transition, the heat capacity can also be calculated, since
U
N
2
zJs
2

3N
2
zJ (1 T/T
c
) . (5.15)
The heat capacity is nite, with a value C = 3Nk/2 at the transition. The
numerical solution of Eqs. (5.10) and (5.15) yields a heat capacity that decreases
to zero for T < T
c
.
122
It is also instructive to calculate the susceptibility within the mean eld
approximation as well. Above the transition (i.e., for T > T
c
), we nd for small
applied magnetic elds h that
s =
h
kT zJ
. (5.16)
This comes from minimizing f in Eq. (5.12) with respect to s for small h. Then,
the susceptibility diverges above the transition as

M
h
=
N
kT zJ
|T T
c
|

, (5.17)
where = 1.
Although we may, and should question the validity of the mean-eld ap-
proximation above, the results derived above are illustrative of rather general
123
behavior of systems near critical points. In particular, the continuous power-law
growth of magnetization below the transition, as well as the power-law diver-
gence of the susceptibility are characteristic of critical points. Nevertheless, it
is worth thinking ahead about what important physics this kind of approach
leaves out. By assuming that the environment of each spin s
i
can be treated by a
single number representing the same average eld, we are clearly leaving out any
possibility of uctuations. This can be an important eect, since a uctuation
that gives more up spins in one region will tend be self-reinforcing by making
neighboring spins also ip up. In fact, we might expect that this reinforcing
eect becomes most important near the critical point, since the susceptibility
, which measures how easy it is to change the magnetization, diverges! More
subtly, we will also nd that precisely near the critical point, the spatial extent
of correlations in the magnetization gets very large. I.e., the uctuations involve
more and more spins the closer one gets to the critical point. With hindsight, it
is, in fact, surprising that mean-eld theory works as well as it does. It practice,
it is still the rst approach that most people take in dealing with a new problem
in phase transitions.
5.2 Landau Theory
Arguably the rst reasonably successful and general approach to understanding
phase transitions goes back to Landau in the 1930s. The basic idea here is
simple enough; one assumes that that one can write the free energy (e.g., f
above) as a smooth function of a measurable parameter (the so-called order
parameter) describing the transition. In the case of the Ising model, the natural
order parameter is the magnetization M, or equivalently, the average spin s.
This is a particularly convenient choice, since it is zero in the disordered phase
(the paramagnet) and non-zero in the ordered phase (the ferromagnet).
By assuming that the free energy is smooth, we expect to be able to expand
this free energy in a Taylor series in the order parameter, which we will denote
as m. So, for instance, for a magnet at temperature T in the absence of a
magnetic eld, we expect to have
f(T, m) a(T) +
1
2
b(T)m
2
+
1
4
c(T)m
4
+ , (5.18)
were we have used the fact that the Ising model is symmetric under ipping
all spins (changing the sign of m), even if the equilibrium phase may develop a
spontaneous magnetization that breaks this symmetry. The free energy of any
state should be the same if m m, unless we have a non-zero magnetic eld
that breaks this symmetry. This is, of course, the (approximate!) form of the
free energy we derived above for the Ising model. Here, the various coecients
in the Taylor series are unknown functions of the remaining variable T. We do
not know what they are, but by taking a cue from the Ising model above, we
see that if c(T) > 0, then the transition will occur when b(T) changes sign. So,
if we now imagine that b(T) can also be expressed as a Taylor series about T
c
,
124
at which it changes sign, then we nd
b(T) b
0
(T T
c
) + , (5.19)
and similarly,
c(T) = c
0
+c
1
(T T
c
) + . (5.20)
Near the critical point, provided that c
0
> 0, we can nd the equilibrium by
minimizing the free energy to nd
f
m
= 0 = bm+ cm
3
+ , (5.21)
which gives
m =

b
0
/c
0
|T T
c
|

, (5.22)
where = 1/2. This is valid for temperatures just slightly below T
c
. (Apologies
for the use of for two things. But, there is simply no way around this: both uses
are too deeply imbedded in the statistical physics literature to do otherwise.)
We could also calculate the behavior of both the heat capacity C and the
susceptibility within this model (assuming an additional term hm in the free
energy for small external magnetic elds). We would nd essentially the same
results as before, in that C remains nite with a discontinuity, and diverges
with the same exponent as above. Since all we have really assumed here is
that the free energy is smooth (analytic) as a function of the order parameter
(here, m), this suggests that all phase transitions (with the same symmetry as
the magnet) should have the same behavior, e.g., in m, C, and , at least near
a critical point, where the order parameter is small, allowing for Taylor series
expansions similar to those above.
You will probably not be surprised to hear that this is false, in that = 1/2
and = 1; we have, after all, made very strong assumptions/approximations
above. What you probably will be more surprised to hear is that experiments
on phase transitions in many dierent materials show THE SAME values of
and . So, while Landau theory may fail in detail (e.g., in predicting values of
and ), it seems to get one very important qualitative feature right: critical
behavior is remarkably universal from one system to another. This universality
in itself begs an explanation. We will begin to see why this is true. And, in the
process, we will also learn something very deep about the nature of condensed
phases of matter with many degrees of freedom: most of the microscopic details,
such as the precise way in which molecules interact with each other, become
largely irrelevant in determining the way systems behave at the macroscopic
scale. In the development of the statistical mechanics approach to try to derive
thermodynamic properties, this is simply something that we hope for. Critical
phenomena provide a wonderful window into the origins of the emergence of
macroscopic simplicity out of microscopic complexityi.e., the fact that only a
few thermodynamic parameters are enough to describe the macroscopic behavior
of systems that are horribly complicated in microscopic detail. This is one of
the reasons we will spend so much time studying relatively rare critical points!
125
Before we try to go beyond Landau theory, it is worth noting a few more
things about it. First of all, it is a sort of mean-eld theory, in that we have
only accounted for a single, average value of the order parameter m, and have
not allowed for this order parameter to uctuate, either from point to point
within the sample, or from time to time at the same point. We can improve on
this limitation by allowing for a spatially varying order parameter m(x). This
will not, in the end, cure the above failures of Landau theory, although we will
learn something useful about spatial correlations.
There is another important success of Landau theory. It shows us that
symmetry alone can determine whether a phase transition is rst-order (like
the melting of ice, in which there is a latent heat) or second order (like the
development of ferromagnetism below the Curie temperature). Notice that if
we do not have the m m symmetry of the magnet, we could not assume all
odd terms in the Taylor series expansion of f vanish above. You can check for
yourselves, for instance, that with a free energy of the form
f(T, m) a(T) +
1
2
b(T)m
2
+
1
3
c(T)m
3
+
1
4
d(T)m
4
, (5.23)
the order parameter jumps discontinuously from one value to another as b(t)
decreases below the value 2c
2
/(9d). (Here, d is assumed to be positive for
stability.) While this may seem like a curiosity, it explains why the transition
between isotropic (I) and nematic (N) phases of liquid crystals is rst order: the
symmetry of liquid crystals is dierent from magnets, and permits odd terms
in the Landau theory (Chaikin and Lubensky 2000).
Even when the symmetry of the problem does not permit odd terms in the
free energy, it is still possible to have a rst-order transition. An example of
such a situation is what happens in the model above when the 4th order term
is not positive:
f(T, m) a(T) +
1
2
b(T)m
2
+
1
4
c(T)m
4
+
1
6
d(T)m
6
+ . (5.24)
Here, we assume that d > 0, which ensures that the model is stable (m does
not diverge.) By plotting this function for various values of the parameters b,
c, and d > 0, you can convince yourself that there can be a transition with
a discontinuous jump in m (sign of a rst-order phase transition). The other
interesting thing about this model is that it exhibits a line of critical points
that becomes a rst-order phase boundary. This so-called, tricritical behavior
is seen in the superuid transition of
4
He-
3
He mixtures. As
3
He is added to
4
He, the continuous (critical) transition to a superuid still occurs, but at a
lower temperature. As more
3
He is added, however, this transition becomes
rst-order (discontinuous), with coexisting
4
He-rich superuid (S) and
3
He-rich
normal uid (N). The dotted line indicates a line of critical points.
126
These examples show why Landau theory is so useful, even if it is not rigor-
ous or accurate in its predictions of critical behavior. We will later return to the
origins of this failure, and why it is particularly apparent at the critical point.
But, it is worth reecting one of the key assumptions of Landau theory, that
the free energy is analytic. Although this seems like a very innocent assump-
tion, it is actually dangerous. Imagine building up a description of increasingly
complicated systems beginning a particle at a time. The kinetic energy of a
particle is certainly analytic as a function of the momentum. Lets assume that
the interaction potential describing the way particles interact is also analytic.
As we add more and more particles, it would seem that the Hamiltonian for the
system will be analytic for any nite number of particles. It might get very hard
or even impossible to solve this problem in practice for more than a few particles
(like 3, even). But, how can the free energy not be analytic? It will turn out
that this failure of analyticity occurs (in general) in the thermodynamic limit,
N . In fact, in a sense, there IS not real critical behavior for any nite
system. More on this later. Of course, in practice, N
A
is a big number!
5.3 Critical Phenomena and Scaling
In the middle of the 20th century, a number of groups were performing ever more
precise measurements of the behavior of magnets and other systems exhibiting
critical behavior. These experiments kept pushing to temperatures closer and
closer to the critical temperature, from above (t (T T
c
)/T
c
0
+
) and below
(t 0

). In fact, these experiments can go to very low reduced temperatures

t of order 10
3
, or even 10
5
in some cases.
These experiments consistently showed behavior very similar to what mean-
127
eld theory or Landau theory predicted. Specically, it was seen that
C
h=0,t0
+ |t|

C
h=0,t0
|t|

m
h=0,t0
|t|

h0,t0
+ |t|

h0,t0
|t|

m
h0,t=0
|h|
1/
sign(h).
These are written in terms of magnetic properties. But, this generalizes easily to
other systems, such as the liquid-gas transition, where, for instance, the density
dierence between liquid and gas plays the role of the order parameter m, while
pressure is analogous to the magnetic eld. Then, for instance, expresses the
degree of pressure as a function of volume along the critical isotherm.
Mean-eld theory predicts that that =

= 0, = 1/2, =

= 1, and
= 3. Experiments on a variety of systems shows that, indeed, the exponents
above and below the transition seem to be the same, but that the values are
very dierent from those predicted by mean-eld theory. For instance, 0.1
(i.e., the heat capacity diverges!), 0.3, 1.3, and 4 5.
In order to understand this, lets begin by trying to x Landau theory. Lets
write the free energy in a simple, canonical form:
f(t, h) = hm+
1
2
btm
2
+
1
4
cm
4
, (5.25)
where we shall treat b and c as constants. The equation of state can be obtained
by dierentiation with respect to m:
h = btm+cm
3
= m(bt +cm
2
). (5.26)
As t 0
+
, we nd that m = h/(bt), which means that
=
+
t

, (5.27)
where
+
= 1/b, and = 1. Also, as t 0

, m
2
= b|t|/c, so that
|m| =

b/c|t|

, (5.28)
where = 1/2. Clearly, it is the exponent 2 in Eq. (5.15) that leads directly to
= 1/2. Perhaps we should try
h = btm+cm
3
= m(bt +cm
1/
) (5.29)
as an equation of state. Obviously, unless 1/ happens to be an even integer,
this cannot make any sense unless we replace m
1/
by |m|
1/
. Likewise, perhaps
we could correct the exponent = 1 by the equation of state
h = m(bt

+c|m|
1/
). (5.30)
128
This would give us the right divergence of as t 0
+
. But, again, if this
is to make sense, we must take the absolute value of t here in order to describe
thing below the transition. In that case, however, we need to also account for
the sign of t. Hence, we presumably should try
h = m(b|t|

+c|m|
1/
). (5.31)
Unfortunately, there are real problems with all these attempts to correct
Landau theory. Not only do we have nonanalytic behavior at the critical point
t = h = 0, but we even have nonanalytic behavior away from the critical point
along the critical isotherm (t = 0)! While divergences and even nonanalytic
behavior are expected AT the transition, there is NO experimental evidence of
any singularities on the critical isotherm for nite h.
There is another way to do this that DOES work. Beginning again with
the original, mean-eld equation of state, we can rescale the whole equation by
b
3/2
|t|
3/2
/c
1/2
:
D
h
|t|

=
m
B|t|

1 +

|m|
B|t|

1/

, (5.32)
where B =

b/c, D = c
1/2
/b
3/2
, = 1/2, and = 3/2. As strange as this
may appear, this CAN be generalized to arbitrary values of the two exponents
and in a consistent way, without introducing any singularities away from
the critical point. For general values of these exponents, as t 0
+
,
D
h
|t|

m
B|t|

. (5.33)
Thus, the susceptibility scales according to |t|

, or = .
Equation (5.32) can be inverted to obtain the reduced or normalized mag-
netization m =
m
B|t|

as a function of the reduced eld h = D

h
|t|

. This can
be done graphically for the mean-eld case of = 1/2, for which we can solve

h = m

1 + m
2

.
129
In general, we write
m = W

, (5.34)
where refer to the t > 0 and t < 0 branches of these solutions of m in terms
of

h. I.e.,
m
B|t|

= W

D
h
|t|

(5.35)
This suggests that if we plot m/|t|

versus h/|t|

, we should see all of the data

for T > T
c
collapse onto a single curve dening the function W
+
, while all of
the data for T < T
c
should collapse onto another curve dening W

, provided
that we choose the right exponents and . In fact, this works remarkably
well. Starting with raw data of the form shown below,
rescaling the data in this way results in two curves similar to the following.
You should try this yourself with classic data obtained by Weiss and Forrer
in 1926 (next page).
130
T (K) h (oe) M (emu/g)
678.86 4,160 0.649
678.86 10,070 1.577
678.86 17,775 2.776
653.31 2,290 0.845
653.31 4,160 1.535
653.31 6,015 2.215
653.31 10,070 3.665
653.31 14,210 5.034
653.31 17,775 6.145
653.31 21,315 7.211
629.53 1,355 6.099
629.53 3,230 9.287
629.53 6,015 11.793
629.53 10,070 13.945
629.53 14,210 15.405
629.53 17,775 16.474
614.41 1,820 19.43
614.41 3,230 20
614.41 6,015 20.77
614.41 10,070 21.67
614.41 14,210 22.36
614.41 17,775 22.89
601.14 3,230 25.34
601.14 6,015 25.83
601.14 10,070 26.4
601.14 14,210 26.83
601.14 17,775 27.21
Data taken from Statistical Mechanics, K. Huang (Wiley, New York, 1987).
Original data taken from P. Weiss and R. Forrer, Ann. Phys. (Paris) 5, 153
(1926). These data were reanalyzed by Kouvel and Fisher (J.S. Kouvel and
M.E. Fisher, Phys. Rev. 136, A1626 (1964)), who found = 1.35 and = 4.22.
The transition temperature is T
c
= 627.2.
Unfortunately, this is not the full data set. These data have been analyzed
and re-analyzed many times in the literature. One such example is the follow-
ing plot of scaled eld versus scaled magnetization (Arrott and Noakes, Phys.
Rev. Lett. 19, 786 (1967)). Although these particular data were obtained
for Nickel, when experiments are performed on other ferromagnets, the same
exponents AND scaling functions W

are found, apart from possibly dierent

scaling factors B and D! If that is not surprising enough, when experiments are
performed on uids (liquid-liquid phase separation, or liguid-gas transitions),
not only is similar scaling behavior found, but again, the exponents and even
scaling functions W

are the same as for magnets!

131
The fact that our scaling hypothesis (actually, due to Widom) works has
important implications for other exponents, such as . Since we can write m
in terms of h and one of two universal functions W

, means that we can now

calculate the susceptibility. For t > 0, in particular,
=

m
h

h0

|t|

+
(0) . (5.36)
Here, W

+
(0) refers to the derivative of W
+
. Thus, must be . In other
words,
= +. (5.37)
A similar argument for t < 0 tells us also that

= . (5.38)
What about the critical isotherm, where t = 0? As we noted above, it is well
established that the magnetization and magnetic eld are both well-dened and
132
smooth away from the critical point, along the critical isotherm. But here, both
m and

h appear to be ill-dened (innite!) at t = 0. Thus, the critical isotherm
corresponds to the asymptotic regime of very large arguments of the functions
W

. How do these functions behave for large arguments? What if W(x) itself
exhibits a power-law here? If W(x) W

as x , then
m |t|

/|t|

. (5.39)
But, if this is to make any sense as t 0, then the t-dependence in the numer-
ator and denominator must just cancel. I.e., must be /. We can then nd
the relationship between m and h along the critical isotherm: m h

, so that
= 1/ = 1 +/. (5.40)
Moreover, the fact that the data collapse for m and

h for the same above and
below the critical temperature, we must also have that the exponent , as well
as the amplitude W

must be the same for W

+
and W

. In other words, the

functions W
+
and W

must converge for large arguments.

What we learn from all of this is that basic thermodynamics imposes con-
straints on some of the various exponents. Specically, there only appear to be
two independent exponents so far. In fact, the exponent is also completely
determined by and . In order to see this, we note that m and h are conjugate
thermodynamic variables, like volume and pressure for a gas. Thus

f
h
= m |t|

Dh/|t|

. (5.41)
We can integrate this (with respect to h), to nd that
f =

mdh |t|
+

md

h |t|
+
. (5.42)
Since
C = T

2
f
T
2
, (5.43)
we nd that
C |t|
+2
, (5.44)
meaning that
+ 2 + = 2. (5.45)
It can also be shown that

= . (5.46)
Thus, there appear to be at most just two independent critical exponents, say
and , from which the other four exponents can be derived. In other words,
although the various critical exponents are not so simple as mean-eld would
predict, there seems to be some order among these exponents.
133
5.4 Why does scaling theory work?
We are left with the question of why our basic scaling hypothesis works. One
very basic observation we can make from the fact that we observe power-law
behavior, e.g., the divergence of near t = 0, is that things look essentially
the same as one moves closer to the critical point, i.e., as t decreases toward
zero. This property is especially clear from Widoms approach to scaling: a
magnet close to the critical point behaves just like one far from the critical
point, apart from a simple rescaling of various physical quantities like h and
M. This mathematical property, known as scale invariance, appears to be a
key physical property of systems near critical points. In fact, scale invariance is
a very important property of many physical systems, and scaling theories and
approaches such as we have employed above can be applied rather widely.
Consider a simple mathematical property of a function such as the suscepti-
bility as a function of t. The fact that we observe a power-law dependence tells
us that there is no intrinsic scale, e.g., for t. Contrast this with, for instance,
an exponential dependence, for which there IS a characteristic scale. If were
exponential in t, we would necessarily have a characteristic scale t
0
for t:
e
t/t0
.
Instead, is a homogeneous function of t of degree :
(t) =

(t), (5.47)
for > 0. Our magnetic system, however, depends on two thermodynamic
variables, t and h. What if the free energy f itself is also homogenous in both
of these variables? Well examine the consequences of
f(
a
t,
b
h) = f(t, h). (5.48)
The reason for assuming this form is that, since our critical point is at t = h = 0,
for any nite t or h, > 1 moves us farther away from the critical point (at
least for a, b > 0). But, it may do so at dierent rates for the t and h. The
assumption we are making here is that, apart from an overall scaling factor,
the free energy is the unchanged. (By the way, our choice of exponent 1 for
on the right-hand side of this equation is not essential. If we replace this by
another exponent c = 0, we could redene a and b so that the equation above
is valid. There are, thus, only two independent exponents above, and we have
chosen this particular form without any loss of generality.)
The free energy can then be expressed as
f(t, h) =
1
f(
a
t,
b
h). (5.49)
From this, we obtain the entropy
S(t, h) =
F
T

f
t

1+a
f
1,0
(
a
t,
b
h), (5.50)
134
where f
i,j
represents the i-th derivative of f with respect to its rst argument,
and j-th derivative with respect to its second argument. The heat capacity is
then
C(t, h) = T

2
F
T
2

2
f
t
2

1+2a
f
2,0
(
a
t,
b
h). (5.51)
The magnetization is given by
m(t, h) =
f
h

1+b
f
0,1
(
a
t,
b
h). (5.52)
The susceptibility involves one more derivative with respect to h:
(t, h) =
m
h

1+2b
f
0,2
(
a
t,
b
h). (5.53)
These relations can be used to nd the various critical exponents as follows.
For h = 0, if we let = |t|
1/a
, then
C(t, 0) |t|
(1+2a)/a
f
2,0
(1, 0), (5.54)
where refer to t > 0 and t < 0. Thus,
= (2a 1)/a. (5.55)
Similarly, for m and we nd that
m(t, 0) |t|
(1b)/a
f
0,1
(1, 0) for t < 0, (5.56)
and
(t, 0) |t|
(12b)/a
f
0,2
(1, 0). (5.57)
This means that
= (1 b)/a (5.58)
and
= (2b 1)/a. (5.59)
Also, for t = 0, by considering = |h|
1/b
we nd that
m(0, h) |h|
(1b)/b
f
0,1
(0, 1). (5.60)
Thus,
= b/(1 b). (5.61)
You should check that the above scaling relationships among these exponents
(Eqs. 5.40, 5.45) are satised.
135
5.4.1 The story so far...
So, we learn, yet again that there appear only to be two independent exponents,
on which all the others depend. Beyond that, we now see that the number of
these key independent parameters describing the critical behavior is connected
to the fact that we have two (intensive) physical parameters describing the
system. For the magnet, these are the temperature (t) and magnetic eld (h).
For liquid-gas systems, the corresponding parameters are the temperature and
the pressure.
So far, I have made things sound a bit simpler than they really are. In
particular, I implied that the critical behavior of magnets is the same as for all
other systems. This is not the case. For instance, uniaxial magnets, for which
the magnetic moment points preferentially along one axis, dier in their critical
behavior (exponents and scaling functions) from Heisenberg-like magnets, in
which spins can point in any direction. Nevertheless, it IS true (and very sur-
prising) that uniaxial ferromagnets and uid systems such as liquid-gas systems
should behave in the same way! What is now known (and understood) is that
all systems with the same underlying symmetry have the same critical behavior.
A liquid-gas system, for instance, can be described by the presence/absence of a
particle, just like an up/down spin in the Ising model for (uniaxial!) ferromag-
netism. So, the critical exponents depend on the symmetry of the problem, but
NOT on any details of how the components interact: what could be more dif-
ferent in detail, after all, than magnetic spins and particles colliding! This lack
of dependence on molecular details is a very striking aspect of critical behavior
that begs for some explanation.
While we shall focus our attention on scaling theory as an approach, it
is worth noting that some relationships between the various critical exponents
were found by other means. For instance, based on very general thermodynamic
principles, Rushbrooke showed that

+ 2 +

2 (5.62)
must be true. Similarly, Griths showed that

+(1 +) 2. (5.63)
Predictions of scaling theory do not conict with these general conditions. But,
both of these inequalities become equalities within scaling theory. Not surpris-
ingly, many people attempted to derive exact expressions for various exponents
directly from thermodynamics or rigorous statistical mechanics. Occasionally,
people even reported derivations that were at odds with the above general in-
equalities. But, these all suered the same fate as various reports of violations
of the second law!
So, we are left with two burning questions: (1) why does scaling theory work;
and (2) why is the behavior universal, in that molecular details do not matter.
Perhaps the most important conceptual breakthrough that actually lead to a
simultaneous answer to both of these questions was due to Kadano in mid-
1960s. The beauty of this idea lies in its simplicity and physical intuition. It
136
does not require, for instance, complicated analysis and calculation. The down-
side of this approach is that is is essentially conceptual. It was not possible
to turn it directly into a practical procedure to actually calculate an exponent.
That had to wait for the work of Wilson in the 1970s, for which he (Wilson) was
awarded the Nobel prize in 1982. We shall NOT go through the latter approach.
But, we can learn a lot from the insights of Kadano, which we can understand
based on just one more property of critical phenomena that we have not focused
on so far: uctuations.
5.5 Critical Fluctuations
One of the most striking features of a critical point is the large uctuations
that become apparent there. This is often refered to as critical opalescence,
because of the opalescent or milky appearance of otherwise transparent systems
at a critical point. For instance, in the case of the liquid-gas transition, (very!)
near the critical temperature and pressure, the sample appears cloudy. This
is because the system tends to uctuate wildly between the two phases, since
they are nearly indistinguishable near the critical point. These uctuations
between phases amount to uctuations in density. These uctuations are both
large in magnitude (of density), as well as of large spatial extent. The density
uctuations result in uctuations in the index of refraction. This, together with
the large length scale of the uctuations (i.e., comparable to the wavelength of
light) means that the sample scatters light strongly, as do clouds.
Not only are these uctuations strong at the critical point, but in a sense
they provide the key to understanding the key features of critical behavior that
we have mentioned so far: (1) the scaling behavior, (2) the anomalous (non-
mean-eld) exponents, and even (3) universality and the fact that microscopic
details of the system become irrelevant.
Lets try to generalize Landau theory to allow for uctuations. The result
is usually known as Ginsburg-Landau theory. We consider a spatially varying
magnetization m(x), and generalize our free energy to include a penalty for
these spatial variations. (If there were no penalty, why would we have any
macroscopically homogeneous phases at all?) We write the free energy as
F[m(x)]

d
d
x

1
2
tm
2
(x) +
1
4
um
4
+
1
2

2
0
(m)
2

, (5.64)
where we have used a slightly simpler normalization of the various terms in the
original Landau theory. We have normalized by b and let u = c/b. We have
also introduced a new parameter with dimensions of a length,
0
. This new free
energy is actually a functional of the spatially varying m(x). Otherwise, this is
still in the spirit of Landau theory, in that we are expanding the free energy in
small quantities, now including gradients or derivatives of the eld m. We are
keeping only the lowest order derivatives, with the assumption that these will
dominate at long length scales.
137
We are not so interested in a precise function m(x) so much as how the
magnetization, or order parameter is correlated from one point to another. With
this in mind, we begin with replacing m by its Fourier series/transform:
m(x) =

k
m
k
e
ikx
. (5.65)
Furthermore, we shall focus on the case of t > 0, where we can ignore the
fourth-order term, at least if m remains small enough. We can then write the
free energy
F

d
d
x
1
2

k,k

t +
2
0
(ik)(ik

e
i(k+k

)x
m
k
m
k

= V

k
1
2

t +
2
0
k
2

|m
k
|
2
. (5.66)
Here, we have used the fact that

d
d
x e
i(k+k

)x
= V
k,k
, (5.67)
and the fact that m
k
is the complex conjugate of m
k
. By using the equiparti-
tion theorem, we then nd that
|m
k
|
2

kT
t +
2
0
k
2
. (5.68)
We can measure the correlations of the uctuating m(x) eld by the corre-
lation function
(x) = m(x +x

)m(x

). (5.69)
138
This correlation function must vanish for large separations x, since m = 0.
Furthermore, it must also be independent of x

. Thus,
m(x + x

)m(x

) =

k,k

m
k
e
ik(x+x

)
m
k
e
ikx

k
|m
k
|
2
e
ikx
. (5.70)
Thus, we can nd (x) by Fourier transform. You might recognize the corre-
sponding problem with t = 0, since the Fourier transformof the electric potential
about a point-charge is proportional to 1/k
2
. For t > 0, the result is
(x)
e
|x|/
|x|
d2
. (5.71)
While writing most of the expressions above for one dimension (d = 1), this
nal expression exhibits the expected inverse dependence on separation x that
is characteristic of the the electrostatic problem mentioned above. For nite
t > 0, the correlation function also exhibits an exponential decay with distance
|x|. The decay or screening length is =
0
|t|
1/2
. This correlation length
represents the typical size of a uctuating domain (with correlated value of
the order parameter, such as a magnetic domain in magnets). This correlation
length diverges near the critical point with an exponent 1/2. Of course, this is
what one would expect within Landau(-Ginsburg) theory. Real systems exhibit
a somewhat stronger dependence
=
0
|t|

, (5.72)
where the new (no pun intended) exponent 0.6. In fact, it is also observed
that the exponent in the denominator of Eq. 5.71 is not the predicted one from
Landau-Ginsburg theory:
(x)
e
|x|/
|x|
d2+
, (5.73)
where is observed to be non-zero, but small.
So, what do these uctuations tell us about the above puzzles of critical
behavior? Well, the most important thing to note is that the correlation length
not only can be large, but strictly diverges at the critical point.
139
5.6 Real-space Renormalization
Since the correlation length diverges near the critical point, if we look at the
Ising model near the critical point, we expect to see neighboring spins to be
highly correlated. This suggests that we can consider larger blocks of spins
as eectively acting in concert. We introduce the concept of block spins, where
groups of spins are thought of as acting like a single spin. Consider, for instance,
a 2-d lattice of spins indicated by s
i
, which take on values 1. If we now consider
squares of L spins on a side, which we dene to be a block, then we can treat
the resulting lattice of block spins as a magnet. (In the gure, L = 2.) We
dene these block spins S
I
= (

i
s
i
) /L
2
, which takes on values between -1 and
1. Provided that L is small enough so that the block spins are still smaller than
, we expect that the system is not fundamentally altered, i.e., that it still looks
like an Ising model. This is because the spins in each block are highly correlated.
This procedure amounts to a transformation of the original Hamiltonian from
H = h

i
s
i
J

i,j
s
i
s
j
(5.74)
to
H

= h

I
S
I
J

I,J
S
I
S
J
. (5.75)
Here, we might expect that h

L
2
h, since there are L
2
spins in a single block.
Note that on scaling up by a factor of L, this amounts to moving AWAY from
the critical point. This also makes sense from the point of view of the correlation
length , since under this rescaling of the system, decreases by a factor of L:

= /L. Given that =

0
|t|

, this also corresponds to moving away

from the critical point, as characterized by t t

= L
1/
t.
All of this can be generalized to d dimensions, by noting that upon rescaling
the system by a factor of L, we expect that the system behaves as though both
t and h increase according to
t t

= L
x
t, (5.76)
140
and
h h

= L
y
h. (5.77)
Here, as noted above, we expect that x = 1/ 2 and y d, the number of
dimensions.
This is an expression of the scale invariance we discussed above. We really
have not derived anything yet. We are simply arguing for a simple rescaling
of the parameters t and h by factors that we might expect to be powers of L.
We argued for this by appealing to the fact that near the critical point, since
the range of correlations is large, we expect to be able to lump large groups of
spins together to form blocks, resulting in a simple geometric increase in the
eective eld strength h. Likewise, the rescaling results in a simple decrease of
the correlation length, as measured in the new lattice units. This is equivalent
to an increase in t by a factor of L
1/
. These arguments really only apply near
the critical point, where the correlation length is suciently large that we still
have correlated block spins after scaling the system up:

is still larger than a

lattice spacing.
We can also see that the scale factors relating t to t

and h to h

must be
simple powers of L. This is really a consequence of scale invariance. If our system
really is scale invariant, meaning that it still looks like an Ising ferromagnet near
the critical point as we rescale our system by a factor of L, then we expect to
have the same form for the Hamiltonian H and free energy F after rescaling.
Consider, for instance, a rescaling of the system by a factor of L
1
followed by
another rescaling of the system by a factor of L
2
. Assuming that the eective
temperature simply rescales by some factor that depends on L
1
:
t t

= (L
1
)t. (5.78)
If we then follow this with another rescaling of the system by a factor of L
2
,
then
t

= (L
2
)t

= (L
2
)(L
1
)t. (5.79)
But, this should be the same as if we simply rescale the original system by a
factor of L
1
L
2
:
t

= (L
2
L
1
)t. (5.80)
This is satised by any simple power (L) = L
x
. Likewise, we also expect
the scaling of h to involve a (possibly dierent) power L
y
. This composition
property of these rescaling/renormalization transformations suggests a group-
like structure, where following one transformation with another corresponds to a
single transformation by a product of the scale factors. In fact, these operations
do NOT form a group in the mathematical sense, since these transformation
cannot, in general, be inverted: we only lose information in a renormalization
transformation.
141
It is very interesting to think about how this loss of information happens.
Let us say, for instance, that in our original system/lattice we have some free
energy that depends on a collection of parameters: t, h
1
, h
2
, . . .. Then, the
arguments above suggest that near the critical point, these all transform as
h
j
h

j
= L
yj
upon rescaling. If the exponent here y
j
happens to be less than
zero, then the corresponding parameter becomes irrelevant near the critical
point. In fact, it becomes increasingly irrelevant near the transition, since the
large correlation length means that we can rescale by a large factor, making
h

j
small. This is one of the reasons why most of the microscopic details of
the system become increasingly irrelevant near the transition. We actually only
care about those parameters for which y > 0. This also begins to explain why
we have universality, and why it is valid only close to the critical point.
So, what does this tell us about scaling? Consider the free energy f(t, h).
Under a renormalization transformation, we have argued that t L
x
t and
h L
y
h. Thus, we expect that f(t, h) f(L
x
t, L
y
h). But, we have to be
careful here. Since the free energy is an extensive quantity, depending on the
size of our sample, we must account for the fact that our sample eectively
shrinks under this renormalization: there are fewer block spins than original
spins. Thus, it is more appropriate to describe the free energy per spin. Since
we have L
d
original spins in a block, the free energy per spin in the new lattice
must be
f(L
x
t, L
y
h) = L
d
f(t, h). (5.81)
This is just like Eq. (5.48) above, with = L
d
. Thus, a = x/d and b = y/d.
Using Eqs. (5.55), (5.58), (5.59) and (5.61), we also obtain the various critical
exponents in terms of x and y. We also have, from above, that
= 1/x, (5.82)
which also means that
= 2 d. (5.83)
This illustrates once again that there are only two independent critical ex-
ponents. This is even true if we consider the exponent above, since it can be
shown that = (2 ).
As it stands, we now have a concrete theoretical procedure/approach to
quantitatively understand critical behavior. If we can learn about how Hamil-
tonians and free energies transform under the rescaling/renormalization trans-
formations outlined above, then we can, for instance, extract x and y. This is
142
more or less what is done. But, this procedure is complicated to do in prac-
tice (one needs to consider transformations in the space of Hamiltonians), and
the so-called real-space procedure of Kadano is not actually a practical one.
Nevertheless, we have at least conceptually identied the origin of scaling, and
closely associated aspects of critical phenomena such as universality. It remains
now to explain just why/when mean-eld theory fails, and when/if it can ever
be valid.
Before doing so, however, we an already learn something about this from
what we have just derived. In particular, there is something very odd about the
scaling relation in Eq. (5.83). It depends on dimensionality. All of the mean-eld
exponents are independent of dimension. On the one hand, this would appear
to tell us that mean-eld theory CANNOT be right. On the other hand, we also
learn that there is a special dimension, d = 4, for which Eq. (5.83) is satised
for mean-eld theory. This is an important point that we shall return to: four
dimensions is a special case. In fact, for any dimension d > 4, mean-eld theory
works! We are either incredibly unfortunate, or lucky that we live in a world
with funny critical exponents. I would argue that it is LUCK. The developments
of phase transitions and critical phenomena would have concluded with Landau
theory, and we would not have learned anything about scaling, renormalization,
etc..
5.7 The Breakdown of Mean-eld Theory
We have seen that uctuations play an important role in the behavior of systems
near a critical point. These uctuations are also the cause of the breakdown of
mean-eld theory. We can see this by making an estimate of the size of, e.g.,
magnetization uctuations in a magnet. If these uctuations are larger than or
comparable to the average value of the magnetization, then clearly the whole
idea behind mean-eld theory is questionable at best.
In order to estimate the magnitude of the uctuations, we consider the case
of t < 0, where we expect to nd a non-zero average of the order parameter m.
From the Ginsburg-Landau theory in Eq. (5.64), for t < 0 we expect to nd
|m
2
| = m
2
0
= t/u = |t|/u. About this value, we expect to nd uctuations
m(x). The free-energy can then be expressed as
F[m(x)] F
0
+

d
d
x

|t||m(x)|
2
+
1
2

2
0
(m)
2

, (5.84)
where we have expanded the integrand about the minimum at m = m
0
. We
have also indicated by F
0
the mean-eld value of the free energy, i.e., the free
energy for a uniform m(x) = m
0
. Note that
d
2
dm
2

1
2
tm
2
+
1
4
um
4

m=m0
= 2|t|. (5.85)
We can estimate the remaining integral by taking very seriously (of course,
a little too seriously in detail) the idea of the correlation length , and the gure
143
above. Well assume that at any given time, the sample can be decomposed into
regions of size that uctuate independently. Because the eld m is strongly
correlated over distances smaller than , we will treat the eld as uniform within
a single region. Furthermore, these regions eectively behave independently of
each other since the the eld m is weakly correlated over distances larger than
. Thus, we can calculate the contribution to the free energy above of one of
these regions as
F
d
|t|m
2
. (5.86)
We have neglected the gradient term in this free energy, but you can convince
yourself that this is of the same order as the above, since |m|
2
m
2
/
2
.
This means that the probability distribution of m is given by the exponen-
tial of F/kT. In fact, we have actually normalized the free energy by kT.
Thus,
m
2
|t|
1

d
|t|
d1

d
0
. (5.87)
But, for mean-eld theory to make sense, we have to have that m
2
/m
2
0
1.
Thus, we must have
|t|
d2

d
0
/u, (5.88)
since m
2
0
= |t|/u. Of course, = 1/2 within mean-eld theory, so that this
condition becomes
|t|
d4

2d
0
/u
2
. (5.89)
This can always be satised close to the critical point for d > 4, but cannot be
satised close to the critical point for d < 4. In three dimensions, this can only
be satised for |t| > t
0
, there
t
0
(u/
d
0
)
2
. (5.90)
This shows us that mean-eld theory cannot be valid arbitrarily close to the
the critical point. As t 0, the uctuations will always dominate the average,
144
or mean eld. But, depending on the system in question, it may be possible to
satisfy the inequality above for even very small |t|. This can happen especially if
the bare correlation length
0
is large enough. This is the case, for instance, for
the transition to superconductivity in many metals. This is because the natural
small distance, and therefore
0
, in this problem is the size of a Cooper pair,
which can be much larger than the atomic/molecular scale.
5.8 Finite-size Scaling
We have found that physical quantities such as the heat capacity and suscep-
tibility diverge, seemingly to innity at the critical point. It is natural to ask,
in what sense can these things be innite. In fact, in any nite system, the
heat capacity and susceptibility remain nite and continuous through the tran-
sition. The divergence is, strictly speaking, only in the thermodynamic limit
that N , where N is the number of degrees of freedom. But, Avagadros
number N
A
is BIG. For all practical purposes in most experiments, these quan-
tities appear to get ever larger near the critical point.
If one were to perform experiments on much smaller samples, say with N
10
6
or less, one would see that these divergences are rounded o. One place
where one denitely deals with small systems is numerical simulation. Here,
one must deal with systems that have many fewer degrees of freedom than
N
A
. In fact, simulation systems are usually so limited in size it is even hard
to recognize what would be critical phenomena in much larger systems. This
raises a challenge for numerical simulations to see critical phenomena. This is
a particular problem with second-order phase transitions, since the correlation
length diverges. When becomes comparable to the size of the system, the
behavior begins to deviate strongly from the thermodynamic limit.
One can actually use these nite-size limitations to ones advantage. Since
the correlation length diverges as
0
|t|

, for a nite system of linear di-

145
mension L, once becomes comparable to L, the system cannot get any more
critical than it already is. Again, we identify the divergence of as the dening
characteristic of criticality, and once reaches the full system size, we have cor-
related uctuations that span the whole system: is as big as we can measure!
We expect this to happen when |t| t
L
= (L/
0
)
1/
. In practice, this will
mean that the divergence of is rounded o and rendered nite in a region of
order t
L
on either side of the transition. Thus, we expect that for
0
|t|

in
an innite system, one should observe a maximum value
max
L
/
. This is
something that can easily be checked in numerical simulations on dierent size
samples. What is more, if we consider the normalized susceptibility /L
/
,
it should approach a constant, independent both of t and the size of the system,
below |t| of order t
L
. Only for t > t
L
should we see the critical behavior, and
here should be independent of L. Thus, (t, L) should satisfy
L
/
=

L
1/
t

, (5.91)
where (x) is a function that approaches a constant for small x and must vary
as x

for large x. Thus, by plotting L

/
versus L
1/
|t|, one should see a
collapse of data obtained for dierent values of L onto a single scaling curve
(actually, two curves

for t > 0 and t < 0), provided that and are

chosen appropriately. A very similar approach is also possible with the order
parameter m. If L
/
m is plotted versus L
1/
|t| (for t < 0), one obtains data
collapse onto a universal function that approaches a constant for small argument
and increases as a power law with exponent for large arguments.
146
This approach of nite-size scaling, and its generalizations, provides a very
useful tool for studying critical phenomena in numerical simulations, which nec-
essarily deal with small systems. In fact, simulations have played an important
role in the eld of critical phenomena over the past few decades. The particu-
larities of the the critical point, and especially the divergent correlation length
have helped to make this possible.
147