Physics 127b: Statistical Mechanics

Fluctuations at a Second Order Transition

We can use the Landau free energy to investigate fluctuations of the order parameter and so the validity of
mean field theory, and the expansion itself.
Remember that in the canonical ensemble the probability of a macroscopic configuration is proportional to
exp(−β1A), where 1A is the change in the Helmholtz free energy arising from the change in configuration.
Thus we have for a configuration m(Er ) of the Ising ferromagnet order parameter

P (m(Er )) ∝ e−BA(m(Er )) (1)

with A(m(Er )) given by the Landau expansion.

Let’s hope that fluctuations are small about the mean we calculated, i.e. δm = m − m̄ is small. Then we
have  2
T > Tc , m̄ = 0, f (δm) = a(T − Tc )δm2 + γ ∇δmE
√  2 . (2)
T < Tc , m̄ = a(Tc − T )/b, f (δm) = 2a(T − Tc )δm2 + γ ∇δm E

Note f (δm) provides the effective potential for small fluctuations in the quadratic minimum. Since the forms
above and below Tc are basically the same, lets work with the expression
 2
f (δm) = a1 δm2 + γ ∇δm E . (3)

It is useful to go to Fourier notation:

X
δm = mqE ei qE·Er (4a)
qE
Z
1
mqE = δm(Er )e−i qE·Er (4b)
V
and then
Z Z XX 0
(δm) d r =
2 3
d 3r mqE mqE0 ei(Eq +Eq )·Er (5a)
qE qE0
X X 2
=V mqE m−Eq = V mqE (5b)
qE qE

where we have used m−q = m∗q (δm real) in the last step .
This gives for the free energy for a magnetization configuration
X 2
A = A(m̄) + V (a1 + γ q 2 ) mqE . (6)
qE

This is now used in Eq. (1) to give the probability of mqE and so m(Er ). We notice that the free energy is
the sum of quadratic terms, so that the probability distribution e−βA is the product of Gaussians in the mq ,
yielding for the ensemble average of the fluctuations:
D E kB T
2
mqE = . (7)
2V (a1 + γ q 2 )

1
 D E
2
The quantity mqE is proportional to the static structure factor S(E
q ) that would be measured, for example,
in spin dependent X-ray or neutron scattering. Fourier transforming back gives us the correlation function.
The correlation function G(Er ) is defined by

G(Er ) = hm(Er )m(0)i − m̄2 (8a)

= hδm(Er )δm(0)i . (8b)

Now introducing the Fourier transform

* +
X X
G(Er ) = mqE ei qE·Er mqE0 (9a)
qE qE0
DX E
mqE ei qE·Er
2
= (9b)

where we have used mqE m−Eq 0 ∝ δqEqE0 since fluctuations in different modes will be uncorrelated.

To see this it is best to go to real and imaginary parts—usually the best way to be really sure of
fluctuations in Fourier space (are mq and m∗q independent or not?)—i.e.

δmq = Rq + iIq (10a)

δm−q = Rq − iIq (10b)

where we have used δm(Er ) real to get R−q = Rq and I−q = −Iq . Then we can write Eq. (6) as
X
A = A(m̄) + 2V (a1 + γ q 2 )(Rq2 + Iq2 ). (11)
qE>0
P P
Note that we have replaced qE by 2 qE>0 , i.e. a sum over some conveniently defined positive
half space, since the fluctuations at qE and −E
q are not independent. Now Eq. (11) is truly the
sum of independent quadratic terms. This means RqE and IqE (for all qE > 0) are independent.
Gaussian, fluctuating variables, and in particular equipartition gives
kB T
Rq2 = Iq2 = . (12)
4V (a1 + γ q 2 )

q | 6= qE0 and
It is easy to see mq mq 0 = 0 if |E

mq mq = Rq2 − Iq2 = 0 (13)

since RqE IqE0 = 0 for any qE, qE0 , and

kB T
mq m−q = mq m∗q = Rq2 + Iq2 = , (14)
2V (a1 + γ q 2 )
as surmised in Eq. (7).

Using the Eq. (7) and converting the qE-sum to an integral in the usual way gives
Z
kB T V ei qE·Er
G(Er ) = d 3q . (15)
2V (2π )3 a1 + γ q 2

2
The integral may be done (we did a similar integral in the Thomas-Fermi model of a screened charge in an
electron gas) to give
kB T e−r/ξ
G(Er ) = (16)
8π γ r
defining the correlation length ξ = (γ /a1 )1/2 , or explicitly
(
γ
1/2
a

(T − Tc )−1/2 for T > Tc
ξ(T ) = γ 1/2 . (17)
2a
(Tc − T )−1/2 for T < Tc

Fluctuations occur over correlated regions of size ξ(T ) which diverges approaching the critical temperature
from above or below.
These results for S(E
q ), G(Er ), and the correlation length are known as Ornstein-Zernike theory.

When is mean field theory good?

The expression (7) shows that the mean square fluctuations at long wavelengths diverge towards Tc
D E kB T
2
mqE'0 = . (18)
2V a1
Close enough to Tc the fluctuations will become as important as the mean, and the simple expressions in terms
of m̄ will break down. We estimate the temperature TG when this happens by asking when the fluctuations
over a correlation volume will be comparable to the mean, i.e. by setting

hδm(ξ )δm(0)i = m̄2 , (19)

which gives
kB TG e−1 a(Tc − TG )
= (20)
8π γ ξ b
or since Tc − TG will usually be small

Tc − TG b2 kB2 Tc
= . (21)
Tc 32π 2 e2 aγ 3
This is known as the Ginzburg criterion. For temperatures closer to Tc than TG fluctuations cannot be
neglected, and mean field theory will be unreliable. A similar region above Tc is expected to be dominated by
fluctuations. The region near Tc where fluctuations are important is known as the critical region Outside of
this range, mean field theory should be a good approximation. I have used the Ginzburg approach to calculate
the range, since the numerical factors are quite large, and give a more reliable estimate of the critical region.
Perhaps a better intuitive understanding is given by demanding that the free energy of the ordering over a
correlation volume be greater than kB Tc for fluctuations to be small, i.e.

δf ξ 3 & kB Tc (22a)
 3/2
a 2 (Tc − TG )2 γ
& k B Tc . (22b)
2b 2a(Tc − TG )

We conclude that close enough to Tc mean field theory will always break down, and fluctuations become
important. Since the Landau expansion itself is also only useful “near Tc ” there may in fact be no range
of temperatures over which mean field theory is quantitatively accurate, although even in these cases it
remains a useful qualitative guide. However the Ginzburg criterion depends on physical parameters of the

3
system, and some systems are accurately described by mean field theory except very close to Tc . One
example is superconductivity, where the correlation length is long compared to atomic scales (roughly
ξ ∼ ξ0 (1 − T /Tc )−1/2 where ξ0 is larger than atomic scales by a factor TF /Tc ∼ 104 ). Since ξ appears cubed
in the Ginzburg criterion (we are interested in fluctuations over the correlation volume), the critical region is
unobservably small, e.g. (1 − TG /Tc ) ∼ 10−9 − 10−12 , and mean field theory is highly accurate for practical
temperature ranges.