Lecture Notes On Thermodynamics and Statistical Mechanics (A Work in Progress)
Lecture Notes On Thermodynamics and Statistical Mechanics (A Work in Progress)
Lecture Notes On Thermodynamics and Statistical Mechanics (A Work in Progress)
(A Work in Progress)
Daniel Arovas
Department of Physics
University of California, San Diego
1 Probability 1
1.1 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Thermodynamics 17
2.1 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.2 Heat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
i
ii CONTENTS
2.2.3 Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.8.1 Energy E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.8.3 Enthalpy H . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3.2 Dynamical systems and the evolution of phase space volumes . . . . . . . . . . . . . . . . . 111
5.2.1 Bose and Fermi systems in the grand canonical ensemble . . . . . . . . . . . . . . . . . . . . 184
5.7.7 Example problem from Fall 2004 UCSD graduate written exam . . . . . . . . . . . . . . . . . 218
6.2.3 H = 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
6.9 Appendix II : One-Particle Irreducible Clusters and the Virial Expansion . . . . . . . . . . . . . . . . 291
7.2.2 Analytic form of the coexistence curve near the critical point . . . . . . . . . . . . . . . . . . 299
7.4.1 h = 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
7.4.3 h 6= 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
0.1 Preface
This is a proto-preface. A more complete preface will be written after these notes are completed.
These lecture notes are intended to supplement a course in statistical physics at the upper division undergraduate
or beginning graduate level.
I was fortunate to learn this subject from one of the great statistical physicists of our time, John Cardy.
I am grateful to my wife Joyce and to my children Ezra and Lily for putting up with all the outrageous lies I’ve
told them about getting off the computer ‘in just a few minutes’ while working on these notes.
These notes are dedicated to the only two creatures I know who are never angry with me: my father and my dog.
Probability
1.1 References
– F. Reif, Fundamentals of Statistical and Thermal Physics (McGraw-Hill, 1987)
This has been perhaps the most popular undergraduate text since it first appeared in 1967, and with good
reason.
– E. T. Jaynes, Probability Theory (Cambridge, 2007)
The bible on probability theory for physicists. A strongly Bayesian approach.
– C. Gardiner, Stochastic Methods (Springer-Verlag, 2010)
Very clear and complete text on stochastic mathematics.
1
2 CHAPTER 1. PROBABILITY
Consider the mechanical system depicted in Fig. 1.1, a version of which is often sold in novelty shops. A ball
is released from the top, which cascades consecutively through N levels. The details of each ball’s motion are
governed by Newton’s laws of motion. However, to predict where any given ball will end up in the bottom row is
difficult, because the ball’s trajectory depends sensitively on its initial conditions, and may even be influenced by
random vibrations of the entire apparatus. We therefore abandon all hope of integrating the equations of motion
and treat the system statistically. That is, we assume, at each level, that the ball moves to the right with probability
p and to the left with probability q = 1 − p. If there is no bias in the system, then p = q = 21 . The position XN after
N steps may be written
XN
X= σj , (1.1)
j=1
where σj = +1 if the ball moves to the right at level j, and σj = −1 if the ball moves to the left at level j. At each
level, the probability for these two outcomes is given by
(
p if σ = +1
Pσ = p δσ,+1 + q δσ,−1 = (1.2)
q if σ = −1 .
This is a normalized discrete probability distribution of the type discussed in section 1.4 below. The multivariate
distribution for all the steps is then
N
Y
P(σ1 , . . . , σN ) = P (σj ) . (1.3)
j=1
Our system is equivalent to a one-dimensional random walk. Imagine an inebriated pedestrian on a sidewalk
taking steps to the right and left at random. After N steps, the pedestrian’s location is X.
Now let’s compute the average of X:
N
X X
hXi = σj = N hσi = N σ P (σ) = N (p − q) = N (2p − 1) . (1.4)
j=1 σ=±1
This could be identified as an equation of state for our system, as it relates a measurable quantity X to the number
of steps N and the local bias p. Next, let’s compute the average of X 2 :
N X
X N
hX 2 i = hσj σj ′ i = N 2 (p − q)2 + 4N pq . (1.5)
j=1 j ′ =1
Figure 1.1: The falling ball system, which mimics a one-dimensional random walk.
fluctuations ∆Xrms are proportional to N 1/2 . In the limit N → ∞ then, the ratio ∆Xrms /hXi vanishes as N −1/2 .
This is a consequence of the central limit theorem (see §1.4.2 below), and we shall meet up with it again on several
occasions.
We can do even better. We can find the complete probability distribution for X. It is given by
N
PN,X = pN R q N L , (1.8)
NR
where NR/L are the numbers of steps taken to the right/left, with N = NR + NL , and X = NR − NL . There are
many independent ways to take NR steps to the right. For example, our first NR steps could all be to the right, and
the remaining NL = N − NR steps would then all be to the left. Or our final NR steps could all be to the right. For
each of these independent possibilities, the probability is pNR q NL . How many possibilities are there? Elementary
combinatorics tells us this number is
N N!
= . (1.9)
NR NR ! NL !
Note that N ± X = 2NR/L , so we can replace NR/L = 21 (N ± X). Thus,
N!
PN,X = N +X
N −X p(N +X)/2 q (N −X)/2 . (1.10)
2 ! 2 !
Consider the limit N → ∞ but with x ≡ X/N finite. This is analogous to what is called the thermodynamic limit
in statistical mechanics. Since N is large, x may be considered a continuous variable. We evaluate ln PN,X using
Stirling’s asymptotic expansion
ln N ! ≃ N ln N − N + O(ln N ) . (1.11)
4 CHAPTER 1. PROBABILITY
We then have
h i
ln PN,X ≃ N ln N − N − 21 N (1 + x) ln 21 N (1 + x) + 21 N (1 + x)
h i
− 12 N (1 − x) ln 12 N (1 − x) + 21 N (1 − x) + 12 N (1 + x) ln p + 21 N (1 − x) ln q
h i h i
= −N 1+x 2 ln 1+x
2 + 1−x
2 ln 1−x
2 + N 1+x
2 ln p + 1−x
2 ln q . (1.12)
Notice that the terms proportional to N ln N have all cancelled, leaving us with a quantity which is linear in N .
We may therefore write ln PN,X = −N f (x) + O(ln N ), where
h i h i
1+x 1+x 1−x 1−x 1+x 1−x
f (x) = 2 ln 2 + 2 ln 2 − 2 ln p + 2 ln q . (1.13)
where C is a normalization constant2 . Since N is by assumption large, the function PN,X is dominated by the
minimum (or minima) of f (x), where the probability is maximized. To find the minimum of f (x), we set f ′ (x) = 0,
where
q 1+x
f ′ (x) = 12 ln · . (1.15)
p 1−x
∞ Z∞ " #
X (X − X̄)2 p
1
PN,X ≈ 2 dX C exp − = 2πN pq C , (1.20)
8N pq
X=−∞ −∞
√
and thus C = 1/ 2πN pq. Why don’t we go beyond second order in the Taylor expansion of f (x)? We will find
out in §1.4.2 below.
origin of C lies in the O(ln N ) and O(N 0 ) terms in the asymptotic expansion of ln N !. We have ignored these terms here. Accounting
2 The
Figure 1.2: Comparison of exact distribution of eqn. 1.10 (red squares) with the Gaussian distribution of eqn. 1.19
(blue line).
The function f (x) can be written as a sum of two contributions, f (x) = e(x) − s(x), where
1+x
1+x
s(x) = − 2 ln 2 − 1−x
2 ln 1−x
2
(1.21)
e(x) = − 21 ln(pq) − 12 x ln(p/q) .
The function S(N, x) ≡ N s(x) is analogous to the statistical entropy of our system3 . We have
N N
S(N, x) = N s(x) = ln = ln 1 . (1.22)
NR 2 N (1 + x)
Thus, the statistical entropy is the logarithm of the number of ways the system can be configured so as to yield the same value
of X (at fixed N ). The second contribution to f (x) is the energy term. We write
The energy term biases the probability PN,X = exp(S − E) so that low energy configurations are more probable than
high energy configurations. For our system, we see that when p < q (i.e. p < 12 ), the energy is minimized by taking x
as small as possible (meaning as negative as possible). The smallest possible allowed value of x = X/N is x = −1.
Conversely, when p > q (i.e. p > 12 ), the energy is minimized by taking x as large as possible, which means x = 1.
The average value of x, as we have computed explicitly, is x̄ = p − q = 2p − 1, which falls somewhere in between
these two extremes.
In actual thermodynamic systems, as we shall see, entropy and energy are not dimensionless. What we have
called S here is really S/kB , which is the entropy in units of Boltzmann’s constant. And what we have called E
here is really E/kB T , which is energy in units of Boltzmann’s constant times temperature.
It was shown in the classic 1948 work of Claude Shannon that entropy is in fact a measure of information4 . Suppose
we observe that a particular event occurs with probability p. We associate with this observation an amount of
information I(p). The information I(p) should satisfy certain desiderata:
From these four properties, it is easy to show that the only possible function I(p) is
I(p) = −A ln p , (1.24)
where A is an arbitrary constant that can be absorbed into the base of the logarithm, since logb x = ln x/ ln b. We
will take A = 1 and use e as the base, so I(p) = − ln p. Another common choice is to take the base of the logarithm
to be 2, so I(p) = − log2 p. In this latter case, the units of information are known as bits. Note that I(0) = ∞. This
means that the observation of an extremely rare event carries a great deal of information.
Now suppose we have a set of events labeled by an integer n which occur with probabilities {pn }. What is
the expected amount of information in N observations? Since event n occurs an average of N pn times, and the
information content in pn is − ln pn , we have that the average information per observation is
hI i X
S= N =− pn ln pn , (1.25)
N n
which is known as the entropy of the distribution. Thus, maximizing S is equivalent to maximizing the information
content per observation.
Consider, for example, the information content of course grades. As we have seen, if the only constraint on the
probability distribution is that of overall normalization, then S is maximized when all the probabilities pn are
equal. The binary entropy is then S = log2 Γ , since pn = 1/Γ . Thus, for pass/fail grading, the maximum average
information per grade is − log2 ( 12 ) = log2 2 = 1 bit. If only A, B, C, D, and F grades are assigned, then the
maximum average information per grade is log2 5 = 2.32 bits. If we expand the grade options to include {A+, A,
A-, B+, B, B-, C+, C, C-, D, F}, then the maximum average information per grade is log2 11 = 3.46 bits.
Equivalently, consider, following the discussion in vol. 1 of Kardar, a random sequence {n1 , n2 , . . . , nN } where
each element nj takes one of K possible values. There are then K N such possible sequences, and to specify one of
them requires log2 (K N ) = N log2 K bits of information. However, if the value n occurs with probability pn , then
on average it will occur Nn = N pn times in a sequence of length N , and the total number of such sequences will
be
N!
g(N ) = QK . (1.26)
n=1 Nn !
In general, this is far less that the total possible number K N , and the number of bits necessary to specify one from
among these g(N ) possibilities is
K
X K
X
log2 g(N ) = log2 (N !) − log2 (Nn !) ≈ −N pn log2 pn , (1.27)
n=1 n=1
4 See
‘An Introduction to Information Theory and Entropy’ by T. Carter, Santa Fe Complex Systems Summer School, June 2011. Available
online at http://astarte.csustan.edu/ tom/SFI-CSSS/info-theory/info-lec.pdf.
1.3. PROBABILITY DISTRIBUTIONS FROM MAXIMUM ENTROPY 7
1
where we have invoked Stirling’s approximation. If the distribution is uniform, then we have pn = K for all
n ∈ {1, . . . , K}, and log2 g(N ) = N log2 K.
We have shown how one can proceed from a probability distribution and compute various averages. We now
seek to go in the other direction, and determine the full probability distribution based on a knowledge of certain
averages.
At first, this seems impossible. Suppose we want to reproduce the full probability distribution for an N -step
random walk from knowledge of the average hXi = (2p − 1)N . The problem seems ridiculously underdeter-
mined, since there are 2N possible configurations for an N -step random walk: σj = ±1 for j = 1, . . . , N . Overall
normalization requires X
P (σ1 , . . . , σN ) = 1 , (1.28)
{σj }
but this just imposes one constraint on the 2 probabilities P (σ1 , . . . , σN ), leaving 2N −1 overall parameters. What
N
where here we take e as the base of the logarithm. The entropy may therefore be regarded as a function of the
probability distribution: S = S {pn } . One special
property
of the entropy is the following. Suppose we have two
independent normalized distributions pAa and pBb . The joint probability for events a and b is then Pa,b = pAa pBb .
The entropy of the joint distribution is then
XX XX XX
S=− Pa,b ln Pa,b = − pAa pBb ln pAa pBb = − pAa pBb ln pAa + ln pBb
a b a b a b
X X X X X X
=− pAa ln pAa · pBb − pBb ln pBb · pAa =− pAa ln pAa − pBb ln pBb
a b b a a b
A B
=S +S .
Thus, the entropy of a joint distribution formed from two independent distributions is additive.
P
Suppose all we knew about {pn } was that it was normalized. Then n pn = 1. This is a constraint on the values
{pn }. Let us now extremize the entropy S with respect to the distribution {pn }, but subject to the normalization
constraint. We do this using Lagrange’s method of undetermined multipliers. We define
X X
S ∗ {pn }, λ = − pn ln pn − λ pn − 1 (1.31)
n n
8 CHAPTER 1. PROBABILITY
and we freely extremize S ∗ over all its arguments. Thus, for all n we have
∂S ∗
= − ln pn + 1 + λ = 0 (1.32)
∂pn
as well as
∂S ∗ X
= pn − 1 = 0 . (1.33)
∂λ n
From the first of these equations, we obtain pn = e−(1+λ) , and from the second we obtain
X X
pn = e−(1+λ) · 1 = Γ e−(1+λ) , (1.34)
n n
P
where Γ ≡ n 1 is the total number of possible events. Thus,
1
pn = , (1.35)
Γ
which says that all events are equally probable.
Now suppose we know one other piece of information, which is the average value of some quantity X =
P
n Xn pn . We now extremize S subject to two constraints, and so we define
X X X
S ∗ {pn }, λ0 , λ1 = − pn ln pn − λ0 pn − 1 − λ1 Xn pn − X . (1.36)
n n n
We then have
∂S ∗
= − ln pn + 1 + λ0 + λ1 Xn = 0 , (1.37)
∂pn
which yields the two-parameter distribution
General formulation
The generalization to K extra pieces of information (plus normalization) is immediately apparent. We have
X
Xa = Xna pn , (1.41)
n
X K
X X
∗
S {pn }, {λa } = − pn ln pn − λa Xna pn − X a , (1.42)
n a=0 n
1.3. PROBABILITY DISTRIBUTIONS FROM MAXIMUM ENTROPY 9
(a=0)
with Xn ≡ X (a=0) = 1. Then the optimal distribution which extremizes S subject to the K + 1 constraints is
( K
)
X
a
pn = exp − 1 − λa Xn
a=0
( ) (1.43)
K
X
1
= exp − λa Xna ,
Z a=1
P
where Z = e1+λ0 is determined by normalization: n pn = 1. This is a (K + 1)-parameter distribution, with
{λ0 , λ1 , . . . , λK } determined by the K + 1 constraints in eqn. 1.41.
Example
As an example, consider the random walk problem. We have two pieces of information:
X X
··· P (σ1 , . . . , σN ) = 1 (1.44)
σ1 σN
X X N
X
··· P (σ1 , . . . , σN ) σj = X . (1.45)
σ1 σN j=1
Here the discrete label n from §1.3.1 ranges over 2N possible values, and may be written as an N digit binary
number rN · · · r1 , where rj = 12 (1 + σj ) is 0 or 1. Extremizing S subject to these constraints, we obtain
( ) N
X Y
P (σ1 , . . . , σN ) = C exp −λ σj =C e−λ σj , (1.46)
j j=1
where
e−λ eλ
p= , q =1−p= . (1.49)
eλ + e−λ eλ + e−λ
We then have X = (2p − 1)N , which determines p = 12 (N + X), and we have recovered the correct distribution.
Suppose we have a continuous probability density P (ϕ) defined over some set Ω. We have observables
Z
X = dµ X a (ϕ) P (ϕ) ,
a
(1.50)
Ω
10 CHAPTER 1. PROBABILITY
QD
where dµ is the appropriate integration measure. We assume dµ = j=1 dϕj , where D is the dimension of Ω.
Then we extremize the functional
Z K Z !
X
∗ a a
S P (ϕ), {λa } = − dµ P (ϕ) ln P (ϕ) − λa dµ P (ϕ) X (ϕ) − X (1.51)
Ω a=0 Ω
with respect to P (ϕ) and with respect to {λa }. Again, X 0 (ϕ) ≡ X 0 ≡ 1. This yields the following result:
K
X
ln P (ϕ) = −1 − λa X a (ϕ) . (1.52)
a=0
The K + 1 Lagrange multipliers {λa } are then determined from the K + 1 constraint equations in eqn. 1.50.
As an example, consider a distribution P (x) over the real numbers R. We constrain
Z∞ Z∞ Z∞
dx P (x) = 1 , dx x P (x) = µ , dx x2 P (x) = µ2 + σ 2 . (1.53)
−∞ −∞ −∞
2 2 √
In other words, λ1 = −µ/σ 2 and λ2 = 1/2σ 2 , with C = e−µ /2σ / 2πσ 2 .
Consider a system whose possible configurations | n i can be labeled by a discrete variable n ∈ C, where C is the
set of possible configurations. The total number of possible configurations, which is to say the order of the set C,
may be finite or infinite. Next, consider an ensemble of such systems, and let Pn denote the probability that a
given random element from that ensemble is in the state (configuration) | n i. The collection {Pn } forms a discrete
probability distribution. We assume that the distribution is normalized, meaning
X
Pn = 1 . (1.56)
n∈C
Now let An be a quantity which takes values depending on n. The average of A is given by
X
hAi = Pn An . (1.57)
n∈C
Typically, C is the set of integers (Z) or some subset thereof, but it could be any countable set. As an example,
consider the throw of a single six-sided die. Then Pn = 61 for each n ∈ {1, . . . , 6}. Let An = 0 if n is even and 1 if n
is odd. Then find hAi = 12 , i.e. on average half the throws of the die will result in an even number.
1.4. GENERAL ASPECTS OF PROBABILITY DISTRIBUTIONS 11
It may be that the system’s configurations are described by several discrete variablesP{n1 , n2 , n3 , . . .}. We can
combine these into a vector n and then we write Pn for the discrete distribution, with n Pn = 1.
Another possibility is that the system’s configurations are parameterized by a collection of continuous variables,
ϕ = {ϕ1 , . . . , ϕn }. We write ϕ ∈ Ω, where Ω is the phase space (or configuration space) of the system. Let dµ be a
measure on this space. In general, we can write
The phase space measure used in classical statistical mechanics gives equal weight W to equal phase space vol-
umes:
Yr
dµ = C dqσ dpσ , (1.59)
σ=1
5
where C is a constant we shall discuss later on below .
Any continuous probability distribution P (ϕ) is normalized according to
Z
dµ P (ϕ) = 1 . (1.60)
Ω
1 2 2
P (x) = √ e−(x−µ) /2σ . (1.62)
2πσ 2
hxi = µ
(1.64)
hx i − hxi2 = σ 2 .
2
We call µ the mean and σ the standard deviation of the distribution, eqn. 1.62.
The quantity P (ϕ) is called the distribution or probability density. One has
and momenta which leave Hamilton’s equations of motion invariant, and which preserve phase space volumes under Hamiltonian evolution.
For this reason dµ is called an invariant phase space measure. See the discussion in appendix II of chapter 4.
6 Memorize this!
12 CHAPTER 1. PROBABILITY
If x is distributed according to P1 (x), then the probability distribution on the product space (x1 , x2 ) is simply the
product of the distributions:
Suppose we have a function φ(x1 , . . . , xN ). How is it distributed? Let Q(φ) be the distribution for φ. We then have
Z∞ Z∞
P(φ) = dx1 · · · dxN PN (x1 , . . . , xN ) δ φ(x1 , . . . , xN ) − φ
−∞ −∞
(1.66)
Z∞ Z∞
= dx1 · · · xN P1 (x1 ) · · · P1 (xN ) δ φ(x1 , . . . , xN ) − φ ,
−∞ −∞
where the second line is appropriate if the {xj } are themselves distributed independently. Note that
Z∞
dφ P(φ) = 1 , (1.67)
−∞
N
X
X= xi . (1.68)
i=1
We will be particularly interested in the case where N is large. For general N , though, we have
Z∞ Z∞
PN (X) = dx1 · · · dxN P1 (x1 ) · · · P1 (xN ) δ x1 + x2 + . . . + xN − X . (1.69)
−∞ −∞
1.4. GENERAL ASPECTS OF PROBABILITY DISTRIBUTIONS 13
Z∞
P̂N (k) = dX PN (X) e−ikX (1.70)
−∞
Z∞ Z∞ Z∞
= dX dx1 · · · xN P1 (x1 ) · · · P1 (xN ) δ x1 + . . . + xN − X) e−ikX
−∞ −∞ −∞
N
= P̂1 (k) ,
where
Z∞
P̂1 (k) = dx P1 (x) e−ikx (1.71)
−∞
is the Fourier transform of the single variable distribution P1 (x). The distribution PN (X) is a convolution of the
individual P1 (xi ) distributions. We have therefore proven that the Fourier transform of a convolution is the product of
the Fourier transforms.
Z∞
P̂1 (k) = dx P1 (x) 1 − ikx − 1
2 k 2 x2 + 1
6 i k 3 x3 + . . .
(1.72)
−∞
= 1 − ikhxi − 1
2 k 2 hx2 i + 1
6 i k 3 hx3 i + . . . .
Thus,
ln P̂1 (k) = −iµk − 21 σ 2 k 2 + 1
6 i γ 3 k3 + . . . , (1.73)
where
µ = hxi
σ 2 = hx2 i − hxi2 (1.74)
3 3 2 3
γ = hx i − 3 hx i hxi + 2 hxi
7 Jean Baptiste Joseph Fourier (1768-1830) had an illustrious career. The son of a tailor, and orphaned at age eight, Fourier’s ignoble status
rendered him ineligible to receive a commission in the scientific corps of the French army. A Benedictine minister at the École Royale Militaire
of Auxerre remarked, ”Fourier, not being noble, could not enter the artillery, although he were a second Newton.” Fourier prepared for the priesthood,
but his affinity for mathematics proved overwhelming, and so he left the abbey and soon thereafter accepted a military lectureship position.
Despite his initial support for revolution in France, in 1794 Fourier ran afoul of a rival sect while on a trip to Orléans and was arrested and very
nearly guillotined. Fortunately the Reign of Terror ended soon after the death of Robespierre, and Fourier was released. He went on Napoleon
Bonaparte’s 1798 expedition to Egypt, where he was appointed governor of Lower Egypt. His organizational skills impressed Napoleon, and
upon return to France he was appointed to a position of prefect in Grenoble. It was in Grenoble that Fourier performed his landmark studies
of heat, and his famous work on partial differential equations and Fourier series. It seems that Fourier’s fascination with heat began in Egypt,
where he developed an appreciation of desert climate. His fascination developed into an obsession, and he became convinced that heat could
promote a healthy body. He would cover himself in blankets, like a mummy, in his heated apartment, even during the middle of summer.
On May 4, 1830, Fourier, so arrayed, tripped and fell down a flight of stairs. This aggravated a developing heart condition, which he refused
to treat with anything other than more heat. Two weeks later, he died. Fourier’s is one of the 72 names of scientists, engineers and other
luminaries which are engraved on the Eiffel Tower. Source: http://www.robertnowlan.com/pdfs/Fourier,%20Joseph.pdf
14 CHAPTER 1. PROBABILITY
3 3
Now for the inverse transform. In computing PN (X), we will expand the term eiN γ k /6
and all subsequent terms
in the above product as a power series in k. We then have
Z∞
dk ik(X−N µ) −N σ2 k2 /2 n o
PN (X) = e e 1+ 1
6 i N γ 3k3 + . . .
2π
−∞
3
3 ∂ 1 2 2 (1.76)
1
= 1 − 6Nγ + ... √ e−(X−N µ) /2N σ
∂X 3 2πN σ 2
1 2 2
= √ e−(X−N µ) /2N σ (N → ∞) .
2πN σ 2
√ ∂3 −1/2 ∂ 3
In going from the second line to the third, we have written X = N ξ, in which case N ∂X 3 = N ∂ξ 3 , which
gives a subleading contribution which vanishes in the N → ∞ limit. We have just proven the central limit theorem:
in the limit N → ∞, the distribution
√ of a sum of N independent random variables xi is a Gaussian with mean N µ
and standard deviation N σ. Our only assumptions are that the mean µ and standard deviation σ exist for the
distribution P1 (x). Note that P1 (x) itself need not be a Gaussian – it could be a very peculiar distribution indeed,
but so long as its first and second moment exist, where the k th moment is simply hxk i, the distribution of the sum
PN
X = i=1 xi is a Gaussian.
Here, the vector b = (b1 , . . . , bn ) is identified as a source. Since Z(0) = 1, we have that the distribution P (x) is
normalized. Now consider averages of the form
Z
h xj1 · · · xj2k i = dnx P (x) xj1 · · · xj2k
∂ nZ(b)
= (1.79)
∂bj · · · ∂bj b=0
1 2k
X
= A−1
j j · · · A−1
j j .
σ(1) σ(2) σ(2k−1) σ(2k)
contractions
The sum in the last term is over all contractions of the indices {j1 , . . . , j2k }. A contraction is an arrangement of
the 2k indices into k pairs. There are C2k = (2k)!/2k k! possible such contractions. To obtain this result for Ck ,
we start with the first index and then find a mate among the remaining 2k − 1 indices. Then we choose the next
unpaired index and find a mate among the remaining 2k − 3 indices. Proceeding in this manner, we have
(2k)!
C2k = (2k − 1) · (2k − 3) · · · 3 · 1 = . (1.80)
2k k!
1.5. APPENDIX : BAYESIAN STATISTICS 15
Equivalently, we can take all possible permutations of the 2k indices, and then divide by 2k k! since permuta-
tion within a given pair results in the same contraction and permutation among the k pairs results in the same
contraction. For example, for k = 2, we have C4 = 3, and
Let the probability of a discrete event A be P (A). We now introduce two additional probabilities. The joint
probability for events A and B together is written P (A ∩ B). The conditional probability of B given A is P (B|A). We
can compute the joint probability P (A ∩ B) = P (B ∩ A) in two ways:
Thus,
P (B|A) · P (A)
P (A|B) = , (1.83)
P (B)
a result known as Bayes’ theorem. Now suppose the ‘event space’ is partitioned as {Ai }. Then
X
P (B) = P (B|Ai ) · P (Ai ) . (1.84)
i
We then have
P (B|Ai ) · P (Ai )
P (Ai |B) = P , (1.85)
j P (B|Aj ) · P (Aj )
a result sometimes known as the extended form of Bayes’ theorem. When the event space is a ‘binary partition’
{A, ¬A}, we have
P (B|A) · P (A)
P (A|B) = . (1.86)
P (B|A) · P (A) + P (B|¬A) · P (¬A)
Note that P (A|B) + P (¬A|B) = 1 (which follows from ¬¬A = A).
As an example, consider the following problem in epidemiology. Suppose there is a rare but highly contagious
disease A which occurs in 0.01% of the general population. Suppose further that there is a simple test for the
disease which is accurate 99.99% of the time. That is, out of every 10,000 tests, the correct answer is returned 9,999
times, and the incorrect answer is returned only once8 . Now let us administer the test to a large group of people
from the general population. Those who test positive are quarantined. Question: what is the probability that
someone chosen at random from the quarantine group actually has the disease? We use Bayes’ theorem with the
binary partition {A, ¬A}. Let B denote the event that an individual tests positive. Anyone from the quarantine
group has tested positive. Given this datum, we want to know the probability that that person has the disease.
That is, we want P (A|B). Applying eqn. 1.86 with
we find P (A|B) = 12 . That is, there is only a 50% chance that someone who tested positive actually has the disease,
despite the test being 99.99% accurate! The reason is that, given the rarity of the disease in the general population,
the number of false positives is statistically equal to the number of true positives.
8 Epidemiologists define the sensitivity of a binary classification test as the fraction of actual positives which are correctly identified, and
the specificity as the fraction of actual negatives that are correctly identified. In our example in the text, the sensitivity and specificity are both
0.9999.
16 CHAPTER 1. PROBABILITY
and
P (y|x) · P (x)
P (x|y) = R . (1.88)
dx′ P (y|x′ ) · P (x′ )
The range of integration may depend on the specific application.
The quantities P (Ai ) are called the prior distribution. Clearly in order to compute P (B) or P (Ai |B) we must know
the priors, and this is usually the weakest link in the Bayesian chain of reasoning. If our prior distribution is not
accurate, Bayes’ theorem will generate incorrect results. One approach to obtaining the prior probabilities P (Ai )
is to obtain them from a maximum entropy construction.
Chapter 2
Thermodynamics
2.1 References
– E. Fermi, Thermodynamics (Dover, 1956)
This outstanding and inexpensive little book is a model of clarity.
– A. H. Carter, Classical and Statistical Thermodynamics
(Benjamin Cummings, 2000)
A very relaxed treatment appropriate for undergraduate physics majors.
– H. B. Callen, Thermodynamics and an Introduction to Thermostatistics
(2nd edition, Wiley, 1985)
A comprehensive text appropriate for an extended course on thermodynamics.
– D. V. Schroeder, An Introduction to Thermal Physics (Addison-Wesley, 2000)
An excellent thermodynamics text appropriate for upper division undergraduates. Contains many illustra-
tive practical applications.
– D. Kondepudi and I. Prigogine, Modern Thermodynamics: From Heat Engines to Dissipative Structures
(Wiley, 1998)
Lively modern text with excellent choice of topics and good historical content. More focus on chemical and
materials applications than in Callen.
– L. E. Reichl, A Modern Course in Statistical Physics (2nd edition, Wiley, 1998)
A graduate level text with an excellent and crisp section on thermodynamics.
17
18 CHAPTER 2. THERMODYNAMICS
Thermodynamics is the study of relations among the state variables describing a thermodynamic system, and of
transformations of heat into work and vice versa.
Thermodynamic systems contain large numbers of constituent particles, and are described by a set of state variables
which describe the system’s properties in an average sense. State variables are classified as being either extensive
or intensive.
Extensive variables, such as volume V , particle number N , total internal energy E, magnetization M , etc., scale
linearly with the system size, i.e. as the first power of the system volume. If we take two identical thermodynamic
systems, place them next to each other, and remove any barriers between them, then all the extensive variables
will double in size.
Intensive variables, such as the pressure p, the temperature T , the chemical potential µ, the electric field E, etc., are
independent of system size, scaling as the zeroth power of the volume. They are the same throughout the system,
if that system is in an appropriate state of equilibrium. The ratio of any two extensive variables is an intensive
variable. For example, we write n = N/V for the number density, which scales as V 0 . Intensive variables may
also be inhomogeneous. For example, n(r) is the number density at position r, and is defined as the limit of ∆N/∆V
of the number of particles ∆N inside a volume ∆V which contains the point r, in the limit V ≫ ∆V ≫ V /N .
Classically, the full motion of a system of N point particles requires 6N variables to fully describe it (3N positions
and 3N velocities or momenta, in three space dimensions)1 . Since the constituents are very small, N is typically
very large. A typical solid or liquid, for example, has a mass density on the order of ̺ ∼ 1 g/cm3 ; for gases,
̺ ∼ 10−3 g/cm3 . The constituent atoms have masses of 100 to 102 grams per mole, where one mole of X contains
NA of X, and NA = 6.0221415 × 1023 is Avogadro’s number. Thus, for solids and liquids we roughly expect
number densities n of 10−2 − 100 mol/cm3 for solids and liquids, and 10−5 − 10−3 mol/cm3 for gases. Clearly
we are dealing with fantastically large numbers of constituent particles in a typical thermodynamic system. The
underlying theoretical basis for thermodynamics, where we use a small number of state variables to describe a
system, is provided by the microscopic theory of statistical mechanics, which we shall study in the weeks ahead.
Intensive quantities such as p, T , and n ultimately involve averages over both space and time. Consider for
example the case of a gas enclosed in a container. We can measure the pressure (relative to atmospheric pressure)
by attaching a spring to a moveable wall, as shown in Fig. 2.2. From the displacement of the spring and the value
of its spring constant k we determine the force F . This force is due to the difference in pressures, so p = p0 + F/A.
Microscopically, the gas consists of constituent atoms or molecules, which are constantly undergoing collisions
with each other and with the walls of the container. When a particle bounces off a wall, it imparts an impulse
2n̂(n̂ · p), where p is the particle’s momentum and n̂ is the unit vector normal to the wall. (Only particles with
p · n̂ > 0 will hit the wall.) Multiply this by the number of particles colliding with the wall per unit time, and one
finds the net force on the wall; dividing by the area gives the pressure p. Within the gas, each particle travels for
a distance ℓ, called the mean free path, before it undergoes a collision. We can write ℓ = v̄τ , where v̄ is the average
particle speed and τ is the mean free time. When we study the kinetic theory of gases, we will derive formulas
for ℓ and v̄ (and hence τ ). For now it is helpful to quote some numbers to get an idea of the relevant distance
and time scales. For O2 gas at standard temperature and pressure (T = 0◦ C, p = 1 atm), the mean free path is
ℓ ≈ 1.1 × 10−5 cm, the average speed is v̄ ≈ 480 m/s, and the mean free time is τ ≈ 2.5 × 10−10 s. Thus, particles in
the gas undergo collisions at a rate τ −1 ≈ 4.0 × 109 s−1 . A measuring device, such as our spring, or a thermometer,
1 For
a system of N molecules which can freely rotate, we must then specify 3N additional orientational variables – the Euler angles – and
their 3N conjugate momenta. The dimension of phase space is then 12N .
2.2. WHAT IS THERMODYNAMICS? 19
effectively performs time and space averages. If there are Nc collisions with a particular patch of wall during some
time interval on which our measurement device responds, then the root mean square relative fluctuations in the
−1/2
local pressure will be on the order of Nc times the average. Since Nc is a very large number, the fluctuations
are negligible.
If the system is in steady state, the state variables do not change with time. If furthermore there are no macroscopic
currents of energy or particle number flowing through the system, the system is said to be in equilibrium. A
continuous succession of equilibrium states is known as a thermodynamic path, which can be represented as a
smooth curve in a multidimensional space whose axes are labeled by state variables. A thermodynamic process
is any change or succession of changes which results in a change of the state variables. In a cyclic process, the
initial and final states are the same. In a quasistatic process, the system passes through a continuous succession of
equilibria. A reversible process is one where the external conditions and the thermodynamic path of the system can
be reversed (at first this seems to be a tautology). All reversible processes are quasistatic, but not all quasistatic
processes are reversible. For example, the slow expansion of a gas against a piston head, whose counter-force is
always infinitesimally less than the force pA exerted by the gas, is reversible. To reverse this process, we simply
add infinitesimally more force to pA and the gas compresses. A quasistatic process which is not reversible: slowly
dragging a block across the floor, or the slow leak of air from a tire. Irreversible processes, as a rule, are dissipative.
Other special processes include isothermal (dT = 0) isobaric (dp = 0), isochoric (dV = 0), and adiabatic (d̄Q = 0,
i.e. no heat exchange):
reversible: d̄Q = T dS isothermal: dT = 0
spontaneous: d̄Q < T dS isochoric: dV = 0
adiabatic: d̄Q = 0 isobaric: dp = 0
quasistatic: infinitely slowly
We shall discuss later the entropy S and its connection with irreversibility.
How many state variables are necessary to fully specify the equilibrium state of a thermodynamic system? For
a single component system, such as water which is composed of one constituent molecule, the answer is three.
These can be taken to be T , p, and V . One always must specify at least one extensive variable, else we cannot
determine the overall size of the system. For a multicomponent system with g different species, we must specify
g + 2 state variables, which may be {T, p, N1, . . . , Ng }, where Na is the number of particles of species a. P Another
g
possibility is the set (T, p, V, x1 , . . . , xg−1 },P
where the concentration of species a is xa = Na /N . Here, N = a=1 Na
g
is the total number of particles. Note that a=1 xa = 1.
If then follows that if we specify more than g + 2 state variables, there must exist a relation among them. Such
relations are known as equations of state. The most famous example is the ideal gas law,
pV = N kB T , (2.1)
20 CHAPTER 2. THERMODYNAMICS
Figure 2.2: The pressure p of a gas is due to an average over space and time of the impulses due to the constituent
particles.
relating the four state variables T , p, V , and N . Here kB = 1.3806503 × 10−16 erg/K is Boltzmann’s constant.
Another example is the van der Waals equation,
aN 2
p + 2 (V − bN ) = N kB T , (2.2)
V
where a and b are constants which depend on the molecule which forms the gas. For a third example, consider a
paramagnet, where
M CH
= , (2.3)
V T
where M is the magnetization, H the magnetic field, and C the Curie constant.
Any quantity which, in equilibrium, depends only on the state variables is called a state function. For example,
the total internal energy E of a thermodynamics system is a state function, and we may write E = E(T, p, V ).
State functions can also serve as state variables, although the most natural state variables are those which can be
directly measured.
2.2.2 Heat
Once thought to be a type of fluid, heat is now understood in terms of the kinetic theory of gases, liquids, and
solids as a form of energy stored in the disordered motion of constituent particles. The units of heat are therefore
units of energy, and it is appropriate to speak of heat energy, which we shall simply abbreviate as heat:2
1 J = 107 erg = 6.242 × 1018 eV = 2.390 × 10−4 kcal = 9.478 × 10−4 BTU . (2.4)
We will use the symbol Q to denote the amount of heat energy absorbed by a system during some given ther-
modynamic process, and d̄Q to denote a differential amount of heat energy. The symbol d̄ indicates an ‘inexact
differential’, about which we shall have more to say presently. This means that heat is not a state function: there
is no ‘heat function’ Q(T, p, V ).
calorie (cal) is the amount of heat needed to raise 1 g of H2 O from T0 = 14.5◦ C to T1 = 15.5◦ C at a pressure of p0 = 1 atm. One
2 One
British Thermal Unit (BTU) is the amount of heat needed to raise 1 lb. of H2 O from T0 = 63◦ F to T1 = 64◦ F at a pressure of p0 = 1 atm.
2.2. WHAT IS THERMODYNAMICS? 21
2.2.3 Work
In general we will write the differential element of work d̄W done by the system as
X
d̄W = Fi dXi , (2.5)
i
where Fi is a generalized force and dXi a generalized displacement3 . The generalized forces and displacements are
themselves state variables, and by convention we will take the generalized forces to be intensive and the general-
ized displacements to be extensive. As an example, in a simple one-component system, we have d̄W = p dV . More
generally, we write
P P
− j yj dXj a µa dNa
z }| { z }| {
d̄W = p dV − H · dM − E · dP − σ dA + . . . − µ1 dN1 + µ2 dN2 + . . . (2.6)
Here we distinguish between two types of work. The first involves changes in quantities such as volume, mag-
netization, electric polarization, area, etc. The conjugate forces yi applied to the system are then −p, the magnetic
field H, the electric field E, the surface tension σ, respectively. The second type of work involves changes in the
number of constituents of a given species. For example, energy is required in order to dissociate two hydrogen
atoms in an H2 molecule. The effect of such a process is dNH2 = −1 and dNH = +2.
As with heat, d̄W is an inexact differential, and work W is not a state variable, since it is path-dependent. There is
no ‘work function’ W (T, p, V ).
The units of pressure (p) are force per unit area. The SI unit is the Pascal (Pa): 1 Pa = 1 N/m2 = 1 kg/m s2 . Other
units of pressure we will encounter:
1 bar ≡ 105 Pa
1 atm ≡ 1.01325 × 105 Pa
1 torr ≡ 133.3 Pa .
Temperature (T ) has a very precise definition from the point of view of statistical mechanics, as we shall see. Many
physical properties depend on the temperature – such properties are called thermometric properties. For example,
the resistivity of a metal ρ(T, p) or the number density of a gas n(T, p) are both thermometric properties, and can
be used to define a temperature scale. Consider the device known as the ‘constant volume gas thermometer’
depicted in Fig. 2.3, in which the volume or pressure of a gas may be used to measure temperature. The gas
is assumed to be in equilibrium at some pressure p, volume V , and temperature T . An incompressible fluid of
density ̺ is used to measure the pressure difference ∆p = p − p0 , where p0 is the ambient pressure at the top of
the reservoir:
p − p0 = ̺g(h2 − h1 ) , (2.7)
where g is the acceleration due to gravity. The height h1 of the left column of fluid in the U-tube provides a
measure of the change in the volume of the gas:
where A is the (assumed constant) cross-sectional area of the left arm of the U-tube. The device can operate in two
modes:
3 We use the symbol d¯ in the differential dW
¯ to indicate that this is not an exact differential. More on this in section 2.4 below.
22 CHAPTER 2. THERMODYNAMICS
Figure 2.3: The constant volume gas thermometer. The gas is placed in thermal contact with an object of temper-
ature T . An incompressible fluid of density ̺ is used to measure the pressure difference ∆p = pgas − p0 .
• Constant pressure mode : The height of the reservoir is adjusted so that the height difference h2 − h1 is held
constant. This fixes the pressure p of the gas. The gas volume still varies with temperature T , and we can
define
T V
= , (2.9)
Tref Vref
where Tref and Vref are the reference temperature and volume, respectively.
• Constant volume mode : The height of the reservoir is adjusted so that h1 = 0, hence the volume of the gas
is held fixed, and the pressure varies with temperature. We then define
T p
= , (2.10)
Tref pref
where Tref and pref are the reference temperature and pressure, respectively.
What should we use for a reference? One might think that a pot of boiling water will do, but anyone who has
gone camping in the mountains knows that water boils at lower temperatures at high altitude (lower pressure).
This phenomenon is reflected in the phase diagram for H2 O, depicted in Fig. 2.4. There are two special points in the
phase diagram, however. One is the triple point, where the solid, liquid, and vapor (gas) phases all coexist. The
second is the critical point, which is the terminus of the curve separating liquid from gas. At the critical point, the
latent heat of transition between liquid and gas phases vanishes (more on this later on). The triple point temperature
Tt at thus unique and is by definition Tt = 273.16 K. The pressure at the triple point is 611.7 Pa = 6.056 × 10−3 atm.
A question remains: are the two modes of the thermometer compatible? E.g. it we boil water at p = p0 = 1 atm,
do they yield the same value for T ? And what if we use a different gas in our measurements? In fact, all these
measurements will in general be incompatible, yielding different results for the temperature T . However, in the
limit that we use a very low density gas, all the results converge. This is because all low density gases behave as
ideal gases, and obey the ideal gas equation of state pV = N kB T .
2.2. WHAT IS THERMODYNAMICS? 23
Figure 2.4: A sketch of the phase diagram of H2 O (water). Two special points are identified: the triple point (Tt , pt )
at which there is three phase coexistence, and the critical point (Tc , pc ), where the latent heat of transformation
from liquid to gas vanishes. Not shown are transitions between several different solid phases.
It is customary in the physical sciences to define certain standard conditions with respect to which any arbitrary
conditions may be compared. In thermodynamics, there is a notion of standard temperature and pressure, abbre-
viated STP. Unfortunately, there are two different definitions of STP currently in use, one from the International
Union of Pure and Applied Chemistry (IUPAC), and the other from the U.S. National Institute of Standards and
Technology (NIST). The two standards are:
To make matters worse, in the past it was customary to define STP as T0 = 0◦ C and p0 = 1 atm. We will use the
NIST definition in this course. Unless I slip and use the IUPAC definition. Figuring out what I mean by STP will
keep you on your toes.
The volume of one mole of ideal gas at STP is then
(
NA kB T0 22.711 ℓ (IUPAC)
V = = (2.11)
p0 24.219 ℓ (NIST) ,
where 1 ℓ = 106 cm3 = 10−3 m3 is one liter. Under the old definition of STP as T0 = 0◦ C and p0 = 1 atm, the
volume of one mole of gas at STP is 22.414 ℓ, which is a figure I remember from my 10th grade chemistry class
with Mr. Lawrence.
24 CHAPTER 2. THERMODYNAMICS
Figure 2.5: As the gas density tends to zero, the readings of the constant volume gas thermometer converge.
Equilibrium is established by the exchange of energy, volume, or particle number between different systems or
subsystems:
Equilibrium is transitive, so
The differential
k
X
dF = Ai dxi (2.12)
i=1
is called exact if there is a function F (x1 , . . . , xk ) whose differential gives the right hand side of eqn. 2.12. In this
case, we have
∂F ∂Ai ∂Aj
Ai = ⇐⇒ = ∀ i, j . (2.13)
∂xi ∂xj ∂xi
4 As we shall see further below, thermomechanical equilibrium in fact leads to constant p/T , and thermochemical equilibrium to constant
µ/T . If there is thermal equilibrium, then T is already constant, and so thermomechanical and thermochemical equilibria then guarantee the
constancy of p and µ.
2.4. MATHEMATICAL INTERLUDE : EXACT AND INEXACT DIFFERENTIALS 25
ZB
dF = F (xB1 , . . . , xBk ) − F (xA1 , . . . , xAk ) , (2.14)
A
from which it follows that the integral of dF around any closed path must vanish:
I
dF = 0 . (2.15)
When the cross derivatives are not identical, i.e. when ∂Ai /∂xj 6= ∂Aj /∂xi , the differential is inexact. In this case,
the integral of dF is path dependent, and does not depend solely on the endpoints.
As an example, consider the differential
dF = K1 y dx + K2 x dy . (2.16)
Let’s evaluate the integral of dF , which is the work done, along each of the two paths in Fig. 2.6:
ZxB ZyB
(I)
W = K1 dx yA + K2 dy xB = K1 yA (xB − xA ) + K2 xB (yB − yA ) (2.17)
xA yA
ZxB ZyB
(II)
W = K1 dx yB + K2 dy xA = K1 yB (xB − xA ) + K2 xA (yB − yA ) . (2.18)
xA yA
Note that in general W (I) 6= W (II) . Thus, if we start at point A, the kinetic energy at point B will depend on the
path taken, since the work done is path-dependent.
26 CHAPTER 2. THERMODYNAMICS
The difference between the work done along the two paths is
I
W (I) − W (II) = dF = (K2 − K1 ) (xB − xA ) (yB − yA ) . (2.19)
Thus, we see that if K1 = K2 , the work is the same for the two paths. In fact, if K1 = K2 , the work would be
path-independent, and would depend only on the endpoints. This is true for any path, and not just piecewise
linear paths of the type depicted in Fig. 2.6. Thus, if K1 = K2 , we are justified in using the notation dF for the
differential in eqn. 2.16; explicitly, we then have F = K1 xy. However, if K1 6= K2 , the differential is inexact, and
we will henceforth write d̄F in such cases.
The first law is a statement of energy conservation, and is depicted in Fig. 2.7. It says, quite simply, that during
a thermodynamic process, the change in a system’s internal energy E is given by the heat energy Q added to the
system, minus the work W done by the system:
∆E = Q − W . (2.20)
We use the symbol d̄ in the differentials d̄Q and d̄W to remind us that these are inexact differentials. The energy
E, however, is a state function, hence dE is an exact differential.
Consider a volume V of fluid held in a flask, initially at temperature T0 , and held at atmospheric pressure. The
internal energy is then E0 = E(T0 , p, V ). Now let us contemplate changing the temperature in two different ways.
The first method (A) is to place the flask on a hot plate until the temperature of the fluid rises to a value T1 . The
second method (B) is to stir the fluid vigorously. In the first case, we add heat QA > 0 but no work is done, so
WA = 0. In the second case, if we thermally insulate the flask and use a stirrer of very low thermal conductivity,
then no heat is added, i.e. QB = 0. However, the stirrer does work −WB > 0 on the fluid (remember W is the work
done by the system). If we end up at the same temperature T1 , then the final energy is E1 = E(T1 , p, V ) in both
cases. We then have
∆E = E1 − E0 = QA = −WB . (2.22)
It also follows that for any cyclic transformation, where the state variables are the same at the beginning and the
end, we have
∆Ecyclic = Q − W = 0 =⇒ Q = W (cyclic) . (2.23)
2.5. THE FIRST LAW OF THERMODYNAMICS 27
A single component system is specified by three state variables. In many applications, the total number of particles
N is conserved, so it is useful to take N as one of the state variables. The remaining two can be (T, V ) or (T, p) or
(p, V ). The differential form of the first law says
dE = d̄Q − d̄W
= d̄Q − p dV + µ dN . (2.24)
The quantity µ is called the chemical potential. Here we shall be interested in the case dN = 0 so the last term will
not enter into our considerations. We ask: how much heat is required in order to make an infinitesimal change in
temperature, pressure, or volume? We start by rewriting eqn. 2.24 as
d̄Q = dE + p dV − µ dN . (2.25)
We now must roll up our sleeves and do some work with partial derivatives.
• (T, V, N ) systems : If the state variables are (T, V, N ), we write
∂E ∂E ∂E
dE = dT + dV + dN . (2.26)
∂T V,N ∂V T,N ∂N T,V
We also write
∂V ∂V ∂V
dV = dT + dp + dN . (2.29)
∂T p,N ∂p T,N ∂N T,p
Then
" # " #
∂E ∂V ∂E ∂V
d̄Q = +p dT + +p dp
∂T p,N ∂T p,N ∂p T,N ∂p T,N
" # (2.30)
∂E ∂V
+ +p − µ dN .
∂N T,p ∂N T,p
cp c̃p cp c̃p
SUBSTANCE (J/mol K) (J/g K) SUBSTANCE (J/mol K) (J/g K)
Air 29.07 1.01 H2 O (25◦ C) 75.34 4.181
Aluminum 24.2 0.897 H2 O (100◦+ C) 37.47 2.08
Copper 24.47 0.385 Iron 25.1 0.450
CO2 36.94 0.839 Lead 26.4 0.127
Diamond 6.115 0.509 Lithium 24.8 3.58
Ethanol 112 2.44 Neon 20.786 1.03
Gold 25.42 0.129 Oxygen 29.38 0.918
Helium 20.786 5.193 Paraffin (wax) 900 2.5
Hydrogen 28.82 5.19 Uranium 27.7 0.116
H2 O (−10◦ C) 38.09 2.05 Zinc 25.3 0.387
Table 2.1: Specific heat (at 25◦ C, unless otherwise noted) of some common substances. (Source: Wikipedia.)
The heat capacity of a body, C, is by definition the ratio d̄Q/dT of the amount of heat absorbed by the body to the
associated infinitesimal change in temperature dT . The heat capacity will in general be different if the body is
heated at constant volume or at constant pressure. Setting dV = 0 gives, from eqn. 2.27,
d̄Q ∂E
CV,N = = . (2.33)
dT V,N ∂T V,N
Unless explicitly stated as otherwise, we shall assume that N is fixed, and will write CV for CV,N and Cp for Cp,N .
The units of heat capacity are energy divided by temperature, e.g. J/K. The heat capacity is an extensive quantity,
scaling with the size of the system. If we divide by the number of moles N/NA , we obtain the molar heat capacity,
sometimes called the molar specific heat: c = C/ν, where ν = N/NA is the number of moles of substance. Specific
heat is also sometimes quoted in units of heat capacity per gram of substance. We shall define
Here m is the mass per particle and M is the mass per mole: M = NA m.
Suppose we raise the temperature of a body from T = TA to T = TB . How much heat is required? We have
ZTB
Q = dT C(T ) , (2.36)
TA
where C = CV or C = Cp depending on whether volume or pressure is held constant. For ideal gases, as we shall
discuss below, C(T ) is constant, and thus
Q
Q = C(TB − TA ) =⇒ TB = TA + . (2.37)
C
2.5. THE FIRST LAW OF THERMODYNAMICS 29
Figure 2.8: Heat capacity CV for one mole of hydrogen (H2 ) gas. At the lowest temperatures, only translational
degrees of freedom are relevant, and f = 3. At around 200 K, two rotational modes are excitable and f = 5. Above
1000 K, the vibrational excitations begin to contribute. Note the logarithmic temperature scale. (Data from H. W.
Wooley et al., Jour. Natl. Bureau of Standards, 41, 379 (1948).)
In metals at very low temperatures one finds C = γT , where γ is a constant5 . We then have
ZTB
Q = dT C(T ) = 21 γ TB2 − TA2 (2.38)
TA
p
TB = TA2 + 2γ −1 Q . (2.39)
The ideal gas equation of state is pV = N kB T . In order to invoke the formulae in eqns. 2.27, 2.30, and 2.32, we
need to know the state function E(T, V, N ). A landmark experiment by Joule in the mid-19th century established
that the energy of a low density gas is independent of its volume6 . Essentially, a gas at temperature T was allowed
to freely expand from one volume V to a larger volume V ′ > V , with no added heat Q and no work W done.
Therefore the energy cannot change. What Joule found was that the temperature also did not change. This means
that E(T, V, N ) = E(T, N ) cannot be a function of the volume.
Since E is extensive, we conclude that
E(T, V, N ) = ν ε(T ) , (2.40)
where ν = N/NA is the number of moles of substance. Note that ν is an extensive variable. From eqns. 2.33 and
2.34, we conclude
CV (T ) = ν ε′ (T ) , Cp (T ) = CV (T ) + νR , (2.41)
where we invoke the ideal gas law to obtain the second of these. Empirically it is found that CV (T ) is temperature
independent over a wide range of T , far enough from boiling point. We can then write CV = ν cV , where ν ≡ N/NA
is the number of moles, and where cV is the molar heat capacity. We then have
cp = cV + R , (2.42)
5 In most metals, the difference between CV and Cp is negligible.
6 See the description in E. Fermi, Thermodynamics, pp. 22-23.
30 CHAPTER 2. THERMODYNAMICS
Figure 2.9: Molar heat capacities cV for three solids. The solid curves correspond to the predictions of the Debye
model, which we shall discuss later.
where R = NA kB = 8.31457 J/mol K is the gas constant. We denote by γ = cp /cV the ratio of specific heat at
constant pressure and at constant volume.
From the kinetic theory of gases, one can show that
monatomic gases: cV = 32 R , cp = 52 R , γ= 5
3
diatomic gases: cV = 25 R , cp = 72 R , γ= 7
5
4
polyatomic gases: cV = 3R , cp = 4R , γ= 3 .
We will conclude in general from noninteracting classical statistical mechanics that the specific heat of a substance
is cv = 21 f R, where f is the number of phase space coordinates, per particle, for which there is a quadratic kinetic
or potential energy function. For example, a point particle has three translational degrees of freedom, and the
kinetic energy is a quadratic function of their conjugate momenta: H0 = (p2x + p2y + p2z )/2m. Thus, f = 3. Diatomic
molecules have two additional rotational degrees of freedom – we don’t count rotations about the symmetry axis
– and their conjugate momenta also appear quadratically in the kinetic energy, leading to f = 5. For polyatomic
molecules, all three Euler angles and their conjugate momenta are in play, and f = 6.
The reason that f = 5 for diatomic molecules rather than f = 6 is due to quantum mechanics. While translational
eigenstates form a continuum, or are quantized in a box with ∆kα = 2π/Lα being very small, since the dimensions
Lα are macroscopic, angular momentum, and hence rotational kinetic energy, is quantized. For rotations about a
principal axis with very low moment of inertia I, the corresponding energy scale ~2 /2I is very large, and a high
temperature is required in order to thermally populate these states. Thus, degrees of freedom with a quantization
energy on the order or greater than ε0 are ‘frozen out’ for temperatures T <
∼ ε0 /kB .
In solids, each atom is effectively connected to its neighbors by springs; such a potential arises from quantum
mechanical and electrostatic consideration of the interacting atoms. Thus, each degree of freedom contributes
to the potential energy, and its conjugate momentum contributes to the kinetic energy. This results in f = 6.
Assuming only lattice vibrations, then, the high temperature limit for cV (T ) for any solid is predicted to be 3R =
24.944 J/mol K. This is called the Dulong-Petit law. The high temperature limit is reached above the so-called Debye
2.5. THE FIRST LAW OF THERMODYNAMICS 31
d̄Q = CV dT + p dV . (2.43)
Invoking the ideal gas law to write p = νRT /V , and remembering CV = ν cV , we have, setting d̄Q = 0,
dT R dV
+ =0. (2.44)
T cV V
We can immediately integrate to obtain
γ−1
T V = constant
d̄Q = 0 =⇒ γ (2.45)
pV = constant
γ 1−γ
T p = constant
where the second two equations are obtained from the first by invoking the ideal gas law. These are all adiabatic
equations of state. Note the difference between the adiabatic equation of state d(pV γ ) = 0 and the isothermal
equation of state d(pV ) = 0. Equivalently, we can write these three conditions as
It turns out that air is a rather poor conductor of heat. This suggests the following model for an adiabatic atmosphere.
The hydrostatic pressure decrease associated with an increase dz in height is dp = −̺g dz, where ̺ is the density
and g the acceleration due to gravity. Assuming the gas is ideal, the density can be written as ̺ = M p/RT , where
M is the molar mass. Thus,
dp Mg
=− dz . (2.47)
p RT
If the height changes are adiabatic, then, from d(T γ p1−γ ) = 0, we have
γ − 1 T dp γ − 1 Mg
dT = =− dz , (2.48)
γ p γ R
with the solution
γ − 1 Mg γ−1 z
T (z) = T0 − z= 1− T0 , (2.49)
γ R γ λ
where T0 = T (0) is the temperature at the earth’s surface, and
RT0
λ= . (2.50)
Mg
With M = 28.88 g and γ = 57 for air, and assuming T0 = 293 K, we find λ = 8.6 km, and dT /dz = −(1−γ −1 ) T0 /λ =
−9.7 K/km. Note that in this model the atmosphere ends at a height zmax = γλ/(γ − 1) = 30 km.
32 CHAPTER 2. THERMODYNAMICS
Recall that
x k
ex = lim 1+ . (2.52)
k→∞ k
Thus, in the limit γ → 1, where k = γ/(γ − 1) → ∞, we have p(z) = p0 exp(−z/λ). Finally, since ̺ ∝ p/T from the
ideal gas law, we have
1
̺(z) γ − 1 z γ−1
= 1− . (2.53)
̺0 γ λ
Consider the situation depicted in Fig. 2.10. A quantity (ν moles) of gas in equilibrium at temperature T and
volume V1 is allowed to expand freely into an evacuated chamber of volume V2 by the removal of a barrier.
Clearly no work is done on or by the gas during this process, hence W = 0. If the walls are everywhere insulating,
so that no heat can pass through them, then Q = 0 as well. The First Law then gives ∆E = Q − W = 0, and there
is no change in energy.
If the gas is ideal, then since E(T, V, N ) = N cV T , then ∆E = 0 gives ∆T = 0, and there is no change in tem-
perature. (If the walls are insulating against the passage of heat, they must also prevent the passage of particles,
so ∆N = 0.) There is of course a change in volume: ∆V = V2 , hence there is a change in pressure. The initial
pressure is p = N kB T /V1 and the final pressure is p′ = N kB T /(V1 + V2 ).
If the gas is nonideal, then the temperature will in general change. Suppose, for example, that E(T, V, N ) =
α V x N 1−x T y , where α, x, and y are constants. This form is properly extensive: if V and N double, then E doubles.
If the volume changes from V to V ′ under an adiabatic free expansion, then we must have, from ∆E = 0,
x y x/y
V T′ V
= =⇒ T′ = T · . (2.54)
V′ T V′
If x/y > 0, the temperature decreases upon the expansion. If x/y < 0, the temperature increases. Without an
equation of state, we can’t say what happens to the pressure.
Adiabatic free expansion of a gas is a spontaneous process, arising due to the natural internal dynamics of the system.
It is also irreversible. If we wish to take the gas back to its original state, we must do work on it to compress it. If
the gas is ideal, then the initial and final temperatures are identical, so we can place the system in thermal contact
with a reservoir at temperature T and follow a thermodynamic path along an isotherm. The work done on the gas
during compression is then
ZVi
dV Vf V2
W = −N kB T = N kB T ln = N kB T ln 1 + (2.55)
V Vi V1
Vf
R
The work done by the gas is W = p dV = −W. During the compression, heat energy Q = W < 0 is transferred to
the gas from the reservoir. Thus, Q = W > 0 is given off by the gas to its environment.
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 33
Figure 2.10: In the adiabatic free expansion of a gas, there is volume expansion with no work or heat exchange
with the environment: ∆E = Q = W = 0.
A heat engine is a device which takes a thermodynamic system through a repeated cycle which can be represented
as a succession of equilibrium states: A → B → C · · · → A. The net result of such a cyclic process is to convert
heat into mechanical work, or vice versa.
For a system in equilibrium at temperature T , there is a thermodynamically large amount of internal energy
stored in the random internal motion of its constituent particles. Later, when we study statistical mechanics, we
will see how each ‘quadratic’ degree of freedom in the Hamiltonian contributes 21 kB T to the total internal energy.
An immense body in equilibrium at temperature T has an enormous heat capacity C, hence extracting a finite
quantity of heat Q from it results in a temperature change ∆T = −Q/C which is utterly negligible. Such a body
is called a heat bath, or thermal reservoir. A perfect engine would, in each cycle, extract an amount of heat Q from the
bath and convert it into work. Since ∆E = 0 for a cyclic process, the First Law then gives W = Q. This situation is
depicted schematically in Fig. 2.11. One could imagine running this process virtually indefinitely, slowly sucking
energy out of an immense heat bath, converting the random thermal motion of its constituent molecules into
useful mechanical work. Sadly, this is not possible:
A transformation whose only final result is to extract heat from a source at fixed temperature and
transform that heat into work is impossible.
This is known as the Postulate of Lord Kelvin. It is equivalent to the postulate of Clausius,
A transformation whose only result is to transfer heat from a body at a given temperature to a body at
higher temperature is impossible.
These postulates which have been repeatedly validated by empirical observations, constitute the Second Law of
Thermodynamics.
34 CHAPTER 2. THERMODYNAMICS
Figure 2.11: A perfect engine would extract heat Q from a thermal reservoir at some temperature T and convert it
into useful mechanical work W . This process is alas impossible, according to the Second Law of thermodynamics.
The inverse process, where work W is converted into heat Q, is always possible.
While it is not possible to convert heat into work with 100% efficiency, it is possible to transfer heat from one
thermal reservoir to another one, at lower temperature, and to convert some of that heat into work. This is what
an engine does. The energy accounting for one cycle of the engine is depicted in the left hand panel of Fig. 2.12.
An amount of heat Q2 > 0 is extracted- from the reservoir at temperature T2 . Since the reservoir is assumed to
be enormous, its temperature change ∆T2 = −Q2 /C2 is negligible, and its temperature remains constant – this
is what it means for an object to be a reservoir. A lesser amount of heat, Q1 , with 0 < Q1 < Q2 , is deposited
in a second reservoir at a lower temperature T1 . Its temperature change ∆T1 = +Q1 /C1 is also negligible. The
difference W = Q2 − Q1 is extracted as useful work. We define the efficiency, η, of the engine as the ratio of the
work done to the heat extracted from the upper reservoir, per cycle:
W Q
η= =1− 1 . (2.56)
Q2 Q2
This is a natural definition of efficiency, since it will cost us fuel to maintain the temperature of the upper reservoir
over many cycles of the engine. Thus, the efficiency is proportional to the ratio of the work done to the cost of the
fuel.
A refrigerator works according to the same principles, but the process runs in reverse. An amount of heat Q1 is
extracted from the lower reservoir – the inside of our refrigerator – and is pumped into the upper reservoir. As
Clausius’ form of the Second Law asserts, it is impossible for this to be the only result of our cycle. Some amount
of work W must be performed on the refrigerator in order for it to extract the heat Q1 . Since ∆E = 0 for the cycle,
a heat Q2 = W + Q1 must be deposited into the upper reservoir during each cycle. The analog of efficiency here
is called the coefficient of refrigeration, κ, defined as
Q1 Q1
κ= = . (2.57)
W Q2 − Q 1
Thus, κ is proportional to the ratio of the heat extracted to the cost of electricity, per cycle.
Please note the deliberate notation here. I am using symbols Q and W to denote the heat supplied to the engine
(or refrigerator) and the work done by the engine, respectively, and Q and W to denote the heat taken from the
engine and the work done on the engine.
A perfect engine has Q1 = 0 and η = 1; a perfect refrigerator has Q1 = Q2 and κ = ∞. Both violate the Second
Law. Sadi Carnot7 (1796 – 1832) realized that a reversible cyclic engine operating between two thermal reservoirs
7 Carnot died during cholera epidemic of 1832. His is one of the 72 names engraved on the Eiffel Tower.
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 35
Figure 2.12: An engine (left) extracts heat Q2 from a reservoir at temperature T2 and deposits a smaller amount of
heat Q1 into a reservoir at a lower temperature T1 , during each cycle. The difference W = Q2 − Q1 is transformed
into mechanical work. A refrigerator (right) performs the inverse process, drawing heat Q1 from a low tempera-
ture reservoir and depositing heat Q2 = Q1 + W into a high temperature reservoir, where W is the mechanical (or
electrical) work done per cycle.
must produce the maximum amount of work W , and that the amount of work produced is independent of the
material properties of the engine. We call any such engine a Carnot engine.
The efficiency of a Carnot engine may be used to define a temperature scale. We know from Carnot’s observations
that the efficiency ηC can only be a function of the temperatures T1 and T2 : ηC = ηC (T1 , T2 ). We can then define
T1
≡ 1 − ηC (T1 , T2 ) . (2.58)
T2
Below, in §2.6.4, we will see that how, using an ideal gas as the ‘working substance’ of the Carnot engine, this
temperature scale coincides precisely with the ideal gas temperature scale from §2.2.4.
The Carnot engine is the most efficient engine possible operating between two thermal reservoirs. To see this,
let’s suppose that an amazing wonder engine has an efficiency even greater than that of the Carnot engine. A key
feature of the Carnot engine is its reversibility – we can just go around its cycle in the opposite direction, creating
a Carnot refrigerator. Let’s use our notional wonder engine to drive a Carnot refrigerator, as depicted in Fig. 2.13.
We assume that
W W′
= ηwonder > ηCarnot = ′ . (2.59)
Q2 Q2
But from the figure, we have W = W ′ , and therefore the heat energy Q′2 − Q2 transferred to the upper reservoir is
positive. From
W = Q2 − Q1 = Q′2 − Q′1 = W ′ , (2.60)
we see that this is equal to the heat energy extracted from the lower reservoir, since no external work is done on
the system:
Q′2 − Q2 = Q′1 − Q1 > 0 . (2.61)
36 CHAPTER 2. THERMODYNAMICS
Therefore, the existence of the wonder engine entails a violation of the Second Law. Since the Second Law is
correct – Lord Kelvin articulated it, and who are we to argue with a Lord? – the wonder engine cannot exist.
We further conclude that all reversible engines running between two thermal reservoirs have the same efficiency, which is
the efficiency of a Carnot engine. For an irreversible engine, we must have
W Q T
η= = 1 − 1 ≤ 1 − 1 = ηC . (2.62)
Q2 Q2 T2
Thus,
Q2 Q
− 1 ≤0. (2.63)
T2 T1
Let us now consider a specific cycle, known as the Carnot cycle, depicted in Fig. 2.14. The cycle consists of two
adiabats and two isotherms. The work done per cycle is simply the area inside the curve on our p − V diagram:
I
W = p dV . (2.64)
The gas inside our Carnot engine is called the ‘working substance’. Whatever it may be, the system obeys the First
Law,
dE = d̄Q − d̄W = d̄Q − p dV . (2.65)
We will now assume that the working material is an ideal gas, and we compute W as well as Q1 and Q2 to find
the efficiency of this cycle. In order to do this, we will rely upon the ideal gas equations,
νRT
E= , pV = νRT , (2.66)
γ−1
where γ = cp /cv = 1 + f2 , where f is the effective number of molecular degrees of freedom contributing to the
internal energy. Recall f = 3 for monatomic gases, f = 5 for diatomic gases, and f = 6 for polyatomic gases. The
finite difference form of the first law is
∆E = Ef − Ei = Qif − Wif , (2.67)
where i denotes the initial state and f the final state.
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 37
Figure 2.14: The Carnot cycle consists of two adiabats (dark red) and two isotherms (blue).
AB: This stage is an isothermal expansion at temperature T2 . It is the ‘power stroke’ of the engine. We have
ZVB
νRT2 VB
WAB = dV = νRT2 ln (2.68)
V VA
VA
νRT2
EA = EB = , (2.69)
γ−1
hence
VB
QAB = ∆EAB + WAB = νRT2 ln . (2.70)
VA
BC: This stage is an adiabatic expansion. We have
QBC = 0 (2.71)
νR
∆EBC = EC − EB = (T − T2 ) . (2.72)
γ−1 1
The energy change is negative, and the heat exchange is zero, so the engine still does some work during this
stage:
νR
WBC = QBC − ∆EBC = (T − T1 ) . (2.73)
γ−1 2
CD: This stage is an isothermal compression, and we may apply the analysis of the isothermal expansion, mutatis
mutandis:
ZVD
νRT2 VD
WCD = dV = νRT1 ln (2.74)
V VC
VC
νRT1
EC = ED = , (2.75)
γ−1
hence
VD
QCD = ∆ECD + WCD = νRT1 ln . (2.76)
VC
38 CHAPTER 2. THERMODYNAMICS
DA: This last stage is an adiabatic compression, and we may draw on the results from the adiabatic expansion in
BC:
QDA = 0 (2.77)
νR
∆EDA = ED − EA = (T − T1 ) . (2.78)
γ−1 2
The energy change is positive, and the heat exchange is zero, so work is done on the engine:
νR
WDA = QDA − ∆EDA = (T − T2 ) . (2.79)
γ−1 1
We now add up all the work values from the individual stages to get for the cycle
Since we are analyzing a cyclic process, we must have ∆E = 0, we must have Q = W , which can of course be
verified explicitly, by computing Q = QAB + QBC + QCD + QDA. To finish up, recall the adiabatic ideal gas equation
of state, d(T V γ−1 ) = 0. This tells us that
W T
η= =1− 1 . (2.86)
QAB T2
Many other engine cycles are possible. The Stirling cycle, depicted in Fig. 2.15, consists of two isotherms and two
isochores. Recall the isothermal ideal gas equation of state, d(pV ) = 0. Thus, for an ideal gas Stirling cycle, we
have
pA V1 = pB V2 , pD V1 = pC V2 , (2.87)
which says
pB p V
= C = 1 . (2.88)
pA pD V2
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 39
Figure 2.15: A Stirling cycle consists of two isotherms (blue) and two isochores (green).
AB: This isothermal expansion is the power stroke. Assuming ν moles of ideal gas throughout, we have pV =
νRT2 = p1 V1 , hence
ZV2
νRT2 V2
WAB = dV = νRT2 ln . (2.89)
V V1
V1
νR(T1 − T2 )
∆EBC = EC − EB = , (2.90)
γ−1
which is negative. Since WBC = 0, we have QBC = ∆EBC .
CD: Isothermal compression. Clearly
ZV1
νRT1 V2
WCD = dV = −νRT1 ln . (2.91)
V V1
V2
Figure 2.16: An Otto cycle consists of two adiabats (dark red) and two isochores (green).
The Otto cycle is a rough approximation to the physics of a gasoline engine. It consists of two adiabats and two
isochores, and is depicted in Fig. 2.16. Assuming an ideal gas, along the adiabats we have d(pV γ ) = 0. Thus,
which says γ
pB p V1
= C = . (2.96)
pA pD V2
AB: Adiabatic expansion, the power stroke. The heat transfer is QAB = 0, so from the First Law we have WAB =
−∆EAB = EA − EB , thus " γ−1 #
pA V1 − pB V2 pA V1 V1
WAB = = 1− . (2.97)
γ−1 γ−1 V2
Note that this result can also be obtained from the adiabatic equation of state pV γ = pA V1γ :
BC: Isochoric cooling (exhaust); dV = 0 hence WBC = 0. The heat QBC absorbed is then
V2
QBC = EC − EB = (p − pB ) . (2.99)
γ−1 C
In a realistic engine, this is the stage in which the old burned gas is ejected and new gas is inserted.
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 41
Figure 2.17: A Diesel cycle consists of two adiabats (dark red), one isobar (light blue), and one isochore (green).
DA: Isochoric heating, i.e. the combustion of the gas. As with BC we have dV = 0, and thus WDA = 0. The heat
QDA absorbed by the gas is then
V1
QDA = EA − ED = (p − pD ) . (2.101)
γ−1 A
The ratio V2 /V1 is called the compression ratio. We can make our Otto cycle more efficient simply by increasing the
compression ratio. The problem with this scheme is that if the fuel mixture becomes too hot, it will spontaneously
‘preignite’, and the pressure will jump up before point D in the cycle is reached. A Diesel engine avoids preignition
by compressing the air only, and then later spraying the fuel into the cylinder when the air temperature is sufficient
for fuel ignition. The rate at which fuel is injected is adjusted so that the ignition process takes place at constant
pressure. Thus, in a Diesel engine, step DA is an isobar. The compression ratio is r ≡ VB /VD , and the cutoff ratio
is s ≡ VA /VD . This refinement of the Otto cycle allows for higher compression ratios (of about 20) in practice, and
greater engine efficiency.
42 CHAPTER 2. THERMODYNAMICS
pA VA − pB VB p V − pD VD
W = pA (VA − VD ) + + C C
γ −1 γ−1
(2.104)
γ pA (VA − VD ) (pB − pC )VB
= −
γ−1 γ−1
and
γ pA (VA − VD )
QDA = . (2.105)
γ−1
To find the efficiency, we will need to eliminate pB and pC in favor of pA using the adiabatic equation of state
d(pV γ ) = 0. Thus, γ γ
VA VD
pB = pA · , pC = pA · , (2.106)
VB VB
where we’ve used pD = pA and VC = VB . Putting it all together, the efficiency of the Diesel cycle is
W 1 r1−γ (sγ − 1)
η= =1− . (2.107)
QDA γ s−1
Our final example is the Joule-Brayton cycle, depicted in Fig. 2.18, consisting of two adiabats and two isobars.
Along the adiabats we have Thus,
which says
γ −1
VD V p2
= C = . (2.109)
VA VB p1
ZVB
WAB = dV p2 = p2 (VB − VA ) (2.110)
VA
p2 (VB − VA )
∆EAB = EB − EA = (2.111)
γ−1
γ p2 (VB − VA )
QAB = ∆EAB + WAB = . (2.112)
γ−1
BC: Adiabatic expansion; QBC = 0 and WBC = EB − EC . The work done by the gas is
p2 VB − p1 VC p2 VB p1 VC
WBC = = 1− ·
γ−1 γ−1 p2 VB
" 1−γ −1 # (2.113)
p2 VB p1
= 1− .
γ−1 p2
2.6. HEAT ENGINES AND THE SECOND LAW OF THERMODYNAMICS 43
Figure 2.18: A Joule-Brayton cycle consists of two adiabats (dark red) and two isobars (light blue).
ZVD 1−γ −1
p1
WCD = dV p1 = p1 (VD − VC ) = −p2 (VB − VA ) (2.114)
p2
VC
p1 (VD − VC )
∆ECD = ED − EC = (2.115)
γ−1
1−γ −1
γ p2 p1
QCD = ∆ECD + WCD =− (V − VA ) . (2.116)
γ−1 B p2
DA: Adiabatic expansion; QDA = 0 and WDA = ED − EA . The work done by the gas is
p1 VD − p2 VA p V p V
WDA = =− 2 A 1− 1 · D
γ−1 γ−1 p2 VA
" 1−γ −1 # (2.117)
p V p1
=− 2 A 1− .
γ−1 p2
While the Carnot engine described above in §2.6.4 has maximum efficiency, it is practically useless, because the
isothermal processes must take place infinitely slowly in order for the working material to remain in thermal
equilibrium with each reservoir. Thus, while the work done per cycle is finite, the cycle period is infinite, and the
engine power is zero.
A modification of the ideal Carnot cycle is necessary to create a practical engine. The idea8 is as follows. During
the isothermal expansion stage, the working material is maintained at a temperature T2w < T2 . The temperature
difference between the working material and the hot reservoir drives a thermal current,
d̄Q2
= κ2 (T2 − T2w ) . (2.120)
dt
Here, κ2 is a transport coefficient which describes the thermal conductivity of the chamber walls, multiplied by a
geometric parameter (which is the ratio of the total wall area to its thickness). Similarly, during the isothermal
compression, the working material is maintained at a temperature T1w > T1 , which drives a thermal current to
the cold reservoir,
d̄Q1
= κ1 (T1w − T1 ) . (2.121)
dt
Now let us assume that the upper isothermal stage requires a duration ∆t2 and the lower isotherm a duration
∆t1 . Then
We optimize the engine by maximizing P with respect to the temperatures T1w and T2w . This yields
p
T2 − T1 T2
T2w = T2 − p (2.128)
1 + κ2 /κ1
p
T1 T2 − T1
T1w = T1 + p . (2.129)
1 + κ1 /κ2
8 See F. L. Curzon and B. Ahlborn, Am. J. Phys. 43, 22 (1975).
2.7. THE ENTROPY 45
Table 2.2: Observed performances of real heat engines, taken from table 1 from Curzon and Albhorn (1975).
Table 2.2, taken from the article of Curzon and Albhorn (1975), shows how the efficiency of this practical Carnot
cycle, given by eqn. 2.130, rather accurately predicts the efficiencies of functioning power plants.
The Second Law guarantees us that an engine operating between two heat baths at temperatures T1 and T2 must
satisfy
Q1 Q
+ 2 ≤0, (2.133)
T1 T2
with the equality holding for reversible processes. This is a restatement of eqn. 2.63, after writing Q1 = −Q1 for
the heat transferred to the engine from reservoir #1. Consider now an arbitrary curve in the p − V plane. We can
describe such a curve, to arbitrary accuracy, as a combination of Carnot cycles, as shown in Fig. 2.19. Each little
Carnot cycle consists of two adiabats and two isotherms. We then conclude
XQ I
i d̄Q
−→ ≤0, (2.134)
i
T i T
C
with equality holding if all the cycles are reversible. Rudolf Clausius, in 1865, realized that one could then define a
new state function, which he called the entropy, S, that depended only on the initial and final states of a reversible
46 CHAPTER 2. THERMODYNAMICS
Figure 2.19: An arbitrarily shaped cycle in the p − V plane can be decomposed into a number of smaller Carnot
cycles. Red curves indicate isotherms and blue curves adiabats, with γ = 35 .
process:
ZB
d̄Q d̄Q
dS = =⇒ SB − SA = . (2.135)
T T
A
Eqn. 2.135 determines the entropy up to a constant. By choosing a standard state Υ, we can define SΥ = 0, and
then by taking A = Υ in the above equation, we can define the absolute entropy S for any state. However, it
turns out that this seemingly arbitrary constant SΥ in the entropy does have consequences, for example in the
theory of gaseous equilibrium. The proper definition of entropy, from the point of view of statistical mechanics,
will lead us to understand how the zero temperature entropy of a system is related to its quantum mechanical
ground state degeneracy. Walther Nernst, in 1906, articulated a principle which is sometimes called the Third
Law of Thermodynamics,
Again, this is not quite correct, and quantum mechanics tells us that S(T = 0) = kB ln g, where g is the ground
state degeneracy. Nernst’s law holds when g = 1.
We can combine the First and Second laws to write
For a cyclic process, whether reversible or not, the change in entropy around a cycle is zero: ∆SCYC = 0. This
is because the entropy S is a state function, with a unique value for every equilibrium state. A cyclical process
returns to the same equilibrium state, hence S must return as well to its corresponding value from the previous
cycle.
Consider now a general engine, as in Fig. 2.12. Let us compute the total entropy change in the entire Universe
over one cycle. We have
(∆S)TOTAL = (∆S)ENGINE + (∆S)HOT + (∆S)COLD , (2.137)
written as a sum over entropy changes of the engine itself, the hot reservoir, and the cold reservoir9 . Clearly
(∆S)ENGINE = 0. The changes in the reservoir entropies are
Z
d̄QHOT Q
(∆S)HOT = =− 2 < 0 (2.138)
T T2
T =T2
Z
d̄QCOLD Q Q
(∆S)COLD = = 1 =− 1 > 0, (2.139)
T T1 T1
T =T1
because the hot reservoir loses heat Q2 > 0 to the engine, and the cold reservoir gains heat Q1 = −Q1 > 0 from
the engine. Therefore,
Q1 Q
(∆S)TOTAL = − + 2 ≥0. (2.140)
T1 T2
Thus, for a reversible cycle, the net change in the total entropy of the engine plus reservoirs is zero. For an
irreversible cycle, there is an increase in total entropy, due to spontaneous processes.
This says that the energy E is a function of the entropy S, the generalized displacements {Xj }, and the particle
numbers {Na }:
E = E S, {Xj }, {Na } . (2.143)
Furthermore, we have
!
∂E ∂E ∂E
T = , yj = , µa = (2.144)
∂S {X ∂Xj ∂Na
j ,Na } S,{Xi(6=j) ,Na } S,{Xj ,Nb(6=a) }
We now differentiate the LHS and RHS above with respect to λ, setting λ = 1 afterward. The result is
∂E X ∂E X ∂E
E=S + Xj + Na
∂S j
∂X j a
∂N a
X X (2.146)
= TS + yj Xj + µa Na .
j a
Mathematically astute readers will recognize this result as an example of Euler’s theorem for homogeneous func-
tions. Taking the differential of eqn. 2.146, and then subtracting eqn. 2.142, we obtain
X X
S dT + Xj dyj + Na dµa = 0 . (2.147)
j a
This is called the Gibbs-Duhem relation. It says that there is one equation of state which may be written in terms of
all the intensive quantities alone. For example, for a single component system, we must have p = p(T, µ), which
follows from
S dT − V dp + N dµ = 0 . (2.148)
1 p µ
dS = dE + dV − dN
T T T (2.149)
1 dT p µ
= 2 f N kB + dV + 12 f kB − dN .
T T T
Integrating, we obtain
S(T, V, N ) = 12 f N kB ln T + N kB ln V + ϕ(N ) , (2.151)
where ϕ(N ) is an arbitrary function. Extensivity of S places restrictions on ϕ(N ), so that the most general case is
V
S(T, V, N ) = 12 f N kB ln T + N kB ln + Na , (2.152)
N
where b = a − 12 f kB ln( 21 f kB ) is another constant. When we study statistical mechanics, we will find that for the
monatomic ideal gas the entropy is
" #
V
S(T, V, N ) = N kB 52 + ln , (2.154)
N λ3T
p
where λT = 2π~2 /mkB T is the thermal wavelength, which involved Planck’s constant. Let’s now contrast two
illustrative cases.
2.7. THE ENTROPY 49
• Adiabatic free expansion – Suppose the volume freely expands from Vi to Vf = r Vi , with r > 1. Such an
expansion can be effected by a removal of a partition between two chambers that are otherwise thermally
insulated (see Fig. 2.10). We have already seen how this process entails
∆E = Q = W = 0 . (2.155)
∆S = Sf − Si = N kB ln r . (2.156)
• Reversible adiabatic expansion – If the gas expands quasistatically and reversibly, then S = S(E, V, N ) holds
everywhere along the thermodynamic path. We then have, assuming dN = 0,
dE dV
0 = dS = 12 f N kB + N kB
E V (2.157)
f /2
= N kB d ln V E .
Integrating, we find
2/f
E V0
= . (2.158)
E0 V
Thus,
Ef = r−2/f Ei ⇐⇒ Tf = r−2/f Ti . (2.159)
aS 3
E(S, V, N ) = , (2.160)
NV
where a is a constant. We have
dE = T dS − p dV + µ dN , (2.161)
and therefore
∂E 3aS 2
T = = (2.162)
∂S V,N NV
∂E aS 3
p=− = (2.163)
∂V S,N NV 2
∂E aS 3
µ= =− 2 . (2.164)
∂N S,V N V
Choosing any two of these equations, we can eliminate S, which is inconvenient for experimental purposes. This
yields three equations of state,
T3 V T3 N p N
= 27a , = 27a , =− , (2.165)
p2 N µ2 V µ V
1/2
NV T
S= . (2.166)
3a
We then have
1/2
∂S 1 NV T N T2
CV = T = = , (2.167)
∂T V,N 2 3a 18a p
where the last equality on the RHS follows upon invoking the first of the equations of state in eqn. 2.165. To find
Cp , we eliminate V from eqns. 2.162 and 2.163, obtaining T 2 /p = 9aS/N . From this we obtain
∂S 2N T 2
Cp = T = . (2.168)
∂T p,N 9a p
Thus, Cp /CV = 4.
We can derive still more. To find the isothermal compressibility κT = − V1 ∂V
∂p T,N , use the first of the equations of
state in eqn. 2.165. To derive the adiabatic compressibility κS = − V1 ∂V
∂p S,N , use eqn. 2.163, and then eliminate
the inconvenient variable S.
Suppose we use this system as the working substance for a Carnot engine. Let’s compute the work done and
the engine efficiency. To do this, it is helpful to eliminate S in the expression for the energy, and to rewrite the
equation of state:
r r
N 1/2 3/2 N T 3/2
E = pV = V T , p= . (2.169)
27a 27a V 1/2
E
dT = 0 : √ = constant (2.170)
V
Furthermore, since
r
N 3/2 dV
d̄W T = T 1/2
= 2 dE T , (2.171)
27a V
we conclude that
dT = 0 : Wif = 2(Ef − Ei ) , Qif = Ef − Ei + Wif = 3(Ef − Ei ) . (2.172)
E
d̄Q = 0 : T V = constant , = constant , EV = constant (2.173)
T
Since
s !
VB
QAB = 3(EB − EA ) = 23 WAB = 3 − 1 EA , (2.180)
VA
If we can measure the heat capacity CV (T ) or Cp (T ) of a substance as a function of temperature down to the
lowest temperatures, then we can measure the entropy. At constant pressure, for example, we have T dS = Cp dT ,
hence
ZT
Cp (T ′ )
S(p, T ) = S(p, T = 0) + dT ′ . (2.182)
T′
0
The zero temperature entropy is S(p, T = 0) = kB ln g where g is the quantum ground state degeneracy at pressure
p. In all but highly unusual cases, g = 1 and S(p, T = 0) = 0.
Thermodynamic systems may do work on their environments. Under certain constraints, the work done may be
bounded from above by the change in an appropriately defined thermodynamic potential.
52 CHAPTER 2. THERMODYNAMICS
2.8.1 Energy E
Suppose we wish to create a thermodynamic system from scratch. Let’s imagine that we create it from scratch in
a thermally insulated box of volume V . The work we must to to assemble the system is then
W=E. (2.183)
After we bring all the constituent particles together, pulling them in from infinity (say), the system will have total
energy E. After we finish, the system may not be in thermal equilibrium. Spontaneous processes will then occur
so as to maximize the system’s entropy, but the internal energy remains at E.
We have, from the First Law, dE = d̄Q − d̄W . For equilibrium systems, we have
dE = T dS − p dV + µ dN , (2.184)
dE ≤ T dS − p dV + µ dN . (2.186)
This form is valid for single component systems and is easily generalized to multicomponent systems, or magnetic
systems, etc. Now consider a process at fixed (S, V, N ). We then have dE ≤ 0. This says that spontaneous
processes in a system with dS = dV = dN = 0 always lead to a reduction in the internal energy E. Therefore,
spontaneous processes drive the internal energy E to a minimum in systems at fixed (S, V, N ).
Allowing for other work processes, we have
d̄W ≤ T dS − dE . (2.187)
Hence, the work done by a thermodynamic system under conditions of constant entropy is bounded above by −dE, and the
maximum d̄W is achieved for a reversible process.
It is useful to define the quantity
d̄Wfree = d̄W − p dV , (2.188)
which is the differential work done by the system other than that required to change its volume. Then
d̄Wfree ≤ T dS − p dV − dE , (2.189)
and we conclude that for systems at fixed (S, V ) that d̄Wfree ≤ −dE.
Suppose that when we spontaneously create our system while it is in constant contact with a thermal reservoir
at temperature T . Then as we create our system, it will absorb heat from the reservoir. Therefore, we don’t have
to supply the full internal energy E, but rather only E − Q, since the system receives heat energy Q from the
reservoir. In other words, we must perform work
W = E − TS (2.190)
2.8. THERMODYNAMIC POTENTIALS 53
to create our system, if it is constantly in equilibrium at temperature T . The quantity E − T S is known as the
Helmholtz free energy, F , which is related to the energy E by a Legendre transformation,
F = E − TS . (2.191)
The general properties of Legendre transformations are discussed in Appendix II, §2.16.
Under equilibrium conditions, we have
dF = −S dT − p dV + µ dN . (2.192)
dF ≤ −S dT − p dV + µ dN . (2.194)
The equality holds for reversible processes, and the inequality for spontaneous processes. Therefore, spontaneous
processes drive the Helmholtz free energy F to a minimum in systems at fixed (T, V, N ).
We may also write
d̄W ≤ −S dT − dF , (2.195)
In other words, the work done by a thermodynamic system under conditions of constant temperature is bounded above by
−dF , and the maximum d̄W is achieved for a reversible process. We also have
d̄Wfree ≤ −S dT − p dV − dF , (2.196)
2.8.3 Enthalpy H
Suppose that when we spontaneously create our system while it is thermally insulated, but in constant mechanical
contact with a ‘volume bath’ at pressure p. For example, we could create our system inside a thermally insulated
chamber with one movable wall where the external pressure is fixed at p. Thus, when creating the system, in
addition to the system’s internal energy E, we must also perform work pV in order to make room for the it. In
other words, we must perform work
W = E + pV . (2.197)
The quantity E + pV is known as the enthalpy, H. (We use the sans-serif symbol H for enthalpy to avoid confusing
it with magnetic field, H.)
The enthalpy is obtained from the energy via a different Legendre transformation:
H = E + pV . (2.198)
In equilibrium, then,
dH = T dS + V dp + µ dN , (2.199)
which says H = H(S, p, N ), with
∂H ∂H ∂H
T = , V = , µ= . (2.200)
∂S p,N ∂p S,N ∂N S,p
54 CHAPTER 2. THERMODYNAMICS
In general, we have
dH ≤ T dS + V dp + µ dN , (2.201)
hence spontaneous processes drive the enthalpy H to a minimum in systems at fixed (S, p, N ).
For general systems,
dH ≤ T dS − d̄W + p dV + V dp , (2.202)
hence
d̄Wfree ≤ T dS + V dp − dH , (2.203)
and we conclude, for systems at fixed (S, p), that d̄Wfree ≤ −dH.
If we create a thermodynamic system at conditions of constant temperature T and constant pressure p, then it
absorbs heat energy Q = T S from the reservoir and we must expend work energy pV in order to make room for
it. Thus, the total amount of work we must do in assembling our system is
W = E − T S + pV . (2.204)
G = E − T S + pV (2.205)
dG = −S dT + V dp + µ dN , (2.206)
therefore
G = µN . (2.209)
hence spontaneous processes drive the Gibbs free energy G to a minimum in systems at fixed (T, p, N ). For general
systems,
d̄Wfree ≤ −S dT + V dp − dG . (2.211)
Accordingly, we conclude, for systems at fixed (T, p), that d̄Wfree ≤ −dG.
2.9. MAXWELL RELATIONS 55
The grand potential, sometimes called the Landau free energy, is defined by
Ω = E − T S − µN . (2.212)
Its differential is
dΩ = −S dT − p dV − N dµ , (2.213)
hence
∂Ω ∂Ω ∂Ω
−S = , −p = , −N = . (2.214)
∂T V,µ ∂V T,µ ∂µ T,V
Maxwell relations are conditions equating certain derivatives of state variables which follow from the exactness
of the differentials of the various state functions.
dE = T dS − p dV + µ dN , (2.218)
and therefore
∂E ∂E ∂E
T = , −p = , µ= . (2.219)
∂S V,N ∂V S,N ∂N S,V
dF = −S dT − p dV + µ dN , (2.223)
and therefore
∂F ∂F ∂F
−S = , −p = , µ= . (2.224)
∂T V,N ∂V T,N ∂N T,V
dG = −S dT + V dp + µ dN , (2.233)
dΩ = −S dT − p dV − N dµ , (2.238)
hence
∂Ω ∂Ω ∂Ω
−S = , −p = , −N = . (2.239)
∂T V,µ ∂V T,µ ∂µ T,V
We can also derive Maxwell relations based on the entropy S(E, V, N ) itself. For example, we have
1 p µ
dS = dE + dV − dN . (2.243)
T T T
et cetera.
58 CHAPTER 2. THERMODYNAMICS
We have up until now assumed a generalized force-displacement pair (y, X) = (−p, V ). But the above results also
generalize to e.g. magnetic systems, where (y, X) = (H, M ). In general, we have
F = E − TS dF = −S dT + y dX + µ dN (2.246)
H = E − yX dH = T dS − X dy + µ dN (2.247)
G = E − T S − yX dG = −S dT − X dy + µ dN (2.248)
Ω = E − T S − µN dΩ = −S dT + y dX − N dµ . (2.249)
Generalizing (−p, V ) → (y, X), we also obtain, mutatis mutandis, the following Maxwell relations:
∂T ∂y ∂T ∂µ ∂y ∂µ
= = =
∂X S,N ∂S X,N ∂N S,X ∂S X,N ∂N S,X ∂X S,N
∂T ∂X ∂T ∂µ ∂X ∂µ
=− = =−
∂y S,N ∂S y,N ∂N S,y ∂S y,N ∂N S,y ∂y S,N
∂S ∂y ∂S ∂µ ∂y ∂µ
=− =− =
∂X T,N ∂T X,N ∂N T,X ∂T X,N ∂N T,X ∂X T,N
∂S ∂X ∂S ∂µ ∂X ∂µ
= =− =−
∂y T,N ∂T y,N ∂N T,y ∂T y,N ∂N T,y ∂y T,N
∂S ∂y ∂S ∂N ∂y ∂N
=− = =− .
∂X T,µ ∂T X,µ ∂µ T,X ∂T X,µ ∂µ T,X ∂X T,µ
Suppose we have two systems, A and B, which are free to exchange energy, volume, and particle number, subject
to overall conservation rules
EA + EB = E , VA + VB = V , NA + NB = N , (2.250)
where E, V , and N are fixed. Now let us compute the change in the total entropy of the combined systems when
they are allowed to exchange energy, volume, or particle number. We assume that the entropy is additive, i.e.
" # " #
∂SA ∂SB ∂SA ∂SB
dS = − dEA + − dVA
∂EA V ,N ∂EB V ,N ∂VA E ,N ∂VB E ,N
A A B B A A B B
" #
∂SA ∂SB
+ − dNA . (2.251)
∂NA E ,V ∂NB E ,V
A A B B
Note that we have used dEB = −dEA , dVB = −dVA , and dNB = −dNA . Now we know from the Second Law that
spontaneous processes result in T dS > 0, which means that S tends to a maximum. If S is a maximum, it must
2.10. EQUILIBRIUM AND STABILITY 59
Figure 2.20: To check for an instability, we compare the energy of a system to its total energy when we reapportion
its energy, volume, and particle number slightly unequally.
be that the coefficients of dEA , dVA , and dNA all vanish, else we could increase the total entropy of the system by
a judicious choice of these three differentials. From T dS = dE + p dV − µ, dN , we have
1 ∂S p ∂S µ ∂S
= , = , =− . (2.252)
T ∂E V,N T ∂V E,N T ∂N E,V
Thus, we conclude that in order for the system to be in equilibrium, so that S is maximized and can increase no
further under spontaneous processes, we must have
Now consider a uniform system with energy E ′ = 2E, volume V ′ = 2V , and particle number N ′ = 2N . We wish
to check that this system is not unstable with respect to spontaneously becoming inhomogeneous. To that end,
we imagine dividing the system in half. Each half would have energy E, volume V , and particle number N . But
suppose we divided up these quantities differently, so that the left half had slightly different energy, volume, and
particle number than the right, as depicted in Fig. 2.20. Does the entropy increase or decrease? We have
∆S = S(E + ∆E, V + ∆V, N + ∆N ) + S(E − ∆E, V − ∆V, N − ∆N ) − S(2E, 2V, 2N )
∂ 2S ∂ 2S ∂ 2S
= 2
(∆E)2 + 2
(∆V )2 + (∆N )2 (2.256)
∂E ∂V ∂N 2
∂ 2S ∂ 2S ∂ 2S
+2 ∆E ∆V + 2 ∆E ∆N + 2 ∆V ∆N .
∂E ∂V ∂E ∂N ∂V ∂N
Thus, we can write X
∆S = Qij Ψi Ψj , (2.257)
i,j
where
∂ 2S ∂ 2S ∂ 2S
∂E 2 ∂E ∂V ∂E ∂N
∂ 2S 2
∂ S ∂ S 2
Q = ∂E ∂V 2
(2.258)
∂V ∂V ∂N
∂ 2S ∂ 2S ∂ 2S
∂E ∂N ∂V ∂N ∂N 2
is the matrix of second derivatives, known in mathematical parlance as the Hessian, and Ψ = (∆E, ∆V, ∆N ). Note
that Q is a symmetric matrix.
60 CHAPTER 2. THERMODYNAMICS
Since S must be a maximum in order for the system to be in equilibrium, we are tempted to conclude that the
homogeneous system is stable if and only if all three eigenvalues of Q are negative. If one or more of the eigen-
values is positive, then it is possible to choose a set of variations Ψ such that ∆S > 0, which would contradict
the assumption that the homogeneous state is one of maximum entropy. A matrix with this restriction is said
to be negative definite. While it is true that Q can have no positive eigenvalues, it is clear from homogeneity of
S(E, V, N ) that one of the three eigenvalues must be zero, corresponding to the eigenvector Ψ = (E, V, N ). Ho-
mogeneity means S(λE, λV, λN ) = λS(E, V, N ). Now let us take λ = 1 + η, where η is infinitesimal. Then
∆E = ηE, ∆V = ηV , and ∆N = ηN , and homogeneity says S(E ± ∆E, V ± ∆V, N ± ∆N ) = (1 ± η) S(E, V, N )
and ∆S = (1+η)S +(1−η)S −2S = 0. We then have a slightly weaker characterization of Q as negative semidefinite.
However, if we fix one of the components of (∆E, ∆V, ∆N ) to be zero, then Ψ must have some component orthog-
onal to the zero eigenvector, in which case ∆S > 0. Suppose we set ∆N = 0 and we just examine the stability with
respect to inhomogeneities in energy and volume. We then restrict our attention to the upper left 2 × 2 submatrix
of Q. A general symmetric 2 × 2 matrix may be written
a b
Q= (2.259)
b c
It is easy to solve for the eigenvalues of Q. One finds
s 2
a+c a−c
λ± = ± + b2 . (2.260)
2 2
In order for Q to be negative definite, we require λ+ < 0 and λ− < 0. Clearly we must have a + c < 0, or else
λ+ > 0 for sure. If a + c < 0 then clearly λ− < 0, but there still is a possibility that λ+ > 0, if the radical is larger
than − 21 (a + c). Demanding that λ+ < 0 therefore yields two conditions:
The condition for equilibrium is that ∆G > 0 for all (∆S, ∆V ). The linear terms vanish by the definition since
T = T0 and p = p0 . Stability then requires that the Hessian matrix Q be positive definite, with
∂ 2E ∂ 2E
∂S 2 ∂S ∂V
Q= . (2.266)
∂ 2E ∂ 2E
∂S ∂V ∂V 2
A discussion of various useful mathematical relations among partial derivatives may be found in the appendix in
§2.17. Some facility with the differential multivariable calculus is extremely useful in the analysis of thermody-
namics problems.
Consider once again the adiabatic free expansion of a gas from initial volume Vi to final volume Vf = rVi . Since
the system is not in equilibrium during the free expansion process, the initial and final states do not lie along an
adiabat, i.e. they do not have the same entropy. Rather, as we found, from Q = W = 0, we have that Ei = Ef ,
which means they have the same energy, and, in the case of an ideal gas, the same temperature (assuming N is
constant). Thus, the initial and final states lie along an isotherm. The situation is depicted in Fig. 2.21. Now let us
compute the change in entropy ∆S = Sf − Si by integrating along this isotherm. Note that the actual dynamics
are irreversible and do not quasistatically follow any continuous thermodynamic path. However, we can use what
is a fictitious thermodynamic path as a means of comparing S in the initial and final states.
We have
ZVf
∂S
∆S = Sf − Si = dV . (2.270)
∂V T,N
Vi
hence
ZVf
∂p
∆S = dV . (2.272)
∂T V,N
Vi
62 CHAPTER 2. THERMODYNAMICS
Figure 2.21: Adiabatic free expansion via a thermal path. The initial and final states do not lie along an adabat!
Rather, for an ideal gas, the initial and final states lie along an isotherm.
as we found before, in eqn. 2.156 What is different about this derivation? Previously, we derived the entropy
change from the explicit formula for S(E, V, N ). Here, we did not need to know this function. The Maxwell
relation allowed us to compute the entropy change using only the equation of state.
We saw how E(T, V, N ) = 21 f N kB T for an ideal gas, independent of the volume. In general we should have
V
E(T, V, N ) = N φ T, N . (2.275)
V
For the ideal gas, φ T, N = 12 f kB T is a function of T alone and is independent on the other intensive quantity
V /N . How does energy vary with volume? At fixed temperature and particle number, we have, from E = F +T S,
∂E ∂F ∂S ∂p
= +T = −p + T , (2.276)
∂V T,N ∂V T,N ∂V T,N ∂T V,N
∂S
∂p
2
where we have used the Maxwell relation ∂V T .N
= ∂T V,N
, derived from the mixed second derivative ∂T∂ ∂V
F
.
Another way to derive this result is as follows. Write dE = T dS − p dV + µ dN and then express dS in terms of
dT , dV , and dN , resulting in
" # " #
∂S ∂S ∂µ
dE = T dT + T − p dV − T + µ dN . (2.277)
∂T V,N ∂V T,N ∂T V,N
2.11. APPLICATIONS OF THERMODYNAMICS 63
∂E
Now read off ∂V V,N
and use the same Maxwell relation as before to recover eqn. 2.276. Applying this result to
the ideal gas law pV = N kB T results in the vanishing of the RHS, hence for any substance obeying the ideal gas
law we must have
E(T, V, N ) = ν ε(T ) = N ε(T )/NA . (2.278)
It is clear that the same conclusion follows for any equation of state of the form p(T, V, N ) = T · f (V /N ), where
f (V /N ) is an arbitrary function of its argument: the ideal gas law remains valid10 . This is not true, however, for
the van der Waals equation of state,
a
p + 2 v − b) = RT , (2.279)
v
where v = NA V /N is the molar volume. We then find (always assuming constant N ),
∂E ∂ε ∂p a
= =T −p= , (2.280)
∂V T ∂v T ∂T V v2
What about cp ? This requires a bit of work. We start with eqn. 2.34,
∂ε ∂v
cp = +p
∂T p ∂T p
(2.283)
′ a ∂v
= ω (T ) + p + 2
v ∂T p
We can now read off the result for the volume expansion coefficient,
1 ∂v 1 R
αp = = · a 2ab
. (2.285)
v ∂T p v p − v2 + v3
10 Note V /N = v/NA .
64 CHAPTER 2. THERMODYNAMICS
cVDW
V = 12 f R (2.287)
2 3
R Tv
cVDW
p = 12 f R + . (2.288)
RT v 3 − 2a(v − b)2
Appealing to a Maxwell relation derived from F (T, V, N ), and then appealing to eqn. 2.492, we have
∂S ∂p ∂p ∂V
= =− . (2.291)
∂V T ∂T V ∂V T ∂T p
Thus,
T α2p
Cp − CV = V , (2.296)
κT
or, in terms of intensive quantities,
v T α2p
cp − cV = , (2.297)
κT
where, as always, v = V NA /N is the molar volume.
This above relation generalizes to any conjugate force-displacement pair (−p, V ) → (y, X):
∂y ∂X
Cy − CX = −T
∂T X ∂T y
2 (2.298)
∂y ∂X
=T .
∂X T ∂T y
we obtain
v T α2p
κT − κS = . (2.303)
cp
Comparing eqns. 2.297 and 2.303, we find
Remark: The previous discussion has assumed an isotropic magnetic system where M and H are collinear, hence
H ·M = HM .
∂M α 1 ∂ 2G
χαβ
T = = − (2.311)
∂H β T ν ∂H α ∂H β
∂M α 1 ∂2H
χαβ
S = = − . (2.312)
∂H β S ν ∂H α ∂H β
Previously we considered the adiabatic free expansion of an ideal gas. We found that Q = W = 0 hence ∆E = 0,
which means the process is isothermal, since E = νε(T ) is volume-independent. The entropy changes, however,
since S(E, V, N ) = N kB ln(V /N ) + 21 f N kB ln(E/N ) + N s0 . Thus,
V
Sf = Si + N kB ln f . (2.315)
Vi
What happens if the gas is nonideal?
We integrate along a fictitious thermodynamic path connecting initial and final states, where dE = 0 along the
path. We have
∂E ∂E
0 = dE = dV + dT (2.316)
∂V T ∂T V
hence
∂T (∂E/∂V )T 1 ∂E
=− =− . (2.317)
∂V E (∂E/∂T )V CV ∂V T
We also have
∂E ∂S ∂p
=T −p=T −p. (2.318)
∂V T ∂V T ∂T V
Thus, "
#
∂T 1 ∂p
= p−T . (2.319)
∂V E CV ∂T V
2.11. APPLICATIONS OF THERMODYNAMICS 67
L2 ·bar L
gas a mol2
b mol pc (bar) Tc (K) vc (L/mol)
Acetone 14.09 0.0994 52.82 505.1 0.2982
Argon 1.363 0.03219 48.72 150.9 0.0966
Carbon dioxide 3.640 0.04267 7404 304.0 0.1280
Ethanol 12.18 0.08407 63.83 516.3 0.2522
Freon 10.78 0.0998 40.09 384.9 0.2994
Helium 0.03457 0.0237 2.279 5.198 0.0711
Hydrogen 0.2476 0.02661 12.95 33.16 0.0798
Mercury 8.200 0.01696 1055 1723 0.0509
Methane 2.283 0.04278 46.20 190.2 0.1283
Nitrogen 1.408 0.03913 34.06 128.2 0.1174
Oxygen 1.378 0.03183 50.37 154.3 0.0955
Water 5.536 0.03049 220.6 647.0 0.0915
Table 2.3: Van der Waals parameters for some common gases. (Source: Wikipedia.)
Note that the term in square brackets vanishes for any system obeying the ideal gas law. For a nonideal gas,
ZVf
∂T
∆T = dV , (2.320)
∂V E
Vi
a
p + 2 (v − b) = RT .
v
We then have
∂p a aν 2
p−T =− = − . (2.321)
∂T V v2 V2
In §2.11.3 we concluded that CV = 12 f νR for the van der Waals gas, hence
ZVf
2aν dV 2a 1 1
∆T = − = − . (2.322)
fR V2 f R vf vi
Vi
Thus, if Vf > Vi , we have Tf < Ti and the gas cools upon expansion.
Consider O2 gas with an initial specific volume of vi = 22.4 L/mol, which is the STP value for an ideal gas, freely
expanding to a volume vf = ∞ for maximum cooling. According to table 2.3, a = 1.378 L2 · bar/mol2 , and we
have ∆T = −2a/f Rvi = −0.296 K, which is a pitifully small amount of cooling. Adiabatic free expansion is a
very inefficient way to cool a gas.
68 CHAPTER 2. THERMODYNAMICS
Figure 2.22: In a throttle, a gas is pushed through a porous plug separating regions of different pressure. The
change in energy is the work done, hence enthalpy is conserved during the throttling process.
In a throttle, depicted in Fig. 2.22, a gas is forced through a porous plug which separates regions of different
pressures. According to the figure, the work done on a given element of gas is
ZVf ZVi
W = dV pf − dV pi = pf Vf − pi Vi . (2.323)
0 0
Now we assume that the system is thermally isolated so that the gas exchanges no heat with its environment, nor
with the plug. Then Q = 0 so ∆E = −W , and
Ei + pi Vi = Ef + pf Vf (2.324)
Hi = Hf , (2.325)
where H is enthalpy. Thus, the throttling process is isenthalpic. We can therefore study it by defining a fictitious
thermodynamic path along which dH = 0. The, choosing T and p as state variables,
∂H ∂H
0 = dH = dp + dT (2.326)
∂T p ∂p T
hence
∂T (∂H/∂p)T
=− . (2.327)
∂p H (∂H/∂T )p
The numerator on the RHS is computed by writing dH = T dS + V dp and then dividing by dp, to obtain
∂H ∂S ∂V
=V +T =V −T . (2.328)
∂p T ∂p T ∂T p
The denominator is
∂H ∂H ∂S
=
∂T p ∂S p ∂T p
(2.329)
∂S
=T = Cp .
∂T p
Thus,
" #
∂T 1 ∂v
= T −v
∂p H cp ∂T p
(2.330)
v
= T αp − 1 ,
cp
2.11. APPLICATIONS OF THERMODYNAMICS 69
Figure 2.23: Inversion temperature T ∗ (p) for the van der Waals gas. Pressure and temperature are given in terms
of pc = a/27b2 and Tc = 8a/27bR, respectively.
1 ∂V
where αp = V ∂T p is the volume expansion coefficient.
From the van der Waals equation of state, we obtain, from eqn. 2.285,
T ∂v RT /v v−b
T αp = = a 2ab
= 2 . (2.331)
v ∂T p p − v2 + v3 v − RT v−b
2a
v
a
Assuming v ≫ , b, we have
RT
∂T 1 2a
= −b . (2.332)
∂p H cp RT
2a ∂T
Thus, for T > T ∗ = Rb , we have ∂p H < 0 and the gas heats up upon an isenthalpic pressure decrease. For
∗
T < T , the gas cools under such conditions.
∗
In fact, there are two inversion temperatures T1,2 for the van der Waals gas. To see this, we set T αp = 1, which is
the criterion for inversion. From eqn. 2.331 it is easy to derive
r
b bRT
=1− . (2.333)
v 2a
We insert this into the van der Waals equation of state to derive a relationship T = T ∗ (p) at which T αp = 1 holds.
After a little work, we find r
3RT 8aRT a
p=− + − 2 . (2.334)
2b b3 b
This is a quadratic equation for T , the solution of which is
r !2
2a 3b 2p
T ∗ (p) = 2± 1− . (2.335)
9 bR a
70 CHAPTER 2. THERMODYNAMICS
∂T
In Fig. 2.23 we plot pressure versus temperature in scaled units, showing the curve along which ∂p H = 0. The
volume, pressure, and temperature scales defined are
a 8a
vc = 3b , pc = , Tc = . (2.336)
27 b2 27 bR
Values for pc , Tc , and vc are provided in table 2.3. If we define v = v/vc , p = p/pc , and T = T /Tc , then the van der
Waals equation of state may be written in dimensionless form:
3
p + 2 3v − 1) = 8T . (2.337)
v
∂T
In terms of the scaled parameters, the equation for the inversion curve ∂p = 0 becomes
H
q 2 q 2
p = 9 − 36 1 − 13 T ⇐⇒ T =3 1± 1− 1
9 p . (2.338)
Thus, there is no inversion for p > 9 pc . We are usually interested in the upper inversion temperature, T2∗ ,
corresponding to the upper sign in eqn. 2.335. The maximum inversion temperature occurs for p = 0, where
2a
∗
Tmax = bR = 27 ∗
4 Tc . For H2 , from the data in table 2.3, we find Tmax (H2 ) = 224 K, which is within 10% of the
experimentally measured value of 205 K.
What happens when H2 gas leaks from a container with T > T2∗ ? Since ∂T ∂p < 0 and ∆p < 0, we have ∆T > 0.
H
The gas warms up, and the heat facilitates the reaction 2 H2 + O2 −→ 2 H2 O, which releases energy, and we have
a nice explosion.
A typical phase diagram of a p-V -T system is shown in the Fig. 2.24(a). The solid lines delineate boundaries
between distinct thermodynamic phases. These lines are called coexistence curves . Along these curves, we can
have coexistence of two phases, and the thermodynamic potentials are singular. The order of the singularity is
often taken as a classification of the phase transition. I.e. if the thermodynamic potentials E, F , G, and H have
discontinuous or divergent mth derivatives, the transition between the respective phases is said to be mth order.
Modern theories of phase transitions generally only recognize two possibilities: first order transitions, where the
order parameter changes discontinuously through the transition, and second order transitions, where the order param-
eter vanishes continuously at the boundary from ordered to disordered phases11 . We’ll discuss order parameters
during Physics 140B.
For a more interesting phase diagram, see Fig. 2.24(b,c), which displays the phase diagrams for 3 He and 4 He. The
only difference between these two atoms is that the former has one fewer neutron: (2p + 1n + 2e) in 3 He versus (2p
+ 2n + 2e) in 4 He. As we shall learn when we study quantum statistics, this extra neutron makes all the difference,
because 3 He is a fermion while 4 He is a boson.
Figure 2.24: (a) Typical thermodynamic phase diagram of a single component p-V -T system, showing triple point
(three phase coexistence) and critical point. (Source: Univ. of Helsinki.) Also shown: phase diagrams for 3 He (b)
and 4 He (c). What a difference a neutron makes! (Source: Brittanica.)
This may in principle be inverted to yield p = p(v, T ) or v = v(T, p) or T = T (p, v). The single constraint f (p, v, T )
on the three state variables defines a surface in {p, v, T } space. An example of such a surface is shown in Fig. 2.25,
for the ideal gas.
Real p-v-T surfaces are much richer than that for the ideal gas, because real systems undergo phase transitions in
which thermodynamic properties are singular or discontinuous along certain curves on the p-v-T surface. An
example is shown in Fig. 2.26. The high temperature isotherms resemble those of the ideal gas, but as one cools
below the critical temperature Tc , the isotherms become singular. Precisely at T = Tc , the isotherm p = p(v, Tc )
becomes perfectly horizontal at v = vc , which is the critical molar volume. This means that the isothermal com-
pressibility, κT = − v1 ∂v
∂p T diverges at T = Tc . Below Tc , the isotherms have a flat portion, as shown in Fig. 2.28,
corresponding to a two-phase region where liquid and vapor coexist. In the (p, T ) plane, sketched for H2 O in Fig. 2.4
and shown for CO2 in Fig. 2.29, this liquid-vapor phase coexistence occurs along a curve, called the vaporization
(or boiling) curve. The density changes discontinuously across this curve; for H2 O, the liquid is approximately
1000 times denser than the vapor at atmospheric pressure. The density discontinuity vanishes at the critical point.
Note that one can continuously transform between liquid and vapor phases, without encountering any phase
transitions, by going around the critical point and avoiding the two-phase region.
In addition to liquid-vapor coexistence, solid-liquid and solid-vapor coexistence also occur, as shown in Fig. 2.26.
The triple point (Tt , pt ) lies at the confluence of these three coexistence regions. For H2 O, the location of the triple
point and critical point are given by
Tt = 273.16 K Tc = 647 K
−3
pt = 611.7 Pa = 6.037 × 10 atm pc = 22.06 MPa = 217.7 atm
Recall that the homogeneity of E(S, V, N ) guaranteed E = T S −pV +µN , from Euler’s theorem. It also guarantees
a relation between the intensive variables T , p, and µ, according to eqn. 2.148. Let us define g ≡ G/ν = NA µ, the
Gibbs free energy per mole. Then
dg = −s dT + v dp , (2.340)
72 CHAPTER 2. THERMODYNAMICS
Figure 2.25: The surface p(v, T ) = RT /v corresponding to the ideal gas equation of state, and its projections onto
the (p, T ), (p, v), and (T, v) planes.
where s = S/ν and v = V /ν are the molar entropy and molar volume, respectively. Along a coexistence curve
between phase #1 and phase #2, we must have g1 = g2 , since the phases are free to exchange energy and particle
number, i.e. they are in thermal and chemical equilibrium. This means
dg1 = −s1 dT + v1 dp = −s2 dT + v2 dp = dg2 . (2.341)
Therefore, along the coexistence curve we must have
dp s − s1 ℓ
= 2 = , (2.342)
dT coex v2 − v1 T ∆v
where
ℓ ≡ T ∆s = T (s2 − s1 ) (2.343)
is the molar latent heat of transition. A heat ℓ must be supplied in order to change from phase #1 to phase #2, even
without changing p or T . If ℓ is the latent heat per mole, then we write ℓ̃ as the latent heat per gram: ℓ̃ = ℓ/M ,
where M is the molar mass.
Along the liquid-gas coexistence curve, we typically have vgas ≫ vliquid , and assuming the vapor is ideal, we may
write ∆v ≈ vgas ≈ RT /p. Thus,
dp ℓ pℓ
= ≈ . (2.344)
dT liq−gas T ∆v RT 2
If ℓ remains constant throughout a section of the liquid-gas coexistence curve, we may integrate the above equation
to get
dp ℓ dT
= =⇒ p(T ) = p(T0 ) eℓ/RT0 e−ℓ/RT . (2.345)
p R T2
2.12. PHASE TRANSITIONS AND PHASE EQUILIBRIA 73
Figure 2.26: A p-v-T surface for a substance which contracts upon freezing. The red dot is the critical point and the
red dashed line is the critical isotherm. The yellow dot is the triple point at which there is three phase coexistence of
solid, liquid, and vapor.
Life on planet earth owes much of its existence to a peculiar property of water: the solid is less dense than the
liquid along the coexistence curve. For example at T = 273.1 K and p = 1 atm,
The latent heat of the transition is ℓ̃ = 333 J/g = 79.5 cal/g. Thus,
dp ℓ̃ 333 J/g
= =
dT liq−sol T ∆ṽ (273.1 K) (−9.05 × 10−2 cm3 /g)
(2.347)
dyn
8 atm
= −1.35 × 10 2
= −134 ◦ .
cm K C
The negative slope of the melting curve is invoked to explain the movement of glaciers: as glaciers slide down
a rocky slope, they generate enormous pressure at obstacles12 Due to this pressure, the story goes, the melting
temperature decreases, and the glacier melts around the obstacle, so it can flow past it, after which it refreezes.
But it is not the case that the bottom of the glacier melts under the pressure, for consider a glacier of height
h = 1 km. The pressure at the bottom is p ∼ gh/ṽ ∼ 107 Pa, which is only about 100 atmospheres. Such a pressure
can produce only a small shift in the melting temperature of about ∆Tmelt = −0.75◦ C.
Does the Clausius-Clapeyron relation explain how we can skate on ice? My seven year old daughter has a mass
of about M = 20 kg. Her ice skates have blades of width about 5 mm and length about 10 cm. Thus, even on one
12 Themelting curve has a negative slope at relatively low pressures, where the solid has the so-called Ih hexagonal crystal structure. At
pressures above about 2500 atmospheres, the crystal structure changes, and the slope of the melting curve becomes positive.
74 CHAPTER 2. THERMODYNAMICS
Figure 2.27: Equation of state for a substance which expands upon freezing, projected to the (v, T ) and (v, p) and
(T, p) planes.
Mg 20 kg × 9.8 m/s2
∆p = ≈ = 3.9 × 105 Pa = 3.9 atm . (2.348)
A (5 × 10−3 m) × (10−1 m)
Figure 2.28: Projection of the p-v-T surface of Fig. 2.26 onto the (v, p) plane.
movie, directed by Sidney Lumet). My point here is that Hans’ crappy wooden skates can’t compare to the metal
ones, even though the surface melt between the ice and the air is the same. The skate blade material also makes a
difference, both for the interface energy and, perhaps more importantly, for the generation of friction as well.
Suppose we have an ice cube initially at temperature T0 < Θ ≡ 273.15 K (i.e. Θ = 0◦ C) and we toss it into a pond
of water. We regard the pond as a heat bath at some temperature T1 > Θ. Let the mass of the ice be M . How much
heat Q is absorbed by the ice in order to raise its temperature to T1 ? Clearly
where c̃S and c̃L are the specific heats of ice (solid) and water (liquid), respectively15 , and ℓ̃ is the latent heat
of melting per unit mass. The pond must give up this much heat to the ice, hence the entropy of the pond,
discounting the new water which will come from the melted ice, must decrease:
Q
∆Spond = − . (2.350)
T1
Now we ask what is the entropy change of the H2 O in the ice. We have
Z ZΘ ZT1
d̄Q M c̃S M ℓ̃ M c̃L
∆Sice = = dT + + dT
T T Θ T
T0 Θ (2.351)
Θ M ℓ̃ T
= M c̃S ln + + M c̃L ln 1 .
T0 Θ Θ
15 We assume c̃S (T ) and c̃L (T ) have no appreciable temperature dependence, and we regard them both as constants.
76 CHAPTER 2. THERMODYNAMICS
Figure 2.29: Phase diagram for CO2 in the (p, T ) plane. (Source: www.scifun.org.)
Equilibrium between two phases means that p, T , and µ(p, T ) are identical. From
µ1 (p, T ) = µ2 (p, T ) , (2.355)
we derive an equation for the slope of the coexistence curve, the Clausius-Clapeyron relation. Note that we have
one equation in two unknowns (T, p), so the solution set is a curve. For three phase coexistence, we have
µ1 (p, T ) = µ2 (p, T ) = µ3 (p, T ) , (2.356)
which gives us two equations in two unknowns. The solution is then a point (or a set of points). A critical point
also is a solution of two simultaneous equations:
critical point =⇒ v1 (p, T ) = v2 (p, T ) , µ1 (p, T ) = µ2 (p, T ) . (2.357)
2.12. PHASE TRANSITIONS AND PHASE EQUILIBRIA 77
Figure 2.30: Left panel: data from R. R. Gilpin, J. Colloid Interface Sci. 77, 435 (1980) showing measured thickness
of the surface melt on ice at temperatures below 0◦ C. The straight line has slope − 31 , as predicted by theory. Right
panel: phase diagram of H2 O, showing various high pressure solid phases. (Source : Physics Today, December
2005).
∂µ
Recall v = NA ∂p T . Note that there can be no four phase coexistence for a simple p-V -T system.
Now for the general result. Suppose we have σ species, with particle numbers Na , where a = 1, . . . , σ. It is
useful to briefly recapitulate the derivation of the Gibbs-Duhem relation. The energy E(S, V, N1 , . . . , Nσ ) is a
homogeneous function of degree one:
E(λS, λV, λN1 , . . . , λNσ ) = λE(S, V, N1 , . . . , Nσ ) . (2.358)
From Euler’s theorem for homogeneous functions (just differentiate with respect to λ and then set λ = 1), we have
σ
X
E = TS − pV + µa Na . (2.359)
a=1
(j)
If there are ϕ different phases, then in each phase j, with j = 1, . . . , ϕ, there is a chemical potential µa for each
species a. We then have
(j) (j)
µσ(j) = µ(j)
σ T, p, µ1 , . . . , µσ−1 . (2.363)
78 CHAPTER 2. THERMODYNAMICS
(j)
Here µa is the chemical potential of the ath species in the j th phase. Thus, there are ϕ such equations relating the
(j)
2 + ϕ σ variables T, p, µa , meaning that only 2 + ϕ (σ − 1) of them may be chosen as independent. This, then,
is the dimension of ‘thermodynamic space’ containing a maximal number of intensive variables:
dTD (σ, ϕ) = 2 + ϕ (σ − 1) . (2.364)
To completely specify the state of our system, we of course
P introduce a single extensive variable, such as the total
volume V . Note that the total particle number N = σa=1 Na may not be conserved in the presence of chemical
reactions!
Now suppose we have equilibrium among ϕ phases. We have implicitly assumed thermal and mechanical equi-
librium among all the phases, meaning that p and T are constant. Chemical equilibrium applies on a species-by-
species basis. This means
(j ′ )
µ(j)
a = µa (2.365)
where j, j ′ ∈ {1, . . . , ϕ}. This gives σ(ϕ − 1) independent equations equations16 . Thus, we can have phase equilib-
rium among the ϕ phases of σ species over a region of dimension
dPE (σ, ϕ) = 2 + ϕ (σ − 1) − σ (ϕ − 1)
(2.366)
=2+σ−ϕ.
Since dPE ≥ 0, we must have ϕ ≤ σ + 2. Thus, with two species (σ = 2), we could have at most four phase
coexistence.
If the various species can undergo ρ distinct chemical reactions of the form
(r) (r)
ζ1 A1 + ζ2 A2 + · · · + ζσ(r) Aσ = 0 , (2.367)
(r)
where Aa is the chemical formula for species a, and ζa is the stoichiometric coefficient for the ath species in the
rth reaction, with r = 1, . . . , ρ, then we have an additional ρ constraints of the form
σ
X
ζa(r) µ(j)
a = 0. (2.368)
a=1
Therefore,
dPE (σ, ϕ, ρ) = 2 + σ − ϕ − ρ . (2.369)
One might ask what value of j are we to use in eqn. 2.368, or do we in fact have ϕ such equations for each r? The
answer is that eqn. 2.365 guarantees that the chemical potential of species a is the same in all the phases, hence it
doesn’t matter what value one chooses for j in eqn. 2.368.
P
Let us assume that no reactions take place, i.e. ρ = 0, so the total number of particles σb=1 Nb is conserved.
(j) (j)
Instead of choosing (T, p, µ1 , . . . , µσ−1 ) as dTD intensive variables, we could have chosen (T, p, µ1 , . . . , xσ−1 ), where
xa = Na /N is the concentration of species a.
Why do phase diagrams in the (p, v) and (T, v) plane look different than those in the (p, T ) plane?17 For example,
Fig. 2.27 shows projections of the p-v-T surface of a typical single component substance onto the (T, v), (p, v), and
(p, T ) planes. Coexistence takes place along curves in the (p, T ) plane, but in extended two-dimensional regions
in the (T, v) and (p, v) planes. The reason that p and T are special is that temperature, pressure, and chemical
potential must be equal throughout an equilibrium phase if it is truly in thermal, mechanical, and chemical equi-
librium. This is not the case for an intensive variable such as specific volume v = NA V /N or chemical concentration
xa = Na /N .
j = 1 and let j ′ range over the ϕ − 1 values 2, . . . , ϕ.
16 Set
17 Thesame can be said for multicomponent systems: the phase diagram in the (T, x) plane at constant p looks different than the phase
diagram in the (T, µ) plane at constant p.
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 79
Entropy is widely understood as a measure of disorder. Of course, such a definition should be supplemented by
a more precise definition of disorder – after all, one man’s trash is another man’s treasure.
P To gain some intuition
about entropy, let us explore the mixing of a multicomponent ideal gas. Let N = aP Na be the total number of
particles of all species, and let xa = Na /N be the concentration of species a. Note that a xa = 1.
For any substance obeying the ideal gas law pV = N kB T , the entropy is
Now let us imagine removing all the barriers separating the different gases and letting the particles mix thor-
oughly. The result is that each component gas occupies the full volume V , so the entropy is
X Xh i
Smixed = Sa = Na kB ln(V /Na ) + Na φa (T ) . (2.372)
a a
Na Va
where xa = N = V is the fraction of species a. Note that ∆Smix = 0.
What if all the components were initially identical? It seems absurd that the entropy should increase simply by
removing some invisible barriers. This is again the Gibbs paradox. In this case, the resolution of the paradox is to
note that the sum in the expression for Smixed is a sum over distinct species. Hence if the particles are all identical,
we have Smixed = N kB ln(V /N ) + N φ(T ) = Sunmixed, hence ∆Smix = 0.
80 CHAPTER 2. THERMODYNAMICS
Figure 2.31: A multicomponent system consisting of isolated gases, each at temperature T and pressure p. Then
system entropy increases when all the walls between the different subsystems are removed.
As we shall learn when we study statistical mechanics, the entropy may be interpreted in terms of the number of
ways W (E, V, N ) a system at fixed energy and volume can arrange itself. One has
Consider a system consisting of σ different species of particles. Now let it be that for each species label a, Na
particles of that species are confined among Qa little boxes such that at most one particle can fit in a box (see Fig.
2.32). How many ways W are there to configure N identical particles among Q boxes? Clearly
Q Q!
W = = . (2.375)
N N ! (Q − N )!
Q!
Were the particles distinct, we’d have Wdistinct = (Q−N )! , which is N ! times greater. This is because permuting
distinct particles results in a different configuration, and there are N ! ways to permute N particles.
The entropy for species a is then Sa = kB ln Wa = kB ln Q N
a
. We then use Stirling’s approximation,
a
ln(K!) = K ln K − K + 1
2 ln K + 1
2 ln(2π) + O(K −1 ) , (2.376)
which is an asymptotic expansion valid for K ≫ 1. One then finds for Q, N ≫ 1, with x = N/Q ∈ [0, 1],
Q
ln = Q ln Q − Q − xQ ln(xQ) − xQ − (1 − x)Q ln (1 − x)Q − (1 − x)Q
N
h i
= −Q x ln x + (1 − x) ln(1 − x) . (2.377)
This is valid up to terms of order Q in Stirling’s expansion. Since ln Q ≪ Q, the next term is small and we are safe
to stop here. Summing up the contributions from all the species, we get
σ
X σ
X h i
Sunmixed = kB ln Wa = −kB Qa xa ln xa + (1 − xa ) ln(1 − xa ) , (2.378)
a=1 a=1
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 81
Figure 2.32: Mixing among three different species of particles. The mixed configuration has an additional entropy,
∆Smix .
where xea = Na /Q is the fraction of all boxes containing a particle of species a, and N0 is the number of empty
boxes. Note that
N N Q
xea = a = a · a = xa fa , (2.381)
Q Qa Q
P
where fa ≡ Qa /Q. Note that σa=1 fa = 1.
Qa Na
Let’s assume all the densities are initially the same, so xa = x∀a, so x
ea = x fa . In this case, fa = Q = N is the
fraction of species a among all the particles. We then have xe0 = 1 − x, and
σ
X
Smixed = −kB Q xfa ln(xfa ) − kB Q x
e0 ln x
e0
a=1
h i σ (2.382)
X
= −kB Q x ln x + (1 − x) ln(1 − x) − kB x Q fa ln fa .
a=1
Suppose one of the species is much more plentiful than all the others, and label it with a = 0. We will call this the
solvent. The entropy of mixing is then
" X σ #
N0 Na
∆Smix = −kB N0 ln + Na ln , (2.384)
N0 + N ′ a=1
N0 + N ′
Pσ
where N ′ = a=1 Na is the total number of solvent molecules, summed over all species. We assume the solution
is weak, which means Na ≤ N ′ ≪ N0 . Expanding in powers of N ′ /N0 and Na /N0 , we find
σ
" #
X Na 2
∆Smix = −kB Na ln − Na + O N ′ /N0 . (2.385)
a=1
N0
Consider now a solution consisting of N0 molecules of a solvent and Na molecules of species a of solute, where
a = 1, . . . , σ. We begin by expanding the Gibbs free energy G(T, p, N0 , N1 , . . . , Nσ ), where there are σ species of
solutes, as a power series in the small quantities Na . We have
X
Na
G T, p, N0 , {Na } = N0 g0 (T, p) + kB T Na ln
a
eN0
X (2.386)
1 X
+ Na ψa (T, p) + Aab (T, p) Na Nb .
a
2N0
a,b
The first term on the RHS corresponds to the Gibbs free energy of the solvent. The second term is due to the
entropy of mixing. The third term is the contribution to the total free energy from the individual species. Note
the factor of e in the denominator inside the logarithm, which accounts for the second term in the brackets on the
RHS of eqn. 2.385. The last term is due to interactions between the species; it is truncated at second order in the
solute numbers.
The chemical potential for the solvent is
∂G X X
1
µ0 (T, p) = = g0 (T, p) − kB T xa − 2 Aab (T, p) xa xb , (2.387)
∂N0 a a,b
where xa = Na /N0 is the concentrations of solute species a. PBy assumption, the last term on the RHS of each of
σ
these equations is small, since Nsolute ≪ N0 , where Nsolute = a=1 Na is the total number of solute molecules. To
lowest order, then, we have
Figure 2.33: Osmotic pressure causes the column on the right side of the U-tube to rise higher than the column on
the left by an amount ∆h = π/̺ g.
pure solvent. There is energy exchange through the membrane, so the temperature is T throughout. There is no
volume exchange, however: dV = dV ′ = 0, hence the pressure need not be the same. Since the membrane is
permeable to the solvent, we have that the chemical potential µ0 is the same on each side. This means
g0 (T, pR ) − xkB T = g0 (T, pL ) , (2.391)
where pL,R is the pressure on the left and right sides of the membrane, and x = N/N0 is again the total solute
concentration. This equation once again tells us that the pressure p cannot be the same on both sides of the
membrane. If the pressure difference is small, we can expand in powers of the osmotic pressure, π ≡ pR − pL , and
we find
∂µ0
π = x kB T . (2.392)
∂p T
But a Maxwell relation (§2.9) guarantees
∂µ ∂V
= = v(T, p)/NA , (2.393)
∂p T,N ∂N T,p
Along the coexistence curve separating liquid and vapor phases, the chemical potentials of the two phases are
identical:
µ0L (T, p) = µ0V (T, p) . (2.395)
84 CHAPTER 2. THERMODYNAMICS
Here we write µ0 for µ to emphasize that we are talking about a phase with no impurities present. This equation
provides a single constraint on the two variables T and p, hence one can, in principle, solve to obtain T = T0∗ (p),
which is the equation of the liquid-vapor coexistence curve in the (T, p) plane. Now suppose there is a solute
present in the liquid. We then have
µL (T, p, x) = µ0L (T, p) − xkB T , (2.396)
where x is the dimensionless solute concentration, summed over all species. The condition for liquid-vapor coex-
istence now becomes
µ0L (T, p) − xkB T = µ0V (T, p) . (2.397)
in the boiling temperature at fixed p. Assuming this shift is small, let us expand to lowest
This will lead to a shift
order in T − T0∗ (p) , writing
∂µ0L ∂µ0V
µ0L (T0∗ , p) + T − T0∗ − xkB T = µ0V (T0∗ , p) + T − T0∗ . (2.398)
∂T p ∂T p
Note that
∂µ ∂S
=− (2.399)
∂T p,N ∂N T,p
from a Maxwell relation deriving from exactness of dG. Since S is extensive, we can write S = (N/NA ) s(T, p),
where s(T, p) is the molar entropy. Solving for T , we obtain
2
∗ xR T0∗ (p)
T (p, x) = T0∗ (p) + , (2.400)
ℓv (p)
where ℓv = T0∗ · (sV − sL ) is the latent heat of the liquid-vapor transition19 . The shift ∆T ∗ = T ∗ − T0∗ is called the
boiling point elevation.
As an example, consider seawater, which contains approximately 35 g of dissolved Na+ Cl− per kilogram of H2 O.
The atomic masses of Na and Cl are 23.0 and 35.4, respectively, hence the total ionic concentration in seawater
(neglecting everything but sodium and chlorine) is given by
2 · 35 1000
x= ≈ 0.022 . (2.401)
23.0 + 35.4 18
19 We shall discuss latent heat again in §2.12.2 below.
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 85
Put another way, the boiling point elevation of H2 O at atmospheric pressure is about 0.28◦C per percent solute.
We can express this as ∆T ∗ = Km, where the molality m is the number of moles of solute per kilogram of solvent.
For H2 O, we find K = 0.51◦C kg/mol.
Similar considerations apply at the freezing point, when we equate the chemical potential of the solvent plus
solute to that of the pure solid. The latent heat of fusion for H2 O is about ℓf = Tf0 · (sLIQUID − sSOLID ) = 6.01 kJ/mol20
2
We thus predict a freezing point depression of ∆T ∗ = −xR T0∗ /ℓf = 1.03◦ C · x[%]. This can be expressed once
again as ∆T ∗ = −Km, with K = 1.86◦ C kg/mol21 .
Consider a binary solution, and write the Gibbs free energy G(T, p, NA , NB ) as
NA
G(T, p, NA , NB ) = NA µ0A (T, p) + NB µ0B (T, p) + NA kB T ln
NA + NB
(2.403)
NB NA NB
+ NB kB T ln +λ .
NA + NB NA + NB
The first four terms on the RHS represent the free energy of the individual component fluids and the entropy of
mixing. The last term is an interaction contribution. With λ > 0, the interaction term prefers that the system be
either fully A or fully B. The entropy contribution prefers a mixture, so there is a competition. What is the stable
thermodynamic state?
It is useful to write the Gibbs free energy per particle, g(T, p, x) = G/(NA + NB ), in terms of T , p, and the concen-
tration x ≡ xB = NB /(NA + NB ) of species B (hence xA = 1 − x is the concentration of species A). Then
h i
g(T, p, x) = (1 − x) µ0A + x µ0B + kB T x ln x + (1 − x) ln(1 − x) + λ x (1 − x) . (2.404)
In order for the system to be stable against phase separation into relatively A-rich and B-rich regions, we must have
that g(T, p, x) be a convex function of x. Our first check should be for a local instability, i.e. spinodal decomposition.
We have
∂g x
= µ0B − µ0A + kB T ln + λ (1 − 2x) (2.405)
∂x 1−x
and
∂ 2g k T k T
= B + B − 2λ . (2.406)
∂x2 x 1−x
∂ 2g
The spinodal is given by the solution to the equation ∂x2 = 0, which is
2λ
T ∗ (x) = x (1 − x) . (2.407)
kB
1
Since x (1 − x) achieves its maximum value of 4 at x = 21 , we have T ∗ ≤ kB /2λ.
20 See table 2.4, and recall M = 18 g is the molar mass of H2 O.
21 It is more customary to write ∆T ∗ = Tpure∗ ∗ ∗
solvent − Tsolution in the case of the freezing point depression, in which case ∆T is positive.
86 CHAPTER 2. THERMODYNAMICS
Figure 2.34: Gibbs free energy per particle for a binary solution as a function of concentration x = xB of the B
species (pure A at the left end x = 0 ; pure B at the right end x = 1), in units of the interaction parameter λ. Dark
red curve: T = 0.65 λ/kB > Tc ; green curve: T = λ/2kB = Tc ; blue curve: T = 0.40 λ/kB < Tc . We have chosen
µ0A = 0.60 λ − 0.50 kB T and µ0B = 0.50 λ − 0. 50 kB T . Note that the free energy g(T, p, x) is not convex in x for
T < Tc , indicating an instability and necessitating a Maxwell construction.
In Fig. 2.34 we sketch the free energy g(T, p, x) versus x for three representative temperatures. For T > λ/2kB , the
free energy is everywhere convex in λ. When T < λ/2kB , there free energy resembles the blue curve in Fig. 2.34,
and the system is unstable to phase separation. The two phases are said to be immiscible, or, equivalently, there
exists a solubility gap. To determine the coexistence curve, we perform a Maxwell construction, writing
g(x2 ) − g(x1 ) ∂g ∂g
= = . (2.408)
x2 − x1 ∂x x ∂x x
1 2
Here, x1 and x2 are the boundaries of the two phase region. These equations admit a symmetry of x ↔ 1 − x,
hence we can set x = x1 and x2 = 1 − x. We find
g(1 − x) − g(x) = (1 − 2x) µ0B − µ0A , (2.409)
and invoking eqns. 2.408 and 2.405 we obtain the solution
λ 1 − 2x
Tcoex(x) = · . (2.410)
kB ln 1−x
x
The phase diagram for the binary system is shown in Fig. 2.36. For T < T ∗ (x), the system is unstable, and
spinodal decomposition occurs. For T ∗ (x) < T < Tcoex(x), the system is metastable, just like the van der Waals gas
in its corresponding regime. Real binary solutions behave qualitatively like the model discussed here, although
the coexistence curve is generally not symmetric under x ↔ 1 − x, and the single phase region extends down to
low temperatures for x ≈ 0 and x ≈ 1. If λ itself is temperature-dependent, there can be multiple solutions to
eqns. 2.407 and 2.410. For example, one could take
λ0 T 2
λ(T ) = . (2.411)
T 2 + T02
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 87
Figure 2.35: Upper panels: chemical potential shifts ∆µ± = ∆µA ± ∆µB versus concentration x = xB . The dashed
black line is the spinodal, and the solid black line the coexistence boundary. Temperatures range from T = 0 (dark
blue) to T = 0.6 λ/kB (red) in units of 0.1 λ/kB . Lower panels: phase diagram in the (T, ∆µ± ) planes. The black
dot is the critical point.
In this case, kB T > λ at both high and low temperatures, and we expect the single phase region to be reentrant.
Such a phenomenon occurs in water-nicotine mixtures, for example.
It is instructive to consider the phase diagram in the (T, µ) plane. We define the chemical potential shifts,
∆µA ≡ µA − µ0A = kB T ln(1 − x) + λ x2 (2.412)
∆µB ≡ µB − µ0B 2
= kB T ln x + λ (1 − x) , (2.413)
and their sum and difference, ∆µ± ≡ ∆µA ± ∆µB . From the Gibbs-Duhem relation, we know that we can write µB
as a function of T , p, and µA . Alternately, we could write ∆µ± in terms of T , p, and ∆µ∓ , so we can choose which
among ∆µ+ and ∆µ− we wish to use in our phase diagram. The results are plotted in Fig. 2.35. It is perhaps
easiest to understand the phase diagram in the (T, ∆µ− ) plane. At low temperatures, below T = Tc = λ/2kB ,
there is a first order phase transition at ∆µ− = 0. For T < Tc = λ/2kB and ∆µ− = 0+ , i.e. infinitesimally positive,
the system is in the A-rich phase, but for ∆µ− = 0− , i.e. infinitesimally negative, it is B-rich. The concentration
x = xB changes discontinuously across the phase boundary. The critical point lies at (T, ∆µ− ) = (λ/2kB , 0).
If we choose N = NA + NB to be the extensive variable, then fixing N means dNA + dNB = 0. So st fixed T and p,
dGT,p = µA dNA + µB dNB ⇒ dg T,p = −∆µ− dx . (2.414)
88 CHAPTER 2. THERMODYNAMICS
Figure 2.36: Phase diagram for the binary system. The black curve is the coexistence curve, and the dark red curve
is the spinodal. A-rich material is to the left and B-rich to the right.
Since ∆µ− (x, T ) = ϕ(x, T ) − ϕ(1 − x, T ) = −∆µ− (1 − x, T ), where ϕ(x, T ) = λx − kB T ln x, we have that the
R
1−x
coexistence boundary in the (x, ∆− ) plane is simply the line ∆µ− = 0, because dx′ ∆µ− (x′ , T ) = 0.
x
Note also that there is no two-phase region in the (T, ∆µ) plane; the phase boundary in this plane is a curve
which terminates at a critical point. As we saw in §2.12, the same situation pertains in single component (p, v, T )
systems. That is, the phase diagram in the (p, v) or (T, v) plane contains two-phase regions, but in the (p, T ) plane
the boundaries between phases are one-dimensional curves. Any two-phase behavior is confined to these curves,
where the thermodynamic potentials are singular.
The phase separation can be seen in a number of systems. A popular example involves mixtures of water and
ouzo or other anise-based liqueurs, such as arak and absinthe. Starting with the pure liqueur (x = 1), and at a
temperature below the coexistence curve maximum, the concentration is diluted by adding water. Follow along
on Fig. 2.36 by starting at the point (x = 1 , kB T /λ = 0.4) and move to the left. Eventually, one hits the boundary
of the two-phase region. At this point, the mixture turns milky, due to the formation of large droplets of the pure
phases on either side of coexistence region which scatter light, a process known as spontaneous emulsification22 . As
one continues to dilute the solution with more water, eventually one passes all the way through the coexistence
region, at which point the solution becomes clear once again, and described as a single phase.
What happens if λ < 0? In this case, both the entropy and the interaction energy prefer a mixed phase, and there is
no instability to phase separation. The two fluids are said to be completely miscible. An example would be benzene,
C6 H6 , and toluene, C7 H8 (i.e. C6 H5 CH3 ). The phase diagram would be blank, with no phase boundaries below
the boiling transition, because the fluid could exist as a mixture in any proportion.
Any fluid will eventually boil if the temperature is raised sufficiently high. Let us assume that the boiling points
∗
of our A and B fluids are TA,B , and without loss of generality let us take TA∗ < TB∗ at some given fixed pressure23 .
This means µLA (TA∗ , p) = µVA (TA∗ , p) and µLB (TB∗ , p) = µVB (TB∗ , p). What happens to the mixture? We begin by writing
22 An emulsion is a mixture of two or more immiscible liquids.
23 We assume the boiling temperatures are not exactly equal!
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 89
Figure 2.37: Gibbs free energy per particle g for an ideal binary solution for temperatures T ∈ [TA∗ , TB∗ ]. The
Maxwell construction is shown for the case TA∗ < T < TB∗ . Right: phase diagram, showing two-phase region and
distillation sequence in (x, T ) space.
the free energies of the mixed liquid and mixed vapor phases as
h i
gL (T, p, x) = (1 − x) µLA (T, p) + x µLB (T, p) + kB T x ln x + (1 − x) ln(1 − x) + λL x(1 − x) (2.415)
h i
gV (T, p, x) = (1 − x) µVA (T, p) + x µVB (T, p) + kB T x ln x + (1 − x) ln(1 − x) + λV x(1 − x) . (2.416)
Typically λV ≈ 0. Consider these two free energies as functions of the concentration x, at fixed T and p. If the
curves never cross, and gL (x) < gV (x) for all x ∈ [0, 1], then the liquid is always the state of lowest free energy.
This is the situation in the first panel of Fig. 2.37. Similarly, if gV (x) < gL (x) over this range, then the mixture is in
the vapor phase throughout. What happens if the two curves cross at some value of x? This situation is depicted
in the second panel of Fig. 2.37. In this case, there is always a Maxwell construction which lowers the free energy
throughout some range of concentration, i.e. the system undergoes phase separation.
In an ideal fluid, we have λL = λV = 0, and setting gL = gV requires
where ∆µA/B (T, p) = µLA/B (T, p) − µVA/B (T, p). Expanding the chemical potential about a given temperature T ∗ ,
cp (T ∗ , p)
µ(T, p) = µ(T ∗ , p) − s(T ∗ , p) (T − T ∗ ) − (T − T ∗ )2 + . . . , (2.418)
2T
∂µ
∂S
∂s
where we have used ∂T p,N
= − ∂N T,p
= −s(T, p), the entropy per particle, and ∂T p,N = cp /T . Thus,
∗
expanding ∆µA/B about TA/B , we have
cVpA − cLpA
∆µA ≡ µLA − µVA = (sVA − sLA )(T − TA∗ ) + (T − TA∗ )2 + . . .
2TA∗
(2.419)
cVpB − cLpB
∆µB ≡ µLB − µVB = (sVB − sLB )(T − TB∗ ) + (T − TB∗ )2 + . . .
2TB∗
90 CHAPTER 2. THERMODYNAMICS
Figure 2.38: Negative (left) and positive (right) azeotrope phase diagrams. From Wikipedia.
We assume sVA/B > sLA/B , i.e. the vapor phase has greater entropy per particle. Thus, ∆µA/B (T ) changes sign from
∗
negative to positive as T rises through TA/B . If we assume that these are the only sign changes for ∆µA/B (T ) at
fixed p, then eqn. 2.417 can only be solved for T ∈ [TA∗ , TB∗ ]. This immediately leads to the phase diagram in the
rightmost panel of Fig. 2.37.
According to the Gibbs phase rule, with σ = 2, two-phase equilibrium (ϕ = 2) occurs along a subspace of dimen-
sion dPE = 2 + σ − ϕ = 2. Thus, if we fix the pressure p and the concentration x = xB , liquid-gas equilibrium
occurs at a particular temperature T ∗ , known as the boiling point. Since the liquid and the vapor with which it
is in equilibrium at T ∗ may have different composition, i.e. different values of x, one may distill the mixture to
separate the two pure substances, as follows. First, given a liquid mixture of A and B, we bring it to boiling, as
shown in the rightmost panel of Fig. 2.37. The vapor is at a different concentration x than the liquid (a lower value
of x if the boiling point of pure A is less than that of pure B, as shown). If we collect the vapor, the remaining fluid
is at a higher value of x. The collected vapor is then captured and then condensed, forming a liquid at the lower
x value. This is then brought to a boil, and the resulting vapor is drawn off and condensed, etc The result is a
purified A state. The remaining liquid is then at a higher B concentration. By repeated boiling and condensation,
A and B can be separated. For liquid-vapor transitions, the upper curve, representing the lowest temperature at a
given concentration for which the mixture is a homogeneous vapor, is called the dew point curve. The lower curve,
representing the highest temperature at a given concentration for which the mixture is a homogeneous liquid,
is called the bubble point curve. The same phase diagram applies to liquid-solid mixtures where both phases are
completely miscible. In that case, the upper curve is called the liquidus, and the lower curve the solidus.
When a homogeneous liquid or vapor at concentration x is heated or cooled to a temperature T such that (x, T )
lies within the two-phase region, the mixture phase separates into the the two end components (x∗L , T ) and (x∗V , T ),
which lie on opposite sides of the boundary of the two-phase region, at the same temperature. The locus of points
at constant T joining these two points is called the tie line. To determine how much of each of these two homo-
geneous phases separates out, we use particle number conservation. If ηL,V is the fraction of the homogeneous
liquid and homogeneous vapor phases present, then ηL x∗L + ηV x∗V = x, which says ηL = (x − x∗V )/(x∗L − x∗V ) and
ηV = (x − x∗L )/(x∗V − x∗L ). This is known as the lever rule.
2.13. ENTROPY OF MIXING AND THE GIBBS PARADOX 91
Figure 2.39: Free energies before Maxwell constructions for a binary fluid mixture in equilibrium with a vapor
(λV = 0). Panels show (a) λL = 0 (ideal fluid), (b) λL < 0 (miscible fluid; negative azeotrope), (c) λLAB > 0
(positive azeotrope), (d) λLAB > 0 (heteroazeotrope). Thick blue and red lines correspond to temperatures TA∗ and
TB∗ , respectively, with TA∗ < TB∗ . Thin blue and red curves are for temperatures outside the range [TA∗ , TB∗ ]. The
black curves show the locus of points where g is discontinuous, i.e. where the liquid and vapor free energy curves
cross. The yellow curve in (d) corresponds to the coexistence temperature for the fluid mixture. In this case the
azeotrope forms within the coexistence region.
For many binary mixtures, the boiling point curve is as shown in Fig. 2.38. Such cases are called azeotropes. For
∗
negative azeotropes, the maximum of the boiling curve lies above both TA,B . The free energy curves for this case
∗ ∗
are shown in panel (b) of Fig. 2.39. For x < x , where x is the azeotropic composition, one can distill A but
not B. Similarly, for x > x∗ one can distill B but not A. The situation is different for positive azeotropes, where the
∗
minimum of the boiling curve lies below both TA,B , corresponding to the free energy curves in panel (c) of Fig.
2.39. In this case, distillation (i.e. condensing and reboiling the collected vapor) from either side of x∗ results in
the azeotrope. One can of course collect the fluid instead of the vapor. In general, for both positive and negative
azeotropes, starting from a given concentration x, one can only arrive at pure A plus azeotrope (if x < x∗ ) or
pure B plus azeotrope (if x > x∗ ). Ethanol (C2 H5 OH) and water (H2 O) form a positive azeotrope which is 95.6%
ethanol and 4.4% water by weight. The individual boiling points are TC∗2 H5 OH = 78.4◦C , TH∗ 2 O = 100◦ C, while
∗
the azeotrope boils at TAZ = 78.2◦C. No amount of distillation of this mixture can purify ethanol beyond the
95.6% level. To go beyond this level of purity, one must resort to azeotropic distillation, which involved introducing
another component, such as benzene (or a less carcinogenic additive), which alters the molecular interactions.
92 CHAPTER 2. THERMODYNAMICS
Figure 2.40: Phase diagram for a eutectic mixture in which a liquid L is in equilibrium with two solid phases α and
β. The same phase diagram holds for heteroazeotropes, where a vapor is in equilibrium with two liquid phases.
To model the azeotrope system, we need to take λL 6= 0, in which case one can find two solutions to the energy
crossing condition gV (x) = gL (x). With two such crossings come two Maxwell constructions, hence the phase dia-
grams in Fig. 2.38. Generally, negative azeotropes are found in systems with λL < 0 , whereas positive azeotropes
are found when λL > 0. As we’ve seen, such repulsive interactions between the A and B components in general
lead to a phase separation below a coexistence temperature TCOEX (x) given by eqn. 2.410. What happens if the
minimum boiling point lies within the coexistence region? This is the situation depicted in panel (d) of Fig. 2.39.
The system is then a liquid/vapor version of the solid/liquid eutectic (see Fig. 2.40), and the minimum boiling
point mixture is called a heteroazeotrope.
ζ1 A1 + ζ2 A2 + · · · + ζσ Aσ = 0 , (2.420)
where
Aa = chemical formula
ζa = stoichiometric coefficient .
for which
ζ(H2 ) = −3 , ζ(N2 ) = −1 , ζ(NH3 ) = 2 . (2.422)
2.14. SOME CONCEPTS IN THERMOCHEMISTRY 93
When ζa > 0, the corresponding Aa is a product; when ζa < 0, the corresponding Aa is a reactant. The bookkeeping
of the coefficients ζa which ensures conservation of each individual species of atom in the reaction(s) is known as
stoichiometry24
Now we ask: what are the conditions for equilibrium?
At constant T and p, which is typical for many chemical
reactions, the conditions are that G T, p, {Na} be a minimum. Now
X
dG = −S dT + V dp + µa dNa , (2.423)
i
so if we let the reaction go forward, we have dNa = ζa , and if it runs in reverse we have dNa = −ζa . Thus, setting
dT = dp = 0, we have the equilibrium condition
σ
X
ζa µa = 0 . (2.424)
a=1
Let us investigate the consequences of this relation for ideal gases. The chemical potential of the ath species is
µa (T, p) = kB T φa (T ) + kB T ln pa . (2.425)
P
Here pa = p xa is the partial pressure of species a, where xa = Na / b Nb the dimensionless concentration of
species a. Chemists sometimes write xa = [Aa ] for the concentration of species a. In equilibrium we must have
X h i
ζa ln p + ln xa + φa (T ) = 0 , (2.426)
a
which says X X X
ζa ln xa = − ζa ln p − ζa φa (T ) . (2.427)
a a a
The quantity κ(p, T ) is called the equilibrium constant. When κ is large, the LHS of the above equation is large.
This favors maximal concentration xa for the products (ζa > 0) and minimal concentration xa for the reactants
(ζa < 0). This means that the equation REACTANTS ⇋ PRODUCTS is shifted to the right, i.e. the products are
plentiful and the reactants are scarce. When κ is small, the LHS is small and the reaction is shifted to the left, i.e.
the reactants are plentiful and the products P are scarce. Remember we are describing equilibrium conditions here.
Now we observe that reactions for which a ζa > 0 shift to the left with increasing pressure and shift to the right
24 Antoine Lavoisier, the ”father of modern chemistry”, made pioneering contributions in both chemistry and biology. In particular, he is
often credited as the progenitor of stoichiometry. An aristocrat by birth, Lavoisier was an administrator of the Ferme générale, an organization
in pre-revolutionary France which collected taxes on behalf of the king. At the age of 28, Lavoisier married Marie-Anne Pierette Paulze, the
13-year-old daughter of one of his business partners. She would later join her husband in his research, and she played a role in his disproof of
the phlogiston theory of combustion. The phlogiston theory was superseded by Lavoisier’s work, where, based on contemporary experiments
by Joseph Priestley, he correctly identified the pivotal role played by oxygen in both chemical and biological processes (i.e. respiration). Despite
his fame as a scientist, Lavoisier succumbed to the Reign of Terror. His association with the Ferme générale, which collected taxes from the poor
and the downtrodden, was a significant liability in revolutionary France (think Mitt Romney vis-a-vis Bain Capital). Furthermore – and let
this be a lesson to all of us – Lavoisier had unwisely ridiculed a worthless pseudoscientific pamphlet, ostensibly on the physics of fire, and
its author, Jean-Paul Marat. Marat was a journalist with scientific pretensions, but apparently little in the way of scientific talent or acumen.
Lavoisier effectively blackballed Marat’s candidacy to the French Academy of Sciences, and the time came when Marat sought revenge. Marat
was instrumental in getting Lavoisier and other members of the Ferme générale arrested on charges of counterrevolutionary activities, and on
May 8, 1794, after a trial lasting less than a day, Lavoisier was guillotined. Along with Fourier and Carnot, Lavoisier’s name is one of the 72
engraved on the Eiffel Tower. Source: http://www.vigyanprasar.gov.in/scientists/ALLavoisier.htm.
94 CHAPTER 2. THERMODYNAMICS
P
with decreasing pressure, while reactions for which a ζa > 0 the situation
P is reversed: they shift to the right with
increasing pressure and to the left with decreasing pressure. When a ζa = 0 there is no shift upon increasing or
decreasing pressure.
The rate at which the equilibrium constant changes with temperature is given by
X
∂ ln κ
=− ζa φ′a (T ) . (2.429)
∂T p a
Now from eqn. 2.425 we have that the enthalpy per particle for species i is
∂µa
ha = µa − T , (2.430)
∂T p
∂G
since H = G + T S and S = − ∂T p . We find
and thus P
∂ ln κ i ζa ha ∆h
= = , (2.432)
∂T p kB T 2 kB T 2
where ∆h is the enthalpy of the reaction, which is the heat absorbed or emitted as a result of the reaction.
When ∆h > 0 the reaction is endothermic and the yield increases with increasing T . When ∆h < 0 the reaction is
exothermic and the yield decreases with increasing T .
As an example, consider the reaction H2 + I2 ⇋ 2 HI. We have
Suppose our initial system consists of ν10 moles of H2 , ν20 = 0 moles of I2P
, and ν30 moles of undissociated HI . These
0
mole numbers determine the initial concentrations xa , where xa = νa / b νb . Define
x03 − x3
α≡ , (2.434)
x3
in which case we have
Most chemical reactions take place under constant pressure. The heat Qif associated with a given isobaric process
is
Zf Zf
Qif = dE + p dV = (Ef − Ei ) + p (Vf − Vi ) = Hf − Hi , (2.437)
i i
2.14. SOME CONCEPTS IN THERMOCHEMISTRY 95
∆H0f ∆H0f
Formula Name State kJ/mol Formula Name State kJ/mol
Ag Silver crystal 0.0 NiSO4 Nickel sulfate crystal -872.9
C Graphite crystal 0.0 Al2 O3 Aluminum oxide crystal -1657.7
C Diamond crystal 1.9 Ca3 P2 O8 Calcium phosphate gas -4120.8
O3 Ozone gas 142.7 HCN Hydrogen cyanide liquid 108.9
H2 O Water liquid -285.8 SF6 Sulfur hexafluoride gas -1220.5
H3 BO3 Boric acid crystal -1094.3 CaF2 Calcium fluoride crystal -1228.0
ZnSO4 Zinc sulfate crystal -982.8 CaCl2 Calcium chloride crystal -795.4
The enthalpy of formation of any substance in its standard state is zero at all temperatures, by definition: ∆H0f [O2 ] =
∆H0f [He] = ∆H0f [K] = ∆H0f [Mn] = 0, etc.
Suppose now we have a reaction
∆H
a A + b B −−−−−→ c C + d D . (2.444)
To compute the reaction enthalpy ∆H, we can imagine forming the components A and B from their standard state
constituents. Similarly, we can imagine doing the same for C and D. Since the number of atoms of a given kind is
conserved in the process, the constituents of the reactants must be the same as those of the products, we have
Figure 2.41: Left panel: reaction enthalpy and activation energy (exothermic case shown). Right panel: reaction
enthalpy as a difference between enthalpy of formation of reactants and products.
A list of a few enthalpies of formation is provided in table 2.5. Note that the reaction enthalpy is independent of
the actual reaction path. That is, the difference in enthalpy between A and B is the same whether the reaction is
A −→ B or A −→ X −→ (Y + Z) −→ B. This statement is known as Hess’s Law.
Note that
dH = dE + p dV + V dp = d̄Q + V dp , (2.446)
hence
d̄Q ∂H
Cp = = . (2.447)
dT p ∂T p
We therefore have
ZT
H(T, p, ν) = H(T0 , p, ν) + ν dT ′ cp (T ′ ) . (2.448)
T0
For ideal gases, we have cp (T ) = (1+ 21 f ) R. For real gases, over a range of temperatures, there are small variations:
cp (T ) = α + β T + γ T 2 . (2.449)
J J J
O2 : α = 25.503 , β = 13.612 × 10−3 , γ = −42.553 × 10−7
mol K mol K2 mol K3
J J J
H2 O : α = 30.206 , β = 9.936 × 10−3 , γ = 11.14 × 10−7
mol K mol K2 mol K3
If all the gaseous components in a reaction can be approximated as ideal, then we may write
X
(∆H)rxn = (∆E)rxn + ζa RT , (2.450)
a
where the subscript ‘rxn’ stands for ‘reaction’. Here (∆E)rxn is the change in energy from reactants to products.
2.14. SOME CONCEPTS IN THERMOCHEMISTRY 97
Table 2.6: Average bond enthalpies for some common bonds. (Source: L. Pauling, The Nature of the Chemical Bond
(Cornell Univ. Press, NY, 1960).)
The enthalpy needed to break a chemical bond is called the bond enthalpy, h[ • ]. The bond enthalpy is the energy
required to dissociate one mole of gaseous bonds to form gaseous atoms. A table of bond enthalpies is given in
Tab. 2.6. Bond enthalpies are endothermic, since energy is required to break chemical bonds. Of course, the actual
bond energies can depend on the location of a bond in a given molecule, and the values listed in the table reflect
averages over the possible bond environment.
The bond enthalpies in Tab. 2.6 may be used to compute reaction enthalpies. Consider, for example, the reaction
2 H2 (g) + O2 (g) −→ 2 H2 O(l). We then have, from the table,
(∆H)rxn = 2 h[H−H] + h[O = O] − 4 h[H−O]
(2.451)
= −483 kJ/mol O2 .
Thus, 483 kJ of heat would be released for every two moles of H2 O produced, if the H2 O were in the gaseous phase.
Since H2 O is liquid at STP, we should also include the condensation energy of the gaseous water vapor into liquid
water. At T = 100◦ C the latent heat of vaporization is ℓ̃ = 2270 J/g, but at T = 20◦ C, one has ℓ̃ = 2450 J/g, hence
with M = 18 we have ℓ = 44.1 kJ/mol. Therefore, the heat produced by the reaction 2 H2 (g) + O2 (g) −⇀ ↽− 2 H2 O(l)
is (∆H)rxn = −571.2 kJ / mol O2 . Since the reaction produces two moles of water, we conclude that the enthalpy
of formation of liquid water at STP is half this value: ∆H0f [H2 O] = 285.6 kJ/mol.
Consider next the hydrogenation of ethene (ethylene): C2 H4 + H2 ↽− −⇀ C2 H6 . The product is known as ethane.
The energy accounting is shown in Fig. 2.42. To compute the enthalpies of formation of ethene and ethane from
the bond enthalpies, we need one more bit of information, which is the standard enthalpy of formation of C(g)
from C(s), since the solid is the standard state at STP. This value is ∆H0f [C(g)] = 718 kJ/mol. We may now write
−2260 kJ
2 C(g) + 4 H(g) −−−−−−−−−→ C2 H4 (g)
1436 kJ
2 C(s) −−−−−−−−−→ 2 C(g)
872 kJ
2 H2 (g) −−−−−−−−−→ 4 H(g) .
Thus, using Hess’s law, i.e. adding up these reaction equations, we have
48 kJ
2 C(s) + 2 H2 (g) −−−−−−−−−→ C2 H4 (g) .
98 CHAPTER 2. THERMODYNAMICS
Figure 2.42: Calculation of reaction enthalpy for the hydrogenation of ethene (ethylene), C2 H4 .
If there are K independent variables {x1 , . . . , xK }, then there are 12 K(K − 1) independent equations of the above
form – one for each distinct (i, j) pair. These equations can be written compactly as
∂L
Ωijk = Fij , (2.456)
∂xk
where
∂Aj ∂Ai
Fij = − . (2.458)
∂xi ∂xj
Note that Fij is antisymmetric, and resembles a field strength tensor, and that Ωijk = −Ωjik is antisymmetric in
the first two indices (but is not totally antisymmetric in all three).
Can we solve these 21 K(K − 1) coupled equations to find an integrating factor L? In general the answer is no.
However, when K = 2 we can always find an integrating factor. To see why, let’s call x ≡ x1 and y ≡ x2 . Consider
now the ODE
dy A (x, y)
=− x . (2.459)
dx Ay (x, y)
This equation can be integrated to yield a one-parameter set of integral curves, indexed by an initial condition.
The equation for these curves may be written as Uc (x, y) = 0, where c labels the curves. Then along each curve
we have
dUc ∂Ux ∂Uc dy
0= = +
dx ∂x ∂y dx
(2.460)
∂Uc Ax ∂Uc
= − .
∂x Ay ∂y
Thus,
∂Uc ∂Uc
A = A ≡ e−L Ax Ay . (2.461)
∂x y ∂y x
This equation defines the integrating factor L :
1 ∂Uc 1 ∂Uc
L = − ln = − ln . (2.462)
Ax ∂x Ay ∂y
A convex function of a single variable f (x) is one for which f ′′ (x) > 0 everywhere. The Legendre transform of a
convex function f (x) is a function g(p) defined as follows. Let p be a real number, and consider the line y = px,
as shown in Fig. 2.43. We define the point x(p) as the value of x for which the difference F (x, p) = px − f (x) is
100 CHAPTER 2. THERMODYNAMICS
greatest. Then define g(p) = F x(p), p .25 The value x(p) is unique if f (x) is convex, since x(p) is determined by
the equation
f ′ x(p) = p . (2.465)
Note that from p = f ′ x(p) we have, according to the chain rule,
d ′ h i−1
f x(p) = f ′′ x(p) x′ (p) =⇒ x′ (p) = f ′′ x(p) . (2.466)
dp
From this, we can prove that g(p) is itself convex:
dh i
g ′ (p) = p x(p) − f x(p)
dp (2.467)
= p x′ (p) + x(p) − f ′ x(p) x′ (p) = x(p) ,
hence h i−1
g ′′ (p) = x′ (p) = f ′′ x(p) >0. (2.468)
In higher dimensions, the generalization of the definition f ′′ (x) > 0 is that a function F (x1 , . . . , xn ) is convex if
the matrix of second derivatives, called the Hessian,
∂ 2F
Hij (x) = (2.469)
∂xi ∂xj
is positive definite. That is, all the eigenvalues of Hij (x) must be positive for every x. We then define the Legendre
transform G(p) as
G(p) = p · x − F (x) (2.470)
where
p = ∇F . (2.471)
25 Note that g(p) may be a negative number, if the line y = px lies everywhere below f (x).
2.16. APPENDIX II : LEGENDRE TRANSFORMATIONS 101
Note that
dG = x · dp + p · dx − ∇F · dx = x · dp , (2.472)
which establishes that G is a function of p and that
∂G
= xj . (2.473)
∂pj
Note also that the Legendre transformation is self dual, which is to say that the Legendre transform of G(p) is F (x):
F → G → F under successive Legendre transformations.
We can also define a partial Legendre transformation as follows. Consider a function of q variables F (x, y), where
x = {x1 , . . . , xm } and y = {y1 , . . . , yn }, with q = m + n. Define p = {p1 , . . . , pm }, and
where
∂F
pa = (a = 1, . . . , m) . (2.475)
∂xa
These equations are then to be inverted to yield
∂G
xa = xa (p, y) = . (2.476)
∂pa
Note that
∂F
pa = x(p, y), y . (2.477)
∂xa
Thus, from the chain rule,
∂pa ∂ 2F ∂xc ∂ 2F ∂ 2G
δab = = = , (2.478)
∂pb ∂xa ∂xc ∂pb ∂xa ∂xc ∂pc ∂pb
which says
∂ 2G ∂xa
= = K−1
ab , (2.479)
∂pa ∂pb ∂pb
where the m × m partial Hessian is
∂ 2F ∂p
= a = Kab . (2.480)
∂xa ∂xb ∂xb
Note that Kab = Kba is symmetric. And with respect to the y coordinates,
∂ 2G ∂ 2F
=− = −Lµν , (2.481)
∂yµ ∂yν ∂yµ ∂yν
where
∂ 2F
Lµν = (2.482)
∂yµ ∂yν
is the partial Hessian in the y coordinates. Now it is easy to see that if the full q × q Hessian matrix Hij is positive
definite, then any submatrix such as Kab or Lµν must also be positive definite. In this case, the partial Legendre
transform is convex in {p1 , . . . , pm } and concave in {y1 , . . . , yn }.
102 CHAPTER 2. THERMODYNAMICS
Consider a set of n independent variables {x1 , . . . , xn }, which can be thought of as a point in n-dimensional space.
Let {y1 , . . . , yn } and {z1 , . . . , zn } be other choices of coordinates. Then
Note that this entails a matrix multiplication: Aik = Bij Cjk , where Aik = ∂xi /∂zk , Bij = ∂xi /∂yj , and Cjk =
∂yj /∂zk . We define the determinant
∂xi ∂(x1 , . . . , xn )
det ≡ . (2.484)
∂zk ∂(z1 , . . . , zn )
Such a determinant is called a Jacobean. Now if A = BC, then det(A) = det(B) · det(C). Thus,
We also have
∂(x, y) ∂(u, v) ∂(x, y)
· = . (2.488)
∂(u, v) ∂(r, s) ∂(r, s)
From this simple mathematics follows several very useful results.
1) First, write " #−1
∂(x, y) ∂(u, v)
= . (2.489)
∂(u, v) ∂(x, y)
Now let y = v:
∂(x, y) ∂x 1
= = ∂u
. (2.490)
∂(u, y) ∂u y ∂x y
Thus,
∂x . ∂u
=1 (2.491)
∂u y ∂x y
2) Second, we have
∂(x, y) ∂x ∂(x, y) ∂(x, u) ∂y ∂x
= = · =− ,
∂(u, y) ∂u y ∂(x, u) ∂(u, y) ∂u x ∂y u
which is to say
∂x ∂y ∂x
=− . (2.492)
∂y u ∂u x ∂u y
2.17. APPENDIX III : USEFUL MATHEMATICAL RELATIONS 103
3) Third, we have
∂(x, v) ∂(x, v) ∂(y, v)
= · , (2.494)
∂(u, v) ∂(y, v) ∂(u, v)
which says
∂x ∂x ∂y
= (2.495)
∂u v ∂y v ∂u v
This is simply the chain rule of partial differentiation.
4) Fourth, we have
(∂x, ∂y) (∂x, ∂y) (∂u, ∂v)
= ·
(∂u, ∂y) (∂u, ∂v) (∂u, ∂y)
(2.496)
∂x ∂y ∂v ∂x ∂y ∂v
= − ,
∂u v ∂v u ∂y u ∂v u ∂u v ∂y u
which says
∂x ∂x ∂x ∂y
= − (2.497)
∂u y ∂u v ∂y u ∂u v
5) Fifth, whenever we differentiate one extensive quantity with respect to another, holding only intensive quanti-
ties constant, the result is simply the ratio of those extensive quantities. For example,
∂S S
= . (2.498)
∂V p,T V
The reason should be obvious. In the above example, S(p, V, T ) = V φ(p, T ), where φ is a function of the two
intensive quantities p and T . Hence differentiating S with respect to V holding p and T constant is the same as
dividing S by V . Note that this implies
∂S ∂S ∂S S
= = = , (2.499)
∂V p,T ∂V p,µ ∂V n,T V
dE = x dy + u dv . (2.500)
That is,
∂E ∂E
x= ≡ Ey , u= ≡ Ev . (2.501)
∂y v ∂v y
Writing
Next, we use eqn. 2.503 with du = 0 to eliminate dy in favor of dv, and then substitute into eqn. 2.502. This yields
∂x Eyy Evv
= Eyv − . (2.506)
∂v u Evy
Combining the results of eqns. 2.504, 2.505, 2.506, and 2.507, we have
∂(x, y) ∂x∂y ∂x ∂y
= −
∂(u, v) ∂u
v ∂v u ∂v u ∂u v
(2.508)
Eyy E Eyy Evv 1
= − vv − Eyv − = −1 .
Evy Evy Evy Evy
Thus,
∂(T, S)
=1. (2.509)
∂(p, V )
Nota bene: It is important to understand what other quantities are kept constant, otherwise we can run into
trouble. For example, it would seem that eqn. 2.508 would also yield
∂(µ, N )
=1. (2.510)
∂(p, V )
∂(T, S, N ) ∂(T, S, µ)
= −1 = −1
∂(y, X, N ) ∂(y, X, µ)
∂(T, S, X) ∂(T, S, y)
= −1 = −1 (2.512)
∂(µ, N, X) ∂(µ, N, y)
∂(y, X, S) ∂(y, X, T )
= −1 = −1
∂(µ, N, S) ∂(µ, N, T )
If we are careful, then the general result in eq. 2.512, where (y, X) = (−p, V ) or (H α , M α ) or (E α , P α ), can be
quite handy, especially when used in conjunction with eqn. 2.485. For example, we have
=1
z }| {
∂S ∂(T, S, N ) ∂(T, S, N ) ∂(p, V, N ) ∂p
= = · = , (2.514)
∂V T,N ∂(T, V, N ) ∂(p, V, N ) ∂(T, V, N ) ∂T V,N
which is one of the Maxwell relations derived from the exactness of dF . Some other examples:
∂V ∂(V, p, N ) ∂(V, p, N ) ∂(S, T, N ) ∂T
= = · = (2.515)
∂S p,N ∂(S, p, N ) ∂(S, T, N ) ∂(S, p, N ) ∂p S,N
∂S ∂(S, T, p) ∂(S, T, p) ∂(µ, N, p) ∂µ
= = · =− . (2.516)
∂N T,p ∂(N, T, p) ∂(µ, N, p) ∂(N, T, p) ∂T p,N
Note that due to the alternating nature of the determinant – it is antisymmetric under interchange of any two rows
or columns – we have
∂(x, y, z) ∂(y, x, z) ∂(y, x, z)
=− = = ... . (2.517)
∂(u, v, w) ∂(u, v, w) ∂(w, v, u)
In general, it is usually advisable to eliminate S from a Jacobean. If we have a Jacobean involving T , S, and N , we
can write
∂(T, S, N ) ∂(T, S, N ) ∂(p, V, N ) ∂(p, V, N )
= = , (2.518)
∂( • , • , N ) ∂(p, V, N ) ∂( • , • , N ) ∂( • , • , N )
where each • is a distinct arbitrary state variable other than N .
For example,
∂T ∂(T, S, N ) ∂(T, S, N ) ∂(p, V, N ) ∂(p, T, N ) T ∂V
= = · · = (2.521)
∂p S,N ∂(p, S, N ) ∂(p, V, N ) ∂(p, T, N ) ∂(p, S, N ) Cp ∂T p,N
∂V ∂(V, S, N ) ∂(V, S, N ) ∂(V, T, N ) ∂(p, T, N ) C ∂V
= = · · = V . (2.522)
∂p S,N ∂(p, S, N ) ∂(V, T, N ) ∂(p, T, N ) ∂(p, S, N ) Cp ∂p T,N
Chapter 3
3.1 References
– R. Balescu, Equilibrium and Nonequilibrium Statistical Mechanics (Wiley, 1975)
An advanced text with an emphasis on fluids and kinetics.
– R. Balian, From Macrophysics to Microphysics (2 vols., Springer-Verlag, 2006)
A very detailed discussion of the fundamental postulates of statistical mechanics and their implications.)
107
108 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
3.2.1 Equilibrium
A thermodynamic system typically consists of an enormously large number of constituent particles, a typical
‘large number’ being Avogadro’s number, NA = 6.02 × 1023 . Nevertheless, in equilibrium, such a system is char-
acterized by a relatively small number of thermodynamicstate variables. Thus, while a complete description of a
(classical) system would require us to account for O 1023 evolving degrees of freedom, with respect to the phys-
ical quantities in which we are interested, the details of the initial conditions are effectively forgotten over some
microscopic time scale τ , called the collision time, and over some microscopic distance scale, ℓ, called the mean
free path1 . The equilibrium state is time-independent.
Relaxation to equilibrium is often modeled with something called the master equation. Let Pi (t) be the probability
that the system is in a quantum or classical state i at time t. Then write
dPi X
= Wij Pj − Wji Pi . (3.1)
dt j
Here, Wij is the rate at which j makes a transition to i. Note that we can write this equation as
dPi X
=− Γij Pj , (3.2)
dt j
where (
−W if i 6= j
Γij = P′ ij (3.3)
k Wkj if i = j ,
where the prime on the sum indicates that k = j is to be excluded. The constraints on the Wij are that Wij ≥ 0 for
all i, j, and we may take Wii ≡ 0 (no sum on i). Fermi’s Golden Rule of quantum mechanics says that
2π 2
Wij = h i | V̂ | j i ρ(Ej ) , (3.4)
~
where Ĥ0 i = Ei i , V̂ is an additional potential which leads to transitions, and ρ(Ei ) is the density of final
states at energy Ei . The fact that Wij ≥ 0 means that if each Pi (t = 0) ≥ 0, then Pi (t) ≥ 0 for all t ≥ 0. To see this,
suppose that at some time t > 0 one of the probabilities Pi is crossing zero and about to become negative. But
P
then eqn. 3.1 says that Ṗi (t) = j Wij Pj (t) ≥ 0. So Pi (t) can never become negative.
If the transition rates Wij are themselves time-independent, then we may formally write
Pi (t) = e−Γ t ij
Pj (0) . (3.5)
1 Exceptions involve quantities which are conserved by collisions, such as overall particle number, momentum, and energy. These quantities
relax to equilibrium in a special way called hydrodynamics.
3.2. MODELING THE APPROACH TO EQUILIBRIUM 109
Here we have used the Einstein ‘summation convention’ in which repeated indices are summed over (in this case,
the j index). Note that X
Γij = 0 , (3.6)
i
P
which says that the total probability i Pi is conserved:
d X X X X
Pi = − Γij Pj = − Pj Γij = 0 . (3.7)
dt i i,j j i
We conclude that φ ~ = (1, 1, . . . , 1) is a left eigenvector of Γ with eigenvalue λ = 0. The corresponding right
eigenvector, which we write as Pieq , satisfies Γij Pjeq = 0, and is a stationary (i.e. time independent) solution to the
master equation. Generally, there is only one right/left eigenvector pair corresponding to λ = 0, in which case
any initial probability distribution Pi (0) converges to Pieq as t → ∞, as shown in Appendix I (§3.7).
In equilibrium, the net rate of transitions into a state | i i is equal to the rate of transitions out of | i i. If, for each
state | j i the transition rate from | i i to | j i is equal to the transition rate from | j i to | i i, we say that the rates
satisfy the condition of detailed balance. In other words,
Wij Pjeq = Wji Pieq . (3.8)
Assuming Wij 6= 0 and Pjeq 6= 0, we can divide to obtain
Wji Pjeq
= eq . (3.9)
Wij Pi
Note that detailed balance is a stronger condition than that required for a stationary solution to the master equa-
tion.
If Γ = Γ t is symmetric, then the right eigenvectors and left eigenvectors are transposes of each other, hence
P eq = 1/N , where N is the dimension of Γ . The system then satisfies the conditions of detailed balance. See
Appendix II (§3.8) for an example of this formalism applied to a model of radioactive decay.
Suppose for the moment that Γ is a symmetric matrix, i.e. Γij = Γji . Then construct the function
X
H(t) = Pi (t) ln Pi (t) . (3.10)
i
Then
dH X dPi X dP
i
= 1 + ln Pi ) = ln Pi
dt i
dt i
dt
X
=− Γij Pj ln Pi (3.11)
i,j
X
= Γij Pj ln Pj − ln Pi ,
i,j
P
where we have used i Γij = 0. Now switch i ↔ j in the above sum and add the terms to get
dH 1X
= Γij Pi − Pj ln Pi − ln Pj . (3.12)
dt 2 i,j
110 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
Note that the i = j term does not contribute to the sum. For i 6= j we have Γij = −Wij ≤ 0, and using the result
(x − y) (ln x − ln y) ≥ 0 , (3.13)
we conclude
dH
≤0. (3.14)
dt
In equilibrium, Pieq is a constant, independent of i. We write
1 X
Pieq = , Ω= 1 =⇒ H = − ln Ω . (3.15)
Ω i
If Γij 6= Γji , we can still prove a version of the H-theorem. Define a new symmetric matrix
Then
" #
dH 1 X Pi Pj Pi Pj
=− W ij − ln − ln ≤0. (3.18)
dt 2 i,j Pieq Pjeq Pieq Pjeq
The master equation provides us with a semi-phenomenological description of a dynamical system’s relaxation to
equilibrium. It explicitly breaks time reversal symmetry. Yet the microscopic laws of Nature are (approximately)
time-reversal symmetric. How can a system which obeys Hamilton’s equations of motion come to equilibrium?
Let’s start our investigation by reviewing the basics of Hamiltonian
dynamics.
R Recall the Lagrangian L = L(q, q̇, t) =
T − V . The Euler-Lagrange equations of motion for the action S q(t) = dt L are
d ∂L ∂L
ṗσ = = , (3.19)
dt ∂ q̇σ ∂qσ
∂L
pσ = . (3.20)
∂ q̇σ
Note that
r
X
∂L ∂L ∂L
dH = pσ dq̇σ + q̇σ dpσ − dqσ − dq̇ − dt
σ=1
∂qσ ∂ q̇σ σ ∂t
r (3.22)
X ∂L ∂L
= q̇σ dpσ − dqσ − dt .
σ=1
∂qσ ∂t
∂H ∂H ∂L
= q̇σ , =− = −ṗσ (3.23)
∂pσ ∂qσ ∂qσ
and
dH ∂H ∂L
= =− . (3.24)
dt ∂t ∂t
∂H
ϕ̇i = Jij , (3.26)
∂ϕj
where
0r×r 1r×r
J= (3.27)
−1r×r 0r×r
is a rank 2r matrix. Note that J t = −J, i.e. J is antisymmetric, and that J 2 = −12r×2r .
where
dµ = dϕ1 dϕ2 · · · dϕn , (3.31)
for an n-dimensional phase space. We then have
Z Z
∂ϕi (t + dt)
Ω(t + dt) = ′
dµ =
dµ , (3.32)
∂ϕj (t)
R(t+dt) R(t)
where
∂ϕi (t + dt) ∂(ϕ′1 , . . . , ϕ′n )
∂ϕ (t) ≡ ∂(ϕ , . . . , ϕ ) (3.33)
j 1 n
is a determinant,
′ which is the Jacobean of the transformation from the set of coordinates ϕi = ϕi (t) to the
coordinates ϕi = ϕi (t + dt) . But according to the dynamics, we have
ϕi (t + dt) = ϕi (t) + Vi ϕ(t) dt + O(dt2 ) (3.34)
and therefore
∂ϕi (t + dt) ∂Vi
= δij + dt + O(dt2 ) . (3.35)
∂ϕj (t) ∂ϕj
We now make use of the equality
ln det M = Tr ln M , (3.36)
3
for any matrix M , which gives us , for small ε,
2
det 1 + εA = exp Tr ln 1 + εA = 1 + ε Tr A + 1
2 ε2 Tr A − Tr (A2 ) + . . . (3.37)
Thus, Z
Ω(t + dt) = Ω(t) + dµ ∇·V dt + O(dt2 ) , (3.38)
R(t)
which says Z Z
dΩ
= dµ ∇·V = dS n̂ · V (3.39)
dt
R(t) ∂R(t)
Xn
∂Vi
∇·V = , (3.40)
i=1
∂ϕi
and we have used the divergence theorem to convert the volume integral of the divergence to a surface integral
of n̂ · V , where n̂ is the surface normal and dS is the differential element of surface area, and ∂R denotes the
boundary of the region R. We see that if ∇·V = 0 everywhere in phase space, then Ω(t) is a constant, and phase
space volumes are preserved by the evolution of the system.
For an alternative derivation, consider a function ̺(ϕ, t) which is defined to be the density of some collection of
points in phase space at phase space position ϕ and time t. This must satisfy the continuity equation,
∂̺
+ ∇·(̺V ) = 0 . (3.41)
∂t
3 The equality ln det M = Tr ln M is most easily proven by bringing the matrix to diagonal form via a similarity transformation, and
proving the equality for diagonal matrices.
3.3. PHASE FLOWS IN CLASSICAL MECHANICS 113
Figure 3.1: Time evolution of two immiscible fluids. The local density remains constant.
This is called the continuity equation. It says that ‘nobody gets lost’. If we integrate it over a region of phase space
R, we have Z Z Z
d
dµ ̺ = − dµ ∇·(̺V ) = − dS n̂ · (̺V ) . (3.42)
dt
R R ∂R
It is perhaps helpful to think of ̺ as a charge density, in which case J = ̺V is the current density. The above
equation then says Z
dQR
= − dS n̂ · J , (3.43)
dt
∂R
where QR is the total charge contained inside the region R. In other words, the rate of increase or decrease of the
charge within the region R is equal to the total integrated current flowing in or out of R at its boundary.
The Leibniz rule lets us write the continuity equation as
∂̺
+ V ·∇̺ + ̺ ∇·V = 0 . (3.44)
∂t
But now suppose that the phase flow is divergenceless, i.e. ∇·V = 0. Then we have
D̺ ∂
≡ + V ·∇ ̺ = 0 . (3.45)
Dt ∂t
The combination inside the brackets above is known as the convective derivative. It tells us the total rate of change
of ̺ for an observer moving with the phase flow. That is
d ∂̺ dϕi ∂̺
̺ ϕ(t), t = +
dt ∂ϕi dt ∂t
Xn (3.46)
∂ρ ∂̺ D̺
= Vi + = .
i=1
∂ϕi ∂t Dt
114 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
If D̺/Dt = 0, the local density remains the same during the evolution of the system. If we consider the ‘charac-
teristic function’ (
1 if ϕ ∈ R0
̺(ϕ, t = 0) = (3.47)
0 otherwise
then the vanishing of the convective derivative means that the image of the set R0 under time evolution will
always have the same volume.
Hamiltonian evolution in classical mechanics is volume preserving. The equations of motion are
∂H ∂H
q̇i = + , ṗi = − (3.48)
∂pi ∂qi
A point in phase space is specified by r positions qi and r momenta pi , hence the dimension of phase space is
n = 2r:
q q̇ ∂H/∂p
ϕ= , V = = . (3.49)
p ṗ −∂H/∂q
Hamilton’s equations of motion guarantee that the phase space flow is divergenceless:
r
X
∂ q̇i ∂ ṗi
∇·V = +
i=1
∂qi ∂pi
( ) (3.50)
X ∂ ∂H
r
∂
∂H
= + − =0.
i=1
∂qi ∂pi ∂pi ∂qi
D̺ ∂̺
≡ + V ·∇̺ = 0 , (3.51)
Dt ∂t
for any distribution ̺(ϕ, t) on phase space. Thus, the value of the density ̺(ϕ(t), t) is constant, which tells us that
the phase flow is incompressible. In particular, phase space volumes are preserved.
Let ̺(ϕ) = ̺(q, p) be a distribution on phase space. Assuming the evolution is Hamiltonian, we can write
r
X
∂̺ ∂ ∂
= −ϕ̇ · ∇̺ = − q̇k + ṗk ̺ = −iL̺̂ , (3.52)
∂t ∂qk ∂pk
k=1
Eqn. 3.52, known as Liouville’s equation, bears an obvious resemblance to the Schrödinger equation from quantum
mechanics.
Suppose that Λa (ϕ) is conserved by the dynamics of the system. Typical conserved quantities include the com-
ponents of the total linear momentum (if there is translational invariance), the components of the total angular
3.4. IRREVERSIBILITY AND POINCARÉ RECURRENCE 115
momentum (if there is rotational invariance), and the Hamiltonian itself (if the Lagrangian is not explicitly time-
dependent). Now consider a distribution ̺(ϕ, t) = ̺(Λ1 , Λ2 , . . . , Λk ) which is a function only of these various
conserved quantities. Then from the chain rule, we have
X ∂̺
ϕ̇ · ∇̺ = ϕ̇ · ∇Λa = 0 , (3.54)
a
∂Λa
The dynamics of the master equation describe an approach to equilibrium. These dynamics are irreversible:
dH/dt ≤ 0, where H is Boltzmann’s H-function. However, the microscopic laws of physics are (almost) time-
reversal invariant4 , so how can we understand the emergence of irreversibility? Furthermore, any dynamics
which are deterministic and volume-preserving in a finite phase space exhibits the phenomenon of Poincaré recur-
rence, which guarantees that phase space trajectories are arbitrarily close to periodic if one waits long enough.
The proof of the recurrence theorem is simple. Let gτ be the ‘τ -advance mapping’ which evolves points in phase
space according to Hamilton’s equations. Assume that gτ is invertible and volume-preserving, as is the case for
Hamiltonian flow. Further assume that phase space volume is finite. Since energy is preserved in the case of
time-independent Hamiltonians, we simply ask that the volume of phase space at fixed total energy E be finite, i.e.
Z
dµ δ E − H(q, p) < ∞ , (3.57)
Theorem: In any finite neighborhood R0 of phase space there exists a point ϕ0 which will return to R0 after m
applications of gτ , where m is finite.
Proof: Assume the theorem fails; we will show this assumption results in a contradiction. Consider the set Υ
formed from the union of all sets gτk R for all m:
∞
[
Υ= gτk R0 (3.58)
k=0
4 Actually, the microscopic laws of physics are not time-reversal invariant, but rather are invariant under the product P CT , where P is
parity, C is charge conjugation, and T is time reversal.
116 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
Figure 3.2: Successive images of a set R0 under the τ -advance mapping gτ , projected onto a two-dimensional
phase plane. The Poincaré recurrence theorem guarantees that if phase space has finite volume, and gτ is invertible
and volume preserving, then for any set R0 there exists an integer m such that R0 ∩ gτm R0 6= ∅.
We assume that the set {gτk R0 | k ∈ N} is disjoint5 . The volume of a union of disjoint sets is the sum of the
individual volumes. Thus,
∞
X
vol(Υ) = vol gτk R0
k=0
∞ (3.59)
X
= vol(R0 ) · 1=∞,
k=0
since vol gτk R0 = vol R0 from volume preservation. But clearly Υ is a subset of the entire phase space, hence
we have a contradiction, because by assumption phase space is of finite volume.
Thus, the assumption that the set {gτk R0 | k ∈ Z+ } is disjoint fails. This means that there exists some pair of integers
k and l, with k 6= l, such that gτk R0 ∩ gτl R0 6= ∅. Without loss of generality we may assume k < l. Apply the
inverse gτ−1 to this relation k times to get gτl−k R0 ∩ R0 6= ∅. Now choose any point ϕ1 ∈ gτm R0 ∩ R0 , where
m = l − k, and define ϕ0 = gτ−m ϕ1 . Then by construction both ϕ0 and gτm ϕ0 lie within R0 and the theorem is
proven.
Poincaré recurrence has remarkable implications. Consider a bottle of perfume which is opened in an otherwise
evacuated room, as depicted in fig. 3.3. The perfume molecules evolve according to Hamiltonian evolution.
The positions are bounded because physical space is finite. The momenta are bounded because the total energy is
conserved, hence no single particle can have a momentum such that T (p) > ETOT , where T (p) is the single particle
kinetic energy function6 . Thus, phase space, however large, is still bounded. Hamiltonian evolution, as we have
seen, is invertible and volume preserving, therefore the system is recurrent. All the molecules must eventually
return to the bottle. What’s more, they all must return with momenta arbitrarily close to their initial momenta!7
5 The natural numbers N is the set of non-negative integers {0, 1, 2, . . .}.
6 Inthe nonrelativistic limit, T = p2 /2m. For relativistic particles, we have T = (p2 c2 + m2 c4 )1/2 − mc2 .
7 Actually, what the recurrence theorem guarantees is that there is a configuration arbitrarily close to the initial one which recurs, to within
Figure 3.3: Poincaré recurrence guarantees that if we remove the cap from a bottle of perfume in an otherwise
evacuated room, all the perfume molecules will eventually return to the bottle! (Here H is the Hubble constant.)
The implications of the Poincaré recurrence theorem are surprising – even shocking. If one takes a bottle of
perfume in a sealed, evacuated room and opens it, the perfume molecules will diffuse throughout the room. The
recurrence theorem guarantees that after some finite time T all the molecules will go back inside the bottle (and
arbitrarily close to their initial velocities as well). The hitch is that this could take a very long time, e.g. much much
longer than the age of the Universe.
On less absurd time scales, we know that most systems come to thermodynamic equilibrium. But how can a
system both exhibit equilibration and Poincaré recurrence? The two concepts seem utterly incompatible!
118 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
Figure 3.4: Left: A configuration of the Kac ring with N = 16 sites and F = 4 flippers. The flippers, which live
on the links, are represented by blue dots. Right: The ring system after one time step. Evolution proceeds by
clockwise rotation. Spins passing through flippers are flipped.
A beautifully simple model due to Kac shows how a recurrent system can exhibit the phenomenon of equilibra-
tion. Consider a ring with N sites. On each site, place a ‘spin’ which can be in one of two states: up or down.
Along the N links of the system, F of them contain ‘flippers’. The configuration of the flippers is set at the outset
and never changes. The dynamics of the system are as follows: during each time step, every spin moves clockwise
a distance of one lattice spacing. Spins which pass through flippers reverse their orientation: up becomes down,
and down becomes up.
The ‘phase space’ for this system consists of 2N discrete configurations. Since each configuration maps onto a
unique image under the evolution of the system, phase space ‘volume’ is preserved. The evolution is invertible;
the inverse is obtained simply by rotating the spins counterclockwise. Figure 3.4 depicts an example configuration
for the system, and its first iteration under the dynamics.
Suppose the flippers were not fixed, but moved about randomly. In this case, we could focus on a single spin and
determine its configuration probabilistically. Let pn be the probability that a given spin is in the up configuration
at time n. The probability that it is up at time (n + 1) is then
pn+1 = (1 − x) pn + x (1 − pn ) , (3.61)
where x = F/N is the fraction of flippers in the system. In words: a spin will be up at time (n + 1) if it was up at
time n and did not pass through a flipper, or if it was down at time n and did pass through a flipper. If the flipper
locations are randomized at each time step, then the probability of flipping is simply x = F/N . Equation 3.61 can
be solved immediately:
pn = 21 + (1 − 2x)n (p0 − 12 ) , (3.62)
1
which decays exponentially to the equilibrium value of peq = 2 with time scale
1
τ (x) = − . (3.63)
ln |1 − 2x|
We identify τ (x) as the microscopic relaxation time over which local equilibrium is established. If we define the
magnetization m ≡ (N↑ − N↓ )/N , then m = 2p − 1, so mn = (1 − 2x)n m0 . The equilibrium magnetization is
meq = 0. Note that for 12 < x < 1 that the magnetization reverses sign each time step, as well as decreasing
exponentially in magnitude.
3.4. IRREVERSIBILITY AND POINCARÉ RECURRENCE 119
Figure 3.5: Three simulations of the Kac ring model with N = 2500 sites and three different concentrations of
flippers. The red line shows the magnetization as a function of time, starting from an initial configuration in which
100% of the spins are up. The blue line shows the prediction of the Stosszahlansatz, which yields an exponentially
decaying magnetization with time constant τ .
The assumption that leads to equation 3.61 is called the Stosszahlansatz8 , a long German word meaning, approx-
imately, ‘assumption on the counting of hits’. The resulting dynamics are irreversible: the magnetization inex-
orably decays to zero. However, the Kac ring model is purely deterministic, and the Stosszahlansatz can at best
be an approximation to the true dynamics. Clearly the Stosszahlansatz fails to account for correlations such as the
following: if spin i is flipped at time n, then spin i + 1 will have been flipped at time n − 1. Also if spin i is flipped
at time n, then it also will be flipped at time n + N . Indeed, since the dynamics of the Kac ring model are invertible
and volume preserving, it must exhibit Poincaré recurrence. We see this most vividly in figs. 3.5 and 3.6.
The model is trivial to simulate. The results of such a simulation are shown in figure 3.5 for a ring of N = 1000
sites, with F = 100 and F = 24 flippers. Note how the magnetization decays and fluctuates about the equilibrium
value meq = 0, but that after N iterations m recovers its initial value: mN = m0 . The recurrence time for this
system is simply N if F is even, and 2N if F is odd, since every spin will then have flipped an even number of
8 Unfortunately, many important physicists were German and we have to put up with a legacy of long German words like Gedankenexperi-
ment, Zitterbewegung, Brehmsstrahlung, Stosszahlansatz, Kartoffelsalat, etc.
120 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
Figure 3.6: Simulations of the Kac ring model. Top: N = 2500 sites with F = 201 flippers. After 2500 iterations,
each spin has flipped an odd number of times, so the recurrence time is 2N . Middle: N = 2500 with F = 2400,
resulting in a near-complete reversal of the population with every iteration. Bottom: N = 25000 with N = 1000,
showing long time equilibration and dramatic resurgence of the spin population.
times.
In figure 3.6 we plot two other simulations. The top panel shows what happens when x > 12 , so that the magneti-
zation wants to reverse its sign with every iteration. The bottom panel shows a simulation for a larger ring, with
N = 25000 sites. Note that the fluctuations in m about equilibrium are smaller than in the cases with N = 1000
sites. Why?
A mechanical system evolves according to Hamilton’s equations of motion. We have seen how such a system is
recurrent in the sense of Poincaré.
3.5. REMARKS ON ERGODIC THEORY 121
There is a level beyond recurrence called ergodicity. In an ergodic system, time averages over intervals [0, T ] with
T → ∞ may be replaced by phase space averages. The time average of a function f (ϕ) is defined as
ZT
1
f (ϕ) T = lim dt f ϕ(t) . (3.64)
T →∞ T
0
For a Hamiltonian system, the phase space average of the same function is defined by
Z Z
f (ϕ) S = dµ f (ϕ) δ E − H(ϕ) dµ δ E − H(ϕ) , (3.65)
where H(ϕ) = H(q, p) is the Hamiltonian, and where δ(x) is the Dirac δ-function. Thus,
ergodicity ⇐⇒ f (ϕ) T = f (ϕ) S , (3.66)
for all smooth functions f (ϕ) for which f (ϕ) S exists and is finite. Note that we do not average over all of phase
space. Rather, we average only over a hypersurface along which H(ϕ) = E is fixed, i.e. over one of the level
sets of the Hamiltonian function. This is because the dynamics preserves the energy. Ergodicity means that almost
all points ϕ will, upon Hamiltonian evolution, move in such a way as to eventually pass through every finite
neighborhood on the energy surface, and will spend equal time in equal regions of phase space.
Let χR (ϕ) be the characteristic function of a region R:
(
1 if ϕ ∈ R
χR (ϕ) = (3.67)
0 otherwise,
called the density of states, is the surface area of phase space at energy E, and
Z
DR (E) = dµ δ E − H(ϕ) . (3.71)
R
is the density of states for the phase space subset R. Note that
Z Z
dS
D(E) ≡ dµ δ E − H(ϕ) = (3.72)
|∇H|
SE
Z
d dΩ(E)
= dµ Θ E − H(ϕ) = . (3.73)
dE dE
122 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
Figure 3.7: Constant phase space velocity at an irrational angle over a toroidal phase space is ergodic, but not
mixing. A circle remains a circle, and a blob remains a blob.
Here, dS is the differential surface element, SE is the constant H hypersurface H(ϕ) = E, and Ω(E) is the volume
of phase space over which H(ϕ) < E. Note also that we may write
dµ = dE dΣE , (3.74)
where
dS
dΣE = (3.75)
|∇H| H(ϕ)=E
is the the invariant surface element.
The distribution,
δ E − H(ϕ) δ E − H(ϕ)
̺E (ϕ) = =R , (3.76)
D(E) dµ δ E − H(ϕ)
defines the microcanonical ensemble (µCE) of Gibbs.
We could also write Z
1
f (ϕ) S = dΣE f (ϕ) , (3.77)
D(E)
SE
integrating over the hypersurface SE rather than the entire phase space.
Figure 3.8: The baker’s transformation is a successive stretching, cutting, and restacking.
Now consider the average of some function f (q, p). We can write f (q, p) in terms of its Fourier transform,
X
f (q, p) = fˆmn e2πi(mq+np) . (3.81)
m,n
We have, then, X
f q(t), p(t) = fˆmn e2πi(mq0 +np0 ) e2πi(m+αn)t . (3.82)
m,n
The situation is depicted in fig. 3.7. If we start with the characteristic function of a disc,
̺(q, p, t = 0) = Θ a2 − (q − q0 )2 − (p − p0 )2 , (3.85)
then it remains the characteristic function of a disc:
̺(q, p, t) = Θ a2 − (q − q0 − t)2 − (p − p0 − αt)2 , (3.86)
A stronger condition one could impose is the following. Let A and B be subsets of SE . Define the measure
Z Z
D (E)
ν(A) = dΣE χA (ϕ) dΣE = A , (3.87)
D(E)
124 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
Figure 3.9: The multiply iterated baker’s transformation. The set A covers half the phase space and its area is
preserved under the map. Initially, the fraction of B covered by A is zero. After many iterations, the fraction of B
covered by g nA approaches 12 .
where χA (ϕ) is the characteristic function of A. The measure of a set A is the fraction of the energy surface SE
covered by A. This means ν(SE ) = 1, since SE is the entire phase space at energy E. Now let g be a volume-
preserving map on phase space. Given two measurable sets A and B, we say that a system is mixing if
mixing ⇐⇒ lim ν g nA ∩ B = ν(A) ν(B) . (3.88)
n→∞
In other words, the fraction of B covered by the nth iterate of A, i.e. g nA, is, as n → ∞, simply the fraction of SE
covered by A. The iterated map g n distorts the region A so severely that it eventually spreads out ‘evenly’ over
the entire energy hypersurface. Of course by ‘evenly’ we mean ‘with respect to any finite length scale’, because at
the very smallest scales, the phase space density is still locally constant as one evolves with the dynamics.
Mixing means that
Z
f (ϕ) = dµ ̺(ϕ, t) f (ϕ)
Z Z
−−−−−→ dµ f (ϕ) δ E − H(ϕ) dµ δ E − H(ϕ) (3.89)
t→∞
h i. h i
≡ Tr f (ϕ) δ E − H(ϕ) Tr δ E − H(ϕ) .
Physically, we can imagine regions of phase space being successively stretched and folded. During the stretching
process, the volume is preserved, so the successive stretch and fold operations map phase space back onto itself.
An example of a mixing system is the baker’s transformation, depicted in fig. 3.8. The baker map is defined by
1
2q , 2 p if 0 ≤ q < 21
g(q, p) = (3.90)
1 1
1
2q − 1 , 2 p + 2 if 2 ≤ q < 1 .
3.5. REMARKS ON ERGODIC THEORY 125
Figure 3.10: The Arnold cat map applied to an image of 150 × 150 pixels. After 300 iterations, the image repeats
itself. (Source: Wikipedia)
Note that g is invertible and volume-preserving. The baker’s transformation consists of an initial stretch in which q
is expanded by a factor of two and p is contracted by a factor of two, which preserves the total volume. The system
is then mapped back onto the original area by cutting and restacking, which we can call a ‘fold’. The inverse
transformation is accomplished by stretching first in the vertical (p) direction and squashing in the horizontal (q)
direction, followed by a slicing and restacking. Explicitly,
1
2 q , 2p if 0 ≤ p < 12
g −1 (q, p) = (3.91)
1 1
1
2 q + 2 , 2p − 1 if 2 ≤ p < 1 .
where [x] denotes the fractional part of x. One can write this in matrix form as
M
′ z }| {
q 1 1 q
= mod Z2 . (3.93)
p′ 1 2 p
The matrix M is very special because it has integer entries and its determinant is det M = 1. This means that the
inverse also has integer entries. The inverse transformation is then
M −1
z }| {
q 2 −1 q′
= mod Z2 . (3.94)
p −1 1 p′
Now for something cool. Suppose that our image consists of a set of discrete points located at (n1 /k , n2 /k),
where the denominator k ∈ Z is fixed, and where n1 and n2 range over the set {1, . . . , k}. Clearly g and its inverse
9 The cat map gets its name from its initial application, by Arnold, to the image of a cat’s face.
126 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
preserve this set, since the entries of M and M −1 are integers. If there are two possibilities for each pixel (say off
2
and on, or black and white), then there are 2(k ) possible images, and the cat map will map us invertibly from
one image to another. Therefore it must exhibit Poincaré recurrence! This phenomenon is demonstrated vividly
in fig. 3.10, which shows a k = 150 pixel (square) image of a cat subjected to the iterated cat map. The image is
stretched and folded with each successive application of the cat map, but after 300 iterations the image is restored!
How can this be if the cat map is mixing? The point is that only the discrete set of points (n1 /k , n2 /k) is periodic.
Points with different denominators will exhibit a different periodicity, and points with irrational coordinates will
in general never return to their exact initial conditions, although recurrence says they will come arbitrarily close,
given enough iterations. The baker’s transformation is also different in this respect, since the denominator of the
p coordinate is doubled upon each successive iteration.
The student should now contemplate the hierarchy of dynamical systems depicted in fig. 3.11, understanding the
characteristic features of each successive refinement10 .
Thermalization of quantum systems is fundamentally different from that of classical systems. Whereas time evolu-
tion in classical mechanics is in general a nonlinear dynamical system, the Schrödinger equation for time evolution
in quantum mechanics is linear:
∂Ψ
i~ = ĤΨ , (3.95)
∂t
where Ĥ is a many-body Hamiltonian. In classical mechanics, the thermal state is constructed by time evolution
– this is the content of the ergodic theorem. In quantum mechanics, as we shall see, the thermal distribution must
be encoded in the eigenstates themselves.
Let us assume an initial condition at t = 0,
X
|Ψ(0)i = Cα |Ψα i , (3.96)
α
10 There is something beyond mixing, called a K-system. A K-system has positive Kolmogorov-Sinai entropy. For such a system, closed
orbits separate exponentially in time, and consequently the Liouvillian L has a Lebesgue spectrum with denumerably infinite multiplicity.
3.6. THERMALIZATION OF QUANTUM SYSTEMS 127
where | Ψα i is an orthonormal
P eigenbasis for Ĥ satisfying Ĥ |Ψα i = Eα |Ψα iṪhe expansion coefficients satisfy
Cα = hΨα |Ψ(0)i and α |Cα |2 = 1. Normalization requires
X
h Ψ(0) | Ψ(0) i = |Cα |2 = 1 . (3.97)
α
Typically we assume that the distribution P (E) is narrowly peaked about hEi, such that (∆E)rms ≪ E − E0 ,
where E0 is the ground state energy. Note that P (E) = 0 for E < E0 , i.e. the eigenspectrum of Ĥ is bounded from
below.
Now consider a general quantum observable described by an operator A. We have
X
hA(t)i = h Ψ(t) | A | Ψ(t) i = Cα∗ Cβ ei(Eα −Eβ )t/~ Aαβ , (3.102)
α,β
ZT X
1
hAiT ≡ lim dt hA(t)i = |Cα |2 Aαα . (3.103)
T →∞ T
0 α
Note that this implies that all coherence between different eigenstates is lost in the long time limit, due to dephasing.
The essential ideas behind the eigenstate thermalization hypothesis (ETH) were described independently by J. Deutsch
(1991) and by M. Srednicki (1994). The argument goes as follows. If the total energy is the only conserved quan-
tity, and if A is a local, translationally-invariant, few-body operator, then the time average hAi is given by its
microcanonical value, P
X
2 α Aαα Θ(Eα ∈ I)
hAiT = |Cα | Aαα = P ≡ hAiE , (3.104)
α α Θ(Eα ∈ I)
128 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
where I = E, E + ∆E is an energy interval of width ∆E. So once again, time averages are micro canonical
averages.
But how is it that this is the case? The hypothesis of Deutsch and of Srednicki is that thermalization in isolated and
bounded quantum systems occurs at the level of individual eigenstates. That is, for all eigenstates |Ψα i with Eα ∈ I,
one has
Aαα = hAiEα . (3.105)
This means that thermal information is encoded in each eigenstate. This is called the eigenstate thermalization hypothesis
(ETH).
An equivalent version of the ETH is the following scenario. Suppose we have an infinite or extremely large
quantum system U (the ‘universe’) fixed in an eigenstate |Ψα i. Then form the projection operator Pα = |Ψα ihΨα |.
Projection operators satisfy P 2 = P and their eigenspectrum consists of one eigenvalue 1 and the rest of the
eigenvalues are zero11 . Now consider a partition of U = W ∪ S, where W ≫ S. We imagine S to be the ‘system’
and W the ‘world’. We can always decompose the state |Ψα i in a complete product basis for W and S, viz.
NW NS
X X
|Ψα i = Qα W S
pj |ψp i ⊗ |ψj i . (3.106)
p=1 j=1
Here NW/S is the size of the basis for W/S. The reduced density matrix for S is defined as
NS NW !
X X
ρS = Tr Pα = Qα
pj Qα∗
pj ′ |ψjS ihψjS′ | . (3.107)
W
j,j ′ =1 p=1
1 −β ĤS
ρS = e , (3.108)
ZS
where ĤS is the Hamiltonian restricted to S, and ZS = Tr e−β ĤS , so that Tr ρS = 1 and ρS is properly normalized.
Here β = 1/kB T with T the temperature. The temperature is fixed by the requirement that Tr (ρS ĤS ) = Eα ·
(VS /VU ), where the last factor is a ratio of volumes.
There is no rigorous proof of the ETH. Deutsch showed that the ETH holds for the case of an integrable Hamilto-
nian weakly perturbed by a single Gaussian random matrix. Horoi et al. (1995) showed that nuclear shell model
wavefunctions reproduce thermodynamic predictions. Recent numerical work by M. Rigol and collaborators has
verified the applicability of the ETH in small interacting boson systems. ETH fails for so-called integrable models,
where there are a large number of conserved quantities, which commute with the Hamiltonian. Integrable models
are, however, quite special, and as Deutsch showed, integrability is spoiled by weak perturbations, in which case
ETH then applies.
Note again that in contrast to the classical case, time evolution of a quantum state does not create the thermal state.
Rather, it reveals the thermal distribution which is encoded in all eigenstates after sufficient time for dephasing to
occur, so that correlations between all the wavefunction expansion coefficients {Cα } for α 6= α′ are all lost.
11 More generally, we could project onto a K-dimensional subspace, in which case there would be K eigenvalues of +1 and N − K eigen-
values of 0, where N is the dimension of the entire vector space.
3.7. APPENDIX I : FORMAL SOLUTION OF THE MASTER EQUATION 129
Recall the master equation Ṗi = −Γij Pj . The matrix Γij is real but not necessarily symmetric. For such a matrix,
β
the left eigenvectors φα
i and the right eigenvectors ψj are not the same: general different:
φα α
i Γij = λα φj
(3.109)
Γij ψjβ = λβ ψiβ .
Note that the eigenvalue equation for the right eigenvectors is Γ ψ = λψ while that for the left eigenvectors is
Γ t φ = λφ. The characteristic polynomial is the same in both cases:
We have seen that φ ~ = (1, 1, . . . , 1) is a left eigenvector with eigenvalue λ = 0, since P Γ = 0. We do not know
i ij
a priori the corresponding right eigenvector, which depends on other details of Γij . Now let’s expand Pi (t) in the
right eigenvectors of Γ , writing X
Pi (t) = Cα (t) ψiα . (3.115)
α
Then
dPi X dC
α α
= ψi
dt α
dt
X
= −Γij Pj = − Cα Γij ψjα (3.116)
α
X
=− λα Cα ψiα .
α
It is now easy to see that Re (λα ) ≥ 0 for all λ, or else the probabilities will become negative. For suppose
Re (λα ) < 0 for some α. Then as t → ∞, the sum in eqn. 3.118 will be dominated by the term P for which λα has
the largest negative real part; all other contributions will be subleading. But we must have i ψiα = 0 since ψ α
must be orthogonal to the left eigenvector φ ~α=0 = (1, 1, . . . , 1). Therefore, at least one component of ψ α (i.e. for
i
some value of i) must have a negative real part, which means a negative probability!12 As we have already proven
that an initial nonnegative distribution {Pi (t = 0)} will remain nonnegative under the evolution of the master
equation, we conclude that Pi (t) → Pieq as t → ∞, relaxing to the λ = 0 right eigenvector, with Re (λα ) ≥ 0 for all
α.
Consider a group of atoms, some of which are in an excited state which can undergo nuclear decay. Let Pn (t) be
the probability that n atoms are excited at some time t. We then model the decay dynamics by
0 if m ≥ n
Wmn = nγ if m = n − 1 (3.119)
0 if m < n − 1 .
Here, γ is the decay rate of an individual atom, which can be determined from quantum mechanics. The master
equation then tells us
dPn
= (n + 1) γ Pn+1 − n γ Pn . (3.120)
dt
2
The interpretation here is as follows: let n denote a state in which n atoms are excited. Then Pn (t) = h ψ(t) | n i .
Then Pn (t) will increase due to spontaneous transitions from | n+1 i to | n i, and will decrease due to spontaneous
transitions from | n i to | n−1 i.
The average number of particles in the system is
∞
X
N (t) = n Pn (t) . (3.121)
n=0
Note that
dN X∞ h i
= n (n + 1) γ Pn+1 − n γ Pn
dt n=0
∞ h
X i
=γ n(n − 1) Pn − n2 Pn (3.122)
n=0
X∞
= −γ n Pn = −γ N .
n=0
Thus,
N (t) = N (0) e−γt . (3.123)
−1
The relaxation time is τ = γ , and the equilibrium distribution is
∂P X∞ h i
=γ z n (n + 1) Pn+1 − n Pn
∂t n=0 (3.126)
∂P ∂P
=γ − γz .
∂z ∂z
Thus,
1 ∂P ∂P
− (1 − z) =0. (3.127)
γ ∂t ∂z
We now see that any function f (ξ) satisfies the above equation, where ξ = γt − ln(1 − z). Thus, we can write
P (z, t) = f γt − ln(1 − z) . (3.128)
Setting t = 0 we have P (z, 0) = f −ln(1 − z) , and inverting this result we obtain f (u) = P (1 − e−u , 0), i.e.
P (z, t) = P 1 + (z − 1) e−γt , 0 . (3.129)
P∞
The total probability is P (z = 1, t) = n=0 Pn , which clearly is conserved: P (1, t) = P (1, 0). The average particle
number is
∞
X ∂P
N (t) = n Pn (t) = = e−γt P (1, 0) = N (0) e−γt . (3.130)
n=0
∂z z=1
The Euler-Lagrange equations of motion of classical mechanics are invariant under a redefinition of generalized
coordinates,
Qσ = Qσ (q1 , . . . , qr , t) , (3.131)
called a point transformation. That is, if we express the new Lagrangian in terms of the new coordinates and their
time derivatives, viz.
L̃ Q, Q̇, t) = L q(Q, t) , q̇(Q, Q̇, t) , t , (3.132)
then the equations of motion remain of the form
∂ L̃ d ∂ L̃
= . (3.133)
∂Qσ dt ∂ Q̇σ
Hamilton’s equations13 ,
∂H ∂H
q̇σ = , ṗσ = − (3.134)
∂pσ ∂qσ
13 We revert to using H for the Hamiltonian in this section, rather than Ĥ as before.
132 CHAPTER 3. ERGODICITY AND THE APPROACH TO EQUILIBRIUM
are invariant under a much broader class of transformations which mix all the q ′ s and p′ s, called canonical trans-
formations. The general form for a canonical transformation is
qσ = qσ Q1 , . . . , Qr , P1 , . . . , Pr , t (3.135)
pσ = pσ Q1 , . . . , Qr , P1 , . . . , Pr , t , (3.136)
∂ H̃ ∂ H̃
Q̇σ = , Ṗσ = − , (3.139)
∂Pσ ∂Qσ
which gives
∂ Q̇σ ∂ Ṗσ ∂ Ξ̇i
+ =0= . (3.140)
∂Qσ ∂Pσ ∂Ξi
I.e. the flow remains incompressible in the new (Q, P ) variables. We will also require that phase space volumes
are preserved by the transformation, i.e.
∂Ξi ∂(Q, P )
det = = 1 . (3.141)
∂ξj ∂(q, p)
This last condition guarantees the invariance of the phase space measure
r
Y
dµ = h−r dqσ dpσ , (3.142)
σ=1
Statistical Ensembles
4.1 References
– F. Reif, Fundamentals of Statistical and Thermal Physics (McGraw-Hill, 1987)
This has been perhaps the most popular undergraduate text since it first appeared in 1967, and with good
reason.
– A. H. Carter, Classical and Statistical Thermodynamics
(Benjamin Cummings, 2000)
A very relaxed treatment appropriate for undergraduate physics majors.
– D. V. Schroeder, An Introduction to Thermal Physics (Addison-Wesley, 2000)
This is the best undergraduate thermodynamics book I’ve come across, but only 40% of the book treats
statistical mechanics.
– C. Kittel, Elementary Statistical Physics (Dover, 2004)
Remarkably crisp, though dated, this text is organized as a series of brief discussions of key concepts and
examples. Published by Dover, so you can’t beat the price.
– M. Kardar, Statistical Physics of Particles (Cambridge, 2007)
A superb modern text, with many insightful presentations of key concepts.
– M. Plischke and B. Bergersen, Equilibrium Statistical Physics (3rd edition, World Scientific, 2006)
An excellent graduate level text. Less insightful than Kardar but still a good modern treatment of the subject.
Good discussion of mean field theory.
– E. M. Lifshitz and L. P. Pitaevskii, Statistical Physics (part I, 3rd edition, Pergamon, 1980)
This is volume 5 in the famous Landau and Lifshitz Course of Theoretical Physics . Though dated, it still
contains a wealth of information and physical insight.
133
134 CHAPTER 4. STATISTICAL ENSEMBLES
We have seen how in an ergodic dynamical system, time averages can be replaced by phase space averages:
ergodicity ⇐⇒ f (ϕ) T = f (ϕ) S , (4.1)
where
ZT
1
f (ϕ) T = lim dt f ϕ(t) . (4.2)
T →∞ T
0
and Z
Z
f (ϕ) S = dµ f (ϕ) δ E − Ĥ(ϕ) dµ δ E − Ĥ(ϕ) . (4.3)
Here Ĥ(ϕ) = Ĥ(q, p) is the Hamiltonian, and where δ(x) is the Dirac δ-function1 . Thus, averages are taken over
a constant energy hypersurface which is a subset of the entire phase space.
We’ve also seen how any phase space distribution ̺(Λ1 , . . . , Λk ) which is a function of conserved quantitied Λa (ϕ)
is automatically a stationary (time-independent) solution to Liouville’s equation. Note that the microcanonical
distribution,
Z
̺E (ϕ) = δ E − Ĥ(ϕ) dµ δ E − Ĥ(ϕ) , (4.4)
is of this form, since Ĥ(ϕ) is conserved by the dynamics. Linear and angular momentum conservation generally
are broken by elastic scattering off the walls of the sample.
So averages in the microcanonical ensemble are computed by evaluating the ratio
Tr A δ(E − Ĥ)
A = , (4.5)
Tr δ(E − Ĥ)
where Tr means ‘trace’, which entails an integration over all phase space:
N Z
1 Y ddpi ddqi
Tr A(q, p) ≡ A(q, p) . (4.6)
N ! i=1 (2π~)d
Here N is the total number of particles and d is the dimension of physical space in which each particle moves.
The factor of 1/N !, which cancels in the ratio between numerator and denominator, is present for indistinguishable
particles. The normalization factor (2π~)−N d renders the trace dimensionless. Again, this cancels between numer-
ator and denominator. These factors may then seem arbitrary in the definition of the trace, but we’ll see how they
in fact are required from quantum mechanical considerations. So we now adopt the following metric for classical
phase space integration:
N
1 Y ddpi ddqi
dµ = . (4.7)
N ! i=1 (2π~)d
1 We write the Hamiltonian as Ĥ (classical or quantum) in order to distinguish it from magnetic field (H) or enthalpy (H).
4.2. MICROCANONICAL ENSEMBLE (µCE) 135
The denominator,
D(E) = Tr δ(E − Ĥ) , (4.8)
is called the density of states. It has dimensions of inverse energy, such that
Z
E+∆E Z Z
D(E) ∆E = dE ′ dµ δ(E ′ − Ĥ) = dµ (4.9)
E E<Ĥ<E+∆E
Let us now compute D(E) for the nonrelativistic ideal gas. The Hamiltonian is
XN
p2i
Ĥ(q, p) = . (4.10)
i=1
2m
We assume that the gas is enclosed in a region of volume V , and we’ll do a purely classical calculation, neglecting
discreteness of its quantum spectrum. We must compute
Z YN XN
1 ddpi ddqi p2i
D(E) = δ E − . (4.11)
N! i=1
(2π~)d i=1
2m
√
We’ll do this calculation in two ways. First, let’s rescale pα
i ≡ 2mE uα
i . We then have
√ !N d Z
VN 2mE 1
D(E) = dMu δ u21 + u22 + . . . + u2M − 1 . (4.12)
N! h E
Here we have written u = (u1 , u2 , . . . , uM ) with M = N d as a M -dimensional vector. We’ve also used the rule
δ(Ex) = E −1 δ(x) for δ-functions. We can now write
where dΩM is the M -dimensional differential solid angle. We now have our answer:2
√ !N d
VN 2m 1
N d−1 1
D(E) = E2 · 2 ΩN d . (4.14)
N! h
What remains is for us to compute ΩM , the total solid angle in M dimensions. We do this by a nifty mathematical
trick. Consider the integral
Z Z∞
−u2 2
IM = M
d ue = ΩM du uM−1 e−u
0
(4.15)
Z∞
1
1
M−1 −s
= 2 ΩM ds s 2 e = 21 ΩM Γ 1
2M ,
0
1 1 1
2 Thefactor of 2
preceding ΩM in eqn. 4.14 appears because δ(u2 − 1) = 2
δ(u − 1) + 2
δ(u + 1). Since u = |u| ≥ 0, the second term can
be dropped.
136 CHAPTER 4. STATISTICAL ENSEMBLES
is the Gamma function, which satisfies z Γ(z) = Γ(z + 1).3 On the other hand, we can compute IM in Cartesian
coordinates, writing
∞ M
Z
2 √ M
IM = du1 e−u1 = π . (4.17)
−∞
Therefore
2π M/2
ΩM = . (4.18)
Γ(M/2)
We thereby obtain Ω2 = 2π, Ω3 = 4π, Ω4 = 2π 2 , etc., the first two of which are familiar.
Our final result, then, is
N d/2 1
N d−1
VN m E2
D(E, V, N ) = . (4.19)
N! 2π~2 Γ(N d/2)
Here we have emphasized that the density of states is a function of E, V , and N . Using Stirling’s approximation,
ln N ! = N ln N − N + 1
2 ln N + 1
2 ln(2π) + O N −1 , (4.20)
where
E V d E V d m
φ , = ln + ln + ln 2
+ 1 + 21 d . (4.22)
N N 2 N N 2 dπ~
Z∞
Z(β) = L D(E) ≡ dE e−βE D(E) = Tr e−β Ĥ . (4.23)
0
Z
c+i∞
dβ βE
D(E) = L−1 Z(β) ≡ e Z(β) , (4.24)
2πi
c−i∞
where c is such that the integration contour is to the right of any singularities of Z(β) in the complex β-plane. We
Figure 4.1: Complex integration contours C for inverse Laplace transform L−1 Z(β) = D(E). When the product
dN is odd, there is a branch cut along the negative Re β axis.
then have
N Z
1 Y ddxi ddpi −β p2i /2m
Z(β) = e
N ! i=1 (2π~)d
∞ N d
Z
VN dp −βp2 /2m (4.25)
= e
N! 2π~
−∞
N
N d/2
V m
= β −N d/2 .
N! 2π~2
exactly as before. The integration contour for the inverse Laplace transform is extended in an infinite semicircle in
the left half β-plane. When N d is even, the function β −N d/2 has a simple pole of order N d/2 at the origin. When
N d is odd, there is a branch cut extending along the negative Re β axis, and the integration contour must avoid
the cut, as shown in fig. 4.1.
For a general system, the Laplace transform, Z(β) = L D(E) also is called the partition function. We shall again
meet up with Z(β) when we discuss the ordinary canonical ensemble.
Note that D(E) has dimensions of inverse energy, so one might ask how we are to take the logarithm of a di-
mensionful quantity in eqn. 4.21. We must introduce an energy scale, such as ∆E in eqn. 4.9, and define
138 CHAPTER 4. STATISTICAL ENSEMBLES
D̃(E; ∆E) = D(E) ∆E and S(E; ∆E) ≡ kB ln D̃(E; ∆E). The definition of statistical entropy then involves the
arbitrary parameter ∆E, however this only affects S(E) in an additive way. That is,
∆E1
S(E, V, N ; ∆E1 ) = S(E, V, N ; ∆E2 ) + kB ln . (4.27)
∆E2
Note that the difference between the two definitions of S depends only on the ratio ∆E1 /∆E2 , and is independent
of E, V , and N .
Consider an ultrarelativistic ideal gas, with single particle dispersion ε(p) = cp. We then have
∞ N
Z
V N ΩdN
Z(β) = dp pd−1 e−βcp
N ! hN d
0 (4.28)
N
N
V Γ(d) Ωd
= .
N! cd h d β d
E V
The statistical entropy is S(E, V, N ) = kB ln D(E, V, N ) = N kB φ N , N , with
E V E V Ωd Γ(d)
φ , = d ln + ln + ln + (d + 1) (4.29)
N N N N (dhc)d
For classical systems where the energy levels are discrete, the states of the system | σ i are labeled by a set of
discrete quantities {σ1 , σ2 , . . .}, where each variable σi takes discrete values. The number of ways of configuring
the system at fixed energy E is then X
Ω(E, N ) = δĤ(σ),E , (4.30)
σ
where the sum is over all possible configurations. Here N labels the total number of particles. For example, if
we have N spin- 12 particles on a lattice which are placed in a magnetic field H, so the individual particle energy
is εi = −µ0 Hσ, where σ = ±1, then in a configuration in which N↑ particles have σi = +1 and N↓ = N − N↑
particles have σi = −1, the energy is E = (N↓ − N↑ )µ0 H. The number of configurations at fixed energy E is
N N!
Ω(E, N ) = = N E
N E
, (4.31)
N↑ 2 − 2µ0 H ! 2 + 2µ0 H !
N E
since N↑/↓ = 2 ∓ 2µ0 H . The statistical entropy is S(E, N ) = kB ln Ω(E, N ).
Thus far our understanding of ergodicity is rooted in the dynamics of classical mechanics. A Hamiltonian flow
which is ergodic is one in which time averages can be replaced by phase space averages using the microcanonical
ensemble. What happens, though, if our system is quantum mechanical, as all systems ultimately are?
4.3. THE QUANTUM MECHANICAL TRACE 139
S
First, let us consider that our system will in general be in contact with a world W . We call
the union of S and
W the universe, U = W ∪ S. Let N denote a quantum mechanical state of W , and let n denote a quantum
mechanical state of S. Then the most general wavefunction we can write is of the form
X
Ψ = Ψ N ⊗ n . (4.32)
N,n
N,n
Now let us compute the expectation value of some operator  which acts as the identity within W , meaning
N  N ′ =  δN N ′ , where  is the ‘reduced’ operator which acts within S alone. We then have
X X
Ψ Â Ψ = Ψ∗N,n ΨN ′ ,n′ δN N ′ n  n′
N,N ′ n,n′ (4.33)
= Tr ̺ˆ Â ,
where XX
̺ˆ = Ψ∗N,n ΨN,n′ n′ n (4.34)
N n,n′
∂ ̺ˆ
i~ = Ĥ, ̺ˆ . (4.36)
∂t
Note that the density matrix evolves according to a slightly different equation than an operator in the Heisenberg
picture, for which
∂ Â
Â(t) = e+iHt/~ A e−iĤt/~ =⇒ i~ = Â, Ĥ = − Ĥ, Â . (4.37)
∂t
For Hamiltonian systems, we found that the phase space distribution ̺(q, p, t) evolved according to the Liouville
equation,
∂̺
i = L̺ , (4.38)
∂t
where the Liouvillian L is the differential operator
Nd
( )
X ∂ Ĥ ∂ ∂ Ĥ ∂
L = −i − . (4.39)
j=1
∂pj ∂qj ∂qj ∂pj
Accordingly, any distribution ̺(Λ1 , . . . , Λk ) which is a function of constants of the motion Λa (q, p) is a station-
ary solution to the Liouville equation: ∂t ̺(Λ1 , . . . , Λk ) = 0. Similarly, any quantum mechanical density matrix
which commutes with the Hamiltonian is a stationary solution to eqn. 4.36. The corresponding microcanonical
distribution is
̺ˆE = δ E − Ĥ . (4.40)
140 CHAPTER 4. STATISTICAL ENSEMBLES
Figure 4.2: A system S in contact with a ‘world’ W . The union of the two, universe U = W ∪ S, is said to be the
‘universe’.
If our quantum mechanical system is placed in a finite volume, the energy levels will be discrete, rather than
continuous, and the density of states (DOS) will be of the form
X
D(E) = Tr δ E − Ĥ = δ(E − El ) , (4.41)
l
where {El } are the eigenvalues of the Hamiltonian Ĥ. In the thermodynamic limit, V → ∞, and the discrete
spectrum of kinetic energies remains discrete for all finite V but must approach the continuum result. To recover
the continuum result, we average the DOS over a window of width ∆E:
Z
E+∆E
1
D(E) = dE ′ D(E ′ ) . (4.42)
∆E
E
If we take the limit ∆E → 0 but with ∆E ≫ δE, where δE is the spacing between successive quantized levels, we
recover a smooth function, as shown in fig. 4.3. We will in general drop the bar and refer to this function as D(E).
Note that δE ∼ 1/D(E) = e−N φ(ε,v) is (typically) exponentially small in the size of the system, hence if we took
∆E ∝ V −1 which vanishes in the thermodynamic limit, there are still exponentially many energy levels within an
interval of width ∆E.
The quantum-classical correspondence is elucidated with the use of coherent states. Recall that the one-dimensional
harmonic oscillator Hamiltonian may be written
p2
Ĥ0 = + 1 m ω02 q 2
2m 2 (4.43)
= ~ω0 a† a + 12 ,
where a and a† are ladder operators satisfying a, a† = 1, which can be taken to be
∂ q ∂ q
a=ℓ + , a† = −ℓ + , (4.44)
∂q 2ℓ ∂q 2ℓ
4.3. THE QUANTUM MECHANICAL TRACE 141
Figure 4.3: Averaging the quantum mechanical discrete density of states yields a continuous curve.
p
with ℓ = ~/2mω0 . Note that
~
q = ℓ a + a† , p= a − a† . (4.45)
2iℓ
hence different coherent states are not orthogonal. Despite this nonorthogonality, the coherent states allow a
simple resolution of the identity,
Z 2
dz d2z d Rez d Imz
1= | z ih z | ; ≡ (4.49)
2πi 2πi π
which is straightforward to establish.
To gain some physical intuition about the coherent states, define
Q iℓP
z≡ + (4.50)
2ℓ ~
and write | z i ≡ | Q, P i. One finds (exercise!)
2
/4ℓ2
ψQ,P (q) = h q | z i = (2πℓ2 )−1/4 e−iP Q/2~ eiP q/~ e−(q−Q) , (4.51)
hence the coherent state ψQ,P (q) is a wavepacket Gaussianly localized about q = Q, but oscillating with average
momentum P .
142 CHAPTER 4. STATISTICAL ENSEMBLES
and ∆q · ∆p = 21 ~. Thus we learn that the coherent state ψQ,P (q) is localized in phase space, i.e. in both position
and momentum. If we have a general operator Â(q, p), we can then write
Q, P Â(q, p) Q, P = A(Q, P ) + O(~) , (4.57)
Z∞ Z∞
1
Tr  = dQ dP Q, P  Q, P . (4.59)
2π~
−∞ −∞
We now can understand the origin of the factor 2π~ in the denominator of each (qi , pi ) integral over classical phase
space in eqn. 4.6.
Note that ω0 is arbitrary in our discussion. By increasing ω0 , the states become more localized in q and more plane
wave like in p. However, so long as ω0 is finite, the width of the coherent state in each direction is proportional to
~1/2 , and thus vanishes in the classical limit.
Consider two systems in thermal contact, as depicted in fig. 4.4. The two subsystems #1 and #2 are free to exchange
energy, but their respective volumes and particle numbers remain fixed. We assume the contact is made over a
surface, and that the energy associated with that surface is negligible when compared with the bulk energies E1
and E2 . Let the total energy be E = E1 + E2 . Then the density of states D(E) for the combined system is
Z
D(E) = dE1 D1 (E1 ) D2 (E − E1 ) . (4.60)
4.4. THERMAL EQUILIBRIUM 143
a result familiar from thermodynamics. The difference is now we have a more rigorous definition of the entropy.
When the total entropy S is maximized, we have that T1 = T2 . Once again, two systems in thermal contact and
can exchange energy will in equilibrium have equal temperatures.
According to eqns. 4.22 and 4.29, the entropies of nonrelativistic and ultrarelativistic ideal gases in d space dimen-
sions are given by
1 E V
SNR = 2 N d kB ln + N kB ln + const. (4.66)
N N
E V
SUR = N d kB ln + N kB ln + const. . (4.67)
N N
Invoking eqn. 4.65, we then have
We saw that the probability distribution P1 (E1 ) is maximized when T1 = T2 , but how sharp is the peak in the
distribution? Let us write E1 = E1∗ + ∆E1 , where E1∗ is the solution to eqn. 4.62. We then have
1 ∂ 2S1 1 ∂ 2S2
ln P1 (E1∗ + ∆E1 ) = ln P1 (E1∗ ) + (∆E )2
+ (∆E1 )2 + . . . , (4.69)
2kB ∂E12 E ∗ 1
2kB ∂E22 E ∗
1 2
where
CV,1 CV,2
C̄V = . (4.72)
CV,1 + CV,2
The distribution is therefore a Gaussian, and the fluctuations in ∆E1 can now be computed:
q
(∆E1 )2 = kB T 2 C̄V =⇒ (∆E1 )RMS = kB T C̄V /kB . (4.73)
The individual heat capacities CV,1 and CV,2 scale with the volumes V1 and V2 , respectively. If V2 ≫ V1 , then
CV,2 ≫ CV,1 , in which case C̄V ≈ CV,1 . Therefore the RMS fluctuations in ∆E1 are proportional to the square
root of the system size, whereas E1 itself is extensive. Thus, the ratio (∆E1 )RMS /E1 ∝ V −1/2 scales as the inverse
square root of the volume. The distribution P1 (E1 ) is thus extremely sharp.
Consider a system S in contact with a world W , and let their union U = W ∪ S be called the ‘universe’. The
situation is depicted in fig. 4.2. The volume VS and particle number NS of the system are held fixed, but the
energy is allowed to fluctuate by exchange with the world W . We are interested in the limit NS → ∞, NW → ∞,
with NS ≪ NW , with similar relations holding for the respective volumes and energies. We now ask what is the
probability that S is in a state | n i with energy En . This is given by the ratio
DW (EU − En ) ∆E
Pn = lim
∆E→0 DU (EU ) ∆E
(4.74)
# of states accessible to W given that ES = En
= .
total # of states in U
Then
ln Pn = ln DW (EU − En ) − ln DU (EU )
∂ ln DW (E)
= ln DW (EU ) − ln DU (EU ) − En + ... (4.75)
∂E E=E U
≡ −α − βEn .
4.5. ORDINARY CANONICAL ENSEMBLE (OCE) 145
P
Thus, we find Pn = e−α e−βEn . The constant α is fixed by the requirement that n Pn = 1:
1 −βEn X
Pn = e , Z(T, V, N ) = e−βEn = Tr e−β Ĥ . (4.77)
Z n
We’ve already met Z(β) in eqn. 4.23 – it is the Laplace transform of the density of states. It is also called the
partition function of the system S. Quantum mechanically, we can write the ordinary canonical density matrix as
e−β Ĥ
̺ˆ = . (4.78)
Tr e−β Ĥ
Note that ̺ˆ, Ĥ = 0, hence the ordinary canonical distribution is a stationary solution to the evolution equation
for the density matrix. Note that the OCE is specified by three parameters: T , V , and N .
Let the total energy of the Universe be fixed at EU . The joint probability density P (ES , EW ) for the system to have
energy ES and the world to have energy EW is
P (ES , EW ) = DS (ES ) DW (EW ) δ(EU − ES − EW ) DU (EU ) , (4.79)
where
Z∞
DU (EU ) = dES DS (ES ) DW (EU − ES ) , (4.80)
−∞
R R
which ensures that dES dEW P (ES , EW ) = 1. The probability density P (ES ) is defined such that P (ES ) dES is
the (differential) probability for the system to have an energy in the range [ES , ES + dES ]. The units of P (ES ) are
E −1 . To obtain P (ES ), we simply integrate the joint probability density P (ES , EW ) over all possible values of EW ,
obtaining
D (E ) D (E − ES )
P (ES ) = S S W U , (4.81)
DU (EU )
as we have in eqn. 4.74.
Now suppose we wish to know the probability Pn that the system is in a particular state | n i with energy En .
Clearly
probability that ES ∈ [En , En + ∆E] P (En ) ∆E D (E − En )
Pn = lim = = W U . (4.82)
∆E→0 # of S states with ES ∈ [En , En + ∆E] DS (En ) ∆E DU (EU )
where we have conveniently taken the trace in a basis of energy eigenstates. In the classical limit, we have
Z
1 −β Ĥ(ϕ)
̺(ϕ) = e , Z = Tr e −β Ĥ
= dµ e−β Ĥ(ϕ) , (4.84)
Z
QN
with dµ = N1 ! j=1 (dd qj dd pj /hd ) for identical particles (‘Maxwell-Boltzmann statistics’). Thus,
R
dµ A(ϕ) e−β Ĥ(ϕ)
hAi = Tr (̺A) = R . (4.85)
dµ e−β Ĥ(ϕ)
The Boltzmann entropy and the statistical entropy S = kB ln D(E) are identical in the thermodynamic limit.
We define the Helmholtz free energy F (T, V, N ) as
hence
Pn = eβF e−βEn , ln Pn = βF − βEn . (4.88)
Therefore the entropy is
X
S = −kB Pn βF − βEn
n
(4.89)
F h Ĥ i
=− + ,
T T
which is to say
F = E − TS , (4.90)
where
X Tr Ĥ e−β Ĥ
E= Pn En = (4.91)
n Tr e−β Ĥ
is the average energy. We also see that
X P −βEn
−β Ĥ −βEn n En e ∂ ∂
Z = Tr e = e =⇒ E= P −βE
=− ln Z = βF . (4.92)
ne ∂β ∂β
n
n
dF = −S dT − p dV + µ dN , (4.93)
which means
∂F ∂F ∂F
S=− , p=− , µ=+ . (4.94)
∂T V,N ∂V T,N ∂N T,V
4.5. ORDINARY CANONICAL ENSEMBLE (OCE) 147
In the OCE, the energy is not fixed. It therefore fluctuates about its average value E = hĤi. Note that
∂E ∂E ∂ 2 ln Z
− = kB T 2 =
∂β ∂T ∂β 2
!2
Tr Ĥ 2 e−β Ĥ Tr Ĥ e−β Ĥ (4.95)
= −
Tr e−β Ĥ Tr e−β Ĥ
2
= Ĥ 2 − Ĥ .
Thus, the heat capacity is related to the fluctuations in the energy, just as we saw at the end of §4.4:
∂E 1
2
2
CV = = Ĥ − Ĥ (4.96)
∂T V,N kB T 2
For the nonrelativistic ideal gas, we found CV = d2 N kB , hence the ratio of RMS fluctuations in the energy to the
energy itself is q
p r
(∆Ĥ)2 k T 2 CV 2
= dB = , (4.97)
hĤi 2 N kB T N d
and the ratio of the RMS fluctuations to the mean value vanishes in the thermodynamic limit.
The full distribution function for the energy is
Tr δ(E − Ĥ) e−β Ĥ 1
P (E) = δ(E − Ĥ) = = D(E) e−βE . (4.98)
Tr e−β Ĥ Z
Thus,
e−β[E−T S(E)]
P (E) = R , (4.99)
dẼ e−β [Ẽ−T S(Ẽ)]
where S(E) = kB ln D(E) is the statistical entropy. Let’s write E = E + δE, where E extremizes the combination
E − T S(E), i.e. the solution to T S ′ (E) = 1, where the energy derivative of S is performed at fixed volume V and
particle number N . We now expand S(E + δE) to second order in δE, obtaining
2
δE δE
S(E + δE) = S(E) + − + ... (4.100)
T 2T 2 CV
∂ 1
1
Recall that S ′′ (E) = ∂E T = − T 2 C . Thus,
V
(δE)2
E − T S(E) = E − T S(E) + + O (δE)2 . (4.101)
2T CV
Applying this to both numerator and denominator of eqn. 4.99, we obtain4
" #
(δE)2
P (E) = N exp − , (4.102)
2kB T 2 CV
R
where N = (2πkB T 2 CV )−1/2 is a normalization constant which guarantees dE P (E) = 1. Once again, we see that
p
the distribution is a Gaussian centered at hEi = E, and of width (∆E)RMS = kB T 2 CV . This is a consequence of
the Central Limit Theorem.
4 In
applying eqn. 4.101 to the denominator of eqn. 4.99, we shift Ẽ by E and integrate over the difference δẼ ≡ Ẽ − E, retaining terms up
to quadratic order in δE˜ in the argument of the exponent.
148 CHAPTER 4. STATISTICAL ENSEMBLES
and therefore
X X
dE = En dPn + Pn dEn
n n (4.104)
= d̄Q − d̄W ,
where
X
d̄W = − Pn dEn (4.105)
n
X
d̄Q = En dPn . (4.106)
n
En = −kB T ln Z − kB T ln Pn , (4.107)
4.5. ORDINARY CANONICAL ENSEMBLE (OCE) 149
This is the force acting on the system5 . In the chapter on thermodynamics, we defined the generalized force
conjugate to Xi as yi ≡ −Fi .
Thus we see from eqn. 4.104 that there are two ways that the average energy can change; these are depicted in
the sketch of fig. 4.5. Starting from a set of energy levels {En } and probabilities {Pn }, we can shift the energies
to {En′ }. The resulting change in energy (∆E)I = −W is identified with the work done on the system. We could
also modify the probabilities to {Pn′ } without changing the energies. The energy change in this case is the heat
absorbed by the system: (∆E)II = Q. This provides us with a statistical and microscopic interpretation of the First
Law of Thermodynamics.
and
1 ∂Z 1
= β · Tr Q̂ e−β Ĥ(λ) = β hQ̂i . (4.113)
Z ∂λ Z
But then from Z = e−βF we have
∂F
Q(λ, T ) = h Q̂ i = − . (4.114)
∂λ T
5 Inderiving eqn. 4.110, we have used the so-called Feynman-Hellman theorem of quantum mechanics: dhn|Ĥ|ni = hn| dĤ |ni, if |ni is an
energy eigenstate.
150 CHAPTER 4. STATISTICAL ENSEMBLES
Typically we will take Q to be an extensive quantity. We can now define the susceptibility χ as
1 ∂Q 1 ∂ 2F
χ= =− . (4.115)
V ∂λ V ∂λ2
The volume factor in the denominator ensures that χ is intensive.
It is important to realize that we have assumed here that Ĥ0 , Q̂ = 0, i.e. the ‘bare’ Hamiltonian Ĥ0 and the
operator Q̂ commute. If they do not commute, then the response functions must be computed within a proper
quantum mechanical formalism, which we shall not discuss here.
Note also that we can imagine an entire family of observables Q̂k satisfying Q̂k , Q̂k′ = 0 and Ĥ0 , Q̂k = 0,
for all k and k ′ . Then for the Hamiltonian
X
Ĥ (~λ) = Ĥ0 − λk Q̂k , (4.116)
k
we have that
∂F
Qk (~λ, T ) = h Q̂k i = − (4.117)
∂λk T, Na , λk′ 6=k
1 ∂Qk 1 ∂ 2F
χkl = =− . (4.118)
V ∂λl V ∂λk ∂λl
Consider once again the situation depicted in fig. 4.2, where a system S is in contact with a world W , their union
U = W ∪ S being called the ‘universe’. We assume that the system’s volume VS is fixed, but otherwise it is allowed
to exchange energy and particle number with W . Hence, the system’s energy ES and particle number NS will
fluctuate. We ask what is the probability that S is in a state | n i with energy En and particle number Nn . This is
given by the ratio
DW (EU − En , NU − Nn ) ∆E
Pn = lim
∆E→0 DU (EU , NU ) ∆E
(4.119)
# of states accessible to W given that ES = En and NS = Nn
= .
total # of states in U
Then
ln Pn = ln DW (EU − En , NU − Nn ) − ln DU (EU , NU )
= ln DW (EU , NU ) − ln DU (EU , NU )
(4.120)
∂ ln DW (E, N ) ∂ ln DW (E, N )
− En E=EU − Nn E=EU + . . .
∂E ∂N
N =N N =N
U U
≡ −α − βEn + βµNn .
4.6. GRAND CANONICAL ENSEMBLE (GCE) 151
The quantity µ has dimensions of energy and is called the chemical potential. Nota bene: Some texts define the
‘grand canonical Hamiltonian’ K̂ as
K̂ ≡ Ĥ − µN̂ . (4.123)
−α −β(En −µNn )
P
Thus, Pn = e e . Once again, the constant α is fixed by the requirement that n Pn = 1:
1 −β(En−µNn ) X
Pn = e , Ξ(β, V, µ) = e−β(En−µNn ) = Tr e−β(Ĥ−µN̂ ) = Tr e−β K̂ . (4.124)
Ξ n
e−β K̂
̺ˆ = . (4.125)
Tr e−β K̂
Note that ̺ˆ, K̂ = 0.
The quantity Ξ(T, V, µ) is called the grand partition function. It stands in relation to a corresponding free energy in
the usual way:
Ξ(T, V, µ) ≡ e−βΩ(T,V,µ) ⇐⇒ Ω = −kB T ln Ξ , (4.126)
where Ω(T, V, µ) is the grand potential, also known as the Landau free energy. The dimensionless quantity z ≡ eβµ
is called the fugacity.
If Ĥ, N̂ = 0, the grand potential may be expressed as a sum over contributions from each N sector, viz.
X
Ξ(T, V, µ) = eβµN Z(T, V, N ) . (4.127)
N
When there is more than one species, we have several chemical potentials {µa }, and accordingly we define
X
K̂ = Ĥ − µa N̂a , (4.128)
a
Ω hĤi µ hN̂ i
=− + − ,
T T T
152 CHAPTER 4. STATISTICAL ENSEMBLES
which says
Ω = E − T S − µN , (4.130)
where
X
E= En Pn = Tr ̺ˆ Ĥ (4.131)
n
X
N= Nn Pn = Tr ̺ˆ N̂ . (4.132)
n
dΩ = −S dT − p dV − N dµ , (4.133)
which entails
∂Ω ∂Ω ∂Ω
S=− , p=− , N =− . (4.134)
∂T V,µ ∂V T,µ ∂µ T,V
Since Ω(T, V, µ) is an extensive quantity, we must be able to write Ω = V ω(T, µ). We identify the function ω(T, µ)
as the negative of the pressure:
∂Ω kB T ∂Ξ 1 X ∂En −β(En −µNn )
=− = e
∂V Ξ ∂V T,µ Ξ n ∂V
(4.135)
∂E
= = −p(T, µ) .
∂V T,µ
Therefore,
Ω = −pV , p = p(T, µ) (equation of state) . (4.136)
This is consistent with the result from thermodynamics that G = E − T S + pV = µN . Taking the differential, we
obtain the Gibbs-Duhem relation,
dΩ = −S dT − p dV − N dµ = −p dV − V dp ⇒ S dT − V dp + N dµ = 0 . (4.137)
We can appropriate the results from §4.5.7 and apply them, mutatis mutandis, to the GCE. Suppose we have a
family of observables Q̂k satisfying Q̂k , Q̂k′ = 0 and Ĥ0 , Q̂k = 0 and N̂a , Q̂k = 0 for all k, k ′ , and a.
Then for the grand canonical Hamiltonian
X X
K̂ (~λ) = Ĥ0 − µa N̂a − λk Q̂k , (4.138)
a k
we have that
∂Ω
Qk (~λ, T ) = h Q̂k i = − (4.139)
∂λk T,µa , λk′ 6=k
1 ∂Qk 1 ∂2Ω
χkl = =− . (4.140)
V ∂λl V ∂λk ∂λl
4.6. GRAND CANONICAL ENSEMBLE (GCE) 153
Both energy and particle number fluctuate in the GCE. Let us compute the fluctuations in particle number. We
have
Tr N̂ e−β(Ĥ−µN̂ ) 1 ∂
N = h N̂ i = = ln Ξ . (4.141)
Tr e−β(Ĥ−µN̂ ) β ∂µ
Therefore,
!2
1 ∂N Tr N̂ 2 e−β(Ĥ−µN̂ ) Tr N̂ e−β(Ĥ−µN̂ )
= −
β ∂µ Tr e−β(Ĥ−µN̂ ) Tr e−β(Ĥ−µN̂ )
(4.142)
2
2
= N̂ − N̂ .
Let the system’s particle number N be fixed, but let it exchange energy and volume with the world W . Mutatis
mutandis , we have
DW (EU − En , VU − Vn ) ∆E ∆V
Pn = lim lim . (4.146)
∆E→0 ∆V →0 DU (EU , VU ) ∆E ∆V
Then
ln Pn = ln DW (EU − En , VU − Vn ) − ln DU (EU , VU )
= ln DW (EU , VU ) − ln DU (EU , VU )
(4.147)
∂ ln DW (E, V ) ∂ ln DW (E, V )
− En E=EU − Vn E=EU + . . .
∂E ∂V
V =V V =V
U U
≡ −α − βEn − βp Vn .
154 CHAPTER 4. STATISTICAL ENSEMBLES
where V0 is a constant which has dimensions of volume. The factor V0−1 in front of the integral renders Y di-
mensionless. Note that G(V0′ ) = G(V0 ) + kB T ln(V0′ /V0 ), so the difference is not extensive and can be neglected
in the thermodynamic limit. In other words, it doesn’t matter what constant we choose for V0 since it contributes
subextensively to G. Moreover, in computing averages, the constant V0 divides out in the ratio of numerator
and denominator. Like the Helmholtz free energy, the Gibbs free energy G(T, p, N ) is also a double Legendre
transform of the energy E(S, V, N ), viz.
G = E − T S + pV
(4.151)
dG = −S dT + V dp + µ dN ,
which entails
∂G ∂G ∂G
S=− , V =+ , µ=+ . (4.152)
∂T p,N ∂p T,N ∂N T,p
4.7.1 µCE
1 X
Pn = , Ω= Θ(E + ∆E − En ) Θ(En − E) . (4.158)
Ω n
Note that Ω is the number of states with energies between E and E + ∆E.
4.7.2 OCE
X 2
X
δS ∗ = δS − kB λ1 + λ2 En δPn − kB Cj δλj
n j=1
(4.161)
Xh i 2
X
= −kB ln Pn + 1 + λ1 + λ2 En δPn − kB Cj δλj ≡ 0 .
n j=1
Thus, C1 = C2 = 0 and
ln Pn = − 1 + λ1 + λ2 En . (4.162)
We define λ2 ≡ β and we fix λ1 by normalization. This yields
1 −βEn X
Pn = e , Z= e−βEn . (4.163)
Z n
4.7.3 GCE
and hence
X 3
X
δS ∗ = δS − kB λ1 + λ2 En + λ3 Nn δPn − kB Cj δλj
n j=1
(4.166)
Xh i 3
X
= −kB ln Pn + 1 + λ1 + λ2 En + λ3 Nn δPn − kB Cj δλj ≡ 0 .
n j=1
Thus, C1 = C2 = C3 = 0 and
ln Pn = − 1 + λ1 + λ2 En + λ3 Nn . (4.167)
We define λ2 ≡ β and λ3 ≡ −βµ, and we fix λ1 by normalization. This yields
1 X
Pn = e−β(En−µNn ) , Ξ= e−β(En−µNn ) . (4.168)
Ξ n
The ordinary canonical partition function for the ideal gas was computed in eqn. 4.25. We found
N Z
1 Y ddxi ddpi −β p2i /2m
Z(T, V, N ) = e
N ! i=1 (2π~)d
∞ N d
N Z
V dp −βp2 /2m (4.169)
= e
N! 2π~
−∞
N
1 V
= ,
N ! λdT
where λT is the thermal wavelength: p
λT = 2π~2 /mkB T . (4.170)
The physical interpretation of λT is that it is the de Broglie wavelength for a particle of mass m which has a kinetic
energy of kB T .
In the GCE, we have
∞
X
Ξ(T, V, µ) = eβµN Z(T, V, N )
N =0
!N ! (4.171)
∞
X 1 V eµ/kB T V eµ/kB T
= = exp .
N!
N =1
λdT λdT
Combined, the last two equations recapitulate the ideal gas law, pV = N kB T .
4.8. IDEAL GAS STATISTICAL MECHANICS 157
If we are only interested in averaging functions of v = |v| which are isotropic, then we can define the Maxwell
speed distribution, f˜(v), as
3/2
m 2
f˜(v) = 4π v 2 f (v) = 4π v 2 e−mv /2kB T . (4.180)
2πkB T
Note that f˜(v) is normalized according to
Z∞
dv f˜(v) = 1 . (4.181)
0
p
It is convenient to represent v in units of v0 = kB T /m, in which case
q
1 2
f˜(v) = ϕ(v/v0 ) , ϕ(s) = π2 s2 e−s /2 . (4.182)
v0
The distribution ϕ(s) is shown in fig. 4.6. Computing averages, we have
Z∞
2
k
Ck ≡ hs i = ds sk ϕ(s) = 2k/2 · √ Γ 3
2 + k
2 . (4.183)
π
0
q
8
Thus, C0 = 1, C1 = π, C2 = 3, etc. The speed averages are
k/2
k kB T
v = Ck . (4.184)
m
p
Note that the average velocity is hvi = 0, but the average speed is hvi = 8kB T /πm. The speed distribution is
plotted in fig. 4.6.
158 CHAPTER 4. STATISTICAL ENSEMBLES
√
Figure 4.6:qMaxwell distribution of speeds ϕ(v/v0 ). The most probable speed is vMAX = 2 v0 . The average speed
√
is vAVG = π8 v0 . The RMS speed is vRMS = 3 v0 .
4.8.2 Equipartition
The Hamiltonian for ballistic (i.e. massive nonrelativistic) particles is quadratic in the individual components of
each momentum pi . There are other cases in which a classical degree of freedom appears quadratically in Ĥ as
well. For example, an individual normal mode ξ of a system of coupled oscillators has the Lagrangian
L = 12 ξ˙2 − 1
2 ω02 ξ 2 , (4.185)
where the dimensions of ξ are [ξ] = M 1/2 L by convention. The Hamiltonian for this normal mode is then
p2
Ĥ = + 1
2 ω02 ξ 2 , (4.186)
2
from which we see that both the kinetic as well as potential energy terms enter quadratically into the Hamiltonian.
The classical rotational kinetic energy is also quadratic in the angular momentum components.
Let us compute the contribution of a single quadratic degree of freedom in Ĥ to the partition function. We’ll
call this degree of freedom ζ – it may be a position or momentum or angular momentum – and we’ll write its
contribution to Ĥ as
Ĥζ = 12 Kζ 2 , (4.187)
where K is some constant. Integrating over ζ yields the following factor in the partition function:
Z∞ 1/2
2 2π
dζ e−βKζ /2 = . (4.188)
Kβ
−∞
∂ 1
∆Eζ = β ∆Fζ = = 12 kB T . (4.190)
∂β 2β
We have thus derived what is commonly called the equipartition theorem of classical statistical mechanics:
To each degree of freedom which enters the Hamiltonian quadratically is associated a contribution
1 1
2 kB T to the internal energy of the system. This results in a concomitant contribution of 2 kB to the heat
capacity.
We now see why the internal energy of a classical ideal gas with f degrees of freedom per molecule is E =
1 1
2 f N kB T , and CV = 2 N kB . This result also has applications in the theory of solids. The atoms in a solid possess
kinetic energy due to their motion, and potential energy due to the spring-like interatomic potentials which tend
to keep the atoms in their preferred crystalline positions. Thus, for a three-dimensional crystal, there are six
quadratic degrees of freedom (three positions and three momenta) per atom, and the classical energy should
be E = 3N kB T , and the heat capacity CV = 3N kB . As we shall see, quantum mechanics modifies this result
considerably at temperatures below the highest normal mode (i.e. phonon) frequency, but the high temperature
limit is given by the classical value CV = 3νR (where ν = N/NA is the number of moles) derived here, known as
the Dulong-Petit limit.
N
X
Ĥ = ĥj . (4.191)
j=1
The single particle Hamiltonian ĥ has eigenstates | α i with corresponding energy eigenvalues εα . What is the
partition function? Is it
?
X X −β ε + ε + ... + ε
H= ··· e α
1
α
2
α
N = ζN , (4.192)
α1 αN
For systems where the individual particles are distinguishable, such as spins on a lattice which have fixed positions,
this is indeed correct. But for particles free to move in a gas, this equation is wrong. The reason is that for
indistinguishable particles the many particle quantum mechanical states are specified by a collection of occupation
numbers nα , which tell us how many particles are in the single-particle state | α i. The energy is
X
E= nα ε α (4.194)
α
That is, each collection of occupation numbers {nα } labels a unique many particle state {nα } . In the product
ζ N , the collection {nα } occurs many times. We have therefore overcounted the contribution to ZN due to this state.
By what factor have we overcounted? It is easy to see that the overcounting factor is
N!
degree of overcounting = Q ,
α nα !
which is the number of ways we can rearrange the labels αj to arrive at the same collection {nα }. This follows
from the multinomial theorem,
K
!N
X XX X N! n n n
xα = ··· x1 1 x2 2 · · · xKK δN,n1 + ... + nK . (4.196)
α=1
n 1 ! n 2 ! · · · n K !
n1 n2 nK
When we study quantum statistics, we shall learn how to handle these constrained sums. For now it suffices to
note that in the high temperature limit, almost all the nα are either 0 or 1, hence
ζN
ZN ≈ . (4.198)
N!
This is the classical Maxwell-Boltzmann limit of quantum statistical mechanics. We now see the origin of the 1/N !
term which is so important in the thermodynamics of entropy of mixing.
Consider a system of Ns spins , each of which can be either up (σ = +1) or down (σ = −1). The Hamiltonian for
this system is
Ns
X
Ĥ = −µ0 H σj , (4.199)
j=1
where now we write Ĥ for the Hamiltonian, to distinguish it from the external magnetic field H, and µ0 is the
magnetic moment per particle. We treat this system within the ordinary canonical ensemble. The partition func-
tion is X X
Z= ··· e−β Ĥ = ζ Ns , (4.200)
σ1 σN
s
The magnetization is
∂F µ0 H
M =− = Ns µ0 tanh . (4.203)
∂H T, Ns kB T
The energy is
∂ µ H
E= βF = −Ns µ0 H tanh 0 . (4.204)
∂β kB T
Hence, E = −HM , which we already knew, from the form of Ĥ itself.
Each spin here is independent. The probability that a given spin has polarization σ is
eβµ0 Hσ
Pσ = . (4.205)
eβµ0 H + e−βµ0 H
The total probability is unity, and the average polarization is a weighted average of σ = +1 and σ = −1 contribu-
tions:
µ0 H
P↑ + P↓ = 1 , hσi = P↑ − P↓ = tanh . (4.206)
kB T
At low temperatures T ≪ µ0 H/kB , we have P↑ ≈ 1 − e−2µ0 H/kB T . At high temperatures T > µ0 H/kB , the two
σµ H
polarizations are equally likely, and Pσ ≈ 21 1 + k 0T .
B
(Typically this is computed per unit volume rather than per particle.) At H = 0, we have χT = µ20 /kB T , which is
known as the Curie law.
Aside
The energy E = −HM here is not the same quantity we discussed in our study of thermodynamics. In fact,
the thermodynamic energy for this problem vanishes! Here is why. To avoid confusion, we’ll need to invoke a
new symbol for the thermodynamic energy, E. Recall that the thermodynamic energy E is a function of exten-
sive quantities, meaning E = E(S, M, Ns ). It is obtained from the free energy F (T, H, Ns ) by a double Legendre
transform:
E(S, M, Ns ) = F (T, H, Ns ) + T S + HM . (4.208)
Now from eqn. 4.202 we derive the entropy
" #
∂F µ0 H µ0 H µ0 H
S=− = Ns kB ln 2 cosh − Ns tanh . (4.209)
∂T kB T T kB T
The potential confusion here arises from our use of the expression F (T, H, Ns ). In thermodynamics, it is the Gibbs
free energy G(T, p, N ) which is a double Legendre transform of the energy: G = E − T S + pV . By analogy, with
magnetic systems we should perhaps write G = E − T S − HM , but in keeping with many textbooks we shall
use the symbol F and refer to it as the Helmholtz free energy. The quantity we’ve called E in eqn. 4.204 is in fact
E = E − HM , which means E = 0. The energy E(S, M, Ns ) vanishes here because the spins are noninteracting.
Consider again a system of Ns spins, each of which can be either up (+) or down (−). Let Nσ be the number of
sites with spin σ, where σ = ±1. Clearly N+ + N− = Ns . We now treat this system within the microcanonical
ensemble.
The energy of the system is
E = −HM , (4.210)
where H is an external magnetic field, and M = (N+ − N− )µ0 is the total magnetization. We now compute S(E)
using the ordinary canonical ensemble. The number of ways of arranging the system with N+ up spins is
Ns
Ω= , (4.211)
N+
E M E0 − E
=− = 1 − 2x =⇒ x= . (4.213)
E0 Ns µ0 2E0
We therefore have
" #
E0 − E E0 − E E0 + E E0 + E
S(E, Ns ) = −Ns kB ln + ln . (4.214)
2E0 2E0 2E0 2E0
We now have
1 ∂S ∂S ∂x Ns kB E0 − E
= = = ln . (4.215)
T ∂E Ns ∂x ∂E 2E0 E0 + E
We see that the temperature is positive for −E0 ≤ E < 0 and is negative for 0 < E ≤ E0 .
What has gone wrong? The answer is that nothing has gone wrong – all our calculations are perfectly correct. This
system does exhibit the possibility of negative temperature. It is, however, unphysical in that we have neglected
kinetic degrees of freedom, which result in an entropy function S(E, Ns ) which is an increasing function of energy.
In this system, S(E, Ns ) achieves a maximum of Smax = Ns kB ln 2 at E = 0 (i.e. x = 12 ), and then turns over and
starts decreasing. In fact, our results are completely consistent with eqn. 4.204 : the energy E is an odd function
of temperature. Positive energy requires negative temperature! Another example of this peculiarity is provided
in the appendix in §4.11.2.
4.9. SELECTED EXAMPLES 163
Figure 4.7: When entropy decreases with increasing energy, the temperature is negative. Typically, kinetic degrees
of freedom prevent this peculiarity from manifesting in physical systems.
4.9.3 Adsorption
PROBLEM: A surface containing Ns adsorption sites is in equilibrium with a monatomic ideal gas. Atoms adsorbed
on the surface have an energy −∆ and no kinetic energy. Each adsorption site can accommodate at most one
atom. Calculate the fraction f of occupied adsorption sites as a function of the gas density n, the temperature T ,
the binding energy ∆, and physical constants.
The grand partition function for the surface is
Ns
Ξsurf = e−Ωsurf /kB T = 1 + e∆/kB T eµ/kB T . (4.216)
Since the surface is in equilibrium with the gas, its fugacity z = exp(µ/kB T ) and temperature T are the same as in
the gas.
p
SOLUTION: For a monatomic ideal gas, the single particle partition function is ζ = V λ−3
T , where λT = 2π~2 /mkB T
is the thermal wavelength. Thus, the grand partition function, for indistinguishable particles, is
Ξgas = exp V λ−3
T e µ/kB T
. (4.218)
We can now solve for the fugacity: z = eµ/kB T = nλ3T . Thus, the fraction of occupied adsorption sites is
nλ3T
f= . (4.220)
nλ3T + e−∆/kB T
Wool consists of interlocking protein molecules which can stretch into an elongated configuration, but reversibly
so. This feature gives wool its very useful elasticity. Let us model a chain of these proteins by assuming they can
exist in one of two states, which we will call A and B, with energies εA and εB and lengths ℓA and ℓB . The situation
is depicted in fig. 4.8. We model these conformational degrees of freedom by a spin variable σ = ±1 for each
molecule, where σ = +1 in the A state and σ = −1 in the B state. Suppose a chain consisting of N monomers is
placed under a tension τ . We then have
N h
X
1
1
i
Ĥ = 2 εA + εB + 2 εA − εB σj . (4.221)
j=1
N h
X
1
1
i
K̂ = 2 ε̃A + ε̃B + 2 ε̃A − ε̃B σj , (4.223)
j=1
where ε̃A ≡ εA − τ ℓA and ε̃B ≡ εB − τ ℓB . At τ = 0 the A state is preferred for each monomer, but when τ exceeds
τ ∗ , defined by the relation ε̃A = ε̃B , the B state is preferred. One finds
εB − εA
τ∗ = . (4.224)
ℓB − ℓA
Figure 4.8: The monomers in wool are modeled as existing in one of two states. The low energy undeformed state
is A, and the higher energy deformed state is B. Applying tension induces more monomers to enter the B state.
4.9. SELECTED EXAMPLES 165
Figure 4.9: Upper panel: length L(τ, T ) for kB T /ε̃ = 0.01 (blue), 0.1 (green), 0.5 (dark red), and 1.0 (red). Bottom
panel: dimensionless force constant k/N (∆ℓ)2 versus temperature.
Once again, we have a set of N noninteracting spins. The partition function is Y = ζ N , where ζ is the single
monomer partition function,
ζ = Tr e−β ĥ = e−β ε̃A + e−β ε̃B , (4.225)
where
1 1
ĥ = 2 ε̃A + ε̃B + 2 ε̃A − ε̃B σ , (4.226)
is the single spin Hamiltonian. It is convenient to define the differences
The polymer behaves as a spring, and for small τ the spring constant is
∂τ 4kB T 2 ∆ε
k= = cosh . (4.231)
∂L τ =0 N (∆ℓ)2 2kB T
The results are shown in fig. 4.9. Note that length increases with temperature for τ < τ ∗ and decreases with
temperature for τ > τ ∗ . Note also that k diverges at both low and high temperatures. At low T , the energy gap
∆ε dominates and L = N ℓA , while at high temperatures kB T dominates and L = 21 N (ℓA + ℓB ).
Consider a system of noninteracting spin dimers as depicted in fig. 4.10. Each dimer contains two spins, and is
described by the Hamiltonian
Ĥdimer = −J σ1 σ2 − µ0 H (σ1 + σ2 ) . (4.232)
Here, J is an interaction energy between the spins which comprise the dimer. If J > 0 the interaction is ferromagnetic,
which prefers that the spins are aligned. That is, the lowest energy states are | ↑↑ i and | ↓↓ i. If J < 0 the interaction
is antiferromagnetic, which prefers that spins be anti-aligned: | ↑↓ i and | ↓↑ i.6
Suppose there are Nd dimers. Then the OCE partition function is Z = ζ Nd , where ζ(T, H) is the single dimer
partition function. To obtain ζ(T, H), we sum over the four possible states of the two spins, obtaining
ζ = Tr e−Ĥdimer /kB T
2µ0 H
= 2 e−J/kB T + 2 eJ/kB T cosh .
kB T
Thus, the free energy is
" #
−J/kB T J/kB T 2µ0 H
F (T, H, Nd ) = −Nd kB T ln 2 − Nd kB T ln e +e cosh . (4.233)
kB T
The magnetization is
2µ H
eJ/kB T sinh k 0T
∂F B
M =− = 2Nd µ0 · (4.234)
∂H T,Nd e −J/kB T +e J/k B T 2µ H
cosh k 0T
B
The quantity µ20 /kB T is simply the Curie susceptibility for noninteracting classical spins. Note that we correctly
recover the Curie result when J = 0, since then the individual spins comprising each dimer are in fact noninter-
acting. For the ferromagnetic case, if J ≫ kB T , then we obtain
2µ20
χT (J ≫ kB T ) ≈ . (4.236)
kB T
This has the following simple interpretation. When J ≫ kB T , the spins of each dimer are effectively locked in
parallel. Thus, each dimer has an effective magnetic moment µeff = 2µ0 . On the other hand, there are only half as
many dimers as there are spins, so the resulting Curie susceptibility per spin is 12 × (2µ0 )2 /kB T .
6 Nota bene we are concerned with classical spin configurations only – there is no superposition of states allowed in this model!
4.10. STATISTICAL MECHANICS OF MOLECULAR GASES 167
Figure 4.10: A model of noninteracting spin dimers on a lattice. Each red dot represents a classical spin for which
σj = ±1.
When −J ≫ kB T , the spins of each dimer are effectively locked in one of the two antiparallel configurations. We
then have
2µ20 −2|J|/kB T
χT (−J ≫ kB T ) ≈ e . (4.237)
kB T
In this case, the individual dimers have essentially zero magnetic moment.
The states of a noninteracting atom or molecule are labeled by its total momentum p and its internal quantum
numbers, which we will simply write with a collective index α, specifying rotational, vibrational, and electronic
degrees of freedom. The single particle Hamiltonian is then
p2
ĥ = + ĥint , (4.238)
2m
with
~2 k 2
ĥ k , α = + εα k , α . (4.239)
2m
The partition function is
X 2 X
ζ = Tr e−β ĥ = e−β p /2m
gj e−βεj . (4.240)
p j
Here we have replaced the internal label α with a label j of energy eigenvalues, with gj being the degeneracy of
the internal state with energy εj . To do the p sum, we quantize in a box of dimensions L1 × L2 × · · · × Ld , using
periodic boundary conditions. Then
2π~n1 2π~n2 2π~nd
p= , , ... , , (4.241)
L1 L2 Ld
168 CHAPTER 4. STATISTICAL ENSEMBLES
where each ni is an integer. Since the differences between neighboring quantized p vectors are very tiny, we can
replace the sum over p by an integral:
X Z
ddp
−→ (4.242)
p
∆p1 · · · ∆pd
(2π~)d (2π~)d
∆p1 · · · ∆pd = = . (4.243)
L1 · · · Ld V
Thus,
Z X
ddp 2
ζ=V d
e−p /2mkB T gj e−εj /kB T = V λ−d
T ξ (4.244)
(2π~) j
X
ξ(T ) = gj e−εj /kB T . (4.245)
j
Here, ξ(T ) is the internal coordinate partition function. The full N -particle ordinary canonical partition function is
then
N
1 V
ZN = ξ N (T ) . (4.246)
N ! λdT
where
ϕ(T ) = −kB T ln ξ(T ) (4.248)
is the internal coordinate contribution to the single particle free energy. We could also compute the partition
function in the Gibbs (T, p, N ) ensemble:
Z∞
1
Y (T, p, N ) = e−βG(T,p,N ) = dV e−βpV Z(T, V, N )
V0
0 (4.249)
N
kB T kB T
= ξ N (T ) .
pV0 p λdT
Since the internal coordinate contribution to the free energy is volume-independent, we have
∂G N kB T
V = = , (4.251)
∂p T,N p
Thus, any temperature variation in Cp must be due to the internal degrees of freedom.
At energy scales of interest we can separate the internal degrees of freedom into distinct classes, writing
as a sum over internal Hamiltonians governing rotational, vibrational, and electronic degrees of freedom. Then
Associated with each class of excitation is a characteristic temperature Θ. Rotational and vibrational temperatures
of a few common molecules are listed in table tab. 4.1.
4.10.4 Rotations
Consider a class of molecules which can be approximated as an axisymmetric top. The rotational Hamiltonian is
then
L2a + L2b L2
ĥrot = + c
2I1 2I3
(4.257)
~2 L(L + 1) 1 1
= + − L2c ,
2I1 2I3 2I1
where n̂a.b,c (t) are the principal axes, with n̂c the symmetry axis, and La,b,c are the components of the angular
momentum vector L about these instantaneous body-fixed principal axes. The components of L along space-fixed
axes {x, y, z} are written as Lx,y,z . Note that
µ
L , Lc = nνc Lµ , Lν + Lµ , nνc Lν = iǫµνλ nνc Lλ + iǫµνλ nλc Lν = 0 , (4.258)
170 CHAPTER 4. STATISTICAL ENSEMBLES
which is equivalent to the statement that Lc = n̂c · L is a rotational scalar. We can therefore simultaneously specify
the eigenvalues of {L2 , Lz , Lc }, which form a complete set of commuting observables (CSCO)7 . The eigenvalues
of Lz are m~ with m ∈ {−L, . . . , L}, while those of Lc are k~ with k ∈ {−L, . . . , L}. There is a (2L + 1)-fold
degeneracy associated with the Lz quantum number.
We assume the molecule is prolate, so that I3 < I1 . We can the define two temperature scales,
~2 e= ~2
Θ= , Θ . (4.259)
2I1 kB 2I3 kB
e > Θ. We conclude that the rotational partition function for an axisymmetric molecule
Prolateness then means Θ
is given by
∞
X XL
2 e
ξrot (T ) = (2L + 1) e−L(L+1) Θ/T e−k (Θ−Θ)/T (4.260)
L=0 k=−L
In the high temperature limit, we have a slowly varying summand. The Euler-MacLaurin summation formula may
be used to evaluate such a series:
Zn
n
X X∞
B2j h (2j−1) i
Fk = dk F (k) + 21 F (0) + F (n) + F (n) − F (2j−1) (0) (4.263)
(2j)!
k=0 0 j=1
B0 = 1 , B1 = − 21 , B2 = 1
6 , 1
B4 = − 30 , B6 = 1
42 . (4.264)
Thus,
∞
X Z∞
′ 1
Fk = dx F (x) + 12 F (0) − 1
12 F (0) − F ′′′ (0) + . . . . (4.265)
720
k=0 0
7 Note that while we cannot simultaneously specify the eigenvalues of two components of L along axes fixed in space, we can simulta-
neously specify the components of L along one axis fixed in space and one axis rotating with a body. See Landau and Lifshitz, Quantum
Mechanics, §103.
4.10. STATISTICAL MECHANICS OF MOLECULAR GASES 171
R∞ T
We have F (x) = (2x + 1) e−x(x+1)Θ/T , for which dx F (x) = Θ , hence
0
2
T 1 1 Θ 4 Θ
ξrot = + + + + ... . (4.266)
Θ 3 15 T 315 T
Recall that ϕ(T ) = −kB T ln ξ(T ). We conclude that ϕrot (T ) ≈ −3kB T e−2Θ/T for T ≪ Θ and ϕrot (T ) ≈ −kB T ln(T /Θ)
for T ≫ Θ. We have seen that the internal coordinate contribution to the heat capacity is ∆CV = −N T ϕ′′ (T ). For
diatomic molecules, then, this contribution is exponentially suppressed for T ≪ Θ, while for high temperatures
we have ∆CV = N kB . One says that the rotational excitations are ‘frozen out’ at temperatures much below Θ.
Including the first few terms, we have
2
Θ
∆CV (T ≪ Θ) = 12 N kB e−2Θ/T + . . . (4.267)
T
( 2 3 )
1 Θ 16 Θ
∆CV (T ≫ Θ) = N kB 1 + + + ... . (4.268)
45 T 945 T
Note that CV overshoots its limiting value of N kB and asymptotically approaches it from above.
Special care must be taken in the case of homonuclear diatomic molecules, for then only even or odd L states are
allowed, depending on the total nuclear spin. This is discussed below in §4.10.7.
For polyatomic molecules, the moments of inertia generally are large enough that the molecule’s rotations can be
considered classically. We then have
L2a L2 L2
ε(La , Lb , Lc ) = + b + c . (4.269)
2I1 2I2 2I3
We then have Z
1 dLa dLb dLc dφ dθ dψ −ε(La Lb Lc )/kB T
ξrot (T ) = e , (4.270)
grot (2π~)3
where (φ, θ ψ) are the Euler angles. Recall φ ∈ [0, 2π], θ ∈ [0, π], and ψ ∈ [0, 2π]. The factor grot accounts for
physically indistinguishable orientations of the molecule brought about by rotations, which can happen when
more than one of the nuclei is the same. We then have
3/2
2kB T p
ξrot (T ) = πI1 I2 I3 . (4.271)
~2
4.10.5 Vibrations
Vibrational frequencies are often given in units of inverse wavelength, such as cm−1 , called a wavenumber. To
convert to a temperature scale T ∗ , we write kB T ∗ = hν = hc/λ, hence T ∗ = (hc/kB ) λ−1 , and we multiply by
hc
= 1.436 K · cm . (4.272)
kB
For example, infrared absorption (∼ 50 cm−1 to 104 cm−1 ) reveals that the ‘asymmetric stretch’ mode of the H2 O
molecule has a vibrational frequency of ν = 3756 cm−1 . The corresponding temperature scale is T ∗ = 5394 K.
172 CHAPTER 4. STATISTICAL ENSEMBLES
Vibrations are normal modes of oscillations. A single normal mode Hamiltonian is of the form
p2
ĥ = + 12 mω 2 q 2 = ~ω a† a + 21 . (4.273)
2m
In general there are many vibrational modes, hence many normal mode frequencies ωα . We then must sum over
all of them, resulting in
Y (α)
ξvib = ξvib . (4.274)
α
Consider now a two-level system, with energies ε0 and ε1 . We define ∆ ≡ ε1 − ε0 and assume without loss of
generality that ∆ > 0. The partition function is
ζ = e−βε0 + e−βε1 = e−βε0 1 + e−β∆ . (4.278)
∂f ∆ 1
s=− = kB ln 1 + e−∆/kB T + · ∆/k T (4.280)
∂T T e B +1
∆2 e∆/kB T
c(T ) = · . (4.281)
kB T 2 e∆/kB T + 1 2
4.10. STATISTICAL MECHANICS OF MOLECULAR GASES 173
Figure 4.11: Heat capacity per molecule as a function of temperature for (a) heteronuclear diatomic gases, (b) a
single vibrational mode, and (c) a single two-level system.
Thus,
∆2 −∆/kB T
c (T ≪ ∆) = e (4.282)
kB T 2
∆2
c (T ≫ ∆) = . (4.283)
4kB T 2
We find that c(T ) has a characteristic peak at T ∗ ≈ 0.42 ∆/kB . The heat capacity vanishes in both the low tem-
perature and high temperature limits. At low temperatures, the gap to the excited state is much greater than kB T ,
and it is not possible to populate it and store energy. At high temperatures, both ground state and excited state
are equally populated, and once again there is no way to store energy.
If we have a distribution of independent two-level systems, the heat capacity of such a system is a sum over the
individual Schottky functions:
X Z∞
C(T ) = c (∆i /kB T ) = N d∆ P (∆) e
e c(∆/T ) , (4.284)
i 0
where N is the number of two level systems, e c(x) = kB x2 ex /(ex + 1)2 , and where P (∆) is the normalized distri-
bution function, which satisfies the normalization condition
Z∞
d∆ P (∆) = 1 . (4.285)
0
Ns is the total number of two level systems. If P (∆) ∝ ∆r for ∆ → 0, then the low temperature heat capacity
behaves as C(T ) ∝ T 1+r . Many amorphous or glassy systems contain such a distribution of two level systems,
with r ≈ 0 for glasses, leading to a linear low-temperature heat capacity. The origin of these two-level sys-
tems is not always so clear but is generally believed to be associated with local atomic configurations for which
there are two low-lying states which are close in energy. The paradigmatic example is the mixed crystalline solid
(KBr)1−x (KCN)x which over the range 0.1 < ∼x< ∼ 0.6 forms an ‘orientational glass’ at low temperatures. The two
level systems are associated with different orientation of the cyanide (CN) dipoles.
174 CHAPTER 4. STATISTICAL ENSEMBLES
For a monatomic gas, the internal coordinate partition function arises due to electronic and nuclear degrees of
freedom. Let’s first consider the electronic degrees of freedom. We assume that kB T is small compared with
energy differences between successive electronic shells. The atomic ground state is then computed by filling up
the hydrogenic orbitals until all the electrons are used up. If the atomic number is a ‘magic number’ (A = 2 (He),
10 (Ne), 18 (Ar), 36 (Kr), 54 (Xe), etc.) then the atom has all shells filled and L = 0 and S = 0. Otherwise the last
shell is partially filled and one or both of L and S will be nonzero. The atomic ground state configuration 2J+1 LS
is then determined by Hund’s rules:
The last of Hund’s rules distinguishes between the (2S + 1)(2L + 1) states which result upon fixing S and L as
per rules #1 and #2. It arises due to the atomic spin-orbit coupling, whose effective Hamiltonian may be written
Ĥ = ΛL · S, where Λ is the Russell-Saunders coupling. If the last shell is less than or equal to half-filled, then
Λ > 0 and the ground state has J = |L − S|. If the last shell is more than half-filled, the coupling is inverted, i.e.
Λ < 0, and the ground state has J = L + S.8
The electronic contribution to ξ is then
L+S
X
ξelec = (2J + 1) e−∆ε(L,S,J)/kBT (4.286)
J=|L−S|
where h i
∆ε(L, S, J) = 21 Λ J(J + 1) − L(L + 1) − S(S + 1) . (4.287)
At high temperatures, kB T is larger than the energy difference between the different J multiplets, and we have
ξelec ∼ (2L + 1)(2S + 1) e−βε0 , where ε0 is the ground state energy. At low temperatures, a particular value of
J is selected – that determined by Hund’s third rule – and we have ξelec ∼ (2J + 1) e−βε0 . If, in addition, there
is a nonzero nuclear spin I, then we also must include a factor ξnuc = (2I + 1), neglecting the small hyperfine
splittings due to the coupling of nuclear and electronic angular momenta.
For heteronuclear diatomic molecules, i.e. molecules composed from two different atomic nuclei, the internal par-
(1) (2)
tition function simply receives a factor of ξelec · ξnuc · ξnuc , where the first term is a sum over molecular electronic
states, and the second two terms arise from the spin degeneracies of the two nuclei. For homonuclear diatomic
molecules, the exchange of nuclear centers is a symmetry operation, and does not represent a distinct quantum
state. To correctly count the electronic states, we first assume that the total electronic spin is S = 0. This is gen-
erally a very safe assumption. Exchange symmetry now puts restrictions on the possible values of the molecular
angular momentum L, depending on the total nuclear angular momentum Itot . If Itot is even, then the molecular
angular momentum L must also be even. If the total nuclear angular momentum is odd, then L must be odd. This
is so because the molecular ground state configuration is 1Σg+ .9
The total number of nuclear states for the molecule is (2I + 1)2 , of which some are even under nuclear exchange,
and some are odd. The number of even states, corresponding to even total nuclear angular momentum is written
8 See e.g. §72 of Landau and Lifshitz, Quantum Mechanics, which, in my humble estimation, is the greatest physics book ever written.
9 See Landau and Lifshitz, Quantum Mechanics, §86.
4.10. STATISTICAL MECHANICS OF MOLECULAR GASES 175
2I gg gu
odd I(2I + 1) (I + 1)(2I + 1)
even (I + 1)(2I + 1) I(2I + 1)
Table 4.2: Number of even (gg ) and odd (gu ) total nuclear angular momentum states for a homonuclear diatomic
molecule. I is the ground state nuclear spin.
as gg , where the subscript conventionally stands for the (mercifully short) German word gerade, meaning ‘even’.
The number of odd (Ger. ungerade) states is written gu . Table 4.2 gives the values of gg,u corresponding to half-
odd-integer I and integer I.
The final answer for the rotational component of the internal molecular partition function is then
ξrot (T ) = gg ζg + gu ζu , (4.288)
where
X
ζg = (2L + 1) e−L(L+1) Θrot /T
L even
X (4.289)
ζu = (2L + 1) e−L(L+1) Θrot /T .
L odd
For hydrogen, the molecules with the larger nuclear statistical weight are called orthohydrogen and those with the
smaller statistical weight are called parahydrogen. For H2 , we have I = 12 hence the ortho state has gu = 3 and the
para state has gg = 1. In D2 , we have I = 1 and the ortho state has gg = 6 while the para state has gu = 3. In
equilibrium, the ratio of ortho to para states is then
NHortho
2
gu ζu 3 ζu NDortho
2
gg ζg 2 ζg
= = , = = . (4.290)
NHpara gg ζg ζg NDpara gu ζu ζu
2 2
Incidentally, how do we derive the results in Tab. 4.10.7 ? The total nuclear angular momentum Itot is the quan-
tum mechanical sum of the two individual nuclear angular momenta, each of which are of magnitude I. From
elementary addition of angular momenta, we have
I ⊗ I = 0 ⊕ 1 ⊕ 2 ⊕ · · · ⊕ 2I . (4.291)
The right hand side of the above equation lists all the possible multiplets. Thus, Itot ∈ {0, 1, . . . , 2I}. Now let us
count the total number of states with even Itot . If 2I is even, which is to say if I is an integer, we have
I n
X o
gg(2I=even) = 2 · (2n) + 1 = (I + 1)(2I + 1) , (4.292)
n=0
It follows that
gu(2I=odd) = (2I + 1)2 − gg = (I + 1)(2I + 1) . (4.295)
176 CHAPTER 4. STATISTICAL ENSEMBLES
Consider a spin-1 particle where σ = −1, 0, +1. We model this with the single particle Hamiltonian
We can also interpret this as describing a spin if σ = ±1 and a vacancy if σ = 0. The parameter ∆ then represents
the vacancy formation energy. The single particle partition function is
With Ns distinguishable noninteracting spins (e.g. at different sites in a crystalline lattice), we have Z = ζ Ns and
h i
F ≡ Nsf = −kB T ln Z = −Ns kB T ln e−β∆ + 2 cosh(βµ0 H) , (4.298)
∂ ĥ
n̂V = 1 − σ 2 = (4.299)
∂∆
∂ ĥ
m̂ = µ0 σ = − (4.300)
∂H
are the vacancy number and magnetization, respectively. Thus,
∂f e−∆/kB T
nV = n̂V = = −∆/k T (4.301)
∂∆ e B + 2 cosh(µ0 H/kB T )
and
∂f 2µ sinh(µ0 H/kB T )
m = m̂ = − = −∆/k 0T . (4.302)
∂H e B + 2 cosh(µ0 H/kB T )
At weak fields we can compute
∂m µ20 2
χT = = · . (4.303)
∂H H=0 kB T 2 + e−∆/kB T
We thus obtain a modified Curie law. At temperatures T ≪ ∆/kB , the vacancies are frozen out and we recover the
usual Curie behavior. At high temperatures, where T ≫ ∆/kB , the low temperature result is reduced by a factor
of 23 , which accounts for the fact that one third of the time the particle is in a nonmagnetic state with σ = 0.
PROBLEM: A collection of spin- 21 particles is confined to a surface with N sites. For each site, let σ = 0 if there
is a vacancy, σ = +1 if there is particle present with spin up, and σ = −1 if there is a particle present with spin
down. The particles are non-interacting, and the energy for each site is given by ε = −W σ 2 , where −W < 0 is the
binding energy.
(a) Let Q = N↑ + N↓ be the number of spins, and N0 be the number of vacancies. The surface magnetization is
M = N↑ − N↓ . Compute, in the microcanonical ensemble, the statistical entropy S(Q, M ).
4.11. APPENDIX I : ADDITIONAL EXAMPLES 177
(b) Let q = Q/N and m = M/N be the dimensionless particle density and magnetization density, respectively.
Assuming that we are in the thermodynamic limit, where N , Q, and M all tend to infinity, but with q and m
finite, Find the temperature T (q, m). Recall Stirling’s formula
ln(N !) = N ln N − N + O(ln N ) .
(c) Show explicitly that T can be negative for this system. What does negative T mean? What physical degrees
of freedom have been left out that would avoid this strange property?
N↑ + N0 + N↓ = Q + N0 = N . (4.304)
Fixing Q and M , along with the above constraint, is enough to completely determine {N↑ , N0 , N↓ }:
1 1
N↑ = 2 (Q + M ) , N0 = N − Q , N↓ = 2 (Q − M ) , (4.306)
whence
N!
Ω(Q, M ) = 1 1 . (4.307)
2 (Q + M ) ! 2 (Q − M ) ! (N − Q)!
The statistical entropy is S = kB ln Ω:
1
S(Q, M ) = kB ln(N !) − kB ln 2 (Q + M )! − kB ln 12 (Q − M )! − kB ln (N − Q)! . (4.308)
Thus,
W/kB
T = p . (4.314)
ln 2(1 − q)/ q 2 − m2
(c) We have 0 ≤ q ≤ 1 and −q ≤ m ≤ q, so T is real (thank heavens!). But it is easy to choose {q, m} such that
T < 0. For example, when m = 0 we have T = W/kB ln(2q −1 − 2) and T < 0 for all q ∈ 32 , 1 . The reason for
this strange state of affairs is that the entropy S is bounded, and is not an monotonically increasing function
of the energy E (or the dimensionless quantity Q). The entropy is maximized for N ↑= N0 = N↓ = 13 , which
says m = 0 and q = 23 . Increasing q beyond this point (with m = 0 fixed) starts to reduce the entropy, and
hence (∂S/∂E) < 0 in this range, which immediately gives T < 0. What we’ve left out are kinetic degrees of
freedom, such as vibrations and rotations, whose energies are unbounded, and which result in an increasing
S(E) function.
Consider an interface between two dissimilar fluids. In equilibrium, in a uniform gravitational field, the denser
fluid is on the bottom. Let z = z(x, y) be the height the interface between the fluids, relative to equilibrium. The
potential energy is a sum of gravitational and surface tension terms, with
Z Zz
Ugrav = d x dz ′ ∆ρ g z ′
2
(4.315)
0
Z
Usurf = d2 x 12 σ (∇z)2 . (4.316)
We won’t need the kinetic energy in our calculations, but we can include it just for completeness. It isn’t so clear
how to model it a priori so we will assume a rather general form
Z Z
∂z(x, t) ∂z(x′ , t)
T = d x d2x′ 12 µ(x, x′ )
2
. (4.317)
∂t ∂t
We assume that the (x, y) plane is a rectangle of dimensions Lx × Ly . We also assume µ(x, x′ ) = µ |x − x′ | . We
can then Fourier transform −1/2 X
z(x) = Lx Ly zk eik·x , (4.318)
k
where the wavevectors k are quantized according to
2πnx 2πny
k= x̂ + ŷ , (4.319)
Lx Ly
with integer nx and ny , if we impose periodic boundary conditions (for calculational convenience). The La-
grangian is then
1 X h 2 2 i
L= µk żk − g ∆ρ + σk2 zk , (4.320)
2
k
4.11. APPENDIX I : ADDITIONAL EXAMPLES 179
where Z
µk = d2x µ |x| e−ik·x . (4.321)
Since z(x, t) is real, we have the relation z−k = zk∗ , therefore the Fourier coefficients at k and −k are not indepen-
dent. The canonical momenta are given by
∂L ∂L
pk = = µk żk , p∗k = = µk żk∗ (4.322)
∂ żk∗ ∂ żk
The Hamiltonian is then
X′ h i
Ĥ = pk zk∗ + p∗k zk − L (4.323)
k
X′ |p |2
= k + g ∆ρ + σk2 |z |2 , (4.324)
µk k
k
where the prime on the k sum indicates that only one of the pair {k, −k} is to be included, for each k.
We may now compute the ordinary canonical partition function:
Y′ Z d2p d2z
Z= k k e−|pk |2 /µk kB T e−(g ∆ρ+σk2 ) |zk |2 /kB T
(2π~)2
k
(4.325)
Y′ k T 2 µk
B
= .
2~ g ∆ρ + σk2
k
Thus,
X kB T
F = −kB T ln , (4.326)
2~Ωk
k
where10 1/2
g ∆ρ + σk2
Ωk = . (4.327)
µk
is the normal mode frequency for surface oscillations at wavevector k. For deep water waves, it is appropriate to
take µk = ∆ρ |k|, where ∆ρ = ρL − ρG ≈ ρL is the difference between the densities of water and air.
It is now easy to compute the thermal average
Z Z
2 2 2 2
|zk |2 = d2zk |zk |2 e−(g ∆ρ+σk ) |zk | /kB T d2zk e−(g ∆ρ+σk ) |zk | /kB T
(4.328)
kB T
= .
g ∆ρ + σk2
Note that this result does not depend on µk , i.e. on our choice of kinetic energy. One defines the correlation function
Z 2
1 X
dk kB T
C(x) ≡ z(x) z(0) = |zk |2 eik·x = eik·x
Lx Ly (2π)2 g ∆ρ + σk2
k
Z∞ (4.329)
k T eik|x| k T
= B dq p = B K0 |x|/ξ ,
4πσ q2 + ξ2 4πσ
0
10 Note that there is no prime on the k sum for F , as we have divided the logarithm of Z by two and replaced the half sum by the whole
sum.
180 CHAPTER 4. STATISTICAL ENSEMBLES
p
where ξ = g ∆ρ/σ is the correlation length, and where K0 (z) is the Bessel function of imaginary argument. The
asymptotic behavior of K0 (z) for small z is K0 (z) ∼ ln(2/z), whereas for large z one has K0 (z) ∼ (π/2z)1/2 e−z . We
see that on large length scales the correlations decay exponentially, but on small length scales they diverge. This
divergence is due to the improper energetics we have assigned to short wavelength fluctuations of the interface.
Roughly, it can cured by imposing a cutoff on the integral, or by insisting that the shortest distance scale is a
molecular diameter.
nH = (1 − x) n , np = xn , ne = xn . (4.332)
where g is the degeneracy and εint the internal energy for a given species. We have εint = 0 for p and e, and
εint = −∆ for H, where ∆ = e2 /2aB = 13.6 eV, the binding energy of hydrogen. Neglecting hyperfine splittings11 ,
we have gH = 4, while ge = gp = 2 because each has spin S = 12 . Thus, the associated grand potentials are
ΩH = −gH V kB T λ−3
T,H e
(µH +∆)/kB T
(4.334)
Ωp = −gp V kB T λ−3
T,p e
µp /kB T
(4.335)
Ωe = −ge V kB T λ−3
T,e e
µe /kB T
, (4.336)
where s
2π~2
λT,a = (4.337)
ma kB T
which yields
x2
nλ̃3T = e−∆/kB T , (4.341)
1−x
p
where λ̃T = 2π~2 /m∗ kB T , with m∗ = mp me /mH ≈ me . Note that
s s
4πmH ∆
λ̃T = aB , (4.342)
mp kB T
where T0 = ∆/kB = 1.578×105 K and ν = na3B . Consider for example a temperature T = 3000 K, for which T0 /T =
52.6, and assume that x = 12 . We then find ν = 1.69×10−27, corresponding to a density of n = 1.14×10−2 cm−3 . At
this temperature, the fraction of hydrogen molecules in their first excited (2s) state is x′ ∼ e−T0 /2T = 3.8 × 10−12 .
This is quite striking: half the hydrogen atoms are completely dissociated, which requires an energy of ∆, yet
the number in their first excited state, requiring energy 12 ∆, is twelve orders of magnitude smaller. The student
should reflect on why this can be the case.
182 CHAPTER 4. STATISTICAL ENSEMBLES
Chapter 5
5.1 References
– F. Reif, Fundamentals of Statistical and Thermal Physics (McGraw-Hill, 1987)
This has been perhaps the most popular undergraduate text since it first appeared in 1967, and with good
reason.
– A. H. Carter, Classical and Statistical Thermodynamics
(Benjamin Cummings, 2000)
A very relaxed treatment appropriate for undergraduate physics majors.
– D. V. Schroeder, An Introduction to Thermal Physics (Addison-Wesley, 2000)
This is the best undergraduate thermodynamics book I’ve come across, but only 40% of the book treats
statistical mechanics.
– C. Kittel, Elementary Statistical Physics (Dover, 2004)
Remarkably crisp, though dated, this text is organized as a series of brief discussions of key concepts and
examples. Published by Dover, so you can’t beat the price.
– R. K. Pathria, Statistical Mechanics (2nd edition, Butterworth-Heinemann, 1996)
This popular graduate level text contains many detailed derivations which are helpful for the student.
– M. Plischke and B. Bergersen, Equilibrium Statistical Physics (3rd edition, World Scientific, 2006)
An excellent graduate level text. Less insightful than Kardar but still a good modern treatment of the subject.
Good discussion of mean field theory.
– E. M. Lifshitz and L. P. Pitaevskii, Statistical Physics (part I, 3rd edition, Pergamon, 1980)
This is volume 5 in the famous Landau and Lifshitz Course of Theoretical Physics . Though dated, it still
contains a wealth of information and physical insight.
183
184 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
where n̂α is the number of particles in the quantum state α with energy εα . This form is called the second quantized
representation of the Hamiltonian. The number eigenbasis is therefore also an energy eigenbasis.
Any eigenstate of
Ĥ may be labeled by the integer eigenvalues of the n̂α number operators, and written as n1 , n2 , . . . . We then
have
n̂α ~n = nα ~n (5.2)
and
X
Ĥ ~n = nα εα ~n . (5.3)
α
The eigenvalues nα take on different possible values depending on whether the constituent particles are bosons or
fermions, viz.
bosons : nα ∈ 0 , 1 , 2 , 3 , . . . (5.4)
fermions : nα ∈ 0 , 1 . (5.5)
In other words, for bosons, the occupation numbers are nonnegative integers. For fermions, the occupation num-
bers are either 0 or 1 due to the Pauli principle, which says that at most one fermion can occupy any single particle
quantum state. There is no Pauli principle for bosons.
The N -particle partition function ZN is then
X P
ZN = e−β α nα εα
δN,Pα nα , (5.6)
{nα }
where the sum is over all allowed values of the set {nα }, which depends on the statistics of the particles. Bosons
satisfy Bose-Einstein (BE) statistics, in which nα ∈ {0 , 1 , 2 , . . .}. Fermions satisfy Fermi-Dirac (FD) statistics, in
which nα ∈ {0 , 1}.
P
The OCE partition sum is difficult to perform, owing to the constraint α nα = N on the total number of particles.
This constraint is relaxed in the GCE, where
X
Ξ= eβµN ZN
N
X P P
= e−β α nα εα
eβµ α nα
{nα }
!
Y X
−β(εα −µ) nα
= e . (5.7)
α nα
Note that the grand partition function Ξ takes the form of a product over contributions from the individual single
particle states.
5.2. STATISTICAL MECHANICS OF NONINTERACTING QUANTUM SYSTEMS 185
Therefore we have
Y 1
ΞBE = (5.10)
α
1− e−(εα −µ)/kB T
X
ΩBE = kB T ln 1 − e−(εα −µ)/kB T (5.11)
α
and
Y
ΞFD = 1 + e−(εα −µ)/kB T (5.12)
α
X
ΩFD = −kB T ln 1 + e−(εα −µ)/kB T . (5.13)
α
where we take the upper sign for Bose-Einstein statistics and the lower sign for Fermi-Dirac statistics. Note that
the average occupancy of single particle state α is
∂Ω 1
hn̂α i = = (ε −µ)/k T , (5.15)
∂εα e α B ∓1
and the total particle number is then
X 1
N (T, V, µ) = . (5.16)
α
e(εα −µ)/kB T ∓1
We will henceforth write nα (µ, T ) = hn̂α i for the thermodynamic average of this occupancy.
This is the Maxwell-Boltzmann limit of quantum statistical mechanics. The occupation number average is then
in this limit.
186 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
The single particle density of states per unit volume g(ε) is defined as
1 X
g(ε) = δ(ε − εα ) . (5.19)
V α
In order to obtain g(ε) as a function of the energy ε one must invert the dispersion relation ε = ε(k) to obtain
k = k(ε).
Note that we can equivalently write
ddk g Ωd d−1
g(ε) dε = g d
= k dk (5.23)
(2π) (2π)d
to derive g(ε).
For a spin-S particle with ballistic dispersion ε(k) = ~2 k2 /2m, we have
d/2
2S +1 m d
g(ε) = ε 2 −1 Θ(ε) , (5.24)
Γ(d/2) 2π~2
where Θ(ε) is the step function, which takes the value 0 for ε < 0 and 1 for ε ≥ 0. The appearance of Θ(ε) simply
says that all the single particle energy eigenvalues are nonnegative. Note that we are assuming a box of volume
V but we are ignoring the quantization of kinetic energy, and assuming that the difference between successive
quantized single particle energy eigenvalues is negligible so that g(ε) can be replaced by the average in the above
expression. Note that
1
n(ε, T, µ) = (ε−µ)/k T . (5.25)
e B ∓1
This result holds true independent of the form of g(ε). The average total number of particles is then
Z∞
1
N (T, V, µ) = V dε g(ε) , (5.26)
e(ε−µ)/kB T ∓1
−∞
5.3. QUANTUM IDEAL GASES : LOW DENSITY EXPANSIONS 187
Z∞
g(ε)
n(T, z) = dε
z −1 eε/kB T ∓ 1
−∞
(5.27)
X∞
j−1 j
= (±1) Cj (T ) z ,
j=1
Z∞
Cj (T ) = dε g(ε) e−jε/kB T . (5.28)
−∞
From Ω = −pV and our expression above for Ω(T, V, µ), we have
Z∞
p(T, z) = ∓ kB T dε g(ε) ln 1 ∓ z e−ε/kB T
−∞
(5.29)
∞
X
= kB T (±1)j−1 j −1 Cj (T ) z j .
j=1
Eqns. 5.27 and 5.29 express n(T, z) and p(T, z) as power series in the fugacity z, with T -dependent coefficients. In
principal, we can eliminate z using eqn. 5.27, writing z = z(T, n) as a power series in the number density n, and
substitute this into eqn. 5.29 to obtain an equation of state p = p(T, n) of the form
p(T, n) = n kB T 1 + B2 (T ) n + B3 (T ) n2 + . . . . (5.30)
Note that the low density limit n → 0 yields the ideal gas law independent of the density of states g(ε). This
follows from expanding n(T, z) and p(T, z) to lowest order in z, yielding n = C1 z+O(z 2 ) and p = kB T C1 z+O(z 2 ).
Dividing the second of these equations by the first yields p = n kB T + O(n2 ), which is the ideal gas law. Note that
z = n/C1 + O(n2 ) can formally be written as a power series in n.
Unfortunately, there is no general analytic expression for the virial coefficients Bj (T ) in terms of the expansion
coefficients nj (T ). The only way is to grind things out order by order in our expansions. Let’s roll up our sleeves
and see how this is done. We start by formally writing z(T, n) as a power series in the density n with T -dependent
coefficients Aj (T ):
z = A1 n + A2 n2 + A3 n3 + . . . . (5.31)
188 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
n = C1 z ± C2 z 2 + C3 z 3 + . . . (5.32)
2 3
2 3
2
= C1 A1 n + A2 n + A3 n + . . . ± C2 A1 n + A2 n + A3 n + . . .
3
+ C3 A1 n + A2 n2 + A3 n3 + . . . + . . . .
In order for this equation to be true we require that the coefficient of n on the RHS be unity, and that the coefficients
of nj for all j > 1 must vanish. Thus,
C1 A1 = 1
C1 A2 ± C2 A21 = 0 (5.34)
C1 A3 ± 2C2 A1 A2 + C3 A31 =0.
C2
A2 = ∓ . (5.36)
C13
Next, insert the expressions for A1 and A2 into the third equation to obtain A3 :
2C22 C
A3 = − 34 . (5.37)
C15 C1
= n + B2 n2 + B3 n3 + . . .
It is easy to derive the general result that BjF = (−1)j−1 BjB , where the superscripts denote Fermi (F) or Bose (B)
statistics.
5.3. QUANTUM IDEAL GASES : LOW DENSITY EXPANSIONS 189
We remark that the equation of state for classical (and quantum) interacting systems also can be expanded in terms
of virial coefficients. Consider, for example, the van der Waals equation of state,
aN 2
p+ 2 V − N b) = N kB T . (5.41)
V
This may be recast as
nkB T
p= − an2
1 − bn (5.42)
= nkB T + b kB T − a n2 + kB T b2 n3 + kB T b3 n4 + . . . ,
where n = N/V . Thus, for the van der Waals system, we have B2 = (b kB T − a) and Bk = kB T bk−1 for all k ≥ 3.
For the ballistic dispersion ε(p) = p2 /2m we computed the density of states in eqn. 5.24. One finds
Z∞
g λ−d d
Cj (T ) = S T dt t 2 −1 e−jt = gS λ−d
T j
−d/2
. (5.43)
Γ(d/2)
0
We then have
Note that B2 (T ) is negative for bosons and positive for fermions. This is because bosons have a tendency to bunch
and under certain circumstances may exhibit a phenomenon known as Bose-Einstein condensation (BEC). Fermions,
on the other hand, obey the Pauli principle, which results in an extra positive correction to the pressure in the low
density limit.
We may also write
n(T, z) = ±gS λ−d
T Li d (±z) (5.46)
2
and
p(T, z) = ±gS kB T λ−d
T Li d +1 (±z) , (5.47)
2
where
X∞
zn
Liq (z) ≡ (5.48)
n=1
nq
is the polylogarithm function1 . Note that Liq (z) obeys a recursion relation in its index, viz.
∂
z Li (z) = Liq−1 (z) , (5.49)
∂z q
and that
X∞
1
Liq (1) = = ζ(q) . (5.50)
n=1
nq
1 Several texts, such as Pathria and Reichl, write gq (z) for Liq (z). I adopt the latter notation since we are already using the symbol g for the
density of states function g(ε) and for the internal degeneracy g.
190 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
Suppose we are to partition N particles among J possible distinct single particle states. How many ways Ω are
there of accomplishing this task?
The answer depends on the statistics of the particles. If the particles are fermions,
J
the answer is easy: ΩFD = N . For bosons, the number of possible partitions can be evaluated via the following
argument. Imagine that we line up all the N particles in a row, and we place J − 1 barriers among the particles,
as shown below in Fig. 5.1. The number of partitions is then the total number of ways of placing the N particles
among these N + J − 1 objects (particles plus barriers), hence we have ΩBE = N +J−1 N . For Maxwell-Boltzmann
statistics, we take ΩMB = J N /N ! Note that ΩMB is not necessarily an integer, so Maxwell-Boltzmann statistics
does not represent any actual state counting. Rather, it manifests itself as a common limit of the Bose and Fermi
distributions, as we have seen and shall see again shortly.
Figure 5.1: Partitioning N bosons into J possible states (N = 14 and J = 5 shown). The N black dots represent
bosons, while the J − 1 white dots represent markers separating the different single particle populations. Here
n1 = 3, n2 = 1, n3 = 4, n4 = 2, and n5 = 4.
The entropy in each case is simply S = kB ln Ω. We assume N ≫ 1 and J ≫ 1, with n ≡ N/J finite. Then using
Stirling’s approximation, ln(K!) = K ln K − K + O(ln K), we have
SMB = −JkB n ln n
SBE = −JkB n ln n − (1 + n) ln(1 + n) (5.51)
SFD = −JkB n ln n + (1 − n) ln(1 − n) .
In the Maxwell-Boltzmann limit, n ≪ 1, and all three expressions agree. Note that
∂SMB
= −kB 1 + ln n
∂N J
∂SBE
= kB ln n−1 + 1 (5.52)
∂N J
∂SFD
= kB ln n−1 − 1 .
∂N J
Now let’s imagine grouping the single particle spectrum into intervals of J consecutive energy states. If J is finite
and the spectrum is continuous and we are in the thermodynamic limit, then these states will all be degenerate.
Therefore, using α as a label for the energies, we have that the grand potential Ω = E − T S − µN is given in each
case by
Xh i
ΩMB = J (εα − µ) nα + kB T nα ln nα
α
Xh i
ΩBE = J (εα − µ) nα + kB T nα ln nα − kB T (1 + nα ) ln(1 + nα ) (5.53)
α
Xh i
ΩFD = J (εα − µ) nα + kB T nα ln nα + kB T (1 − nα ) ln(1 − nα ) .
α
Now - lo and behold! - treating Ω as a functionPof the distribution {nα } and extremizing in each case, subject
to the constraint of total particle number N = J α nα , one obtains the Maxwell-Boltzmann, Bose-Einstein, and
5.5. PHOTON STATISTICS 191
As long as J is finite, so the states in each block all remain at the same energy, the results are independent of J.
There exists a certain class of particles, including photons and certain elementary excitations in solids such as
phonons (i.e. lattice vibrations) and magnons (i.e. spin waves) which obey bosonic statistics but with zero chemical
potential. This is because their overall number is not conserved (under typical conditions) – photons can be
emitted and absorbed by the atoms in the wall of a container, phonon and magnon number is also not conserved
due to various processes, etc. In such cases, the free energy attains its minimum value with respect to particle
number when
∂F
µ= =0. (5.55)
∂N T.V
The number distribution, from eqn. 5.15, is then
1
n(ε) = . (5.56)
eβε − 1
2g π d/2 εd−1
g(ε) = Θ(ε) . (5.59)
Γ(d/2) (hc)d
In d = 3 dimensions the degeneracy is g = 2, the number of independent polarization states. The pressure p(T ) is
then obtained using Ω = −pV . We have
Z∞
p(T ) = −kB T dε g(ε) ln 1 − e−ε/kB T
−∞
Z∞
2 g π d/2
=− (hc) kB T dε εd−1 ln 1 − e−ε/kB T
−d
(5.60)
Γ(d/2)
0
Z∞
2 g π d/2 (kB T ) d+1
=− dt td−1 ln 1 − e−t .
Γ(d/2) (hc)d
0
Z∞
Id ≡ − dt td−1 ln 1 − e−t
0
∞
X Z∞
1 (5.61)
= dt td−1 e−nt
n=1
n
0
∞
X 1
= Γ(d) = Γ(d) ζ(d + 1) .
n=1
nd+1
Finally, we invoke a result from the mathematics of the gamma function known as the doubling formula,
2z−1 z
z+1
Γ(z) = √ Γ 2 Γ 2 . (5.62)
π
− 12 (d+1) d+1
(kB T )d+1
p(T ) = g π Γ 2 ζ(d + 1) . (5.63)
(~c)d
Z∞
g(ε)
n(T ) = dε
eε/kB T −1
−∞ (5.64)
d
− 12 (d+1) d+1
kB T
= gπ Γ 2 ζ(d) .
~c
5.5. PHOTON STATISTICS 193
A number of thermodynamic properties of the photon gas can be determined from purely classical arguments.
Here we recapitulate a few important ones.
1. Suppose our photon gas is confined to a rectangular box of dimensions Lx × Ly × Lz . Suppose further
that the dimensions are all expanded by a factor λ1/3 , i.e. the volume is isotropically expanded by a factor
of λ. The cavity modes of the electromagnetic radiation have quantized wavevectors, even within classical
electromagnetic theory, given by
2πnx 2πny 2πnz
k= , , . (5.71)
Lx Ly Lz
Since the energy for a given mode is ε(k) = ~c|k|, we see that the energy changes by a factor λ−1/3 under an
adiabatic volume expansion V → λV , where the distribution of different electromagnetic mode occupancies
remains fixed. Thus,
∂E ∂E
V =λ = − 31 E . (5.72)
∂V S ∂λ S
Thus,
∂E E
p=− = , (5.73)
∂V S 3V
as we found in eqn. 5.69. Since E = E(T, V ) is extensive, we must have p = p(T ) alone.
194 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
Zπ/2 Z2π
dP cE cE
= dθ dφ sin θ · cos θ = = 3
4 c p(T ) ≡ σ T 4 , (5.77)
dA 4πV 4V
0 0
where σ = 43 cA, with p(T ) = A T 4 as we found above. From quantum statistical mechanical considerations,
we have
π 2 kB4 W
σ= = 5.67 × 10−8 2 4 (5.78)
60 c2 ~3 m K
is Stefan’s constant.
We derived the result P = σT 4 · A where σ = 5.67 × 10−8 W/m2 K4 for the power emitted by an electromagnetic
‘black body’. Let’s apply this result to the earth-sun system. We’ll need three lengths: the radius of the sun
R⊙ = 6.96 × 108 m, the radius of the earth Re = 6.38 × 106 m, and the radius of the earth’s orbit ae = 1.50 × 1011 m.
Let’s assume that the earth has achieved a steady state temperature of Te . We balance the total power incident
upon the earth with the power radiated by the earth. The power incident upon the earth is
Figure 5.2: Spectral density ρε (ν, T ) for blackbody radiation at three temperatures.
Recall that the frequency of an electromagnetic wave of wavevector k is ν = c/λ = ck/2π. Therefore the number
of photons NT (ν, T ) per unit frequency in thermodynamic equilibrium is (recall there are two polarization states)
2V d3k V k 2 dk
N (ν, T ) dν = 3
· ~ck/k T = 2 · ~ck/k T . (5.82)
8π e B −1 π e B −1
We therefore have
8πV ν2
N (ν, T ) = · . (5.83)
c3 ehν/kB T − 1
Since a photon of frequency ν carries energy hν, the energy per unit frequency E(ν) is
8πhV ν3
E(ν, T ) = 3
· hν/k T . (5.84)
c e B −1
Note what happens if Planck’s constant h vanishes, as it does in the classical limit. The denominator can then be
written
hν
ehν/kB T − 1 = + O(h2 ) (5.85)
kB T
and
8πkB T 2
ECL (ν, T ) = lim E(ν) = V · ν . (5.86)
h→0 c3
In classical electromagnetic theory, then, the total energy integrated over all frequencies diverges. This is known
as the ultraviolet catastrophe, since the divergence comes from the large ν part of the integral, which in the optical
spectrum is the ultraviolet portion. With quantization, the Bose-Einstein factor imposes an effective ultraviolet
cutoff kB T /h on the frequency integral, and the total energy, as we found above, is finite:
Z∞
π 2 (kB T )4
E(T ) = dν E(ν) = 3pV = V · . (5.87)
15 (~c)3
0
196 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
We saw in eqn. 5.76 that the power emitted per unit surface area by a blackbody is σT 4 . The power law here
follows from the ultrarelativistic dispersion ε = ~ck of the photons. Suppose that we replace this dispersion with
the general form ε = ε(k). Now consider a large box in equilibrium at temperature T . The energy current incident
on a differential area dA of surface normal to ẑ is
Z 3
dk 1 ∂ε(k) 1
dP = dA · Θ(cos θ) · ε(k) · · . (5.90)
(2π)3 ~ ∂kz eε(k)/kB T − 1
Let us assume an isotropic power law dispersion of the form ε(k) = Ck α . Then after a straightforward calculation
we obtain
dP 2
= σ T 2+ α , (5.91)
dA
where 2
2+ α 2
2
2
g kB C− α
σ =ζ 2+ α Γ 2+ α · . (5.92)
8π 2 ~
One can check that for g = 2, C = ~c, and α = 1 that this result reduces to that of eqn. 5.78.
Crystalline solids support propagating waves called phonons, which are quantized vibrations of the lattice. Recall
p2
that the quantum mechanical Hamiltonian for a single harmonic oscillator, Ĥ = 2m + 21 mω02 q 2 , may be written as
1
Ĥ = ~ω0 (a a + 2 ), where a and a are ‘ladder operators’ satisfying commutation relations a , a† = 1.
† †
Consider the linear chain of masses and springs depicted in fig. 5.3. We assume that our system consists of N mass
points on a large ring of circumference L. In equilibrium, the masses are spaced evenly by a distance b = L/N .
That is, x0n = nb is the equilibrium position of particle n. We define un = xn − x0n to be the difference between the
position of mass n and The Hamiltonian is then
X p2
n 1 2
Ĥ = + κ (xn+1 − xn − a)
n
2m 2
X p2 (5.93)
n 1 2 1 2
= + κ (un+1 − un ) + 2 N κ(b − a) ,
n
2m 2
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 197
where a is the unstretched length of each spring, m is the mass of each mass point, κ is the force constant of each
spring, and N is the total number of mass points. If b 6= a the springs are under tension in equilibrium, but as we
see this only leads to an additive constant in the Hamiltonian, and hence does not enter the equations of motion.
The classical equations of motion are
∂ Ĥ p
u̇n = = n (5.94)
∂pn m
∂ Ĥ
ṗn = − = κ un+1 + un−1 − 2un . (5.95)
∂un
Taking the time derivative of the first equation and substituting into the second yields
κ
ün = u + un−1 − 2un . (5.96)
m n+1
We now write
1 X
un = √ ũk eikna , (5.97)
N k
where periodicity uN +n = un requires that the k values are quantized so that eikN a = 1, i.e. k = 2πj/N a where
j ∈ {0, 1, . . . , N −1}. The inverse of this discrete Fourier transform is
1 X
ũk = √ un e−ikna . (5.98)
N n
Note that ũk is in general complex, but that ũ∗k = ũ−k . In terms of the ũk , the equations of motion take the form
2κ
ũ¨k = − 1 − cos(ka) ũk ≡ −ωk2 ũk . (5.99)
m
Thus, each ũk is a normal mode, and the normal mode frequencies are
r
κ
ωk = 2 sin 21 ka . (5.100)
m
The density of states for this band of phonon excitations is
Zπ/a
dk
g(ε) = δ(ε − ~ωk )
2π
−π/a (5.101)
2 −1/2
= J 2 − ε2 Θ(ε) Θ(J − ε) ,
πa
p
where J = 2~ κ/m is the phonon bandwidth. The step functions require 0 ≤ ε ≤ J; outside this range there are
no phonon energy levels and the density of states accordingly vanishes.
The entire theory can be quantized, taking pn , un′ = −i~δnn′ . We then define
1 X 1 X
pn = √ p̃k eikna , p̃k = √ pn e−ikna , (5.102)
N k N n
in which case p̃k , ũk′ = −i~δkk′ . Note that ũ†k = ũ−k and p̃†k = p̃−k . We then define the ladder operator
1/2 1/2
1 mωk
ak = p̃k − i ũk (5.103)
2m~ωk 2~
198 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
Figure 5.3: A linear chain of masses and springs. The black circles represent the equilibrium positions of the
masses. The displacement of mass n relative to its equilibrium value is un .
The most general model of a harmonic solid is described by a Hamiltonian of the form
X p2 (R) 1 XX X α
Ĥ = i
+ ui (R) Φαβ ′ β ′
ij (R − R ) uj (R ) , (5.105)
2Mi 2 i,j
R,i ′α,β R,R
where U is the potential energy of interaction among all the atoms. Here we have simply expanded the potential
to second order in the local displacements uα i (R). The lattice sites R are elements of a Bravais lattice. The indices i
and j specify basis elements with respect to this lattice, and the indices α and β range over {1, . . . , d}, the number
of possible directions in space. The subject of crystallography is beyond the scope of these notes, but, very briefly,
a Bravais lattice in d dimensions is specified by a set of d linearly independent primitive direct lattice vectors al , such
that any point in the Bravais lattice may be written as a sum over the primitive vectors with integer coefficients:
P
R = dl=1 nl al . The set of all such vectors {R} is called the direct lattice. The direct lattice is closed under the
operation of vector addition: if R and R′ are points in a Bravais lattice, then so is R + R′ .
A crystal is a periodic arrangement of lattice sites. The fundamental repeating unit is called the unit cell. Not
every crystal is a Bravais lattice, however. Indeed, Bravais lattices are special crystals in which there is only one
atom per unit cell. Consider, for example, the structure in fig. 5.4. The blue dots form a square Bravais lattice
with primitive direct lattice vectors a1 = a x̂ and a2 = a ŷ, where a is the lattice constant, which is the distance
between any neighboring pair of blue dots. The red squares and green triangles, along with the blue dots, form
a basis for the crystal structure which label each sublattice. Our crystal in fig. 5.4 is formally classified as a square
Bravais lattice with a three element basis. To specify an arbitrary site in the crystal, we must specify both a direct
lattice vector R as well as a basis index j ∈ {1, . . . , r}, so that the location is R + ηj . The vectors {ηj } are the basis
vectors for our crystal structure. We see that a general crystal structure consists of a repeating unit, known as a
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 199
Figure 5.4: A crystal structure with an underlying square Bravais lattice and a three element basis.
unit cell. The centers (or corners, if one prefers) of the unit cells form a Bravais lattice. Within a given unit cell, the
individual sublattice sites are located at positions ηj with respect to the unit cell position R.
Upon diagonalization, the Hamiltonian of eqn. 5.105 takes the form
X
Ĥ = ~ωa (k) A†a (k) Aa (k) + 12 , (5.107)
k,a
where
Aa (k) , A†b (k′ ) = δab δkk′ . (5.108)
The eigenfrequencies are solutions to the eigenvalue equation
X αβ (a) (a)
Φ̃ij (k) ejβ (k) = Mi ωa2 (k) eiα (k) , (5.109)
j,β
where X
Φ̃αβ
ij (k) = Φαβ
ij (R) e
−ik·R
. (5.110)
R
Here, k lies within the first Brillouin zone, which is the unit cell of the reciprocal lattice of points G satisfying
eiG·R = 1 for all G and R. The reciprocal lattice is also a Bravais lattice, with primitive reciprocal lattice vectors
Pd
bl , such that any point on the reciprocal lattice may be written G = l=1 ml bl . One also has that al · bl′ = 2πδll′ .
(a)
The index a ranges from 1 to d · r and labels the mode of oscillation at wavevector k. The vector eiα (k) is the
polarization vector for the ath phonon branch. In solids of high symmetry, phonon modes can be classified as
longitudinal or transverse excitations.
For a crystalline lattice with an r-element basis, there are then d · r phonon modes for each wavevector k lying
in the first Brillouin zone. If we impose periodic boundary conditions, then the k points within the first Brillouin
200 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
zone are themselves quantized, as in the d = 1 case where we found k = 2πn/N . There are N distinct k points in
the first Brillouin zone – one for every direct lattice site. The total number of modes is than d · r · N , which is the
total number of translational degrees of freedom in our system: rN total atoms (N unit cells each with an r atom
basis) each free to vibrate in d dimensions. Of the d · r branches of phonon excitations, d of them will be acoustic
modes whose frequency vanishes as k → 0. The remaining d(r − 1) branches are optical modes and oscillate at finite
frequencies. Basically, in an acoustic mode, for k close to the (Brillouin) zone center k = 0, all the atoms in each
unit cell move together in the same direction at any moment of time. In an optical mode, the different basis atoms
move in different directions.
There is no number conservation law for phonons – they may be freely created or destroyed in anharmonic pro-
cesses, where two photons with wavevectors k and q can combine into a single phonon with wavevector k + q,
and vice versa. Therefore the chemical potential for phonons is µ = 0. We define the density of states ga (ω) for the
ath phonon mode as
Z d
1 X dk
ga (ω) = δ ω − ωa (k) = V0 δ ω − ωa (k) , (5.111)
N (2π)d
k BZ
where N is the number of unit cells, V0 is the unit cell volume of the direct lattice, and the k sum and integral are
over the first Brillouin zone only. Note that ω here has dimensions of frequency. The functions ga (ω) is normalized
to unity:
Z∞
dω ga (ω) = 1 . (5.112)
0
The total phonon density of states per unit cell is given by2
dr
X
g(ω) = ga (ω) . (5.113)
a=1
∂Ω
Note that V = N V0 since there are N unit cells, each of volume V0 . The entropy is given by S = − ∂T V and thus
the heat capacity is
Z∞ 2
∂ 2Ω ~ω 2 ~ω
CV = −T = N kB dω g(ω) csch (5.115)
∂T 2 2kB T 2kB T
0
~ω
2kB T
Note that as T → ∞ we have csch 2kB T → ~ω , and therefore
Z∞
lim CV (T ) = N kB dω g(ω) = rdN kB . (5.116)
T →∞
0
2 Note the dimensions of g(ω) are (frequency)−1 . By contrast, the dimensions of g(ε) in eqn. 5.24 are (energy)−1 · (volume)−1 . The
difference lies in the a factor of V0 · ~, where V0 is the unit cell volume.
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 201
Figure 5.5: Upper panel: phonon spectrum in elemental rhodium (Rh) at T = 297 K measured by high precision
inelastic neutron scattering (INS) by A. Eichler et al., Phys. Rev. B 57, 324 (1998). Note the three acoustic branches
and no optical branches, corresponding to d = 3 and r = 1. Lower panel: phonon spectrum in gallium arsenide
(GaAs) at T = 12 K, comparing theoretical lattice-dynamical calculations with INS results of D. Strauch and B.
Dorner, J. Phys.: Condens. Matter 2, 1457 (1990). Note the three acoustic branches and three optical branches,
corresponding to d = 3 and r = 2. The Greek letters along the x-axis indicate points of high symmetry in the
Brillouin zone.
This is the classical Dulong-Petit limit of 12 kB per quadratic degree of freedom; there are rN atoms moving in d
dimensions, hence d · rN positions and an equal number of momenta, resulting in a high temperature limit of
CV = rdN kB .
HIstorically, two models of lattice vibrations have received wide attention. First is the so-called Einstein model, in
which there is no dispersion to the individual phonon modes. We approximate ga (ω) ≈ δ(ω − ωa ), in which case
X ~ω 2
~ωa
a 2
CV (T ) = N kB csch . (5.117)
a
2kB T 2kB T
202 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
~ωa
At low temperatures, the contribution from each branch vanishes exponentially, since csch2 2kB T ≃ 4 e−~ωa /kB T →
0. Real solids don’t behave this way.
A more realistic model. due to Debye, accounts for the low-lying acoustic phonon branches. Since the acoustic
phonon dispersion vanishes linearly with |k| as k → 0, there is no temperature at which the acoustic phonons
‘freeze out’ exponentially, as in the case of Einstein phonons. Indeed, the Einstein model is appropriate in describ-
ing the d (r−1) optical phonon branches, though it fails miserably for the acoustic branches.
In the vicinity of the zone center k = 0 (also called Γ in crystallographic notation) the d acoustic modes obey a
linear dispersion, with ωa (k) = ca (k̂) k. This results in an acoustic phonon density of states in d = 3 dimensions
of
Z
V ω 2 X dk̂ 1
g̃(ω) = 0 2 Θ(ωD − ω)
2π 4π c3a (k)
a (5.118)
3V0 2
= ω Θ(ωD − ω) ,
2π 2 c̄3
where c̄ is an average acoustic phonon velocity (i.e. speed of sound) defined by
X Z
3 dk̂ 1
= (5.119)
c̄3 a
4π c3a (k)
and ωD is a cutoff known as the Debye frequency. The cutoff is necessary because the phonon branch does not
extend forever, but only to the boundaries of the Brillouin zone. Thus, ωD should roughly be equal to the energy
of a zone boundary phonon. Alternatively, we can define ωD by the normalization condition
Z∞
dω g̃(ω) = 3 =⇒ ωD = (6π 2 /V0 )1/3 c̄ . (5.120)
0
This allows us to write g̃(ω) = 9ω 2 /ωD3 Θ(ωD − ω).
The specific heat due to the acoustic phonons is then
ZωD 2
9N kB ~ω ~ω
CV (T ) = dω ω 2 csch2
ωD3 2kB T 2kB T
0 (5.121)
3
2T
= 9N kB φ ΘD /2T ,
ΘD
Therefore, 3
12π 4 T
T ≪ ΘD
5 N kB ΘD
CV (T ) = (5.123)
3N kB T ≫ ΘD .
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 203
Element Ag Al Au C Cd Cr Cu Fe Mn
ΘD (K) 227 433 162 2250 210 606 347 477 409
Tmelt (K) 962 660 1064 3500 321 1857 1083 1535 1245
Element Ni Pb Pt Si Sn Ta Ti W Zn
ΘD (K) 477 105 237 645 199 246 420 383 329
Tmelt (K) 1453 327 1772 1410 232 2996 1660 3410 420
Table 5.1: Debye temperatures (at T = 0) and melting points for some common elements (carbon is assumed to be
diamond and not graphite). (Source: the internet!)
Thus, the heat capacity due to acoustic phonons obeys the Dulong-Petit rule in that CV (T → ∞) = 3N kB , cor-
responding to the three acoustic degrees of freedom per unit cell. The remaining contribution of 3(r − 1)N kB to
the high temperature heat capacity comes from the optical modes not considered in the Debye model. The low
temperature T 3 behavior of the heat capacity of crystalline solids is a generic feature, and its detailed description
is a triumph of the Debye model.
Here we sum over all wavevectors k in the first Brilliouin zone, and over all normal modes a. There are dr normal
modes per unit cell i.e. d branches of the phonon dispersion ωa (k). (For the one-dimensional chain with d = 1 and
r = 1 there was only one such branch to consider). Note also the quantity Mia (k), which has units of mass and is
(a)
defined in terms of the polarization vectors eiα (k) as
d
X
1 (a) 2
= e (k) . (5.127)
Mia (k) µ=1 iµ
204 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
The dimensions of the polarization vector are [mass]−1/2 , since the generalized orthonormality condition on the
normal modes is X (a) ∗ (b)
Mi eiµ (k) eiµ (k) = δ ab , (5.128)
i,µ
where Mi is the mass of the atom of species i within the unit cell (i ∈ {1, . . . , r}). For our purposes we can replace
Mia (k) by an appropriately averaged quantity which we call Mi ; this ‘effective mass’ is then independent of the
mode index a as well as the wavevector k. We may then write
Z∞ ( )
~ 1 1
h u2i i ≈ dω g(ω) · + , (5.129)
Mi ω e~ω/kB T − 1 2
0
where we have dropped the site label R since translational invariance guarantees that the fluctuations are the
same from one unit cell to the next. Note that the fluctuations h u2i i can be divided into a temperature-dependent
part h u2i ith and a temperature-independent quantum contribution h u2i iqu , where
Z∞
~ g(ω) 1
h u2i ith = dω · ~ω/k T (5.130)
Mi ω e B −1
0
Z∞
~ g(ω)
h u2i iqu = dω . (5.131)
2Mi ω
0
Let’s evaluate these contributions within the Debye model, where we replace g(ω) by
d2 ω d−1
ḡ(ω) = Θ(ωD − ω) . (5.132)
ωDd
We then find
d−1
d2 ~ kB T
h u2i ith = Fd (~ωD /kB T ) (5.133)
Mi ωD ~ωD
d2 ~
h u2i iqu = · , (5.134)
d − 1 2Mi ωD
where d−2
x
Zx d−2 d−2 x→0
s
Fd (x) = ds s = . (5.135)
e −1
0 ζ(d − 1) x → ∞
1) The T = 0 contribution to the the fluctuations, h u2i iqu , diverges in d = 1 dimensions. Therefore there are no
one-dimensional quantum solids.
2) The thermal contribution to the fluctuations, h u2i ith , diverges for any T > 0 whenever d ≤ 2. This is because
the integrand of Fd (x) goes as sd−3 as s → 0. Therefore, there are no two-dimensional classical solids.
3) Both the above conclusions are valid in the thermodynamic limit. Finite size imposes a cutoff on the fre-
quency integrals, because there is a smallest wavevector kmin ∼ 2π/L, where L is the (finite) linear dimen-
sion of the system. This leads to a low frequency cutoff ωmin = 2πc̄/L, where c̄ is the appropriately averaged
acoustic phonon velocity from eqn. 5.119, which mitigates any divergences.
5.6. LATTICE VIBRATIONS : EINSTEIN AND DEBYE MODELS 205
An old phenomenological theory of melting due to Lindemann says that a crystalline solid melts when the RMS
fluctuations in the atomic positions exceeds a certain fraction η of the lattice constant a. We therefore define the
ratios
h u2i ith ~2 T d−1
x2i,th ≡ 2
= d2
· 2
· · F (ΘD /T ) (5.136)
a Mi a kB ΘDd
h u2i iqu d2 ~2 1
x2i,qu ≡ 2
= · 2
· , (5.137)
a 2(d − 1) Mi a kB ΘD
q p .
with xi = x2i,th + x2i,qu = h u2i i a.
Let’s now work through an example of a three-dimensional solid. We’ll assume a single element basis (r = 1). We
have that
9~2 /4kB
2 = 109 K . (5.138)
1 amu Å
According to table 5.1, the melting temperature always exceeds the Debye temperature, and often by a great
amount. We therefore assume T ≫ ΘD , which puts us in the small x limit of Fd (x). We then find
s
2 Θ⋆ 2 Θ⋆ 4T 4T Θ⋆
xqu = , xth = · , x= 1+ . (5.139)
ΘD ΘD ΘD ΘD ΘD
where
109 K
Θ∗ = 2 . (5.140)
M [amu] · a[Å]
The total position fluctuation is of course the sum x2 = x2i,th + x2i,qu . Consider for example the case of copper, with
M = 56 amu and a = 2.87 Å. The Debye temperature is ΘD = 347 K. From this we find xqu = 0.026, which says
that at T = 0 the RMS fluctuations of the atomic positions are not quite three percent of the lattice spacing (i.e.
the distance between neighboring copper atoms). At room temperature, T = 293 K, one finds xth = 0.048, which
is about twice as large as the quantum contribution. How big are the atomic position fluctuations at the melting
point? According to our table, Tmelt = 1083 K for copper, and from our formulae we obtain xmelt = 0.096. The
Lindemann criterion says that solids melt when x(T ) ≈ 0.1.
We were very lucky to hit the magic number xmelt = 0.1 with copper. Let’s try another example. Lead has
M = 208 amu and a = 4.95 Å. The Debye temperature is ΘD = 105 K (‘soft phonons’), and the melting point is
Tmelt = 327 K. From these data we obtain x(T = 0) = 0.014, x(293 K) = 0.050 and x(T = 327 K) = 0.053. Same
ballpark.
We can turn the analysis around and predict a melting temperature based on the Lindemann criterion x(Tmelt ) = η,
where η ≈ 0.1. We obtain 2
η ΘD Θ
TL = −1 · D . (5.141)
Θ⋆ 4
We call TL the Lindemann temperature. Most treatments of the Lindemann criterion ignore the quantum correction,
which gives the −1 contribution inside the above parentheses. But if we are more careful and include it, we see
that it may be possible to have TL < 0. This occurs for any crystal where ΘD < Θ⋆ /η 2 .
Consider for example the case of 4 He, which at atmospheric pressure condenses into a liquid at Tc = 4.2 K and
remains in the liquid state down to absolute zero. At p = 1 atm, it never solidifies! Why? The number density of
liquid 4 He at p = 1 atm and T = 0 K is 2.2 × 1022 cm−3 . Let’s say the Helium atoms want to form a crystalline
206 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
lattice. We don’t know a priori what the lattice structure will be, so let’s for the sake of simplicity assume a simple
cubic lattice. From the number density we obtain a lattice spacing of a = 3.57 Å. OK now what do we take for
the Debye temperature? Theoretically this should depend on the microscopic force constants which enter the
small oscillations problem (i.e. the spring constants between pairs of helium atoms in equilibrium). We’ll use the
expression we derived for the Debye frequency, ωD = (6π 2 /V0 )1/3 c̄, where V0 is the unit cell volume. We’ll take
c̄ = 238 m/s, which is the speed of sound in liquid helium at T = 0. This gives ΘD = 19.8 K. We find Θ⋆ = 2.13 K,
and if we take η = 0.1 this gives Θ⋆ /η 2 = 213 K, which significantly exceeds ΘD . Thus, the solid should melt
because the RMS fluctuations in the atomic positions at absolute zero are huge: xqu = (Θ⋆ /ΘD )1/2 = 0.33. By
applying pressure, one can get 4 He to crystallize above pc = 25 atm (at absolute zero). Under pressure, the unit
cell volume V0 decreases and the phonon velocity c̄ increases, so the Debye temperature itself increases.
It is important to recognize that the Lindemann criterion does not provide us with a theory of melting per se.
Rather it provides us with a heuristic which allows us to predict roughly when a solid should melt.
The vanishing of the acoustic phonon dispersion at k = 0 is a consequence of Goldstone’s theorem which says that
associated with every broken generator of a continuous symmetry there is an associated bosonic gapless excitation
(i.e. one whose frequency ω vanishes in the long wavelength limit). In the case of phonons, the ‘broken generators’
are the symmetries under spatial translation in the x, y, and z directions. The crystal selects a particular location
for its center-of-mass, which breaks this symmetry. There are, accordingly, three gapless acoustic phonons.
Magnetic materials support another branch of elementary excitations known as spin waves, or magnons. In
isotropic magnets, there is a global symmetry associated with rotations in internal spin space, described by the
group SU(2). If the system spontaneously magnetizes, meaning there is long-ranged ferromagnetic order (↑↑↑
· · · ), or long-ranged antiferromagnetic order (↑↓↑↓ · · · ), then global spin rotation symmetry is broken. Typically
a particular direction is chosen for the magnetic moment (or staggered moment, in the case of an antiferromag-
net). Symmetry under rotations about this axis is then preserved, but rotations which do not preserve the selected
axis are ‘broken’. In the most straightforward case, that of the antiferromagnet, there are two such rotations for
SU(2), and concomitantly two gapless magnon branches, with linearly vanishing dispersions ωa (k). The situation
is more subtle in the case of ferromagnets, because the total magnetization is conserved by the dynamics (unlike
the total staggered magnetization in the case of antiferromagnets). Another wrinkle arises if there are long-ranged
interactions present.
For our purposes, we can safely ignore the deep physical reasons underlying the gaplessness of Goldstone bosons
and simply posit a gapless dispersion relation of the form ω(k) = A |k|σ . The density of states for this excitation
branch is then d
−1
g(ω) = C ω σ Θ(ωc − ω) , (5.142)
where C is a constant and ωc is the cutoff, which is the bandwidth for this excitation branch.3 Normalizing the
density of states for this branch results in the identification ωc = (d/σC)σ/d .
The heat capacity is then found to be
Zωc 2
d
σ −1
~ω 2 ~ω
CV = N kB C dω ω csch
kB T 2kB T
0 (5.143)
d/σ
d 2T
= N kB φ Θ/2T ,
σ Θ
−d d
−σ
3 If ω(k) = Ak σ , then C = 21−d π σ−1 A
2
‹
g Γ(d/2) .
5.7. THE IDEAL BOSE GAS 207
which is a generalization of our earlier results. Once again, we recover Dulong-Petit for kB T ≫ ~ωc , with CV (T ≫
~ωc /kB ) = N kB .
In an isotropic ferromagnet, i.e.a ferromagnetic material where there is full SU(2) symmetry in internal ‘spin’
space, the magnons have a k 2 dispersion. Thus, a bulk three-dimensional isotropic ferromagnet will exhibit a heat
capacity due to spin waves which behaves as T 3/2 at low temperatures. For sufficiently low temperatures this will
overwhelm the phonon contribution, which behaves as T 3 .
Recall that the grand partition function for noninteracting bosons is given by
!
Y X ∞ Y −1
β(µ−εα )nα
Ξ= e = 1 − eβ(µ−εα ) , (5.145)
α nα =0 α
In order for the sum to converge to the RHS above, we must have µ < εα for all single-particle states |αi. The
density of particles is then
Z ∞
1 ∂Ω 1 X 1 g(ε)
n(T, µ) = − = = dε β(ε−µ) , (5.146)
V ∂µ T,V V α eβ(εα −µ) − 1 e −1
ε0
P
where g(ε) = V1 α δ(ε − εα ) is the density of single particle states per unit volume. We assume that g(ε) = 0 for
ε < ε0 ; typically ε0 = 0, as is the case for any dispersion of the form ε(k) = A|k|r , for example. However, in the
presence of a magnetic field, we could have ε(k, σ) = A|k|r − gµ0 Hσ, in which case ε0 = −gµ0 |H|.
Clearly n(T, µ) is an increasing function of both T and µ. At fixed T , the maximum possible value for n(T, µ),
called the critical density nc (T ), is achieved for µ = ε0 , i.e.
Z∞
g(ε)
nc (T ) = dε .
eβ(ε−ε0 ) −1
ε0
The above integral converges provided g(ε0 ) = 0, assuming g(ε) is continuous4 . If g(ε0 ) > 0, the integral diverges,
and nc (T ) = ∞. In this latter case, one can always invert the equation for n(T, µ) to obtain the chemical potential
µ(T, n). In the former case, where the nc (T ) is finite, we have a problem – what happens if n > nc (T ) ?
In the former case, where nc (T ) is finite, we can equivalently restate the problem in terms of a critical temperature
Tc (n), defined by the equation nc (Tc ) = n. For T < Tc , we apparently can no longer invert to obtain µ(T, n), so
4 OK, that isn’t quite true. For example, if g(ε) ∼ 1/ ln ε, then the integral has a very weak ln ln(1/η) divergence, where η is the lower
cutoff. But for any power law density of states g(ε) ∝ εr with r > 0, the integral converges.
208 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
clearly something has gone wrong. The remedy is to recognize that the single particle energy levels are discrete,
and separate out the contribution from the lowest energy state ε0 . I.e. we write
n0 n′
z }| { z }| {
Z∞
1 g0 g(ε)
n(T, µ) = β(ε −µ)
+ dε , (5.147)
V e 0 −1 eβ(ε−µ) −1
ε0
where g0 is the degeneracy of the single particle state with energy ε0 . We assume that n0 is finite, which means
that N0 = V n0 is extensive. We say that the particles have condensed into the state with energy ε0 . The quantity n0
is the condensate density. The remaining particles, with density n′ , are said to comprise the overcondensate. With the
total density n fixed, we have n = n0 + n′ . Note that n0 finite means that µ is infinitesimally close to ε0 :
g g k T
µ = ε0 − kB T ln 1 + 0 ≈ ε0 − 0 B . (5.148)
V n0 V n0
Note also that if ε0 − µ is finite, then n0 ∝ V −1 is infinitesimal.
Thus, for T < Tc (n), we have µ = ε0 with n0 > 0, and
Z∞
g(ε)
n(T, n0 ) = n0 + dε . (5.149)
e(ε−ε0 )/kB T − 1
ε0
We already derived, in §5.3.3, expressions for n(T, z) and p(T, z) for the ideal Bose gas (IBG) with ballistic disper-
sion ε(p) = p2 /2m, We found
n(T, z) = g λ−d
T Li d (z) (5.151)
2
p(T, z) = g kB T λ−d
T Li d +1 (z), (5.152)
2
where g is the internal (e.g. spin) degeneracy of each single particle energy level. Here z = eµ/kB T is the fugacity
and
X∞
zm
Lis (z) = (5.153)
m=1
ms
is the polylogarithm function. For bosons with a spectrum bounded below by ε0 = 0, the fugacity takes values on
the interval z ∈ [0, 1]5 .
5 Itis easy to see that the chemical potential for noninteracting bosons can never exceed the minimum value ε0 of the single particle
dispersion.
5.7. THE IDEAL BOSE GAS 209
Figure 5.6: The polylogarithm function Lis (z) versus z for s = 21 , s = 32 , and s = 52 . Note that Lis (1) = ζ(s) diverges
for s ≤ 1.
Clearly n(T, z) = g λT−d Li d (z) is an increasing function of z for fixed T . In fig. 5.6 we plot the function Lis (z)
2
versus z for three different values of s. We note that the maximum value Lis (z = 1) is finite if s > 1. Thus, for
d > 2, there is a maximum density nmax (T ) = g Li d (z) λ−d
T which is an increasing function of temperature T . Put
2
another way, if we fix the density n, then there is a critical temperature Tc below which there is no solution to the
equation n = n(T, z). The critical temperature Tc (n) is then determined by the relation
2/d
d
mkB Tc d/2 2π~2 n
n = gζ 2 =⇒ kB Tc = . (5.154)
2π~2 m g ζ d2
What happens for T < Tc ?
As shown above in §5.7, we must separate out the contribution from the lowest energy single particle mode, which
for ballistic dispersion lies at ε0 = 0. Thus writing
1 1 1 X 1
n= + , (5.155)
V z −1 − 1 V α
−1
z e εα /kBT − 1
(εα >0)
where we have taken g = 1. Now V −1 is of course very small, since V is thermodynamically large, but if µ → 0
then z −1 − 1 is also very small and their ratio can be finite, as we have seen. Indeed, if the density of k = 0 bosons
n0 is finite, then their total number N0 satisfies
1 1
N 0 = V n0 = =⇒ z= . (5.156)
z −1 − 1 1 + N0−1
The chemical potential is then
k T
µ = kB T ln z = −kB T ln 1 + N0−1 ≈ − B → 0− . (5.157)
N0
In other words, the chemical potential is infinitesimally negative, because N0 is assumed to be thermodynamically
large.
According to eqn. 5.14, the contribution to the pressure from the k = 0 states is
kB T k T
p0 = − ln(1 − z) = B ln(1 + N0 ) → 0+ . (5.158)
V V
210 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
So the k = 0 bosons, which we identify as the condensate, contribute nothing to the pressure.
Having separated out the k = 0 mode, we can now replace the remaining sum over α by the usual integral over
k. We then have
T < Tc : n = n0 + g ζ d2 λ−d
T (5.159)
p = gζ d
2 +1 kB T λ−d
T (5.160)
and
p = g Li d +1 (z) kB T λ−d
T . (5.162)
2
The condensate fraction n0 /n is unity at T = 0, when all particles are in the condensate with k = 0, and decreases
with increasing T until T = Tc , at which point it vanishes identically. Explicitly, we have
d/2
n0 (T ) g ζ d2 T
=1− = 1 − . (5.163)
n n λdT Tc (n)
Let us compute the internal energy E for the ideal Bose gas. We have
∂ ∂Ω ∂Ω
(βΩ) = Ω + β =Ω−T = Ω + TS (5.164)
∂β ∂β ∂T
and therefore
∂
E = Ω + T S + µN = µN + (βΩ)
∂β
∂
= V µn− (βp) (5.165)
∂β
= 12 d gV kB T λ−d
T Li d +1 (z) .2
This expression is valid at all temperatures, both above and below Tc . Note that the condensate particles do not
contribute to E, because the k = 0 condensate particles carry no energy.
We now investigate the heat capacity CV,N = ∂E ∂T V,N . Since we have been working in the GCE, it is very
important to note that N is held constant when computing CV,N . We’ll also restrict our attention to the case d = 3
since the ideal Bose gas does not condense at finite T for d ≤ 2 and d > 3 is unphysical. While we’re at it, we’ll
also set g = 1.
The number of particles is
N 0 + ζ
3
2 V λ−3
T (T < Tc )
N= (5.166)
V λ−3
T Li3/2 (z) (T > Tc ) ,
Figure 5.7: Molar heat capacity of the ideal Bose gas (units of R). Note the cusp at T = Tc .
CV,N −1
cV,N (T, n) = NA · = 15
4 ζ 5
2 R · n λ3T . (5.169)
N
V dT V dz
dN V = 3
Li3/2 (z) · + Li1/2 (z) 3 · . (5.171)
2 λ3T T λT z
n(T, z) = λ−3
T Li3/2 (z) (5.173)
in order to obtain z(T, n), and then insert this into eqn. 5.172. The results are shown in fig. 5.7. There are several
noteworthy features of this plot. First of all, by dimensional analysis the function cV,N (T, n) is R times a function
of the dimensionless ratio T /Tc(n) ∝ T n−2/3 . Second, the high temperature limit is 32 R, which is the classical
value. Finally, there is a cusp at T = Tc (n).
212 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
Figure 5.8: Phase diagrams for the ideal Bose gas. Left panel: (p, v) plane. The solid blue curves are isotherms,
and the green hatched region denotes v < vc (T ), where the system is partially condensed. Right panel: (p, T )
plane. The solid red curve is the coexistence curve pc (T ), along which Bose condensation occurs. No distinct
thermodynamic phase exists in the yellow hatched region above p = pc (T ).
Figure 5.9: Phase diagram of 4 He. All phase boundaries are first order transition lines, with the exception of the
normal liquid-superfluid transition, which is second order. (Source: University of Helsinki)
where ℓ = T (s2 − s1 ) is the latent heat per mole, and ∆v = v2 − v1 . For ideal gas Bose condensation, the
coexistence curve resembles the red curve in the right hand panel of fig. 5.8. There is no meaning to the shaded
region where p > pc (T ). Nevertheless, it is tempting to associate the curve p = pc (T ) with the coexistence of the
k = 0 condensate and the remaining uncondensed (k 6= 0) bosons7 .
The entropy in the coexistence region is given by
5
1 ∂Ω ζ 52 n0
s=− = 25 ζ 5
2 kB v λ−3
T = 2 k 1 − . (5.179)
N ∂T V ζ 23 B
n
All the entropy is thus carried by the uncondensed bosons, and the condensate carries zero entropy. The Clausius-
Clapeyron relation can then be interpreted as describing a phase equilibrium between the condensate, for which
s0 = v0 = 0, and the uncondensed bosons, for which s′ = s(T ) and v ′ = vc (T ). So this identification forces us to
conclude that the specific volume of the condensate is zero. This is certainly false in an interacting Bose gas!
While one can identify, by analogy, a ‘latent heat’ ℓ = T ∆s = T s in the Clapeyron equation, it is important to
understand that there is no distinct thermodynamic phase associated with the region p > pc (T ). Ideal Bose gas
condensation is a second order transition, and not a first order transition.
Helium has two stable isotopes. 4 He is a boson, consisting of two protons, two neutrons, and two electrons (hence
an even number of fermions). 3 He is a fermion, with one less neutron than 4 He. Each 4 He atom can be regarded
as a tiny hard sphere of mass m = 6.65 × 10−24 g and diameter a = 2.65 Å. A sketch of the phase diagram is shown
in fig. 5.9. At atmospheric pressure, Helium liquefies at Tl = 4.2 K. The gas-liquid transition is first order, as usual.
However, as one continues to cool, a second transition sets in at T = Tλ = 2.17 K (at p = 1 atm). The λ-transition,
so named for the λ-shaped anomaly in the specific heat in the vicinity of the transition, as shown in fig. 5.10, is
continuous (i.e. second order).
If we pretend that 4 He is a noninteracting Bose gas, then from the density of the liquid n = 2.2 × 1022 cm−3 , we
2 2/3
obtain a Bose-Einstein condensation temperature Tc = 2π~ m n/ζ( 32 ) = 3.16 K, which is in the right ballpark.
7 The k 6= 0 particles are sometimes called the overcondensate.
214 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
Figure 5.10: Specific heat of liquid 4 He in the vicinity of the λ-transition. Data from M. J. Buckingham and W. M.
Fairbank, in Progress in Low Temperature Physics, C. J. Gortner, ed. (North-Holland, 1961). Inset at upper right:
more recent data of J. A. Lipa et al., Phys. Rev. B 68, 174518 (2003) performed in zero gravity earth orbit, to within
∆T = 2 nK of the transition.
−α
Cp (T ) = A T − Tλ (p) . (5.180)
α is an example of a critical exponent. We shall study the physics of critical phenomena later on in this course.
For now, note that a cusp singularity of the type found in fig. 5.7 corresponds to α = −1. The behavior of
Cp (T ) in 4 He is very nearly logarithmic in |T − Tλ |. In fact, both theory (renormalization group on the O(2)
model) and experiment concur that α is almost zero but in fact slightly negative, with α = −0.0127 ± 0.0003 in
the best experiments (Lipa et al., 2003). The λ transition is most definitely not an ideal Bose gas condensation.
Theoretically, in the parlance of critical phenomena, IBG condensation and the λ-transition in 4 He lie in different
universality classes8 . Unlike the IBG, the condensed phase in 4 He is a distinct thermodynamic phase, known as a
superfluid.
Note that Cp (T < Tc ) for the IBG is not even defined, since for T < Tc we have p = p(T ) and therefore dp = 0
requires dT = 0.
8 IBG condensation is in the universality class of the spherical model. The λ-transition is in the universality class of the XY model.
5.7. THE IDEAL BOSE GAS 215
Figure 5.11: The fountain effect. In each case, a temperature gradient is maintained across a porous plug through
which only superfluid can flow. This results in a pressure gradient which can result in a fountain or an elevated
column in a U-tube.
At temperatures T < Tλ , liquid 4 He has a superfluid component which is a type of Bose condensate. In fact, there
is an important difference between condensate fraction Nk=0 /N and superfluid density, which is denoted by the
symbol ρs . In 4 He, for example, at T = 0 the condensate fraction is only about 8%, while the superfluid fraction
ρs /ρ = 1. The distinction between N0 and ρs is very interesting but lies beyond the scope of this course.
One aspect of the superfluid state is its complete absence of viscosity. For this reason, superfluids can flow through
tiny cracks called microleaks that will not pass normal fluid. Consider then a porous plug which permits the
passage of superfluid but not of normal fluid. The key feature of the superfluid component is that it has zero
energy density. Therefore even though there is a transfer of particles across the plug, there is no energy exchange,
and therefore a temperature gradient across the plug can be maintained9 .
The elementary excitations in the superfluid state are sound waves called phonons. They are compressional waves,
just like longitudinal phonons in a solid, but here in a liquid. Their dispersion is acoustic, given by ω(k) = ck
where c = 238 m/s.10 The have no internal degrees of freedom, hence g = 1. Like phonons in a solid, the phonons
in liquid helium are not conserved. Hence their chemical potential vanishes and these excitations are described
by photon statistics. We can now compute the height difference ∆h in a U-tube experiment.
Clearly ∆h = ∆p/ρg. so we must find p(T ) for the helium. In the grand canonical ensemble, we have
Z
d3k
p = −Ω/V = −kB T ln 1 − e−~ck/kB T
(2π)3
Z∞
(kB T )4 4π
=− du u2 ln(1 − e−u ) (5.181)
(~c)3 8π 3
0
π 2 (kB T )4
= .
90 (~c)3
9 Recall that two bodies in thermal equilibrium will have identical temperatures if they are free to exchange energy.
10 The phonon velocity c is slightly temperature dependent.
216 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
Let’s assume T = 1 K. We’ll need the density of liquid helium, ρ = 148 kg/m3 .
3
dh 2π 2 kB T kB
=
dT 45 ~c ρg
2
3
2π (1.38 × 10−23 J/K)(1 K) (1.38 × 10−23 J/K) (5.182)
= ×
45 (1.055 × 10−34 J · s)(238 m/s) (148 kg/m3 )(9.8 m/s2 )
≃ 32 cm/K ,
a very noticeable effect!
The 2001 Nobel Prize in Physics was awarded to Weiman, Cornell, and Ketterle for the experimental observation
of Bose condensation in dilute atomic gases. The experimental techniques required to trap and cool such systems
are a true tour de force, and we shall not enter into a discussion of the details here11 .
The optical trapping of neutral bosonic atoms, such as 87 Rb, results in a confining potential V (r) which is quadratic
in the atomic positions. Thus, the single particle Hamiltonian for a given atom is written
~2 2 1
Ĥ = − ∇ + 2 m ω12 x2 + ω22 y 2 + ω32 z 2 , (5.183)
2m
where ω1,2,3 are the angular frequencies of the trap. This is an anisotropic three-dimensional harmonic oscillator,
the solution of which is separable into a product of one-dimensional harmonic oscillator wavefunctions. The
eigenspectrum is then given by a sum of one-dimensional spectra, viz.
En1 ,n2 ,n3 = n1 + 12 ) ~ω1 + n2 + 21 ) ~ω2 + n3 + 12 ) ~ω3 . (5.184)
1
≈ (5.187)
1 − e−k~ωj /kB T
∗
1 k > k (T )
We then have
∗ 3 X
k∗
y k +1 kB T yk
N (T, y) ≈ + , (5.189)
1−y ~ω̄ k3
k=1
where the first term on the RHS is due to k > k ∗ and the second term from k ≤ k ∗ in the previous sum. Since
k ∗ ≫ 1 and since the sum of inverse cubes is convergent, we may safely extend the limit on the above sum to
−1
infinity. To help make more sense of the first term, write N0 = y −1 − 1 for the number of particles in the
(n1 , n2 , n3 ) = (0, 0, 0) state. Then
N0
y= . (5.190)
N0 + 1
This is true always. The issue vis-a-vis Bose-Einstein condensation is whether N0 ≫ 1. At any rate, we now see
that we can write 3
∗
−1 −k kB T
N ≈ N0 1 + N0 + Li3 (y) . (5.191)
~ω̄
As for the first term, we have
0 N0 ≪ k ∗
−1 −k
∗
N0 1 + N0 = (5.192)
∗
N0 N0 ≫ k
Thus, as in the case of IBG condensation of ballistic particles, we identify the critical temperature by the condition
y = N0 /(N0 + 1) ≈ 1, and we have
1/3
~ω̄ N ν̄
Tc = = 4.5 N 1/3 [ nK ] , (5.193)
kB ζ(3) 100 Hz
where ν̄ = ω̄/2π. We see that kB Tc ≫ ~ω̄ if the number of particles in the trap is large: N ≫ 1. In this regime, we
have
3
k T
T < Tc : N = N0 + ζ(3) B (5.194)
~ω̄
3
kB T
T > Tc : N= Li3 (y) . (5.195)
~ω̄
It is interesting to note that BEC can also occur in two-dimensional traps, which is to say traps which are very
anisotropic, with oblate equipotential surfaces V (r) = V0 . This happens when ~ω3 ≫ kB T ≫ ω1,2 . We then have
1/2
~ω̄ 6N
Tc(d=2) = · (5.196)
kB π2
1/2
with ω̄ = ω1 ω2 . The particle number then obeys a set of equations like those in eqns. 5.194 and 5.195, mutatis
12
mutandis .
For extremely prolate traps, with ω3 ≪ ω1,2 , the situation is different because Li1 (y) diverges for y = 1. We then
have
k T
N = N0 + B ln 1 + N0 . (5.197)
~ω3
Here we have simply replaced y by the equivalent expression N0 /(N0 + 1). If our criterion for condensation is that
N0 = αN , where α is some fractional value, then we have
~ω3 N
Tc (α) = (1 − α) · . (5.198)
kB ln N
π2
12 Explicitly,
` ´3 ` ´2
one replaces ζ(3) with ζ(2) = 6
, Li3 (y) with Li2 (y), and kB T /~ω̄ with kB T /~ω̄ .
218 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
5.7.7 Example problem from Fall 2004 UCSD graduate written exam
1/2
PROBLEM: A three-dimensional gas of noninteracting bosonic particles obeys the dispersion relation ε(k) = A k .
(a) Obtain an expression for the density n(T, z) where z = exp(µ/kB T ) is the fugacity. Simplify your expres-
sion as best you can, adimensionalizing any integral or infinite sum which may appear. You may find it
convenient to define
Z∞ ∞
X
1 tν−1 zk
Liν (z) ≡ dt −1 t = . (5.199)
Γ(ν) z e −1 kν
0 k=1
where we have changed integration variables from k to s = Ak 1/2 /kB T , and we have defined the functions
Liν (z) as above, in eqn. 5.199. Note Liν (1) = ζ(ν), the Riemann zeta function.
(b) Bose condensation sets in for z = 1, i.e. µ = 0. Thus, the critical temperature Tc and the density n are related
by
6
120 ζ(6) kB Tc
n= , (5.201)
π2 A
or 1/6
A π2 n
Tc (n) = . (5.202)
kB 120 ζ(6)
(c) For T < Tc , we have
6
120 ζ(6) kB T
n = n0 +
π2 A
6 (5.203)
T
= n0 + n,
Tc
5.8. THE IDEAL FERMI GAS 219
1
where n0 is the condensate density. Thus, at T = 2 Tc ,
n0 T = 21 Tc = 63
64 n. (5.204)
We expect B2 (T ) < 0 for noninteracting bosons, reflecting the tendency of the bosons to condense. (Corre-
spondingly, for noninteracting fermions we expect B2 (T ) > 0.)
For the curious, we compute B2 (T ) by eliminating the fugacity z from the equations for n(T, z) and p(T, z).
First, we find p(T, z):
Z 3
dk 1/2
p(T, z) = −kB T ln 1 − z exp(−Ak /k T )
(2π)3 B
6 Z∞
kB T kB T
=− 2 ds s5 ln 1 − z e−s (5.206)
π A
0
6
120 kB T kB T
= Li7 (z).
π2 A
Expanding in powers of the fugacity, we have
6
120 kB T n z2 z3 o
n= 2 z+ + + . . . (5.207)
π A 26 36
6
p 120 kB T n z2 z3 o
= 2 z+ + + . . . . (5.208)
kB T π A 27 37
Solving for z(n) using the first equation, we obtain, to order n2 ,
2
π 2 A6 n 1 π 2 A6 n
z= − 6 + O(n3 ) . (5.209)
120 (kB T )6 2 120 (kB T )6
Plugging this into the equation for p(T, z), we obtain the first nontrivial term in the virial expansion, with
6
π2 A
B2 (T ) = − , (5.210)
15360 kB T
which is negative, as expected. Note also that the ideal gas law is recovered for T → ∞, for fixed n.
The grand potential of the ideal Fermi gas is, per eqn. 5.14,
X
Ω(T, V, µ) = −V kB T ln 1 + eµ/kB T e−εα /kB T
α
Z∞ (5.211)
= −V kB T dε g(ε) ln 1 + e(µ−ε)/kB T .
−∞
220 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
−1
Figure 5.12: The Fermi distribution, f (ǫ) = exp(ǫ/kB T ) + 1 . Here we have set kB = 1 and taken µ = 2, with
1
T = 20 (blue), T = 34 (green), and T = 2 (red). In the T → 0 limit, f (ǫ) approaches a step function Θ(−ǫ).
1
n(ε) = , (5.212)
e(ε−µ)/kB T +1
hence the total number of particles is
Z∞
1
N = V dε g(ε) . (5.213)
e(ε−µ)/kB T +1
−∞
At T = 0, we therefore have n(ε) = Θ(µ − ε), which says that all single particle energy states up to ε = µ are
filled, and all energy states above ε = µ are empty. We call µ(T = 0) the Fermi energy: εF = µ(T = 0). If the
single particle dispersion ε(k) depends only on the wavevector k, then the locus of points in k-space for which
ε(k) = εF is called the Fermi surface. For isotropic systems, ε(k) = ε(k) is a function only of the magnitude k = |k|,
and the Fermi surface is a sphere in d = 3 or a circle in d = 2. The radius of this circle is the Fermi wavevector, kF .
When there is internal (e.g. spin) degree of freedom, there is a Fermi surface and Fermi wavevector (for isotropic
systems) for each polarization state of the internal degree of freedom.
5.8. THE IDEAL FERMI GAS 221
Let’s compute the Fermi wavevector kF and Fermi energy εF for the IFG with a ballistic dispersion ε(k) =
~2 k2 /2m. The number density is
g kF /π
(d = 1)
Z
g Ωd kFd
d
n = g d k Θ(kF − k) = · = g kF2 /4π (d = 2) (5.215)
(2π)d d
g k 3 /6π 2 (d = 3) .
F
Note that the form of n(kF ) is independent of the dispersion relation, so long as it remains isotropic. Inverting the
above expressions, we obtain kF (n):
πn/g (d = 1)
1/d
dn
kF = 2π = (4πn/g)1/2 (d = 2) (5.216)
g Ωd
(6π 2 n/g)1/3 (d = 3) .
Another useful result for the ballistic dispersion, which follows from the above, is that the density of states at the
Fermi level is given by
g Ωd mkFd−2 d n
g(εF ) = · = · . (5.218)
(2π)d ~2 2 εF
−1 −1
For the electron gas, we have g = 2. In a metal, one typically has kF ∼ 0.5 Å to 2 Å , and εF ∼ 1 eV − 10 eV.
Due to the effects of the crystalline lattice, electrons in a solid behave as if they had an effective mass m∗ which
is typically on the order of the electron mass but very often about an order of magnitude smaller, particularly in
semiconductors.
Nonisotropic dispersions ε(k) are more interesting in that they give rise to non-spherical Fermi surfaces. The
simplest example is that of a two-dimensional ‘tight-binding’ model of electrons hopping on a square lattice, as
may be appropriate in certain layered materials. The dispersion relation is then
If we only work to quadratic order in kx and ky , the dispersion is isotropic, and the Fermi surface is a circle, with
kF2 = (εF + 4t)/ta2 . As the energy increases further, the continuous O(2) rotational invariance is broken down to
222 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
Figure 5.13: Fermi surfaces for two and three-dimensional structures. Upper left: free particles in two dimensions.
Upper right: ‘tight binding’ electrons on a square lattice. Lower left: Fermi surface for cesium, which is predom-
inantly composed of electrons in the 6s orbital shell. Lower right: the Fermi surface of yttrium has two parts.
One part (yellow) is predominantly due to 5s electrons, while the other (pink) is due to 4d electrons. (Source:
www.phys.ufl.edu/fermisurface/)
the discrete group of rotations of the square, C4v . The Fermi surfaces distort and eventually, at εF = 0, the Fermi
π π
is itself a square. As εF increases further, the square turns
surface
2π
back into a circle, but centered about the point
,
a a . Note that everything is periodic in kx and ky modulo a . The Fermi surfaces for this model are depicted
in the upper right panel of fig. 5.13.
Fermi surfaces in three dimensions can be very interesting indeed, and of great importance in understanding the
electronic properties of solids. Two examples are shown in the bottom panels of fig. 5.13. The electronic configu-
ration of cesium (Cs) is [Xe] 6s1 . The 6s electrons ‘hop’ from site to site on a body centered cubic (BCC) lattice, a
generalization of the simple two-dimensional square lattice hopping model discussed above. The elementary unit
cell in k space, known as the first Brillouin zone, turns out to be a dodecahedron. In yttrium, the electronic structure
is [Kr] 5s2 4d1 , and there are two electronic energy bands at the Fermi level, meaning two Fermi surfaces. Yttrium
forms a hexagonal close packed (HCP) crystal structure, and its first Brillouin zone is shaped like a hexagonal
pillbox.
Consider an electron gas in an external magnetic field H. The single particle Hamiltonian is then
p2
Ĥ = + µB H σ , (5.221)
2m
5.8. THE IDEAL FERMI GAS 223
In dealing with the ideal Fermi gas, we will repeatedly encounter integrals of the form
Z∞
I(T, µ) ≡ dε f (ε − µ) φ(ε) . (5.227)
−∞
The Sommerfeld expansion provides a systematic way of expanding these expressions in powers of T and is an
important analytical tool in analyzing the low temperature properties of the ideal Fermi gas (IFG).
We start by defining
Zε
Φ(ε) ≡ dε′ φ(ε′ ) (5.228)
−∞
Thus,
I(T, µ) = πD csc(πD) Φ(µ) , (5.235)
which is to be understood as the differential operator πD(csc πD) = πD/ sin(πD) acting on the function Φ(µ).
Appealing once more to Taylor’s theorem, we have
π2 d2 7π 4 d4
πD csc(πD) = 1 + (kB T )2 2 + (kB T )4 4 + . . . . (5.236)
6 dµ 360 dµ
Thus,
Z∞
I(T, µ) = dε f (ε − µ) φ(ε)
−∞
Zµ
π2 7π 4
= dε φ(ε) + (kB T )2 φ′ (µ) + (k T )4 φ′′′ (µ) + . . . . (5.237)
6 360 B
−∞
If φ(ε) is a polynomial function of its argument, then each derivative effectively reduces the order of the polyno-
mial by one degree, and the dimensionless parameter of the expansion is (T /µ)2 . This procedure is known as the
Sommerfeld expansion.
13 Note that writing v = (2n + 1) iπ + ǫ we have e±v = −1 ∓ ǫ − 1 ǫ2 + . . . , so (ev + 1)(e−v + 1) = −ǫ2 + . . . We then expand
2
evD = e(2n+1)iπD 1 + ǫD + . . .) to find the residue: Res = −D e(2n+1)iπD .
`
5.8. THE IDEAL FERMI GAS 225
As our first application of the Sommerfeld expansion formalism, let us compute µ(n, T ) for the ideal Fermi gas.
The number density n(T, µ) is
Z∞
n = dε g(ε) f (ε − µ)
−∞
(5.238)
Zµ
π2
= dε g(ε) + (kB T )2 g ′ (µ) + . . . .
6
−∞
Let us write µ = εF + δµ, where εF = µ(T = 0, n) is the Fermi energy, which is the chemical potential at T = 0.
We then have
εFZ+δµ
π2
n= dε g(ε) + (kB T )2 g ′ (εF + δµ) + . . .
6
−∞
(5.239)
ZεF
π2
= dε g(ε) + g(εF ) δµ + (kB T )2 g ′ (εF ) + . . . ,
6
−∞
1
Thus, g(ε) ∝ ε1/2 and (ln g)′ = 2 ε−1 , so
π 2 (kB T )2
µ(n, T ) = εF − + ... , (5.242)
12 εF
226 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
~2 2 2/3
where εF (n) = 2m (3π n) .
Magnetism has two origins: (i) orbital currents of charged particles, and (ii) intrinsic magnetic moment. The
intrinsic magnetic moment m of a particle is related to its quantum mechanical spin via
q~
m = gµ0 S/~ , µ0 = = magneton , (5.249)
2mc
5.8. THE IDEAL FERMI GAS 227
1 q 2
Ĥ = p− A − H ·m
2m∗ c
(5.250)
1 e 2 g
= p+ A + µB H σ ,
2m∗ c 2
where in the last line we’ve restricted our attention to the electron, for which q = −e. The g-factor for an electron
is g = 2 at tree level, and when radiative corrections are accounted for using quantum electrodynamics (QED) one
finds g = 2.0023193043617(15). For our purposes we can take g = 2, although we can always absorb the small
difference into the definition of µB , writing µB → µ̃B = ge~/4mc. We’ve chosen the ẑ-axis in spin space to point
in the direction of the magnetic field, and we wrote the eigenvalues of S z as 21 ~σ, where σ = ±1. The quantity
m∗ is the effective mass of the electron, which we mentioned earlier. An important distinction is that it is m∗ which
enters into the kinetic energy term p2 /2m∗ , but it is the electron mass m itself (m = 511 keV) which enters into the
definition of the Bohr magneton. We shall discuss the consequences of this further below.
In the absence of orbital magnetic coupling, the single particle dispersion is
~2 k 2
εσ (k) = + µ̃B H σ . (5.251)
2m∗
At T = 0, we have the results of §5.8.4. At finite T , we once again use the Sommerfeld expansion. We then have
Z∞ Z∞
n = dε g↑ (ε) f (ε − µ) + dε g↓ (ε) f (ε − µ)
−∞ −∞
Z∞ n o
1
= 2 dε g(ε − µ̃B H) + g(ε + µ̃B H) f (ε − µ) (5.252)
−∞
Z∞ n o
= dε g(ε) + (µ̃B H)2 g ′′ (ε) + . . . f (ε − µ) .
−∞
where g(ε) is the total density of states per unit volume, for both spin species, in the absence of a magnetic field.
We conclude that the chemical potential shift in an external field is
2 ′
π g (εF )
δµ(T, n, H) = − (kB T )2 + (µ̃B H)2 + ... . (5.255)
6 g(εF )
228 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
Figure 5.15: Fermi distributions in the presence of an external Zeeman-coupled magnetic field.
We next compute the difference n↑ − n↓ in the densities of up and down spin electrons:
Z∞ n o
n↑ − n↓ = dε g↑ (ε) − g↓ (ε) f (ε − µ)
−∞
Z∞ n o (5.256)
1
= 2 dε g(ε − µ̃B H) − g(ε + µ̃B H) f (ε − µ)
−∞
We needn’t go beyond the trivial lowest order term in the Sommerfeld expansion, because H is already assumed
to be small. Thus, the magnetization density is
When orbital effects are included, the single particle energy levels are given by
~2 kz2
ε(n, kz , σ) = (n + 21 )~ωc + + µ̃B H σ . (5.259)
2m∗
Here n is a Landau level index, and ωc = eH/m∗ c is the cyclotron frequency. Note that
µ̃B H ge~H m∗ c g m∗
= · = · . (5.260)
~ωc 4mc ~eH 4 m
5.8. THE IDEAL FERMI GAS 229
Accordingly, we define the ratio r ≡ (g/2) × (m∗ /m). We can then write
1
~2 kz2
ε(n, kz , σ) = n + 2 + 21 rσ ~ωc + . (5.261)
2m∗
The grand potential is then given by
Z∞
dkz X X h i
∞
HA 1 1 2 2
Ω=− · Lz · kB T ln 1 + eµ/kB T e−(n+ 2 + 2 rσ)~ωc /kB T e−~ kz /2mkB T . (5.262)
φ0 2π n=0 σ=±1
−∞
A few words are in order here regarding the prefactor. In the presence of a uniform magnetic field, the energy
levels of a two-dimensional ballistic charged particle collapse into Landau levels. The number of states per Landau
level scales with the area of the system, and is equal to the number of flux quanta through the system: Nφ =
HA/φ0 , where φ0 = hc/e is the Dirac flux quantum. Note that
HA V
· Lz · kB T = ~ωc · 3 , (5.263)
φ0 λT
hence we can write
∞ X
X
1
Ω(T, V, µ, H) = ~ωc Q (n + 2 + 12 rσ)~ωc − µ , (5.264)
n=0 σ=±1
where
Z∞
V dkz h 2 2 ∗
i
Q(ε) = − 2 ln 1 + e−ε/kB T e−~ kz /2m kB T . (5.265)
λT 2π
−∞
resulting in
( Z∞
X
1 1
Ω= dε Q(ε − µ) + 2 ~ωc Q 2 (1 + rσ)~ωc − µ
σ=±1 1
2 (1+rσ)~ωc (5.267)
)
1
− 12 (~ωc )2 Q′ 12 (1 + rσ)~ωc − µ + . . .
1 ∂ 2Ω 2 2
χ=−
2
= r2 − 13 · µ̃2B · m/m∗ · − Q′ (−µ)
V ∂H V
2 (5.269)
g m2
= − · µ̃2B · n2 κT ,
4 3m∗ 2
230 CHAPTER 5. NONINTERACTING QUANTUM SYSTEMS
where κT is the isothermal compressibility14 . In most metals we have m∗ ≈ m and the term in brackets is positive
(recall g ≈ 2). In semiconductors, however, we can have m∗ ≪ m; for example in GaAs we have m∗ = 0.067 .
Thus, semiconductors can have a diamagnetic response. If we take g = 2 and m∗ = m, we see that the orbital
currents give rise to a diamagnetic contribution to the magnetic susceptibility which is exactly − 31 times as large
as the contribution arising from Zeeman coupling. The net result is then paramagnetic (χ > 0) and 32 as large as
the Pauli susceptibility. The orbital currents can be understood within the context of Lenz’s law.
Exercise : Show that − V2 Q′ (−µ) = n2 κT .
There is a nice discussion of this material in R. K. Pathria, Statistical Mechanics . As a model, consider a mass
M ∼ 1033 g of helium at nuclear densities of ρ ∼ 107 g/cm3 and temperature T ∼ 107 K. This temperature is
much larger than the ionization energy of 4 He, hence we may safely assume that all helium atoms are ionized. If
there are N electrons, then the number of α particles (i.e. 4 He nuclei) must be 12 N . The mass of the α particle is
mα ≈ 4mp . The total stellar mass M is almost completely due to α particle cores.
The electron density is then
N 2 · M/4mp ρ
n= = = ≈ 1030 cm−3 , (5.270)
V V 2mp
since M = N · me + 21 N · 4mp . From the number density n we find for the electrons
Since pF ∼ mc, we conclude that the electrons are relativistic. The Fermi temperature will then be TF ∼ mc2 ∼
106 eV ∼ 1012 K. Thus, T ≪ Tf which says that the electron gas is degenerate and may be considered to be at
T ∼ 0. So we need to understand the ground state properties of the relativistic electron gas.
The kinetic energy is given by p
ε(p) = p2 c2 + m2 c4 − mc2 . (5.274)
The velocity is
∂ε pc2
v= =p . (5.275)
∂p p 2 c2 + m 2 c4
The pressure in the ground state is
p0 = 13 nhp · vi
ZpF
1 p 2 c2
= dp p2 · p
3π 2 ~3 p 2 c2 + m 2 c4
0
ZθF (5.276)
m 4 c5
= dθ sinh4 θ
3π 2 ~3
0
4 5
m c
= 2 3
sinh(4θF ) − 8 sinh(2θF ) + 12 θF ,
96π ~
∂ 2Ω
14 We’ve used − V2 Q′ (µ) = − V1 ∂µ2
= n2 κ T .
5.8. THE IDEAL FERMI GAS 231
Figure 5.16: Mass-radius relationship for white dwarf stars. (Source: Wikipedia).
~2
R= 3
40γ (9π)2/3 5/3
∝ M −1/3 . (5.285)
G mp m M 1/3
To find the R dependence, we must go beyond the lowest order expansion of eqn. 5.283, in which case we find
1/3 1/3 " 2/3 #1/2
9π ~ M M
R= 1− . (5.287)
8 mc mp M0
The value M0 is the limiting size for a white dwarf. It is called the Chandrasekhar limit.
Chapter 6
6.1 References
– M. Kardar, Statistical Physics of Particles (Cambridge, 2007)
A superb modern text, with many insightful presentations of key concepts.
– L. E. Reichl, A Modern Course in Statistical Physics (2nd edition, Wiley, 1998)
A comprehensive graduate level text with an emphasis on nonequilibrium phenomena.
– M. Plischke and B. Bergersen, Equilibrium Statistical Physics (3rd edition, World Scientific, 2006)
An excellent graduate level text. Less insightful than Kardar but still a good modern treatment of the subject.
Good discussion of mean field theory.
– E. M. Lifshitz and L. P. Pitaevskii, Statistical Physics (part I, 3rd edition, Pergamon, 1980)
This is volume 5 in the famous Landau and Lifshitz Course of Theoretical Physics . Though dated, it still
contains a wealth of information and physical insight.
– J.-P Hansen and I. R. McDonald, Theory of Simple Liquids (Academic Press, 1990)
An advanced, detailed discussion of liquid state physics.
233
234 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
6.2.1 Definition
The simplest model of an interacting system consists of a lattice L of sites, each of which contains a spin σi which
may be either up (σi = +1) or down (σi = −1). The Hamiltonian is
X X
Ĥ = −J σi σj − µ0 H σi . (6.1)
hiji i
When J > 0, the preferred (i.e. lowest energy) configuration of neighboring spins is that they are aligned, i.e.
σi σj = +1. The interaction is then called ferromagnetic. When J < 0 the preference is for anti-alignment, i.e.
σi σj = −1, which is antiferromagnetic.
This model is not exactly solvable in general. In one dimension, the solution is quite straightforward. In two
dimensions, Onsager’s solution of the model (with H = 0) is among the most celebrated results in statistical
physics. In higher dimensions the system has been studied by numerical simulations (the Monte Carlo method)
and by field theoretic calculations (renormalization group), but no exact solutions exist.
Consider a one-dimensional ring of N sites. The ordinary canonical partition function is then
Zring = Tr e−β Ĥ
N
X Y
= eβJσn σn+1 eβµ0 Hσn (6.2)
{σn } n=1
= Tr RN ,
where σN +1 ≡ σ1 owing to periodic (ring) boundary conditions, and where R is a 2 × 2 transfer matrix,
′ ′
Rσσ′ = eβJσσ eβµ0 H(σ+σ )/2
βJ βµ H
e e 0 e−βJ
= (6.3)
e−βJ eβJ e−βµ0 H
= eβJ cosh(βµ0 H) + eβJ sinh(βµ0 H) τ z + e−βJ τ x ,
where τ α are the Pauli matrices. Since the trace of a matrix is invariant under a similarity transformation, we have
Z(T, H, N ) = λN N
+ + λ− , (6.4)
where q
λ± (T, H) = eβJ cosh(βµ0 H) ± e2βJ sinh2 (βµ0 H) + e−2βJ (6.5)
are the eigenvalues of R. In the thermodynamic limit, N → ∞, and the λN + term dominates exponentially. We
therefore have
F (T, H, N ) = −N kB T ln λ+ (T, H) . (6.6)
From the free energy, we can compute the magnetization,
∂F N µ0 sinh(βµ0 H)
M =− = q (6.7)
∂H T,N
sinh2 (βµ0 H) + e−4βJ
6.2. ISING MODEL 235
6.2.3 H = 0
Consider the case H = 0, where R = eβJ + e−βJ τ x , where τ x is the Pauli matrix. Then
| ± i = √12 | ↑i ± | ↓i , (6.14)
This result is also valid for n < 0, provided |n| ≤ N . We see that we may write
When the chain has free ends, there are (N −1) links, and the partition function is
X
Zchain = RN −1 σσ′
σ,σ′
Xn o (6.20)
= λN
+
−1
ψ+ (σ) ψ+ (σ ′ ) + λN
−
−1
ψ− (σ) ψ− (σ ′ ) ,
σ,σ′
and therefore
Zchain = 2N coshN −1 (βJ) . (6.22)
There’s a nifty trick to obtaining the partition function for the Ising chain which amounts to a change of variables.
We define
νn ≡ σn σn+1 (n = 1 , . . . , N − 1) . (6.23)
Thus, ν1 = σ1 σ2 , ν2 = σ2 σ3 , etc. Note that each νj takes the values ±1. The Hamiltonian for the chain is
N
X −1 N
X −1
Hchain = −J σn σn+1 = −J νn . (6.24)
n=1 n=1
The state of the system is defined by the N Ising variables {σ1 , ν1 , . . . , νN −1 }. Note that σ1 doesn’t appear in the
Hamiltonian. Thus, the interacting model is recast as N −1 noninteracting Ising spins, and the partition function
is
Zchain = Tr e−βHchain
XX X
= ··· eβJν1 eβJν2 · · · eβJνN −1
σ1 ν1 νN −1 (6.25)
!N −1
X X
= eβJν = 2N coshN −1 (βJ) .
σ1 ν
6.2. ISING MODEL 237
We have just seen how in one dimension, the Ising model never achieves long-ranged spin order. That is, the
spin-spin correlation function decays asymptotically as an exponential function of the distance with a correlation
length ξ(T ) which is finite for all > 0. Only for T = 0 does the correlation length diverge. At T = 0, there are two
ground states, | ↑↑↑↑ · · · ↑ i and | ↓↓↓↓ · · · ↓ i. To choose between these ground states, we can specify a boundary
condition at the ends of our one-dimensional chain, where we demand that the spins are up. Equivalently, we can
apply a magnetic field H of order 1/N , which vanishes in the thermodynamic limit, but which at zero temperature
will select the ‘all up’ ground state. At finite temperature, there is always a finite probability for any consecutive
pair of sites (n, n+1) to be in a high energy state, i.e. either | ↑↓ i or | ↓↑ i. Such a configuration is called a domain
wall, and in one-dimensional systems domain walls live on individual links. Relative to the configurations | ↑↑ i
and | ↓↓ i, a domain wall costs energy 2J. For a system with M = xN domain walls, the free energy is
N
F = 2M J − kB T ln
M
( ) (6.26)
h i
= N · 2Jx + kB T x ln x + (1 − x) ln(1 − x) ,
Minimizing the free energy with respect to x, one finds x = 1 e2J/kB T + 1 , so the equilibrium concentration of
domain walls is finite, meaning there can be no long-ranged spin order. In one dimension, entropy wins and there
is always a thermodynamically large number of domain walls in equilibrium. And since the correlation length for
T > 0 is finite, any boundary conditions imposed at spatial infinity will have no thermodynamic consequences
since they will only be ‘felt’ over a finite range.
As we shall discuss in the following chapter, this consideration is true for any system with sufficiently short-
ranged interactions and a discrete global symmetry. Another example is the q-state Potts model,
X X
H = −J δσ ,σ − h δσ ,1 . (6.27)
i j i
hiji i
Here, the spin variables σi take values in the set {1, 2, . . . , q} on each site. The equivalent of an external magnetic
field in the Ising case is a field h which prefers a particular value of σ (σ = 1 in the above Hamiltonian). See the
appendix in §6.8 for a transfer matrix solution of the one-dimensional Potts model.
What about higher dimensions? A nifty argument due to R. Peierls shows that there will be a finite temperature
phase transition for the Ising model on the square lattice1 . Consider the Ising model, in zero magnetic field, on
a Nx × Ny square lattice, with Nx,y → ∞ in the thermodynamic limit. Along the perimeter of the system we
impose the boundary condition σi = +1. Any configuration of the spins may then be represented uniquely in
the following manner. Start with a configuration in which all spins are up. Next, draw a set of closed loops on
the lattice. By definition, the loops cannot share any links along their boundaries, i.e. each link on the lattice is
associated with at most one such loop. Now flip all the spins inside each loop from up to down. Identify each
such loop configuration with a label Γ . The partition function is
X
Z = Tr e−β Ĥ = e−2βJLΓ , (6.28)
Γ
where LΓ is the total perimeter of the loop configuration Γ . The domain walls are now loops, rather than individ-
ual links, but as in the one-dimensional case, each link of each domain wall contributes an energy +2J relative to
the ground state.
Now we wish to compute the average magnetization of the central site (assume Nx,y are both odd, so there is a
unique central site). This is given by the difference P+ (0) − P− (0), where Pµ (0) = δσ0 , µ is the probability that
1 Here we modify slightly the discussion in chapter 5 of the book by L. Peliti.
238 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Figure 6.1: Clusters and boundaries for the square lattice Ising model. Left panel: a configuration Γ where the
central spin is up. Right panel: a configuration Cγ ◦ Γ where the interior spins of a new loop γ containing the
central spin have been flipped.
the central spin has spin polarization µ. If P+ (0) > P− (0), then the magnetization per site m = P+ (0) − P− (0) is
finite in the thermodynamic limit, and the system is ordered. Clearly
1 X −2βJLΓ
P+ (0) = e , (6.29)
Z
Γ ∈Σ+
where the restriction on the sum indicates that only those configurations where the central spin is up (σ0 = +1)
are to be included. (see fig. 6.1a). Similarly,
1 X −2βJL e
P− (0) = e Γ , (6.30)
Z
e∈Σ
Γ −
where only configurations in which σ0 = −1 are included in the sum. Here we have defined
n o
Σ± = Γ σ0 = ± . (6.31)
I.e. Σ+ (Σ− ) is the set of configurations Γ in which the central spin is always up (down). Consider now the
construction in fig. 6.1b. Any loop configuration Γe ∈ Σ− may be associated with a unique loop configuration
Γ ∈ Σ+ by reversing all the spins within the loop of Γe which contains the origin. Note that the map from Γe to
Γ is many-to-one. That is, we can write Γe = Cγ ◦ Γ , where Cγ overturns the spins within the loop γ, with the
conditions that (i) γ contains the origin, and (ii) none of the links in the perimeter of γ coincide with any of the
links from the constituent loops of Γ . Let us denote this set of loops as ΥΓ :
n o
ΥΓ = γ : 0 ∈ int(γ) and γ ∩ Γ = ∅ . (6.32)
Then
1 X −2βJLΓ X
−2βJLγ
m = P+ (0) − P− (0) = e 1− e . (6.33)
Z
Γ ∈Σ+ γ∈ΥΓ
6.2. ISING MODEL 239
P
If we can prove that γ∈Υ e−2βJLγ < 1, then we will have established that m > 0. Let us ask: how many loops
Γ
γ are there in ΥΓ with perimeter L? We cannot answer this question exactly, but we can derive a rigorous upper
bound for this number, which, following Peliti, we call g(L). We claim that
2
2 L L L
g(L) < · 3L · = ·3 . (6.34)
3L 4 24
To establish this bound, consider any site on such a loop γ. Initially we have 4 possible directions to proceed
to the next site, but thereafter there are only 3 possibilities for each subsequent step, since the loop cannot run
into itself. This gives 4 · 3L−1 possibilities. But we are clearly overcounting, since any point on the loop could
have been chosen as the initial point, and moreover we could have started by proceeding either clockwise or
counterclockwise. So we are justified in dividing this by 2L. We are still overcounting, because we have not
accounted for the constraint that γ is a closed loop, nor that γ ∩ Γ = ∅. We won’t bother trying to improve our
estimate to account for these constraints. However, we are clearly undercounting due to the fact that a given loop
can be translated in space so long as the origin remains within it. To account for this, we multiply by the area
of a square of side length L/4, which is the maximum area that can be enclosed by a loop of perimeter L. We
therefore arrive at eqn. 6.34. Finally, we note that the smallest possible value of L is L = 4, corresponding to a
square enclosing the central site alone. Therefore
X ∞
1 X 2k x4 (2 − x2 )
e−2βJLγ < k · 3 e−2βJ = ≡r, (6.35)
12 12 (1 − x2 )2
γ∈ΥΓ k=2
where x = 3 e−2βJ . Note that we have accounted for the fact that the perimeter L of each loop γ must be an even
integer. The sum is smaller than unity provided x < x0 = 0.869756 . . ., hence the system is ordered provided
kB T 2
< = 1.61531 . (6.36)
J ln(3/x0 )
The exact result is kB Tc = 2J/ sinh−1 (1) = 2.26918 . . . The Peierls argument has been generalized to higher
dimensional lattices as well2 .
With a little more work we can derive a bound for the magnetization. We have shown that
Thus,
1 = P+ (0) + P− (0) < (1 + r) P+ (0) (6.38)
and therefore
1−r
m = P+ (0) − P− (0) > (1 − r) P+ (0) > , (6.39)
1+r
where r(T ) is given in eqn. 6.35.
We showed that the one-dimensional Ising model has no finite temperature phase transition, and is disordered
at any finite temperature T , but in two dimensions on the square lattice there is a finite critical temperature Tc
below which there is long-ranged order. Consider now the construction depicted in fig. 6.2, where the sites of
2 See. e.g. J. L. Lebowitz and A. E. Mazel, J. Stat. Phys. 90, 1051 (1998).
240 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
-31 18 19 20 21 22 23 24 25
-29 16 -7 2 3 4 5 -14 27
-28 15 -6 1 0 -1 6 -15 28
-27 14 -5 -4 -3 -2 7 -16 29
-26 13 12 11 10 9 8 -17 30
35 34 33 32
a two-dimensional square lattice are mapped onto those of a linear chain3 . Clearly we can elicit a one-to-one
mapping between the sites of a two-dimensional square lattice and those of a one-dimensional chain. That is, the
two-dimensional square lattice Ising model may be written as a one-dimensional Ising model, i.e.
square linear
X
lattice X
chain
How can this be consistent with the results we have just proven?
The fly in the ointment here is that the interaction along the chain Jn,n′ is long-ranged. This is apparent from
inspecting the site labels in fig. 6.2. Note that site n = 15 is linked to sites n′ = 14 and n′ = 16, but also to sites n′ =
−6 and n′ = −28. With each turn of the concentric spirals in the figure, the range of the interaction increases. To
complicate matters further, the interactions are no longer translationally invariant, i.e. Jnn′ 6= J(n−n′ ). But it is the
long-ranged nature of the interactions on our contrived one-dimensional chain which spoils our previous energy-
entropy argument, because now the domain walls themselves interact via a long-ranged potential. Consider for
example the linear chain with Jn,n′ = J |n − n′ |−α , where α > 0. Let us compute the energy of a domain wall
configuration where σn = +1 if n > 0 and σn = −1 if n ≤ 0. The domain wall energy is then
∞ X
X ∞
2J
∆= . (6.41)
m=0 n=1
|m + n|α
Here we have written one of the sums in terms of m = −n′ . For asymptotically large m and n, we can write
R = (m, n) and we obtain an integral over the upper right quadrant of the plane:
Z∞ Zπ/2 Zπ/4 Z∞
2J dφ dR
dR R dφ = 2−α/2 . (6.42)
Rα (cos φ + sin φ)α cosα φ Rα−1
1 0 −π/4 1
The φ integral is convergent, but the R integral diverges for α ≤ 2. For a finite system, the upper bound on the
R integral becomes the system size L. For α > 2 the domain wall energy is finite in the thermodynamic limit
3A corresponding mapping can be found between a cubic lattice and the linear chain as well.
6.2. ISING MODEL 241
L → ∞. In this case, entropy again wins. I.e. the entropy associated with a single domain wall is kB ln L, and
therefore F = E − kB T is always lowered by having a finite density of domain walls. For α < 2, the energy of a
single domain wall scales as L2−α . It was first proven by F. J. Dyson in 1969 that this model has a finite temperature
phase transition provided 1 < α < 2. There is no transition for α < 1 or α > 2. The case α = 2 is special, and is
discussed as a special case in the beautiful renormalization group analysis by J. M. Kosterlitz in Phys. Rev. Lett.
37, 1577 (1976).
Consider once again the ferromagnetic Ising model in zero field (H = 0), but on an arbitrary lattice. The partition
function is ( )
P
βJ hiji σi σj NL Y
Z = Tr e = cosh βJ Tr 1 + x σi σj , (6.43)
hiji
where x = tanh βJ and NL is the number of links. For regular lattices, NL = 21 zN , where N is the number of lattice
sites and z is the lattice coordination number, i.e. the number of nearest neighbors for each site. We have used
(
′
n o e+βJ if σσ ′ = +1
eβJσσ = cosh βJ · 1 + σσ ′ tanh βJ = (6.44)
e−βJ if σσ ′ = −1 .
We expand eqn. 6.43 in powers of x, resulting in a sum of 2NL terms, each of which can be represented graphically
in terms of so-called lattice animals. A lattice animal is a distinct (including reflections and rotations) arrangement
of adjacent plaquettes on a lattice. In order that the trace not vanish, only such configurations and their compo-
sitions are permitted. This is because each σi for every given site i must occur an even number of times in order
for a given term in the sum not to vanish. For all such terms, the trace is 2N . Let Γ represent a collection of lattice
animals, and gΓ the multiplicity of Γ . Then
N X L
Z = 2N cosh βJ L gΓ tanh βJ Γ , (6.45)
Γ
where LΓ is the total number of sites in the diagram Γ , and gΓ is the multiplicity of Γ . Since x vanishes as T → ∞,
this procedure is known as the high temperature expansion (HTE).
For the square lattice, he enumeration of all lattice animals with up to order eight is given in fig. 6.3. For the
diagram represented as a single elementary plaquette, there are N possible locations for the lower left vertex. For
the 2 × 1 plaquette animal, one has g = 2N , because there are two inequivalent orientations as well as N trans-
lations. For two disjoint elementary squares, one has g = 21 N (N − 5), which arises from subtracting 5N ‘illegal’
configurations involving double lines (remember each link in the partition sum appears only once!), shown in the
figure, and finally dividing by two because the individual squares are identical. Note that N (N − 5) is always
even for any integer value of N . Thus, to lowest interesting order on the square lattice,
2N n o
Z = 2N cosh βJ 1 + N x4 + 2N x6 + 7 − 52 N x8 + 12 N 2 x8 + O(x10 ) . (6.46)
again with x = tanh βJ. Note that we’ve substituted cosh2 βJ = 1/(1−x2 ) to write the final result as a power series
in x. Notice that the O(N 2 ) factor in Z has cancelled upon taking the logarithm, so the free energy is properly
extensive.
242 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Figure 6.3: HTE diagrams on the square lattice and their multiplicities.
Note that the high temperature expansion for the one-dimensional Ising chain yields
in agreement with the transfer matrix calculations. In higher dimensions, where there is a finite temperature phase
transition, one typically computes the specific heat c(T ) and tries to extract its singular behavior in the vicinity of
Tc , where c(T ) ∼ A (T − Tc )−α . Since x(T ) = tanh(J/kB T ) is analytic in T , we have c(x) ∼ A′ (x − xc )−α , where
xc = x(Tc ). One assumes xc is the singularity closest to the origin and corresponds to the radius of convergence
of the high temperature expansion. If we write
∞
X −α
n x ′′
c(x) = an x ∼ A 1 − , (6.49)
n=0
xc
an 1 1−α
= 1− . (6.50)
an−1 xc n
Thus, by plotting an /an−1 versus 1/n, one extracts 1/xc as the intercept, and (α − 1)/xc as the slope.
6.3. NONIDEAL CLASSICAL GASES 243
Figure 6.4: HTE diagrams for the numerator Ykl of the correlation function Ckl . The blue path connecting sites k
and l is the string. The remaining red paths are all closed loops.
Can we also derive a high temperature expansion for the spin-spin correlation function Ckl = hσk σl i ? Yes we
can. We have h P i
Tr σk σl eβJ hiji σi σj Y
Ckl = h P i ≡ kl . (6.51)
Tr eβJ hiji σi σj Z
Recall our analysis of the partition function Z. We concluded that in order for the trace not to vanish, the spin
variable σi on each site i must occur an even number of times in the expansion of the product. Similar considera-
tions hold for Ykl , except now due to the presence of σk and σl , those variables now must occur an odd number of
times when expanding the product. It is clear that the only nonvanishing diagrams will be those in which there
is a finite string connecting sites k and l, in addition to the usual closed HTE loops. See fig. 6.4 for an instructive
sketch. One then expands both Ykl as well as Z in powers of x = tanh βJ, taking the ratio to obtain the correlator
Ckl . At high temperatures (x → 0), both numerator and denominator are dominated by the configurations Γ with
the shortest possible total perimeter. For Z, this means the trivial path Γ = {∅}, while for Ykl this means finding
the shortest length path from k to l. (If there is no straight line path from k to l, there will in general be several such
minimizing paths.) Note, however, that the presence of the string between sites k and l complicates the analysis
of gΓ for the closed loops, since none of the links of Γ can intersect the string. It is worth stressing that this does
not mean that the string and the closed loops cannot intersect at isolated sites, but only that they share no common
links; see once again fig. 6.4.
Let’s switch gears now and return to the study of continuous classical systems described by a Hamiltonian
Ĥ {xi }, {pi } . In the next chapter, we will see how the critical properties of classical fluids can in fact be modeled
by an appropriate lattice gas Ising model, and we’ll derive methods for describing the liquid-gas phase transition
in such a model.
244 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Consider the ordinary canonical partition function for a nonideal system of identical point particles interacting via
a central two-body potential u(r). We work in the ordinary canonical ensemble. The N -particle partition function
is
Z YN
1 ddpi ddxi −Ĥ/kB T
Z(T, V, N ) = e
N! i=1
hd
(6.52)
dZ Y N
λ−N
T d 1 X
= d xi exp − u |xi − xj | .
N! i=1
kB T i<j
XN X
p2i
Ĥ = + u |xi − xj | , (6.53)
i=1
2m i<j
p
in which massive nonrelativistic particles interact via a two-body central potential. As before, λT = 2π~2 /mkB T
is the thermal wavelength. We can now write
Z(T, V, N ) = λ−N
T
d
QN (T, V ) , (6.54)
There are no general methods for evaluating the configurational integral exactly.
The Tonks gas is a one-dimensional generalization of the hard sphere gas. Consider a one-dimensional gas of
indistinguishable particles of mass m interacting via the potential
(
′ ∞ if |x − x′ | < a
u(x − x ) = (6.56)
0 if |x − x′ | ≥ a .
Let the gas be placed in a finite volume L. The hard sphere nature of the particles means that no particle can get
within a distance 12 a of the ends at x = 0 and x = L. That is, there is a one-body potential v(x) acting as well,
where
1
∞ if x < 2 a
v(x) = 0 if 21 a ≤ x ≤ L − 21 a (6.57)
∞ if x > L − 12 a .
ZL ZL
λ−N
Z(T, L, N ) = T dx1 · · · dxN χ(x1 , . . . , xN ) , (6.58)
N!
0 0
6.3. NONIDEAL CLASSICAL GASES 245
where χ = e−U/kB T is zero if any two ‘rods’ (of length a) overlap, or if any rod overlaps with either boundary at
x = 0 and x = L, and χ = 1 otherwise. Note that χ does not depend on temperature. Without loss of generality,
we can integrate over the subspace where x1 < x2 < · · · < xN and then multiply the result by N ! . Clearly xj must
lie to the right of xj−1 + a and to the left of Yj ≡ L − (N − j)a − 12 a. Thus,
YN −1
ZY1 ZY2 Z
−N
= λT dx1 dx2 · · · dxN −1 YN −1 − xN −1
a/2 x1 +a xN −2 +a
YN −2
ZY1 ZY2 Z
2
= λ−N
T dx1 dx2 · · · dxN −2 1
2 YN −2 − xN −2 = ···
a/2 x1 +a xN −3 +a
λ−N λ−NN
= T
X1 − 21 a T
(L − N a)N .
= (6.59)
N! N!
p
The λN
T factor comes from integrating over the momenta; recall λT = 2π~2 /mkB T .
The free energy is ( )
L
F = −kB T ln Z = −N kB T − ln λT + 1 + ln −a , (6.60)
N
where we have used Stirling’s rule to write ln N ! ≈ N ln N − N . The pressure is
∂F k T nkB T
p=− = LB = , (6.61)
∂L N −a
1 − na
where n = N/L is the one-dimensional density. Note that the pressure diverges as n approaches 1/a. The usual
one-dimensional ideal gas law, pL = N kB T , is replaced by pLeff = N kB T , where Leff = L − N a is the ‘free’ volume
obtained by subtracting the total ‘excluded volume’ N a from the original volume L.
Let us return to the general problem of computing the configuration integral. Consider the function e−βuij , where
uij ≡ u(|xi −xj |). We assume that at very short distances there is a strong repulsion between particles, i.e. uij → ∞
as rij = |xi − xj | → 0, and that uij → 0 as rij → ∞. Thus, e−βuij vanishes as rij → 0 and approaches unity as
rij → ∞. For our purposes, it will prove useful to define the function
Figure 6.5: Bottom panel: Lennard-Jones potential u(r) = 4ǫ x−12 − x−6 , with x = r/σ and ǫ = 1. Note the weak
attractive tail and the strong repulsive core. Top panel: Mayer function f (r, T ) = e−u(r)/kB T − 1 for kB T = 0.8 ǫ
(blue), kB T = 1.5 ǫ (green), and kB T = 5 ǫ (red).
This accounts for a long-distance attraction due to mutually induced electric dipole fluctuations, and a strong
short-ranged repulsion, phenomenologically modelled with a r−12 potential, which mimics a hard core due to
overlap of the atomic electron distributions. Setting u′ (r) = 0 we obtain r∗ = 21/6 σ ≈ 1.12246 σ at the minimum,
where u(r∗ ) = −ǫ. In contrast to the Boltzmann weight e−βu(r) , the Mayer function f (r) vanishes as r → ∞,
behaving as f (r) ∼ −βu(r). The Mayer function also depends on temperature. Sketches of u(r) and f (r) for the
Lennard-Jones model are shown in fig. 6.5.
The Lennard-Jones potential4 is realistic for certain simple fluids, but it leads to a configuration integral which is
in general impossible to evaluate. Indeed, even a potential as simple as that of the hard sphere gas is intractable in
more than one space dimension. We can however make progress by deriving a series expansion for the equation of
state in powers of the particle density. This is known as the virial expansion. As was the case when we investigated
noninteracting quantum statistics, it is convenient to work in the grand canonical ensemble and to derive series
expansions for the density n(T, z) and the pressure p(T, z) in terms of the fugacity z, then solve for z(T, n) to
obtain p(T, n). These expansions in terms of fugacity have a nifty diagrammatic interpretation, due to Mayer.
We begin by expanding the product in eqn. 6.63 as
Y X X
1 + fij = 1 + fij + fij fkl + . . . . (6.65)
i<j i<j i<j , k<l
(ij)6=(kl)
As there are 12 N (N − 1) possible pairings, there are 2N (N −1)/2 terms in the expansion of the above product. Each
such term may be represented by a graph, as shown in fig. 6.7. For each such term, we draw a connection
between dots representing different particles i and j if the factor fij appears in the term under consideration. The
contribution for any given graph may be written as a product over contributions from each of its disconnected
4 Disambiguation footnote: Take care not to confuse Philipp Lenard (Hungarian-German, cathode ray tubes, Nazi), Alfred-Marie Liénard
(French, Liénard-Wiechert potentials, not a Nazi), John Lennard-Jones (British, molecular structure, definitely not a Nazi), and Lynyrd Skynyrd
(American, ”Free Bird”, possibly killed by Nazis in 1977 plane crash). I thank my colleague Oleg Shpyrko for setting me straight on this.
6.3. NONIDEAL CLASSICAL GASES 247
Figure 6.6: Left: John Lennard-Jones. Center: Catherine Zeta-Jones. Right: James Earl Jones.
component clusters. For example, in the case of the term in fig. 6.7, the contribution to the configurational integral
would be
Z
1
∆Q = ddx1 ddx4 ddx7 ddx9 f1,4 f4,7 f4,9 f7,9
N!
Z Z Z (6.66)
d d d d d d d
× d x2 d x5 d x6 f2,5 f2,6 × d x3 d x10 f3,10 × d x8 d x11 f8,11 .
We will refer to a given product of Mayer functions which arises from this expansion as a term.
Figure 6.7: Diagrammatic interpretation of a term involving a product of eight Mayer functions.
The particular labels we assign to each vertex of a given graph don’t affect the overall value of the graph. Now
a given unlabeled graph consists of a certain number of connected subgraphs. For a system with N particles, we
may then write X
N= mγ n γ , (6.67)
γ
Note that the single vertex • counts as a connected subgraph, with n• = 1. We now ask: how many ways are
there of assigning the N labels to the N vertices of a given unlabeled graph? One might first thing the answer is
simply N !, however this is too big, because different assignments of the labels to the vertices may not result in a
distinct graph. To see this, consider the examples in fig. 6.8. In the first example, an unlabeled graph with four
vertices consists of two identical connected subgraphs. Given any assignment of labels to the vertices, then, we
248 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Figure 6.8: Different assignations of labels to vertices may not result in a distinct term in the expansion of the
configuration integral.
Q
can simply exchange the two subgraphs and get the same term. So we should divide N ! by the product γ mγ !.
But even this is not enough, because within each connected subgraph γ there may be permutations which leave
the integrand unchanged, as shown in the second and third examples in fig. 6.8. We define the symmetry factor
sγ as the number of permutations of the labels which leaves a given connected subgraphs γ invariant. Examples
of symmetry factors are shown in fig. 6.9. Consider, for example, the third subgraph in the top row. Clearly one
can rotate the figure about its horizontal symmetry axis to obtain a new labeling which represents the same term.
This twofold axis is the only symmetry the diagram possesses, hence sγ = 2. For the first diagram in the second
row, one can rotate either of the triangles about the horizontal symmetry axis. One can also rotate the figure in the
plane by 180◦ so as to exchange the two triangles. Thus, there are 2 × 2 × 2 = 8 symmetry operations which result
in the same term, and sγ = 8. Finally, the last subgraph in the second row consists of five vertices each of which is
connected to the other four. Therefore any permutation of the labels results in the same term, and sγ = 5! = 120.
Q Q m
In addition to dividing by the product γ mγ !, we must then also divide by γ sγ γ .
We can now write the partition function as
Z γ
!mγ
λ−N d X N! Y
d d
Y
Z= T Q mγ · d x1 · · · d xnγ fij · δN , P m , (6.68)
N! γ nγ
{mγ } m γ ! sγ γ i<j
Pthe last product is over all links in the subgraph γ. The final Kronecker delta enforces the constraint
where
N = γ mγ nγ . We next define the cluster integral bγ as
Z γ
Y
1 1 d d
bγ (T ) ≡ · d x1 · · · d xnγ fij . (6.69)
sγ V i<j
Qγ
Since fij = f |xi − xj | , the product i<j fij is invariant under simultaneous translation of all the coordinate
vectors by any constant vector, and hence the integral over the nγ position variables contains exactly one factor of
the volume, which cancels with the prefactor in the above definition of bγ . Thus, each cluster integral is intensive,
scaling as V 0 .5
If we compute the grand partition function, then the fixed N constraint is relaxed, and we can do the sums:
X P mγ nγ Y 1 m
Ξ = e−βΩ = eβµ λ−d
T V bγ γ
γ
mγ !
{mγ }
1 βµ −d mγ nγ
∞
Y X m
= e λT V bγ γ (6.70)
γ mγ =0
m γ !
X
nγ
= exp V eβµ λ−d
T b γ .
γ
Thus, X nγ
Ω(T, V, µ) = −V kB T eβµ λ−d
T bγ (T ) , (6.71)
γ
where z = exp(βµ) is the fugacity, and where b• ≡ 1. As in the case of ideal quantum gas statistical mechanics, we
can systematically invert the relation n = n(z, T ) to obtain z = z(n, T ), and then insert this into the equation for
p(z, T ) to obtain the equation of state p = p(n, T ). This yields the virial expansion of the equation of state,
n o
p = nkB T 1 + B2 (T ) n + B3 (T ) n2 + . . . . (6.73)
The virial coefficients Bj (T ) are obtained by summing a restricted set of cluster integrals, viz.
X
Bj (T ) = −(j − 1) bγ (T ) , (6.74)
γ∈Γj
250 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
where Γj is the set of one-particle irreducible clusters with j vertices. A one-particle irreducible cluster is one which
remains connected (i.e. it does not ‘fall apart’) if any of its sites are removed. Thus, the j = 3 cluster which is a
linear chain is one-particle reducible, because it becomes disconnected if the middle site is removed. The j = 3
cluster which is a triangle, however, is one-particle irreducible.
• The pressure and number density are written as an expansion over unlabeled connected clusters γ, viz.
X nγ
βp = zλ−d
T bγ
γ
X nγ
n= nγ zλ−d
T bγ .
γ
• For each term in each of these sums, draw the unlabeled connected cluster γ.
• Assign labels 1 , 2 , . . . , nγ to the vertices, where nγ is the total number of vertices in the cluster γ. It doesn’t
matter how you assign the labels.
Qγ
• Write down the product i<j fij . The factor fij appears in the product if there is a link in your (now labeled)
cluster between sites i and j.
• The symmetry factor sγ is the number of elements of the symmetric group Snγ which leave the product
Qγ
i<j fij invariant. The identity permutation always leaves the product invariant, so sγ ≥ 1.
Due to translation invariance, bγ (T ) ∝ V 0 . One can therefore set xnγ ≡ 0, eliminate the volume factor from
the denominator, and perform the integral over the remaining nγ −1 coordinates.
• This procedure generates expansions for p(T, z) and n(T, z) in powers of the fugacity z = eβµ . To obtain
something useful like p(T, n), we invert the equation
n = n(T, z) to find z = z(T, n), and then substitute into
the equation p = p(T, z) to obtain p = p T, z(T, n) = p(T, n). The result is the virial expansion,
n o
p = nkB T 1 + B2 (T ) n + B3 (T ) n2 + . . . ,
where
X
Bj (T ) = −(j − 1) bγ (T ) ,
γ∈Γj
We have
Z Z
1
b− (T ) = d
d x1 ddx2 f |x1 − x2 |
2V
Z (6.75)
= 2 ddr f (r)
1
and
Z Z Z
1
b∧ (T ) = ddx1 ddx2 ddx3 f |x1 − x2 | f |x1 − x3 |
2V
Z Z (6.76)
2
= 1
2 d r ddr′ f (r) f (r′ ) = 2 b−
d
and
Z Z Z
1
b△ (T ) = ddx1 ddx2 ddx3 f |x1 − x2 | f |x1 − x3 | f |x2 − x3 |
6V
Z Z (6.77)
= 1
6 d r ddr′ f (r) f (r′ ) f |r − r ′ | .
d
We invert by writing
zλ−d 2 3
T = n + α2 n + α3 n + . . . (6.80)
and substituting into the equation for n(z, T ), yielding
n = (n + α2 n2 + α3 n3 ) + 2(n + α2 n2 )2 b− + 3n3 b∧ + b△ + O(n4 ) . (6.81)
Thus,
0 = (α2 + 2b− ) n2 + (α3 + 4α2 b− + 3b∧ + 3b△ ) n3 + . . . . (6.82)
We therefore conclude
α2 = −2b− (6.83)
α3 = −4α2 b− − 3b∧ − 3b△
= 8b2− − 6b2− − 3b△ = 2b2− − 3b△ . (6.84)
We now insert eqn. 6.80 with the determined values of α2,3 into the equation for p(z, T ), obtaining
p
= n − 2b− n2 + (2b2− − 3b△ ) n3 + (n − 2b− n2 )2 b− + n3 (2b2− + b△ ) + O(n4 )
kB T (6.85)
= n − b− n2 − 2b△ n3 + O(n4 ) .
Thus,
B2 (T ) = −b− (T ) , B3 (T ) = −2b△ (T ) . (6.86)
Note that △ is the sole one-particle irreducible cluster with three vertices.
252 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Figure 6.10: The overlap of hard sphere Mayer functions. The shaded volume is V.
We must first compute the volume of overlap for spheres of radius a (recall a is the diameter of the constituent hard
sphere particles) centered at 0 and at ρ:
Z
V = d3r f (r) f |r − ρ|
Za (6.91)
=2 dz π(a2 − z 2 ) = 4π 3
3 a − πa2 ρ + π
12 ρ3 .
ρ/2
Thus, n o
2π 3 5π 2 6 2
p = nkB T 1 + 3 a n + 18 a n + O(n3 ) . (6.93)
Suppose (
∞ if r ≤ a
u(r) = (6.94)
−u0 (r) if r > a .
Then the corresponding Mayer function is
(
−1 if r ≤ a
f (r) = βu0 (r)
(6.95)
e − 1 if r > a .
Thus,
Z Z∞ h i
3 3
b− (T ) = 1
2 d r f (r) = − 2π
3 a + 2π dr r2 eβu0 (r) − 1 . (6.96)
a
Thus, the second virial coefficient is
Z∞
2π 3 2π
B2 (T ) = −b− (T ) ≈ 3 a − dr r2 u0 (r) , (6.97)
kB T
a
where we have assumed kB T ≪ u0 (r). We see that the second virial coefficient changes sign at some temperature
T0 , from a negative low temperature value to a positive high temperature value.
Figure 6.11: An attractive spherical well with a repulsive core u(r) and its associated Mayer function f (r).
To find the temperature T0 where B2 (T ) changes sign, we set B2 (T0 ) = 0 and obtain
3
s
kB T0 = ǫ ln 3 . (6.101)
s −1
Recall in our study of the thermodynamics of the Joule-Thompson effect in §1.10.6 that the throttling process is
isenthalpic. The temperature change, when a gas is pushed (or escapes) through a porous plug from a high pressure
region to a low pressure one is
Zp2
∂T
∆T = dp , (6.102)
∂p H
p1
where " #
∂T 1 ∂V
= T −V . (6.103)
∂p H Cp ∂T p
Appealing to the virial expansion, and working to lowest order in corrections to the ideal gas law, we have
N N2
p= kB T + 2 kB T B2 (T ) + . . . (6.104)
V V
∂V
and we compute ∂T p
by seting
N kB T N kB 2N 2 N2
0 = dp = − 2
dV + dT − 3 kB T B2 (T ) dV + 2 d kB T B2 (T ) + . . . . (6.105)
V V V V
B B 2B
∗
=A− ∗ =⇒ T∗ = . (6.108)
T T A
6.3. NONIDEAL CLASSICAL GASES 255
Consider a hard sphere gas in three dimensions in the presence of a hard wall at z = 0. The gas is confined to the
region z > 0. The total potential energy is now
X X
W (x1 , . . . , xN ) = v(xi ) + u(xi − xj ) , (6.109)
i i<j
where (
∞ if z ≤ 21 a
v(r) = v(z) = (6.110)
0 if z > 21 a ,
and u(r) is given in eqn. 6.87. The grand potential is written as a series in the total particle number N , and is
given by Z Z Z
′ ′
Ξ=e −βΩ
= 1+ξ d re 3 −βv(z)
+ 2 ξ d r d3r′ e−βv(z) e−βv(z ) e−βu(r−r ) + . . . ,
1 2 3
(6.111)
where ξ = z λ−3
T , with z = e
µ/kB T
the fugacity. Taking the logarithm, and invoking the Taylor series ln(1 + δ) =
1 2 1 3
δ − 2 δ + 3 δ − . . ., we obtain
Z Z Z h i
′
−βΩ = ξ d r + 2 ξ d r d3r′ e−βu(r−r ) − 1 + . . .
3 1 2 3
(6.112)
z> a
2 z> a ′ a
2 z >2
R
The volume is V = d3r. Dividing by V , we have, in the thermodynamic limit,
z>0
Z Z h i
βΩ 1 ′
− = βp = ξ + 21 ξ 2 d3r d3r′ e−βu(r−r ) − 1 + . . .
V V
z> a ′ a
2 z >2
(6.113)
= ξ − 32 πa3 ξ 2 + O(ξ 3 ) .
The number density is
∂
n=ξ(βp) = ξ − 34 πa3 ξ 2 + O(ξ 3 ) , (6.114)
∂ξ
and inverting to obtain ξ(n) and then substituting into the pressure equation, we obtain the lowest order virial
expansion for the equation of state, n o
p = kB T n + 32 πa3 n2 + . . . . (6.115)
As expected, the presence of the wall does not affect a bulk property such as the equation of state.
Next, let us compute the number density n(z), given by
X
n(z) = δ(r − ri ) . (6.116)
i
Due to translational invariance in the (x, y) plane, we know that the density must be a function of z alone. The
presence of the wall at z = 0 breaks translational symmetry in the z direction. The number density is
XN
β(µN̂−Ĥ)
n(z) = Tr e δ(r − ri ) Tr eβ(µN̂−Ĥ)
i=1
( Z )
=Ξ −1
ξe −βv(z) 2 −βv(z)
+ξ e 3 ′ −βv(z ′ ) −βu(r −r ′ )
dr e e + ... (6.117)
Z h i
′ ′
= ξe −βv(z) 2 −βv(z)
+ξ e d3r′ e−βv(z ) e−βu(r−r ) − 1 + . . . .
256 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Figure 6.12: In the presence of a hard wall, the Mayer sphere is cut off on the side closest to the wall. The resulting
density n(z) vanishes for z < 12 a since the center of each sphere must be at least one radius ( 21 a) away from the
wall. Between z = 21 a and z = 32 a there is a density enhancement. If the calculation were carried out to higher
order, n(z) would exhibit damped spatial oscillations with wavelength λ ∼ a.
′
Note that the term in square brackets in the last line is the Mayer function f (r − r ′ ) = e−βu(r−r ) − 1. Consider
the function
0 if z < 21 a or z ′ < 12 a
−βv(z) −βv(z ′ ) ′
e e f (r − r ) = 0 if |r − r ′ | > a (6.118)
1 ′ 1 ′
−1 if z > 2 a and z > 2 a and |r − r | < a .
Now consider the integral of the above function with respect to r ′ . Clearly the result depends on the value of z.
If z > 32 a, then there is no excluded region in r ′ and the integral is (−1) times the full Mayer sphere volume, i.e.
− 43 πa3 . If z < 12 a the integral vanishes due to the e−βv(z) factor. For z infinitesimally
larger than 12 a, the integral
is (−1) times half the Mayer sphere volume, i.e. − 3 πa . For z ∈ 2 , 2 the integral interpolates between − 23 πa3
2 3 a 3a
After substituting ξ = n + 34 πa3 n2 + O(n3 ) to relate ξ to the bulk density n = n∞ , we obtain the desired result:
0 h
if z < 12 a
3 z 1
1 z 1 3 i 2 3 2
n(z) = n + 1 − 2 a − 2 + 2 a − 2 · 3 πa n if 21 a < z < 32 a (6.120)
n if z > 3 a . 2
A sketch is provided in the right hand panel of fig. 6.12. Note that the density n(z) vanishes identically for z < 12
due to the exclusion of the hard spheres by the wall. For z between 12 a and 32 a, there is a density enhancement,
the origin of which has a simple physical interpretation. Since the wall excludes particles from the region z < 21 ,
there is an empty slab of thickness 21 z coating the interior of the wall. There are then no particles in this region to
exclude neighbors to their right, hence the density builds up just on the other side of this slab. The effect vanishes
to the order of the calculation past z = 23 a, where n(z) = n returns to its bulk value. Had we calculated to higher
order, we’d have found damped oscillations with spatial period λ ∼ a.
6.4. LEE-YANG THEORY 257
How can statistical mechanics describe phase transitions? This question was addressed in some beautiful mathe-
matical analysis by Lee and Yang6 . Consider the grand partition function Ξ,
∞
X
Ξ(T, V, z) = z N QN (T, V ) λ−dN
T , (6.121)
N =0
where Z Z
1
QN (T, V ) = dd x1 · · · dd xN e−U(x1 , ... , xN )/kB T (6.122)
N!
is the contribution to the N -particle partition function from the potential energy U (assuming no momentum-
dependent potentials). For two-body central potentials, we have
X
U (x1 , . . . , xN ) = v |xi − xj | . (6.123)
i<j
Suppose further that these classical particles have hard cores. Then for any finite volume, there must be some
maximum number NV such that QN (T, V ) vanishes for N > NV . This is because if N > NV at least two spheres
must overlap, in which case the potential energy is infinite. The theoretical maximum packing density for hard
π
spheres is achieved for a hexagonal close packed (HCP) lattice7 , for which fHCP = 3√ 2
= 0.74048. If the spheres
√ 3
have radius r0 , then NV = V /4 2r0 is the maximum particle number.
Thus, if V itself is finite, then Ξ(T, V, z) is a finite degree polynomial in z, and may be factorized as
NV NV
X Y z
Ξ(T, V, z) = N
z QN (T, V ) λ−dN
T = 1− , (6.124)
zk
N =0 k=1
where zk (T, V ) is one of the NV zeros of the grand partition function. Note that the O(z 0 ) term is fixed to be unity.
Note also that since the configuration integrals QN (T, V ) are all positive, Ξ(z) is an increasing function along the
positive real z axis. In addition, since the coefficients of z N in the polynomial Ξ(z) are all real, then Ξ(z) = 0
implies Ξ(z) = Ξ(z̄) = 0, so the zeros of Ξ(z) are either real and negative or else come in complex conjugate pairs.
For finite NV , the situation is roughly as depicted in the left panel of fig. 6.13, with a set of NV zeros arranged in
complex conjugate pairs (or negative real values). The zeros aren’t necessarily distributed along a circle as shown
in the figure, though. They could be anywhere, so long as they are symmetrically distributed about the Re(z) axis,
and no zeros occur for z real and nonnegative.
Lee and Yang proved the existence of the limits
p 1
= lim ln Ξ(T, V, z) (6.125)
kB T V →∞ V
∂ 1
n = lim z ln Ξ(T, V, z) , (6.126)
V →∞ ∂z V
6 See C. N. Yang and R. D. Lee, Phys. Rev. 87, 404 (1952) and ibid, p. 410
7 See e.g. http://en.wikipedia.org/wiki/Close-packing . For randomly close-packed hard spheres, one finds, from numerical simulations,
fRCP = 0.644.
258 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Figure 6.13: In the thermodynamic limit, the grand partition function can develop a singularity at positive real
fugacity z. The set of discrete zeros fuses into a branch cut.
where
1
φ(z) = − ln(z) (6.129)
V
is the complex potential due to a line charge of linear density λ = V −1 located at origin. The number density is
then
NV
∂ p ∂ X
n=z = −z φ(z − zk ) , (6.130)
∂z kB T ∂z
k=1
If we decompose the complex potential φ = φ1 + iφ2 into real and imaginary parts, the condition of analyticity is
recast as the Cauchy-Riemann equations,
Thus,
∂φ 1 ∂φ i ∂φ
− =− +
∂z 2 ∂x 2 ∂y
1 ∂φ1 ∂φ2 i ∂φ1 ∂φ2
=− + + −
2 ∂x ∂y 2 ∂y ∂x (6.133)
∂φ1 ∂φ1
=− +i
∂x ∂y
= Ex − iEy ,
where E = −∇φ1 is the electric field. Suppose, then, that as V → ∞ a continuous charge distribution develops,
which crosses the positive real z axis at a point x ∈ R+ . Then
n+ − n−
= Ex (x+ ) − Ex (x− ) = 4πσ(x) , (6.134)
x
where σ is the linear charge density (assuming logarithmic two-dimensional potentials), or the two-dimensional
charge density (if we extend the distribution along a third axis).
6.4.3 Example
(1 + z)M (1 − z M )
Ξ(z) =
1−z (6.135)
= (1 + z)M 1 + z + z 2 + . . . + z M−1 .
The (2M − 1) degree polynomial has an M th order zero at z = −1 and (M − 1) simple zeros at z = e2πik/M , where
k ∈ {1, . . . , M −1}. Since M serves as the maximum particle number NV , we may assume that V = M v0 , and the
V → ∞ limit may be taken as M → ∞. We then have
p 1
= lim ln Ξ(z)
kB T V →∞ V
1 1
= lim ln Ξ(z) (6.136)
v0 M→∞ M
1 1 M
= lim M ln(1 + z) + ln 1 − z − ln(1 − z) .
v0 M→∞ M
The limit depends on whether |z| > 1 or |z| < 1, and we obtain
|z| < 1
p v0 ln(1 + z)
if
= h i (6.137)
kB T
ln(1 + z) + ln z if |z| > 1 .
260 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Figure 6.14: Fugacity z and pv0 /kB T versus dimensionless specific volume v/v0 for the example problem discussed
in the text.
Thus,
1 z
∂
p
v0 ·
1+z if |z| < 1
n=z = h i (6.138)
∂z kB T
1
· z
+1 if |z| > 1 .
v0 1+z
The virial expansion is typically applied to low-density systems. When the density is high, i.e. when na3 ∼ 1,
where a is a typical molecular or atomic length scale, the virial expansion is impractical. There are to many terms
to compute, and to make progress one must use sophisticated resummation techniques to investigate the high
density regime.
To elucidate the physics of liquids, it is useful to consider the properties of various correlation functions. These
6.5. LIQUID STATE PHYSICS 261
The quantity
ddx1 ddp1 ddxN ddpN
· ·
f (x1 , . . . , xN ; p1 , . . . , pN ) · (6.143)
hd hd
is the propability of finding N particles in the system, with particle #1 lying within d3x1 of x1 and having momen-
tum within ddp1 of p1 , etc. If we compute averages of quantities which only depend on the positions {xj } and not
on the momenta {pj }, then we may integrate out the momenta to obtain, in the OCE,
1 −βW (x1 , ... , xN )
P (x1 , . . . , xN ) = Q−1
N · e , (6.144)
N!
where W is the total potential energy,
X X X
W (x1 , . . . , xN ) = v(xi ) + u(xi − xj ) + w(xi − xj , xj − xk ) + . . . , (6.145)
i i<j i<j<k
As in the case of the one-particle density matrix, i.e. the local density n1 (r), the two-particle density matrix satisfies
a sum rule: Z Z
d r1 ddr2 n2 (r1 , r2 ) = N (N − 1) .
d
(6.152)
The average potential energy can be expressed in terms of the distribution functions. Assuming only two-body
interactions, we have
X
hW i = u(xi − xj )
i<j
Z Z
X
= 1
2 ddr1 ddr2 u(r1 − r2 ) δ(r1 − xi ) δ(r2 − xj ) (6.155)
i6=j
Z Z
= 1
2 ddr1 ddr2 u(r1 − r2 ) n2 (r1 , r2 ) .
As the separations rij = |ri − rj | get large, we expect the correlations to vanish, in which case
X′
nk (r1 , . . . , rk ) = δ(r1 − xi1 ) · · · δ(rk − xik )
i1 ···ik
X ′
−−−−−→ δ(r1 − xi1 ) · · · δ(rk − xik )
rij →∞
i1 ···ik
(6.156)
N! 1
= · n (r ) · · · n1 (rk )
(N − k)! N k 1 1
1 2 k−1
= 1− 1− ··· 1 − n1 (r1 ) · · · n1 (rk ) .
N N N
6.5. LIQUID STATE PHYSICS 263
nk (r1 , . . . , rk )
gk (r1 , . . . , rk ) ≡ . (6.157)
n1 (r1 ) · · · n1 (rk )
For isotropic systems, the two-particle distribution function g2 (r1 , r2 ) depends only on the magnitude |r1 − r2 |.
As a function of this scalar separation, the function is known as the radial distribution function:
1
X
g(r) ≡ g2 (r) = 2
δ(r − xi ) δ(xj )
n
i6=j
(6.159)
1
X
= 2
δ(r − xi + xj ) .
Vn
i6=j
The radial distribution function is of great importance in the physics of liquids because
Z Z
= 12 n2 ddr1 ddr2 u(r1 − r2 ) g |r1 − r2 | (6.160)
Z
N2
= ddr u(r) g(r) .
2V
For a three-dimensional system, the average internal (i.e. potential) energy per particle is
Z∞
hW i
= 2πn dr r2 g(r) u(r) . (6.161)
N
0
Intuitively, f (r) dr ≡ 4πr2 n g(r) dr is the average number of particles lying at a radial distance between r and
r + dr from a given reference particle. The total potential energy of interaction with the reference particle is then
f (r) u(r) dr. Now integrate over all r and divide by two to avoid double-counting. This recovers eqn. 6.161.
In the OCE, g(r) obeys the sum rule
Z
V V
ddr g(r) = · N (N − 1) = V − , (6.162)
N2 N
hence Z
n ddr g(r) − 1 = −1 (OCE) . (6.163)
Figure 6.15: Pair distribution functions for hard spheres of diameter a at filling fraction η = π6 a3 n = 0.49 (left)
and for liquid Argon at T = 85 K (right). Molecular dynamics data for hard spheres (points) is compared with
the result of the Percus-Yevick approximation (see below in §6.5.8). Reproduced (without permission) from J.-P.
Hansen and I. R. McDonald, Theory of Simple Liquids, fig 5.5. Experimental data on liquid argon are from the
neutron scattering work of J. L. Yarnell et al., Phys. Rev. A 7, 2130 (1973). The data (points) are compared with
molecular dynamics calculations by Verlet (1967) for a Lennard-Jones fluid.
Z
"
#
3 N N (N − 1)
n d r h(r) = · V −V
V hN i2
2
2
N − N (6.164)
=
−1
N
= nkB T κT − 1 (GCE) ,
where κT is the isothermal compressibility. Note that in an ideal gas we have h(r) = 0 and κT = κ0T ≡ 1/nkB T .
R liquids and solids far from criticality, are nearly incompressible, hence 0 <
Self-condensed systems, such as
nkB T κT ≪ 1, and therefore n d3r h(r) ≈ −1. For incompressible systems, where κT = 0, this becomes an
equality.
As we shall see below in §6.5.4, the function h(r), or rather its Fourier transform ĥ(k), is directly measured in
a scattering experiment. The question then arises as to which result applies: the OCE result from eqn. 6.163 or
the GCE result from eqn. 6.164. The answer is that under almost all experimental conditions it is the GCE result
which applies. The reason for this is that the scattering experiment typically illuminates only a subset of the entire
system. This subsystem is in particle equilibrium with the remainder of the system, hence it is appropriate to use
the grand canonical ensemble. The OCE results would only apply if the scattering experiment were to measure
the entire system.
6.5. LIQUID STATE PHYSICS 265
Figure 6.16: Monte Carlo pair distribution functions for liquid water. From A. K. Soper, Chem Phys. 202, 295
(1996).
where Fi is the total force acting on particle i. If we average G over time, we obtain
ZT X
1
hGi = lim dt xi · Fi
T →∞ T
0 i
ZT X (6.166)
1
= − lim dt m ẋ2i
T →∞ T
0 i
= −3N kB T .
8 To derive this expression, note that F (surf) is directed inward and vanishes away from the surface. Each Cartesian direction α = (x, y, z)
(surf) (surf)
then contributes −Fα Lα , where Lα is the corresponding linear dimension. But Fα = pAα , where Aα is the area of the corresponding
face and p. is the pressure. Summing over the three possibilities for α, one obtains eqn. 6.168.
266 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Thus,
1 ∂Ω(ℓ3 V ) kB T 1 ∂Ξ(ℓ3 V )
p=− =
3V ∂ℓ 3V Ξ ∂ℓ
ℓ=1
" #
∞
kB T 1 X 1 Z Z X ∂W
N −βW (x1 , ... , xN )
= zλ−3
T d3
x1 · · · d3
xN e 3N − β x i ·
3V Ξ N! ∂xi
N =0 V V i
1 D ∂W E
= nkB T − . (6.175)
3V ∂ℓ ℓ=1
P
Finally, from W = i<j u(ℓxij ) we have
D ∂W E X
= xij · ∇u(xij )
∂ℓ ℓ=1
i<j
Z∞ (6.176)
2πN 2
= dr r3 g(r) u′ (r) ,
V
0
and hence
Z∞
2
p = nkB T − 2
3 πn dr r3 g(r) u′ (r) . (6.177)
0
Note that the density n enters the equation of state explicitly on the RHS of the above equation, but also implicitly
through the pair distribution function g(r), which has implicit dependence on both n and T .
6.5. LIQUID STATE PHYSICS 267
Figure 6.17: In a scattering experiment, a beam of particles interacts with a sample and the beam particles scatter
off the sample particles. A momentum ~q and energy ~ω are transferred to the beam particle during such a
collision. If ω = 0, the scattering is said to be elastic. For ω 6= 0, the scattering is inelastic.
Consider the scattering of a light or particle beam (i.e. photons or neutrons) from a liquid. We label the states of
the beam particles by their wavevector k and we assume a general dispersion εk . For photons, εk = ~c|k|, while
2 2
for neutrons εk = ~2mk . We assume a single scattering process with the liquid, during which the total momentum
n
and energy of the liquid plus beam are conserved. We write
k′ = k + q (6.178)
εk′ = εk + ~ω , (6.179)
where k′ is the final state of the scattered beam particle. Thus, the fluid transfers momentum ∆p = ~q and energy
~ω to the beam.
Now consider the scattering process between an initial state | i, k i and a final state | j, k′ i, where these states
describe both the beam and the liquid. According to Fermi’s Golden Rule, the scattering rate is
2π 2
Γik→j k′ = h j, k′ | V | i, k i δ(Ej − Ei + ~ω) , (6.180)
~
where V is the scattering potential and Ei is the initial internal energy of the liquid. If r is the position of the beam
particle and {xl } are the positions of the liquid particles, then
N
X
V(r) = v(r − xl ) . (6.181)
l=1
The differential scattering cross section (per unit frequency per unit solid angle) is
∂ 2σ ~ g(εk′ ) X
= P Γ ′ , (6.182)
∂Ω ∂ω 4π |vk | i,j i ik→j k
where Z
ddk
g(ε) = δ(ε − εk ) (6.183)
(2π)d
268 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
where we have assumed that the incident and scattered beams are plane waves. We then have
N
∂ 2σ ~ g(εk+q ) |v̂(q)|2 X X
X
2
= 2
Pi j e−iq·xl i δ(Ej − Ei + ~ω)
∂Ω ∂ω 2 |∇ ε | V
k k i j l=1
(6.186)
g(εk+q ) N
= 2
|v̂(q)|2 S(q, ω) ,
4π |∇ ε | V
k k
where S(q, ω) is the dynamic structure factor,
N
2π~ X X
X
j
2
S(q, ω) = Pi e−iq·xl i δ(Ej − Ei + ~ω) (6.187)
N i j l=1
Thus,
Z∞ X
X
1
S(q, ω) = dt eiωt Pi i eiq·xl (0) e−iq·xl′ (t) i
N
−∞ i l,l′
(6.189)
Z∞
1
X
= dt e iωt
eiq·xl (0) e−iq·xl′ (t)
,
N
−∞ l,l′
where the angular brackets in the last line denote a thermal expectation value of a quantum mechanical operator.
If we integrate over all frequencies, we obtain the equal time correlator,
Z∞
dω 1 X
iq·(xl −x ′ )
S(q) = S(q, ω) = e l
2π N ′
−∞ l,l (6.190)
Z
= N δq,0 + 1 + n ddr e−iq·r g(r) − 1 .
6.5. LIQUID STATE PHYSICS 269
Figure 6.18: Comparison of the static structure factor as determined by neutron scattering work of J. L. Yarnell et
al., Phys. Rev. A 7, 2130 (1973) with molecular dynamics calculations by Verlet (1967) for a Lennard-Jones fluid.
known as the static structure factor9 . Note that S(q = 0) = N , since all the phases eiq·(xi −xj ) are then unity. As
q → ∞, the phases oscillate rapidly with changes in the distances |xi − xj |, and average out to zero. However, the
‘diagonal’ terms in the sum, i.e. those with i = j, always contribute a total of 1 to S(q). Therefore in the q → ∞
limit we have S(q → ∞) = 1.
In general, the detectors used in a scattering experiment are sensitive to the energy of the scattered beam particles,
although there is always a finite experimental resolution, both in q and ω. This means that what is measured is
actually something like Z Z
Smeas (q, ω) = ddq ′ dω ′ F (q − q ′ ) G(ω − ω ′ ) S(q ′ , ω ′ ) , (6.191)
where F and G are essentially Gaussian functions of their argument, with width given by the experimental reso-
lution. If one integrates over all frequencies ω, i.e. if one simply counts scattered particles as a function of q but
without any discrimination of their energies, then one measures the static structure factor S(q). Elastic scattering
is determined by S(q, ω = 0, i.e. no energy transfer.
where Z Z
1 P
QN [v] = ddx1 · · · ddxN e−βW (x1 , ... , xN ) e−β i v(xi ) . (6.193)
N!
1
9 We may write δq ,0 = V
(2π)d δ(q ).
270 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Using Z
X X
v(xi ) = ddr v(r) δ(r − xi ) , (6.196)
i i
hence
Z Z X
δF
= ddx1 · · · ddxN P (x1 , . . . , xN ) δ(r − xi )
δv(r) i
= n1 (r) , (6.197)
which is the local density at r.
Next, consider the response function,
δn1 (r) δ 2F [v]
χ(r, r ′ ) ≡ =
δv(r ′ ) δv(r) δv(r ′ )
1 1 δQN δQN 1 1 δ 2QN (6.198)
= · 2 − ·
β QN δv(r) δv(r ′ ) β QN δv(r) δv(r ′ )
= β n1 (r) n1 (r ′ ) − β n1 (r) δ(r − r ′ ) − β n2 (r, r ′ ) .
In an isotropic system, χ(r, r ′ ) = χ(r − r ′ ) is a function of the coordinate separation, and
−kB T χ(r − r ′ ) = −n2 + n δ(r − r ′ ) + n2 g |r − r ′ |
(6.199)
= n h |r − r | + n δ(r − r ′ ) .
2 ′
Note that if v(r) > 0 it becomes energetically more costly for a particle to be at r. Accordingly, the density response
is negative, and proportional to the ratio v(r)/kB T – this is the first term in the above equation. If there were no
correlations between the particles, then h = 0 and this would be the entire story. However, the particles in general
are correlated. Consider, for example, the case of hard spheres of diameter a, and let there be a repulsive potential
at r = 0. This means that it is less likely for a particle to be centered anywhere within a distance a of the origin.
But then it will be more likely to find a particle in the next ‘shell’ of radial thickness a.
The distribution functions satisfy a hierarchy of integro-differential equations known as the BBGKY hierarchy10 . In
homogeneous systems, we have
Z Z
N! 1
gk (r1 , . . . , rk ) = d xk+1 · · · ddxN P (r1 , . . . , rk , xk+1 , . . . , xN ) ,
d
(6.203)
(N − k)! nk
where
1 1 −βW (x1 , ... , xN )
P (x1 , . . . , xN ) = · e . (6.204)
QN N !
X N
X −1 N
X
u(xij ) ≡ u xi − xj
k<i<j i=k+1 j=i+1
X k−1
X k
X
u(rij ) ≡ u ri − rj
i<j≤k i=1 j=i+1
X k
X N
X
u(ri − xj ) = u(ri − xj ) .
i≤k<j i=1 j=k+1
Now
P P
∂
e−β i<j≤k u(rij ) · e−β i≤k<j u(ri −xj ) = (6.206)
∂r1
( )
X ∂u(r1 − rj ) X ∂u(r1 − rj ) P P
−β i<j≤k u(rij ) −β i≤k<j u(ri −xj )
β + · e ·e ,
∂r1 ∂r1
1<j≤k k<j
hence
Xk
∂ ∂u(r1 − rj )
gk (r1 , . . . , rk ) = −β gk (r1 , . . . , rk ) (6.207)
∂r1 j=2
∂r1
Z
∂u(r1 − xk+1 )
− β(N − k) ddxk+1 P (r1 , . . . , rk , xk+1 , . . . , xN )
∂r1
k
X ∂u(r1 − rj )
= −β gk (r1 , . . . , rk ) (6.208)
j=2
∂r1
Z
∂u(r1 − xk+1 )
+ n ddxk+1 gk+1 (r1 , . . . , rk , xk+1 )
∂r1
Thus, we obtain the BBGKY hierarchy:
k
X ∂u(r1 − rj )
∂
−kB T gk (r1 , . . . , rk ) = gk (r1 , . . . , rk ) (6.209)
∂r1 j=2
∂r1
Z
∂u(r1 − r ′ )
+ n ddr′ gk+1 (r1 , . . . , rk , r ′ ) .
∂r1
The BBGKY hierarchy is an infinite tower of coupled integro-differential equations, relating gk to gk+1 for all k. If
we approximate gk at some level k in terms of equal or lower order distributions, then we obtain a closed set of
equations which in principle can be solved, at least numerically. For example, the Kirkwood approximation closes
the hierarchy at order k = 2 by imposing the condition
This is known as the Born-Green-Yvon (BGY) equation. In practice, the BGY equation, which is solved numerically,
gives adequate results only at low densities.
where h(r) = g(r) − 1 and we assume an isotropic system. This is called the Ornstein-Zernike equation. The first
term, c(r), accounts for local correlations, which are then propagated in the second term to account for long-
ranged correlations.
The OZ equation is an integral equation, but it becomes a simple algebraic one upon Fourier transforming:
1 + n ĥ(0) 1 1 κ0T
κT = = · =⇒ n ĉ(0) = 1 − , (6.216)
nkB T nkB T 1 − n ĉ(0) κT
where κ0T = 1/nkB T is the ideal gas isothermal compressibility.
At this point, we have merely substituted one unknown function, h(r), for another, namely c(r). To close the
system, we need to relate c(r) to h(r) again in some way. There are various approximation schemes which do just
this.
This is the Percus-Yevick equation. Remarkably, the Percus-Yevick (PY) equation can be solved analytically for the
case of hard spheres, where u(r) = ∞ for r ≤ a and u(r) = 0 for r > a, where a is the hard sphere diameter. Define
the function y(r) = eβu(r) g(r), in which case
(
−y(r) , r ≤ a
c(r) = y(r) f (r) = (6.219)
0 , r>a.
Here, f (r) = e−βu(r) − 1 is the Mayer function. We remark that the definition of y(r) may cause some concern for
the hard sphere system, because of the eβu(r) term, which diverges severely for r ≤ a. However, g(r) vanishes in
this limit, and their product y(r) is in fact finite! The PY equation may then be written for the function y(r) as
Z Z
y(r) = 1 + n d3r′ y(r ′ ) − n d3r′ y(r ′ ) y(r − r ′ ) . (6.220)
r ′ <a r′ <a
|r −r ′ |>a
This has been solved using Laplace transform methods by M. S. Wertheim, J. Math. Phys. 5, 643 (1964). The final
result for c(r) is r r 3
c(r) = − λ1 + 6η λ2 + 21 η λ1 · Θ(a − r) , (6.221)
a a
where η = 16 πa3 n is the packing fraction and
(1 + 2η)2 (1 + 21 η)2
λ1 = , λ2 = − . (6.222)
(1 − η)4 (1 − η)4
This leads to the equation of state
1 + η + η2
p = nkB T · . (6.223)
(1 − η)3
274 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Table 6.1: Comparison of exact (Monte Carlo) results to those of the Percus-Yevick (PY) and hypernetted chains
approximation (HCA) for hard spheres in three dimensions. Sources: Hansen and McDonald (1990) and Reichl
(1998)
This gets B2 and B3 exactly right. The accuracy of the PY approximation for higher order virial coefficients is
shown in table 6.1.
To obtain the equation of state from eqn. 6.221, we invoke the compressibility equation,
∂n 1
nkB T κT = = . (6.224)
∂p T 1 − n ĉ(0)
We therefore need
Z
ĉ(0) = d3r c(r)
Z1
(6.225)
3
= −4πa dx x2 λ1 + 6 η λ2 x + 1
2 η λ1 x3
0
1
= −4πa3 3 λ1 + 3
2 η λ2 + 1
12 η λ1 .
With η = 61 πa3 n and using the definitions of λ1,2 in eqn. 6.222, one finds
1 + 4η + 4η 2
1 − n ĉ(0) = . (6.226)
(1 − η)4
We then have, from the compressibility equation,
6kB T ∂p 1 + 4η + 4η 2
= . (6.227)
πa3 ∂η (1 − η)4
Integrating, we obtain p(η) up to a constant. The constant is set so that p = 0 when n = 0. The result is eqn. 6.223.
Another commonly used scheme is the hypernetted chains (HNC) approximation, for which
c(r) = −βu(r) + h(r) − ln 1 + h(r) . (6.228)
The rationale behind the HNC and other such approximation schemes is rooted in diagrammatic approaches,
which are extensions of the Mayer cluster expansion to the computation of correlation functions. For details and
references to their application in the literature, see Hansen and McDonald (1990) and Reichl (1998).
Let’s expand the direct correlation function ĉ(q) in powers of the wavevector q, viz.
ĉ(q) = ĉ(0) + c2 q 2 + c4 q 4 + . . . . (6.229)
6.5. LIQUID STATE PHYSICS 275
In higher dimensions d > 1 we can obtain the result asymptotically in two limits:
At criticality, ξ → ∞, and clearly our results in d = 1 and d = 2 dimensions are nonsensical, as they are divergent.
To correct this behavior, M. E. Fisher in 1963 suggested that the OZ correlation functions in the r ≪ ξ limit be
replaced by
ξη e−r/ξ
χ(r) ≃ −Cd′′ n · · , (6.240)
kB T R2 rd−2+η
a result known as anomalous scaling. Here, η is the anomalous scaling exponent.
Recall that the isothermal compressibility is given by κT = −χ̂(0). Near criticality, the integral in χ̂(0) is dominated
by the r ≪ ξ part, since ξ → ∞. Thus, using Fisher’s anomalous scaling,
Z
κT = −χ̂(0) = − ddr χ(r)
Z (6.241)
e−r/ξ −(2−η)ν
∼ A ddr d−2+η ∼ B ξ 2−η ∼ C T − Tc ,
r
where A, B, and C are temperature-dependent constants which are nonsingular at T = Tc . Thus, since κT ∝
|T − Tc |−γ , we conclude
γ = (2 − η) ν , (6.242)
a result known as hyperscaling.
Coulomb systems are particularly interesting in statistical mechanics because of their long-ranged forces, which
result in the phenomenon of screening. Long-ranged forces wreak havoc with the Mayer cluster expansion, since
the Mayer function is no longer integrable. Thus, the virial expansion fails, and new techniques need to be applied
to reveal the physics of plasmas.
The potential energy of a Coulomb system is
Z Z
U = 21 ddr ddr′ ρ(r) u(r − r ′ ) ρ(r ′ ) , (6.243)
where ρ(r) is the charge density and u(r), which has the dimensions of (energy)/(charge)2, satisfies
Thus,
−2π |x − x′ | , d=1
u(r) = −2 ln |r − r ′ | , d = 2 (6.245)
|r − r ′ |−1 , d=3.
where qi is the charge of the ith particle. We will assume two types of charges: q = ±e, with e > 0. The electric
potential is
Z
φ(r) = ddr′ u(r − r ′ ) ρ(r ′ )
X (6.247)
= qi u(r − xi ) .
i
We then have
Z
U= 1
2 ddr ρav (r) + δρ(r) · φav (r) + δφ(r)
≡ U0 ignore fluctuation term (6.254)
z Z }| { Z zZ }| {
1 d av av d av 1 d
= − 2 d r ρ (r) φ (r) + d r φ (r) ρ(r)+ 2 d r δρ(r) δφ(r) .
We apply the mean field approximation in each region of space, which leads to
Z
e φav (r)
Ω(T, V, µ+ , µ− ) = −kB T λ−d
+ z + d d
r exp −
kB T
Z (6.255)
e φav (r)
− kB T λ−d
− z − dd
r exp + ,
kB T
where
2π~2 µ±
λ± = , z± = exp . (6.256)
m± kB T kB T
278 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Debye-Hückel theory is valid provided n∞ λ3D ≫ 1, so that the statistical assumption of many charges in a screen-
ing volume is justified.
6.6. COULOMB SYSTEMS : PLASMAS AND THE ELECTRON GAS 279
Assuming kB T ≪ εF , thermal fluctuations are unimportant and we may assume T = 0. In the same spirit as the
Debye-Hückel approach, we assume a slowly varying mean electrostatic potential φ(r). Locally, we can write
~2 kF2
εF = − eφ(r) . (6.268)
2m
Thus, the Fermi wavevector kF is spatially varying, according to the relation
1/2
2m
kF (r) = ε + eφ(r) . (6.269)
~2 F
The local electron number density is
3/2
kF3 (r) eφ(r)
n(r) = = n∞ 1 + . (6.270)
3π 2 εF
In the presence of a uniform compensating positive background charge ρ+ = en∞ , Poisson’s equation takes the
form " 3/2 #
eφ(r)
∇2 φ = 4πe n∞ · 1+ − 1 − 4πρext (r) . (6.271)
εF
If eφ ≪ εF , we may expand in powers of the ratio, obtaining
6πn∞ e2
∇2 φ = φ ≡ κ2TF φ − 4πρext (r) . (6.272)
εF
Here, κTF is the Thomas-Fermi wavevector,
1/2
6πn∞ e2
κTF = . (6.273)
εF
Thomas-Fermi theory is valid provided n∞ λ3TF ≫ 1, where λTF = κ−1
TF , so that the statistical assumption of many
electrons in a screening volume is justified.
One important application of Thomas-Fermi screening is to the theory of metals. In a metal, the outer, valence
electrons of each atom are stripped away from the positively charged ionic core and enter into itinerant, plane-
wave-like states. These states disperse with some ε(k) function (that is periodic in the Brillouin zone, i.e. under
k → k + G, where G is a reciprocal lattice vector), and at T = 0 this energy band is filled up to the Fermi level εF ,
as Fermi statistics dictates. (In some cases, there may be several bands at the Fermi level, as we saw in the case
of yttrium.) The set of ionic cores then acts as a neutralizing positive background. In a perfect crystal, the ionic
cores are distributed periodically, and the positive background is approximately uniform. A charged impurity in
a metal, such as a zinc atom in a copper matrix, has a different nuclear charge and a different valency than the
host. The charge of the ionic core, when valence electrons are stripped away, differs from that of the host ions,
and therefore the impurity acts as a local charge impurity. For example, copper has an electronic configuration of
[Ar] 3d10 4s1 . The 4s electron forms an energy band which contains the Fermi surface. Zinc has a configuration of
[Ar] 3d10 4s2 , and in a Cu matrix the Zn gives up its two 4s electrons into the 4s conduction band, leaving behind
a charge +2 ionic core. The Cu cores have charge +1 since each copper atom contributed only one 4s electron to
the conduction band. The conduction band electrons neutralize the uniform positive background of the Cu ion
cores. What is left is an extra Q = +e nuclear charge at the Zn site, and one extra 4s conduction band electron.
The Q = +e impurity is, however, screened by the electrons, and at distances greater than an atomic radius the
potential that a given electron sees due to the Zn core is of the Yukawa form,
Q −κTF r
φ(r) = ·e . (6.274)
r
280 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
We should take care, however, that the dispersion ε(k) for the conduction band in a metal is not necessarily of the
free electron form ε(k) = ~2 k2 /2m. To linear order in the potential, however, the change in the local electronic
density is
δn(r) = eφ(r) g(εF ) , (6.275)
where g(εF ) is the density of states at the Fermi energy. Thus, in a metal, we should write
∇2 φ = (−4π)(−e δn)
(6.276)
= 4πe2 g(εF ) φ = κ2TF φ ,
where q
κTF = 4πe2 g(εF ) . (6.277)
The value of g(εF ) will depend on the form of the dispersion. For ballistic bands with an effective mass m∗ , the
formula in eqn. 6.272 still applies.
Consider an ion formed of a nucleus of charge +Ze and an electron cloud of charge −N e. The net ionic charge is
then (Z − N )e. Since we will be interested in atomic scales, we can no longer assume a weak field limit and we
must retain the full nonlinear screening theory, for which
(2m)3/2 3/2
∇2 φ(r) = 4πe · 2 3
εF + eφ(r) − 4πZe δ(r) . (6.278)
3π ~
We assume an isotropic solution. It is then convenient to define
Ze2
εF + eφ(r) = · χ(r/r0 ) , (6.279)
r
where r0 is yet to be determined. As r → 0 we expect χ → 1 since the nuclear charge is then unscreened. We then
have 2
2 Ze 1 Ze2 ′′
∇ · χ(r/r0 ) = 2 χ (r/r0 ) , (6.280)
r r0 r
thus we arrive at the Thomas-Fermi equation,
1
χ′′ (t) = √ χ3/2 (t) , (6.281)
t
χ(0) = 1 . (6.283)
6.7. POLYMERS 281
Figure 6.19: The Thomas-Fermi atom consists of a nuclear charge +Ze surrounded by N electrons distributed in a
cloud. The electric potential φ(r) felt by any electron at position r is screened by the electrons within this radius,
resulting in a self-consistent potential φ(r) = φ0 + (Ze2 /r) χ(r/r0 ).
• For positive ions, with N < Z, there is perfect screening at the ionic boundary R = t∗ r0 , where χ(t∗ ) = 0.
This requires
Ze2 Ze2 ′ (Z − N ) e
E = −∇φ = − 2 χ(R/r0 ) + χ (R/r0 ) r̂ = r̂ . (6.284)
R R r0 R2
This requires
N
−t∗ χ′ (t∗ ) = 1 − . (6.285)
Z
For an atom, with N = Z, the asymptotic solution to the TF equation is a power law, and by inspection is found
to be χ(t) ∼ C t−3 , where C is a constant. The constant follows from the TF equation, which yields 12 C = C 3/2 ,
hence C = 144. Thus, a neutral TF atom has a density with a power law tail, with ρ ∼ r−6 . TF ions with N > Z
are unstable.
6.7 Polymers
Linear chain polymers are repeating structures with the chemical formula (A)x , where A is the formula unit and
x is the degree of polymerization. In many cases (e.g. polystyrene), x > 5
∼ 10 is not uncommon. For a very readable
introduction to the subject, see P. G. de Gennes, Scaling Concepts in Polymer Physics.
Quite often a given polymer solution will contain a distribution of x values; this is known as polydispersity. Various
preparation techniques, such as chromatography, can mitigate the degree of polydispersity. Another morpholog-
ical feature of polymers is branching, in which the polymers do not form linear chains.
Polymers exhibit a static flexibility which can be understood as follows. Consider a long chain hydrocarbon with a
−C − C − C− backbone. The angle between successive C − C bonds is fixed at θ ≈ 68◦ , but the azimuthal angle
282 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
ϕ can take one of three possible low-energy values, as shown in the right panel of fig. 6.21. Thus, the relative
probabilities of gauche and trans orientations are
Prob (gauche)
= 2 e−∆ε/kB T (6.286)
Prob (trans)
where ∆ε is the energy difference between trans and gauche configurations. This means that the polymer chain is
in fact a random coil with a persistence length
ℓp = ℓ0 e∆ε/kB T (6.287)
where ℓ0 is a microscopic length scale, roughly given by the length of a formula unit, which is approximately a
few Ångstroms (see fig. 6.22). Let L be the total length of the polymer when it is stretched into a straight line. If
ℓp > L, the polymer is rigid. If ℓp ≪ L, the polymer is rigid on the length scale ℓp but flexible on longer scales. We
have
ℓp 1 ∆ε/kB T
= e , (6.288)
L N
where we now use N (rather than x) for the degree of polymerization.
In the time domain, the polymer exhibits a dynamical flexibility on scales longer than a persistence time. The persis-
tence time τp is the time required for a trans-gauche transition. The rate for such transitions is set by the energy
barrier B separating trans from gauche configurations:
τp = τ0 eB/kB T (6.289)
where τ0 ∼ 10−11 s. On frequency scales ω ≪ τp−1 the polymer is dynamically flexible. If ∆ε ∼ kB T ≪ B the
polymer is flexible from a static point of view, but dynamically rigid. That is, there are many gauche orientations
of successive carbon bonds which reflect a quenched disorder. The polymer then forms a frozen random coil, like
a twisted coat hanger.
6.7. POLYMERS 283
Figure 6.21: Left” trans and gauche orientations in carbon chains. Right: energy as a function of azimuthal angle ϕ.
There are three low energy states: trans (ϕ = 0) and gauche (ϕ = ±ϕ0 ).
A polymer can be modeled by a self-avoiding random walk (SAW). That is, on scales longer than ℓp , it twists about
randomly in space subject to the constraint that it doesn’t overlap itself. Before we consider the mathematics of
SAWs, let’s first recall some aspects of ordinary random walks which are not self-avoiding.
We’ll simplify matters further by considering random walks on a hypercubic lattice of dimension d. Such a lattice
has coordination number 2d, i.e. there are 2d nearest neighbor separations, δ = ±a ê1 , ±a ê2 , . . . , ±a êd , where a
is the lattice spacing. Consider now a random walk of N steps starting at the origin. After N steps the position is
N
X
RN = δj (6.290)
j=1
where δj takes on one of 2d possible values. Now N is no longer the degree of polymerization, but somthing
approximating L/ℓp , which is the number of persistence lengths in the chain. We assume each step is independent,
hence hδjα δjβ′ i = (a2 /d) δjj ′ δ αβ and R2N = N a2 . The full distribution PN (R) is given by
X X
PN (R) = (2d)−N ··· δ R, P δ
j j
δ1 δN
Zπ/a π/a
Z " d #N
d dk1 dkd −ik·R 1 X
=a ··· e cos(kµ a)
2π 2π d µ=1
−π/a −π/a
" (6.291)
Z d
#
d k −ik·R 1 2 2
= ad e exp N ln 1 − k a + ...
(2π)d 2d
Ω̂
d Z d/2
a −N k2 a2 /2d −ik·R d 2
/2N a2
≈ dke d
e = e−dR .
2d 2πN
This is a simple Gaussian, with width R2 = d·(N a2 /d) = N a2 , as we have already computed. The quantity R
284 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
√
defined here is the end-to-end vector of the chain. The RMS end-to-end distance is then hR2 i1/2 = N a ≡ R0 .
A related figure of merit is the radius of gyration, Rg , defined by
1 DX 2 E
N
Rg2 = Rn − RCM , (6.292)
N n=1
1 PN
where RCM = N j=1 Rj is the center of mass position. A brief calculation yields
N a2
Rg2 = N + 3 − 4N −1 a2 ∼ , (6.293)
6
in all dimensions.
The total number of random walk configurations with end-to-end vector R is then (2d)N PN (R), so the entropy of
a chain at fixed elongation is
h i dk R2
S(R, N ) = kB ln (2d)N PN (R) = S(0, N ) − B 2 . (6.294)
2N a
If we assume that the energy of the chain is conformation independent, then E = E0 (N ) and
dkB T R2
F (R, N ) = F (0, N ) + . (6.295)
2N a2
In the presence of an external force Fext , the Gibbs free energy is the Legendre transform
Structure factor
Pm
Now for m > n we have Rm − Rn = j=n+1 δj , and therefore
D 2 E Xm
1
k · (Rm − Rn ) = (k · δj )2 = (m − n) k2 a2 , (6.301)
j=n+1
d
where µk = k2 a2 /2d. In the limit where N → ∞ and a → 0 with N a2 = R02 constant, the structure factor has a
scaling form, S(k) = N f (N µk ) = (R0 /a) f (k 2 R02 /2d) , where
2 −x x x2
f (x) = 2
e −1+x = 1− + + ... . (6.303)
x 3 12
What is missing from the random walk free energy is the effect of steric interactions. An argument due to Flory
takes these interactions into account in a mean field treatment. Suppose we have a chain of radius R. Then the
average monomer density within the chain is c = N/Rd. Assuming short-ranged interactions, we should then add
a term to the free energy which effectively counts the number of near self-intersections of the chain. This number
should be roughly N c. Thus, we write
N2 1 R2
F (R, N ) = F0 + u(T ) + 2 dkB T . (6.304)
Rd N a2
The effective interaction u(T ) is positive in the case of a so-called ‘good solvent’.
The free energy is minimized when
∂F dvN 2 R
0= = − d+1 + dkB T , (6.305)
∂R R N a2
which yields the result
1/(d+2)
ua2
RF (N ) = N 3/(d+2) ∝ N ν . (6.306)
kB T
Thus, we obtain ν = 3/(d + 2). In d = 1 this says ν = 1, which is exactly correct because a SAW in d = 1 has no
option but to keep going in the same direction. In d = 2, Flory theory predicts ν = 43 , which is also exact. In d = 3,
286 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
3
10
Rg / nm
2
10
1
10
5 6 7 8
10 10 10 10
M / (g/mol)
Figure 6.23: Radius of gyration Rg of polystyrene in a toluene and benzene solvent, plotted as a function of
molecular weight of the polystyrene. The best fit corresponds to a power law Rg ∝ M ν with ν = 0.5936. From J.
Des Cloizeaux and G. Jannink, Polymers in Solution: Their Modeling and Structure (Oxford, 1990).
we have νd=3 = 53 , which is extremely close to the numerical value ν = 0.5880. Flory theory is again exact at the
SAW upper critical dimension, which is d = 4, where ν = 21 , corresponding to a Gaussian random walk11 . Best.
Mean. Field. Theory. Ever.
How well are polymers described as SAWs? Fig. 6.23 shows the radius of gyration Rg versus molecular weight M
for polystyrene chains in a toluene and benzene solvent. The slope is ν = d ln Rg /d ln M = 0.5936. Experimental
results can vary with concentration and temperature, but generally confirm the validity of the SAW model.
For a SAW under an external force, we compute the Gibbs partition function,
Z Z
Y (Fext , N ) = ddR PN (R) eFext ·R/kB T = ddx f (x) esn̂·x , (6.307)
where x = R/RF and s = kB T /RF Fext and n̂ = F̂ext . One than has R(Fext ) = RF Φ(RF /ξ), where ξ = kB T /Fext
2
and R(Fext ) = Fext RF /kB T . For small values of its argument one has Φ(u) ∝ u. For large u it can be shown that
R(Fext ) ∝ (Fext RF /kB T )2/3 .
On a lattice of coordination number z, the number of N -step random walks starting from the origin is ΩN = z N .
If we constrain our random walks to be self-avoiding, the number is reduced to
γ−1 N
N = CN
ΩSAW y , (6.308)
where C and γ are dimension-dependent constants, and we expect y < ∼ z − 1, since at the very least a SAW cannot
immediately double back on itself. In fact, on the cubic lattice one has z = 6 but y = 4.68, slightly less than z − 1.
One finds γd=2 ≃ 43 and γd=3 ≃ 76 . The RMS end-to-end distance of the SAW is
RF = a N ν , (6.309)
where a and ν are d-dependent constants,with νd=1 = 1, νd=2 ≃ 43 , and νd=3 ≃ 53 . The distribution PN (R) has a
scaling form,
1 R
PN (R) = d f (a ≪ R ≪ N a) . (6.310)
RF RF
11 There 1
are logarithmic corrections to the SAW result exactly at d = 4, but for all d > 4 one has ν = 2
.
6.7. POLYMERS 287
One finds (
xg x≪1
f (x) ∼ δ
(6.311)
exp(−x ) x ≫ 1 ,
with g = (γ − 1)/ν and δ = 1/(1 − ν).
Consider a solution of monodisperse polymers of length N in a solvent. Let φ be the dimensionless monomer
concentration, so φ/N is the dimensionless polymer concentration and φs = 1 − φ is the dimensionless solvent
concentration. (Dimensionless concentrations are obtained by dividing the corresponding dimensionful concen-
tration by the overall density.) The entropy of mixing for such a system is given by eqn. 2.352. We have
Vk 1
Smix = − B · φ ln φ + (1 − φ) ln(1 − φ) , (6.312)
v0 N
where v0 ∝ a3 is the volume per monomer. Accounting for an interaction between the monomer and the solvent,
we have that the free energy of mixing is
v0 Fmix 1
= φ ln φ + (1 − φ) ln(1 − φ) + χ φ(1 − φ) . (6.313)
V kB T N
where χ is the dimensionless polymer-solvent interaction, called the Flory parameter. This provides a mean field
theory of the polymer-solvent system.
The osmotic pressure Π is defined by
∂Fmix
Π=− , (6.314)
∂V N
p
which is the variation of the free energy of mixing with respect to volume holding the number of polymers constant.
The monomer concentration is φ = N Np v0 /V , so
∂ φ2 ∂
=− . (6.315)
∂V Np N Np v0 ∂φ Np
Now we have
1
Fmix = N Np kB T ln φ + (φ−1 − 1) ln(1 − φ) + χ (1 − φ) , (6.316)
N
and therefore
kB T h −1 i
Π= (N − 1) φ − ln(1 − φ) − χ φ2 . (6.317)
v0
Note that Π > 0 only if χ < 12 , which is the condition for a ’good solvent’.
In fact, eqn. 6.319 is only qualitatively correct. In the limit where χ ≪ 21 , Flory showed that the individual polymer
coils behave much as hard spheres of radius RF . The osmotic pressure then satisfies something analogous to a
virial equation of state:
2
Π φ φ 3
= +A RF + ...
kB T N v0 N v0
(6.320)
φ
= h(φ/φ∗ ) .
N v0
This is generalized to a scaling form in the second line, where h(x) is a scaling function, and φ∗ = N v0 /RF3
∝
−4/5 3 ∗
N , assuming d = 3 and ν = 5 from Flory theory. As x = φ/φ → 0, we must recover the ideal gas law, so
h(x) = 1+O(x) in this limit. For x → ∞, we require that the result be independent of the degree of polymerization
N . This means h(x) ∝ xp with 45 p = 1, i.e. p = 45 . The result is known as the des Cloiseaux law:
v0 Π
= C φ9/4 , (6.321)
kB T
where C is a constant. This is valid for what is known as semi-dilute solutions, where φ∗ ≪ φ ≪ 1. In the dense
limit φ ∼ 1, the results do not exhibit this universality, and we must appeal to liquid state theory, which is no fun
at all.
6.8.1 Definition
Here, the spin variables σi take values in the set {1, 2, . . . , q} on each site. The equivalent of an external magnetic
field in the Ising case is a field h which prefers a particular value of σ (σ = 1 in the above Hamiltonian). Once
again, it is not possible to compute the partition function on general lattices, however in one dimension we may
once again find Z using the transfer matrix method.
Z = Tr e−βH
X βhδ βJδ βhδ βJδ
= e σ1 ,1 e σ1 ,σ2 · · · e σN ,1 e σN ,σ1
(6.323)
{σn }
= Tr RN ,
6.8. APPENDIX I : POTTS MODEL IN ONE DIMENSION 289
In matrix form,
eβ(J+h) eβh/2 eβh/2 ··· eβh/2
eβh/2 eβJ 1 ··· 1
βh/2
e 1 eβJ ··· 1
R= . .. .. .. .. (6.325)
.. . . . .
eβh/2 1 1 · · · eβJ 1
βh/2
e 1 1 ··· 1 eβJ
The matrix R has q eigenvalues λj , with j = 1, . . . , q. The partition function for the Potts chain is then
q
X
Z= λN
j . (6.326)
j=1
We can actually find the eigenvalues of R analytically. To this end, consider the vectors
1 eβh/2
0 −1/2
1
φ = . , ψ = q − 1 + eβh . . (6.327)
.. ..
0 1
λ1 = eβJ + q − 1 (6.330)
βJ
λj = e −1 , j ∈ {2, . . . , q} , (6.331)
since | ψ i is an eigenvector with eigenvalue λ = eβJ + q − 1, and any vector orthogonal to | ψ i has eigenvalue
λ = eβJ − 1. The partition function is then
N N
Z = eβJ + q − 1 + (q − 1) eβJ − 1 . (6.332)
When h is nonzero, the calculation becomes somewhat more tedious, but still relatively easy. The problem is that
| ψ i and | φ i are not orthogonal, so we define
| φ i − | ψ ih ψ | φ i
|χi = q , (6.334)
2
1 − hφ|ψi
where 1/2
eβh
x ≡ hφ|ψi = . (6.335)
q − 1 + eβh
Now we have h χ | ψ i = 0, with h χ | χ i = 1 and h ψ | ψ i = 1, with
p
| φ i = 1 − x2 | χ i + x | ψ i . (6.336)
M = a0 I + a · τ
a0 + a3 a1 − ia2 (6.339)
= ,
a1 + ia2 a0 − a3
For the transfer matrix of eqn. 6.337, we obtain, after a little work,
h i
λ1,2 = eβJ − 1 + 21 q − 1 + eβh + eβJ − 1 eβh − 1 (6.342)
rh
1
i2
±2 q − 1 + eβh + eβJ − 1 eβh − 1 − 4(q − 1) eβJ − 1 eβh − 1 .
6.9. APPENDIX II : ONE-PARTICLE IRREDUCIBLE CLUSTERS AND THE VIRIAL EXPANSION 291
There are q − 2 other eigenvalues, however, associated with the (q − 2)-dimensional subspace orthogonal to | χ i
and | ψ i. Clearly all these eigenvalues are given by
λj = eβJ − 1 , j ∈ {3 , . . . , q} . (6.343)
and in the thermodynamic limit N → ∞ the maximum eigenvalue λ1 dominates. Note that we recover the correct
limit as h → 0.
It is convenient to work with dimensionless quantities, using λdT as the unit of volume. To this end, define
nγ −1
ν ≡ nλdT , π ≡ pλdT , cγ (T ) ≡ bγ (T ) λdT , (6.347)
so that
X ∞
X
βπ = cγ z nγ = c̃l z l
γ l=1
∞ (6.348)
X X
nγ l
ν= n γ cγ z = l c̃l z ,
γ l=1
where X
c̃l = cγ δnγ ,l (6.349)
γ
is the sum over all connected clusters with l vertices. Here and henceforth, the functional dependence on T is
implicit; π and ν are regarded here as explicit functions of z. We can, in principle, invert to obtain z(ν). Let us
write this inverse as
X ∞
z(ν) = ν exp αk ν k . (6.350)
k=1
292 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Ultimately we need to obtain expressions for the coefficients αk , but let us first assume the above form and use it
to write π in terms of ν. We have
∞
X Zz ∞
X l−1
βπ = c̃l z = dz ′
l
l c̃l z ′
l=1 0 l=1
Zν Zν Zν ∞
X
dz ′ ν ′ d ln z ′ ′k (6.351)
= dν ′ ′ ′ = dν ′ = dν ′
1 − k αk ν
dν z d ln ν ′
0 0 0 k=1
∞
X ∞
k αk k+1 X e k
=ν− ν = Bk ν ,
k+1
k=1 k=1
The clusters which contribute to c̃l are all connected, by definition. However, it is useful to make a further dis-
tinction based on the topology of connected clusters and define a connected cluster γ to be irreducible if, upon
removing any site in γ and all the links connected to that site, the remaining sites of the cluster are still connected.
The situation is depicted in Fig. 6.24. For a reducible cluster γ, the integral cγ is proportional to a product of
cluster integrals over its irreducible components. Let us define the set Γl as the set of all irreducible clusters of l
vertices, and X
ζl = cγ . (6.353)
γ∈Γl
We may also obtain the cluster integrals c̃l in terms of the αk . To this end, note that l2 c̃l is the coefficient of z l in
the function z dν
dz , hence
I I
2 dz 1 dν dν −l
l c̃l = l
z = z
2πiz z dz 2πi
I ∞ I ∞
dν 1 Y lαk ν k dν 1 X Y (l αk )mk kmk
= e = ν
2πi ν l 2πi ν l mk ! (6.354)
k=1 {mk } k=1
X ∞
Y (l αk )mk
= δPk km , l−1 .
k mk !
{mk } k=1
6.9. APPENDIX II : ONE-PARTICLE IRREDUCIBLE CLUSTERS AND THE VIRIAL EXPANSION 293
Figure 6.24: Connected versus irreducible clusters. Clusters (a) through (d) are irreducible in that they remain
connected if any component site and its connecting links are removed. Cluster (e) is connected, but is reducible. Its
integral cγ may be reduced to a product over its irreducible components, each shown in a unique color.
294 CHAPTER 6. CLASSICAL INTERACTING SYSTEMS
Chapter 7
7.1 References
– M. Kardar, Statistical Physics of Particles (Cambridge, 2007)
A superb modern text, with many insightful presentations of key concepts.
– M. Plischke and B. Bergersen, Equilibrium Statistical Physics (3rd edition, World Scientific, 2006)
An excellent graduate level text. Less insightful than Kardar but still a good modern treatment of the subject.
Good discussion of mean field theory.
– G. Parisi, Statistical Field Theory (Addison-Wesley, 1988)
An advanced text focusing on field theoretic approaches, covering mean field and Landau-Ginzburg theories
before moving on to renormalization group and beyond.
– J. P. Sethna, Entropy, Order Parameters, and Complexity (Oxford, 2006)
An excellent introductory text with a very modern set of topics and exercises.
295
296 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
RT a
p= − . (7.2)
v − b v2
Let us fix the temperature T and examine the function p(v). Clearly p(v) is a decreasing function of volume for
v just above the minimum allowed value v = b, as well as for v → ∞. But is p(v) a monotonic function for all
v ∈ [b, ∞]?
We can answer this by computing the derivative,
∂p 2a RT
= − . (7.3)
∂v T v3 (v − b)2
2a u3
= , (7.4)
bRT (u − 1)2
where u ≡ v/b is dimensionless. It is easy to see that the function f (u) = u3 /(u − 1)2 has a unique minimum
for u > 1. Setting f ′ (u∗ ) = 0 yields u∗ = 3, and so fmin = f (3) = 27
4 . Thus, for T > Tc = 8a/27bR, the LHS of
eqn. 7.4 lies below the minimum value of the RHS, and there is no solution. This means that p(v, T > Tc ) is a
monotonically decreasing function of v.
At T = Tc there is a saddle-node bifurcation. Setting vc = bu∗ = 3b and evaluating pc = p(vc , Tc ), we have that the
location of the critical point for the van der Waals system is
a 8a
pc = , vc = 3b , Tc = . (7.5)
27 b2 27 bR
For T < Tc , there are two solutions to eqn. 7.4, corresponding to a local minimum and a local maximum of the
function p(v). The locus of points in the (v, p) plane for which (∂p/∂v)T = 0 is obtained by setting eqn. 7.3 to zero
and solving for T , then substituting this into eqn. 7.2. The result is
a 2ab
p∗ (v) = − 2 . (7.6)
v3 v
Expressed in terms of dimensionless quantities p̄ = p/pc and v̄ = v/vc , this equation becomes
3 2
p̄∗ (v̄) = 3
− 2 . (7.7)
v̄ v̄
Figure 7.1: Pressure versus molar volume for the van der Waals gas at temperatures in equal intervals from T =
1.10 Tc (red) to T = 0.85 Tc (blue). The purple curve is p̄∗ (v̄).
Along the curve p = p∗ (v), the isothermal compressibility, κT = − v1 ∂v
∂p T diverges, heralding a thermodynamic
instability. To understand better, let us compute the free energy of the van der Waals system, F = E − T S.
Regarding the energy E, we showed back in chapter 2 that
∂ε ∂p a
=T −p= 2 , (7.8)
∂v T ∂T V v
which entails
a
ε(T, v) = 12 f RT −
, (7.9)
v
where ε = E/ν is the molar internal energy. The first term is the molar energy of an ideal gas, where f is
the number of molecular freedoms, which is the appropriate low density limit. The molar specific heat is then
∂ε
cV = ∂T v
= f2 R, which means that the molar entropy is
Z T
cV
s(T, v) = dT ′ = f2 R ln(T /Tc) + s1 (v) . (7.10)
T′
∂f
We then write f = ε − T s, and we fix the function s1 (v) by demanding that p = − ∂v T . This yields s1 (v) =
R ln(v − b) + s0 , where s0 is a constant. Thus2 ,
a
f (T, v) = f2 R T 1 − ln T /Tc − − RT ln(v − b) − T s0 . (7.11)
v
We know that under equilibrium conditions, f is driven to a minimum by spontaneous processes. Now suppose
∂ 2f
that ∂v 2
T
< 0 over some range of v at a given temperature T . This would mean that one mole of the system at
2 Don’t confuse the molar free energy (f ) with the number of molecular degrees of freedom (f )!
298 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
L2 ·bar L
gas a mol2
b mol pc (bar) Tc (K) vc (L/mol)
Acetone 14.09 0.0994 52.82 505.1 0.2982
Argon 1.363 0.03219 48.72 150.9 0.0966
Carbon dioxide 3.640 0.04267 7404 304.0 0.1280
Ethanol 12.18 0.08407 63.83 516.3 0.2522
Freon 10.78 0.0998 40.09 384.9 0.2994
Helium 0.03457 0.0237 2.279 5.198 0.0711
Hydrogen 0.2476 0.02661 12.95 33.16 0.0798
Mercury 8.200 0.01696 1055 1723 0.0509
Methane 2.283 0.04278 46.20 190.2 0.1283
Nitrogen 1.408 0.03913 34.06 128.2 0.1174
Oxygen 1.378 0.03183 50.37 154.3 0.0955
Water 5.536 0.03049 220.6 647.0 0.0915
Table 7.1: Van der Waals parameters for some common gases. (Source: Wikipedia)
Figure 7.2: Molar free energy f (T, v) of the van der Waals system T = 0.85 Tc, with dot-dashed black line showing
Maxwell construction connecting molar volumes v1,2 on opposite sides of the coexistence curve.
volume v and temperature T could lower its energy by rearranging into two half-moles, with respective molar
volumes v ± δv, each at temperature T . The total volume and temperature thus remain fixed, but the free energy
∂ 2f
changes by an amount ∆f = 21 ∂v 2
T
(δv)2 < 0. This means that the system is unstable – it can lower its energy by
dividing up into two subsystems each with different densities (i.e. molar volumes). Note that the onset of stability
occurs when
∂ 2f ∂p 1
= − = =0, (7.12)
∂v 2 T ∂v T vκp
which is to say when κp = ∞. As we saw, this occurs at p = p∗ (v), given in eqn. 7.6.
∂ 2f
However, this condition, ∂v 2
T
< 0, is in fact too strong. That is, the system can be unstable even at molar volumes
∂ 2f
where ∂v 2
T
> 0. The reason is shown graphically in fig. 7.2. At the fixed temperature T , for any molar volume v
between vliquid ≡ v1 and vgas ≡ v2 , the system can lower its free energy by phase separating into regions of different
7.2. THE VAN DER WAALS SYSTEM 299
f = (1 − x) f1 + x f2 , (7.14)
where fj = f (vj , T ). This function is given by the straight black line connecting the points at volumes v1 and v2
in fig. 7.2.
The two equations which give us v1 and v2 are
∂f ∂f
= (7.15)
∂v v ∂v v
1 ,T 2 ,T
and
∂f
f (T, v1 ) − f (T, v2 ) = (v2 − v1 ) . (7.16)
∂v v
1 ,T
In terms of the pressure, p = − ∂f
∂v T , these equations are equivalent to
This procedure is known as the Maxwell construction, and is depicted graphically in Fig. 7.3. When the Maxwell
construction is enforced, the isotherms resemble the curves in Fig. 7.4. In this figure, all points within the purple
∂ 2f
shaded region have ∂v 2 < 0, hence this region is unstable to infinitesimal fluctuations. The boundary of this
region is called the spinodal, and the spontaneous phase separation into two phases is a process known as spinodal
decomposition. The dot-dashed orange curve, called the coexistence curve, marks the instability boundary for nucle-
ation. In a nucleation process, an energy barrier must be overcome in order to achieve the lower free energy state.
There is no energy barrier for spinodal decomposition – it is a spontaneous process.
7.2.2 Analytic form of the coexistence curve near the critical point
where
∂p ∂2p ∂3p ∂2p
pv ≡ , pvv ≡ , pvvv ≡ , pvT ≡ , etc. (7.20)
∂v ∂v 2 ∂v 3 ∂v ∂T
The second equation we write as
ZwG
dw p(vc + w, T ) = 21 (wG − wL ) p(vc + wL , T ) + p(vc + wG , T ) . (7.21)
wL
300 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Figure 7.3: Maxwell construction in the (v, p) plane. The system is absolutely unstable between volumes vd and
ve . For v ∈ [va , vd ] of v ∈ [ve , vc ], the solution is unstable with respect to phase separation. Source: Wikipedia.
1
Subtracting the LHS from the RHS, we find that we can then divide by 6 wG2 − wL2 , resulting in Cleaning this up,
we obtain
0 = pvv (vc , T ) + 1
2 pvvv (vc , T ) (wG + wL ) + 1
20
3
pvvvv (vc , T ) 3wG2 + 4wG wL + 3wL2 + O wG,L . (7.23)
We now define w± ≡ wG ± wL . In terms of these variables, eqns. 7.19 and 7.23 become
1 1 2
2 31
0 = pv (vc , T ) + 2 pvv (vc , T ) w+ + 8 pvvv (vc , T ) w+
w− + O w±
+ 3 (7.24)
2 2
3
0 = pvv (vc , T ) + 21 pvvv (vc , T ) w+ + 18 pvvvv (vc , T ) w+ + 51 w− + O w± . (7.25)
We now evaluate w± to order T . Note that pv (vc , Tc ) = pvv (vc , Tc ) = 0, since the critical point is an inflection point
2
v (vc , T ) = pvT Θ + O(Θ ), where T = Tc + Θ and pvT = pvT (vc , Tc ). We can
in the (v, p) plane. Thus, we have p√
then see that w− is of leading order −Θ , while w+ is of leading order Θ. This allows us to write
2
0 = pvT Θ + 1
24 pvvv w− + O(Θ2 ) (7.26)
2 2
0 = pvvT Θ + 12 pvvv w+ + 1
40 pvvvv w− + O(Θ ) . (7.27)
7.2. THE VAN DER WAALS SYSTEM 301
Figure 7.4: Pressure-volume isotherms for the van der Waals system, as in Fig. 7.1, but corrected to account for
the Maxwell construction. The boundary of the purple shaded region is the spinodal line p̄∗ (v̄). The boundary of
the orange shaded region is the stability boundary with respect to phase separation.
Thus,
1/2
24 pvT √
w− = −Θ + . . . (7.28)
pvvv
6 pvT pvvvv − 10 pvvv pvvT
w+ = Θ + ... . (7.29)
5 p2vvv
We then have
1/2
6 pvT √ 3 pvT pvvvv − 5 pvvv pvvT
wL = − −Θ + 2
Θ + O Θ3/2 (7.30)
pvvv 5 pvvv
1/2
6 pvT √ 3 pvT pvvvv − 5 pvvv pvvT
wG = −Θ + 2
Θ + O Θ3/2 . (7.31)
pvvv 5 pvvv
Suppose we follow along an isotherm starting from the high molar volume (gas) phase. If T > Tc , the volume
v decreases continuously as the pressure p increases. If T < Tc , then at the instant the isotherm first intersects
the orange boundary curve in Fig. 7.4, there is a discontinuous change in the molar volume from high (gas) to
low (liquid). This discontinuous change is the hallmark of a first order phase transition. Note that the volume
discontinuity, ∆v = w− ∝ (Tc − T )1/2 . This is an example of a critical behavior in which the order parameter
φ, which
in this case may be taken to be the difference φ = vG − vL , behaves as a power law in T − Tc, where Tc is the
critical temperature. In this case, we have φ(T ) ∝ (Tc − T )β+ , where β = 21 is the exponent, and where (Tc − T )+ is
defined to be Tc − T if T < Tc and 0 otherwise. The isothermal compressibility is κT = −v/pv (v, T ). This is finite
302 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
along the coexistence curve – it diverges only along the spinodal. It therefore diverges at the critical point, which
lies at the intersection of the spinodal and the coexistence curve.
It is convenient to express the equation of state and the coexistence curve in terms of dimensionless variables.
Write
p v T
p̄ = , v̄ = , T̄ = . (7.32)
pc vc Tc
The van der Waals equation of state then becomes
8T̄ 3
p̄ = − . (7.33)
3v̄ − 1 v̄ 2
Further expressing these dimensionless quantities in terms of distance from the critical point, we write
p̄ = 1 + π , v̄ = 1 + ǫ , T̄ = 1 + t . (7.34)
Thus,
8(1 + t) 3
π(ǫ, t) = − −1. (7.35)
2 + 3ǫ (1 + ǫ)2
Note that the LHS and the RHS of this equation vanish identically for (π, ǫ, t) = (0, 0, 0). We can then write
1/2
6 πǫt 3 πǫt πǫǫǫǫ − 5 πǫǫǫ πǫǫt
ǫL,G = ∓ (−t)1/2 + 2
t + O (−t)3/2 . (7.36)
πǫǫǫ 5 πǫǫǫ
The van der Waals equation of state first appears in van der Waals’ 1873 PhD thesis3 , “Over de Continuı̈teit van den
Gas - en Vloeistoftoestand” (“On the continuity of the gas and liquid state”). In his Nobel lecture4 , van der Waals
writes of how he was inspired by Rudolf Clausius’ 1857 treatise on the nature of heat, where it is posited that a gas
in fact consists of microscopic particles whizzing around at high velocities. van der Waals reasoned that liquids,
which result when gases are compressed, also consist of ’small moving particles’: ”Thus I conceived the idea that
there is no essential difference between the gaseous and the liquid state of matter. . . ”
Clausius’ treatise showed how his kinetic theory of heat was consistent with Boyle’s law for gases (pV = constant
at fixed temperature). van der Waals pondered why this might fail for the non-dilute liquid phase, and he reasoned
that there were two principal differences: inter-particle attraction and excluded volume. These considerations
prompted him to posit his famous equation,
RT a
p= − . (7.37)
v − b v2
The first term on the RHS accounts for excluded volume effects, and the second for mutual attractions.
In the limiting case of p → ∞, the molar volume approaches v = b. On physical grounds it seems that one should
take b = NA · 8v0 , where v0 is the molecular volume. The factor of eight arises from the hard core intermolecular
potential effective for separations less than the molecular diameter (i.e. twice the molecular radius). Expanding in
inverse powers of v, though,
RT a RT
p= + b− · 2 + O v −3 , (7.38)
v RT v
3 Johannes Diderik van der Waals, the eldest of ten children, was the son of a carpenter. As a child he received only a primary school
education. He worked for a living until age 25, and was able to enroll in a three-year industrial evening school for working class youth.
Afterward he continued his studies independently, in his spare time, working as a teacher. By the time he obtained his PhD, he was 36 years
old. He received the Nobel Prize for Physics in 1910.
4 See http://www.nobelprize.org/nobel prizes/physics/laureates/1910/waals-lecture.pdf
7.2. THE VAN DER WAALS SYSTEM 303
Figure 7.5: ‘Universality’ of the liquid-gas transition for eight different atomic and molecular fluids, from E. A.
Guggenheim, J. Chem. Phys. 13, 253 (1945). Dimensionless pressure p/pc versus dimensionless density ρ/ρc =
vc /v is shown. The van der Waals / mean field theory gives ∆v = vgas − vliquid ∝ (−t)1/2 , while experiments show
a result closer to ∆v ∝ (−t)1/3 . Here t ≡ T̄ − 1 = (T − Tc )/Tc is the dimensionless temperature deviation with
respect to the critical point.
a
and we read of the second virial coefficient B2 = b − RT /NA . But we have computed from the Mayer expansion
that B2 = 4v0 , so the naı̈ve guess for b is twice too large.
Another of van der Waals’ great achievements was his articulation of the law of corresponding states. Recall that
the van der Waals equation of state, when written in terms of dimensionless quantities p̄ = p/pc , v̄ = v/vc , and
T̄ = T /Tc, takes the form of eqn. 7.33. Thus, while the a and b parameters are specific to each fluid – see Tab.
7.1 – when written in terms of these scaled dimensionless variables, the equation of state and all its consequent
properties (i.e. the liquid-gas phase transition) are universal.
The van der Waals equation is best viewed as semi-phenomenological. Interaction and excluded volume effects
surely are present, but the van der Waals equation itself only captures them in a very approximate way. It is
applicable to gases, where it successfully predicts features that are not present in ideal systems (e.g. throttling). It is
of only qualitative and pedagogical use in the study of fluids, the essential physics of which lies in the behavior of
quantities like the pair distribution function g(r). As we saw in chapter 6, any adequate first principles derivation
of g(r) - a function which can be measured in scattering experiments - involves rather complicated approximation
304 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
schemes to close the BBGKY hierarchy. Else one must resort to numerical simulations such as the Monte Carlo
method. Nevertheless, the lessons learned from the van der Waals system are invaluable and they provide us
with a first glimpse of what is going on in the vicinity of a phase transition, and how nonanalytic behavior, such
as vG − vL ∝ (Tc − T )β with noninteger exponent β may result due to singularities in the free energy at the critical
point.
7.3. FLUIDS, MAGNETS, AND THE ISING MODEL 305
Figure 7.6: The lattice gas model. An occupied cell corresponds to n = 1 (σ = +1), and a vacant cell to n = 0
(σ = −1).
The usual description of a fluid follows from a continuum Hamiltonian of the form
XN X
p2i
Ĥ(p, x) = + u(xi − xj ) . (7.39)
i=1
2m i<j
The potential u(r) is typically central, depending only on the magnitude |r|, and short-ranged. Now consider a
discretized version of the fluid, in which we divide up space into cells (cubes, say), each of which can accommo-
date at most one fluid particle (due to excluded volume effects). That is, each cube has a volume on the order of
a3 , where a is the diameter of the fluid particles. In a given cube i we set the occupancy ni = 1 if a fluid particle is
present and ni = 0 if there is no fluid particle present. We then have that the potential energy is
X X
U= u(xi − xj ) = 12 VRR′ nR nR′ , (7.40)
i<j R6=R′
where VRR′ ≈ v(R − R′ ), where Rk is the position at the center of cube k. The grand partition function is then
approximated as
X Y X
nR 1
Ξ(T, V, µ) ≈ ξ exp − 2 β VRR′ nR nR′ , (7.41)
{nR } R R6=R′
where
ξ = eβµ λ−d d
T a , (7.42)
where a is the side length of each cube (chosen to be
Pon the order of the hard sphere diameter). The λ−d
T factor
arises from the integration over the momenta. Note R nR = N is the total number of fluid particles, so
Y
ξ nR = ξ N = eβµN λ−N
T
d Nd
a . (7.43)
R
306 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
where σR ≡ 2nR − 1 is a spin variable taking the possible values {−1, +1}, and
JRR′ = − 14 VRR′
X′ (7.45)
H = 12 kB T ln ξ − 1
4 VRR′ ,
R′
where the prime on the sum indicates that R′ = R is to be excluded. For the Lennard-Jones system, VRR′ =
v(R − R′ ) < 0 is due to the attractive tail of the potential, hence JRR′ is positive, which prefers alignment of the
spins σR and σR′ . This interaction is therefore ferromagnetic. The spin Hamiltonian in eqn. 7.44 is known as the
Ising model.
The physics of the liquid-gas transition in fact has a great deal in common with that of the transition between a
magnetized and unmagnetized state of a magnetic system. The correspondences are5
p ←→ H , v ←→ m ,
where m is the magnetization density, defined here to be the total magnetization M divided by the number of
lattice sites Ns :6
M 1 X
m= = hσR i . (7.46)
Ns Ns
R
Sketches of the phase diagrams are reproduced in fig. 7.7. Of particular interest is the critical point, which occurs
at (Tc , pc ) in the fluid system and (Tc , Hc ) in the magnetic system, with Hc = 0 by symmetry.
In the fluid, the coexistence curve in the (p, T ) plane separates high density (liquid) and low density (vapor)
phases. The specific volume v (or the density n = v −1 ) jumps discontinuously across the coexistence curve. In the
magnet, the coexistence curve in the (H, T ) plane separates positive magnetization and negative magnetization
phases. The magnetization density m jumps discontinuously across the coexistence curve. For T > Tc , the latter
system is a paramagnet, in which the magnetization varies smoothly as a function of H. This behavior is most
apparent in the bottom panel of the figure, where v(p) and m(H) curves are shown.
For T < Tc , the fluid exists in a two phase region, which is spatially inhomogeneous, supporting local regions of
high and low density. There is no stable homogeneous thermodynamic phase for (T, v) within the two phase
region shown in the middle left panel. Similarly, for the magnet, there is no stable homogeneous thermodynamic
phase at fixed temperature T and magnetization m if (T, m) lies within the coexistence region. Rather, the system
consists of blobs where the spin is predominantly up, and blobs where the spin is predominantly down.
5 One could equally well identify the second correspondence as n ←→ m between density (rather than specific volume) and magnetization.
One might object that H is more properly analogous to µ. However, since µ = µ(p, T ) it can equally be regarded as analogous to p. Note also
that βp = zλ−d d
T for the ideal gas, in which case ξ = z(a/λT ) is proportional to p.
6 Note the distinction between the number of lattice sites N and the number of occupied cells N . According to our definitions, N =
s
1
2
(M + Ns ).
7.3. FLUIDS, MAGNETS, AND THE ISING MODEL 307
Figure 7.7: Comparison of the liquid-gas phase diagram with that of the Ising ferromagnet.
Note also the analogy between the isothermal compressibility κT and the isothermal susceptibility χT :
1 ∂v
κT = − , κT (Tc , pc ) = ∞
v ∂p T
∂m
χT = , χT (Tc , Hc ) = ∞
∂H T
308 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
The ‘order parameter’ for a second order phase transition is a quantity which vanishes in the disordered phase
and is finite in the ordered phase. For the fluid, the order parameter can be chosen to be Ψ ∝ (vvap − vliq ), the
difference in the specific volumes of the vapor and liquid phases. In the vicinity of the critical point, the system
exhibits power law behavior in many physical quantities, viz.
m(T, Hc ) ∼ Tc − T )β+
χ(T, Hc ) ∼ |T − Tc |−γ
(7.47)
CM (T, Hc ) ∼ |T − Tc |−α
m(Tc , H) ∼ ±|H|1/δ .
The quantities α, β, γ, and δ are the critical exponents associated with the transition. These exponents satisfy certain
equalities, such as the Rushbrooke and Griffiths relations and hyperscaling,7
α + 2β + γ = 2 (Rushbrooke)
β + γ = βδ (Griffiths) (7.48)
2 − α = dν (hyperscaling) .
Originally such relations were derived as inequalities, and only after the advent of scaling and renormalization
group theories it was realized that they held as equalities. We shall have much more to say about critical behavior
later on, when we discuss scaling and renormalization.
Homogeneity of E(S, M, Ns ) means E = T S + HM + µNs , and, after invoking the First Law dE = T dS + H dM +
µ dNs , we have
S dT + M dH + Ns dµ = 0 . (7.49)
Now consider two magnetic phases in coexistence. We must have dµ1 = dµ2 , hence
where m = M/Ns is the magnetization per site and s = S/Ns is the specific entropy. Thus, we obtain the Clapeyron
equation for magnetic systems,
dH s − s2
=− 1 . (7.51)
dT coex m1 − m2
Thus, if m1 6= m2 and dH dT coex = 0, then we must have s1 = s2 , which says that there is no latent heat associated with
the transition. This absence of latent heat is a consequence of the symmetry which guarantees that F (T, H, Ns ) =
F (T, −H, Ns ).
Another application of the Ising model lies in the theory of order-disorder transitions in alloys. Examples include
Cu3 Au, CuZn, and other compounds. In CuZn, the Cu and Zn atoms occupy sites of a body centered cubic (BCC)
7 In the third of the following exponent equalities, d is the dimension of space and ν is the correlation length exponent.
7.4. MEAN FIELD THEORY 309
lattice, forming an alloy known as β-brass. Below Tc ≃ 740 K, the atoms are ordered, with the Cu preferentially
occupying one simple cubic sublattice and the Zn preferentially occupying the other.
The energy is a sum of pairwise interactions, with a given link contributing εAA , εBB , or εAB , depending on whether
it is an A-A, B-B, or A-B/B-A link. Here A and B represent Cu and Zn, respectively. Thus, we can write the energy
of the link hiji as
Eij = εAA PiA PjA + εBB PiB PjB + εAB PiA PjB + PiB PjA , (7.52)
where
(
A 1 1 if site i contains Cu
Pi = 2 (1 + σi ) =
0 if site i contains Zn
(
B 1 1 if site i contains Zn
Pi = 2 (1 − σi ) =
0 if site i contains Cu .
where the exchange constant J and the magnetic field H are given by
1
J= 4 2εAB − εAA − εBB
(7.54)
1
H= 4 εBB − εAA ,
and E0 = 18 N z εAA + εBB + 2εAB , where N is the total number of lattice sites and z = 8 is the lattice coordination
number, which is the number of nearest neighbors of any given site.
Note that
√ 7.8:√Order-disorder transition on the square lattice. Below T = Tc , order develops spontaneously on the
Figure
two 2 × 2 sublattices. There is perfect sublattice order at T = 0 (left panel).
where the first sum on the RHS is over all links of the lattice. Each spin can be either ‘up’ (σ = +1) or ‘down’ (σ =
−1). We further assume that the spins are located on a Bravais lattice8 and that the coupling Jij = J |Ri − Rj | ,
where Ri is the position of the ith spin.
On each site i we decompose σi into a contribution from its thermodynamic average and a fluctuation term, i.e.
We will write hσi i ≡ m, the local magnetization (dimensionless), and assume that m is independent of position i.
Then
σi σj = (m + δσi ) (m + δσj )
= m2 + m (δσi + δσj ) + δσi δσj (7.58)
2
= −m + m (σi + σj ) + δσi δσj .
The last term on the RHS of the second equation above is quadratic in the fluctuations, and we assume this to be
negligibly small. Thus, we obtain the mean field Hamiltonian
X
ĤMF = 12 N zJ m2 − H + zJm σi , (7.59)
i
where N is the total number of lattice sites. The first term is a constant, although the value of m is yet to be
determined. The Boltzmann weights are then completely determined by the second term, which is just what we
would write down for a Hamiltonian of noninteracting spins in an effective ‘mean field’
In other words, Heff = Hext + Hint , where the external field is applied field Hext = H, and the ‘internal field’ is
Hint = zJm. The internal field accounts for the interaction with the average values of all other spins coupled to a
spin at a given site, hence it is often called the ‘mean field’. Since the spins are noninteracting, we have
eβHeff − e−βHeff H + zJm
m = βH = tanh . (7.61)
e eff + e−βHeff kB T
8 A Bravais lattice is one in which any site is equivalent to any other site through an appropriate discrete translation. Examples of Bravais
lattices include the linear chain, square, triangular, simple cubic, face-centered cubic, etc. lattices. The honeycomb lattice is not a Bravais lattice,
because there are two sets of inequivalent sites – those in the center of a Y and those in the center of an upside down Y.
7.4. MEAN FIELD THEORY 311
It is a simple matter to solve for the free energy, given the noninteracting Hamiltonian ĤMF . The partition function
is X N
1 2
Z = Tr e−β ĤMF = e− 2 βN zJ m eβ(H+zJm)σ = e−βF . (7.62)
σ
We now define dimensionless variables:
F kB T H
f≡ , θ≡ , h≡ , (7.63)
N zJ zJ zJ
and obtain the dimensionless free energy
f (m, h, θ) = 12 m2 − θ ln e(m+h)/θ + e−(m+h)/θ . (7.64)
7.4.1 h = 0
Figure 7.9: Results for h = 0. Upper panels: graphical solution to self-consistency equation m = tanh(m/θ) at
temperatures θ = 0.65 (blue) and θ = 1.5 (dark red). Lower panel: mean field free energy, with energy shifted by
θ ln 2 so that f (m = 0, θ) = 0.
Thus, the specific heat is discontinuous at θ = θc . We emphasize that this is only valid near θ = θc = 1. The general
result valid for all θ is9
1 m2 (θ) − m4 (θ)
cV (θ) = · , (7.72)
θ θ − 1 + m2 (θ)
9 To
` ´
obtain this result, one writes f = f θ, m(θ) and then differentiates twice with respect to θ, using the chain rule. Along the way, any
∂f
naked (i.e. undifferentiated) term proportional to ∂m may be dropped, since this vanishes at any θ by the mean field equation.
7.4. MEAN FIELD THEORY 313
7.10: Results for h = 0.1. Upper panels: graphical solution to self-consistency equation m = tanh (m +
Figure
h)/θ at temperatures θ = 0.65 (blue), θ = 0.9 (dark green), and θ = 1.5 (dark red). Lower panel: mean field free
energy, with energy shifted by θ ln 2 so that f (m = 0, θ) = 0.
With this expression one can check both limits θ → 0 and θ → θc . As θ → 0 the magnetization saturates and one
has m2 (θ) ≃ 1 − 4 e−2/θ . The numerator then vanishes as e−2/θ , which overwhelms the denominator that itself
vanishes as θ2 . As a result, cV (θ → 0) = 0, as expected. As θ → 1, invoking m2 ≃ 3(1 − θ) we recover cV (θc− ) = 23 .
In the theory of critical phenomena, cV (θ) ∝ |θ − θc |−α as θ → θc . We see that mean field theory yields α = 0.
7.4.3 h 6= 0
Consider without loss of generality the case h > 0. The minimum of the free energy f (m, h, θ) now lies at m > 0
for any θ. At low temperatures, the double well structure we found in the h = 0 case is tilted so that the right
well lies lower in energy than the left well. This is depicted in fig. 7.10. As the temperature is raised, the local
minimum at m < 0 vanishes, annihilating with the local maximum in a saddle-node bifurcation. To find where
∂f ∂ 2f
this happens, one sets ∂m = 0 and ∂m 2 = 0 simultaneously, resulting in
√
∗
√ θ 1+ 1−θ
h (θ) = 1 − θ − ln √ . (7.73)
2 1− 1−θ
The solutions lie at h = ±h∗ (θ). For θ < θc = 1 and h ∈ −h∗ (θ) , +h∗ (θ) , there are three solutions to the mean
field equation. Equivalently we could in principle invert the above expression to obtain θ∗ (h). For θ > θ∗ (h), there
314 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
is only a single global minimum in the free energy f (m) and there is no local minimum. Note θ∗ (h = 0) = 1.
Assuming h ≪ |θ − 1| ≪ 1, the mean field solution for m(θ, h) will also be small, and we expand the free energy
in m, and to linear order in h:
m4 hm
f (m, h, θ) = −θ ln 2 + 21 (1 − θ−1 ) m2 + −
12 θ3 θ (7.74)
1 2 1 4
= f0 + 2 (θ − 1) m + 12 m − hm + . . . .
∂f
Setting ∂m = 0, we obtain
1 3
3m + (θ − 1) · m − h = 0 . (7.75)
3
If θ > 1 then we have a solution m = h/(θ − 1). The m term can be ignored because it is higher order in h, and
we have assumed h ≪ |θ − 1| ≪ 1. This is known as the Curie-Weiss law10 . The magnetic susceptibility behaves as
∂m 1
χ(θ) =
= ∝ |θ − 1|−γ , (7.76)
∂h θ−1
where the magnetization critical exponent γ is γ = 1. If θ < 1 then while there is still a solution at m = h/(θ − 1),
it lies at a local maximum of the free energy,
√ as shown in fig. 7.10. The minimum of the free energy occurs close
to the h = 0 solution m = m0 (θ) ≡ 3 (1 − θ), and writing m = m0 + δm we find δm to linear order in h as
δm(θ, h) = h/2(1 − θ). Thus,
√ h
m(θ, h) = 3 (1 − θ) + . (7.77)
2(1 − θ)
Once again, we find that χ(θ) diverges as |θ − 1|−γ with γ = 1. The exponent γ on either side of the transition is
the same.
Finally, we can set θ = θc and examine m(h). We find, from eqn. 7.76,
m(θ = θc , h) = (3h)1/3 ∝ h1/δ , (7.78)
where δ is a new critical exponent. Mean field theory gives δ = 3. Note that at θ = θc = 1 we have m = tanh(m+h),
and inverting we find
1 1+m m3 m5
h(m, θ = θc ) = 2 ln −m= + + ... , (7.79)
1−m 3 5
which is consistent with what we just found for m(h, θ = θc ).
How well does mean field theory do in describing the phase transition of the Ising model? In table 7.2 we compare
our mean field results for the exponents α, β, γ, and δ with exact values for the two-dimensional Ising model,
numerical work on the three-dimensional Ising model, and experiments on the liquid-gas transition in CO2 . The
first thing to note is that the exponents are dependent on the dimension of space, and this is something that mean
field theory completely misses. In fact, it turns out that the mean field exponents are exact provided d > du , where
du is the upper critical dimension of the theory. For the Ising model, du = 4, and above four dimensions (which is of
course unphysical) the mean field exponents are in fact exact. We see that all in all the MFT results compare better
with the three dimensional exponent values than with the two-dimensional ones – this makes sense since MFT
does better in higher dimensions. The reason for this is that higher dimensions means more nearest neighbors,
which has the effect of reducing the relative importance of the fluctuations we neglected to include.
10 Pierre Curie was a pioneer in the fields of crystallography, magnetism, and radiation physics. In 1880, Pierre and his older brother Jacques
discovered piezoelectricity. He was 21 years old at the time. It was in 1895 that Pierre made the first systematic studies of the effects of
temperature on magnetic materials, and he formulated what is known as Curie’s Law, χ = C/T , where C is a constant. Curie married Marie
Sklodowska in the same year. Their research turned toward radiation, recently discovered by Becquerel and Röntgen. In 1898, Pierre and
Marie Curie discovered radium. They shared the 1903 Nobel Prize in Physics with Becquerel. Marie went on to win the 1911 Nobel Prize in
Chemistry and was the first person ever awarded two Nobel Prizes. Their daughter Irène Joliot Curie shared the 1935 Prize in Chemistry (with
her husband), also for work on radioactivity. Pierre Curie met an untimely and unfortunate end in the Spring of 1906. Walking across the
Place Dauphine, he slipped and fell under a heavy horse-drawn wagon carrying military uniforms. His skull was crushed by one of the wagon
wheels, killing him instantly. Later on that year, Pierre-Ernest Weiss proposed a modification of Curie’s Law to account for ferromagnetism.
This became known as the Curie-Weiss law, χ = C/(T − Tc ).
7.4. MEAN FIELD THEORY 315
Table 7.2: Critical exponents from mean field theory as compared with exact results for the two-dimensional Ising
model, numerical results for the three-dimensional Ising model, and experiments on the liquid-gas transition in
CO2 . Source: H. E. Stanley, Phase Transitions and Critical Phenomena.
Dissipative processes drive physical systems to minimum energy states. We can crudely model the dissipative
dynamics of a magnet by writing the phenomenological equation
dm ∂f
=− , (7.80)
ds ∂m
where s is a dimensionless time variable. Under these dynamics, the free energy is never increasing:
2
df ∂f ∂m ∂f
= =− ≤0. (7.81)
ds ∂m ∂s ∂m
∂f
Clearly the fixed point of these dynamics, where ṁ = 0, is a solution to the mean field equation ∂m = 0.
The phase flow for the equation ṁ = −f ′ (m) is shown in fig. 7.11. As we have seen, for any value of h there is
a temperature θ∗ below which the free energy f (m) has two local minima and one local maximum. When h = 0
the minima are degenerate, but at finite h one of the minima is a global minimum. Thus, for θ < θ∗ (h) there
are three solutions to the mean field equations. In the language of dynamical systems, under the dynamics of
eqn. 7.80, minima of f (m) correspond to attractive fixed points and maxima to repulsive fixed points. If h > 0,
the rightmost of these fixed points corresponds to the global minimum of the free energy. As θ is increased,
this fixed point evolves smoothly. At θ = θ∗ , the (metastable) local minimum and the local maximum coalesce
and annihilate in a saddle-note bifurcation. However at h = 0 all three fixed points coalesce at θ = θc and the
bifurcation is a supercritical pitchfork. As a function of t at finite h, the dynamics are said to exhibit an imperfect
bifurcation, which is a deformed supercritical pitchfork.
The solution set for the mean field equation is simply expressed by inverting the tanh function to obtain h(θ, m).
One readily finds
θ 1+m
h(θ, m) = ln −m. (7.82)
2 1−m
As we see in the bottom panel of fig. 7.12, m(h) becomes multivalued for h ∈ − h∗ (θ) , +h∗ (θ) , where h∗ (θ)
is given in eqn. 7.73. Now imagine that θ < θc and we slowly ramp the field h from a large negative value to
a large positive value, and then slowly back down to its original value. On the time scale of the magnetization
dynamics, we can regard h(s) as a constant. (Remember the time variable is s here.) Thus, m(s) will flow to the
nearest stable fixed point. Initially the system starts with m = −1 and h large and negative, and there is only
one fixed point, at m∗ ≈ −1. As h slowly increases, the fixed point value m∗ also slowly increases. As h exceeds
−h∗ (θ), a saddle-node bifurcation occurs, and two new fixed points are created at positive m, one stable and one
unstable. The global minimum of the free energy still lies at the fixed point with m∗ < 0. However, when h
crosses h = 0, the global minimum of the free energy lies at the most positive fixed point m∗ . The dynamics,
316 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Figure 7.11: Dissipative magnetization dynamics ṁ = −f ′ (m). Bottom panel shows h∗ (θ) from eqn. 7.73. For
(θ, h) within the blue shaded region, the free energy f (m) has a global minimum plus a local minimum and a
local maximum. Otherwise f (m) has only a single global minimum. Top panels show an imperfect bifurcation
in the magnetization dynamics at h = 0.0215 , for which θ∗ = 0.90. Temperatures shown: θ = 0.65 (blue),
θ = θ∗ (h) = 0.90 (green), and θ = 1.2. The rightmost stable fixed point corresponds to the global minimum of the
free energy. The bottom of the middle two upper panels shows h = 0, where both of the attractive fixed points
and the repulsive fixed point coalesce into a single attractive fixed point (supercritical pitchfork bifurcation).
however, keep the system stuck in what is a metastable phase. This persists until h = +h∗ (θ), at which point
another saddle-note bifurcation occurs, and the attractive fixed point at m∗ < 0 annihilates with the repulsive
fixed point. The dynamics then act quickly to drive m to the only remaining fixed point. This process is depicted
in the top panel of fig. 7.12. As one can see from the figure, the the system follows a stable fixed point until the
fixed point disappears, even though that fixed point may not always correspond to a global minimum of the free
energy. The resulting m(h) curve is then not reversible as a function of time, and it possesses a characteristic shape
known as a hysteresis loop. Etymologically, the word hysteresis derives from the Greek υστ ερησις, which means
‘lagging behind’. Systems which are hysteretic exhibit a history-dependence to their status, which is not uniquely
determined by external conditions. Hysteresis may be exhibited with respect to changes in applied magnetic field,
changes in temperature, or changes in other externally determined parameters.
7.4. MEAN FIELD THEORY 317
Figure 7.12: Top panel : hysteresis as a function of ramping the dimensionless magnetic field h at θ = 0.40. Dark
red arrows below the curve follow evolution of the magnetization on slow increase of h. Dark grey arrows above
the curve follow evolution of the magnetization on slow decrease of h. Bottom panel : solution set for m(θ, h) as
a function of h at temperatures θ = 0.40 (blue), θ = θc = 1.0 (dark green), and t = 1.25 (red).
where Jij is the coupling between spins on sites i and j. In the top equation above, each pair (ij) is counted once
in the interaction term; this may be replaced by a sum over all i and j if we include a factor of 12 .11 The resulting
mean field Hamiltonian is then
X
ˆ m2 − H + J(0)
ĤMF = 21 N J(0) ˆ m σi . (7.84)
i
11 The self-interaction terms with i = j contribute a constant to Ĥ and may be either included or excluded. However, this property only
pertains to the σi = ±1 model. For higher spin versions of the Ising model, say where Si ∈ {−1, 0, +1}, then Si2 is not constant and we
should explicitly exclude the self-interaction terms.
318 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
ˆ
Here, J(q) is the Fourier transform of the interaction matrix Jij :12
X
ˆ
J(q) = J(R) e−iq·R . (7.85)
R
Consider an Ising model where Jij = J/N for all i and j, so that there is a very weak interaction between every
pair of spins. The Hamiltonian is then
2
J X X
Ĥ = − σi − H σk . (7.86)
2N i k
Thus,
" X 2 # 1/2 Z∞
βJ N βJ 1 2 P
exp σi = dm e− 2 N βJm +βJm i σi , (7.89)
2N i
2π
−∞
Since N → ∞, we can perform the integral using the method of steepest descents. Thus, we must set
∗
dA ∗ m +h
= 0 =⇒ m = tanh . (7.92)
dm m∗ θ
12 The sum in the discrete Fourier transform is over all ‘direct Bravais lattice vectors’ and the wavevector q may be restricted to the ‘first
Brillouin zone’. These terms are familiar from elementary solid state physics.
7.5. VARIATIONAL DENSITY MATRIX METHOD 319
F θ
f= = A(m∗ ) + ln A′′ (m∗ ) + O(N −2 ) , (7.95)
NJ 2N
where m∗ is the solution to the mean field equation which minimizes A(m). Mean field theory is exact for this
model!
Suppose we are given a Hamiltonian Ĥ. From this we construct the free energy, F :
F = E − TS
(7.96)
= Tr (̺ Ĥ) + kB T Tr (̺ ln ̺) .
Here, ̺ is the density matrix13 . A physical density matrix must be (i) normalized (i.e. Tr ̺ = 1), (ii) Hermitian, and
(iii) non-negative definite (i.e. all the eigenvalues of ̺ must be non-negative).
Our goal is to extremize the free energy subject to the various constraints on ̺. Let us assume that ̺ is diagonal in
the basis of eigenstates of Ĥ, i.e.
X
̺= Pγ γ γ , (7.97)
γ
where Pγ is the probability that the system is in state γ . Then
X X
F = Eγ Pγ + kB T Pγ ln Pγ . (7.98)
γ γ
Thus, the free energy is a function of the set {Pγ }. We now extremize F subject to the normalization constraint.
This means we form the extended function
X
F ∗ {Pγ }, λ = F {Pγ } + λ Pγ − 1 , (7.99)
γ
13 How do we take the logarithm of a matrix? The rule is this: A = ln B if B = exp(A). The exponential of a matrix may be evaluated via
its Taylor expansion.
320 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
and then freely extremize over both the probabilities {Pγ } as well as the Lagrange multiplier λ. This yields the
Boltzmann distribution,
1
Pγeq = exp(−Eγ /kB T ) , (7.100)
Z
P
where Z = γ e−Eγ /kB T = Tr e−Ĥ/kB T is the canonical partition function, which is related to λ through
λ = kB T (ln Z − 1) . (7.101)
If the spectrum of Ĥ is bounded from below, our extremum should in fact yield a minimum for the free energy
F . Furthermore, since we have freely minimized over all the probabilities, subject to the single normalization
constraint, any distribution {Pγ } other than the equilibrium one must yield a greater value of F .
Alas, the Boltzmann distribution, while exact, is often intractable to evaluate. For one-dimensional systems, there
are general methods such as the transfer matrix approach which do permit an exact evaluation of the free energy.
However, beyond one dimension the situation is in general hopeless. A family of solvable (“integrable”) mod-
els exists in two dimensions, but their solutions require specialized techniques and are extremely difficult. The
idea behind the variational density matrix approximation is to construct a tractable trial density matrix ̺ which
depends on a set of variational parameters {xα }, and to minimize with respect to this (finite) set.
where Y
δγ,γ ′ = δσi ,σi′ . (7.104)
i
Indeed, this is the case for the exact density matrix, which is to say the Boltzmann weight,
1 −β Ĥ(σ1 ,...,σN )
̺N (σ1 , σ2 , . . .) = e . (7.105)
Z
We now write a trial density matrix which is a product over contributions from independent single sites:
Y
̺N (σ1 , σ2 , . . .) = ̺(σi ) , (7.106)
i
where 1 + m 1 − m
̺(σ) = δσ,1 + δσ,−1 . (7.107)
2 2
Note that we’ve changed our notation slightly. We are denoting by ̺(σ) the corresponding diagonal element of
the matrix 1+m
2 0
̺= 1−m , (7.108)
0 2
7.5. VARIATIONAL DENSITY MATRIX METHOD 321
and the full density matrix is a tensor product over the single site matrices:
̺N = ̺ ⊗ ̺ ⊗ · · · ⊗ ̺ . (7.109)
Note that ̺ and hence ̺N are appropriately normalized. The variational parameter here is m, which, if ρ is to be
non-negative definite, must satisfy −1 ≤ m ≤ 1. The quantity m has the physical interpretation of the average
spin on any given site, since X
hσi i = ̺(σ) σ = m. (7.110)
σ
ˆ
We now define the dimensionless free energy per site: f ≡ F/N J(0). We have
1 − m 1 − m
1+m 1+m
f (m, h, θ) = − 12 m2 − hm + θ ln + ln , (7.113)
2 2 2 2
ˆ
where θ ≡ kB T /J(0) ˆ
is the dimensionless temperature, and h ≡ H/J(0) the dimensionless magnetic field, as
before. We extremize f (m) by setting
∂f θ 1 + m
= 0 = −m − h + ln . (7.114)
∂m 2 1−m
Solving for m, we obtain
m+h
m = tanh , (7.115)
θ
which is precisely what we found in eqn. 7.65.
Note that the optimal value of m indeed satisfies the requirement |m| ≤ 1 of non-negative probability. This
nonlinear equation may be solved graphically. For h = 0, the unmagnetized solution p m = 0 always applies.
However, for θ < 1 there are two additional solutions at m = ±mA (θ), with mA (θ) = 3(1 − θ) + O (1 − θ)3/2
for t close to (but less than) one. These solutions, which are related by the Z2 symmetry of the h = 0 model, are
in fact the low energy solutions. This is shown clearly in figure 7.13, where the variational free energy f (m, t) is
plotted as a function of m for a range of temperatures interpolating between ‘high’ and ‘low’ values. At the critical
temperature θc = 1, the lowest energy state changes from being unmagnetized (high temperature) to magnetized
(low temperature).
For h > 0, there is no longer a Z2 symmetry (i.e. σi → −σi ∀ i). The high temperature solution now has m > 0
(or m < 0 if h < 0), and this smoothly varies as t is lowered, approaching the completely polarized limit m = 1
as θ → 0. At very high temperatures, the argument of the tanh function is small, and we may approximate
tanh(x) ≃ x, in which case
h
m(h, θ) = . (7.116)
θ − θc
322 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Figure 7.13: Variational field free energy ∆f = f (m, h, θ) + θ ln 2 versus magnetization m at six equally spaced
temperatures interpolating between ‘high’ (θ = 1.25, red) and ‘low’ (θ = 0.75, blue) values. Top panel: h = 0.
Bottom panel: h = 0.06.
This is called the Curie-Weiss law. One can infer θc from the high temperature susceptibility χ(θ) = (∂m/∂h)h=0
by plotting χ−1 versus θ and extrapolating to obtain the θ-intercept. In our case, χ(θ) = (θ − θc )−1 . For low θ and
weak h, there are two inequivalent minima in the free energy.
When m is small, it is appropriate to expand f (m, h, θ), obtaining
f (m, h, θ) = −θ ln 2 − hm + 1
2 (θ − 1) m2 + θ
12 m4 + θ
30 m6 + θ
56 m8 + . . . . (7.117)
This is known as the Landau expansion of the free energy in terms of the order parameter m. An order parameter is a
thermodynamic variable φ which distinguishes ordered and disordered phases. Typically φ = 0 in the disordered
(high temperature) phase, and φ 6= 0 in the ordered (low temperature) phase. When the order sets in continuously,
i.e. when φ is continuous across θc , the phase transition is said to be second order. When φ changes abruptly, the
transition is first order. It is also quite commonplace to observe phase transitions between two ordered states. For
example, a crystal, which is an ordered state, may change its lattice structure, say from a high temperature tetrag-
onal phase to a low temperature orthorhombic phase. When the high T phase possesses the same symmetries
as the low T phase, as in the tetragonal-to-orthorhombic example, the transition may be second order. When the
two symmetries are completely unrelated, for example in a hexagonal-to-tetragonal transition, or in a transition
between a ferromagnet and an antiferromagnet, the transition is in general first order.
Throughout this discussion, we have assumed that the interactions Jij are predominantly ferromagnetic, i.e. Jij > 0,
7.5. VARIATIONAL DENSITY MATRIX METHOD 323
so that all the spins prefer to align. When Jij < 0, the interaction is said to be antiferromagnetic and prefers anti-
alignment of the spins (i.e. σi σj = −1). Clearly not every pair of spins can be anti-aligned – there are two possible
spin states and a thermodynamically extensive number of spins. But on the square lattice, for example, if the only
interactions Jij are between nearest neighbors and the interactions are antiferromagnetic, then the lowest energy
configuration (T = 0 ground state) will be one in which spins on opposite sublattices are anti-aligned. The square
lattice is bipartite – it breaks up into two interpenetrating sublattices A and B (which are themselves
√ square lattices,
rotated by 45◦ with respect to the original, and with a larger lattice constant by a factor of 2), such that any site in
A has nearest neighbors in B, and vice versa. The honeycomb lattice is another example of a bipartite lattice. So is
the simple cubic lattice. The triangular lattice, however, is not bipartite (it is tripartite). Consequently, with nearest
neighbor antiferromagnetic interactions, the triangular lattice Ising model is highly frustrated. The moral of the
story is this: antiferromagnetic interactions can give rise to complicated magnetic ordering, and, when frustrated
by the lattice geometry, may have finite specific entropy even at T = 0.
Here, σi ∈ {1, . . . , q}, with integer q. This is the so-called ‘q-state Potts model’. The quantity H is analogous to an
external magnetic field, and preferentially aligns (for H > 0) the local spins in the σ = 1 direction. We will assume
H ≥ 0.
The q-component set is conveniently taken to be the integers from 1 to q, but it could be anything, such as
The interaction energy is −Jij if sites i and j contain the same object (q possibilities), and 0 if i and j contain
different objects (q 2 − q possibilities).
The two-state Potts model is equivalent to the Ising model. Let the allowed values of σ be ±1. Then the quantity
δσ,σ′ = 1
2 + 1
2 σσ ′ (7.120)
equals 1 if σ = σ ′ , and is zero otherwise. The three-state Potts model cannot be written as a simple three-state
Ising model, i.e. one with a bilinear interaction σ σ ′ where σ ∈ {−1, 0, +1}. However, it is straightforward to verify
the identity
δσ,σ′ = 1 + 21 σσ ′ + 23 σ 2 σ ′2 − (σ 2 + σ ′2 ) . (7.121)
Thus, the q = 3-state Potts model is equivalent to a S = 1 (three-state) Ising model which includes both bilinear
(σσ ′ ) and biquadratic (σ 2 σ ′2 ) interactions, as well as a local field term which couples to the square of the spin,
σ 2 . In general one can find such correspondences for higher q Potts models, but, as should be expected, the
interactions become increasingly complex, with bi-cubic, bi-quartic, bi-quintic, etc. terms.
Getting back to the mean field theory, we write the single site variational density matrix ̺ as a diagonal matrix
with entries
1−x
̺(σ) = x δσ,1 + 1 − δσ,1 , (7.122)
q−1
with ̺N (σ1 , . . . , σN ) = ̺(σ1 ) · · · ̺(σN ). Note that Tr (̺) = 1. The variational parameter is x. When x = q −1 , all
states are equally probable. But for x > q −1 , the state σ = 1 is preferred, and the other (q − 1) states have identical
324 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
but smaller probabilities. It is a simple matter to compute the energy and entropy:
1 ˆ 2 (1 − x)2
E = Tr (̺N Ĥ) = − 2 N J(0) x + − N Hx
q−1
(7.123)
1−x
S = −kB Tr (̺N ln ̺N ) = −N kB x ln x + (1 − x) ln .
q−1
The dimensionless free energy per site is then
(1 − x)2 1−x
f (x, θ, h) = − 12 x2 + + θ x ln x + (1 − x) ln − hx , (7.124)
q−1 q−1
ˆ
where h = H/J(0). We now extremize with respect to x to obtain the mean field equation,
∂f 1−x 1−x
= 0 = −x + + θ ln x − θ ln −h. (7.125)
∂x q−1 q−1
Note that for h = 0, x = q −1 is a solution, corresponding to a disordered state in which all states are equally
probable. At high temperatures, for small h, we expect x − q −1 ∝ h. Indeed, using Mathematica one can set
x ≡ q −1 + s , (7.126)
q (qθ − 1) q 3 (q − 2) θ 2
h= s+ s + O(s3 ) . (7.127)
q−1 2 (q − 1)2
For weak fields, |h| ≪ 1, and we have
(q − 1) h
s(θ) = + O(h2 ) , (7.128)
q (qθ − 1)
which again is of the Curie-Weiss form. The difference s = x − q −1 is the order parameter for the transition.
Finally, one can expand the free energy in powers of s, obtaining the Landau expansion,
2h + 1 q (qθ − 1) 2 (q − 2) q 3 θ 3
f (s, θ, h) = − − θ ln q − hs + s − s
2q 2 (q − 1) 6 (q − 1)2
q3 θ h i q4 θ h i
(7.129)
+ 1 + (q − 1)−3 s4 − 1 − (q − 1)−4 s5
12 20
q5 θ h i
+ 1 + (q − 1)−5 s6 + . . .
30
Note that, for q = 2, the coefficients of s3 , s5 , and higher order odd powers of s vanish in the Landau expansion.
This is consistent with what we found for the Ising model, and is related to the Z2 symmetry of that model. For
q > 3, there is a cubic term in the mean field free energy, and thus we generically expect a first order transition, as
we shall see below when we discuss Landau theory.
Consider the so-called XY model, in which each site contains a continuous planar spin, represented by an angular
variable φi ∈ [−π, π]: X X
Ĥ = − Jij cos φi − φj − H cos φi . (7.130)
i<j i
7.5. VARIATIONAL DENSITY MATRIX METHOD 325
We write the (diagonal elements of the) full density matrix once again as a product:
Y
̺N (φ1 , φ2 , . . .) = ̺(φi ) . (7.131)
i
Our goal will be to extremize the free energy with respect to the function ̺(φ). To this end, we compute
2
ˆ Tr ̺ eiφ − N H Tr ̺ cos φ .
E = Tr (̺N Ĥ) = − 12 N J(0) (7.132)
The entropy is
S = −N kB Tr (̺ ln ̺) . (7.133)
14
Note that for any function A(φ), we have
Zπ
dφ
Tr ̺ A) ≡ ̺(φ) A(φ) . (7.134)
2π
−π
We now extremize the functional F ̺(φ) = E − T S with respect to ̺(φ), under the condition that Tr ̺ = 1. We
therefore use Lagrange’s method of undetermined multipliers, writing
F̃ = F − N kB T λ Tr ̺ − 1 . (7.135)
Note that F̃ is a function of the Lagrange multiplier λ and a functional of the density matrix ̺(φ). The prefactor
N kB T which multiplies λ is of no mathematical consequence – we could always redefine the multiplier to be λ′ ≡
N kB T λ. It is present only to maintain homogeneity and proper dimensionality of F ∗ with λ itself dimensionless
and of order N 0 . We now have
(
δ F̃ δ 2
= ˆ Tr ̺ eiφ − N H Tr ̺ cos φ
− 12 N J(0)
δ̺(φ) δ̺(φ)
)
+ N kB T Tr ̺ ln ̺ − N kB T λ Tr ̺ − 1 .
Thus, we have
" #
δ F̃ 1 iφ′
−iφ −iφ′
iφ cos φ
1 ˆ
= − 2 N J(0) · Tr′ ̺ e e + Tr′ ̺ e e − NH ·
δ̺(φ) 2π φ φ 2π
(7.137)
1 h i λ
+ N kB T · ln ̺(φ) + 1 − N kB T · .
2π 2π
Now let us define
Zπ
iφ
dφ
Tr ̺ e = ̺(φ) eiφ ≡ m eiφ0 . (7.138)
φ 2π
−π
14 The denominator of 2π in the measure is not necessary, and in fact it is even slightly cumbersome. It divides out whenever we take a ratio
to compute a thermodynamic average. I introduce this factor to preserve the relation Tr 1 = 1. I personally find unnormalized traces to be
profoundly unsettling on purely aesthetic grounds.
326 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
We then have
ˆ
J(0) H
ln ̺(φ) = m cos(φ − φ0 ) + cos φ + λ − 1. (7.139)
kB T kB T
Clearly the free energy will be reduced if φ0 = 0 so that the mean field is maximal and aligns with the external
field, which prefers φ = 0. Thus, we conclude
Heff
̺(φ) = C exp cos φ , (7.140)
kB T
where
ˆ m+H
Heff = J(0) (7.141)
and C = eλ−1 . The value of λ is then determined by invoking the constraint,
Zπ
dφ Heff
Tr ̺ = 1 = C exp cos φ = C I0 (Heff /kB T ) , (7.142)
2π kB T
−π
eε cos φ eε cos φ
̺(φ) = Rπ = . (7.144)
dφ′ ε cos φ′ I0 (ε)
2π e
−π
to obtain
f (ε, h, θ) = 1
4 θ− 1
2 ε2 + 1
64 2 − 3θ ε4 − 1
2 hε + 1
16 hε3 + . . . . (7.150)
This predicts a second order phase transition at θc = 12 .15 Note also the Curie-Weiss form of the susceptibility at
high θ:
∂f h
= 0 =⇒ ε = + ... . (7.151)
∂ε θ − θc
Landau’s theory of phase transitions is based on an expansion of the free energy of a thermodynamic system in
terms of an order parameter, which is nonzero in an ordered phase and zero in a disordered phase. For example,
the magnetization M of a ferromagnet in zero external field but at finite temperature typically vanishes for tem-
peratures T > Tc , where Tc is the critical temperature, also called the Curie temperature in a ferromagnet. A low
order expansion in powers of the order parameter is appropriate sufficiently close to the phase transition, i.e. at
temperatures such that the order parameter, if nonzero, is still small.
The simplest example is the quartic free energy,
where f0 = f0 (θ), a = a(θ), and b = b(θ). Here, θ is a dimensionless measure of the temperature. If for example
the local exchange energy in the ferromagnet is J, then we might define θ = kB T /zJ, as before. Let us assume
b > 0, which is necessary if the free energy is to be bounded from below16 . The equation of state ,
∂f
= 0 = am + bm3 , (7.153)
∂m
p p
has three solutions in the complex m plane: (i) m = 0, (ii) m = −a/b , and (iii) m = − −a/b . The latter two
solutions lie along the (physical) real axis if a < 0. We assume that there exists a unique temperature θc where
a(θc ) = 0. Minimizing f , we find
a2
θ < θc : f (θ) = f0 −
4b (7.154)
θ > θc : f (θ) = f0 .
The free energy is continuous at θc since a(θc ) = 0. The specific heat, however, is discontinuous across the transi-
tion, with
2 2
+
−
∂ 2 a θc a′ (θc )
c θc − c θc = −θc 2 =− . (7.155)
∂θ θ=θc 4b 2b(θc )
The presence of a magnetic field h breaks the Z2 symmetry of m → −m. The free energy becomes
Landau expansion.
328 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Figure 7.14: Phase diagram for the quartic mean field theory f = f0 + 12 am2 + 41 bm4 − hm, with b > 0. There is
a first order line at h = 0 extending from a = −∞ and terminating in a critical point at a = 0. For |h| < h∗ (a)
(dashed red line) there are three solutions to the mean field equation, corresponding to one global minimum, one
local minimum, and one local maximum. Insets show behavior of the free energy f (m).
This is a cubic equation for m with real coefficients, and as such it can either have three real solutions or one real
solution and two complex solutions related by complex conjugation. Clearly we must have a < 0 in order to have
three real roots, since bm3 + am is monotonically increasing otherwise. The boundary between these two classes
of solution sets occurs when two roots coincide, which means f ′′ (m) = 0 as well as f ′ (m) = 0. Simultaneously
solving these two equations, we find
2 (−a)3/2
h∗ (a) = ± , (7.158)
33/2 b1/2
or, equivalently,
3
a∗ (h) = − b1/3 |h|2/3 . (7.159)
22/3
If, for fixed h, we have a < a∗ (h), then there will be three real solutions to the mean field equation f ′ (m) = 0, one
of which is a global minimum (the one for which m · h > 0). For a > a∗ (h) there is only a single global minimum,
at which m also has the same sign as h. If we solve the mean field equation perturbatively in h/a, we find
h b
m(a, h) = − 4 h3 + O(h5 ) (a > 0)
a a
(7.160)
|a|1/2 h 3 b1/2 2
=± 1/2
+ ± h + O(h3 ) (a < 0) .
b 2 |a| 8 |a|5/2
7.6. LANDAU THEORY OF PHASE TRANSITIONS 329
with b > 0 for stability. Without loss of generality, we may assume y > 0 (else send m → −m). Note that we no
longer have m → −m (i.e. Z2 ) symmetry. The cubic term favors positive m. What is the phase diagram in the (a, y)
plane?
Extremizing the free energy with respect to m, we obtain
∂f
= 0 = am − ym2 + bm3 . (7.162)
∂m
This cubic equation factorizes into a linear and quadratic piece, and hence may be solved simply. The three
solutions are m = 0 and r
y y 2 a
m = m± ≡ ± − . (7.163)
2b 2b b
We now see that for y 2 < 4ab there is only one real solution, at m = 0, while for y 2 > 4ab there are three real
solutions. Which solution has lowest free energy? To find out, we compare the energy f (0) with f (m+ )17 . Thus,
we set
f (m) = f (0) =⇒ 21 am2 − 13 ym3 + 41 bm4 = 0 , (7.164)
0 = a − ym + bm2
(7.165)
0 = 21 a − 13 ym + 41 bm2 = 0 .
Eliminating the quadratic term gives m = 3a/y. Finally, substituting m = m+ gives us a relation between a, b, and
y:
y 2 = 29 ab . (7.166)
y2
a> : 1 real root m = 0
4b
y2 2y 2
>a> : 3 real roots; minimum at m = 0 (7.167)
4b 9b r
2
2y y y 2 a
>a : 3 real roots; minimum at m = + −
9b 2b 2b b
The solution m = 0 lies at a local minimum of the free energy for a > 0 and at a local maximum for a < 0. Over
2 2
the range y4b > a > 2y9b , then, there is a global minimum at m = 0, a local minimum at m = m+ , and a local
2
maximum at m = m− , with m+ > m− > 0. For 2y 9b > a > 0, there is a local minimum at a = 0, a global minimum
at m = m+ , and a local maximum at m = m− , again with m+ > m− > 0. For a < 0, there is a local maximum at
m = 0, a local minimum at m = m− , and a global minimum at m = m+ , with m+ > 0 > m− . See fig. 7.15.
17 We needn’t waste our time considering the m = m− solution, since the cubic term prefers positive m.
330 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Figure 7.15: Behavior of the quartic free energy f (m) = 21 am2 − 31 ym3 + 14 bm4 . A: y 2 < 4ab ; B: 4ab < y 2 < 92 ab
; C and D: y 2 > 29 ab. The thick black line denotes a line of first order transitions, where the order parameter is
discontinuous across the transition.
Suppose we now impose some dynamics on the system, of the simple relaxational type
∂m ∂f
= −Γ , (7.168)
∂t ∂m
where Γ is a phenomenological kinetic coefficient. Assuming y > 0 and b > 0, it is convenient to adimensionalize
by writing
y y2 b
m≡ ·u , a≡ ·r , t≡ ·s. (7.169)
b b Γ y2
Then we obtain
∂u ∂ϕ
=− , (7.170)
∂s ∂u
where the dimensionless free energy function is
ϕ(u) = 21 ru2 − 31 u3 + 14 u4 . (7.171)
We see that there is a single control parameter, r. The fixed points of the dynamics are then the stationary points
of ϕ(u), where ϕ′ (u) = 0, with
ϕ′ (u) = u (r − u + u2 ) . (7.172)
The solutions to ϕ′ (u) = 0 are then given by
q
u∗ = 0 , u∗ = 1
2 ± 1
4 −r. (7.173)
For r > 41 there is one fixed point at u = 0, which is attractive under the dynamics u̇ = −ϕ′ (u) since ϕ′′ (0) = r. At
r = 41 there occurs a saddle-node bifurcation and a pair of fixed points is generated, one stable and one unstable.
7.6. LANDAU THEORY OF PHASE TRANSITIONS 331
Figure 7.16: Fixed points for ϕ(u) = 21 ru2 − 13 u3 + 14 u4 and flow under the dynamics u̇ = −ϕ′ (u). Solid curves
represent stable fixed points and dashed curves unstable fixed points. Magenta arrows show behavior under
slowly increasing control parameter r and dark blue arrows show behavior under slowly decreasing r. For u > 0
there is a hysteresis loop. The thick black curve shows the equilibrium thermodynamic value of u(r), i.e. that value
which minimizes the free energy ϕ(u). There is a first order phase transition at r = 92 , where the thermodynamic
value of u jumps from u = 0 to u = 23 .
As we see from fig. 7.14, the interior fixed point is always unstable and the two exterior fixed points are always
stable. At r = 0 there is a transcritical bifurcation where two fixed points of opposite stability collide and bounce
off one another (metaphorically speaking).
At the saddle-node bifurcation, r = 41 and u = 12 , and we find ϕ(u = 21 ; r = 14 ) = 192
1
, which is positive. Thus,
the thermodynamic state of the system remains at u = 0 until the value of ϕ(u+ ) crosses zero. This occurs when
ϕ(u) = 0 and ϕ′ (u) = 0, the simultaneous solution of which yields r = 92 and u = 23 .
Suppose we slowly ramp the control parameter r up and down as a function of the dimensionless time s. Under
the dynamics of eqn. 7.170, u(s) flows to the first stable fixed point encountered – this is always the case for
a dynamical system with a one-dimensional phase space. Then as r is further varied, u follows the position of
whatever locally stable fixed point it initially encountered. Thus, u r(s) evolves smoothly until a bifurcation is
encountered. The situation is depicted by the arrows in fig. 7.16. The equilibrium thermodynamic value for u(r) is
discontinuous; there is a first order phase transition at r = 92 , as we’ve already seen. As r is increased, u(r) follows
a trajectory indicated by the magenta arrows. For an negative initial value of u, the evolution as a function of r
will be reversible. However, if u(0) is initially positive, then the system exhibits hysteresis, as shown. Starting with
a large positive value of r, u(s) quickly evolves to u = 0+ , which means a positive infinitesimal value. Then as r
is decreased, the system remains at u = 0+ even through the first order transition, because u = 0 is an attractive
fixed point. However, once r begins to go q negative, the u = 0 fixed point becomes repulsive, and u(s) quickly
flows to the stable fixed point u+ = 21 + 14 − r. Further decreasing r, the system remains on this branch. If r is
later increased, then u(s) remains on the upper branch past r = 0, until the u+ fixed point annihilates with the
332 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
q
1 1
unstable fixed point at u− = 2 − 4 − r, at which time u(s) quickly flows down to u = 0+ again.
Finally, consider a model with Z2 symmetry, with the Landau free energy
with c > 0 for stability. We seek the phase diagram in the (a, b) plane. Extremizing f with respect to m, we obtain
∂f
= 0 = m (a + bm2 + cm4 ) , (7.175)
∂m
which is a quintic with five solutions over the complex m plane. One solution is obviously m = 0. The other four
are v
u s
u b b
2
a
t
m=± − ± − . (7.176)
2c 2c c
For each ± symbol in the above equation, there are two options, hence four roots in all.
If a > 0 and b > 0, then four of the roots are imaginary and there is a unique minimum at m = 0.
For a < 0, there are only three solutions to f ′ (m) = 0 for real m, since the − choice for the ± sign under the radical
leads to imaginary roots. One of the solutions is m = 0. The other two are
s r
b b 2 a
m=± − + − . (7.177)
2c 2c c
√
The most interesting situation is a > 0 and b < 0. If a > 0 and b < −2 ac, all five roots are real. There must be
three minima, separated by two local maxima. Clearly if m∗ is a solution, then so is −m∗ . Thus, the only question
is whether the outer minima are of lower energy than the minimum at m = 0. We assess this by demanding
f (m∗ ) = f (0), where m∗ is the position of the largest root (i.e. the rightmost minimum). This gives a second
quadratic equation,
0 = 12 a + 41 bm2 + 61 cm4 , (7.178)
which together with equation 7.175 gives √
b = − √43 ac . (7.179)
Thus, we have the following, for fixed a > 0:
√
b > −2 ac : 1 real root m = 0
√ √
−2 ac > b > − √43 ac : 5 real roots; minimum at m = 0 (7.180)
s r
4
√ b b 2 a
− 3 ac > b : 5 real roots; minima at m = ± − +
√ −
2c 2c c
The point (a, b) = (0, 0), which lies at the confluence of a first order line and a second order line, is known as a
tricritical point.
7.6. LANDAU THEORY OF PHASE TRANSITIONS 333
Figure 7.17: Behavior of the sextic free energy f (m) =√21 am2 + 41 bm4 + 61 cm6 . √
A: a > 0 and b√> 0 ; B: a < 0 and
b > 0 ; C: a < 0 and b < 0 ; D: a > 0 and b < − √43 ac ; E: a > 0 and − √43 ac < b < −2 ac ; F: a > 0 and
√
−2 ac < b < 0. The thick dashed line is a line of second order transitions, which meets the thick solid line of first
order transitions at the tricritical point, (a, b) = (0, 0).
Figure 7.18: Free energy ϕ(u) = 21 ru2 − 41 u4 + 61 u6 for several different values of the control parameter r.
New and interesting behavior occurs for b > 0. The fixed points of the dynamics are obtained by setting ϕ′ (u) = 0.
We have
ϕ(u) = 12 ru2 − 41 u4 + 61 u6
(7.184)
ϕ′ (u) = u (r − u2 + u4 ) .
Thus, the equation ϕ′ (u) = 0 factorizes into a linear factor u and a quartic factor u4 − u2 + r which is quadratic in
u2 . Thus, we can easily obtain the roots:
r q
∗ ∗ 1 1
r<0 : u =0, u =± 2 + 4 −r
r q r q
1 ∗ ∗ 1 1 ∗ 1 1 (7.185)
0<r< 4 : u =0, u =± 2 + 4 −r , u =± 2 − 4 −r
r> 1
4 : u∗ = 0 .
In fig. 7.19, we plot the fixed points and the hysteresis loops for this system. At r = 41 , there are two symmetrically
located saddle-node bifurcations at u = ± √12 . We find ϕ(u = ± √12 , r = 14 ) = 48 1
, which is positive, indicating
∗
that the stable fixed point u = 0 remains the thermodynamic minimum for the free energy ϕ(u) √
as r is decreased
through r = 14 . Setting ϕ(u) = 0 and ϕ′ (u) = 0 simultaneously, we obtain r = 16 3
and u = ± 23 . The thermody-
√
namic value for u therefore jumps discontinuously from u = 0 to u = ± 23 (either branch) at r = 16 3
; this is a first
order transition.
Under the dissipative dynamics considered here, the system exhibits hysteresis, as indicated in the figure, where
the arrows show the evolution of u(s) for very slowly varying r(s). When the control parameter r is large and
positive, the flow is toward the sole fixed point at u∗ = 0. At r = 41 , two simultaneous saddle-node bifurcations
take place at u∗ = ± √12 ; the outer branch is stable and the inner branch unstable in both cases. At r = 0 there is a
subcritical pitchfork bifurcation, and the fixed point at u∗ = 0 becomes unstable.
7.6. LANDAU THEORY OF PHASE TRANSITIONS 335
Figure 7.19: Fixed points ϕ′ (u∗ ) = 0 for the sextic potential ϕ(u) = 12 ru2 − 41 u4 + 16 u6 , and corresponding dynamical
flow (arrows) under u̇ = −ϕ′ (u). Solid curves show stable fixed points and dashed curves show unstable fixed
points. The thick solid black and solid grey curves indicate the equilibrium thermodynamic values for u; note the
overall u → −u symmetry. Within the region r ∈ [0, 14 ] the dynamics are irreversible and the system exhibits the
3
phenomenon of hysteresis. There is a first order phase transition at r = 16 .
Suppose one starts off with r ≫ 41 with some value u > 0. The flow u̇ = −ϕ′ (u) then rapidly results in u → 0+ .
This is the ‘high temperature phase’ in which there is no magnetization. Now let r increase slowly, using s as
the dimensionless time variable. The scaled magnetization u(s) = u∗ r(s) will remain pinned at the fixed point
u∗ = 0+ . As r passes through r = 41 , two new stable values of u∗ appear, but our system remains at u = 0+ , since
u∗ = 0 is a stable fixed point. But after the subcritical pitchfork, u∗ = 0 becomes unstable. The magnetization u(s)
1/2
then flows rapidly to the stable fixed point at u∗ = √12 , and follows the curve u∗ (r) = 12 + ( 14 − r)1/2 for all
r < 0.
Now suppose we start increasing r (i.e. increasing temperature). The magnetization follows the stable fixed point
1/2
u∗ (r) = 21 + ( 41 − r)1/2 3
past r = 0, beyond the first order phase transition point at r = 16 , and all the way up
1
to r = 4 , at which point this fixed point is annihilated at a saddle-node bifurcation. The flow then rapidly takes
u → u∗ = 0+ , where it remains as r continues to be increased further.
Within the region r ∈ 0, 14 of control parameter space, the dynamics are said to be irreversible and the behavior
of u(s) is said to be hysteretic.
336 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Expressions such as hσi i, hσi σj i, etc. are in general called correlation functions. For example, we define the spin-spin
correlation function Cij as
Cij ≡ hσi σj i − hσi i hσj i . (7.189)
∂F ∂ 2F
Expressions such as ∂Hi and are called response functions. The above relation between correlation functions
∂Hi ∂Hj
and response functions, Cij = kB T χij , is valid only for the equilibrium distribution. In particular, this relationship
is invalid if one uses an approximate distribution, such as the variational density matrix formalism of mean field
theory.
The question then arises: within mean field theory, which is more accurate: correlation functions or response
functions? A simple argument suggests that the response functions are more accurate representations of the real
physics. To see this, let’s write the variational density matrix ̺var as the sum of the exact equilibrium (Boltzmann)
distribution ̺eq = Z −1 exp(−β Ĥ) plus a deviation δ̺:
̺var = ̺eq + δ̺ . (7.190)
Then if we calculate a correlator using the variational distribution, we have
h i
hσi σj ivar = Tr ̺var σi σj
h i h i (7.191)
= Tr ̺eq σi σj + Tr δ̺ σi σj .
Thus, the variational density matrix gets the correlator right to first order in δ̺. On the other hand, the free energy
is given by
X ∂F 1 X ∂ 2F
F var eq
=F + δ̺ +
∂̺σ ̺eq σ 2 δ̺σ δ̺σ′ + . . . . (7.192)
σ σ ,σ ′ ∂̺σ ∂̺σ ′ ̺eq
Here σ denotes a state of the system, i.e. | σ i = | σ1 , . . . , σN i, where every spin polarization is specified. Since the
free energy is an extremum (and in fact an absolute minimum) with respect to the distribution, the second term
on the RHS vanishes. This means that the free energy is accurate to second order in the deviation δ̺.
7.7. MEAN FIELD THEORY OF FLUCTUATIONS 337
where
1 + mi 1 − mi
̺i (σi ) = δσi ,1 + δσi ,−1 . (7.194)
2 2
The variational energy E = Tr (̺ Ĥ) is
X X
E = − 21 Ji,j mi mj − Hi m i (7.195)
ij i
∂F
Setting the variation ∂mi = 0, with F = E − T S, we obtain the mean field equations,
mi = tanh βJij mj + βHi , (7.197)
P
where we use the summation convention: Jij mj ≡ j Jij mj . Suppose T > Tc and mi is small. Then we can
expand the RHS of the above mean field equations, obtaining
δij − βJij mj = βHi . (7.198)
∂mi −1
χij = = kB T · I − J ij , (7.199)
∂Hj
where I is the identity. Note also that so-called connected averages of the kind in eqn. 7.189 vanish identically if we
compute them using our variational density matrix, since all the sites are independent, hence
hσi σj i = Tr ̺var σi σj = Tr ̺i σi · Tr ̺j σj = hσi i · hσj i , (7.200)
and therefore χij = 0 if we compute the correlation functions themselves from the variational density matrix,
rather than from the free energy F . As we have argued above, the latter approximation is more accurate.
Assuming Jij = J(Ri − Rj ), where Ri is a Bravais lattice site, we can Fourier transform the above equation,
resulting in
Ĥ(q)
m̂(q) = ≡ χ̂(q) Ĥ(q) . (7.201)
ˆ
kB T − J(q)
Once again, our definition of lattice Fourier transform of a function φ(R) is
X
φ̂(q) ≡ φ(R) e−iq·R
R
Z d (7.202)
dq
φ(R) = Ω φ̂(q) eiq·R ,
(2π)d
Ω̂
338 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
where Ω is the unit cell in real space, called the Wigner-Seitz cell, and Ω̂ is the first Brillouin zone, which is the unit
cell in reciprocal space. Similarly, we have
X
ˆ
J(q) = J(R) 1 − iq · R − 21 (q · R)2 + . . .
R (7.203)
n o
ˆ · 1 − q 2 R2 + O(q 4 ) ,
= J(0) ∗
where P 2
R R J(R)
R∗2 = P . (7.204)
2d R J(R)
Here we have assumed inversion symmetry for the lattice, in which case
X 1 µν X 2
Rµ Rν J(R) = ·δ R J(R) . (7.205)
d
R R
√
On cubic lattices with nearest neighbor interactions only, one has R∗ = a/ 2d, where a is the lattice constant and
d is the dimension of space.
ˆ
Thus, with the identification kB Tc = J(0), we have
1
χ̂(q) =
kB (T − Tc ) + kB Tc R∗2 q 2 + O(q 4 )
(7.206)
1 1
= · −2 ,
kB Tc R∗ ξ + q + O(q 4 )
2 2
where
−1/2
T − Tc
ξ = R∗ · (7.207)
Tc
is the correlation length. With the definition
ξ(T ) ∝ |T − Tc |−ν (7.208)
as T → Tc , we obtain the mean field correlation length exponent ν = 21 . The exact result for the two-dimensional
Ising model is ν = 1, whereas ν ≈ 0.6 for the d = 3 Ising model. Note that χ̂(q = 0, T ) diverges as (T − Tc )−1 for
T > Tc .
In real space, we have
X
mi = χij Hj , (7.209)
j
where Z d
dq
χij = Ω χ̂(q) eiq·(Ri −Rj ) . (7.210)
(2π)d
Note that χ̂(q) is properly periodic under q → q + G, where G is a reciprocal lattice vector, which satisfies
eiG·R = 1 for any direct Bravais lattice vector R. Indeed, we have
where δ is a nearest neighbor separation vector, and where in the second line we have assumed nearest neighbor
interactions only. On cubic lattices in d dimensions, there are 2d nearest neighbor separation vectors, δ = ±a êµ ,
where µ ∈ {1, . . . , d}. The real space susceptibility is then
Zπ Zπ
dθ1 dθd ein1 θ1 · · · eind θd
χ(R) = ··· , (7.212)
2π 2π kB T − (2J cos θ1 + . . . + 2J cos θd )
−π −π
Pd
where R = a µ=1 nµ êµ is a general direct lattice vector for the cubic Bravais lattice in d dimensions, and the
{nµ } are integers.
The long distance behavior was discussed in chapter 6 (see §6.4.9 on Ornstein-Zernike theory18 ). For convenience
we reiterate those results:
• In d = 1,
ξ
χd=1 (x) = e−|x|/ξ . (7.213)
2kB Tc R∗2
ξ (3−d)/2 e−r/ξ d−3
χOZ
d (r) ≃ Cd · · · 1+O , (7.214)
kB T R∗2 r(d−1)/2 r/ξ
Cd′ e−r/ξ d−3
χd (r) ≃ · · 1 + O . (7.215)
kB T R∗2 rd−2 r/ξ
In d = 2 dimensions we obtain
C2′ r −r/ξ 1
χd=2 (r) ≃ · ln e · 1 + O , (7.216)
kB T R∗2 ξ ln(r/ξ)
18 There is a sign difference between the particle susceptibility defined in chapter 6 and the spin susceptibility defined here. The origin of the
difference is that the single particle potential v as defined was repulsive for v > 0, meaning the local density response δn should be negative,
while in the current discussion a positive magnetic field H prefers m > 0.
340 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Interacting systems can be broadly classified according to their global symmetry group. Consider the following five
examples:
X
ĤIsing = − Jij σi σj σi ∈ {−1, +1}
i<j
X
2π(ni − nj )
Ĥp−clock = − Jij cos ni ∈ {1, 2, . . . , p}
i<j
p
X
Ĥq−Potts = − Jij δσi ,σj σi ∈ {1, 2, . . . , q} (7.217)
i<j
X
ĤXY = − Jij cos(φi − φj ) φi ∈ 0, 2π
i<j
X
ĤO(n) = − Jij Ω̂i · Ω̂j Ω̂i ∈ S n−1 .
i<j
The Ising Hamiltonian is left invariant by the global symmetry group Z2 , which has two elements, I and η, with
η σi = −σi . (7.218)
I is the identity, and η 2 = I. By simultaneously reversing all the spins σi → −σi , the interactions remain invariant.
The degrees of freedom of the p-state clock model are integer variables ni each of which ranges from 1 to p. The
Hamiltonian is invariant under the discrete group Zp , whose p elements are generated by the single operation η,
where (
ni + 1 if ni ∈ {1, 2, . . . , p − 1}
η ni = (7.219)
1 if ni = p .
Think of a clock with one hand and p ‘hour’ markings consecutively spaced by an angle 2π/p. In each site i, a hand
points to one of the p hour marks; this determines ni . The operation η simply advances all the hours by one tick,
with hour p advancing
to hour 1, just as 23:00 military time is followed one hour later by 00:00. The interaction
cos 2π(ni − nj )/p is invariant under such an operation. The p elements of the group Zp are then
I , η , η 2 , . . . , η p−1 . (7.220)
We’ve already met up with the q-state Potts model, where each site supports a ‘spin’ σi which can be in any of q
possible states, which we may label by integers {1 , . . . , q}. The energy of two interacting sites i and j is −Jij if
σi = σj and zero otherwise. This energy function is invariant under global operations of the symmetric group on q
characters, Sq , which is the group of permutations of the sequence {1 , 2 , 3 , . . . , q}. The group Sq has q! elements.
Note the difference between a Zq symmetry and an Sq symmetry. In the former case, the Hamiltonian is invariant
only under the q-element cyclic permutations, e.g.
1 2 · · · q−1 q
η≡
2 3 ··· q 1
All these models – the Ising, p-state clock, and q-state Potts models – possess a global symmetry group which is
discrete. That is, each of the symmetry groups Z2 , Zp , Sq is a discrete group, with a finite number of elements. The
XY Hamiltonian ĤXY on the other hand is invariant under a continuous group of transformations φi → φi + α,
where φi is the angle variable on site i. More to the point, we could write the interaction term cos(φi − φj ) as
1 ∗ ∗
iφi
2 zi zj + zi zj , where zi = e is a phase which lives on the unit circle, and zi∗ is the complex conjugate of
zi . The model is then invariant under the global transformation zi → eiα zi . The phases eiα form a group under
multiplication, called U(1), which is the same as O(2). Equivalently, we could write the interaction as Ω̂i ·Ω̂j , where
Ω̂i = (cos φi , sin φi ), which explains the O(2), symmetry, since the symmetry operations are global rotations in
the plane, which is to say the two-dimensional orthogonal group. This last representation generalizes nicely to
unit vectors in n dimensions, where
Ω̂ = (Ω 1 , Ω 2 , . . . , Ω n ) (7.221)
with Ω̂ 2 = 1. The dot product Ω̂i · Ω̂j is then invariant under global rotations in this n-dimensional space, which
is the group O(n).
Depending on whether the global symmetry group of a model is discrete or continuous, there exists a lower critical
dimension dℓ at or below which no phase transition may take place at finite temperature. That is, for d ≤ dℓ , the
critical temperature is Tc = 0. Owing to its neglect of fluctuations, mean field theory generally overestimates the
value of Tc because it overestimates the stability of the ordered phase. Indeed, there are many examples where
mean field theory predicts a finite Tc when the actual critical temperature is Tc = 0. This happens whenever
d ≤ dℓ .
Let’s test the stability of the ordered (ferromagnetic) state of the one-dimensional Ising model at low temperatures.
We consider order-destroying domain wall excitations which interpolate between regions of degenerate, symmetry-
related ordered phase, i.e. ↑↑↑↑↑ and ↓↓↓↓↓. For a system with a discrete symmetry at low temperatures, the
domain wall is abrupt, on the scale of a single lattice spacing. If the exchange energy is J, then the energy of a
single domain wall is 2J, since a link of energy −J is replaced with one of energy +J. However, there are N
possible locations for the domain wall, hence its entropy is kB ln N . For a system with M domain walls, the free
energy is
N
F = 2M J − kB T ln
M
( ) (7.222)
h i
= N · 2Jx + kB T x ln x + (1 − x) ln(1 − x) ,
where x = M/N is the density of domain walls, and where we have used Stirling’s approximation for k! when k
is large. Extremizing with respect to x, we find
x 1
= e−2J/kB T =⇒ x= . (7.223)
1−x e2J/kB T + 1
The average distance between domain walls is x−1 , which is finite for finite T . Thus, the thermodynamic state of
the system is disordered, with no net average magnetization.
342 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Figure 7.21: Domain walls in the two-dimensional (left) and three-dimensional (right) Ising model.
Consider next an Ising domain wall in d dimensions. Let the linear dimension of the system be L · a, where L is
a real number and a is the lattice constant. Then the energy of a single domain wall which partitions the entire
system is 2J · Ld−1 . The domain wall entropy is difficult to compute, because the wall can fluctuate significantly,
but for a single domain wall we have S > ∼ kB ln L. Thus, the free energy F = 2JL
d−1
− kB T ln L is dominated by
the energy term if d > 1, suggesting that the system may be ordered. We can do a slightly better job in d = 2 by
writing X
d −2P J/kB T
Z ≈ exp L NP e , (7.224)
P
where the sum is over all closd loops of perimeter P , and NP is the number of such loops. An example of such a
loop circumscribing a domain is depicted in the left panel of fig. 7.21. It turns out that
n o
NP ≃ κP P −θ · 1 + O(P −1 ) , (7.225)
where κ = z − 1 with z the lattice coordination number, and θ is some exponent. We can understand the κP factor
in the following way. At each step along the perimeter of the loop, there are κ = z − 1 possible directions to go
(since one doesn’t backtrack). The fact that the loop must avoid overlapping itself and must return to its original
position to be closed leads to the power law term P −θ , which is subleading since κP P −θ = exp(P ln κ − θ ln P )
and P ≫ ln P for P ≫ 1. Thus,
1 X
F ≈ − Ld P −θ e(ln κ−2βJ)P , (7.226)
β
P
which diverges if ln κ > 2βJ, i.e. if T > 2J/kB ln(z − 1). We identify this singularity with the phase transition. The
high temperature phase involves a proliferation of such loops. The excluded volume effects between the loops,
which we have not taken into account, then enter in an essential way so that the sum converges. Thus, we have
the following picture:
2J
kB Tcapprox = = 1.82 J
ln 3
2J
kB Tcexact = = 2.27 J .
sinh−1 (1)
The agreement is better than we should reasonably expect from such a crude argument.
7.8. GLOBAL SYMMETRIES 343
Nota bene : Beware of arguments which allegedly prove the existence of an ordered phase. Generally speaking,
any approximation will underestimate the entropy, and thus will overestimate the stability of the putative ordered
phase.
When the global symmetry group is continuous, the domain walls interpolate smoothly between ordered phases.
The energy generally involves a stiffness term,
Z
E = 2 ρs ddr (∇θ)2 ,
1
(7.227)
where θ(r) is the angle of a local rotation about a single axis and where ρs is the spin stiffness. Of course, in O(n)
models, the rotations can be with respect to several different axes simultaneously.
In the ordered phase, we have θ(r) = θ0 , a constant. Now imagine a domain wall in which θ(r) rotates by 2π
across the width of the sample. We write θ(r) = 2πnx/L, where L is the linear size of the sample (here with
dimensions of length) and n is an integer telling us how many complete twists the order parameter field makes.
The domain wall then resembles that in fig. 7.22. The gradient energy is
ZL 2
d−1 2πn
E= 1
2 ρs L dx = 2π 2 n2 ρs Ld−2 . (7.228)
L
0
Recall that in the case of discrete symmetry, the domain wall energy scaled as E ∝ Ld−1 . Thus, with S >∼ kB ln L
for a single wall, we see that the entropy term dominates if d ≤ 2, in which case there is no finite temperature
phase transition. Thus, the lower critical dimension dℓ depends on whether the global symmetry is discrete or
continuous, with
Note that all along we have assumed local, short-ranged interactions. Long-ranged interactions can enhance order
and thereby suppress dℓ .
Thus, we expect that for models with discrete symmetries, dℓ = 1 and there is no finite temperature phase transi-
tion for d ≤ 1. For models with continuous symmetries, dℓ = 2, and we expect Tc = 0 for d ≤ 2. In this context
344 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
we should emphasize that the two-dimensional XY model does exhibit a phase transition at finite temperature,
called the Kosterlitz-Thouless transition. However, this phase transition is not associated with the breaking of the
continuous global O(2) symmetry and rather has to do with the unbinding of vortices and antivortices. So there
is still no true long-ranged order below the critical temperature TKT , even though there is a phase transition!
Oftentimes, particularly in condensed matter systems, intrinsic randomness exists due to quenched impurities,
grain boundaries, immobile vacancies, etc. How does this quenched randomness affect a system’s attempt to
order at T = 0? This question was taken up in a beautiful and brief paper by J. Imry and S.-K. Ma, Phys. Rev. Lett.
35, 1399 (1975). Imry and Ma considered models in which there are short-ranged interactions and a random local
field coupling to the local order parameter:
X X
ĤRFI = −J σi σj − Hi σi (7.229)
hiji i
X X
ĤRFO(n) = −J Ω̂i · Ω̂j − Hiα Ωiα , (7.230)
hiji i
where
hh Hiα ii = 0 , hh Hiα Hjβ ii = Γ δ αβ δij , (7.231)
where hh · ii denotes a configurational average over the disorder. Imry and Ma reasoned that a system could try to
lower its free energy by forming domains in which the order parameter took advantage of local fluctuations in the
random field. The size of these domains is assumed to be Ld , a length scale to be determined. See the sketch in
the left panel of fig. 7.23.
There are two contributions to the energy of a given domain: bulk and surface terms. The bulk energy is
where a is the lattice spacing. This is because when we add together (Ld /a)d random fields, the magnitude
√ of the
d/2
result is proportional to the square root of the number of terms, i.e. to (Ld /a) . The quantity Hrms = Γ is the
root-mean-square fluctuation in the random field at a given site. The surface energy is
(
J (Ld /a)d−1 (discrete symmetry)
Esurface ∝ (7.233)
J (Ld /a)d−2 (continuous symmetry) .
We compute the critical dimension dc by balancing the bulk and surface energies,
d − 1 = 12 d =⇒ dc = 2 (discrete)
d − 2 = 12 d =⇒ dc = 4 (continuous) .
The total free energy is F = (V /Ldd ) · ∆E, where ∆E = Ebulk + Esurf . Thus, the free energy per unit cell is
1 dc 1 d
F a 2 a 2
f= ≈J − Hrms . (7.234)
V /ad Ld Ld
If d < dc , the surface term dominates for small Ld and the bulk term dominates for large Ld There is global
minimum at
d 2−d
Ld dc J c
= · . (7.235)
a d Hrms
7.9. GINZBURG-LANDAU THEORY 345
Figure 7.23: Left panel : Imry-Ma domains for an O(2) model. The arrows point in the direction of the local order
parameter field hΩ̂(r)i. Right panel : free energy density as a function of domain size Ld . Keep in mind that the
minimum possible value for Ld is the lattice spacing a.
For d > dc , the relative dominance of the bulk and surface terms is reversed, and there is a global maximum at
this value of Ld .
Sketches of the free energy f (Ld ) in both cases are provided in the right panel of fig. 7.23. We must keep in mind
that the domain size Ld cannot become smaller than the lattice spacing a. Hence we should draw a vertical line on
the graph at Ld = a and discard the portion Ld < a as unphysical. For d < dc , we see that the state with Ld = ∞,
i.e. the ordered state, is never the state of lowest free energy. In dimensions d < dc , the ordered state is always unstable
to domain formation in the presence of a random field.
For d > dc , there are two possibilities, depending on the relative size of J and Hrms . We can see this by evaluating
f (Ld = a) = J − Hrms and f (Ld = ∞) = 0. Thus, if J > Hrms , the minimum energy state occurs for Ld = ∞. In
this case, the system has an ordered ground state, and we expect a finite temperature transition to a disordered
state at some critical temperature Tc > 0. If, on the other hand, J < Hrms , then the fluctuations in H overwhelm
the exchange energy at T = 0, and the ground state is disordered down to the very smallest length scale (i.e. the
lattice spacing a).
Please read the essay, “Memories of Shang-Keng Ma,” at http://sip.clarku.edu/skma.html.
In principle, any term which does not violate the appropriate global symmetry will turn up in such an expansion
of the free energy, with some coefficient. Examples include hm3 (both m and h are odd under time reversal),
346 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
m2 (∇m)2 , etc. We now ask: what function m(x) extremizes the free energy functional F m(x) , h(x) ? The
answer is that m(x) must satisfy the corresponding Euler-Lagrange equation, which for the above functional is
a m + b m3 + c m5 − h − κ ∇2 m = 0 . (7.237)
If a > 0 and h is small (we assume b > 0 and c > 0), we may neglect the m3 and m5 terms and write
a − κ ∇2 m = h , (7.238)
ĥ(q)
m̂(q) = , (7.239)
a + κq 2
which, with h(x) appropriately defined, recapitulates the result in eqn. 7.201. Thus, we conclude that
1
χ̂(q) = , (7.240)
a + κq 2
which should be compared with eqn. 7.206. For continuous functions, we have
Z
m̂(q) = ddx m(x) e−iq·x (7.241)
Z d
dq
m(x) = m̂(q) eiq·x . (7.242)
(2π)d
We can then derive the result Z
m(x) = ddx′ χ(x − x′ ) h(x′ ) , (7.243)
where Z ′
1 ddq eiq·(x−x )
χ(x − x′ ) = , (7.244)
κ (2π)d q 2 + ξ −2
p
where the correlation length is ξ = κ/a ∝ (T − Tc )−1/2 , as before.
If a < 0 then there is a spontaneous magnetization and we write m(x) = m0 + δm(x). Assuming h is weak, we
then have two equations
ĥ(q)
δ m̂(q) = , (7.247)
−2a + κq 2
A particularly interesting application of Ginzburg-Landau theory is its application toward modeling the spatial
profile of defects such as vortices and domain walls. Consider, for example, the case of Ising (Z2 ) symmetry with
h = 0. We expand the free energy density to order m4 :
Z
F m(x) = ddx f0 + 12 a m2 + 14 b m4 + 21 κ (∇m)2 . (7.248)
7.9. GINZBURG-LANDAU THEORY 347
We assume a < 0, corresponding to T < Tc . Consider now a domain wall, where m(x → −∞) = −m0 and
m(x → +∞) = +m0 , where m0 is the equilibrium magnetization, which we obtain from the Euler-Lagrange
equation,
a m + b m3 − κ ∇2 m = 0 , (7.249)
q
assuming a uniform solution where ∇m = 0. This gives m0 = |a| b . It is useful to scale m(x) by m0 , writing
m(x) = m0 φ(x). The scaled order parameter function φ(x) will interpolate between φ(−∞) = −1 and φ(+∞) =
1.
It also proves useful to rescale position, writing x = (2κ/b)1/2 ζ. Then we obtain
2
1
2∇ φ = −φ + φ3 . (7.250)
We assume φ(ζ) = φ(ζ) is only a function of one coordinate, ζ ≡ ζ 1 . Then the Euler-Lagrange equation becomes
d2φ ∂U
2
= −2φ + 2φ3 ≡ − , (7.251)
dζ ∂φ
where
2
U (φ) = − 21 φ2 − 1 . (7.252)
The ‘potential’ U (φ) is an inverted double well, with maxima at φ = ±1. The equation φ̈ = −U ′ (φ), where dot
denotes differentiation with respect to ζ, is simply Newton’s second law with time replaced by space. In order
to have a stationary solution at ζ → ±∞ where φ = ±1, the total energy must be E = U (φ = ±1) = 0, where
E = 12 φ̇2 + U (φ). This leads to the first order differential equation
dφ
= 1 − φ2 , (7.253)
dζ
with solution
φ(ζ) = tanh(ζ) . (7.254)
Restoring the dimensionful constants,
r r
|a| b
m(x) = tanh x . (7.255)
b 2κ
We can make some progress in systematically deriving the Ginzburg-Landau free energy. Consider the Ising
model,
Ĥ X X X
= − 12 Kij σi σj − hi σi + 12 Kii , (7.256)
kB T i,j i i
where now Kij = Jij /kB T and hi = Hi /kB T are the interaction energies and local magnetic fields in units of
kB T . The last term on the RHS above cancels out any contribution from diagonal elements of Kij . Our derivation
makes use of a generalization of the Gaussian integral,
Z∞ 1/2
− 12 ax2 −bx 2π 2
dx e = eb /2a . (7.257)
a
−∞
348 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
The generalization is
Z∞ Z∞
1 (2π)N/2 21 A−1
dx1 · · · dxN e− 2 Aij xi xj −bi xi = √ e ij bi bj , (7.258)
det A
−∞ −∞
where we use the Einstein convention of summing over repeated indices, and where we assume that the matrix A
is positive definite (else the integral diverges). This allows us to write
1 1
Z = e− 2 Kii Tr e 2 Kij σi σj ehi σi
Z∞ Z∞
1 −1
−1/2 − 12 Kii
= det (2πK) e dφ1 · · · dφN e− 2 Kij φi φj Tr e(φi +hi )σi
−∞ −∞
Z∞ Z∞ P
1 1 −1
= det−1/2 (2πK) e− 2 Kii dφ1 · · · dφN e− 2 Kij φi φj
e i ln[2 cosh(φi +hi )]
−∞ −∞
Z∞ Z∞
≡ dφ1 · · · dφN e−Φ(φ1 ,...,φN ) , (7.259)
−∞ −∞
where
X X
1 −1 1 1
Φ= 2 Kij φi φj − ln cosh(φi + hi ) + 2 ln det (2πK) + 2 Tr K − N ln 2 . (7.260)
i,j i
We assume the model is defined on a Bravais lattice, in which case we can write φi = φR . We can then define the
i
Fourier transforms,
1 X
φR = √ φ̂q eiq·R (7.261)
N q
1 X
φ̂q = √ φR e−iq·R (7.262)
N R
and X
K̂(q) = K(R) e−iq·R . (7.263)
R
A few remarks about the lattice structure and periodic boundary conditions are in order. For a Bravais lattice, we
can write each direct lattice vector R as a sum over d basis vectors with integer coefficients, viz.
d
X
R= nµ a µ , (7.264)
µ=1
aµ · bν = 2π δµν , (7.265)
φR+Pd = φR . (7.267)
µ=1 lµ Mµ aµ
This leads to the quantization of the wavevectors, which must then satisfy
and therefore θµ = 2πmµ /Mµ , where mµ is an integer. There are then M1 M2 · · · Md = N independent values of q,
which can be taken to be those corresponding to mµ ∈ {1, . . . , Mµ }.
Let’s now expand the function Φ φ ~ in powers of the φ , and to first order in the external fields h . We obtain
i i
X X X
Φ = 21 K̂ −1 (q) − 1 |φ̂q |2 + 12
1
φ4R − hR φR + O φ6 , h2 (7.269)
q R R
1 1
+ 2 Tr K + 2 Tr ln(2πK) − N ln 2
P
On a d-dimensional lattice, for a model with nearest neighbor interactions K1 only, we have K̂(q) = K1 δ eiq·δ ,
where δ is a nearest neighbor separation vector. These are the eigenvalues of the matrix Kij . We note that Kij is
then not positive definite, since there are negative eigenvalues19 . To fix this, we can add a term K0 everywhere
along the diagonal. We then have X
K̂(q) = K0 + K1 cos(q · δ) . (7.270)
δ
Here we have used the inversion symmetry of the Bravais lattice to eliminate the imaginary term. The eigenvalues
are all positive so long as K0 > zK1 , where z is the lattice coordination number. We can therefore write K̂(q) =
K̂(0) − α q 2 for small q, with α > 0. Thus, we can write
K̂ −1 (q) − 1 = a + κ q 2 + . . . . (7.271)
To lowest order in q the RHS is isotropic if the lattice has cubic symmetry, but anisotropy will enter in higher
order terms. We’ll assume isotropy at this level. This is not necessary but it makes the discussion somewhat less
involved. We can now write down our Ginzburg-Landau free energy density:
F = a φ2 + 12 κ |∇φ|2 + 1
12 φ4 − h φ , (7.272)
ˆ ˆ 1 2
Surely we could expand J(q) = J(0)− 2 aq +. . . and obtain a similar expression for F . However, such a derivation
using the variational density matrix is only approximate. The method outlined in this section is exact.
Let’s return to our complete expression for Φ:
X
~ =Φ φ
Φ φ ~ + v(φR ) , (7.274)
0
R
19 To π
evoke a negative eigenvalue on a d-dimensional cubic lattice, set qµ = a
for all µ. The eigenvalue is then −2dK1 .
350 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
where
X 2 1 2π
~ =
Φ0 φ 1
G−1 (q) φ̂(q) + 1
Tr + 1
Tr ln − N ln 2 . (7.275)
2 2 1 + G−1 2 1 + G−1
q
v(φ) = 12 φ2 − ln cosh φ
(7.276)
= 1
12 φ4 − 1
45 φ6 + 17
2520 φ8 + . . .
and
K̂(q)
G(q) = . (7.277)
1 − K̂(q)
We now want to compute
Z P
~ e−Φ0 (φ) e− ~
Z= Dφ R v(φR ) (7.278)
where
~ ≡ dφ dφ · · · dφ .
Dφ (7.279)
1 2 N
where
Z
~ e−Φ0 (φ) ~
Z0 = Dφ
(7.281)
1 G
ln Z0 = 2 Tr ln(1 + G) − + N ln 2
1+G
and R
~ F e−Φ0
Dφ
~
F φ = R . (7.282)
~ e−Φ0
Dφ
To evaluate the various terms in the expansion of eqn. 7.280, we invoke Wick’s theorem, which says
Z∞ Z∞ , Z∞ Z∞
−1
− 21 Gij xi xj 1 −1
xi xi · · · xi = dx1 · · · dxN e xi xi · · · xi dx1 · · · dxN e− 2 Gij xi xj
1 2 2L 1 2 2L
−∞ −∞ −∞ −∞
X
= Gj Gj · · · Gj , (7.283)
1 j2 3 j4 2L−1 j2L
all distinct
pairings
where the sets {j1 , . . . , j2L } are all permutations of the set {i1 , . . . , i2L }. In particular, we have
4 2
xi = 3 Gii . (7.284)
1
Thus, if we write v(φ) ≈ 12φ4 and retain only the quartic term in v(φ), we obtain
F G 1 2
= − ln Z0 = 21 Tr − ln(1 + G) + Tr G − N ln 2
kB T 1+G 4N
(7.286)
1 2 1
= −N ln 2 + Tr G − Tr G2 + O G3 .
4N 4
Note that if we set Kij to be diagonal, then K̂(q) and hence G(q) are constant functions of q. The O G2 term
then vanishes, which is required since the free energy cannot depend on the diagonal elements of Kij .
Let us define A(T, H, V, N ) to be the usual (i.e. thermodynamic) Helmholtz free energy. Then
Z
e−βA = Dm e−βF [m(x)] , (7.287)
where the functional F [m(x)] is of the Ginzburg-Landau form, given in eqn. 7.248. The integral above is a
functional integral. We can give it a more precise meaning by defining its measure in the case of periodic functions
m(x) confined to a rectangular box. Then we can expand
1 X
m(x) = √ m̂q eiq·x , (7.288)
V q
Note that the fact that m(x) ∈ R means that m̂−q = m̂∗q . We’ll assume T > Tc and H = 0 and we’ll explore limit
T → Tc+ from above to analyze the properties of the critical region close to Tc . In this limit we can ignore all but
the quadratic terms in m, and we have
Z X
e−βA = Dm exp − 12 β (a + κ q 2 ) |m̂q |2
q
(7.290)
Y πkB T
1/2
= .
q
a + κ q2
Thus,
X a + κ q2
1
A= 2 kB T ln . (7.291)
q
πkB T
T − Tc
t= , (7.292)
Tc
CV , which means that we’re only interested in differentiating with respect to T as it appears in a(T ). We divide
by Ns kB where Ns is the number of unit cells of our system, which we presume is a lattice-based model. Note
352 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Ns ∼ V /ad where V is the volume and a the lattice constant. The dimensionless heat capacity per lattice site is
then
ZΛ
CV α2 ad ddq 1
c≡ = 2
, (7.293)
Ns 2κ (2π) (ξ + q 2 )2
d −2
where ξ = (κ/αt)1/2 ∝ |t|−1/2 is the correlation length, and where Λ ∼ a−1 is an ultraviolet cutoff. We define
R∗ ≡ (κ/α)1/2 , in which case
ZΛξ d
−4 d 4−d 1 d q̄ 1
c = R∗ a ξ ·2 , (7.294)
(2π)d (1 + q̄ 2 )2
where q̄ ≡ qξ. Thus,
const. if d > 4
c(t) ∼ − ln t if d = 4 (7.295)
d2 −2
t if d < 4 .
For d > 4, mean field theory is qualitatively accurate, with finite corrections. In dimensions d ≤ 4, the mean
field result is overwhelmed by fluctuation contributions as t → 0+ (i.e. as T → Tc+ ). We see that MFT is sensible
provided the fluctuation contributions are small, i.e. provided
In both the variational density matrix and mean field Hamiltonian methods as applied to the Ising model, we
obtained the same result m = tanh (m + h)/θ . What is perhaps not obvious is whether these theories are in fact
the same, i.e. if their respective free energies agree. Indeed, the two free energy functions,
1+m 1+m 1−m 1−m
fA (m, h, θ) = − 21 m2 − hm + θ ln + ln
2 2 2 2
fB (m, h, θ) = + 21 m2 − θ ln e+(m+h)/θ + e−(m+h)/θ , (7.298)
where fA is the variational density matrix result and fB is the mean field Hamiltonian result, clearly are different
functions of their arguments. However, it turns out that upon minimizing with respect to m in each cast, the
resulting free energies obey fA (h, θ) = fB (h, θ). This agreement may seem surprising. The first method utilizes an
7.10. APPENDIX I : EQUIVALENCE OF THE MEAN FIELD DESCRIPTIONS 353
approximate (variational) density matrix applied to the exact Hamiltonian Ĥ. The second method approximates
the Hamiltonian as ĤMF , but otherwise treats it exactly. The two Landau expansions seem hopelessly different:
fA (m, h, θ) = −θ ln 2 − hm + 1
2 (θ − 1) m2 + θ
12 m4 + θ
30 m6 + . . . (7.299)
2 4 6
(m + h) (m + h) (m + h)
fB (m, h, θ) = −θ ln 2 + 21 m2 − + − + ... . (7.300)
2θ 12 θ3 45 θ5
We shall now prove that these two methods, the variational density matrix and the mean field approach, are in
fact equivalent, and yield the same free energy f (h, θ).
Let us generalize the Ising model and write
X X
Ĥ = − Jij ε(σi , σj ) − Φ(σi ) . (7.301)
i<j i
Here, each ‘spin’ σi may take on any of K possible values, {s1 , . . . , sK }. For the S = 1 Ising model, we would
have K = 3 possibilities, with s1 = −1, s2 = 0, and s3 = +1. But the set {sα }, with α ∈ {1, . . . , K}, is completely
arbitrary20 . The ‘local field’ term Φ(σ) is also a completely arbitrary function. It may be linear, with Φ(σ) = Hσ,
for example, but it could also contain terms quadratic in σ, or whatever one desires.
The symmetric, dimensionless interaction function ε(σ, σ ′ ) = ε(σ ′ , σ) is areal symmetric K × K matrix. According
to the singular value decomposition theorem, any such matrix may be written in the form
Ns
X
ε(σ, σ ′ ) = Ap λp (σ) λp (σ ′ ) , (7.302)
p=1
where the {Ap } are coefficients (the singular values), and the λp (σ) are the singular vectors. The number of
terms Ns in this decomposition is such that Ns ≤ K. This treatment can be generalized to account for continuous
σ.
Thus, xα is the probability for a given site P to be in state α, with σ = sα . The {xα } are the K variational parameters,
subject to the single normalization constraint, α xα = 1. We now have
1
f= Tr (̺Ĥ) + kB T Tr (̺ ln ̺)
ˆ
N J(0)
XX X X (7.304)
= − 21 Ap λp (sα ) λp (sα′ ) xα xα′ − ϕ(sα ) xα + θ xα ln xα ,
p α,α′ α α
and freely extremizing with respect to the (K + 1) parameters {x1 , . . . , xK , ζ}. This yields K nonlinear equations,
∂f ∗ XX
0= =− Ap λp (sα ) λp (sα′ ) xα′ − ϕ(sα ) + θ ln xα + ζ + θ , (7.306)
∂xα p ′ α
for each α, and one linear equation, which is the normalization condition,
∂f ∗ X
0= = xα − 1 . (7.307)
∂ζ α
We cannot solve these nonlinear equations analytically, but they may be recast, by exponentiating them, as
( )
1 1 XX
xα = exp Ap λp (sα ) λp (sα′ ) xα′ + ϕ(sα ) , (7.308)
Z θ p ′ α
with (
X )
(ζ/θ)+1 1 XX
Z =e = exp Ap λp (sα ) λp (sα′ ) xα′ + ϕ(sα ) . (7.309)
α
θ p ′ α
From the logarithm of xα , we may compute the entropy, and, finally, the free energy:
XX
1
f (h, θ) = 2 Ap λp (sα ) λp (sα′ ) xα xα′ − θ ln Z , (7.310)
p α,α′
which is to be evaluated at the solution of 7.306, x∗α (h, θ)
We now derive a mean field approximation in the spirit of that used in the Ising model above. We write
λp (σ) = λp (σ) + δλp (σ) , (7.311)
and abbreviate λ̄p = λp (σ) , the thermodynamic average of λp (σ) on any given site. We then have
λp (σ) λp (σ ′ ) = λ̄2p + λ̄p δλp (σ) + λ̄p δλp (σ ′ ) + δλp (σ) δλp (σ ′ )
(7.312)
= −λ̄2p + λ̄p λp (σ) + λp (σ ′ ) + δλp (σ) δλp (σ ′ ) .
The product δλp (σ) δλp (σ ′ ) is of second order in fluctuations, and we neglect it. This leads us to the mean field
Hamiltonian,
X X X
1
ĤMF = + 2 N J(0)ˆ 2
Ap λ̄p − ˆ
J(0) Ap λ̄p λp (σi ) + Φ(σi ) . (7.313)
p i p
which again agrees with what we found using the variational density matrix.
Thus, whether one extremizes with respect to the set {x1 , . . . , xK , ζ}, or with respect to the set {λ̄p }, the results
are the same, in terms of all these parameters, as well as the free energy f (h, θ). Generically, both approaches may
be termed ‘mean field theory’ since the variational density matrix corresponds to a mean field which acts on each
site independently.
The Blume-Capel model provides a simple and convenient way to model systems with vacancies. The simplest
version of the model is written X X
Ĥ = − 21 Jij Si Sj + ∆ Si2 . (7.318)
i,j i
The spin variables Si range over the values {−1 , 0 , +1}, so this is an extension of the S = 1 Ising model. We
explicitly separate out the diagonal terms, writing Jii ≡ 0, and placing them in the second term on the RHS
above. We say that site i is occupied if Si = ±1 and vacant if Si = 0, and we identify −∆ as the vacancy creation
energy, which may be positive or negative, depending on whether vacancies are disfavored or favored in our
system.
We make the mean field Ansatz, writing Si = m + δSi . This results in the mean field Hamiltonian,
X X
ˆ m2 − J(0)
ĤMF = 12 N J(0) ˆ m Si + ∆ Si2 . (7.319)
i i
ˆ
Once again, we adimensionalize, writing f ≡ F/N J(0), ˆ
θ = kB T /J(0), ˆ
and δ = ∆/J(0). ˆ > 0. The
We assume J(0)
free energy per site is then
f (θ, δ, m) = 12 m2 − θ ln 1 + 2e−δ/θ cosh(m/θ) . (7.320)
2 sinh(m/θ)
m= . (7.321)
exp(δ/θ) + 2 cosh(m/θ)
356 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Figure 7.24: Mean field phase diagram for the Blume-Capel model. The black dot signifies a tricritical point, where
the coefficients of m2 and m4 in the Landau free energy expansion both vanish. The dashed curve denotes a first
order transition, and the solid curve a second order transition. The thin dotted line is the continuation of the θc (δ)
relation to zero temperature.
Note that m = 0 is always a solution. Finding the slope of the RHS at m = 0 and setting it to unity gives us the
critical temperature:
2
θc = . (7.322)
exp(δ/θc ) + 2
This is an implicit equation for θc in terms of the vacancy energy δ.
Let’s now expand the free energy in terms of the magnetization m. We find, to fourth order,
1 2
f = −θ ln 1 + 2e−δ/θ + θ− m2
2θ 2 + exp(δ/θ)
(7.323)
1 6
+ − 1 m4 + . . . .
12 2 + exp(δ/θ) θ3 2 + exp(δ/θ)
Note that setting the coefficient of the m2 term to zero yields the equation for θc . However, upon further exami-
nation, we see that the coefficient of the m4 term can also vanish. As we have seen, when both the coefficients of
the m2 and the m4 terms vanish, we have a tricritical point21 . Setting both coefficients to zero, we obtain
1 2
θt = 3 , δt = 3 ln 2 . (7.324)
At θ = 0, it is easy to see we have a first order transition, simply by comparing the energies of the paramagnetic
(Si = 0) and ferromagnetic (Si = +1 or Si = −1) states. We have
(
EMF 0 if m = 0
= 1 (7.325)
ˆ
N J(0) − ∆ if m = ±1 .
2
These results are in fact exact, and not only valid for the mean field theory. Mean field theory is approximate
because it neglects fluctuations, but at zero temperature, there are no fluctuations to neglect!
21 We should really check that the coefficient of the sixth order term is positive, but that is left as an exercise to the eager student.
7.11. APPENDIX II : ADDITIONAL EXAMPLES 357
The phase diagram is shown in fig. 7.24. Note that for δ large and negative, vacancies are strongly disfavored,
hence the only allowed states on each site have Si = ±1, which is our old friend the two-state Ising model.
Accordingly, the phase boundary there approaches the vertical line θc = 1, which is the mean field transition
temperature for the two-state Ising model.
with J > 0 and σi = ±1. We’ve solved for the mean field phase diagram of the Ising ferromagnet; what happens
if the interactions are antiferromagnetic?
It turns out that under certain circumstances, the ferromagnet and the antiferromagnet behave exactly the same
in terms of their phase diagram, response functions, etc. This occurs when H = 0, and when the interactions are
between nearest neighbors on a bipartite lattice. A bipartite lattice is one which can be divided into two sublattices,
which we call A and B, such that an A site has only B neighbors, and a B site has only A neighbors. The square,
honeycomb, and body centered cubic (BCC) lattices are bipartite. The triangular and face centered cubic lattices
are non-bipartite. Now if the lattice is bipartite and the interaction matrix Jij is nonzero only when i and j are
from different sublattices (they needn’t be nearest neighbors only), then we can simply redefine the spin variables
such that (
′ +σj if j ∈ A
σj = (7.327)
−σj if j ∈ B .
Then σi′ σj′ = −σi σj , and in terms of the new spin variables the exchange constant has reversed. The thermody-
namic properties are invariant under such a redefinition of the spin variables.
We can see why this trick doesn’t work in the presence of a magnetic field, because the field H would have to be
reversed on the B sublattice. In other words, the thermodynamics of an Ising ferromagnet on a bipartite lattice
in a uniform applied field is identical to that of the Ising antiferromagnet, with the same exchange constant (in
magnitude), in the presence of a staggered field HA = +H and HB = −H.
We treat this problem using the variational density matrix method, using two independent variational parameters
mA and mB for the two sublattices:
1 + mA 1 − mA
̺A (σ) = δσ,1 + δσ,−1
2 2 (7.328)
1 + mB 1 − mB
̺B (σ) = δσ,1 + δσ,−1 .
2 2
With the usual adimensionalization, f = F/N zJ, θ = kB T /zJ, and h = H/zJ, we have the free energy
f (mA , mB ) = 21 mA mB − 1
2 h (mA + mB ) − 1
2 θ s(mA ) − 1
2 θ s(mB ) , (7.329)
Note that
ds 1+m d2s 1
= − 21 ln , =− . (7.331)
dm 1−m dm2 1 − m2
358 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Figure 7.25: Graphical solution to the mean field equations for the Ising antiferromagnet in an external field, here
for θ = 0.6. Clockwise from upper left: (a) h = 0.1, (b) h = 0.5, (c) h = 1.1, (d) h = 1.4.
Differentiating f (mA , mB ) with respect to the variational parameters, we obtain two coupled mean field equations:
∂f θ 1 + mA
= 0 =⇒ mB = h − ln
∂mA 2 1 − mA
(7.332)
∂f θ 1 + mB
= 0 =⇒ mA = h − ln .
∂mB 2 1 − mB
h i
Recognizing tanh−1 (x) = 1
2 ln (1 + x)/(1 − x) , we may write these equations in an equivalent but perhaps more
suggestive form:
h − mB h − mA
mA = tanh , mB = tanh . (7.333)
θ θ
In other words, the A sublattice sites see an internal field HA,int = −zJmB from their B neighbors, and the B
sublattice sites see an internal field HB,int = −zJmA from their A neighbors.
We can solve these equations graphically, as in fig. 7.25. Note that there is always a paramagnetic solution with
mA = mB = m, where
θ 1+m h−m
m = h − ln ⇐⇒ m = tanh . (7.334)
2 1−m θ
7.11. APPENDIX II : ADDITIONAL EXAMPLES 359
Figure 7.26: Mean field phase diagram for the Ising antiferromagnet in an external field. The phase diagram is
symmetric under reflection in the h = 0 axis.
However, we can see from the figure that there will be three solutions to the mean field equations provided that
∂mA
∂mB < −1 at the point of the solution where mA = mB = m. This gives us two equations with which to eliminate
mA and mB , resulting in the curve
θ 1+m √
h∗ (θ) = m + ln with m = 1 − θ . (7.335)
2 1−m
Thus, for θ < 1 and |h| < h∗ (θ) there are three solutions to the mean field equations. It is usually the case, the
broken symmetry solutions, which mean those for which mA 6= mB in our case, are of lower energy than the
symmetric solution(s). We show the curve h∗ (θ) in fig. 7.26.
We can make additional progress by defining the average and staggered magnetizations m and ms ,
m ≡ 12 (mA + mB ) , ms ≡ 21 (mA − mB ) . (7.336)
We expand the free energy in terms of ms :
f (m, ms ) = 21 m2 − 21 m2s − h m − 1
θ s(m + ms ) − 21 θ s(m − ms )
2
(7.337)
= 21 m2 − h m − θ s(m) − 12 1 + θ s′′ (m) m2s − 24
1
θ s′′′′ (m) m4s + . . . .
√
The term quadratic in ms vanishes when θ s′′ (m) = −1, i.e. when m = 1 − θ. It is easy to obtain
d3s 2m d4s 2 (1 + 3m2 )
=− , = − , (7.338)
dm3 (1 − m2 )2 dm4 (1 − m2 )3
1
from which we learn that the coefficient of the quartic term, − 24 θ s′′′′ (m), never vanishes. Therefore the transition
remains second order down to θ = 0, where it finally becomes first order.
We can confirm the θ → 0 limit directly. The two competing states are the ferromagnet, with mA = mB = ±1, and
the antiferromagnet, with mA = −mB = ±1. The free energies of these states are
1
f FM = 2 −h , f AFM = − 12 . (7.339)
There is a first order transition when f FM = f AFM , which yields h = 1.
360 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
where σi is the vector of Pauli matrices on site i. The spins live on a square lattice. The second sum is over all
square plaquettes. All the constants J, ∆, and K are positive.
Let’s take a look at the Hamiltonian for a moment. The J term clearly wants the spins to align ferromagnetically
in the (x, y) plane (in internal spin space). The ∆ term prefers antiferromagnetic alignment along the ẑ axis. The
K term discourages any kind of moment along ẑ and works against the ∆ term. We’d like our mean field theory
to capture the physics behind this competition.
√ √
Accordingly, we break up the square lattice into two interpenetrating 2 × 2 square sublattices (each rotated by
45◦ with respect to the original), in order to be able to describe an antiferromagnetic state. In addition, we include
a parameter α which describes the canting angle that the spins on these sublattices make with respect to the x̂-axis.
That is, we write
̺A = 1
2 + 21 m sin α σ x + cos α σ z )
̺B = 1
+ 21 m sin α σ x − cos α σ z ) . (7.341)
2
Note that Tr ̺A = Tr ̺B = 1 so these density matrices are normalized. Note also that the mean direction for a spin
on the A and B sublattices is given by
Thus, when α = 0, the system is an antiferromagnet with its staggered moment lying along the ẑ axis. When
α = 21 π, the system is a ferromagnet with its moment lying along the x̂ axis.
Finally, the eigenvalues of ̺A,B are still λ± = 21 (1 ± m), hence
Note that we have taken mA = mB = m, unlike the case of the antiferromagnet in a uniform field. The reason is
that there remains in our model a symmetry between A and B sublattices.
The free energy is now easily calculated:
F = Tr (̺Ĥ) + kB T Tr (̺ ln ̺)
(7.344)
= −2N J sin2 α + ∆ cos2 α m2 + 14 N Km4 cos4 α − N kB T s(m)
We can adimensionalize by defining δ ≡ ∆/J, κ ≡ K/4J, and θ ≡ kB T /4J. Then the free energy per site is
f ≡ F/4N J is
f (m, α) = − 21 m2 + 21 1 − δ m2 cos2 α + 14 κ m4 cos4 α − θ s(m) . (7.345)
7.11. APPENDIX II : ADDITIONAL EXAMPLES 361
Figure 7.27: Mean field phase diagram for the model of eqn. 7.340 for the case κ = 1.
There are two variational parameters: m and θ. We thus obtain two coupled mean field equations,
∂f 2 3 4 1 1+m
= 0 = −m + 1 − δ m cos α + κ m cos α + 2 θ ln (7.346)
∂m 1−m
∂f
= 0 = 1 − δ + κ m2 cos2 α m2 sin α cos α . (7.347)
∂α
Let’s start with the second of the mean field equations. Assuming m 6= 0, it is clear from eqn. 7.345 that
0 if δ < 1
cos2 α = (δ − 1)/κm2 if 1 ≤ δ ≤ 1 + κ m2 (7.348)
1 if δ ≥ 1 + κ m2 .
Suppose δ < 1. Then we have cos α = 0 and the first mean field equation yields the familiar result
m = tanh m/θ . (7.349)
dimensionless temperature r
m0 δ−1
θ0 = <1 ; m0 = . (7.351)
tanh−1 (m0 ) κ
For θ > θ0 , we have δ > 1 + κm2 , and we must take cos2 α = 1. The first mean field equation then becomes
θ 1+m
δm − κm3 = ln , (7.352)
2 1−m
or, equivalently, m = tanh (δm − κm3 )/θ . A simple graphical analysis shows that a nontrivial solution exists
provided θ < δ. Since cos α = ±1, this solution describes an antiferromagnet, with mA = ±mẑ and mB = ∓mẑ.
The resulting mean field phase diagram is then as depicted in fig. 7.27.
f (m, φ) = 1
2 am m 2 + 1
4 b m m4 + 1
2 aφ φ2 + 1
4 bφ φ4 + 21 Λ m2 φ2 . (7.353)
We write
am ≡ αm θm , aφ = αφ θφ , (7.354)
where
T − Tc,m T − Tc,φ
θm = , θφ = , (7.355)
T0 T0
where T0 is some temperature scale. We assume without loss of generality that Tc,m > Tc,φ . We begin by rescaling:
1/2 1/2
αm αm
m≡ m
e , φ≡ φe . (7.356)
bm bm
We then have
f = ε0 r 1
2 θm m
e + 2 1
4 m
e 4
+r −1 1
2 θφ φe2 + 1
4 φe4 + 1
2 λm
e φ2 e2
, (7.357)
where 1/2
αm αφ αm bφ Λ
ε0 = , r= , λ= . (7.358)
(bm bφ )1/2 αφ bm (bm bφ )1/2
It proves convenient to perform one last rescaling, writing
Then
1 2 1 4 1 −1 2 1 4 1 2 2
f = ε0 2 q θm m + 4 m + 2q θφ ϕ + 4 ϕ + 2 λm ϕ , (7.360)
1
The eigenvalues of the above 2 × 2 matrix are 1 ± λ, with corresponding eigenvectors ±1 . Since ϕ2 > 0, we are
1
only interested in the first eigenvector 1 , corresponding to the eigenvalue 1 + λ. Clearly when λ < 1 the free
energy is unbounded from below, which is unphysical.
We now set
∂f ∂f
=0 , =0, (7.363)
∂m ∂ϕ
and identify four possible phases:
ε0
f= q θm m2 + 1
2 m4 , (7.364)
2
hence we require θm < 0 in this phase, in which case
p ε0 2 2
mII = −q θm , fII = − q θm . (7.365)
4
ε0 −1
f= q θφ ϕ2 + 1
2 ϕ4 , (7.366)
2
hence we require θφ < 0 in this phase, in which case
q ε0 −2 2
ϕIII = −q −1 θφ , fIII = − q θφ . (7.367)
4
with solution
q θm − q −1 θφ λ
m2 = (7.369)
λ2 − 1
q −1 θφ − q θm λ
ϕ2 = . (7.370)
λ2 − 1
Since m2 and ϕ2 must each be nonnegative, phase IV exists only over a yet-to-be-determined subset of the
entire parameter space. The free energy is
q 2 θm
2
+ q −2 θφ2 − 2λ θm θφ
fIV = . (7.371)
4(λ2 − 1)
We now define θ ≡ θm and τ ≡ θφ − θm = (Tc,m − Tc,φ) /T0 . Note that τ > 0. There are three possible temperature
ranges to consider.
364 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
(1) θφ > θm > 0. The only possible phases are I and IV. For phase IV, we must impose the conditions m2 > 0
and φ2 > 0. If λ2 > 1, then the numerators in eqns. 7.369 and 7.370 must each be positive:
2
q 2 θm θφ q θm θφ
λ< , λ< 2 ⇒ λ < min , 2 . (7.372)
θφ q θm θφ q θm
But since either q 2 θm /θφ or its inverse must be less than or equal to unity, this requires λ < −1, which is
unphysical.
If on the other hand we assume λ2 < 1, the non-negativeness of m2 and ϕ2 requires
2
q 2 θm θφ q θm θφ
λ> , λ> 2 ⇒ λ > max , 2 >1. (7.373)
θφ q θm θφ q θm
(2) θφ > 0 > θm . Now the possible phases are I, II, and IV. We can immediately rule out phase I because fII < fI .
To compare phases II and IV, we compute
(q λ θm − q −1 θφ )2
∆f = fIV − fII = . (7.374)
4(λ2 − 1)
Thus, phase II has the lower energy if λ2 > 1. For λ2 < 1, phase IV has the lower energy, but the conditions
m2 > 0 and ϕ2 > 0 then entail
q 2 θm θφ
<λ< 2 ⇒ q 2 |θm | > θφ > 0 . (7.375)
θφ q θm
τ
−τ < θ < − . (7.377)
q2 +1
τ θ+τ
−τ < θ < − and −1<λ<− , (7.378)
r+1 rθ
where r = q 2 .
(3) 0 > θφ > θm . In this regime, any phase is possible, however once again phase I can be ruled out since phases
II and III are of lower free energy. The condition that phase II have lower free energy than phase III is
ε0 −2 2
fII − fIII = q θφ − q 2 θm
2
<0, (7.379)
4
i.e. |θφ | < r|θm |, which means r|θ| > |θ| − τ . If r > 1 this is true for all θ < 0, while if r < 1 phase II is lower
in energy only for |θ| < τ /(1 − r).
7.11. APPENDIX II : ADDITIONAL EXAMPLES 365
We next need to test whether phase IV has an even lower energy than the lower of phases II and III. We have
(q λ θm − q −1 θφ )2
fIV − fII = (7.380)
4(λ2 − 1)
(q θm − q −1 λ θφ )2
fIV − fIII = . (7.381)
4(λ2 − 1)
In both cases, phase IV can only be the true thermodynamic phase if λ2 < 1. We then require m2 > 0 and
ϕ2 > 0, which fixes " 2 #
q θm θφ
λ ∈ − 1 , min , 2 . (7.382)
θφ q θm
The upper limit will be the first term inside the rounded brackets if q 2 |θm | < θφ , i.e. if r|θ| < |θ| − τ . This is
impossible if r > 1, hence the upper limit is given by the second term in the rounded brackets:
θ+τ
r > 1 : λ ∈ −1, (condition for phase IV) . (7.383)
rθ
If r < 1, then the upper limit will be q 2 θm /θφ = rθ/(θ + τ ) if |θ| > τ /(1 − r), and will be θφ /q 2 θm = (θ + τ )/rθ
if |θ| < τ /(1 − r).
#
τ θ+τ
r<1, − < θ < −τ : λ ∈ − 1 , (phase IV) (7.384)
1−r rθ
#
τ rθ
r<1, θ<− : λ ∈ −1, (phase IV) . (7.385)
1−r θ+τ
Representative phase diagrams for the cases r > 1 and r < 1 are shown in fig. 7.28.
366 CHAPTER 7. MEAN FIELD THEORY OF PHASE TRANSITIONS
Figure 7.28: Phase diagram for τ = 0.5, r = 1.5 (top) and τ = 0.5, r = 0.25 (bottom). The hatched purple region is
unphysical, with a free energy unbounded from below. The blue lines denote second order transitions. The thick
red line separating phases II and III is a first order line.
Chapter 8
Nonequilibrium Phenomena
8.1 References
– H. Smith and H. H. Jensen, Transport Phenomena (Oxford, 1989)
An outstanding, thorough, and pellucid presentation of the theory of Boltzmann transport in classical and
quantum systems.
– P. L. Krapivsky, S. Redner, and E. Ben-Naim, A Kinetic View of Statistical Physics (Cambridge, 2010)
Superb, modern discussion of a broad variety of issues and models in nonequilibrium statistical physics.
– E. M. Lifshitz and L. P. Pitaevskii, Physical Kinetics (Pergamon, 1981)
Volume 10 in the famous Landau and Lifshitz Course of Theoretical Physics. Surprisingly readable, and
with many applications (some advanced).
– M. Kardar, Statistical Physics of Particles (Cambridge, 2007)
A superb modern text, with many insightful presentations of key concepts. Includes a very instructive
derivation of the Boltzmann equation starting from the BBGKY hierarchy.
367
368 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
XN N
X XN
p2i
ĤN = + v(xi ) + u(xi − xj ), (8.2)
i=1
2m i=1 i<j
N
X
f (r, p; t) = δ xi (t) − r) δ(pi (t) − p
i=1
Z Y (8.4)
N
=N ddxi ddpi f (r, x2 , . . . , xN ; p, p2 , . . . , pN ; t) .
i=2
In this chapter, we will drop the 1/~ normalization for phase space integration. Thus, f (r, p, t) has dimensions of
h−d , and f (r, p, t) d3r d3p is the average number of particles found within d3r of r and d3p of p at time t.
In the GCE, we sum the RHS above over N . Assuming v = 0 so that there is no one-body potential to break
translational symmetry, the equilibrium distribution is time-independent and space-independent:
2
f 0 (r, p) = n (2πmkB T )−3/2 e−p /2mkB T
, (8.5)
where n = N/V or n = n(T, µ) is the particle density in the OCE or GCE. From the one-body distribution we can
8.3. BOLTZMANN TRANSPORT THEORY 369
compute things like the particle current, j, and the energy current, jε :
Z
p
j(r, t) = ddp f (r, p; t) (8.6)
m
Z
p
jε (r, t) = ddp f (r, p; t) ε(p) , (8.7)
m
where ε(p) = p2 /2m. Clearly these currents both vanish in equilibrium, when f = f 0 , since f 0 (r, p) depends
only on p2 and not on the direction of p. In a steady state nonequilibrium situation, the above quantities are
time-independent.
Thermodynamics says that
dq = T ds = dε − µ dn , (8.8)
where s, ε, and n are entropy density, energy density, and particle density, respectively, and dq is the differential
heat density. This relation may be case as one among the corresponding current densities:
jq = T js = jε − µ j . (8.9)
Thus, in a system with no particle flow, j = 0 and the heat current jq is the same as the energy current jε .
When the individual particles are not point particles, they possess angular momentum as well as linear momen-
tum. Following Lifshitz and Pitaevskii, we abbreviate Γ = (p, L) for these two variables for the case of diatomic
molecules, and Γ = (p, L, n̂ · L) in the case of spherical top molecules, where n̂ is the symmetry axis of the top.
We then have, in d = 3 dimensions,
3
d p point particles
dΓ = d3p L dL dΩL diatomic molecules (8.10)
3 2
d p L dL dΩL d cos ϑ symmetric tops ,
where ϑ = cos−1 (n̂ · L̂). We will call the set Γ the ‘kinematic variables’. The instantaneous number density at r is
then Z
n(r, t) = dΓ f (r, Γ ; t) . (8.11)
One might ask why we do not also keep track of the angular orientation of the individual molecules. There are
two reasons. First, the rotations of the molecules are generally extremely rapid, so we are justified in averaging
over these motions. Second, the orientation of, say, a rotor does not enter into its energy. While the same can be
said of the spatial position in the absence of external fields, (i) in the presence of external fields one must keep
track of the position coordinate r since there is physical transport of particles from one region of space to another,
and (iii) the collision process, which as we shall see enters the dynamics of the distribution function, takes place
in real space.
We now ask how the distribution functions f (r, p, t) evolves in time. It is clear that in the absence of collisions,
the distribution function must satisfy the continuity equation,
∂f
+ ∇·(uf ) = 0 . (8.13)
∂t
This is just the condition of number conservation for particles. Take care to note that ∇ and u are six-dimensional
phase space vectors:
The continuity equation describes a distribution in which each constituent particle evolves according to a pre-
scribed dynamics, which for a mechanical system is specified by
dr ∂H dp ∂H
= = v(p) , =− = Fext , (8.16)
dt ∂p dt ∂r
For example, if the particles are under the influence of gravity, then Uext (r) = mg · r and F = −∇Uext = −mg.
Note that as a consequence of the dynamics, we have ∇·u = 0, i.e. phase space flow is incompressible, provided
that ε(p) is a function of p alone, and not of r. Thus, in the absence of collisions, we have
∂f
+ u ·∇f = 0 . (8.18)
∂t
The differential operator Dt ≡ ∂t + u ·∇ is sometimes called the ‘convective derivative’, because Dt f is the time
derivative of f in a comoving frame of reference.
Next we must consider the effect of collisions, which are not accounted for by the semiclassical dynamics. In a
collision process, a particle with momentum p and one with momentum p̃ can instantaneously convert into a pair
with momenta p′ and p̃′ , provided total momentum is conserved: p + p̃ = p′ + p̃′ . This means that Dt f 6= 0.
Rather, we should write
∂f ∂f ∂f ∂f
+ ṙ · + ṗ · = (8.19)
∂t ∂r ∂p ∂t coll
where the right side is known as the collision integral. The collision integral is in general a function of r, p, and t
and a functional of the distribution f .
After a trivial rearrangement of terms, we can write the Boltzmann equation as
∂f ∂f ∂f
= + , (8.20)
∂t ∂t str ∂t coll
where
∂f ∂f ∂f
≡ −ṙ · − ṗ · (8.21)
∂t str ∂r ∂p
is known as the streaming term. Thus, there are two contributions to ∂f /∂t : streaming and collisions.
8.3. BOLTZMANN TRANSPORT THEORY 371
∂f ∂ε ∂f ∂f
+ · − ∇Uext · =0. (8.22)
∂t ∂p ∂r ∂p
In order to gain some intuition about how the streaming term affects the evolution of the distribution f (r, p, t),
consider a case where Fext = 0. We then have
∂f p ∂f
+ · =0. (8.23)
∂t m ∂r
Clearly, then, any function of the form
f (r, p, t) = ϕ r − v(p) t , p (8.24)
∂ε
will be a solution to the collisionless Boltzmann equation, where v(p) = ∂p . One possible solution would be the
Boltzmann distribution,
2
f (r, p, t) = eµ/kB T e−p /2mkB T , (8.25)
Consider the one-dimensional version, and rescale position, momentum, and time so that
1 2 1 2
f (x, p, t) = A e− 2 (x̄−p̄ t̄) e− 2 p̄ . (8.27)
1 2
Consider the level sets of f , where f (x, p, t) = A e− 2 α . The equation for these sets is
p
x̄ = p̄ t̄ ± α2 − p̄2 . (8.28)
For fixed t̄, these level sets describe the loci in phase space of equal probability densities, with the probability
density decreasing exponentially in the parameter α2 . For t̄ = 0, the initial distribution describes a Gaussian
cloud of particles with a Gaussian momentum distribution. As t̄ increases, the distribution widens in x̄ but not
in p̄ – each particle moves with a constant momentum, so the set of momentum values never changes. However,
the level sets in the (x̄ , p̄) plane become elliptical, with a semimajor axis oriented at an angle θ = ctn −1 (t) with
respect to the x̄ axis. For t̄ > 0, he particles at the outer edges of the cloud are more likely to be moving away from
the center. See the sketches in fig. 8.1
Suppose we add in a constant external force Fext . Then it is easy to show (and left as an exercise to the reader to
prove) that any function of the form
p t Fext t2 F t
f (r, p, t) = A ϕ r − + , p − ext (8.29)
m 2m m
1 2 1 2 1 2
Figure 8.1: Level sets for a sample f (x̄, p̄, t̄) = A e− 2 (x̄−p̄t̄) e− 2 p̄ , for values f = A e− 2 α with α in equally spaced
intervals from α = 0.2 (red) to α = 1.2 (blue). The time variable t̄ is taken to be t̄ = 0.0 (upper left), 0.2 (upper
right), 0.8 (lower right), and 1.3 (lower left).
Consider a function A(r, p) of position and momentum. Its average value at time t is
Z
A(t) = d3r d3p A(r, p) f (r, p, t) . (8.30)
We therefore have that the rate of change of A is determined wholly by the collision integral
Z
dA 3 3 ∂f
= d r d p A(r, p) . (8.33)
dt ∂t coll
Quantities which are then conserved in the collisions satisfy Ȧ = 0. Such quantities are called collisional invariants.
Examples of collisional invariants include the particle number (A = 1), the components of the total momentum
(A = pµ ) (in the absence of broken translational invariance, due e.g. to the presence of walls), and the total energy
(A = ε(p)).
What sort of processes contribute to the collision integral? There are two broad classes to consider. The first
involves potential scattering, where a particle in state |Γ i scatters, in the presence of an external potential, to a
state |Γ ′ i. Recall that Γ is an abbreviation for the set of kinematic variables, e.g. Γ = (p, L) in the case of a diatomic
molecule. For point particles, Γ = (px , py , pz ) and dΓ = d3p.
We now define the function w Γ ′ |Γ such that
(
′
′ rate at which a particle within dΓ of (r, Γ )
w Γ |Γ f (r, Γ ; t) dΓ dΓ = (8.34)
scatters to within dΓ ′ of (r, Γ ′ ) at time t.
The units of w dΓ are therefore 1/T . The differential scattering cross section for particle scattering is then
w Γ ′ |Γ
dσ = dΓ ′ , (8.35)
n |v|
where
X
f2 (r, p ; r ′ , p′ ; t) = δ xi (t) − r) δ(pi (t) − p δ xj (t) − r ′ ) δ(pj (t) − p′ (8.37)
i,j
is the nonequilibrium two-particle distribution for point particles. The differential scattering cross section is
w Γ ′ Γ1′ | Γ Γ1
dσ = dΓ ′ dΓ1′ . (8.38)
|v − v1 |
We assume, in both cases, that any scattering occurs locally, i.e. the particles attain their asymptotic kinematic states
on distance scales small compared to the mean interparticle separation. In this case we can treat each scattering
process independently. This assumption is particular to rarefied systems, i.e. gases, and is not appropriate for
dense liquids. The two types of scattering processes are depicted in fig. 8.2.
374 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
Figure 8.2: Left: single particle scattering process |Γ i → |Γ ′ i. Right: two-particle scattering process |Γ Γ1 i →
|Γ ′ Γ1′ i.
In computing the collision integral for the state |r, Γ i, we must take care to sum over contributions from transitions
out of this state, i.e. |Γ i → |Γ ′ i, which reduce f (r, Γ ), and transitions into this state, i.e. |Γ ′ i → |Γ i, which increase
f (r, Γ ). Thus, for one-body scattering, we have
Z n o
D ∂f
f (r, Γ ; t) = = dΓ ′ w(Γ | Γ ′ ) f (r, Γ ′ ; t) − w(Γ ′ | Γ ) f (r, Γ ; t) . (8.39)
Dt ∂t coll
For two-body scattering, we have
D ∂f
f (r, Γ ; t) =
Dt ∂t coll
Z Z Z n (8.40)
= dΓ1 dΓ ′ dΓ1′ w Γ Γ1 | Γ ′ Γ1′ f2 (r, Γ ′ ; r, Γ1′ ; t)
o
− w Γ ′ Γ1′ | Γ Γ1 f2 (r, Γ ; r, Γ1 ; t) .
Unlike the one-body scattering case, the kinetic equation for two-body scattering does not close, since the LHS
involves the one-body distribution f ≡ f1 and the RHS involves the two-body distribution f2 . To close the
equations, we make the approximation
f2 (r, Γ ′ ; r̃, Γ̃ ; t) ≈ f (r, Γ ; t) f (r̃, Γ̃ ; t) . (8.41)
We then have
Z Z Z n
D ′
f (r, Γ ; t) = dΓ1 dΓ dΓ1′ w Γ Γ1 | Γ ′ Γ1′ f (r, Γ ′ ; t) f (r, Γ1′ ; t)
Dt (8.42)
o
′
−w Γ Γ1′ | Γ Γ1 f (r, Γ ; t) f (r, Γ1 ; t) .
This is because the time reverse of the process |Γ Γ1 i → |Γ ′ Γ1′ i is |Γ ′ T Γ1′ T i → |Γ T Γ1T i.
In equilibrium, we must have
w Γ ′ Γ1′ | Γ Γ1 f 0 (Γ ) f 0 (Γ1 ) d4Γ = w Γ T Γ1T | Γ ′ T Γ1′ T f 0 (Γ ′ T ) f 0 (Γ1′ T ) d4Γ T (8.44)
where
d4Γ ≡ dΓ dΓ1 dΓ ′ dΓ1′ , d4Γ T ≡ dΓ T dΓ1T dΓ ′ T dΓ1′ T . (8.45)
Since dΓ = dΓ etc., we may cancel the differentials above, and after invoking eqn. 8.43 and suppressing the
T
f 0 (Γ ) = A e−ε/kB T , (8.47)
where A is a constant and where ε = ε(Γ ) is the kinetic energy, e.g. ε(Γ ) = p2 /2m in the case of point particles.
Note that ε(Γ T ) = ε(Γ ). Detailed balance is satisfied because the kinematics of the collision requires energy
conservation:
ε + ε1 = ε′ + ε′1 . (8.48)
Since momentum is also kinematically conserved, i.e.
p + p1 = p′ + p′1 , (8.49)
Under this condition, the collision term vanishes for f = f 0 , which is the equilibrium distribution.
376 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
8.3.7 H-theorem
Let’s consider the Boltzmann equation with two particle collisions. We define the local (i.e. r-dependent) quantity
Z
ρϕ (r, t) ≡ dΓ ϕ(Γ, f ) f (Γ, r, t) . (8.57)
At this point, ϕ(Γ, f ) is arbitrary. Note that the ϕ(Γ, f ) factor has r and t dependence through its dependence on
f , which itself is a function of r, Γ , and t. We now compute
Z Z
∂ρϕ ∂(ϕf ) ∂(ϕf ) ∂f
= dΓ = dΓ
∂t ∂t ∂f ∂t
Z Z
∂(ϕf ) ∂f
= − dΓ u · ∇(ϕf ) − dΓ (8.58)
∂f ∂t coll
I Z
∂(ϕf ) ∂f
= − dΣ n̂ · (u ϕf ) − dΓ .
∂f ∂t coll
The first term on the last line follows from the divergence theorem, and vanishes if we assume f = 0 for infinite
values of the kinematic variables, which is the only physical possibility. Thus, the rate of change of ρϕ is entirely
due to the collision term. Thus,
Z Z Z Z n o
∂ρϕ
= dΓ dΓ1 dΓ dΓ1′ w Γ ′ Γ1′ | Γ Γ1 f f1 χ − w Γ Γ1 | Γ ′ Γ1′ f ′ f1′ χ
′
∂t
Z Z Z Z (8.59)
= dΓ dΓ1 dΓ ′ dΓ1′ w Γ ′ Γ1′ | Γ Γ1 f f1 (χ − χ′ ) ,
This shows that ρϕ is preserved by the collision term if χ(Γ ) is a collisional invariant.
Now let us consider ϕ(f ) = ln f . We define h ≡ ρϕ=ln f . We then have
Z Z Z Z
∂h
= − 21 dΓ dΓ1 dΓ ′ dΓ1′ w f ′ f1′ · x ln x , (8.63)
∂t
where w ≡ w Γ ′ Γ1′ | Γ Γ1 and x ≡ f f1 /f ′ f1′ . We next invoke the result
Z Z Z Z
dΓ ′ dΓ1′ w Γ ′ Γ1′ | Γ Γ1 = dΓ ′ dΓ1′ w Γ Γ1 | Γ ′ Γ1′ (8.64)
which is a statement of unitarity of the scattering matrix3 . Multiplying both sides by f (Γ ) f (Γ1 ), then integrating
over Γ and Γ1 , and finally changing variables (Γ, Γ1 ) ↔ (Γ ′ , Γ1′ ), we find
Z Z Z Z Z Z Z Z
0= dΓ dΓ1 dΓ ′ dΓ1′ w f f1 − f ′ f1′ = dΓ dΓ1 dΓ ′ dΓ1′ w f ′ f1′ (x − 1) . (8.65)
1
Multiplying this result by 2 and adding it to the previous equation for ḣ, we arrive at our final result,
Z Z Z Z
∂h 1 ′
= − 2 dΓ dΓ1 dΓ dΓ1′ w f ′ f1′ (x ln x − x + 1) . (8.66)
∂t
Note that w, f ′ , and f1′ are all nonnegative. It is then easy to prove that the function g(x) = x ln x − x + 1 is
nonnegative for all positive x values4 , which therefore entails the important result
∂h(r, t)
≤0. (8.67)
∂t
R
Boltzmann’s H function is the space integral of the h density: H = d3r h.
Thus, everywhere in space, the function h(r, t) is monotonically decreasing or constant, due to collisions. In
equilibrium, ḣ = 0 everywhere, which requires x = 1, i.e.
But this means that ln f 0 is itself a collisional invariant, and if 1, p, and ε are the only collisional invariants, then
ln f 0 must be expressible in terms of them. Thus,
µ V ·p ε
ln f 0 = + − , (8.70)
kB T kB T kB T
where µ, V , and T are constants which parameterize the equilibrium distribution f 0 (p), corresponding to the
chemical potential, flow velocity, and temperature, respectively.
3 See
Lifshitz and Pitaevskii, Physical Kinetics, §2.
function g(x) = x ln x − x + 1 satisfies g ′ (x) = ln x, hence g ′ (x) < 0 on the interval x ∈ [0, 1) and g ′ (x) > 0 on x ∈ (1, ∞]. Thus,
4 The
g(x) monotonically decreases from g(0) = 1 to g(1) = 0, and then monotonically increases to g(∞) = ∞, never becoming negative.
378 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
Consider a gas which is only weakly out of equilibrium. We follow the treatment in Lifshitz and Pitaevskii, §6. As
the gas is only slightly out of equilibrium, we seek a solution to the Boltzmann equation of the form f = f 0 + δf ,
where f 0 is describes a local equilibrium. Recall that such a distribution function is annihilated by the collision
term in the Boltzmann equation but not by the streaming term, hence a correction δf must be added in order to
obtain a solution.
The most general form of local equilibrium is described by the distribution
0 µ − ε(Γ ) + V · p
f (r, Γ ) = C exp , (8.71)
kB T
where µ = µ(r, t), T = T (r, t), and V = V (r, t) vary in both space and time. Note that
!
0 dT ∂f 0
df = dµ + p · dV + (ε − µ − V · p) − dε −
T ∂ε
! (8.72)
1 dT ∂f 0
= dp + p · dV + (ε − h) − dε −
n T ∂ε
∂f 0 f0
− = . (8.74)
∂ε kB T
The RHS of this equation must be of order δf because the local equilibrium distribution f 0 is annihilated by the
collision integral. We therefore wish to evaluate one of the contributions to the LHS of this equation,
( h i
∂f 0 p ∂f 0 ∂f 0 ∂f 0 1 ∂p ε − h ∂T
+ · +F · = − + + mv· (v·∇) V
∂t m ∂r ∂p ∂ε n ∂t T ∂t
) (8.76)
∂V 1 ε−h
+v· m + ∇p + v · ∇T − F · v .
∂t n T
5 Inthe chapter on thermodynamics, we adopted a slightly different definition of cp as the heat capacity per mole. In this chapter cp is the
heat capacity per particle.
8.4. WEAKLY INHOMOGENEOUS GAS 379
To simplify this, first note that Newton’s laws applied to an ideal fluid give ρV̇ = −∇p, where ρ = mn is the mass
density. Corrections to this result, e.g. viscosity and nonlinearity in V , are of higher order.
Next, continuity for particle number means ṅ + ∇ · (nV ) = 0. We assume V is zero on average and that all
derivatives are small, hence ∇·(nV ) = V ·∇n + n ∇·V ≈ n ∇·V . Thus,
∂ ln n ∂ ln p ∂ ln T
= − = −∇·V , (8.77)
∂t ∂t ∂t
where we have invoked the ideal gas law n = p/kB T above.
Next, we invoke conservation of entropy. If s is the entropy per particle, then ns is the entropy per unit volume,
in which case we have the continuity equation
∂(ns) ∂s ∂n
+ ∇ · (nsV ) = n + V ·∇s + s + ∇ · (nV ) = 0 . (8.78)
∂t ∂t ∂t
The second bracketed term on the RHS vanishes because of particle continuity, leaving us with ṡ + V ·∇s ≈ ṡ = 0
(since V = 0 on average, and any gradient is first order in smallness). Now thermodynamics says
∂s ∂s
ds = dT + dp
∂T p ∂p T
(8.79)
cp kB
= dT − dp ,
T p
∂s
∂s
∂v
since T ∂T p
= cp and ∂p T
= ∂T p
, where v = V /N . Thus,
cp ∂ ln T ∂ ln p
− =0. (8.80)
kB ∂t ∂t
∂ ln T ∂ ln p
We now have in eqns. 8.77 and 8.80 two equations in the two unknowns ∂t and ∂t , yielding
∂ ln T k
= − B ∇·V (8.81)
∂t cV
∂ ln p cp
=− ∇·V . (8.82)
∂t cV
where
1 ∂Vα ∂Vβ
Qαβ = + . (8.84)
2 ∂xβ ∂xα
∂ δf ∂ δf
Notice we have dropped the terms v · ∂r and F · ∂p , since δf must already be first order in smallness, and both
the ∂
∂roperator as well as F add a second order of smallness, which is negligible. Typically ∂∂tδf is nonzero if
the applied force F (t) is time-dependent. We use the convention of summing over repeated indices. Note that
δαβ Qαβ = Qαα = ∇·V . For ideal gases in which only translational and rotational degrees of freedom are excited,
h = cp T .
This model is known as the relaxation time approximation. Here, f 0 = f 0 (r, p, t) is a distribution function which
describes a local equilibrium at each position r and time t. The quantity τ is the relaxation time, which can in
principle be momentum-dependent, but which we shall first consider to be constant. In the absence of streaming
terms, we have
∂ δf δf
=− =⇒ δf (r, p, t) = δf (r, p, 0) e−t/τ . (8.87)
∂t τ
The distribution f then relaxes to the equilibrium distribution f 0 on a time scale τ . We note that this approximation
is obviously flawed in that all quantities – even the collisional invariants – relax to their equilibrium values on the
scale τ . In the Appendix, we consider a model for the collision integral in which the collisional invariants are all
preserved, but everything else relaxes to local equilibrium at a single rate.
Consider two particles with velocities v and v ′ . The average of their relative speed is
Z Z
h |v − v | i = d v d3v ′ P (v) P (v ′ ) |v − v ′ | ,
′ 3
(8.88)
which follows from the Boltzmann form of the equilibrium distribution f 0 (p). It is left as an exercise for the
student to verify that
1/2
′ 4 kB T
v̄rel ≡ h |v − v | i = √ . (8.90)
π m
√
Note that v̄rel = 2 v̄, where v̄ is the average particle speed. Let σ be the total scattering cross section, which for
hard spheres is σ = πd2 , where d is the hard sphere diameter. Then the rate at which particles scatter is
1
= n v̄rel σ . (8.91)
τ
8.5. RELAXATION TIME APPROXIMATION 381
Figure 8.3: Graphic representation of the equation n σ v̄rel τ = 1, which yields the scattering time τ in terms of the
number density n, average particle pair relative velocity v̄rel , and two-particle total scattering cross section σ. The
equation says that on average there must be one particle within the tube.
We consider a system with a temperature gradient ∇T and seek a steady state (i.e. time-independent) solution
to the Boltzmann equation. We assume Fα = Qαβ = 0. Appealing to eqn. 8.85, and using the relaxation time
approximation for the collision integral, we have
τ (ε − cp T )
δf = − (v · ∇T ) f 0 . (8.94)
kB T 2
We are now ready to compute the energy and particle currents. In order to compute the local density of any quantity
A(r, p), we multiply by the distribution f (r, p) and integrate over momentum:
Z
ρA (r, t) = d3p A(r, p) f (r, p, t) , (8.95)
R
For the energy (thermal) current, we let A = ε vα = ε pα /m, in which case ρA = jα . Note that d3p p f 0 = 0 since f 0
is isotropic in p even when µ and T depend on r. Thus, only δf enters into the calculation of the various currents.
Thus, the energy (thermal) current is
Z
jεα (r) = d3p ε v α δf
(8.96)
nτ
α β ∂T
=− v v ε (ε − cp T ) ,
kB T 2 ∂xβ
where the repeated index β is summed over, and where momentum averages are defined relative to the equilib-
rium distribution, i.e. Z Z Z
h φ(p) i = d3p φ(p) f 0 (p) d3p f 0 (p) = d3v P (v) φ(mv) . (8.97)
382 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
where 3/2
m 2
P (v) = e−m(v−V ) /2kB T
(8.99)
2πkB T
is the Maxwell velocity distribution.
Note that if φ = φ(ε) is a function of the energy, and if V = 0, then
where
Pe (ε) = √2 (k T )−3/2 ε1/2 e−ε/kB T
π B
, (8.101)
R∞
is the Maxwellian distribution of single particle energies. This distribution is normalized with dε Pe(ε) = 1.
0
Averages with respect to this distribution are given by
Z∞ Z∞
h φ(ε) i = dε φ(ε) Pe(ε) = √2 (k T )−3/2
π B dε ε1/2 φ(ε) e−ε/kB T . (8.102)
0 0
8.5.4 Viscosity
Consider the situation depicted in fig. 8.4. A fluid filling the space between two large flat plates at z = 0 and
z = d is set in motion by a force F = F x̂ applied to the upper plate; the lower plate is fixed. It is assumed that the
fluid’s velocity locally matches that of the plates. Fluid particles at the top have an average x-component of their
momentum hpx i = mV . As these particles move downward toward lower z values, they bring their x-momenta
with them. Therefore there is a downward (−ẑ-directed) flow of hpx i. Since x-momentum is constantly being
drawn away from z = d plane, this means that there is a −x-directed viscous drag on the upper plate. The viscous
drag force per unit area is given by Fdrag /A = −ηV /d, where V /d = ∂Vx /∂z is the velocity gradient and η is the
shear viscosity. In steady state, the applied force balances the drag force, i.e. F + Fdrag = 0. Clearly in the steady
state the net momentum density of the fluid does not change, and is given by 12 ρV x̂, where ρ is the fluid mass
density. The momentum per unit time injected into the fluid by the upper plate at z = d is then extracted by the
8.5. RELAXATION TIME APPROXIMATION 383
Figure 8.4: Gedankenexperiment to measure shear viscosity η in a fluid. The lower plate is fixed. The viscous drag
force per unit area on the upper plate is Fdrag /A = −ηV /d. This must be balanced by an applied force F .
lower plate at z = 0. The momentum flux density Πxz = n h px vz i is the drag force on the upper surface per unit
∂V
area: Πxz = −η ∂zx . The units of viscosity are [η] = M/LT .
We now provide some formal definitions of viscosity. As we shall see presently, there is in fact a second type of
viscosity, called second viscosity or bulk viscosity, which is measurable although not by the type of experiment
depicted in fig. 8.4.
The momentum flux tensor Παβ = n h pα vβ i is defined to be the current of momentum component pα in the
direction of increasing xβ . For a gas in motion with average velocity V, we have
The coefficient of the traceless component is η, known as the shear viscosity. The coefficient of the component
proportional to the identity is ζ, known as the bulk viscosity. The full stress tensor σαβ contains a contribution from
the pressure:
σαβ = −p δαβ + σ̃αβ . (8.109)
The differential force dFα that a fluid exerts on on a surface element n̂ dA is
dFα = −σαβ nβ dA , (8.110)
384 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
Figure 8.5: Left: thermal conductivity (λ in figure) of Ar between T = 800 K and T = 2600 K. The best fit to a
single power law λ = aT b results in b = 0.651. Source: G. S. Springer and E. W. Wingeier, J. Chem Phys. 59, 1747
(1972). Right: log-log plot of shear viscosity (µ in figure) of He between T ≈ 15 K and T ≈ 1000 K. The red line
has slope 21 . The slope of the data is approximately 0.633. Source: J. Kestin and W. Leidenfrost, Physica 25, 537
(1959).
where we are using the Einstein summation convention and summing over the repeated index β. We will now
compute the shear viscosity η using the Boltzmann equation in the relaxation time approximation.
Appealing again to eqn. 8.85, with F = 0 and h = cp T , we find
( )
τ ε − cp T ε
δf = − m vα vβ Qαβ + v · ∇T − ∇·V f0 . (8.111)
kB T T cV /kB
nm2 τ
=− Qαβ h vx vz vα vβ i
kB T
(8.112)
nτ ∂Vx ∂Vz
=− + h mvx2 · mvz2 i
kB T ∂z ∂x
∂Vz ∂Vx
= −nτ kB T + .
∂x ∂z
Why do both κ(T ) as well as η(T ) decrease at low temperatures? The reason is that the heat current which flows
in response to ∇T as well as the momentum current which flows in response to ∂Vx /∂z are due to the presence of
collisions, which result in momentum and energy transfer between particles. This is true even when total energy
and momentum are conserved, which they are not in the relaxation time approximation. Intuitively, we might
think that the viscosity should increase as the temperature is lowered, since common experience tells us that fluids
‘gum up’ as they get colder – think of honey as an extreme example. But of course honey is nothing like an
ideal gas, and the physics behind the crystallization or glass transition which occurs in real fluids when they get
sufficiently cold is completely absent from our approach. In our calculation, viscosity results from collisions, and
with no collisions there is no momentum transfer and hence no viscosity. If, for example, the gas particles were
to simply pass through each other, as though they were ghosts, then there would be no opposition to maintaining
an arbitrary velocity gradient.
Suppose a uniform oscillating external force Fext (t) = F e−iωt is applied. For a system of charged particles, this
force would arise from an external electric field Fext = qE e−iωt , where q is the charge of each particle. We’ll
assume ∇T = 0. The Boltzmann equation is then written
∂f p ∂f ∂f f − f0
+ · + F e−iωt · =− . (8.115)
∂t m ∂r ∂p τ
We again write f = f 0 + δf , and we assume δf is spatially constant. Thus,
∂ δf ∂f 0 δf
+ F e−iωt · v =− . (8.116)
∂t ∂ε τ
If we assume δf (t) = δf (ω) e−iωt then the above differential equation is converted to an algebraic equation, with
solution
τ e−iωt ∂f 0
δf (t) = − F ·v . (8.117)
1 − iωτ ∂ε
We now compute the particle current:
Z
jα (r, t) = d3p v δf
Z
τ e−iωt Fβ
= · d3p f 0 (p) vα vβ
1 − iωτ kB T
Z (8.118)
τ e−iωt nFα
= · d3v P (v) v 2
1 − iωτ 3kB T
nτ Fα e−iωt
= · .
m 1 − iωτ
If the particles are electrons, with charge q = −e, then the electrical current is (−e) times the particle current. We
then obtain
ne2 τ Eα e−iωt
jα(elec) (t) = · ≡ σαβ (ω) Eβ e−iωt , (8.119)
m 1 − iωτ
where
ne2 τ 1
σαβ (ω) = · δ (8.120)
m 1 − iωτ αβ
is the frequency-dependent electrical conductivity tensor. Of course for fermions such as electrons, we should be
using the Fermi distribution in place of the Maxwell-Boltzmann distribution for f 0 (p). This affects the relation
between n and µ only, and the final result for the conductivity tensor σαβ (ω) is unchanged.
386 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
Suppose we have some averaged intensive quantity φ which is spatially dependent through T (r) or µ(r) or V (r).
For simplicity we will write φ = φ(z). We wish to compute the current of φ across some surface whose equation
is dz = 0. If the mean free path is ℓ, then the value of φ for particles crossing this surface in the +ẑ direction is
φ(z − ℓ cos θ), where θ is the angle the particle’s velocity makes with respect to ẑ, i.e. cos θ = vz /v. We perform the
same analysis for particles moving in the −ẑ direction, for which φ = φ(z + ℓ cos θ). The current of φ through this
surface is then
Z Z
jφ = nẑ d3v P (v) vz φ(z − ℓ cos θ) + nẑ d3v P (v) vz φ(z + ℓ cos θ)
vz >0 vz <0
Z (8.121)
∂φ vz2 ∂φ
= −nℓ ẑ d3v P (v) = − 31 nv̄ℓ ẑ ,
∂z v ∂z
q
where v̄ = 8kπmBT
is the average particle speed. If the z-dependence of φ comes through the dependence of φ on
the local temperature T , then we have
∂φ
jφ = − 31 nℓv̄ ∇T ≡ −K ∇T , (8.122)
∂T
where
∂φ
K = 31 nℓv̄ (8.123)
∂T
∂φ
is the transport coefficient. If φ = hεi, then ∂T = cp , where cp is the heat capacity per particle at constant pressure.
We then find jε = −κ ∇T with thermal conductivity
κ = 13 nℓv̄ cp . (8.124)
π
Our Boltzmann equation calculation yielded the same result, but with a prefactor of 8 instead of 31 .
We can make a similar argument for the viscosity. In this case φ = hpx i is spatially varying through its dependence
on the flow velocity V (r). Clearly ∂φ/∂Vx = m, hence
∂Vx
jpzx = Πxz = − 13 nmℓv̄ , (8.125)
∂z
from which we identify the viscosity, η = 31 nmℓv̄. Once again, this agrees in its functional dependences with the
Boltzmann equation calculation in the relaxation time approximation. Only the coefficients differ. The ratio of the
8
coefficients is KQDC/KBRT = 3π = 0.849 in both cases6 .
Suppose, under conditions of constant pressure, we add heat q per unit volume to an ideal gas. We know from
thermodynamics that its temperature will then increase by an amount ∆T = q/ncp . If a heat current jq flows, then
the continuity equation for energy flow requires
∂T
ncp + ∇ · jq = 0 . (8.126)
∂t
6 Here we abbreviate QDC for ‘quick and dirty calculation’ and BRT for ‘Boltzmann equation in the relaxation time approximation’.
8.6. DIFFUSION AND THE LORENTZ MODEL 387
Table 8.1: Viscosities, thermal conductivities, and Prandtl numbers for some common gases at T = 293 K and
p = 1 atm. (Source: Table 1.1 of Smith and Jensen, with data for triatomic gases added.)
In a system where there is no net particle current, the heat current jq is the same as the energy current jε , and
since jε = −κ ∇T , we obtain a diffusion equation for temperature,
∂T κ
= ∇2 T . (8.127)
∂t ncp
The combination
κ
a≡ (8.128)
ncp
is known as the thermal diffusivity. Our Boltzmann equation calculation in the relaxation time approximation
yielded the result κ = nkB T τ cp /m. Thus, we find a = kB T τ /m via this method. Note that the dimensions of a are
the same as for any diffusion constant D, namely [a] = L2 /T .
Another quantity with dimensions of L2 /T is the kinematic viscosity, ν = η/ρ, where ρ = nm is the mass density.
We found η = nkB T τ from the relaxation time approximation calculation, hence ν = kB T τ /m. The ratio ν/a,
called the Prandtl number, Pr = ηcp /mκ, is dimensionless. According to our calculations, Pr = 1. According to
table 8.1, most monatomic gases have Pr ≈ 32 .
As we remarked above, the relaxation time approximation fails to conserve any of the collisional invariants. It is
therefore unsuitable for describing hydrodynamic phenomena such as diffusion. To see this, let f (r, v, t) be the
distribution function, here written in terms of position, velocity, and time rather than position, momentum, and
time as befor7 . In the absence of external forces, the Boltzmann equation in the relaxation time approximation is
∂f ∂f f − f0
+v· =− . (8.129)
∂t ∂r τ
The density of particles in velocity space is given by
Z
ñ(v, t) = d3r f (r, v, t) . (8.130)
∂ ñ ñ − ñ0
=− . (8.131)
∂t τ
Thus, with δñ(v, t) = ñ(v, t) − ñ0 (v), we have
Thus, ñ(v, t) decays exponentially to zero with time constant τ , from which it follows that the total particle number
exponentially relaxes to N0 . This is physically incorrect; local density perturbations can’t just vanish. Rather, they
diffuse.
τ −1 f (k, v, t = 0)
fˆ(k, v, s) = P fˆ(k, v, s) + , (8.136)
s + iv · k + τ −1 s + iv · k + τ −1
which entails Z Z
dv̂ τ −1 dv̂ f (k, v, t = 0)
P fˆ(k, v, s) = P fˆ(k, v, s) + . (8.137)
4π s + iv · k + τ −1 4π s + iv · k + τ −1
Now we have
Z Z1
dv̂ τ −1 τ −1
= dx
4π s + iv · k + τ −1 s + ivkx + τ −1
−1 (8.138)
1 vkτ
= tan−1 .
vk 1 + τs
8 See the excellent discussion in the book by Krapivsky, Redner, and Ben-Naim, cited in §8.1.
8.6. DIFFUSION AND THE LORENTZ MODEL 389
We are interested in the long time limit t ≫ τ for f (r, v, t). This is dominated by s ∼ t−1 , and we assume that τ −1
is dominant over s and iv · k. We then have
1 δ(v − v0 )
fˆ(k, v, s) ≈ 1 2 · . (8.145)
s+ 3 v0 k2 τ 4πv02
2 δ(v − v0 )
f (r, v, t) = (4πDt)−3/2 e−r /4Dt
· , (8.146)
4πv02
Note that Z
d3r n(r, t) = 1 (8.149)
We now return to the classical Boltzmann equation and consider a more formal treatment of the collision term in
the linear approximation. We will assume time-reversal symmetry, in which case
Z Z Z n o
∂f
= d3p1 d3p′ d3p′1 w(p′ , p′1 | p, p1 ) f (p′ ) f (p′1 ) − f (p) f (p1 ) . (8.150)
∂t coll
The collision integral is nonlinear in the distribution f . We linearize by writing
f (p) = f 0 (p) + f 0 (p) ψ(p) , (8.151)
where we assume ψ(p) is small. We then have, to first order in ψ,
∂f
= f 0 (p) L̂ψ + O(ψ 2 ) , (8.152)
∂t coll
where the action of the linearized collision operator is given by
Z Z Z n o
3 3 ′
L̂ψ = d p1 d p d3p′1 w(p′ , p′1 | p, p1 ) f 0 (p1 ) ψ(p′ ) + ψ(p′1 ) − ψ(p) − ψ(p1 )
Z Z n o (8.153)
∂σ 0
= d3p1 dΩ |v − v1 | f (p1 ) ψ(p′ ) + ψ(p′1 ) − ψ(p) − ψ(p1 ) ,
∂Ω
where we have invoked eqn. 8.55 to write the RHS in terms of the differential scattering cross section. In deriving
the above result, we have made use of the detailed balance relation,
f 0 (p) f 0 (p1 ) = f 0 (p′ ) f 0 (p′1 ) . (8.154)
We have also suppressed the r dependence in writing f (p), f 0 (p), and ψ(p).
From eqn. 8.85, we then have the linearized equation
∂
L̂ − ψ = Y, (8.155)
∂t
where, for point particles,
( )
1 ε(p) − cp T k ε(p)
Y = v · ∇T + m vα vβ Qαβ − B ∇·V − F · v . (8.156)
kB T T cV
∂
Eqn. 8.155 is an inhomogeneous linear equation, which can be solved by inverting the operator L̂ − ∂t .
Although L̂ is an integral operator, it shares many properties with other linear operators with which you are
familiar, such as matrices and differential operators. We can define an inner product9,
Z
h ψ1 | ψ2 i ≡ d3p f 0 (p) ψ1 (p) ψ2 (p) . (8.157)
9 The requirements of an inner product hf |gi are symmetry, linearity, and non-negative definiteness.
8.7. LINEARIZED BOLTZMANN EQUATION 391
Note that this is not the usual Hilbert space inner product from quantum mechanics, since the factor f 0 (p) is
included in the metric. This is necessary in order that L̂ be self-adjoint:
We can now define the spectrum of normalized eigenfunctions of L̂, which we write as φn (p). The eigenfunctions
satisfy the eigenvalue equation,
L̂φn = −λn φn , (8.159)
and may be chosen to be orthonormal,
h φm | φn i = δmn . (8.160)
′
Of course, in order to obtain the eigenfunctions φn we must have detailed knowledge of the function w(p , p′1 | p, p1 ).
Recall that there are five collisional invariants, which are the particle number, the three components of the total
particle momentum, and the particle energy. To each collisional invariant, there is an associated eigenfunction φn
with eigenvalue λn = 0. One can check that these normalized eigenfunctions are
1
φn (p) = √ (8.161)
n
p
φpα (p) = p α (8.162)
nmkB T
r
2 ε(p) 3
φε (p) = − . (8.163)
3n kB T 2
If there are no temperature, chemical potential, or bulk velocity gradients, and there are no external forces, then
Y = 0 and the only changes to the distribution are from collisions. The linearized Boltzmann equation becomes
∂ψ
= L̂ψ . (8.164)
∂t
We can therefore write the most general solution in the form
X′
ψ(p, t) = Cn φn (p) e−λn t , (8.165)
n
where the prime on the sum reminds us that collisional invariants are to be excluded. All the eigenvalues λn ,
aside from the five zero eigenvalues for the collisional invariants, must be positive. Any negative eigenvalue
would cause ψ(p, t) to increase without bound, and an initial nonequilibrium distribution would not relax to the
equilibrium f 0 (p), which we regard as unphysical. Henceforth we will drop the prime on the sum but remember
that Cn = 0 for the five collisional invariants.
Recall also the particle, energy, and thermal (heat) currents,
Z Z
j = d p v f (p) = d3p f 0 (p) v ψ(p) = h v | ψ i
3
Z Z
jε = d3p v ε f (p) = d3p f 0 (p) v ε ψ(p) = h v ε | ψ i (8.166)
Z Z
jq = d p v (ε − µ) f (p) = d3p f 0 (p) v (ε − µ) ψ(p) = h v (ε − µ) | ψ i .
3
Note jq = jε − µj.
392 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
Under steady state conditions, there is no time dependence, and the linearized Boltzmann equation takes the form
L̂ψ = Y . (8.167)
P
We may expand ψ in the eigenfunctions φn and write ψ = n Cn φn . Applying L̂ and taking the inner product
with φj , we have
1
Cj = − h φj | Y i . (8.168)
λj
Thus, the formal solution to the linearized Boltzmann equation is
X 1
ψ(p) = − h φn | Y i φn (p) . (8.169)
n
λn
Thermal conductivity
1 ∂T
Y = · Xκ , (8.170)
kB T 2 ∂x
where Xκ ≡ (ε − cp T ) vx . Under the conditions of no particle flow (j = 0), we have jq = −κ ∂x T x̂. Then we have
∂T
h Xκ | ψ i = −κ . (8.171)
∂x
Viscosity
Following the treatment in chapter 1 of Smith and Jensen, define Ĥ ≡ −L̂. We have that Ĥ is a positive semidef-
inite operator, whose only zero eigenvalues correspond to the collisional invariants. We then have the Schwarz
inequality,
2
h ψ | Ĥ | ψ i · h φ | Ĥ | φ i ≥ h φ | Ĥ | ψ i , (8.175)
8.7. LINEARIZED BOLTZMANN EQUATION 393
for any two Hilbert space vectors | ψ i and | φ i. Consider now the above calculation of the thermal conductivity.
We have
1 ∂T
Ĥψ = − X (8.176)
kB T 2 ∂x κ
and therefore
2
kB T 2 1 h φ | Xκ i
κ= 2
h ψ | Ĥ | ψ i ≥ . (8.177)
(∂T /∂x) kB T 2 h φ | Ĥ | φ i
Similarly, for the viscosity, we have
m ∂Vx
Ĥψ = − X , (8.178)
kB T ∂y η
from which we derive
2
kB T m2 h φ | X η i
η= h ψ | Ĥ | ψ i ≥ . (8.179)
(∂Vx /∂y)2 kB T h φ | Ĥ | φ i
In order to get a good lower bound, we want φ in each case to have a good overlap with Xκ,η . One approach then
is to take φ = Xκ,η , which guarantees that the overlap will be finite (and not zero due to symmetry, for example).
We illustrate this method with the viscosity calculation. We have
2
m2 h vx vy | vx vy i
η≥ . (8.180)
kB T h vx vy | Ĥ | vx vy i
Here the kinematics of the collision guarantee total energy and momentum conservation, so p′ and p′1 are deter-
mined as in eqn. 8.56.
Now we have
dΩ = sin χ dχ dϕ , (8.182)
where χ is the scattering angle depicted in Fig. 8.6 and ϕ is the azimuthal angle of the scattering. The differential
scattering cross section is obtained by elementary mechanics and is known to be
2
∂σ d(b /2)
= , (8.183)
∂Ω d sin χ
Z∞
b
χ(b, u) = π − 2 dr q , (8.184)
2U(r)r 4
rp r 4 − b2 r 2 − m̃u2
where m̃ = 21 m is the reduced mass, and rp is the relative coordinate separation at periapsis, i.e. the distance of
closest approach, which occurs when ṙ = 0, i.e.
1 2 ℓ2
2 m̃u =
2m̃rp2
+ U (rp ) , (8.185)
Figure 8.6: Scattering in the CM frame. O is the force center and P is the point of periapsis. The impact parameter
is b, and χ is the scattering angle. φ0 is the angle through which the relative coordinate moves between periapsis
and infinity.
v = V + 21 u v ′ = V + 21 u′ (8.186)
v1 = V − 1
2u v1′ =V − 1 ′
2u , (8.187)
We may write
u′ = u sin χ cos ϕ ê1 + sin χ sin ϕ ê2 + cos χ ê3 , (8.189)
Z2π
dϕ 1
2 uα uβ − u′α u′β = −π sin2 χ u2 δαβ − 3uα uβ . (8.190)
0
Z∞
R(u) ≡ db b sin2 χ(b, u) . (8.192)
0
we have
3 Z Z
m 2 2
h vx vy | L̂ | vx vy i = n 2
d V d3u e−mV
3 /kB T
e−mu /4kB T
·u· 3π
2 ux uy · R(u) · vx vy . (8.194)
2πkB T
This yields
h vx vy | L̂ | vx vy i = π
40 n2 u5 R(u) , (8.195)
where
Z∞ Z∞
2 −mu2 /4kB T 2
F (u) ≡ du u e F (u) du u2 e−mu /4kB T . (8.196)
0 0
The computation for κ is a bit more tedious. One has ψ(p) = (ε − cp T ) vx , in which case
h i
∆(ψ) = 21 m (V · u) ux − (V · u′ ) u′x . (8.199)
Thus, independent of the potential, this variational calculation yields a Prandtl number of
ν η cp 2
Pr = = = 3 , (8.201)
a mκ
which is very close to what is observed in dilute monatomic gases (see Tab. 8.1).
While the variational expressions for η and κ are complicated functions of the potential, for hard sphere scattering
the calculation is simple, because b = d sin φ0 = d cos( 21 χ), where d is the hard sphere diameter. Thus, the impact
parameter b is independent of the relative speed u, and one finds R(u) = 31 d3 . Then
5
128 kB T 5/2 2
u R(u) = 31 d3 u5 = √ d (8.202)
π m
We now derive the equations governing fluid flow. The equations of mass and momentum balance are
∂ρ
+ ∇·(ρ V ) = 0 (8.204)
∂t
∂(ρ Vα ) ∂Παβ
+ =0, (8.205)
∂t ∂xβ
where
σ̃αβ
z
( }| ){
∂Vα ∂Vβ 2
Παβ = ρ Vα Vβ + p δαβ − η + − 3 ∇·V δαβ + ζ ∇·V δαβ . (8.206)
∂xβ ∂xα
Substituting the continuity equation into the momentum balance equation, one arrives at
∂V
ρ + ρ (V ·∇)V = −∇p + η ∇2 V + (ζ + 31 η)∇(∇·V ) , (8.207)
∂t
which, together with continuity, are known as the Navier-Stokes equations. These equations are supplemented by
an equation describing the conservation of energy,
∂s ∂Vα
T + T ∇·(sV ) = σ̃αβ + ∇·(κ∇T ) . (8.208)
∂T ∂xβ
Note that the LHS of eqn. 8.207 is ρ DV /Dt, where D/Dt is the convective derivative. Multiplying by a differential
volume, this gives the mass times the acceleration of a differential local fluid element. The RHS, multiplied by
the same differential volume, gives the differential force on this fluid element in a frame instantaneously moving
with constant velocity V . Thus, this is Newton’s Second Law for the fluid.
Almost everything we have derived thus far can be applied, mutatis mutandis, to quantum systems. The main
difference is that the distribution f 0 corresponding to local equilibrium is no longer of the Maxwell-Boltzmann
form, but rather of the Bose-Einstein or Fermi-Dirac form,
( )−1
0 ε(k) − µ(r, t)
f (r, k, t) = exp ∓1 , (8.209)
kB T (r, t)
where the top sign applies to bosons and the bottom sign to fermions. Here we shift to the more common notation
for quantum systems in which we write the distribution in terms of the wavevector k = p/~ rather than the
momentum p. The quantum distributions satisfy detailed balance with respect to the quantum collision integral
Z 3 Z 3 ′Z 3 ′ n o
∂f d k1 dk d k1 ′ ′ ′ ′
= w f f 1 (1 ± f ) (1 ± f 1 ) − f f 1 (1 ± f ) (1 ± f 1 ) (8.210)
∂t coll (2π)3 (2π)3 (2π)3
where w = w(k, k1 | k′ , k1′ ), f = f (k), f1 = f (k1 ), f ′ = f (k′ ), and f1′ = f (k1′ ), and where we have assumed
time-reversal and parity symmetry. Detailed balance requires
f f1 f′ f1′
· = · , (8.211)
1 ± f 1 ± f1 1 ± f 1 ± f1′
′
8.9. NONEQUILIBRIUM QUANTUM TRANSPORT 397
1 f
f= =⇒ = eβ(µ−ε) , (8.212)
eβ(ε−µ) ∓ 1 1±f
which is the Boltzmann distribution, which we have already shown to satisfy detailed balance. For the streaming
term, we have
∂f 0 ε−µ
df 0 = kB T d
∂ε kB T
0
∂f dµ (ε − µ) dT dε
= kB T − − + (8.213)
∂ε kB T kB T 2 kB T
0
∂f ∂µ ε − µ ∂T ∂ε
=− · dr + · dr − · dk ,
∂ε ∂r T ∂r ∂k
The most important application is to the theory of electron transport in metals and semiconductors, in which case
f 0 is the Fermi distribution. In this case, the quantum collision integral also receives a contribution from one-body
scattering in the presence of an external potential U (r), which is given by Fermi’s Golden Rule:
′
∂f (k) 2π X
′ 2
= | k U k | f (k′ ) − f (k) δ ε(k) − ε(k′ )
∂t coll ~ ′
k ∈ Ω̂
Z 3 (8.215)
2π dk
= 3
| Û (k − k′ )|2 f (k′ ) − f (k) δ ε(k) − ε(k′ ) .
~V (2π)
Ω̂
The wavevectors are now restricted to the first Brillouin zone, and the dispersion ε(k) is no longer the ballistic
form ε = ~2 k2 /2m but rather the dispersion for electrons in a particular energy band (typically the valence band)
of a solid10 . Note that f = f 0 satisfies detailed balance with respect to one-body collisions as well11 .
In the presence of a weak electric field E and a (not necessarily weak) magnetic field B, we have, within the
relaxation time approximation, f = f 0 + δf with
0
∂ δf e ∂ δf ε−µ ∂f δf
− v×B· − v · e E+ ∇T =− , (8.216)
∂t ~c ∂k T ∂ε τ
where E = −∇(φ − µ/e) = E − e−1 ∇µ is the gradient of the ‘electrochemical potential’ φ − e−1 µ. In deriving the
above equation, we have worked to lowest order in small quantities. This entails dropping terms like v· ∂∂δf
r (higher
order in spatial derivatives) and E · ∂∂δf (both E and δf are assumed small). Typically τ is energy-dependent, i.e.
k
τ = τ ε(k) .
10 We neglect interband scattering here, which can be important in practical applications, but which is beyond the scope of these notes.
11 The transition rate from |k′ i to |ki is proportional to the matrix element and to the product f ′ (1 − f ). The reverse process is proportional
to f (1 − f ′ ). Subtracting these factors, one obtains f ′ − f , and therefore the nonlinear terms felicitously cancel in eqn. 8.215.
398 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
We can use eqn. 8.216 to compute the electrical current j and the thermal current jq ,
Z 3
dk
j = −2e v δf (8.217)
(2π)3
Ω̂
Z
d3k
jq = 2 (ε − µ) v δf . (8.218)
(2π)3
Ω̂
Here the factor of 2 is from spin degeneracy of the electrons (we neglect Zeeman splitting).
In the presence of a time-independent temperature gradient and electric field, linearized Boltzmann equation in
the relaxation time approximation has the solution
ε−µ ∂f 0
δf = −τ (ε) v · eE + ∇T − . (8.219)
T ∂ε
We now consider both the electrical current12 j as well as the thermal current density jq . One readily obtains
Z 3
dk
j = −2e v δf ≡ L11 E − L12 ∇ T (8.220)
(2π)3
Ω̂
Z
d3k
jq = 2 (ε − µ) v δf ≡ L21 E − L22 ∇ T (8.221)
(2π)3
Ω̂
The linear relations in eqn. (8.221) may be recast in the following form:
E= ρj + Q∇T
(8.227)
jq = ⊓ j − κ ∇ T ,
where the matrices ρ, Q, ⊓, and κ are given by
−1
ρ = L11 Q = L−1
11 L12 (8.228)
⊓= L21 L−1
11 κ = L22 − L21 L−1
11 L12 , (8.229)
12 In this section we use j to denote electrical current, rather than particle number current as before.
8.9. NONEQUILIBRIUM QUANTUM TRANSPORT 399
Figure 8.7: A thermocouple is a junction formed of two dissimilar metals. With no electrical current passing, an
electric field is generated in the presence of a temperature gradient, resulting in a voltage V = VA − VB .
One practical way to measure the thermopower is to form a junction between two dissimilar metals, A and B. The
junction is held at temperature T1 and the other ends of the metals are held at temperature T0 . One then measures
a voltage difference between the free ends of the metals – this is known as the Seebeck effect. Integrating the
electric field from the free end of A to the free end of B gives
ZB
VA − VB = − E · dl = (QB − QA )(T1 − T0 ) . (8.232)
A
What one measures here is really the difference in thermopowers of the two metals. For an absolute measurement
of QA , replace B by a superconductor (Q = 0 for a superconductor). A device which converts a temperature
gradient into an emf is known as a thermocouple.
400 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
Figure 8.8: A sketch of a Peltier effect refrigerator. An electrical current I is passed through a junction between
two dissimilar metals. If the dotted line represents the boundary of a thermally well-insulated body, then the body
cools when ⊓B > ⊓A , in order to maintain a heat current balance at the junction.
The Peltier effect has practical applications in refrigeration technology. Suppose an electrical current I is passed
through a junction between two dissimilar metals, A and B. Due to the difference in Peltier coefficients, there will
be a net heat current into the junction of W = (⊓A − ⊓B ) I. Note that this is proportional to I, rather than the
familiar I 2 result from Joule heating. The sign of W depends on the direction of the current. If a second junction
is added, to make an ABA configuration, then heat absorbed at the first junction will be liberated at the second. 13
We begin with the continuity equations for charge density ρ and energy density ε:
∂ρ
+∇·j =0 (8.233)
∂t
∂ε
+ ∇ · jε = j ·E , (8.234)
∂t
where E is the electric field14 . Now we invoke local thermodynamic equilibrium and write
∂ε ∂ε ∂n ∂ε ∂T
= +
∂t ∂n ∂t ∂T ∂t
µ ∂ρ ∂T
=− + cV , (8.235)
e ∂t ∂t
13 To create a refrigerator, stick the cold junction inside a thermally insulated box and the hot junction outside the box.
14 Note that it is E · j and not E· j which is the source term in the energy continuity equation.
8.9. NONEQUILIBRIUM QUANTUM TRANSPORT 401
where n is the electron number density (n = −ρ/e) and cV is the specific heat. We may now write
∂T ∂ε µ ∂ρ
cV = +
∂t ∂t e ∂t
µ
= j · E − ∇ · jε − ∇·j
e
= j · E − ∇ · jq . (8.236)
Invoking jq = ⊓j − κ∇ T , we see that if there is no electrical current (j = 0), we obtain the heat equation
∂T ∂ 2T
cV = καβ . (8.237)
∂t ∂xα ∂xβ
This results in a time scale τT for temperature diffusion τT = CL2 cV /κ, where L is a typical length scale and C is
a numerical constant. For a cube of size L subjected to a sudden external temperature change, L is the side length
and C = 1/3π 2 (solve by separation of variables).
We will henceforth assume that sufficient crystalline symmetry exists (e.g. cubic symmetry) to render all the trans-
port coefficients multiples of the identity matrix. Under such conditions, we may write Jnαβ = Jn δαβ with
Z Z
1 n ∂f 0
Jn = dε τ (ε) (ε − µ) − dSε |v| . (8.238)
12π 3 ~ ∂ε
The low-temperature behavior is extracted using the Sommerfeld expansion,
Z∞
∂f 0
I ≡ dε H(ε) − = πD csc(πD) H(ε) (8.239)
∂ε ε=µ
−∞
π2
= H(µ) + (k T )2 H ′′ (µ) + . . . (8.240)
6 B
∂
where D ≡ kB T ∂ε is a dimensionless differential operator.15
Let us now perform some explicit calculations in the case of a parabolic band with an energy-independent scat-
tering time τ . In this case, one readily finds
σ0 −3/2
3/2 n
Jn = µ πD csc πD ε (ε − µ) , (8.241)
e2 ε=µ
and hole bands), and not temperature and chemical potential. An equation of state relating n, µ, and T is then inverted to obtain µ(n, T ), so
that all results ultimately may be expressed in terms of n and T .
402 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
• Hall effect ( ∂T ∂T
∂x = ∂y = jy = 0)
An electrical current j = jx x̂ and a field B = Bz ẑ yield an electric field E. The Hall coefficient is RH =
Ey /jx Bz = −ν.
• Ettingshausen effect ( ∂T
∂x = jy = jq,y = 0)
∂T
An electrical current j = jx x̂ and a field B = Bz ẑ yield a temperature gradient ∂y . The Ettingshausen
coefficient is P = ∂T
∂y jx Bz = −θ/κ.
A stochastic process is one which is partially random, i.e. it is not wholly deterministic. Typically the randomness is
due to phenomena at the microscale, such as the effect of fluid molecules on a small particle, such as a piece of dust
in the air. The resulting motion (called Brownian motion in the case of particles moving in a fluid) can be described
only in a statistical sense. That is, the full motion of the system is a functional of one or more independent random
variables. The motion is then described by its averages with respect to the various random distributions.
Consider a particle of mass M subjected to dissipative and random forcing. We’ll examine this system in one
dimension to gain an understanding of the essential physics. We write
ṗ + γp = F + η(t) . (8.258)
Here, γ is the damping rate due to friction, F is a constant external force, and η(t) is a stochastic random force.
This equation, known as the Langevin equation, describes a ballistic particle being buffeted by random forcing
events. Think of a particle of dust as it moves in the atmosphere; F would then represent the external force due
to gravity and η(t) the random forcing due to interaction with the air molecules. For a sphere of radius a moving
with velocity v in a fluid, the Stokes drag is given by Fdrag = −6πηav, where a is the radius. Thus,
6πηa
γStokes = , (8.259)
M
where M is the mass of the particle. It is illustrative to compute γ in some setting. Consider a micron sized
droplet (a = 10−4 cm) of some liquid of density ρ ∼ 1.0 g/cm3 moving in air at T = 20◦ C. The viscosity of air is
404 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
η = 1.8 × 10−4 g/cm · s at this temperature16 . If the droplet density is constant, then γ = 9η/2ρa2 = 8.1 × 104 s−1 ,
hence the time scale for viscous relaxation of the particle is τ = γ −1 = 12 µs. We should stress that the viscous
damping on the particle is of course due to the fluid molecules, in some average ‘coarse-grained’ sense. The
random component to the force η(t) would then represent the fluctuations with respect to this average.
We can easily integrate this equation:
d
p eγt = F eγt + η(t) eγt
dt
Zt (8.260)
−γt F
p(t) = p(0) e + 1 − e−γt + ds η(s) eγ(s−t)
γ
0
Note that p(t) is indeed a functional of the random function η(t). We can therefore only compute averages in order
to describe the motion of the system.
The first average we will compute is that of p itself. In so doing, we assume that η(t) has zero mean: η(t) = 0.
Then
F
p(t) = p(0) e−γt + 1 − e−γt . (8.261)
γ
−1 −1
time scale γ , the initial conditions p(0) are effectively forgotten, and asymptotically for t ≫ γ we have
On the
p(t) → F/γ, which is the terminal momentum.
Next, consider
Zt Z t
2
2
p (t) = p(t) + ds1 ds2 eγ(s1 −t) eγ(s2 −t) η(s1 ) η(s2 ) . (8.262)
0 0
We now need to know the two-time correlator η(s1 ) η(s2 ) . We assume that the correlator is a function only of
the time difference ∆s = s1 − s2 , so that the random force η(s) satisfies
η(s) = 0 (8.263)
η(s1 ) η(s2 ) = φ(s1 − s2 ) . (8.264)
The function φ(s) is the autocorrelation function of the random force. A macroscopic object moving in a fluid is
constantly buffeted by fluid particles over its entire perimeter. These different fluid particles are almost completely
uncorrelated, hence φ(s) is basically nonzero except on a very small time scale τφ , which is the time a single fluid
particle spends interacting with the object. We can take τφ → 0 and approximate
Zt
2
2
p (t) = p(t) + Γ ds e2γ(s−t)
0 (8.266)
2 Γ
= p(t) + 1 − e−2γt .
2γ
16 The cgs unit of viscosity is the Poise (P). 1 P = 1 g/cm·s.
8.10. STOCHASTIC PROCESSES 405
Consider the case where F = 0 and the limit t ≫ γ −1 . We demand that the object thermalize at temperature T .
Thus, we impose the condition
2
p (t)
= 12 kB T =⇒ Γ = 2γM kB T , (8.267)
2M
The full expressions for this and subsequent expressions, including subleading terms, are contained in an ap-
pendix, §8.14.
Let’s now compute the position x(t). We find
Zt Zs
1
x(t) = x(t) + ds ds1 η(s1 ) eγ(s1 −s) , (8.269)
M
0 0
where
Ft 1 F
x(t) = x(0) + + v(0) − 1 − e−γt . (8.270)
γM γ γM
Note that for γt ≪ 1 we have x(t) = x(0) + v(0) t + 21 M −1 F t2 + O(t3 ), as is appropriate for ballistic particles
moving under the influence of
a constant
force. This long time limit of course agrees with our earlier evaluation
for the terminal velocity, v∞ = p(∞) /M = F/γM . We next compute the position autocorrelation:
Zt Zt′ Zs Zs′
1 ′
x(t) x(t′ ) − x(t) x(t′ ) = 2 ds ds′ e−γ(s+s ) ds1 ds′1 eγ(s1 +s2 ) η(s1 ) η(s2 )
M
0 0 0 0
2k T
= B min(t, t′ ) + O(1) .
γM
In particular, the equal time autocorrelator is
2
2 2k T t
x (t) − x(t) = B ≡ 2D t , (8.271)
γM
at long times, up to terms of order unity. Here,
kB T
D= (8.272)
γM
is the diffusion constant. For a liquid droplet of radius a = 1 µm moving in air at T = 293 K, for which η =
1.8 × 10−4 P, we have
kB T (1.38 × 10−16 erg/K) (293 K)
D= = = 1.19 × 10−7 cm2 /s . (8.273)
6πηa 6π (1.8 × 10−4 P) (10−4 cm)
This result presumes that the droplet is large enough compared to the intermolecular distance in the fluid that one
can adopt a continuum approach and use the Navier-Stokes equations, and then assuming a laminar flow.
406 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
If we consider molecular diffusion, the situation is quite a bit different. As we shall derive below in §8.10.3, the
molecular diffusion constant is D = ℓ2 /2τ , where ℓ is the mean free path and τ is the collision time. As we found
in eqn. 8.91, the mean free path ℓ, collision time τ , number density n, and total scattering cross section σ are related
by
1
ℓ = v̄τ = √ , (8.274)
2 nσ
p
where v̄ = 8kB T /πm is the average particle speed. Approximating the particles as hard spheres, we have
σ = 4πa2 , where a is the hard sphere radius. At T = 293 K, and p = 1 atm, we have n = p/kB T = 2.51 ×
1019 cm−3 . Since air is predominantly composed of N2 molecules, we take a = 1.90 × 10−8 cm and m = 28.0 amu =
4.65 × 10−23 g, which are appropriate for N2 . We find an average speed of v̄ = 471 m/s and a mean free path of
ℓ = 6.21 × 10−6 cm. Thus, D = 21 ℓv̄ = 0.146 cm2 /s. Though much larger than the diffusion constant for large
droplets, this is still too small to explain common experiences. Suppose we set the characteristic distance scale
at d = 10 cm and we ask how much time a point source would take to diffuse out to this radius. The answer is
∆t = d2 /2D = 343 s, which is between five and six minutes. Yet if someone in the next seat emits a foul odor,
your sense the offending emission in on the order of a second. What this tells us is that diffusion isn’t the only
transport process involved in these and like phenomena. More important are convection currents which distribute
the scent much more rapidly.
1
ẍ + γ ẋ + ω02 x = η(t) . (8.276)
M
At this point there are several ways to proceed.
Perhaps the most straightforward is by use of the Laplace transform. Recall:
Z∞
x̂(ν) = dt e−νt η(ν) (8.277)
0
Z
dν +νt
x(t) = e x̂(ν) , (8.278)
2πi
C
where the contour C proceeds from a − i∞ to a + i∞ such that all poles of the integrand lie to the left of C. We
then have
Z∞ Z∞
1 −νt 1
dt e η(t) = dt e−νt ẍ + γ ẋ + ω02 x
M M
0 0
= −(ν + γ) x(0) − ẋ(0) + ν 2 + γν + ω02 x̂(ν) . (8.279)
Thus, we have
Z∞
(ν + γ) x(0) + ẋ(0) 1 1
x̂(ν) = 2 + · 2 dt e−νt η(t) . (8.280)
2
ν + γν + ω0 M ν + γν + ω02
0
8.10. STOCHASTIC PROCESSES 407
x(0) ẋ(0) ν+ t
x(t) = ν+ eν− t − ν− eν+ t + e − eν− t
ν+ − ν− ν+ − ν−
Z∞ (8.283)
+ ds K(t − s) η(s) ,
0
where
Θ(t − s) ν+ (t−s)
K(t − s) = e − eν− (t−s) (8.284)
M (ν+ − ν− )
is the response kernel and Θ(t − s) is the step function which is unity for t > s and zero otherwise. The response is
causal, i.e. x(t) depends on η(s) for all previous times s < t, but not for future times s > t. Note that K(τ ) decays
exponentially for τ → ∞, if Re(ν± ) < 0. The marginal case where ω0 = 0 and ν+ = 0 corresponds to the diffusion
calculation we performed in the previous section.
Consider an object moving on a one-dimensional lattice in such a way that every time step it moves either one
unit to the right or left, at random. If the lattice spacing is ℓ, then after n time steps the position will be
n
X
xn = ℓ σj , (8.285)
j=1
where (
+1 if motion is one unit to right at time step j
σj = (8.286)
−1 if motion is one unit to left at time step j .
Clearly hσj i = 0, so hxn i = 0. Now let us compute
n X
X n
2
xn = ℓ2 hσj σj ′ i = nℓ2 , (8.287)
j=1 j ′ =1
where we invoke
σj σj ′ = δjj ′ . (8.288)
If the length of each time step is τ , then we have, with t = nτ ,
2 ℓ2
x (t) = t, (8.289)
τ
and we identify the diffusion constant
ℓ2
D= . (8.290)
2τ
408 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
Suppose, however, the random walk is biased, so that the probability for each independent step is given by
P (σ) = p δσ,1 + q δσ,−1 , (8.291)
where p + q = 1. Then
hσj i = p − q = 2p − 1 (8.292)
and
hσj σj ′ i = (p − q)2 1 − δjj ′ + δjj ′
(8.293)
= (2p − 1)2 + 4 p (1 − p) δjj ′ .
Then
hxn i = (2p − 1) ℓn (8.294)
2
2
xn − xn = 4 p (1 − p) ℓ2 n . (8.295)
for any value of t1 . This is also known as the Chapman-Kolmogorov equation. In words, what it says is that the
probability density for a particle being at x2 at time t2 , given that it was at x0 at time t0 , is given by the product of
the probability density for being at x2 at time t2 given that it was at x1 at t1 , multiplied by that for being at x1 at t1
given it was at x0 at t0 , integrated over x1 . This should be intuitively obvious, since if we pick any time t1 ∈ [t0 , t2 ],
then the particle had to be somewhere at that time. Indeed, one wonders how Chapman and Kolmogorov got their
names attached to a result that is so obvious. At any rate, a picture is worth a thousand words: see fig. 8.9.
Proceeding, we may write
Z∞
P (x, t + δt | x0 , t0 ) = dx′ P (x, t + δt | x′ , t) P (x′ , t | x0 , t0 ) . (8.300)
−∞
Now
P (x, t + δt | x′ , t) = δ x − δx(t) − x′
d
2 d2
= 1 + δx(t) +2 1
δx(t) + . . . δ(x − x′ ) (8.301)
dx′ dx′ 2
d δ(x − x′ ) 2
′ d δ(x − x )
′
= δ(x − x′ ) + F1 (x′ ) δt + 1
2 F2 (x ) 2 δt + O (δt)2 ,
dx′ dx′
8.10. STOCHASTIC PROCESSES 409
Figure 8.9: Interpretive sketch of the mathematics behind the Chapman-Kolmogorov equation.
where the average is over the random variables. We now insert this result into eqn. 8.300, integrate by parts,
divide by δt, and then take the limit δt → 0. The result is the Fokker-Planck equation,
∂P ∂ 1 ∂2
=− F1 (x) P (x, t) + F2 (x) P (x, t) . (8.302)
∂t ∂x 2 ∂x2
Let’s apply our Fokker-Planck equation to a description of Brownian motion. From our earlier results, we have
F
F1 (x) = , F2 (x) = 2D . (8.303)
γM
A formal proof of these results is left as an exercise for the reader. The Fokker-Planck equation is then
∂P ∂P ∂ 2P
= −u +D 2 , (8.304)
∂t ∂x ∂x
where u = F/γM is the average terminal velocity. If we make a Galilean transformation and define
y = x − ut , s=t (8.305)
∂P ∂ 2P
=D 2 . (8.306)
∂s ∂y
This is known as the diffusion equation. Eqn. 8.304 is also a diffusion equation, rendered in a moving frame.
410 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
While the Galilean transformation is illuminating, we can easily solve eqn. 8.304 without it. Let’s take a look at
this equation after Fourier transforming from x to q:
Z∞
dq iqx
P (x, t) = e P̂ (q, t) (8.307)
2π
−∞
Z∞
P̂ (q, t) = dx e−iqx P (x, t) . (8.308)
−∞
∂
Then as should be well known to you by now, we can replace the operator ∂x with multiplication by iq, resulting
in
∂
P̂ (q, t) = −(Dq 2 + iqu) P̂ (q, t) , (8.309)
∂t
with solution 2
P̂ (q, t) = e−Dq t e−iqut P̂ (q, 0) . (8.310)
We now apply the inverse transform to get back to x-space:
Z∞ Z∞
dq iqx −Dq2 t −iqut ′
P (x, t) = e e e dx′ e−iqx P (x′ , 0)
2π
−∞ −∞
Z∞ Z∞
dq −Dq2 t iq(x−ut−x′ )
= dx′ P (x′ , 0) e e (8.311)
2π
−∞ −∞
Z∞
= dx′ K(x − x′ , t) P (x′ , 0) ,
−∞
where
1 2
K(x, t) = √ e−(x−ut) /4Dt (8.312)
4πDt
is the diffusion kernel. We now have a recipe for obtaining P (x, t) given the initial conditions P (x, 0). If P (x, 0) =
δ(x), describing a particle confined to an infinitesimal region about the origin, then P (x, t) = K(x, t) is the prob-
ability distribution for finding the particle at x at time t. There are two aspects to K(x, t) which merit comment.
√ velocity u. This is due to the presence of the external
The first is that the center of the distribution moves with
force. The second is that the standard deviation σ = 2Dt is increasing in time, so the distribution is not only
shifting its center but it is also getting broader as time evolves. This movement of the center and broadening are
what we have called drift and diffusion, respectively.
Another way to model stochastic processes is via the master equation, which was discussed in chapter 3. Recall
that if Pi (t) is the probability for a system to be in state | i i at time t and Wij is the transition rate from state | j i to
state | i i, then
dPi X
= Wij Pj − Wji Pi . (8.313)
dt j
Consider a birth-death process in which the states | n i are labeled by nonnegative integers. Let αn denote the
rate of transitions from | n i → | n + 1 i and let βn denote the rate of transitions from | n i → | n − 1 i. The master
8.10. STOCHASTIC PROCESSES 411
Let us assume we can write αn = K ᾱ(n/K) and βn = K β̄(n/K), where K ≫ √ 1. We assume the distribution
Pn (t) has a time-dependent maximum
√ at n = Kφ(t) and a width proportional to K. We expand relative to this
maximum, writing n ≡ Kφ(t) + K ξ and we define Pn (t) ≡ Π(ξ, t). We now rewrite the master equation in eqn.
8.314 in terms of Π(ξ, t). Since n is an independent variable, we set
√ √
dn = K φ̇ dt + K dξ ⇒ dξ n = − K φ̇ dt . (8.315)
Therefore
dPn √ ∂Π ∂Π
= − K φ̇ + . (8.316)
dt ∂ξ ∂t
Next, we write, for any function fn ,
fn = Kf φ + K −1/2 ξ
(8.317)
= Kf (φ) + K 1/2 ξ f ′ (φ) + 1
2 ξ 2 f ′′ (φ) + . . . .
Similarly,
fn±1 = Kf φ + K −1/2 ξ ± K −1
(8.318)
= Kf (φ) + K 1/2 ξ f ′ (φ) ± f ′ (φ) + 21 ξ 2 f ′′ (φ) + . . . .
√
Dividing both sides of eqn. 8.314 by K, we have
∂Π −1/2 ∂Π ∂Π −1/2 ′ ′ ∂Π 1 ∂ 2Π ′ ′
− φ̇ + K = (β̄ − ᾱ) +K (β̄ − ᾱ ) ξ + 2 (ᾱ + β̄) 2 + (β̄ − ᾱ )Π + . . . . (8.319)
∂ξ ∂t ∂ξ ∂ξ ∂ξ
∂Π ∂ ∂ 2Φ
= −f ′ φ(t) ξ Π + 21 g φ(t) , (8.321)
∂t ∂ξ ∂ξ 2
where g(φ) ≡ ᾱ(φ) + β̄(φ). If in the limit t → ∞, eqn. 8.320 evolves to a stable fixed point φ∗ , then the stationary
solution of the Fokker-Planck eqn. 8.321, Πeq (ξ) = Π(ξ, t = ∞) must satisfy
∂ ∂ 2Πeq 1 2
/2σ2
−f ′ (φ∗ ) ξ Πeq + 1
2 g(φ∗ ) =0 ⇒ Πeq (ξ) = √ e−ξ , (8.322)
∂ξ ∂ξ 2 2πσ 2
where
g(φ∗ )
σ2 = − . (8.323)
2f ′ (φ∗ )
Now both α and β are rates, hence both are positive and thus g(φ) > 0. We see that the condition σ 2 > 0 , which
is necessary for a normalizable equilibrium distribution, requires f ′ (φ∗ ) < 0, which is saying that the fixed point
in Eqn. 8.320 is stable.
17 We further demand βn=0 = 0 and P−1 (t) = 0 at all times.
412 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
Problem : The linearized Boltzmann operator Lψ is a complicated functional. Suppose we replace L by L, where
3/2 Z
m mu2
Lψ = −γ ψ(v, t) + γ d3u exp −
2πkB T 2kB T
( ) (8.324)
2
m 2 mu 3 mv 2 3
× 1+ u·v+ − − ψ(u, t) .
kB T 3 2kB T 2 2kB T 2
Show that L shares all the important properties of L. What is the meaning of γ? Expand ψ(v, t) in spherical
harmonics and Sonine polynomials,
X
ψ(v, t) = arℓm (t) S r 1 (x) xℓ/2 Ymℓ (n̂), (8.325)
ℓ+ 2
rℓm
with x = mv 2 /2kB T , and thus express the action of the linearized Boltzmann operator algebraically on the expan-
sion coefficients arℓm (t).
The Sonine polynomials Sαn (x) are a complete, orthogonal set which are convenient to use in the calculation of
transport coefficients. They are defined as
n
X Γ(α + n + 1) (−x)m
Sαn (x) = , (8.326)
m=0
Γ(α + m + 1) (n − m)! m!
Solution : The ‘important properties’ of L are that it annihilate the five collisional invariants, i.e. 1, v, and v 2 , and
that all other eigenvalues are negative. That this is true for L can be verified by an explicit calculation.
Plugging the conveniently parameterized form of ψ(v, t) into L, we have
X Z∞
γ X 1/2
Lψ = −γ arℓm (t) S r 1 (x) xℓ/2
Ymℓ (n̂) + arℓm (t) dx1 x1 e−x1
rℓm
ℓ+ 2 2π 3/2 rℓm
0 (8.328)
Z h
1/2 i ℓ/2
× dn̂1 1 + 2 x1/2 x1 n̂· n̂1 + 2
3 x− 3
2 x1 − 3
2 S r 1 (x1 ) x1 Ymℓ (n̂1 ) ,
ℓ+ 2
∗
1 = 4π Y00 (n̂) Y00 (n̂1 ) (8.330)
4π h 1 ∗ ∗ 1 ∗
i
n̂· n̂1 = Y0 (n̂) Y01 (n̂1 ) + Y11 (n̂) Y11 (n̂1 ) + Y−1
1
(n̂) Y−1 (n̂1 ) . (8.331)
3
We can do the integrals by appealing to the orthogonality relations for the spherical harmonics and Sonine poly-
nomials:
Z
′ ∗
dn̂ Ymℓ (n̂) Yml ′ (n̂) = δll′ δmm′ (8.332)
Z∞
′ Γ(n + α + 1)
dx e−x xα Sαn (x) Sαn (x) = δnn′ . (8.333)
Γ(n + 1)
0
Appealing now to the orthogonality of the Sonine polynomials, and recalling that
√
Γ( 21 ) = π , Γ(1) = 1 , Γ(z + 1) = z Γ(z) , (8.336)
we integrate over√x1 . For the first term in brackets, we invoke the orthogonality relation with n = 0 and α = 21 ,
giving Γ( 23 ) = 21 π. For the second bracketed term, we have n = 0 but α = 23 , and we obtain Γ( 25 ) = 23 Γ( 32 ),
while the third bracketed term involves leads to n = 1 and α = 12 , also yielding Γ( 25 ) = 32 Γ( 32 ). Thus, we obtain
the simple and pleasing result
X′
Lψ = −γ arℓm (t) S r 1 (x) xℓ/2 Ymℓ (n̂) (8.337)
ℓ+ 2
rℓm
414 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
are to be excluded from the sum. But these are just the functions which correspond to the five collisional invariants!
Thus, we learn that
ψrℓm (v) = Nrℓm S r 1 (x) xℓ/2 Ymℓ (n̂), (8.339)
ℓ+ 2
is an eigenfunction of L with eigenvalue −γ if (r, ℓ, m) does not correspond to one of the five collisional invariants.
In the latter case, the eigenvalue is zero. Thus, the algebraic action of L on the coefficients arℓm is
(
−γ arℓm if (r, ℓ, m) ∈ / CI
(La)rℓm = (8.340)
=0 if (r, ℓ, m) ∈ CI
Let x ∈ R be a random variable, and P (x) a probability distribution for x. The average of any function φ(x) is then
Z∞ , Z∞
φ(x) = dx P (x) φ(x) dx P (x) . (8.342)
−∞ −∞
Let η(t) be
a random
function of t, with η(t) ∈ R, and let P η(t) be the probability distribution functional for η(t).
Then if Φ η(t) is a functional of η(t), the average of Φ is given by
Z ,Z
Dη P η(t) Φ η(t) Dη P η(t) (8.343)
R
The expression Dη P [η] Φ[η] is a functional integral. A functional integral is a continuum limit of a multivariable
integral. Suppose η(t) were defined on a set of t values tn = nτ . A functional of η(t) becomes a multivariable
function of the values ηn ≡ η(tn ). The metric then becomes
Y
Dη −→ dηn . (8.344)
n
8.12. APPENDIX II : DISTRIBUTIONS AND FUNCTIONALS 415
Figure 8.10: Discretization of a continuous function η(t). Upon discretization, a functional Φ η(t) becomes an
ordinary multivariable function Φ({ηj }).
In fact, for our purposes we will not need to know any details about the functional measure Dη; we will finesse
this delicate issue18 . Consider the generating functional,
Z Z∞ !
Z J(t) = Dη P [η] exp dt J(t) η(t) . (8.345)
−∞
It is clear that
1 δ nZ[J]
= η(t1 ) · · · η(tn ) . (8.346)
Z[J] δJ(t1 ) · · · δJ(tn )
J(t)=0
The function J(t) is an arbitrary source function. We differentiate with respect to it in order to find the η-field
correlators.
Let’s compute the generating function for a class of distributions of the Gaussian form,
Z∞ !
1 2 2 2
P [η] = exp − dt τ η̇ + η (8.347)
2Γ
−∞
Z∞ !
1 dω 2
= exp − 1 + ω 2 τ 2 η̂(ω) . (8.348)
2Γ 2π
−∞
Then Fourier transforming the source function J(t), it is easy to see that
Z∞ !
Γ ˆ 2
dω J(ω)
Z[J] = Z[0] · exp . (8.349)
2 2π 1 + ω 2 τ 2
−∞
Note that with η(t) ∈ R and J(t) ∈ R we have η ∗ (ω) = η(−ω) and J ∗ (ω) = J(−ω). Transforming back to real time,
18 A discussion of measure for functional integrals is found in R. P. Feynman and A. R. Hibbs, Quantum Mechanics and Path Integrals.
416 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
we have !
Z∞ Z∞
1 ′ ′ ′
Z[J] = Z[0] · exp dt dt J(t) G(t − t ) J(t ) , (8.350)
2
−∞ −∞
where
Γ −|s|/τ b Γ
G(s) = e , G(ω) = (8.351)
2τ 1 + ω2τ 2
is the Green’s function, in real and Fourier space. Note that
Z∞
b
ds G(s) = G(0) =Γ . (8.352)
−∞
η(t1 ) η(t2 ) = G(t1 − t2 ) (8.353)
η(t1 ) η(t2 ) η(t3 ) η(t4 ) = G(t1 − t2 ) G(t3 − t4 ) + G(t1 − t3 ) G(t2 − t4 ) (8.354)
+ G(t1 − t4 ) G(t2 − t3 ) .
where the sum is over all distinct contractions of the sequence 1-2 · · · 2n into products of pairs. How many terms
are there? Some simple combinatorics answers this question. Choose the index 1. There are (2n − 1) other time
indices with which it can be contracted. Now choose another index. There are (2n − 3) indices with which that
index can be contracted. And so on. We thus obtain
# of contractions (2n)!
C(n) ≡ = (2n − 1)(2n − 3) · · · 3 · 1 = n . (8.356)
of 1-2-3 · · · 2n 2 n!
We can also solve general autonomous linear inhomogeneous ODEs of the form
dnx dn−1 x dx
+ a n−1 + . . . + a1 + a0 x = ξ(t) . (8.357)
dtn dtn−1 dt
We can write this as
Lt x(t) = ξ(t) , (8.358)
where Lt is the nth order differential operator
dn dn−1 d
Lt = n
+ a n−1 n−1
+ . . . + a1 + a0 . (8.359)
dt dt dt
The general solution to the inhomogeneous equation is given by
Z∞
x(t) = xh (t) + dt′ G(t, t′ ) ξ(t′ ) , (8.360)
−∞
8.13. APPENDIX III : GENERAL LINEAR AUTONOMOUS INHOMOGENEOUS ODES 417
where G(t, t′ ) is the Green’s function. Note that Lt xh (t) = 0. Thus, in order for eqns. 8.358 and 8.360 to be true,
we must have
this vanishes
z }| { Z∞
Lt x(t) = Lt xh (t) + dt′ Lt G(t, t′ ) ξ(t′ ) = ξ(t) , (8.361)
−∞
which means that
Lt G(t, t′ ) = δ(t − t′ ) , (8.362)
where δ(t − t′ ) is the Dirac δ-function.
If the differential equation Lt x(t) = ξ(t) is defined over some finite or semi-infinite t interval with prescribed
boundary conditions on x(t) at the endpoints, then G(t, t′ ) will depend on t and t′ separately. For the case we are
now considering, let the interval be the entire real line t ∈ (−∞, ∞). Then G(t, t′ ) = G(t − t′ ) is a function of the
single variable t − t′ .
d
d
Note that Lt = L dt may be considered a function of the differential operator dt . If we now Fourier transform
the equation Lt x(t) = ξ(t), we obtain
Z∞ Z∞ n
iωt iωt d dn−1 d
dt e ξ(t) = dt e + an−1 n−1 + . . . + a1 + a0 x(t)
dtn dt dt
−∞ −∞
( (8.363)
Z∞
= dt eiωt (−iω)n + an−1 (−iω)n−1 + . . . + a1 (−iω) + a0 x(t) .
−∞
Thus, if we define
n
X
L̂(ω) = ak (−iω)k , (8.364)
k=0
then we have
ˆ
L̂(ω) x̂(ω) = ξ(ω) , (8.365)
where an ≡ 1. According to the Fundamental Theorem of Algebra, the n th
degree polynomial L̂(ω) may be
uniquely factored over the complex ω plane into a product over n roots:
L̂(ω) = (−i)n (ω − ω1 )(ω − ω2 ) · · · (ω − ωn ) . (8.366)
∗ ∗ ∗
If the {ak } are all real, then L̂(ω) = L̂(−ω ), hence if Ω is a root then so is −Ω . Thus, the roots appear in pairs
which are symmetric about the imaginary axis. I.e. if Ω = a + ib is a root, then so is −Ω ∗ = −a + ib.
The general solution to the homogeneous equation is
n
X
xh (t) = Aσ e−iωσ t , (8.367)
σ=1
which involves n arbitrary complex constants Ai . The susceptibility, or Green’s function in Fourier space, Ĝ(ω) is
then
1 in
Ĝ(ω) = = , (8.368)
L̂(ω) (ω − ω1 )(ω − ω2 ) · · · (ω − ωn )
∗
Note that Ĝ(ω) = Ĝ(−ω), which is equivalent to the statement that G(t − t′ ) is a real function of its argument.
The general solution to the inhomogeneous equation is then
Z∞
x(t) = xh (t) + dt′ G(t − t′ ) ξ(t′ ) , (8.369)
−∞
418 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
where xh (t) is the solution to the homogeneous equation, i.e. with zero forcing, and where
Z∞
′ dω −iω(t−t′ )
G(t − t ) = e Ĝ(ω)
2π
−∞
Z∞ ′
n dω e−iω(t−t ) (8.370)
=i
2π (ω − ω1 )(ω − ω2 ) · · · (ω − ωn )
−∞
n
X ′
e−iωσ (t−t )
= Θ(t − t′ ) ,
σ=1 i L′ (ω σ)
where we assume that Im ωσ < 0 for all σ. This guarantees causality – the response x(t) to the influence ξ(t′ ) is
nonzero only for t > t′ .
As an example, consider the familiar case
Now let ′us evaluate the two-point correlation function x(t) x(t′ ) , assuming the noise is correlated according to
ξ(s) ξ(s ) = φ(s − s′ ). We assume t, t′ → ∞ so the transient contribution xh is negligible. We then have
Z∞ Z∞
x(t) x(t′ ) = ds ds′ G(t − s) G(t′ − s′ ) ξ(s) ξ(s′ )
−∞ −∞
(8.374)
Z∞
dω 2 ′
= φ̂(ω) Ĝ(ω) eiω(t−t ) .
2π
−∞
may be represented by the first order system ϕ̇ = V (ϕ). To see this, define ϕk = dk−1x/dtk−1 , with k = 1, . . . , n.
Thus, for k < n we have ϕ̇k = ϕk+1 , and ϕ̇n = F . In other words,
ϕ̇ V (ϕ )
z }| { z }| {
ϕ1 ϕ2
d ... ..
.
= . (8.376)
dt ϕ ϕn
n−1
ϕn F ϕ1 , . . . , ϕp
dnx dn−1x dx
n
+ a n−1 n−1
+ . . . + a1 + a0 x = ξ(t) (8.377)
dt dt dt
may be written in matrix form, as
Q ξ
z }| { z }| {
ϕ1 0 1 0 ··· 0 ϕ1 0
ϕ
d 2 0 0 1 ··· 0 ϕ 0
2
= . .. .. .. . + .. . (8.378)
dt ... .. . . . .
. .
ϕn −a0 −a1 −a2 · · · −an−1 ϕn ξ(t)
Thus,
ϕ̇ = Q ϕ + ξ , (8.379)
and if the coefficients ck are time-independent, i.e. the ODE is autonomous.
For the homogeneous case where ξ(t) = 0, the solution is obtained by exponentiating the constant matrix Qt:
the exponential of a matrix may be given meaning by its Taylor series expansion. If the ODE is not autonomous,
then Q = Q(t) is time-dependent, and the solution is given by the path-ordered exponential,
( Zt )
ϕ(t) = P exp dt′ Q(t′ ) ϕ(0) , (8.381)
0
where P is the path ordering operator which places earlier times to the right. As defined, the equation ϕ̇ = V (ϕ)
is autonomous, since the t-advance mapping gt depends only on t and on no other time variable. However, by
extending the phase space M ∋ ϕ from M → M × R, which is of dimension n + 1, one can describe arbitrary
time-dependent ODEs.
In general, path ordered exponentials are difficult to compute analytically. We will henceforth consider the au-
tonomous case where Q is a constant matrix in time. We will assume the matrix Q is real, but other than that it
has no helpful symmetries. We can however decompose it into left and right eigenvectors:
n
X
Qij = νσ Rσ,i Lσ,j . (8.382)
σ=1
P
Or, in bra-ket notation, Q = σ νσ |Rσ ihLσ |. The normalization condition we use is
Lσ Rσ′ = δσσ′ , (8.383)
420 CHAPTER 8. NONEQUILIBRIUM PHENOMENA
where νσ are the eigenvalues of Q. The eigenvalues may be real or imaginary. Since the characteristic poly-
nomial P (ν) = det (ν I − Q) has real coefficients, we know that the eigenvalues of Q are either real or come in
complex conjugate pairs.
Consider, for example, the n = 2 system we studied earlier. Then
0 1
Q= . (8.384)
−ω02 −γ
q
The eigenvalues are as before: ν± = − 21 γ ± 1 2
4γ − ω02 . The left and right eigenvectors are
±1 1
L± = −ν∓ 1 , R± = . (8.385)
ν+ − ν− ν±
The utility of working in a left-right eigenbasis is apparent once we reflect upon the result
n
X
f (Q) = f (νσ ) Rσ Lσ (8.386)
σ=1
for any function f . Thus, the solution to the general autonomous homogeneous case is
n
X
ϕ(t) = eνσ t Rσ Lσ ϕ(0) (8.387)
σ=1
Xn n
X
ϕi (t) = eνσ t Rσ,i Lσ,j ϕj (0) . (8.388)
σ=1 j=1
If Re (νσ ) ≤ 0 for all σ, then the initial conditions ϕ(0) are forgotten on time scales τσ = νσ−1 . Physicality demands
that this is the case.
Now let’s consider the inhomogeneous case where ξ(t) 6= 0. We begin by recasting eqn. 8.379 in the form
d −Qt
e ϕ = e−Qt ξ(t) . (8.389)
dt
We can integrate this directly:
Zt
Qt
ϕ(t) = e ϕ(0) + ds eQ(t−s) ξ(s) . (8.390)
0
In component notation,
n
X n Zt
X
ϕi (t) = e νσ t
Rσ,i
Lσ ϕ(0) + Rσ,i ds eνσ (t−s) Lσ ξ(s) . (8.391)
σ=1 σ=1 0
Note that the first term on the RHS is the solution to the homogeneous equation, as must be the case when ξ(s) = 0.
The solution in eqn. 8.391 holds for general Q and ξ(s). For the particular form of Q and ξ(s) in eqn. 8.378, we
can proceed further. For starters, hLσ |ξ(s)i = Lσ,n ξ(s). We can further exploit a special feature of the Q matrix to
analytically determine all its left and right eigenvectors. Applying Q to the right eigenvector |Rσ i, we obtain
We are free to choose Rσ,1 = 1 for all σ and defer the issue of normalization to the derivation of the left eigenvec-
tors. Thus, we obtain the pleasingly simple result,
Rσ,k = νσk−1 . (8.393)
Applying Q to the left eigenvector hLσ |, we obtain
−a0 Lσ,n = νσ Lσ,1 (8.394)
Lσ,j−1 − aj−1 Lσ,n = νσ Lσ,j (j > 1) . (8.395)
From these equations we may derive
Lσ,n k−1
X Lσ,n X n
Lσ,k = − aj νσj−k−1 = aj νσj−k−1 . (8.396)
νσ j=0 νσ
j=k
P
The equality in the above equation is derived using the result P (νσ ) = nj=0 aj νσj = 0. Recall also that an ≡ 1.
We now impose the normalization condition,
n
X
Lσ,k Rσ,k = 1 . (8.397)
k=1
This condition determines our last remaining unknown quantity (for a given σ), Lσ,p :
n
X
Lσ Rσ = Lσ,n k ak νσk−1 = P ′ (νσ ) Lσ,n , (8.398)
k=1
where P ′ (ν) is the first derivative of the characteristic polynomial. Thus, we obtain another neat result,
1
Lσ,n = . (8.399)
P ′ (νσ )
Zt
−γt F
p(t) = p(0) e + 1 − e−γt + ds η(s) eγ(s−t) . (8.405)
γ
0
Zt Zt′
′ ′
p(t) p(t′ ) − p(t) p(t′ ) = ds ds′ eγ(s−t) eγ(s −t ) η(s) η(s′ )
0 0
(8.406)
tZmin
′
′ ′
= Γ e−γ(t+t ) ds e2γs = M kB T e−γ|t−t | − e−γ(t+t ) ,
0
where (
′ t if t < t′
tmin = min(t, t ) = (8.407)
t′ if t′ < t
0 0 0 0
(8.408)
tZmin
1 2γtmin
= ds e2γs = e −1 .
2γ
0
One way to intuitively understand this result is as follows. The double integral over s and s′ is over a rectangle
of dimensions t × t′ . Since the δ-function can only be satisfied when s = s′ , there can be no contribution to the
integral from regions where s > t′ or s′ > t. Thus, the only contributions can arise from integration over the
square of dimensions tmin × tmin . Note also
t + t′ − 2 min(t, t′ ) = |t − t′ | . (8.409)
Zt
1
x(t) = x(0) + ds p(s)
M
0
Z t " # Zt Zs
F −γs F 1
= x(0) + ds v(0) − e + + ds ds1 η(s1 ) eγ(s1 −s) (8.410)
γM γM M
0 0 0
Zt Zs
1
= x(t) + ds ds1 η(s1 ) eγ(s1 −s) ,
M
0 0
8.14. APPENDIX IV : CORRELATIONS IN THE LANGEVIN FORMALISM 423
Figure 8.11: Regions for some of the double integrals encountered in the text.
with v = p/M . Since η(t) = 0, we have
Zt " #
F F
x(t) = x(0) + ds v(0) − e−γs +
γM γM
0 (8.411)
Ft 1 F
= x(0) + + v(0) − 1 − e−γt .
γM γ γM
Note that for γt ≪ 1 we have x(t) = x(0) + v(0) t + 21 M −1 F t2 + O(t3 ), as is appropriate for ballistic particles
moving under the influence of
a constant
force. This long time limit of course agrees with our earlier evaluation
for the terminal velocity, v∞ = p(∞) /M = F/γM .
We next compute the position autocorrelation:
′ ′
Zt Zt Zs Zs
′
′ 1 ′ −γ(s+s′ )
x(t) x(t ) − x(t) x(t ) = 2 ds ds e ds1 ds′1 eγ(s1 +s2 ) η(s1 ) η(s2 )
M
0 0 0 0
′
Zt Zt
Γ ′ ′
= ds ds′ e−γ|s−s | − e−γ(s+s ) (8.412)
2γM 2
0 0
We have to be careful in computing the double integral of the first term in brackets on the RHS. We can assume,
without loss of generality, that t ≥ t′ . Then
′ ′ ′ ′
Zt Zt Zt Zt Zt Zs
′ ′ ′
ds ds′ e−γ|s−s | = ds′ eγs ds e−γs + ds′ e−γs ds eγs
(8.413)
0 0 0 s′ 0 0
′ ′
= 2γ −1 t′ + γ −2 e−γt + e−γt − 1 − e −γ(t−t )
.
Suppose χ̂(ω) ≡ Ĝ(ω) is analytic in the UHP19 . Then for all ν, we must have
Z∞
dν χ̂(ν)
=0, (8.417)
2π ν − ω + iǫ
−∞
where ǫ is a positive infinitesimal. The reason is simple: just close the contour in the UHP, assuming χ̂(ω) vanishes
sufficiently rapidly that Jordan’s lemma can be applied. Clearly this is an extremely weak restriction on χ̂(ω),
given the fact that the denominator already causes the integrand to vanish as |ω|−1 .
Let us examine the function
1 ν −ω iǫ
= 2 2
− . (8.418)
ν − ω + iǫ (ν − ω) + ǫ (ν − ω)2 + ǫ2
which we have separated into real and imaginary parts. Under an integral sign, the first term, in the limit ǫ → 0,
is equivalent to taking a principal part of the integral. That is, for any function F (ν) which is regular at ν = ω,
Z∞ Z∞
dν ν −ω dν F (ν)
lim F (ν) ≡ ℘ . (8.419)
ǫ→0 2π (ν − ω)2 + ǫ2 2π ν − ω
−∞ −∞
The principal part symbol ℘ means that the singularity at ν = ω is elided, either by smoothing out the function
1/(ν − ǫ) as above, or by simply cutting out a region of integration of width ǫ on either side of ν = ω.
The imaginary part is more interesting. Let us write
ǫ
h(u) ≡ . (8.420)
u2 + ǫ2
For |u| ≫ ǫ, h(u) ≃ ǫ/u2 , which vanishes as ǫ → 0. For u = 0, h(0) = 1/ǫ which diverges as ǫ → 0. Thus, h(u) has
a huge peak at u = 0 and rapidly decays to 0 as one moves off the peak in either direction a distance greater that
ǫ. Finally, note that
Z∞
du h(u) = π , (8.421)
−∞
a result which itself is easy to show using contour integration. Putting it all together, this tells us that
ǫ
lim = πδ(u) . (8.422)
ǫ→0 u 2 + ǫ2
19 In this section, we use the notation χ̂(ω) for the susceptibility, rather than Ĝ(ω)
8.15. APPENDIX V : KRAMERS-KRÖNIG RELATIONS 425
(In this equation, the primes do not indicate differentiation with respect to argument.) We therefore have, for
every real value of ω,
Z∞ h ih ℘ i
dν ′
0= χ (ν) + iχ′′ (ν) − iπδ(ν − ω) . (8.425)
2π ν −ω
−∞
Taking the real and imaginary parts of this equation, we derive the Kramers-Krönig relations:
Z∞
dν χ̂′′ (ν)
χ′ (ω) = +℘ (8.426)
π ν −ω
−∞
Z∞
dν χ̂′ (ν)
χ′′ (ω) = −℘ . (8.427)
π ν−ω
−∞