Advanced Topics in Equilibrium Statistical

Mechanics

Lecture 1
Reading: 3.1-3.5 Chandler, Chapters 1 and 2 McQuarrie
This course builds on the elementary concepts of statistical mechanics that
were introduced in Physical Chemistry and provides a more in-depth exploration
of topics related to simple and complex uid systems at equilibrium.
* Discuss course info sheet / oce hrs / web site
* Discuss syllabus
Lets now launch into our introductory lecture, reviewing the basic principles
of equilibrium statistical mechanics:

1. Introduction and Review of Equilibrium Statistical Mechanics

A. Classical uid systems of interest
As chemical engineers or complex uids oriented materials scientists, you will
repeatedly encounter a variety of uid systems of varying complexity, e.g.,
Atomic liquids (1
A): e.g. condensed noble gases A, Xe, ...
Diatomic liquids (2-3
A): e.g. O2 , CO, N2 , ...
Molecular liquids (2-5
A): e.g. H2 O, CO2 , CH4 , ...
Polymers (50-500
A): e.g. polyethylene, polystyrene, ...
Colloids (0.1-10 m): e.g. latex particles, clays, ...
All of these systems can be described quite accurately at thermodynamic
equilibrium by using the principles of classical (as opposed to quantum) statistical mechanics. We shall primarily focus on such systems and on the condensed
1

liquid state, rather than the gas state and the solid state (which is treated in
courses on solid state physics). Note the coarse-graining of our cartoons
as we move up in lengthscale.

B. Averages and Ensembles

Let us now review the way in which averages are computed in statistical mechanics. Suppose that within a large volume of a gas or liquid, you had a smaller
measurement volume in which you could count the number of particles.

Imagine that you made a large number M of such measurements over a long
time period and the system was at equilibrium. Then, you could compute the
time average of the number of particles N inside the permeable barrier by
M
1

Ni
M i=1

= lim
N

where Ni is the number of particles in ith observation.

We shall see later in the course that such time averages, constructed by following a particular dynamical trajectory of one system, are the most convenient
way calculating equilibrium averages in computer simulations. It is of course
also what we do in experimental measurements on uids!
There is another convenient way of averaging in statistical mechanics. A
basic tenant of statistical mechanics is that after a long enough time, all microscopic states (quantum or classical) in a given system will be visited. Thus,
we should also be able to construct an ensemble average


P N

N  =

where P is the probability of the th microstate occurring and N is the

number of particles inside the barrier in this microstate. This is called an
ensemble average because the right hand side can be constructed by making

one measurement on a large number (ensemble) of identical equilibrium system at the same time, weighting the observation of a particular microstate by
its frequency of occurrence.
The ergodic hypothesis of statistical mechanics says that
=
N 
N
This is generally met in most systems that you will encounter in your scientic
or engineering careers. However, there are notable exceptionse.g., glasses.
Generally, systems are ergodic when there is largely unrestricted ow and access to the various microstates throughout their dynamical trajectories. This
ow is often closely associated with the trajectories becoming chaotic due to
nonlinearities in the dynamical equations. Thus, it turns out that there is a close
connection between statistical mechanics (a relatively old subject) and one of
the hottest areas of contemporary theoretical research: nonlinear dynamics.
Dierent ensembles are convenient for calculations in statistical mechanics,
depending on what constraints are imposed on a system. To mimick an isolated
system that cannot exchange energy or mass with its surroundings, we use the

C. Microcanonical Ensemble (N , V , E)
To begin we dene a quantity known as the microcanonical partition function:
(N, V, E) = number of states with N, V, and energy E xed
Next, we invoke a fundamental postulate of statistical mechanics that all
states of the same energy are equally likely at equilibrium. In other words, these
states are characterized by a uniform probability distribution, P . It follows
that

1
states of ensemble
P = (N,V,E)
0

/ states of ensemble
Using this distribution function, we can compute ensemble averages of some
property G:



G =
P G

where G is the value of G in microstate . The connection to thermodynamics

is via Boltzmanns formula for the entropy:
S = kB ln (N, V, E)
We shall see that in each ensemble there is such a connection between a thermodynamic quantity (in this case the entropy S), referred to as a thermodynamical
potential, and the logarithm of a partition function (in this case ).

We can now use standard thermodynamics relations to get all the other
equilibrium properties. Recall that the internal energy E in an isolated system
satises
dE = dq dw
1st Law
= T dS p dV
closed system
T =

E
S

=
S
p = T V

V,N
ln
E


E,N

or

1
1 S
= kB
kB T
E


V,N


V,N

gives temperature

ln
= 1 V


E,N

gives pressure

D. The Canonical Ensemble (N, V, T)

This is another very important ensemble. Here we consider systems closed to
mass transfer, but open to energy ow and in contact with a heat bath at
temperature T .

In this ensemble, dierent states can have dierent internal energies and the
probability of occurrence of a state of energy E is now not uniform, but
Boltzmann-distributed:
P
Q(N, V, T )

=e Q

= eE

canonical partition function

Again, notice that T , rather than E, is constant and E varies

with state

in the ensemble. Also, P is evidently normalized such that P = 1. Thus,

G = P G gives the ensemble average of a quantity G as before.

Now, lets make the thermodynamics connections. The average internal

energy should be:

E =
E
= E eE /Q

ln Q(N,V,T )
1 Q
=

= Q

N,V

But, recall that the Helmholtz free energy is dened by

A(N, V, T ) = E T S
dA
S

= dE T dS SdT
= SdT pdV 1st Law
= A
T
V,N

Thus,
E = A + TS = A T

A
T


V,N

= kB T 2

ln Q
T

Notice that the last two expressions can be viewed as a 1st order, inhomogenous,
linear ordinary dierential equation for A as a function of T at xed V and N .
A particular solution is
A(N, V, T ) = kB T ln Q(N, V, T )
which represents the fundamental connection to thermodynamics in the canonical ensemble. In this important ensemble A is thermodynamic potential that is
logarithmically related to the canonical partition function Q. Notice that there
is also a homogeneous solution const. T to the dierential equation, but the
constant coecient must vanish or the third law of thermodynamics would be
violated (vanishing entropy of a perfect crystal at T = 0).
The pressure can be computed in the canonical ensemble according to


ln Q
A
= kB T
P =
V N,T
V
N,T
and similar expressions can be used to calculate all other thermodynamic quantities.
Finally, we return to the entropy expression

A
S=
T V,N
Lets show that this is equivalent to the Gibbs entropy formula
S = kB

P ln P

which holds for any ensemble! The manipulations are as follows:

= Q1 eE [E ln Q]
=
E Q1 (ln Q)Q
= T1
E + kB ln Q
= EA
S
T

 P ln P
 P ln P
kB P ln P

There are many other ensembles that one can choose besides the microcanonical and canonical. For example, it is natural in many experiments to maintain
control over the variables P, T, N . The (P, T, N ) ensemble is referred to as the
isothermal-isobaric ensemble and you will have opportunity to work with it
in your homework exercises. We now turn to consider the most important other
one: the grand canonical ensemble.

Lecture 2
Reading: 3.63.7, 7.1 Chandler, Chapter 3 McQuarrie
Recap: ergodic hypothesis
ensembles, kB T
ensemble
micro. can.
(NVE)
canonical
(NVT)
grand can.
(V T )

part. fnc

(N, V, E) =
Q(N, V, T ) =

eE

QG (, V, T ) =
S = kB

thermo connection
S = kB T ln

e(E N )

P ln P

A = kB T ln Q
pV = kB T ln QG
Gibbs entropy formula

E. The Grand Canonical Ensemble (, V, T )

In the grand canonical ensemble we consider a system open to both mass transfer
and energy ow. This can be most easily realized if our system is part of a larger
system with a mass and energy permeable boundary separating the two.

Thus, N and E both uctuate from state to state. We control the N uctuations by imposing a chemical potential . Recall from the thermodynamics
of open systems
dE = T dS pdV + dN
dA = SdT pdV + dN,
etc.
that is the conjugate thermodynamic variable to N . Indeed, (T, S), (p, V ),
and (, N ) are all conjugate pairs.
One might guess that the appropriate distribution is thus
P = e(E N ) /QG
where
QG (T, V, ) =

e(E N )

is the grand partition function, which is usually given the symbol (but I hate
to draw this on the board!). Indeed, one can prove that
 this is the distribution
that maximizes the Gibbs entropy formula S = kB P ln P subject to the
following constraints:

1.
P = 1

2.
See Chandler 3.7
E P =
E

3.
N P =
N 
You may also want to review the material in Chapter 1 of McQuarrie on
the method of Lagrange multipliers. With this choice of P , the Gibbs formula
gives

S= kB P [E + N ln QG ]
= + T1
E T
N  + kB ln QG
or
kB T ln QG = T S E + N
Now, recall the following denition of the Gibbs free energy:
G = A + pV = E T S + pV
However, G is an extensive function that can be expressed in terms of its partial
molar property .
G 
G = N = N
N T,p
Thus,
kN T ln QG = T S E + G
or
pV = kB T ln QG (T, V, ),
7

which serves as the fundamental connection to thermodynamics in the grand

canonical ensemble. The combination pV plays the role of the thermodynamic
potential in this ensemble.
The average energy is




E P = Q1
E eE +N
E
E =
G

Thus,
Q1
G
Similarly
N=

QG ,V =
E
N 

N P =


1
ln QG T,V

F. Fluctuations
It is important to appreciate that the variables that are not held xed will
uctuate in the various ensembles. We can often relate these uctuations to
higher-order derivatives of the thermodynamic potentials. Such formulas are
particularly useful in computer simulations. As an example, consider energy
uctuations in the canonical ensemble:
E E
E

(E)2  =
(E
E)2 
=
E 2  2
E
E +

E2 
=

E 2 
E2 
= E2 P ( E P )2
1 2
1
2
=Q
2 Q ( Q Q)
2

=
2 ln Q)N,V

=

E

Now, recall that

E
Cv =
T


const. V heat capacity
N,V

T
kB T 2

E
2
kB T Cv =
=
(E)2 

N,V

Thus, in the canonical ensemble, we can calculate the heat capacity via
Cv =

1
2
kB T 2
(E) 

This is referred to as a uctuation formula for the heat capacity. From the
expression on the right hand side, we might navely think that
(E)2  N 2 ,
as one increases the size of a system. However, Cv is extensive (recall that
Cv = 32 N kB for a structureless ideal gas), so

(E)2  N or
(E)2 1/2 N 1/2
Alternatively,

(E)2 1/2
N 1/2
1

1/2 0

E
N
N

as N . Thus, the uctuations become vanishingly small for systems of

macroscopic size, N 1023 !

Advanced Topics in Equilibrium Statistical

Mechanics

2. Classical Fluids
A. Coarse-graining and the classical limit
For concreteness, lets now focus on a uid phase of a simple monatomic substance, e.g. Argon.

We can write generally that

Q(N, V, T ) =

eE

where the sum is over all quantum states that the electrons and nuclei of
the atoms can be in. This is a hard sum to do, since it requires solving a
many-body quantum mechanics problem to nd the states. Instead, we use
our physical intuition that the light electrons are moving much faster than the
heavy nuclei, so that for a given nuclear conguration, the quantum states of the
electrons are nearly at local equilibrium (Born-Oppenheimer approximation).

This suggests the rewriting

 E 
ER,i
Q= e
R
e

i



= e ER
R e ER


where R is a sum over nuclear
 states (i.e. position and momentum treat
with classical mechanics) and i eER,i is a sum over electronic states (use
quantum mechanics).
R , the eective energy (actually free
The second line is a denition of E
energy, since it is dependent) of just the nuclear states with the electronic
states averaged out. This is an example of the notion of coarse-graining in
statistical mechanics
small objects
(electrons, nuclei)
fundamental
interactions

larger objects
(atoms)
eective
interactions

coarsegrain

We have now made a lot of progress in getting rid of the quantum mechanics
in our problem, assuming that we can evaluate the eective interactions between
R . Indeed, this is what ab initio quantum chemistry tries
atoms that enter E
to do. There are now a large number of user-friendly software packages for
computing the eective interactions between atoms and molecules that are very
useful in deducing classical descriptions of uids.

B. Classical phase space averages

What now are the classical states R that we are supposed to sum over? In
classical mechanics, the state of a particle is determined by specifying its position
r = (x, y, z) and momentum p = (px , py , pz ). (We can then gure out its future
state by integrating Newtons equation F = ma = p.)
Thus, for an N -atom
gas,
R (r1 , . . . , rN ; p1 , . . . , pN )
 
 

and R is a 6N -dimensional phase space.

We expect that





Q=
e ER C dr1 . . . drN dp1 . . . dpN e ER
R

where C is a constant prefactor. We adopt the shorthands:

(r1 , . . . , rN ; p1 , . . . , pN ) = (rN , pN )

Note that we use a slightly dierent notation than that of Chandler.

dr1 . . . drN dp1 . . . dpN = drN dpN
Also,

R = E(r
N , pN ) H(rN , pN )
E

where H is known as the Hamiltonian.

The Hamiltonian (dont confuse with enthalpy!) is the sum of the eective
kinetic energy of the atoms (nuclei) and the eective potential energy associated
with their mutual interactions. We can separate the two as
H(rN , pN ) = K(pN ) + U (rN )
Finally, what is C? There are various ways of deriving it, the simplest of
which is the work out Q quantum mechanically for the special case of U = 0
(non-interacting boson or fermion gas) then take the classical limit (h 0,
0). For each momentum degree of freedom this yields a factor of 1/h (Plancks
constant) and an overall factor of 1/N ! to correct overcounting of congurations
corresponding to dierent label permutations of indistinguishable particles. (See
section 4.67 of Chandler for details.) Thus, the canonical partition function in
the classical limit is:


N
N
1
N
dr
dpN eH(r ,p )
Q(N, V, T ) =
N !h3N
Notice that the integrals over particle positions are conned the the volume V of
the uid, while the momentum integrals are unrestricted in the 3N -dimensional
momentum space.
Expressions such as this are easily generalized to multi-component uid systems. For example, the canonical partition function of a binary mixture of A
and B atoms is




NA
NB
1
NA
NB
NA
dr
dr
dp
dpNB eH(r ,...,p )
Q=
3N
3N
A
B
NA !NB !h
h
Returning to the one-component system, it is now convenient to dene a
phase space distribution function
N

eH(r ,p )

drN dpN eH


which is normalized such that drN dpN f = 1. Ensemble averages in the
canonical ensemble are thus dened by


P (rN , pN ) drN dpN f (rN , pN )P (rN , pN )
f (rN , pN ) = 

for any property P (rN , pN ) of interest. For example, the average energy is


E= H(rN ,pN ) = drN dpN f H
1 Q
= Q
as before!

V,N

The expression A(N, V, T ) = kB T ln Q(N, V, T ) remains the fundamental thermodynamic connection for this ensemble.
This is about as far as we can go without specifying the form of the Hamiltonian H. Evaluating any thermodynamic property or Q seems to involve doing
6N integrals, where N 1023 ! Indeed, equilibrium statistical mechanics is really
all about the evaluation of high dimensional integrals. Life gets a bit easier if we
think physically about the form of the eective Hamiltonian that is obtained
by removing the electronic degrees of freedom. We expect
K(pN ) =

N


p2i
, p2i = p2ix + p2iy + p2iz
2m
i=1

for the kinetic energy where m is the eective mass of the atom (m
mnucleus ). Then, noting that
H=

N


p2i
+ U (rN ),
2m
i=1

the phase space distribution function f (rN , pN ) thus factors as

N


N
N
f (r , p ) =
(pi ) P (rN )
i=1

where
prob. of observing
system at conguration
space point rN

P (r ) = 

eU (r )
drN eU (rN )

(pi ) = 

epi /2m
,
2
dpi epi /2m

and
2

Maxwell-Boltzmann
momentum distribution

and where again, pi |pi |.

The Maxwell-Boltzmann distribution gives the probability density of observing a particle with (3-d) vector momentum p in an equilibrium uid. It is very
important to note that all particles (atoms) have their momenta distributed
independently, regardless of the uid density (e.g. liquid, gas, or solid) or form
of U (rN ).
4

Gaussian integrals, such as those appearing in the MB distribution are very

important in statistical mechanics. Here we need
 1/2

2
1
2
dx e 2 ax =
, a>0
a

The denominator of the MB distribution is thus


3/2



2m
(p2x +p2y +p2z )/2m
=
dpx dpy dpz e

Note that all three components of a particles momentum p are themselves

independently distributed (equipartitioned):
3
(p)= =1 g(p )
2

g(px )=
Thus, e.g.
p2x 


=

dpp2x (p)

epx /2m
(2m/)1/2

dpx g(px ) = mkB T

This implies that the mean-squared velocity component of any atom in the uid
is given by vx2  = kB T /m and that
p2  = p2x  + p2y  + p2z  = 3mkB T
We can now draw some conclusions about a classical uid at equilibrium:
At the same T , increasing m implies smaller average RMS velocities v 2 1/2
The kinetic energy is equally partitioned among the three translational
modes at equilibrium
2


K= N p
2m =
3
= 2 N kB T

N
2m

3 mkB T
x, y, z

Notice that when we calculate with the MB distribution, sometimes it is

handy to mix spherical polar and cartesian coordinates. For example, the average speed v = |v| of a molecule is

1
1
v= m
dp|p|(p)
|p| = m

p2 /2m
dp|p|e
1
=m
(2m/)3/2
Since the integrand depends only on p = |p|,



 2
2
dp =
dp p
d
d sin
0
0


0
4

v =

4
m(2m/)3/2




dp p3 ep

2m2 / 2

/2m

= [8kB T /(m)]1/2

Finally, we can use our Gaussian integral formulae to simplify Q:



N
1
N K(pN )
Q=
dp e
drN eU (r )
N !h3N



(2mkB T )1/23N

or
Q=

1
N !3N
T

drN eU (r

where T h/ 2mkB T is the so-called thermal wavelength. It is thus

common to write:
1
Q(N, V, T ) =
Qc (N, V, T )
N !3N
T

N
where Qc (N, V, T ) drN eU (r ) is the congurational partition function.
Evidently, Qc is where the remaining work has to be done in evaluating the 3N
coordinate integrals.

Lecture 3
Recap: Classical Limit, canonical partition function
Q=

eE Q=

1
N !3N
T

Qc (N, V, T ), Qc =

T = h/ 2mkB T

P f (rN , pN )= P (rN )

N 
3

i=1 =1

g(p1x )=

p2 /2m

1x
e
(2m/)1/2

drN eU (r

thermal wavelength
g(pi ), P (rN ) =

1 U (rN )
e
Qc

Maxwell-Boltzmann Distribution

What is the signicance of T ? Recall from quantum mechanics that the

DeBroglie wavelength of a particle with momentum p, = h/p, is the wavelength associated with the wave mechanics picture of a quantum particle. In
quantum mechanics we represent localized particles by wave packets:

where the width of the wave packet is denoted by x. This is the scale
of the uncertainty in position of the quantum particle at some instant in time.
The Fourier transform of such a wave function has a broad peak centered at
1
, which immediately gives the Heisenberg
k = 2/ with width k p/h x
uncertainty principle: xp h. This principle relates the characteristic scales
of position and momentum uncertainty of a quantum particle. In an equilibrium

1/2
system, we could thus estimate x by computing p p2 M B kB T m. It
follows that

length scale over which

x h/ kB T m T
atomic positions are smeared
With this interpretation of T , a reasonable way to assess the importance
of quantum eects is to compare T to the characteristic size (or spacing in a
liquid or solid) of the atoms or molecules that we wish to describe classically.
Thus, quantum eects should be negligible when T
, where is an atomic
or molecular diameter. For example, argon at its triple point temperature,
A, whereas 3.5
A. Thus argon should be quite
T = 84K, has T 0.30
accurately described by by classical mechanics at this temperature.
We will now nish this section by extending the classical limit to the grand canonical ensemble:
In the grand canonical ensemble, the phase space distribution also depends
on N , which uctuates. Thus, in the classical limit:
f (rN , pN ; N ) =

N
N
1
1
eN H(r ,p )
3N
QG (, V, T ) N !h

where the grand partition function can be written:

QG (, V, T )=

eN

N =0

Q(N, V, T )
 

1
N !3N
T

Qc (N,V,T )

zN
Qc (N, V, T )
N!

N =0

where z e3 is referred to as an activity. Note that this is consistent with

T
the thermodynamic sense of the word.
Averages follow naturally
P (rN , pN ; N ) =



drN

dpN f (rN , pN ; N )P (rN , pN ; N )

N =0

and thermodynamic properties follow (as before) from:

pV = kB T ln QG

E=
ln QG

N =

ln QG

,V
,V

+
=

ln QG


ln QG
ln z


,V

,V

C. Intermolecular Potentials
The above results, while restricted to systems that obey classical mechanics, are
exact. However, any calculations based on these formulae require an explicit
form for the eective potential energy U (rN ) of interaction among atoms or
molecules.
It is often the case that the biggest limitation on theoretical calculations
for uids is obtaining an accurate representation of U (rN ). Ab initio quantum
chemical methods are advancing rapidly, but many systems (hydrogen-bonding
uids, molten metals and salts) remain challenging for the purpose of parameterizing U (rN ).
Most calculations on liquids and gases are based on the notion of pairadditive potentials, namely
U (rN )
=

1
2

N

N

i = j




i <

u(ri , rj )

u(ri , rj )

N (N 1)
pairs
2

pairs

In the case of nearly spherical atoms like argon, or molecules like methane, the
pair potential u depends only on the distance rij |ri rj | between a pair
of atoms. Thus, u(ri , rj ) = u(rij ).
This would seem restrictive, but non-spherical molecules can be treated in
an interaction-site model by superposing spherically symmetric potentials at
dierent atomic sites in a molecule. For example, in the case of the nitrogen
molecule we can express the potential energy as a sum of spherically symmetric
site-site interactions:

N

N

2

2


i <

u (|ri rj |),

where u (r) is a 2 2 matrix of site-site potentials.

What are reasonable forms for the pair potential u(rij )? At short distances,
neutral molecules repel strongly due to electronic overlap. The simplest model
is thus

, r < d Hard-sphere
u(r) =
0, r > d uid
where d is a molecular diameter. (Note that it is customary to dene the zero of
potential energy for isolated molecules separated by a large distance.) A problem
with the hard sphere model is that while it has a uid-solid transition, there are
no attractive forces necessary to induce gas-liquid transitions. A slightly more
realistic model, which describes uids with both types of phase transitions, is

, r < d
squareu(r) = , d < r < d well
0, r > d
uid

Mathematically, however, we expect u(r) to be a continuous, smooth function. Neutral molecules at large separations have dipole-induced dipole attractions that vary as r6 . A robust form that ts experiments on argon, methane
and other simple quasi-spherical molecules is
 

12  6
u(r) = 4

r
r

Lennard-Jones
6-12 potential

Finally, we note that eective ion-ion pseudo potentials for liquid metals,
e.g. sodium or potassium, look like

The pair approximation for U (rN ) is also often seriously in question in such
systems!

10

Advanced Topics in Equilibrium Statistical

Mechanics

3. Theory of Gases
A. The Ideal Gas
As we shall see, gases are much easier to deal with than liquids, because the low
density of gases makes collisions and pair interactions of molecules infrequent.
The strategy will be to derive an expansion in powers of particle density known
as the virial expansion. The leading term, accurate at innite dilution, is the
ideal gas and follows by setting U (rN ) = 0.
We shall begin by pursuing the thermodynamic properties of a monatomic
ideal gas in the canonical ensemble.


Qc (N, V, T ) = drN eU = V N
Q=

VN
N !3N
T

A = ln Q = ln N ! + N ln(3T /V )
We are interested in the thermodynamic limit of this expression, which corresponds to considering a system of macroscopic extent, i.e. with Avogadros
number of atoms. We take this limit by simultaneously taking N ,
V , while holding the average particle number density = N
V constant.
Stirlings asymptotic approximation for the logarithm of large factorials, ln N !
N ln N N + . . ., gives
A N ln N N ln V + N (ln 3T 1)
N ln(N/V ) + N (ln 3T 1)
The ideal gas equation of state follows from:

A
p= V
N,T
p= N

1
V

Thus, the ideal gas equation of state can be written in any of the alternate
forms: p = , pV = N kB T, pV = nRT where n is the number of moles of
atoms.
Next, we repeat this calculation in the grand canonical ensemble:



zN
(zV )N
Qc (N, V, T ) =
QG (, V, T )=
N!
N!
N =0

N =0

= ezV
The thermodynamic connection is pV = ln QG = zV or p = z = e 3
T .
We thus have the pressure expressed as p(z, T ), but a conventional equation of
state is of the form p(, T ). Thus, we need an expression relating the activity z
to the average density , i.e. we need the function z(). Recall that
N  =

z 
ln QG 
=V
=Vz
,V
ln z
ln z ,V


So z = N
V is the functional relationship between z and in the ideal gas
limit. It follows immediately that the equation of state for an ideal gas in the
grand canonical ensemble is given as before by p = .
Notice that we needed Stirlings approximation in the canonical ensemble,
but not in the grand canonical ensemble. Thus, results from the two agree only
in the thermodynamic limit! This, however, is as expected, since uctuations in
the two ensembles are dierent for systems of nite size and these will contribute
small corrections to thermodynamic quantities that go to zero in the limit of an
innite system.

Lecture 4
Reading: McQuarrie Chp. 12, Andersen handout

B. Non-ideal gases Cluster & Virial Expansions

Now we are ready to talk about the problem of a real uid with intermolecular interactions.
Even with the pairwise additivity assumption, the congurational integral cannot be done exactly


Qc = drN exp
u(rij )
i<

In the case of a dilute gas, however, we might hope to develop an expansion

in powers of density to correct the ideal gas equation of state. We expect an
expansion of the form
2

p =

virial
expansion

Bi (T )i

i=1

where the Bi (T ) are known as virial coecients and evidently B1 (T ) = 1 so

that the leading term reproduces the ideal gas law.
The virial expansion is a special case of a so-called cluster expansion in
statistical mechanics. Although we will work out the virial case, it is important to know that there is an extensive literature on this subject that you can
read about in the review article by H.C. Andersen (distributed to you), your
McQuarrie text, or in the Hansen-McDonald text listed in the reference list for
the course.
The idea behind cluster expansions is to arrange a perturbation series so
that successive terms involve integrals just involving 2, 3, 4, . . . , particles and
that each term in the series remains nite in the thermodynamic limit. This is
most natural in the grand canonical ensemble:
QG =


zN
drN
N!

N =0

eu(rij )

1i<jN

where we have replaced the exponential of a sum by a product of exponential

factors. Notice that the function eu(r) 1 for r , i.e. when a pair of
particles gets very far apart, which is not good if we want nite integrals for
V . Instead, we write

eu(r) 1 + f (r)
where f (r) = eu(r) 1 is the so-called Mayer f -function. This function has
the property that f (r) 0 for r , so volume integrals over powers of f
will remain nite even for V . With this simple change:



ZN
[1 + f (rij )]
Qc =
drN
N!
N =0

1i<jN

The rst few terms are:

2
Qc = 1 + zV + z2! d1 d2[1 + f12 ]


3
[1 + f12 ][1 + f13 ][1 + f23 ]
+...
+ z3! d1 d2 d3



1 + f12 +f13 + f23 + f12 f13 + f12 f23
+f13 f23 + f12 f13 f23
3



where we have introduced some obvious shorthands. In particular, d1 dr1
and fij f (rij ).
Our next step is to introduce a Mayer cluster diagram shorthand for the
various terms:

Each point (vertex) represents a factor of z and each bond represents a

factor of f . The rules for evaluating one of the diagrams (graphs) are as follows:
1. Assign a set of labels to the dots
2. Assemble the appropriate product of z and f factors
3. Integrate over the labeled positions
4. Divide by the symmetry number S of the graph
The symmetry number S is the number of diagrams generated that are topologically equivalent to the rst, when the labels of the diagram are permuted
all possible ways.
Lets try this out:

S=1

= S1 z V dr1
= zV
1 2

= S1 z 2 V dr1 V dr2
2 1 S=2

z2 2
2 V

1 2
Sz
V

1 3
Sz

dr1

dr1

dr2 f (r12 )

dr2

dr3 f (r12 )

,S = 2

Note that

1
S

1
2

1
3!

3, one for each term. Finally, we look at

1 3
z
dr1 dr2 dr3 f (r12 )f (r23 )f (r13 )
S
All 3! permutations are equivalent, so S = 6.
Now, notice that the disconnected graphs

all grow like V 2 or V 3 for V , while the connected graphs grow only like V
for V . For example,


2
= z2 dr1
dr2 f (r12 )



indep
of
r
for
V

1

z2
2 V

dr2 f (r12 )


nite

The disconnected graphs look like they are going to be a problem in the thermodynamic limit, but when one takes the logarithm and re-expands, magically
all the disconnected diagrams disappear!
ln QG = +

= sum of all distinct, connected

diagrams with z-vertices and f -bonds.
= pV
Notice that, as in the previous series for QG , there is at most one f -bond
between each pair of points. By dividing both sides by a factor of volume V ,
this can be interpreted as an activity expansion:
p =

bi z i

i=1

where bi can be interpreted as 1/V times the sum of all distinct, connected
diagrams with i 1-vertices and f-bonds. The bi are now O(1) for V .

= V1 dr = 1
b1 = V1 {}
b2 =
..
.

1
V

}=

1
V

1
2

1

2V


dr2 f (r12 )

drf (r)

We can turn this into a density expansion by guring out the relation between
z and .



<N >
V

1 ln QG
V ln z


= z p
z
,V

or

z ln QG
V
z

,V

,V


= z ibi z i1
i=1

ibi z i

i=1

Now, to nish the calculation,

we simply invert the series to get z in powers of

, then plug into p = i bi z i and re-expand in powers of . This gives the
virial expansion


p =
Bi (T )i
i=1

where

B1 (T )
B2 (T )

= b1 = 1
= b2 = V1 {

B3 (T )
..
.

= 2 V1 {

} = 12

} = 13

dr2

dr2 f12

dr3 f12 f13 f23

Notice that the b3 diagram has dropped out of the B3 expression. This is
because it contains an articulation vertex, namely a vertex that when removed
disconnects the diagram. Magically, the only diagrams appearing in the Bi are
those that are both connected and free of articulation vertices!
This exercise has illustrated an interesting and elegant feature of cluster
expansions, namely that we can represent a complicated and unwieldy perturbation series by a series of diagrams and that manipulations of the series can
often be described succinctly in terms of topological statements about the remaining diagrams. Similar features are present in perturbative treatments of
eld theories in theoretical physics, where the relevant diagrams are referred to
as Feynmann diagrams.
It turns out to be a lot of work to evaluate the successively higher terms:
Nevertheless, for special cases, e.g. hard spheres, people have worked this
out to very high order. Lets now consider this special case of the hard sphere
potential.
Recall that the hard sphere model has u(r) = , r < d; u(r) = 0, r > d.
This implies

1, r < d
u(r)
1=
f (r) = e
0, r > d

Coe
B2
B3
B4
B5
B6
B7

# diags
1
1
3
10
56
468

# integrals (3-d)
1
2
3
4
5
6

So,
B2 (T )

dr2 f (r12 ) = 12 dr12 f (r12 )

d
2d3
= 12 4
dr r2 (1) =
3
0
= 12

which is 4 times the particle volume Vp = 43 (d/2)3 = d3 /6 and is T independent! Note that for more realistic potential functions, the virial coecients are
normally temperature dependent. The leading term in the virial expansion for
the hard sphere uid is thus
p = +
or

2d3 2
+ O(3 )
3

p
= 1 + 4 + O( 2 )

where
=

Vp N
= d3
V
6

is the particle volume fraction.

Thus, the virial expansion for the hard sphere uid is a series in packing
fraction. Coecients B3 and B4 have been worked out analytically, and B5 B7
numerically.
p
= 1 + 4 + 10 2 + 18.365 3 + 28.24 4 + 39.5 5 + 56.5 6 + . . .

To how high a density can we trust this truncated series? The close-packed

face-centered-cubic
(fcc) crystalline density of hard spheres is 0 = 2/d3 , or

0 = 2/6 0.7405, and the hard sphere uid crystallizes into a solid at
LS 0.670 0.50. Since the virial series is expanded about 0, it is a
representation of the uid phase equation of state, so it is of little interest for
> LS 0.5. Lets see how the series does in the physical range < 0.5:

It appears that the virial series converges surprisingly well over the physical
range of . The solid curve is a very accurate equation of state developed by
Carnahan and Starling. It was derived by noting that the virial series is very
close to


p
1+
(i2 + 3i) i

i=1
(B2 = 4, B3 = 10, B4 18, B5 28, B6 40, B7 54)
This can be summed to
p
1 + + 2 3
=

(1 )3

CarnahanStarling EOS

which is believed very accurate over entire uid range. Now, using

A
A
= ()()
p=
V
V
T,N

or
p




A
N

A ex
N

, 1 =

A id
N

where Aex = A Aid denotes the excess Helmholtz free energy. We can thus
integrate over the packing fraction, starting from the ideal gas (dilute) limit to
obtain:
A ex
=
N

0

d 



1 p 
(4 3)
(
)

1
=

(1 )2

C-S excess
Helmholtz F.E.

A few nal notes about virial expansions:

1. For more realistic potential functions u(r), e.g. the L-J 612 potential,
the Bi (T ) are T -dependent and must be computed numerically. This is a
lot of work to go beyond 2nd or 3rd order in the series.
2. More realistic models of uids also have gas-liquid transitions, which cannot be accurately described by virial expansions with limited radii of convergence.
3. For dilute and moderately concentrated gases, virial (and more generally
cluster) expansions are the way to go. A totally dierent approach is
needed for the liquid state.

Advanced Topics in Equilibrium Statistical

Mechanics

Lecture 5
Reading: Chandler 7.2, 7.3, 7.4, 7.5
Recap:
dilute to moderately dilute gases treated by virial expansions about =
0 (z = 0) state.
easily extended to molecular uids and dilute mixtures.
Also applicable to dilute solutions of solute in solvents, including polymers
and colloids, but not electrolyte solutions as we shall see.

4. Liquid State Theory

A. Distribution functions and radiation scattering
It is clear that virial expansions, i.e. expansions about the = 0 ideal gas
reference state, are not very useful in condensed liquid phases. Instead, we
make note of the following central idea in liquid state theory known as the van
der Waals picture of liquids and exploited heavily by Chandler, Andersen, and
Weeks in the 1970s:
* Due to the harsh repulsions between molecules at close range, the
structure and thermodynamics of liquids are determined primarily
by packing eects that depend on molecular shape.
We shall see that in liquids of nearly spherical molecules like A or CH4 , this
suggests new expansion about a dense hard sphere reference system, where the
expansion is in the strength of the relatively weak attractive interactions.

To discuss liquid structure, we need to introduce the concept of reduced

distribution functions. For simplicity, we work in the canonical ensemble. Recall that at equilibrium, particle positions and momenta are independently distributed, the former described by:
N

P (rN ) =

eU (r )
1 U (rN )
=
e
Qc
drN eU (rN )

This is the joint probability distribution of observing particle 1 at r1 , . . . , and

particle N at rN . Suppose we now integrate over r3 , r4 , . . . , rN to create a
reduced distribution function


P (2) (r1 , r2 ) = dr3 . . . drN P (rN )
This is the joint probability distribution of observing particle 1 at r1 and particle 2 at r2 . However, if we were observing a uid, we would not be able to
distinguish 1 and 2 from any other particles. Thus,
(2) (r1 , r2 ) = N (N 1)P (2) (r1 , r2 )
gives the joint probability distribution of observing any particle at r1 and any
other at r2 .
We can now generalize this to dene the n-particle reduced distribution function by:



N!
(n)
n+1
n+2
dr
dr
. . . drN P (rN )
(r1 , . . . , rn ) =
(N n)!
The functions are normalized such that

drn (n) =

N!
(N n)!

n = 1:
(1) (r1 ) should clearly be independent of position r1 in a uid that is homogeneous (uniform) on average. This is true in the gas and liquid states, but
clearly not in a crystal phase. To be normalized properly, (1) must correspond
to the average density
N
(1) (r1 ) =
V
n = 2:
(2) (r1 , r2 ) should depend only on |r1 r2 | r12 in a homogeneous, isotropic
uid (liquid or gas state). Hence (2) (r1 , r2 ) = (2) (r12 ).

For r12 , the two points in the uid are very far away, so there should
be no correlation in density between the two points. Thus
lim (2) (r12 ) =

r12

N (N 1)
2 .
V
V

It is convenient to normalize according to

g (2) (r12 ) = g (2) (r1 , r2 ) (2) (r1 , r2 )/2
so that g (2) 1 for r12 . This function is known as the radial distribution
function. We can generalize to higher order:
g (n) (r1 , . . . , rn ) (n) (r1 , . . . , rn )/n .
For the special case of n = 2, it is common to drop the superscript (2). Hence
we use the notation g(r) to describe the radial distribution function. Also, it is
common to dene
h(r) = g(r) 1

pair correlation function

as a function that 0 for r , hence focusing attention on particle separations r for which correlations exist. A nice physical way to think about g(r)
is:
average density of particles at r,
g(r) = given that a tagged particle is
placed at the origin

So, how can one measure g(r) in a liquid? In a computer simulation, we

will see that g(r) can be computed directly. Experimentally, we turn to a static
radiation scattering experiment. Because we want to study simple liquids with
d1
A, we choose x-rays or neutrons as the radiation source (wide angle
scattering, WAXS, WANS) since they have short wavelengths. We use SAXS or
SANS (small angle scattering) for studying polymers on 10
A to 100
A scales

and light scattering for studying polymers and colloids on 100

A1m scales.
The setup looks like

For weak, elastic scattering, |kin | |kout | and the momentum transfer or
scattering wavevector is given by
2
2
k = 2kin
2kin
cos
k kout kin , k 2 = k

2
k = 2kin 1 cos = 2kin 2 sin (/2) = 2kin sin /2

or
k=

4
in

sin 2

where in is the wavelength of the incident radiation. Thus, k can be varied by

changing the angle of the detector at xed radiation wavelength in . In the
Born approximation, the intensity of light scattered at angle is:
I()

|f (k)|2
S(k)
  
  
form factor
structure factor

The form factor is related to the Fourier transform of the particle shape, so
contains only single-particle information. The structure factor contains the
information on particle-particle correlations:
S(k) = N 1

N 
N

eik(rj r
)
j=1 =1

Next we separate the j =  terms from the

=  terms. Because the particles
are indistinguishable, each term is the same on average:

S(k)

= N 1 N eik0 + N 1 N (N 1)eikr12 

= 1 + N 1

Finally, noting

S(k)

dr2 eikr12 N (N 1)


dr1


dr3 . . .


drN P (rN )


(2) (r1 ,r2 )

= 1 + N 1 dr1 dr2 eikr12 2 g(r12 )



dr1 dr2 f (r12 )



drf (r)
= dr1 dr12 f (r12 ) = V

= 1 + dreikr g(r)

Thus, the structure factor is related to the radial distribution function by

a Fourier transform! Given S(k) from a scattering measurement, we can thus
invert the Fourier transform to compute g(r):

1
g(r) =
dk eikr [S(k) 1]
(2)3
This three-dimensional integral over wavevectors k can be reduced to a onedimensional integral by performing the integral in spherical polar coordinates
with north pole selected along r:
g(r) =

1
2 2


0

dkk 2 j0 (kr)[S(k) 1]

where j0 (x) = (sin x)/x is the familiar spherical Bessel function of order zero.

B. Some Exact Results

We have just seen that g(r) is related by a Fourier transformation to the observable in scattering experiments. This is reason enough to study it, but it
also turns out that knowledge of g(r) for a system described by a pair-additive
expression for the potential energy U (rN ) is sucient to determine all the thermodynamic properties of a uid!
Consider


 p2
i
+
=
u(rij )
2m
i
i <j
3
1
= N kB T + N (N 1)u(r12 )
2
2 

3
1
dr1 dr2 (2) (r1 r2 ) u(r12 )
= N kB T +
  
2
2

 2 g(r12 )



E

dr1
  

dr12

or
E =

3
1
N kB T + N
2
2

drg(r)u(r)

Thus
E /N = 32 kB T +

1
2

drg(r)u(r)

energy equation

and we have derived an exact formula connecting the average internal energy to
a pair-potential weighted integral over g(r).
Now, what about the equation of state of a uid? In this case we use

A
A = kB T ln Q , p =
V
T,N

A = kB T ln(1/N !3N
T ) kB T ln Qc



A (no V-dep.)

N
(A A ) = ln drN eU (r )
Next, we rescale all of the coordinate integrals to xi = V 1/3 ri , dxi =
V dri , which has the eect of removing V from the limits of integration.


u(V 1/3 xij )

N
i<j
(A A ) = ln V + ln dxN e
1

= +
T,N

dxN e

i<j

u(V 1/3 xij )

dxN e

()

i<j

du
1 2/3
xij
i<j d(V 1/3 xij ) 3 V

u(V 1/3 xij)

Now, restoring the original coordinate scalings leads to

A
V

T,N

1   du(rij )

rij
3 V i <j drij

du(r12 )
11
N (N 1)
r12
3 V 2
dr12

1 1
du
=
r12
dr1 dr2 (2) (r12 )
6 V
dr12
 

V
dr12

1
du
= dr2 g(r)r
6
dr
=

A
V


T,N

A
= V


T,N

= 1 16

drg(r)r du
dr

the virial equation

Thus, knowledge of g(r) also determines the equation of state, and by thermodynamic integration, the various free energies.
There is one more important formula relating g(r) to , the isothermal compressibility:

1 V

V p N,T
If we take k 0+ in the formula for the structure factor in the grand canonical
ensemble, one can show that
S(0+) =

N 2 N 2
N

However, there is a uctuation formula that can be derived in the grand

canonical ensemble relating such particle number uctuations to the isothermal
compressibility:
N 2 N 2
kB T =
N
Thus:
kB T = 1 +

dr[g(r) 1]

is the desired formula.

Lecture 6
Recap:
liquid structure quantied by g(r)
7

compressibility equation

 F.T. of g(r) S(k) scattering measurements.

g(r) thermo

E g
p g
g

energy eqn
virial eqn
compressibility eqn

Now look at approximations for g(r). First, we need to examine w(r)

dened by g(r) ew(r)

C. Potential of Mean Force

We can shed further light on the meaning of g(r) by considering the reversible work
required to move two particles, 1 and 2, together in a liquid, starting from innite separation.

We analyze the work required by considering the force required to move 1

and 2 together. It is sucient to consider the force on particle 1, since the force
on 2 (after averaging over all other particles) is equal and opposite.
The force on particle 1 is r 1 U (rN ) at a given conguration of all particles.
This includes the direct 1 2 force, r 1 U (r12 ), but also all the other particle
interactions with 1. Averaging over all but 1 and 2:






dr3 . . . drN r 1 U eU




 r 1 U (rN ) r1 r2
=
U
dr
.
.
.
dr
fixed
3



N e U

kB T r1 dr3 . . . drN e




=
dr3 . .. drN eU


= kB T
ln dr3 . . . drN eU
r1


Add to both sides kB T r 1 ln[N (N 1)/ drN eU ] = 0,




N (N 1)
U

r
r
 r1 U 1 2
= kB T r1 ln
dr3 . . . drN e
Qc
fixed



p2 g(r12 )

r1

= kB T
ln g(r12 )
r 1 w(r12 )
8

where w is the eective potential.

Clearly, r 1 w(r12 ) is the average force acting on particle 1 due to xed particle 2 and mediated by the other uid particles. The potential of mean force
that generates this force is
w(r12 ) = kB T ln g(r12 )
or
g(r12 ) = ew(r12 )
At innite dilution, in an ideal gas, only the direct force is present, so
w(r, , T ) = u(r) + ()
where w(r) is the work required to bring 1 and 2 together in a uid medium
and u(r) is the work required to bring 1 and 2 together in a vacuum.
In general, at liquid densities, w(r) will be quite dierent than u(r) due to
screening by many intervening particles.

D. Integral Equation Theories

We have discussed the signicance of the RDF g(r), its connection to experiment, and its relation with thermo. properties. How do we compute it; i.e.,
relate g(r) to some u(r)?
It is clear from our treatment of gases that perturbation theory in powers of
density wont work. In the liquid state, a non-perturbative approach is needed.
Integral equations, which sum a given approximation to all orders in , are most
commonly used. The starting point for many of these is the Ornstein-Zernike
equation:

h(r12 ) = c(r12 ) + dr3 c(r13 )h(r32 ) O-Z eqn
Think of this equation as dening the function c(r) in terms of h(r) = g(r) 1.
The function c(r) is known as the direct correlation function.
The most important property of c(r) is that it is of much shorter range than
h(r) (or g(r)). A typical sketch is:

A way to get some insight into why c(r) is shorter-ranged than h(r) is to
employ cluster expansions. Just as we expanded the partition functions and
free energies in our study of gases, one can develop cluster expansions in z or
9

 of the distribution functions. The series for h(r) in powers of is (see HCA
handout):
h(r12 ) = sum of all connected diagrams
with two root vertices (1 and
2), and number of eld vertices and f -bonds, and no articulation vertices.

Notice that the root vertices are not integrated over in evaluting diagrams;
thus, e.g.,

Note also that here an articulation vertex is one that when removed disconnects the diagram with at least one of the pieces having only eld points:
Although the h series has no articulation vertices, the graphs do have connecting
vertices, i.e., vertices that when removed disconnect the diagrams, e.g.,
The series for the direct CF, c(r), turns out to be:
c(r) = subset of h(r) graphs with no connecting
vertices

Note that if we iterate the OZ equation in a Neumann series:




h(r12 ) = c(r12 ) + dr3 c(r13 )c(r32 ) + 2 dr3 dr4 c(r13 )c(r34 )c(r42 ) + . . .
these are the diagrams in h, respectively, with zero, one, two, . . . , connecting
vertices!
10

Finally we point out that many of the c diagrams have a f -bond between
the roots, which makes them short-ranged in r12 . The remaining diagrams also
turn out to be short-ranged, although this is less obvious.
Next, we note that if we view c(r) as our object of approximation, any approximatio nis summed to all orders in . To show this, we solve the OZ equation
by Fourier transformation

h(k)
= dreikr h(r) etc.

h(k)
= c(k) +
c(k)h(k)
c

(k)

h(k)
= 1c(k)


c(k)
1
dkeikr 1
h(r) = (2)
3
c(k)
The RHS has an expansion in to all orders!
Various integral equation theories of liquids are obtained by approximating
c(r) in dierent ways (closure approx.). The most important one is the socalled Percus-Yevick (PY) equation.
The PY equation is obtained by arguing that c(r) can be obtained by subtracting the indirect correlations from the full h(r).
c(r)

= h(r) hindirect
 (r)


ew(r) 1 e[w(r)u(r)] 1

where we approximate hindirect by subtracting the direct interaction potential

u(r) from the full potential of mean force w(r). This can be rewritten as
c(r)

g geu = geu [eu 1]

= [1 + h(r)]eu(r) f (r) PY closure approx.

This approximation is exact up to (p), but approximates the 2 and higher

terms in the cluster expansion of c(r).
PY approx + OZ eqn PY integral eqn
The PY integral equation is nonlinear and must be solved numerically, with an
important exceptionthe hard sphere uid. You can read about the analytical
solution in Hansen and McDonald; it is too complicated to reproduce here.

The PY equation is clearly very accurate for the HS uid and some empirical
corrections exist to make it nearly perfect. As with any approximate theory, it
is important to note that if we substitute into the various exact equations

11

relating g(r) to thermo. props., we will get dierent results. For example,
integrating the compressibility equation gives
1 + + 2
pc
=

(1 )3

PY compressibility EOS

while the virial equation gives

1 + 2 + 3 2
pv
=

(1 )2

PY virial EOS

The nearly exact Carnahan-Starling EOG is bracketed by these:

p=

2 c 1 v
p + p
3
3

as shown in the attached gure.

We nish by noting that there are many other IE theories, e.g., HNC, MSA,
etc. which are more or less accurate for various potentials u(r):
PY
hard spheres,
short-ranged u(r)

HNC
long-ranged
u(r)

MSA
analytical results
for HS + attraction,
e.g., square wall uid

E. Perturbation Theories of Liquids

The basic idea behind perturbation theories of liquids is that the harsh, repulsive
part of the interaction potential plays the dominant role in determining liquid
structure. The attractive (???) serves to provide cohesion, but (???) structure
much. Moreover, (???) for liquids dont vary (???) like gHS (r). Thus, we (???)
using the HS uid as a (???) and perturbating in the attractive (???) of the
potential. (A large Post-It note was photocopied over this section.)
As a simple example of this idea, not that the van der Waals EOS can be
written:
1
1
p
=

a
, = d3


1 4
6
   attractive pert.
HS ref.
We should be able to do even better if we use the best HS ref. the CS
EOS:
modied VDW EOS
1 + + 2 3
p
=
a
(Longuet-Higgins & Wido??)

(1 )3

12

Lecture 7
Notes on HW:
pgm gives gd (r) from analytical solution of PY equation. Parameter in
pgm is redly(?) d.
Vikas will hold a copy of my notes, plus supplemental reading for checkout.
There are two successful approaches in the ilterature for doing this more systematically: the Barker-Henderson method and the Weeks-Chandler-Andersen
(WCA) method. We adopt the latter, which is superior, and illustrate the ideas
in the context of the LJ 612 model:
WCA method for LJ 612:


1. Recall u(r) = 4e (/r)12 (/r)6 has a minimum at r = r0 = 2116 ,
u(r0 ) = . Separate repulsive and attractive parts: u(r) = u0 (r) + u1 (r)

u0 (r) =

u(r) +  ,r < r0
0
,r > r0


u1 (r) =


u(r)

repulsive

,r < r0
,r > r0

attractive

2. Use blip-function theory to relate the repulsive model reference uid

dened by u0 (r) to a hard-sphere uid. The basic idea is nd a d = d(, T )
for the HS uid that does an optimal job of the thermo and structure of
the repulsive model uid:

The blip function expansion works as:

f0 (r) eu0 1 = fd +
b(r)

13

f0 fd
  
blip function

Substituting fd + b = ( + ) in the cluster expansion of A and

collecting all terms rst-order in b:
A = A0 Ad =

1 2

drb(r)eud (r) gd (r) = eud gd yd

The condition that A = 0, leads to a criterion for computing d(T, ):




drgd (r) e[u0 (r)ud (r)] 1 = 0
Also, to (b), g0 (r) is given by:
g0 (r) = e[u0 (r)ud (r)] gd (r)

for gd use d(T, )

Notice that combining these two implies that


dr[g0 (r) gd (r)] = 0
or from the compressibility equation a HS diameter d(T, ) is chosen
that ensures 0 = d .
3. Finally, deal with the attractions by a simple 1st-order perturbation expansion in u, (r):

A0
1
A

+
dru1 (r)g0 (r)
N
N
2



like VDW a term!

AD
CS EOS
from
with d(T, )
N

a = a(T, )

The WCA method works amazingly well for a wide variety of T = kB T /,

= 3 as shown on the attached gures. The theory is weakest for

T >
3 and <
0.6, when the attractive forces start to play a more
signicant role.

14

Advanced Topics in Equilibrium Statistical

Mechanics

5. Computer Simulation Methods

A. General Considerations
We have just seen that successful analytical theories for simple liquids like Argon can be constructed by means of modern perturbation theories of liquids.
However, these calculations become more dicult and tedious when one moves
to consider more complicated atomic and molecular uids. Moreover, computers have gotten much faster since the 1970s, so people today are more likely to
turn to computer simulations of uids.
a1. Choice of Method
There are basically three methods used for liquids and complex uids:
Molecular dynamics (MD) One integrates Newtons equations for a collection of particles, keeping track of their positions and velocities at each time
step. Thermo properties are computed as time averages over dynamical
trajectories.
Brownian dynamics (BD) Similar to MD, except that stochastic force and
frictional drag terms are included. Usually used for colloids or polymers
when solvent can be treated as a continuum.
Monte Carlo (MC) A method for sampling the congurational integral by
means of a pseudo dynamical trajectory.

System
atomic or molecular
liquid

Technique
MC (statics)
MD (dynamics)

Comments
MC faster
for statics

polymer solution
or melt

MC (statics)
BD/MD (dynamics)

coarse-grained
models often used

colloids

MC (statics)
BD (dynamics)

solvent is
continuum.

a2. Periodic BCs and Potential Truncation

3
Unfortunately, we can only aord to run fairly small simulations with N <
10
106 for atoms, much less for polymers! Dierent ensembles can be used; in xed
N, V we choose N, V to establish some . It is also best to choose N consistent
with any known crystal structure, e.g., fcc argon: N = 4n3 , n = 1, 2, 3, . . .
To minimize nite size eects in the simulation cell (mostly surface eorts),
it is typical to impose periodic BCs:

When it comes to computing forces, one also uses the minimum image
convention:
1. translate a cell of the same size (L) and shape to be centered around the
molecule of interest.
2. Sum the forces with the N 1 molecules in the translated cell. These are
the closest periodic images of these molecules.
Frequently, one also truncates the potential at some cuto distance rc (e.g.,
2.5). It is crucial that
rc L/2
for consistency with the minimum image convention as shown in the attached
gure:

B. The MD Method
The basic idea here is to solve Newtons equations of motion:
mri = Fi

i = 1, 2, . . . , N

forward in time, starting from some set of initial positions {ri } and velocities
{ri }. Many nite dierence algorithms are available; the simplest and most
popular is the Verlet algorithm. Start with:
ri (t t) = ri (t) r i (t) +

t2
ri (t) . . .
2!

Adding the + and the :

1.
ri (t + t)= ri (t t) + 2ri (t) + t2 ri (t) + (t4 )
2
4
ri (t t) + 2ri (t) + t
m Fi (t) + (t )
Subtracting the + and the :
2.
1
[ri (t + t) ri (t t)] + (t2 )
r i (t) =
2t
We can thus iterate these to get a series of new positions and velocities at
equally-spaced time increments t, known as the timestep.
The most expensive part of an MD simulation is the force calculation. Lets
see how this works for the LJ potential:


Fij (rij ), where
Fi =
j
=i

Fij =
Now, consider

x u(|u|),

u(rij ) =
u(|rij |)
ri
rij

x = (x1 , x2 , x3 )

x 

x
u=
u(x) =
u (x)
x1
xi x
xi
x2 =

x2i , 2x

xi
u(x) = u (x)
xi
x

Thus,
Fij =

1
rij

du
drij

x
= 2xi
xi
u (x)]
[or u(|x|) = x


rij =

w(tij )
rij
2
rij

where w(r) = r du
dr is the so-called pair virial function. For the LJ 612:
Fij =

24
6
2 [2(/rij ) ]rij
rij

Lecture 8
MD summary
ri =

1
m Fi

1
m

Fij (rij )

j
=i

du
Fij = r1ij dr
rij
ij
24
= r2 [2(/rij )12 [/rij )6 ]rij

for LJ 612

ij

Verlet
ri (t + t)= ri (t t) + 2ri (t) +
r i (t)=

1
2t [ri (t

t2
m

Fij + (t4 )

j
=i

+ t) ri (t t)] + (t2 )

Introduce dimensionless variables rij = rij / etc.:

ri (t + t)= ri (t t)

+ 2ri (t)
  
14
8
2
[2(rij
)
(rij
) ]rij
+24t m2
j
=i

Now, dene

=
For argon,

m 2



1/2
L-J 612 time constant

.34nm,
m = 6.7 1023 g

/kB = 120K

1.7 1012 s picosecond timescale. Finally, non-dimensionalize the

time and velocity according to
t = t/, v = r t/.
Now,

1.

ri (t t )
ri (t + t )= 2ri (t )

8
) ]rij
+24(t )2 [2(rij
j
=i

2.

1
[r (t + t ) ri (t t )]
2t i
For accuracy, we need t  1, typically t <
0.01.
Recall that Newtons equations conserve energy, so as described, this is a
microcanonical algorithm. We can check accuracy by the extent of energy conservation.
Now, how to get thermo properties?
vi (t ) =

E =K +U
and
K =

1
2
3
 mvi  = N kB T
2 i
2

Measuring energies in units of , T = kB T /

1
3
K  =  sumi (vi )2  = N T
2
2
 2
1

or T =  3N i (vi )  use to calculate temperature.

Now

u (rij
)
U =
i<

E = E  = K  + U 



12
6
E = 32 N T + 
4[(rij
)
(rij
) ]
i<

should remain constant.

From the virial equation:

p
1
du
= 1 6 drg(r)r
dr12

= 1 16 N1 dr1 dr2 (2) (r1 , r2 )r12 drdu12



but (2) =
(r1 ri )(r2 rj )
i
=
p
=

j



1
w(rij )
3N
i< j

where w is the virial function r du

dr w(r).
In dimensionless units: w = w/
5

P
T =

w (r )

1 1
3 NT 

i<

w (rij
)

= 24[2(r )12 (r )6 ]

for EOS

Lets try this out in a Mathematica notebook demo2.nb

Thus, we have seen how a basic NVE MD algorithm and simulation works.
We leave the subject with a few comments:
1. The MD equations are sti at (close approach of particles) and numerically unstable. Particle trajectories are not computed accurately (e.g.,
dierent t trajectories exponentially dierent in time) and chaos ensues. However, this is what we want in order to sample constant E phase
space! All that matters is that we conserve E accurately along the overall
trajectory.
2. Potential truncation and neighbor lists can reduce the force calculation
work from 12 N (N 1) N 2 to (N ) for short-ranged potentials.
3. Long-ranged potentials (e.g., Coulomb 1/r) are dicult. Ewald sums
and other specialized techniques have been developed.
4. Other ensembles can also be simulated, e.g., methods for NVT and NPT
MD have been devised. You can read about these in the Allen & Tildesley
reference.

Lecture 9
MC Method
Sample conguration states i rN (i) from Boltzmann distribution P (i) =
U

1
U (i)
/ i eU (i) importance sampling aves f (rN ) = M
e
i=1 f (i).
How to do sampling? Generate Markov chain of states from

P (i, t) =
Wij P (j, t)
Wji P (i, t)
r
i
=1

i
=1

for proper s. state distribution:

Wij P (j) = Wji P (i)

microscopic reversibility

C. The Monte Carlo Method

At its heart, the Monte Carlo Method is a method for evaluating integrals. To
illustrate, imagine that we wanted to do a 1-D integral

I=

dxf (x).

Let (x) be any probability distribution function, normalized such that

1. Then,
 

f (x)
f (x)
=
(x)
(x)

b
a

dx(x) =

Now, suppose we choose a total of M random numbers xi , i = 1, 2, . . . , M from

the distribution (x), where a xi b. Then
M

1
f (xi )
+ (M 112 )
M i=1 (xi )

The simplest choice of (x) is a uniform distribution (x) = 1/(b a):

M

(b a)

f (xi )(M 112 )
M i=1

which gives a simple scheme for evaluating integrals. This Monte Carlo shcme
is poor for low-dimensional integrals. The errors are
error (MC) (M 112 )
while a deterministic quadrature method, e.g., trapezoidal rule, has errors
that are much smaller:
error (TR) (M 2 )
Consider now evaluating the canonical conguration partition function

N
Qc = drN eU (r ) .
This is a 3N-dimensional integral. If we used a deterministic method with say
M 10 points in each dimension, this would require 103N function evaluations,
which is enormous! Instead, we might try the MC method, using a uniform
distribution of points in rN space:
M

Qc

V N
U (rN (i))
e
M i=1
7

This requires the generation of M 3N random number, rN (i), and M function

evaluations, which clearly scales much better with N !
The problem with the method just described is that we give equal weight to
conguration points rN (i) where eU is nearly zero (e.g., a conguration where
particles overlap), as to points where eU is (1). Thus, the method is inecient because we spend a lot of time generating congurations that contribute
little to Qc .
We can make the scheme more ecient by using importance sampling.
Suppose we wanted to calculate the ensemble average (in NVT) of some property
f (rN ):

f (rN )

drN eU (r

N)

=
f (rN )
 NQc N
N
= dr P (r )f
 (r N) N 
 N
)
= dr (rN ) P (r(r)fN(r
)
 N

N
)
= P (r(r)fN(r
)

1
M

M


P (i)f (i)
i=1

(i)

as before. Now, suppose that our random congurations rN (i) are chosen not
from a uniform distribution, but from one with the canonical pdf:
(rN ) = P (rN ).
Then,
f (rN )

M
1

f (rN (i)).
M i=1

This choice gives a simple formula for calculating averages and is the original
choice by Metropolis and coworkers. The drawback of this importance sampling is that we need a scheme for sampling the M congurations rN (i) from
= P.
The method for sampling from the pdf P (rN ) is to invent a ctitious
stochastic dynamics that has as its steady state, P (rN ). To simplify the notation, let i stand for a particular 3N-conguration space point rN (i). Then,
eU (i)
P (i) =  U (i)
ie
is the PDF we would like to sample. Consider the following stochastic dynamics,
known as a master equation:

P (i, t) =
Wij P (j, t)
Wji P (i, t)
t
j(
=i)
j(
=i)


 


gain
loss
8

where P (i, t) is the probabiity of observing state i at time t. Wij is a matrix

that gives the transition probabilities per unit time of a state change from j to
i. Because of the probabilistic nature of the entities in this equation:


P (i, t) = 1
i

and the RHS is consistent with this. However, we have considerable freedom
in selecting the Wij and, hence, dening a dynamical model. In particular, we
would like this dynamics to have the Boltzmann P (i) as its steady-state solution.
For consistency, this requires for t :

P (i, t) = 0 =
Wij P (j)
Wji P (i).
t
j
=i

j(
=i)

One way, but not the only way, for the RHS to be zero is to demand microscopic
reversibility:
Wij P (j) = Wji P (i).
This gives a constraint on the ratios of the corresponding o-diagonal elements
of W :
U (i)
Wij
(i)
= PP (j)
= eeU (j) = e[U (i)U (j)]
Wji
eij
,
ij U (i) U (j)
A suitable model is thus:
Wij =

1
eij

, ij < 0
, ij > 0

where is some rate of making transitions.

We now have rules for making transitions between states (congurations in
a uid), such that the states are Boltzmann distributed. The Metropolis MC
Method is one implementation of these rules:
1. Choose a particle (from among the N ) at random. Call this particle k.
State i is the state where this particle is at its original position, rk (i).
2. General a random, trial displacement of particle k, r = (x, y, z),
where x, y, z are selected from a uniform distribution over (rmax , rmax ).
r is thus uniformly distributed in a cube of side 2rmax about rk (i).

3. Calculate i+1,i = U (i + 1) U (i), where U (i + 1) is the pot energy

if particle k were displaced to rk (i + 1) = rk (i) + r. Note that this only
9

requires summing (N 1) terms:

=
[U (|rk (i + 1)|) U (|rk (i)|)]
(
=k)

4. If < 0 (downhill in energy): accept the move with probability 1.

If > 0 (uphill in energy): accept the move with probability 3 .
To do the latter, choose a random number x over (0,1). If x < e ,
accept; otherwise reject. If accepted, then rk (i + 1) = rk (i) + r. If move
is rejected, then rk (i + 1) = rk (i).
5. Repeat steps 1-4 to generate a Markov chain of M congurational states.
Averages from:
M
1
N
f (r (i))
f (rN )
M i=1
Some comments about the Metropolis method:
i) rmax is typically chosen and adjusted so that roughly 50% of the trial moves
are rejected.
ii) often people move particles sequentially, rather than randomly. This is faster
since less random number generation. N trial moves (1 per particle)
MC cycle. Work is comparable to one MD time step.
iii) More rapid equilibration can be achieved by multi-particle moves rather
than single-particle moves. This is particularly advantageous on a parallel computer!
iv) MC is easily adapted to NPT and VT, PT ensembles.
Lets now try this out on the LJ 612 uid via demo3.nb:

Lecture 11
Recap:
MF theories can be unied via Landau expansions.
10

 Give chemical crit. exponents independent of d.

Notion of universality.

11

Advanced Topics in Equilibrium Statistical

Mechanics
Glenn Fredrickson

6. Statistical Mechanics of Classical Fields

Up to this point we have talked about how to set up and perform equilibrium
statistical mechanical calculations of systems with a nite set of phase space
coordinates, e.g., {rN , pN }. However, often we are faced with objects, e.g.,
polymers, that are better described by continuous elds and have an innite
number of degrees of freedom. We must learn how to do statistical mechanics
for such objects.

A. Transverse Oscillations of a Stretched String

To illustrate the approach, consider a string of a musical instrument that is
stretched between two supports:

On a nanoscale, the same situation can be created by pulling on a DNA

molecule with optical tweezers. One obvious question is the following:
What is the equilibrium distribution of string shapes if the string is
equilibrated with its environment at temperature T ?
There are two ways to tackle this problem. The rst would be to approximate
a continuous elastic string by a bead-spring discrete chain:

Here the bead positions rN are used to approximate the string shape and
1

each bead has a mass

m = a where a is the average bead spacing
and is the mass per length of string.
The tension, , in the string would dene a Hookes Law spring constant k = a ,
N


1
so U (rN ) =
(ri+1 ri )2 k is the potential energy.
2
i
A second approach would be to directly work with the continuous string.
This is the approach that we will follow. For this purpose, it is convenient
to represent the shape of the string by a displacement eld u(x) from the
equilibrium position.

Ths is evidently a problem is the string takes a shape like

, since then
u(x) is not single-valued. For small thermal vibrations this is not a problem,
however.
Next, what is U for a particular shape u(x)? The total P.E. is the work required to stretch the string to that shape. Consider a small element of arclength
of the stretched string:


 2 
ds2 = du2 + dx2 = dx2 1 + du

 dx

2


2
ds= dx 1 + [u (x)] dx 1 + 12 (u
(x))
The work involved in stretching that segment is
dU = (ds dx)
12 [u
(x)]2 dx
So,

2
L

dU
1

dx
2 0
dx
Here we have to use functional notation. U is said to be a functional of the
function u(x). For each shape u(x) is associated a value of U U clearly depends
on the values of u(x) for all D x L.
U [u]

B. Calculus of Functionals
Before studying the statistical mechanics of our string, lets discuss the notion
of a functional derivative. Suppose we perturb the string shape slightly: u(x)
u(x) + u(x), where u(x) is small, arbitrary perturbation.

U [u + u]=

1
2

[u
+ (u)
]2




dx

(u
)2 +2u
(u)
+D(u2 )

L
d

= U [u] +

dx u (x) u +D(u2 )
dx


L

dxu

(x)u(x)
L



+u (x)u(x)
0

However, the string is clamped at the edges, so u(0) = u(L) = 0. Thus,


L 
d2
U U [u + u] U [u] =
dx 2 u du(x) + D(u2 )
dx
0
The quantity in {

} is often denoted
U [u]
u(x)

functional derivative of U with

regard to u(x)

since then

U =

dx

U
= u

(x)
u(x)

The minimum P.E. is achieved when u(x) is replaced by u (x) which gives the
smallest U . At the minimum



U [u] 
u(x) 

=0
uk

We can thus nd the minimum energy conguration by solving

[u (x)]

= 0,

u (0) = 0
u (L) = 0

 u (x) = 0

is minimum energy shape!

This demonstration that the straight path conguration is a minimum energy

state is a simple example of a more general calculus of variations calculation.
*** See the handout for more info about the calculus of functionals
and variations.
Next we turn to statistical mechanics of elds.

C. Statistical Mechanics of Fields

Now, suppose our string is at equilibrium at temperature T . That is the distribution of transverse shape uctuations? First, consider the analog of Qc . For
the discrete bead-spring chain,

N
Qc = drN eU (r )
In the continuum limit, this becomes:
(N , a 0)

Qc = D[u]eU [u]

where the notation D[u] denotes a functional integral, i.e., we integrate over
all possible string shapes u(x), subject to u(0) = u(L) = 0. Similarly,
P (rN ) P [u] 

eU [u]
D[u]eU [u]

and we compute averages of some G[u] according to:

G[u] D[u]P [u]G[u]
Now, how do we actually evaluate a functional integral? We could clearly
retreat back to the discrete bead-spring chain, such that


D[u] aN duN
However, a better approach is to introduce normal modes. Since the string is
clamped at both ends, a Fourier sine series is appropriate:
u(x) =

un sin(nx/L)

n=1

where un =

2
L

dx sin(nx/L)u(x) Now, we interpret

D[u] =

n=1

dun

In this representation,
L
dx[u
(x)]2
U [u]= 12
0
L


 nx 
 mx 
n m
1
cos
cos
= 2
dx
un um
L L
L
L
0
n m
2



nm L
m,n
un um
= 12
L2 2
n m

2 2
n
= 12
n u2n
,
n
2L
n=1
 1
2
Thus, P [u]
e 2 n un and each Fourier component has its own independent
n

Gaussian distribution!
Now we can compute quantities of interest.

2
1
dun un e 2 n un
=0
un = 
1 n u2
dun e

u2n = 1n = 2 2L
n2
un um = n,m 1n
, etc.


Clearly, u(x) =
sin(nx/L)un  = 0
n

u(x)u(x
)=

n

sin(nx/L) sin(mx
/L)un um 

1
=
sin(nx/L) sin(nx
/L)

n
n=1

1
= 12
cos[n(x x
)/L]

n
n

1
12
cos[n(x + x
)/L]

n
m

These are sums that can be evaluated by conversion to contour integrals in the
complex plane. The result is


L
1
1 1
1 1
1

2

2
(x + x
) +
u(x)u(x
) =
(x

x
)

(x
+
x
)
|x x
| +

2L
2L
4 L2
4 L2
Consider the special use of x = x
:
u2 (x) =

x
1 
x 1

L
5

Advanced Topics in Equilibrium Statistical

Mechanics

7. Critical Phenomena
The subject of critical phenomena, i.e., the study of the equilibrium and nonequilibrium properties of systems near a critical point (e.g., a gas-liquid critical
point), has a long history, but many notable recent advances (culminating in a
Nobel prize to Robert Wilson in 1982).

A. Mean-Field Theory: The VDW uid

The simplest theories to construct of critical phenomena are so-called mean-eld
theoriesapproaches that average the environment around a given molecule,
thus decoupling that molecule from its neighbors and making the statistical
mechanics calculation simple, like that of an ideal gas.
Indeed the Van der Waals EOS for a uid can be derived using this sort
of reasoning:


 2 
N
[V N b] = N kB T
p+a
V
and is one of the simplest MF theories of the gas-liquid critical point. Lets
begin by rewriting the equation in terms of = N
V rather than V :
(p + a2 )(1 b) = kB T.
Qualitatively, we have:
p plane:

pT plane:

T plane:

Lets now locate the critical point in the VDW model. Clearly


p
2p
=0 ,
=0 ,
VDW equation

2
T

give three equations that can give pc , Tc , c in terms of a, b. Simpler is to write

out the VDW equation as
ab3 + a2 (bp + kB T ) + p = 0


p kB T
1
p
+
= 0.
3 2 +

b
a
ab
ab
Above Tc , this has 1 real, 2 imaginary roots; these merge to 3 equal real roots
at Tc . Thus, at (Tc , pc ):


pc
kB Tc
1 2
pc
3
+
=0
+

b
a
ab
ab
or

= (p pc )3 = p3 32 c + 32c 3c
c =

1
3b

pc =

a
27b2

kB Tc =

8a
27b

Zc =

c pc
c

3
8

Next, we introduce reduced variables:

r = /c

pr = p/pc

Tr = T /Tc

which allows the VDW equation to be rewritten as:

(pr + 32r )(3 r ) = 8Tr r .
The absence of any material-dependent parameters is the basis for the familiar
Law of Corresponding States, which you have undoubtedly studied in your
thermo classes.
Next, lets zoom in on the critical region by further changing variables and
expanding the VDW equation:
t

T Tc
Tc
c
c

= Tr 1
= r 1

ppc
pc

= pr 1

temp dierence
order parameter
(density dierence)
pressure dierence
2

[ + 1 + 3( + 1)2 ][3 ( + 1)] = 8(t + 1)( + 1)



This reduces to:

1+
1
 1 2 3
4t 1 + 2

= 4t

+ 32 3 111
2
+ . . . + 32 3 + . . .

for |t|  1, ||  1.
Next, we do a thermodynamic integration to get the Helmholtz f.e.: write
A V AV
A
p= V

, AV = AV (, T )
N,T

= AV V

AV
V

= AV V



pc ( + 1)= AV + (1 + )

AV

AV

T

Helmholtz f.e. per unit

volume (intensive)


N,T

V
= AV + A


T

or in terms of a dimensionless f.e. density F AV /pc :


F
F + (1 + )
= 1 + (t, )

Integrate this ODE subject to F (t, 0) F0 (t).

The result can be written in the form of a so-called Landau expansion of

the F.E.:

F (, t) =
an (t) n
n=0

where

a0 (t)
a1 (t)
a2 (t)
a3 (t)
a4 (t)
..
.

= F0 (t)
= 1 + 4t + F0 (t)
= 3t
= 0 + (t)
= 38 + (t)

The a0 and a1 terms constitute constant shifts in the f.e. and chemical potential,
and thus do not contribute to the phase behavior. To see their eort on , note





AV
A
V

N
= V A
= V N
N

V,T
V,T
V
T

pc F
= c
t

so

c
= a1 (t) + 2a2 (t) 2 + 3a3 (t) 3 + . . .
pc

where a1 is constant shift.

Thus, workers in critical phenomena normally ignore the a0 and a1 terms:
F = a2 2 + a4 4 + . . .
a2 = 3t

Landau F.E.
,

a4 = 3/8

This looks like:

In words, above Tc , the disordered phase with = 0 is most stable; below

Tc , there are two coexisting phases with densities 1 , 2 . To nd these, we have
the conditions
p1 = p2
1 = 2 ,
but, because of the symmetry of the f.e., it is easier to apply


= F  = 0 ,
1 = = 2
()

or

6t
2

+ 32 3 = 0
= 4t ,

= 2(t)1/2 .

Thus, close to the critical point, the VDW

Lecture 10
Critical phenomena
t

c
c
= T T
,
= pp
,
=
Tc
pc
3
3 3
4t(1 + 2 + . . .) + 2 + . . .

AV /pc = (3t) 2 + 38 4 + . . .
t < 0 (t)

c
c

Landau
sion
= 1/2

order param
expan-

theory predicts a parabolic shape of the coex. curve:

The limiting shape

(t)

= 1/2

is known as a critical scaling relationship and VDW theory gives us the value
= 1/2.
There are many scaling laws that can be derived, which correspond to
looking at dierent thermo quantities or approaching the critical point from
dierent directions. For example, the isothermal compressibility is calculated
as:




= V1 V

V1 V
p

p
N
T
 N,T 
1

= p
=

T
1
6t+ 92 2 +...

Thus, as we approach the C.P. from the disordered (supercrit.) region at = 0:

and the compressibility diverges at Tc ! This has implications for light scattering
experiments. Recall
kB T =

N 2  N 2
= S(0+)
N 

Small angle (k 0) light scattering intensity should diverge as T Tc . Indeed,

this phenomena is observed and is known as critical opalescence.
We could elaborate many more predictions of the MF theory, but probably
most important is to explore correlations that develop in the liquid structure
over large distances. To study this, recall the OZ equation

h(r12 ) = c(r12 ) + dr3 c(r13 )h(r32 )

which Fourier transforms to:

h(k)
=

c(k)
.
1
c(k)

Recall that

S(k) = 1 + h(k)
=

1
1
c(k)

One nice feature of c(r) you will recall is that it is shorter-ranged than g(r) or
h(r). This remains true near Tc , so we might guess that c(r) is well-behaved in
the critical region. Thus, to study small-angle scattering (long-ranged correlations) in the critical uid, we expand
c(k) c(0) c2 k 2 + (k 4 ).
Now
S(k)
But
S(0+)
S(k)

1
.
[1
c(0)] + c2 k 2

1
= kB T
t
S(0)

2
1+(k)
correlation length

tv

(VDW), so
c2
1
, 2 1
c(0) t
1
S(0) = 1c(0) t1
, v = 12 (VDW)

Next, we invert the F.T.:

h(r) =
=
=

1
dkeikr [S(k) 1]
(2)3

S(0)
1
dkeikr 1+(k)
2 (r)
(2)3
S(0) 1 r/
(r)
4 2 r e

Thus,

Thus, we see that not only does the scattering intensity diverge, but the
correlation range diverges like tv . In particular, at Tc (t = 0):
h(r) r(d2+)

= 0 (VDW)

where d is the space dimension (d = 3 here) and is another critical scaling

exponent.

B. Other Mean-Field Theories and Universality

Many other MF theories have been developed over the years, both for uids and
to describe phase transitions in other types of systems, e.g., ferromagnetism, superconductivity, superuidity, etc. Here we discuss a model for ferromagnetism
that also serves as the canonical model for studying critical phenomenathe
Ising model.
Consider a regular lattice, e.g., a square lattice in 2D, where each site contains a magnetic moment or spin that can point in only two states, up or
down:

Let i = +1 (up), 1 (down) describe the state of the ith spin. A simple
model for the energy of a spin conguration = (1 , 2 , . . . N ) is:
E() = H

i=1

Jij i j .
2
j
i=

The rst term describes the energy of interaction between the spins and an
external magn. eld H and favors the spin up arrangement. The second term
with jJij 0 describes the interactions between spins that tend to cause them to
align in the same direction. Note N V in this model. The partition function
is:

H
i +
Jij i j

2
i
j
i
=
e
Q(, N, H) =

where

1 =1 2 =1

...

N =1

and the free energy is

A(, N, H) = kB T ln Q(, N, H)
Besides the obvious relevance of this model to magnetism, it should be noted
that there is a direct connection to the lattice gas uid model. There an up
spin is associated with a lattic site that is occupied by molecule; a down spin
with a vacant site. The hard core repulsion between molecules is accounted
for since no site may contain more than one molecule. The attractive interactions responsible for condensation are embodied in the Jij . H becomes a
7

chemical potential. Indeed, as discussed in Section 5.2 of Chandler, the GCPF

of the lattice gas exactly maps to the CPF of the Ising model!
Lets now make a MF theory for the Ising model. This is known as the
Weiss theory of ferromagnetism. Consider the force on spin k:

E
=H+
Jkj j Hke ()
k
j(=k)

where H is the direct eld and Jkj is the eld created by neighbors. This is
really a conguration-dependent eective eld that acts on spin k . We make
a MF theory by simply averaging this uctuating eective eld:

Jkj j .
Hke () Hke () = H +
j(=k)

However, j  is independent of j. This is referred to as the average

magnetization per spin, but turns out to be the order parameter, so we give
it the symbol (c.f. in lattice uid). Now, in the usual, short-ranged Ising
model

J i, j are nearest neighbors
Jij
0 otherwise
then
lattice coord. #
z = 2, 4, 6 . . .
Hke H + zJ ,

1D 2D 3D
The energy contributed by spin k is thus Ek = Hke k and the dierent
spins are decoupled. We can determine the order parameter (magnetization)
self-consistently by

k eEk
= k  =

k =1

eEk

= tan h(H + zJ)

k =1

Consider this for H = 0 (analogous to = c ): = tan h(zJ)

The critical point is at (zJ)c = 1. For zJ T 1 , the physical roots are

which correspond to coexistence of two phases, one rich in up spins; the
,
other rich in down spins. This is the ferromagnetic phase. For zJ < 1, i.e.,
8

T > Tc , = 0 is the only root and we have the paramagnetic (or disordered)
c zJ
c
= (zJ)
:
phase. Switching to t = T T
Tc
zJ

Thus, we have a coexistence curve much like that predicted from the VDW
model!
The Helmholtz free energy can be expressed in terms of by (see Handout)

A = kB T N ln




2
1 2
 

1
+ N zJ 2
2

2 cos h(H+zJ)



The rst term is kB T ln Q, Q = e k Ek ; the second term correct for
overcounting the pair interactions.
Perhaps not surprisingly, if we develop a Landau expansion of F = A/N
in powers of and switch to t = (T Tc )/Tc :
F (, t) a2 (t) 2 + a4 (t) 4
then a2 (t) t, a4 (t) 1 as before! Thus, all the same classical or MF
exponents are recovered, e.g.,
(t) , = 1/2
This same result could also have been obtained directly from the relation =
tan h(zJ) by expanding the RHS to ( 3 ).
We summarize with a few observations:
1. MF theories have been developed for many dierent systems including
uids, magnets, superconductors, superuids, etc.
2. These theories can be put in a common language by identifying an order
parameter and developing a Landau expansion of the FE. in the critical
region.
3. MF exponents for similar and seemingly dissimilar systems are in agreement and are independent of the space dimension d (see attached table).
4. MF theory suggests a universality of critical phenomena i.e., that
the molecular details of a uid, the lattice structure of a magnet or superconductor, etc. are irrelevant as far as the scaling behavior near a critical
point is concerned.
9

C. Failure of MF Theory
While the MF predictions are attractive and simple, they prove to be wrong in
general. How do we know this?
1. Experiments Numerous experiments have been carried out on a variety
of uid and magnetic, and other systems. The precision is such that the
MF exponents can be ruled out in many cases. See the attached table.
2. Exact calculations A few nontrivial models with critical points have
been solved exactly in one and two dimensions. For exaple, the shortranged ID Ising model can be exactly solved, as discussed in 5.1 of Chandler. At zero eld,
Q(, N, 0) = [2 cos h(J)]N
which gives ferromagnetic order only at Tc = 0, J ! Thus, the MF
theory is qualitatively wrong for the ID Ising model in that it predicts
ordering at the nite temperature:
(J)c =

MF
(z = 2d)
=2

1
J
1
= =
kB Tc
z
2

The 2D Ising model for H = 0 was also solved exactly by Lars Onsager in
the 1940s by a heroic eort. He found
Exact
2D Ising

(J)c = 0.441 . . .

1
8
which clearly doesnt agree very well with the MF predictions:
m (t) , =

(J)c =

1
1
=
z
4

MF
(z = 2d = 4)

1
2
No one has solved the 3D Ising model exactly (a sure-red way to win a
Nobel prize!), but the best numerical estimates give
m (t) , =

3D Ising
(numerics)

(J)c 0.25

m (t) , 0.313
which is to be contrasted with (J)c =
We can make some observations:

1
6

10

0.166 . . ., = 1/2 from MF theory.

a. MF theory is grossly wrong in ID, but seems to get better as d increases.

b. The scaling exponents apparently depend on d, so one might question the
whole notion of universal scaling behavior, except that the exponents
of dierent systems are comparable.
The fact that MF theory gets better with increasing e is easily understood.

Jkj j
Hke () = H +
j(=k)

will uctuate less if the j sum contains more neighbors. In particular, the error
of writing Hke () Hke () is (z 1/2 ) for z . Thus, we expect MF
theory to be exact in the limit z = 2d .
There is one other limit where MF theory is obviously exact. Consider a
model with long-ranged interactions, e.g., an Ising model where
Jij = Jf (|rij |)
and f (r) decays on some large lengthscale :

Then the uctuations in the e. eld depend on the number of spins Ns

d /ad (ad is volume of a lattice site) within the interaction range. So


Hke () Hke () + (/a)d/2
and MF theory is exact as . So, why is MF theory wrong for nite d, ?
Clearly the answer is that we neglect uctuations that are very important and
can even suppress ordering in lower dimensions.
To understand these uctuations, lets think about what a uid or magnet
looks like as we approach Tc from above or below at the critical composition (or
H = 0):

There is no substitute for an actual picture, so the attached page from Chandler (page 167) shows typical congurations of an Ising model (2D) that has been
sampled by the MC method. In summary:
11

a. Fluctuations are short-ranged and very random for T Tc .

b. Fluctuations are long-ranged and look highly varicose and interconnected
for T >
Tc .
c. A stable interface between domains is present for T < Tc .
d. The interface roughens considerably for T Tc and this roughening has
very long-ranged components.
Overall, the uctuations are strong and long-ranged in the critical region.
What might we do to improve our MF theories? For uids, you might
think about switching to a better EOS than the VDW model, perhaps by using
an integral equation theory or WCA-type approach. The problem with this
is that all the closure of approximation schemes that have been devised, e.g.,
PY closure, only build-in more accuratee short-ranged correlations in the uid
structure. Thus going to a better liquid state theory for the 150
A structure of
a uid does not translate into a better approximation for the critical behavior!
Indeed, IE theories are notoriously bad in the critical region; some are worse
even than the VDW model!!
For magnets, people have devised more sophisticated MF theories (than the
Weiss theory) that build in short-ranged correlatiosn neglected in the simplest
approach. While these methods give more accurate estimates of Tc , they invariably fail at changing the MF exponents; e.g., remains 1/2.
In short, it appears that an entirely new theoretical approach is needed to
analyze critical phenomena!

D. The Scaling Theory

By the mid-1960s it was pretty clear that MF theory was in trouble. Widom,
Kadano and others injected a new idea into the mixthe so-called scaling
hypothesis. Clearly, what MF theory does wrong is to ignore the long-ranged
uctuations that occur on the scale of the correlation length discussed previously. These uctuations apparently change the scaling relations, so lets build
a theory assuming that is the only lengthscale that matters. Recall the t
(pT ) diagram:

The correlation length is an intensive property that can be expressed as

= (, t).

12

If we approach along = 0:
|t|v , = 0
where v is the scaling exponent we saw in our discussion of the VDW uid.
However, we expect a dierent divergence if we approach along t = 0:
||y , t = 0
where y is some new scaling exponent. Since v = y in general, constant
contours in the t plane will not be circles.
Now the basic idea behing the scaling hypothesis is that if only matters,
then the singular part of the free energy density (either Helmholtz or Gibbs)
should scale as
Av Gv [(, t)]d kB T
From this simple statement, all kinds of wonderful things ow. For example,

suppose we want
 the shape of the coexistence curve, |t| . Recall from
thermo that

G
p

= V . One can show that the singular part of the FED,


GV
GV = G/V satises:
. Then
T,N

GV
GV

d 1 .

However, if the system only cres about and not , t separately, then we can
write
d 1 d ( 1/y )1 1/yd
or

|t|v(1/yd)
(1) Therefore, = vd v/y

gives in terms
of v, y

Next, suppose
we want in |t| , expressed in terms of v, y. Recall,


. Then,

t
 G
Gv

d 2
v

t 2
d 1/y 2
2/yd
(
)
|t|v(2/yd)
gives in terms
of v, y
Here is one more example. Recall the exponent dened by the RDF:
(2) Therefore, = 2v/y vd

h(r) r(d2+)
13

It follows that

h() (d2+)

But 2 h(r) (r)(0), so

h() 2 (|t| )2 ( 1/v )2 2w/v
(3) Therefore, d 2 + = 2/v
There are many others that can be derived. The bottom line is that we can use
the scaling hypothesis to calculate all the other exponents, once the fundamental ones v, y are determined. The scaling theory does not help us with values
for v, y.
A convenient way to test the theory is to eliminate y from the various relations, e.g., (1) and (2) give
v/y = vd =

1
( + vd)
2

vd = 2 +
The attached sheet shows that the scaling theory indeed works very well and is
believed correct.
Thus, while the scaling theory does not allow us to compute exponents from
scratch, it does permit one to derive non-trivial relations among the exponents
indicating that only two are truly independent.
Lets nish our discussion by showing how Landau theory fails
Av d (|t|v )d |t|vd
(1/ )vd vd/
Since, vd/ is not an integer, Av () is not an analytic function at = 0! The
Landau theory (all MF theories), incorrectly lead to
Av t 2 + 4 + . . . 4

Lecture 12
Recap: RG Theory A framework for analyzing critical systems
RG Transformation:
1. coarse graining
2. rescaling

14

Analyze how parms evolve, e.g., p = (H, J)

p = Rp
Repeated iteration gives ows
xed points p are CP
critical exponents eigenvalues of linearized R
Methods:
1. Real space Chapter 5, Chandler
2. Momentum space statistical eld theory

E. Renormalization Group Theory

The modern tool for dealing with strong, long-ranged uctautions is the Renormalization Group (RG), introduced by Ken Wilson. This is a very complex
theory, so we will only scratch the surface, but it does allow the direct calculation of critical exponents as well as validates the scaling theory.
The basic idea of RG is to carry out an RG transformation that consists
of two steps:
1. A coarse-graining, or Kadano transformation that removes short-lengthscale
information. For example, in a 2D Ising model we could group sping in
blocks of 4:

2. A rescaling to shrink the lattice back to its original size.

In this transformation, we keep track of how the parameters that characterize

the model, e.g., p = (J, H), are changed to p = (J  , H  ):
p = R[p]

R = RG operator
15

where R is a nonlinear operator in general. Next, we iterate this many times

and look for xed points p , where
p = R[p ].
One or more of these will correspond to the physical critical point and we shall
see that the critical exponents are related to the eigenvalues of the linearized
RG operator about p !
There are two approaches to implementing RG:
1. Real (position) space this is done on a lattice by forming block spins
as described above and is demonstrated in your text (Chandler).
2. Momentum space here the DOF are Fourier components of the order
parameter eld.
We shall follow the momentum space approach here, which is more generally
useful for practical calculations on simple and complex uids.
Our starting point will be a simple eld-theory model of a uid at the
criticl density:

Q=

D[]eH[]

where the eective Hamiltonian has a Landau-type form:




1
1
1
r0 2 (r) + u0 4 (r) + ||2
H[] = dr
2
4
2
v
and is a functional of (r). The parameters of the model are p = (r0 , u0 ),
where u0 > 0, while r0 t vanishes at the MF CP. The term ||2 penalizes
gradients in density, (r) (r)c and can be related to the interfacial tension
below Tc . Finally, D[] is shorthand for a functional integral over the eld
(r). We can dene this a couple of dierent ways. If we discretize the vol. V
into N cubic cells, each of volume a3 , then

D[]

lim
N
Q3 0
N a3 = V

N 


i=1

di

Alternatively, we could Fourier decompose (V L3 )(r) =

x
where kx = 2n
L , nx = 0, 1, 2, . . . and interpret


Notice that k becomes
dk
D[]
continuous for L !

16

1
V

eikr k ,

The latter is more convenient and will be used here. Note that arbitrarily
long-wavelength (small k) density uctuations are allowed, which are important
near the CP!
The rst step in an RG calculation is to do the coarse-graining. In real space
we form block spins at small scales; in mmomentum space we remove (integrate
out) k with the largest |k|!
 

H[]
e

dk eH[] , b >
1

b <|k|<

is coarse-grained Hamiltonian. The parameter is a large k cuto,

where H
which is required to keep the theory nite. In a more complete theory, the
molecular structure would provide a natural cuto at scales 1/. In kspace we are thus sharing a thin shell when b slightly exceeds 1:

Unfortunately, we cannot do this coarse-graining step exactly, but we can

do it in a perturbation expansion in u0 , because the integrals are then Gaussian.
To see this, note that the ( 2 ) part of H is:

1
=
dr (r0 2 + ||2 )
H2
2
V


1
2

1
d
(r0 + k )k k 

dd k(r0 + k 2 )k k
= 2V
2(2) |k|<
k
for
V
so the integrals over k factor into independent Gaussian integrals for each k.
The algebraic details of carrying out the integrals over the momentum shell
are complex, so we simply give the results to (u20 ). Note also that we cite
results for an arbitrary space dimension d: To (u20 ), one nds an expression for

the eective H[]

that has exactly the same form as H[], but with ro r,
u0 u:

1
dd k
+ (u20 )
r
= r0 + 3u0
r0 + k 2
/b

1
dd k
+ (u30 )
u
= u0 9u20
(r0 + k 2 )2
/b
We are thus doe with step 1 of the RG-coarse-graining. Next, we need to
rescale. We do this by amplifying k to restore the radius of the sphere in k-space
to :
q = kb.
17

We shall also want to rescale k so that the coecient of ||2 does not change:
q = z 1 k
where we wil gure out what the scaling factor z needs to be in a moment. Now,
in k-space

1

dd k(r + k 2 )k k .
H2 [] =
2(2)d |k|< b
After the rescalings, we have a new renormalized Hamiltonian:

1


d
d2 2
 
dd q(z 2 b
r + z 2 b
H2 [ ] =
 q )q q .
2(2)d |q|<


r

Clearly we want z 2 bd2 1 or z = bd/2+1 to keep the q 2 coecient unchanged

under the RG. Combining these results, we see that the net eect of the RG is
to change H to H  , where

1
r = z 2
bd{r0 + 3u0
dd k
+ . . .}
r
+
k2
0
/b
2
b

1

4 3d
dd k
+ . . .}
u = z b {u0 9u20
(r
+
k 2 )2
0
/b
This is our rst hint that something is special about 4 dimensions! For d > 4,
u is reduced below u0 (recall b > 1), while for d < 4, u is amplied by b4d .
To study this, we imagine that d can take on any value and expand about 4
dimensions by waiting
d4
and then expanding for   1. (Note this new denitions of !) We thus
expand the wavevector integrals in powers of  (you will be asked to do this in
a homework assignment):


!
3
2
1 b!12 83 2 u0 r0 ln b + . . .
r = b2 r0 + 16
2 u0
u = b
u0 89 2 u20 ln b + . . .
Notice that we treated r0 , r as being () in the expansion! (Justied later.)
Finally, we write b
= e
ln b ln b + . . ., and re-expand:
"
#
9
u = u0 + u0  2 u0 ln b + . . .
8
Our next step is to treat the RG as a dierential transformation by taking
b 1+. We do this by letting b = es , s 0+, and u = u(s + s);
u0 = u(s) etc. Then
$
%
du(s)
= u  89 2 u
ds
dr(s)
= 2r + 83 2 u(2 r)
ds
18

are the RG ow equations, which describe how the eective Hamiltonian transforms under a dierential coarse graining and change of scale.

Lecture 13
RG momentum space recap:
Field theory model



1
1
1
2
4
2
r0 + u0 + ||
H[] = dr
2
4
2
v

SM of order parameter uctuations from: Q =
D eH[]
  
 
dk
|k|<

statistical eld theory

RG transformation
i) Shave a shell from /b < |k| <
ii) Rescale k bk and elds
iii) Go to a dierential transformation b = es 1 + s and expand for
G=4d1
du
ds
dr
ds

$
%
= u  u9 2 u
= 2r + 83 2 u(2 r)

We can now see how the eective Hamiltonian transforms by solving these
ODEs parametrically in s. The relevant IC is
r(0) = r0 , u(0) = u0

initial
model
parameters

First, we look for xed points:

dr
du
=0,
=0
ds
ds
(1) r = 0, u = 0

(r , u )

Gaussian FP
2

2
(2) r = 16
, u =
2 u = 6

8 2
9 

19

both ()!

Wilson-Fisher FP
Now, lets look in the phase plane to see the RG ows near the FPS:
d > 4 ( < 0):

The physical CP is the Gaussian xed point, because there is one eigendirection where the point is stable. To get critical exponents, we linearize about
the GFP:
r(s) = r(s) r = r(s)
u
(s) = u(s) u = u(s)
Let


r(s)
u
(s)

d
2
P = MP , M =
0
ds
P(s)

Solution:

3
2
8 2

P(s) = c1 v1 e,s + c2 v2 e2 s

where the eigensystem is:

Mv
1 = 2

= v
, 2 =  , v1 =

 
 32 
1
1 162
, v2 =
0
1

The eigenvalues have mixed sign for  < 0, so this is indeed a saddle point.
The v1 direction is unstable (r-direction); the v2 direction is stable and is the
critical manifold.
Next, we show how the eigenvalues can be related to critical exponents.
Consider the correlation length relation tr . In our model, = (r, u) or
= (e1 , e2 ), where
e1 = v1 P = r
32
e2 = v2 P = 16
2r + u
Note that the coarse-graining step of the RG does not eect , since we are
integrating out ner scales. Thus,
(e1 (s + s), e2 (s + s)) =

b1


es 1s

20

(e1 (s), e2 (s))

Taylor expanding to (s):

de1
de2
d
+

=
e1 ds
e2 ds
ds
Thus, (e1 (s), e2 (s)) = es (e, (0), e2 (0)) or
(e1 (0), e2 (0)) = es (e, (s), e2 (s)).
Now,

dei
ds

= i ei ei (s) = ei s ei (0)
(e1 (0), e2 (0)) = es (e1 s e1 (0), e2 s e2 (0))

choose s = s s.t. e1 s e1 (0) = e1 s r0 = 1 s = 11 ln r0

1

(r0 , e2 (0)) = r0




2
1, r0 1 e2 (0) .

Now, 21 > 0, so for r0 t 0

(t, u0 ) t1/1 (1, 0) tv
therefore v = 1/1 . This is a general result. For the Gaussian FP, we thus
have
1
v= , d>4
2
and we have proved that the classical MF exponents are exact for d > 4! Thus,
MF theory gives the proper critical scaling relations above four dimensions!
Now, lets see what happens for d < 4:
d < 4( > 0):

The roles are now reversed, so the physical CP is the WF FP! Linearizing
the ow equations about that point:
1
2

= 2 13  + (2 )
=  + (2 )

Thus,
v=

1
1

1
(1 + + . . .)
=
1
2(1 /b)
2
6
21

The two eigenvalues now determine all other critical exponents to (), e.g.

= 1 + 6
+ . . .
= 12 6
+ . . .

Thus, we seem to have a method for computing nontrivial corrections to exponents! Lets see how good they are when we blindly set d = 3( = 1!):

Thus, Wilsons -expansion method at () is already much better than

MF theory! We nish our discussion of RG theory with some comments:
1. Today, the best estimates of critical exponents are obtained by high-order
-expansions, combined with fancy series analysis (Borel transforms).
2. Notice that our momentum-space procedure is consistent because the WF
FP (r , u ) had u (), which justies the perturbative coarse-graining
in powers of u. This is known as a weak-coupling FP. Some very dicult
problems have strong-coupling FPs, where u (1), invalidating the
procedure.
3. Our momentum space approach identied an upper critical dimension
duc = 4, above which MF theory is exact. Note that there often also
exists a lower critical dimension, below which ordering is not possible
at a nonzero temperature. For this model, dc = 1, consistent with the
1D Ising model. Note that it is also possible to do d =c + expansions
about the l.c. dimension where Tc ().
4. The RG theory naturally develops the concept of universality. The shortscale details of our model, e.g., the cuto , did not enter the nal expressions for our exponents. Thus, any ferromagnetic system or uid near
its critical point should demonstrate the same critical scaling exponents,
regardless of the details of its composition!
5. It is important to note, however, that generalizations of the model can
lead to dierent exponents. For example, if our model has n-coupled
order parameter, i.e., i (r), i = 1, 2, . . . , n, a simple generalization is:



1
1
1
2
r0 + u0 ( ) + :
H[] = dr
2
4
2

22

This n-vector model also has duc = 4; an  expansion, however, produces

critical exponents that depend on both d and n: e.g.,
v

n+2
= 12 + 4(n+8)
 + (2 )
3
 + ()2
= 12 2(n+8)
n+2
= 1 + 2(n+8)
 + (2 )

These formulae reduce to those given previously for n = 1. However, generally, we see that we must sub-divide our notion of universality into so-called
universality classes, e.g.:

n=1

Ising Universality Class

(uniaxial magnets, LG
:
CP in uids, LL CP in
binary uids, . . . )

n=2

XY model Universality Class

(magnets with two spin compo:
nents, Langmuir layers, liquid
crystals)

One other interesting point is that dc depends on n! It turns out that dc = 1

for n 1, dc = 2 for n > 1! Many other universality classes can be created
by changing the symmetries or nature of the couplings. How many are there?
We dont know!!

23

Advanced Topics in Equilibrium Statistical

Mechanics

Lecture 14
8. Polymer Statistical Mechanics
A. Coarse-Graining
Polymer science is a particularly rich area of application of statistical mechanics. Before we can talk about the statistical thermo of polymers in solution
or in the melt state, we need to discuss some basic models for summing the
conformational states of a single polymer chain.
Common polymers are:

where N , the degree of polymerization can range from 103 105 . How are
we to describe such molecules within a CM framework?
1. An atomistic approach. Here we construct pair potentials, usually within
an interaction site framework for all the atoms composing the macromolecule. Rotations around backbone C-C bonds are permitted, which
leads to large numbers of conformational states eN . While in principle,
we can apply our MD and MC methods to such models, in practice they
are very computationally demanding! For example, a polymer melt has a
relaxation time that scales as
0 N 3.4

where 0 1 109 sec is the time for a conformational transition of one

monomer. Even for a modest N 103 , this gives 10 sec, which is
nearly 1015 larger than the timestep t that we would like to use in an
MD simulation!
2. A coarse-grained approach. Here, we ignore the atomic details below
about 10 A, but simply preserve the connectivity and architecture of the
polymer. Think of this as a low-magnication view of a chain, e.g.,

In the coarse-grained approach, we need to impose some new eective

interactions among the larger units with which we work.
A disadvantage of the CG approach is that we often cant ask precise questions about properties that depend on monomer shape and interactions, e.g., Tg
or Tm values. An advantage of the CG approach is that if allows one to more
easily identify universal scaling relations and properties; i.e. those independent
of the specic type of polymer. For example, the scaling law
0 N 3.4
holds for all linear polymers, and while the prefactor 0 depends on chemical
details, the scaling exponent 3.4 does not! Another advantage is that coarsegrained models will allow us to extract a number of such relations analytically!

B. CG Models of Single-Chain Statistics

The crudest single-chain model is the freely jointed chain model, where we
describe a polymer as a freely-hinged sequence of N segments of xed length b:

We could use {R1 , . . . , RN +1 } as coordinates, but it is easier to use bond vectors

to describe a particular conformational state of the chain.
ri Ri+1 Ri

Next, we write ri = ni b, where the ni are unit vectors, |ni | = 1. In the FJC
model, all bond orientations are equally likely. Thus, we compute conformational averages as:


N


1

f ({r})
unit dni f ({nb})
4
i=1
sphere


Lets try this out. Notice the end-to-end vector R of a chain is i ri = b i ni :

R

=
sumN
i bni 
=
bni  = 0
i

So, on average, the chain is pointing in no particular direction. The meansquared end-to-end vector is:


R2  R R =  i bni j bnj 


= b2 i j ni nj 

2
=b
N n1 n1  +N (N 1) n1 n2 

 
 

n1 n2 =0

So,
R2 1/2 = bN v , v = 1/2

and the coil size of such a polymer should grow as M W

More sophisticated models than the FJC model have been built and studied.
The most realistic are the so-called rotational isomeric state (RIS) models, which
attempt to capture real bond angles and trans-gauche conformers for specic
polymers. Perhaps surprisingly, the result for R2 1/2 is the same for N
R2 1/2 bN v , v = 1/2
except that now b has a more complex connection to the local geometry of
the chain. In fact, all such models with only short-ranged constraints (along
the chain) on conformations give this universal result. Asymptotically for
N , one can also show that all such models also lead to a Gaussian PDF
for the end-to-end vector R.
3

Because the local details appear not to be crucial, people working in polymer
SM like to work with a simple model that is Gaussian, both locally and globally:

In this bead-spring model of polymers, the mass is centered on the beads.

The beads are connected by linear (Hookian) springs:
U (RN ) =

N
3 
(Ri Ri1 )2 .
2b2 i=1

It is easy to show that b now has the interpretation b = (R2 R1 )2 1/2 ,

i.e. the RMS length of one spring! We could calculate directly with this model,
but it is easier to work with a continuum model of innitely many beads and
springs:

Continuous Gaussian Chain Model a functional of the chain conguration R(s).

The probability density of observing a specic conguration R(s) is thus
P [R] = 

eU [R]
D[R]eU [R]


where D[R] means a functional integral over all polymer conformations
R(s). This is our second example of the SM of elds as opposed to SM of
particles (rst example was critical
phenomena / RG theory).

What is the meaning of D[R]? One approach is simply to undo our
discretization:

N 


dRi .
D[R]
i=1

A better approach is to expand R(s) in normal modes.


R(s) = X0 + 2 Xp cos(ps/N )
p=1

Rouse
modes

These modes have the property that dR

ds = 0 for s = 0 or s = N , so that the
chain is stress-free at both ends. By orthogonality of the cosines:
Xp =

1
N

ds cos(ps/N )R(s)

p = 0, 1, . . .

Now, U transforms to:

U [R]

3
2b2

ds

dR
ds

2


b 2 2
= 12
kp X2p , kp
p
N b2
p=1

and




D[R]

dXp ,

p=0

so
P [R] P [X] =  
p

12

kp X2p

12

dXp e

kp X2p

Thus, to compute averages, we only have to do decoupled Gaussian integrals!

N
Notice that X0 = N1 0 dsR(s) is the center of mass of the polymer. It
does not enter U , because the stretching energy does not depend on the COM
location.
Lets now see how we can calculate with this machinery. First, R2 :


R = R(N ) R(0) = 2 Xp [cos p 1]
p=1

odd
= 4
Xp
p

odd odd
R  = 16
Xp Xq 
 
p
q
3

Xp Xq 
 
=1
2

1
kp

p,q

odd 1
= 16 3
kp
p
odd 1
= 82 N b2
= N b2 !
2
p
p
 
2 /8

Thus, we have set up the Gaussian model with a b denition that agrees with
the FJC. If this were all we could compute, there would be no advantage of
the GCM. However, we can calculate structure factors of polymers, S(k), and
many other quantities of interest. One other interesting quantity is the radiusof-gyration, i.e., the second moment of the mass distribution of the polymer:
Rg2

1
N

1
N

=
=

12
N
12
N

0

N
2

ds[R(s) X0 ]2 

ds4
cos(ps/N ) cos(qs/N ) Xp Xq 
 
p
q

p


p

3p,q k1p

cos2 (ps/N )kp1

kp1 = N b2

1 1
]
2 p=1 p2
 
2
6

Rg2 =

N b2
6

R2 
6

In summary, we see that

R2 1/2 Rg N v , v = 1/2
This is known as ideal-chain scaling behavior.

Lecture 15
Recap: Coarse-grained models
-freely jointed
-Gaussian (discrete and continuous)
R = R2 1/2 = bN v , b = 1/2 ideal chain
long-ranged interferences doesnt account for excluded volume
short-range interferences does account for local stiness, bond constraints,
etc.
Rg2 = R2 /6 , Rg bN v , v = 1/2 ideal chain
Dist. vector of end-to-end
2
2
3R2
P (R) e3R /2N b , N ; F kp T ln P , F 2N
b2 intrinsic
elasticity

C. Conformations and Thermo of Dilute Solutions

The above models might suggest that dilute solutions of polymers, i.e.,

would show ideal chain scaling behavior. This is wrong! Experiments have
shown that
R Rg N v , v 0.59
regardless of the good solvent used or the chemical details of the polymer!
The reason for the departure from ideal behavior is that the CG chain models
have neglected the excluded volume eect, namely that far away monomers
along the chain cannot occupy the same region in space:

self-avoiding random walk

Such repulsions are short-ranged in space, but are long-ranged along the contour
and are hard to deal with. First a crude way to analyze excluded volume eects:
Flory theory of EV
Think of a coil as a gas of hard sphere monomers in a volume R3 :

Estimate the FE as a sum of elastic (entropic) and virial terms:

F

R2
2
Nb 

R3
B2 c2 
 
repulsive coil
f. energy volume
density

entropic
stretching
(connectivity)
 2
N
3
R2
N b2 + b
R3
R3



b3 N 2
R3

Minimize WRT R:

R bN vF , vF 3/5 = 0.6

where 3/5 is the Flory exponent.

The Flory theory gives a quite good estimate of v, but there is a fancy way
of doing better! P.G. DeGennes pointed out in 1972 an interesting connection
between critical phenomena and self-avoiding polymers. Namely, the excluded
volume problem can be mapped exactly to the critical scaling of the n-vector
model for n 0! The correspondence is:
tv R (N 1 )v
crit. phenom.
polymers
so that N corresponds to t 0. Recall that  = 4 d expansions for the
n-vector model gave:
1
n+2
v= +
 + (2 )
2 4(n + 8)
For  = 1 (d = 3) and n 0, this gives v 9/16 0.562. Resummation of
higher order  expansions give the best value currently known:
v = 0.588 0.001
Now lets consider the thermo. of dilute solutions. We should bve able to
model dilute solutions in much the same way as we approached gases, but where
atoms polymer coils and u(r) w(r), where w(r) is a potential of mean
force between two polymer coils (at dilution in the solvent). For the gas, the
Helmholtz FE per volume:
AV ln + (ln 3T 1) + B2 (T )2 + . . .
For the polymer solution, is the density of polymer coils. Rewrite as:
=

# chains
c
=
vol.
N

where c = # monomers/vol. Also, two coils in a good solvent will repel as hard
spheres

Thus,

AV

c
N

ln c+ Nc (ln(3T /N ) 1)
+vN 3v2 c2 + . . .
8

Now, the osmotic pressure is dened as

A


T,np

where p is the # of polymers. Setting A = AV (c, T )V , c =

np N
V

c
V
= AV V V
A
c
AV
= AV + c c
3v2 2
= Nc + vN
 c + ...
A2 (T,N )

Thus, the osmotic pressure of dilute polymer solutions is the sum of an ideal
gas term and an excluded volume term that scales as
A2 N 3vF 2 N 1/5
These predictions have all been conrmed! Note that as the solvent becomes
poorer, w(r) becomes more attractive and at the theta temperature T = ,
A2 (, N ) = 0. At this special point R bN 1/2 , chains are ideal; for T < it
they begin to collapse and aggregate:
Overall, we can classify solvent quality by this scheme:
T > , A2 (T, N ) > 0
T = , A2 (, N ) = 0
T < , A2 (T, N ) < 0

good solvent
theta solvent
poor solvent

For T < 0, because A2 < 0, we need to retain the next term in the virial expansion, A2 < 0. This augmented series can be used to map out a coexistence curve
for T < 0.

Now, what is the meaning of the concentration c ?

D. Semi-Dilute SolutionsScaline Theory

As we increase the concentration of polymers in a good solvent, the coils begin
to touch and then overlap:
9

At the touching point, the so-called overlap concentration c , the concentration of monomers in the solution is the same as in an isolated coil:
c

N
1
3 N 13vF b3 N 4/5
R3
b

For high molecular weight, this can be a very small concentration! Furthermore,
for N large, there can be a big range of concentrations where
c
c
b3 the semi-dilute regime
where b3 is when = cb3 1 - pure polymer.
We have talked about the osmotic pressure of dilute solutions. Lets now try
to gure out how scales with c in the semi-dilute regime. Imagine that we
can represent versus c over both regimes in the sollowing form:
=

c
f (c/c )
N

where f (x) is a dimensionless functio nof x c/c . To recover our dilute results:
f (x) 1 + a1 x + (x2 ) for x
1
In the semi-dilute regime, one might expect to depend on the monomer concentration c, but not on N since chains are strongly overlapping:

We assume then that

f (x) xp for x  1
and deduce the exponent p by demanding that is independent of N !
cN 1 (c/c )p
cp+1 N 1 (N 13vF )p
cp+1 N p1+3pvF
We need 3pvF p 1 = 0 p =

1
3vF 1

5
4

for N -independence. Then

cp+1 c9/4 , c
c
b3 . As you see from the attached experimental
plot (this would predict a slope of 5/4 = 1.25), the data collapse and scaling is
excellent!
Another way to recover this result is informative. Lets try to gure out how
the mesh size or correlation length scales with c. Imagine that
Rg(c/c ) bN v g(c/c )
10

If g(1) (1), then R at overlap. If g(x) xq for x  1 and we demand to

be independent of N , then q = v/(3v 1) = 3/4, so c3/4 c
c
b3 .
We might guess that, like in critical phenomena, the intensive thermo props
of semi-dilute solutions may depend only on . Then, dimensionally
AV

kB T

Indeed

kB T
(c3/4 )3 c9/4 !
3
Lets now think a bit more physically about the meaning of . Suppose we
draw spheres of diameter around the strands of the mesh:

Call these spheres blobs. Within a blob, the piece of chain is nearly unaffected by neighboring bits of chain, so it should obey excluded volume statistics.
Thus, if there are g monomers per blob:
g vF g 3/5
or g 5/3 (c3/4 )5/3 c5/4 is the concentration dependence of g.
Now, each chain can be visualized as a pearl-necklace of blobs:

These pearl-necklace chains pack to form a dense melt of blobs; i.e., the blob
volume fraction, b = 1, but the polymer volume fraction = cb3
1. In
polymer melts (no solvent), we know from experiment and theory that chains
are ideal:
R bN 1/2 Flory theorem for melts
The physical reason for this is that the excluded volume eect is screened out
by the interactions with many other chains.
Now, back to semi-dilute solutions. For the same reason, we expect that our
pearl-necklaces, which are densely packed, should be ideal on large scales:
1/2

R Nb

, Nb

N
# blobs/chain
g

11

where now blobs replace monomers in the melt. This gives

R c3/4 N 1/2 g 1/2 N 1/2 c3/4 (c5/4 )1/2
R(N, c) N 1/2 c1/8 semi-dilute. Thus, chains in semi-dilute solutions are
ideal with regard to N at xed c. Moreover, R decreases with increasing c at
xed N , an interesting prediction that has been conrmed!
Finally, we note that g(r) in a semi-dilute solution looks like:

12

Advanced Topics in Equilibrium Statistical

Mechanics

Lecture 16
Recap: semi-dilute, good solvent

(c) independent of N ! for c
c
b3

b(

cb3 )3/4 c3/4
p 1

So, b (1A )
(20A )
R (100A )

kB T
3

c9/4

9. Interfacial Statistical Mechanics

Thus far, we have restricted our attention to homogeneous, or uniform liquids. However, there are many non-uniform systems of interest (solids, surfaces, interfaces, etc.) and the equilibrium SM of such systems has been studied
in some detail. In this course we will limit our attention to the simplest of
inhomogeneitiesa planar interface between coexisting gas and liquid phases:

The density prole turns out not to be sharp, but has a width of order the
bulk correlation length .

We will prove that this is indeed the proper interfacial width scale shortly,
but now simply focus on the problem of determining the equilibrium (z) and
the interfacial energetics.

A. Thermo of Interfaces
Lets now review surface thermodynamics. Consider the gas-liquid interface as
above in an open system:
dE = T dS pdV + dN +

d
1st Law


work to create
more interface
area

The surface tension is thus = (E/)S,V,N Now, S, V, N are extensive;

T, p, are intensive.
E(S, V, N, ) = E(S, V, N, )
so E is a rst-order homogeneous function of these variables. By Eulers theorem
for such functions:

(E/xi )xi
E({xo }) =
i

so
E = T S pV + N +
Notice that the origin of the z-axis is arbitrary. Lets insert a so-called dividing
surface at zd , such that we call everything for z < zd gas and everything with
z > zd liquid. Then imagine a hypothetical uid with a step function (z):

For this uid,

EL = T SL pV2 + NL + L G
2

and V = VL + VG Now, dene surface excess properties X S for any extensive

property X:
X S = X (XL + XG )
Clearly V S = V (VL + VG ) = 0. We can write
S

N =

dz(z)
L



N

NS
L

zd

dzd (z)
L



zd

NG

dzd (z)



NL

dz[(z) d (z)]

Clearly, by adjusting zd , we can cause this integral to vanish. This choice

Gibbs dividing surface.
Now,
E S = E EL E G
= T S S p(V S ) + (N S ) +
Thus AS = or = AS / surface Helmholtz f.e. per unit area.
Another way to write this is in terms of a z-dependent free energy density
AV (z):

 0
dz[AV (z) AV,G ] +
dz[AV (z) AV,L ]
=

where we have shifted the origin of z to zd , or:


=
dz{AV (z) (z)AV,G (z)AV,L }


where (z) =

1, z > 0
0, z < 0

unit
step fnc.

B. Density Functional Theory

Now we are ready to think about how to calculate and the prole (z). There
exists a theoretical framework known as density functional theory that provides
a powerful way to solve inhomogeneous problems in both classical SM, as well as
many-body QM (c.f. Nobel Prize of Walter Kohn!). In simplest form, DFT says
that the Helmholtz F.E. of an inhomogeneous NVT system with density (r) can
be expressed as a unique functional of , A[]. Moreover, the equilibrium
 density
prole 0 (r) is obtained by minimizing A[] subject to the constraint v dr(r) =
N . Equivalently, we minimize the Grand Potential [] dened by

[] = A[] dr(r) {V T }

subject to variations in (r):

A
=
= 0 equil. condition
(r) 0 (r)
(r) 0 (r)
The chemical potential is thus a LaGrange multiplier that is adjusted to
achieve a desired N . Another interpretation is that the chemical potential is
independent of z through an interface!

Lecture 17
Recap:
= AS / where AS = surface excess Helmh. F.E. = A AL AG

=

dz{AV (z) (z)AV,G (z)AV,L }

Density functional theory: A = A[] , = (r)

A min. wrt. (r) variations, subject to constraint

Equil:

A
(r)

0 (r)


V

dr(r) = N

= chemical potential constant through interface.

0 (r) is equil. density prole.

Lets now think about the form of A[]. If we are near Tc , then L G
will be small. Further, 0 (z) c will be uniformly small. Thus, we can Taylor
expand about a uniform uid with density c :



A
A[] A[c ] + dr (r) [(r) c ]
c

  2 A
1
+ 2! dr dr (r)(r
) [(r) c ][(r ) c ] + . . .
c

A
Now, (r)

is evidently position independent in a uniform uid, so it is a

c

constant that can be lumped in with the chemical potential in []. The

2 A
can be shown to be equal to kB T [1
two-point function (r)(r

)
c (r
c

r )] c(r r )], where c is the direct CF of the uniform uid! Switching to

(r) = [(r) c ]/c :

A[] A[0]+ 21 c  dr2 (r)
21 2c dr dr c(|r r |)(r)(r ) + . . .
In the last term, we now Taylor expand (r ) about (r): (r ) (r) + |r
(r r) + . . . This leads to:


1
1
2
A[] A[0] + c dr[1 c c(0)] (r) + c dr||2 + . . .
2
2
where

dr12 c(r12 = lim dreikr c(r)

k0

2
16 c dr12 r12
c(r12 ) square gradient coe.

c(0)

both of these are nite for short-ranged c(r). Notice that [1c c(0)] t vanishes
at Tc and, indeed, the rst two terms are the beginning of our Landau expansion
of AV for a homogeneous system:


1
A[] = dr
A0V ((r))
+ kB T c dr||2 gradient

 
2
local
a2 (t) 2 (r)
+a4 (t) 4 (r) + . . .
Now, for a 1-D interface along z:

S
1
[] A [] =
dz{A0V ((z)) + kB T c [  (z)]2
2

(z)AV,L (z)AV,G }
Our DFT principle says that the equilibrium tension and prole (z) is given
by:
[]
=0
(z)
and we choose zd = 0 such that N S = 0, or


dz[(z)] ()(z)
(z)]
= 0 odd prole (z) = (z)

Lets now carry out the minimization, using the Landau expanded form of
F ()
A0V () for the VDW uid. Recall A0V () = pc

 
3t 2 + 38 4 +...


Noting that AV,L = AV,G = A0V ():

+ 4 [  (z)]2 }
[] = pc
dz{F () F ()
3

where we have used c kpcB T pcc c = z1c = 83

= 3 ( 2 2 )2 , so
Finally, we note that F () F ()
8

3
4
[] = pc
dz{ ( 2 (z) 2 )2 + [  (z)]2 }
8
3

At equilibrium (Euler-Lagrange equation):

1
3
8
=
= ( 2 (z) 2 )(z)  (z)
pc (z)
pc
2
3
We are interested in the solution of the ODE subject to the BCs: (z) z

and  ,  z
0. To be consistent, clearly = 0 and recall that this is
consistent with the bulk chemical potential of our near-critical uid:

Thus, we must solve

8 
3
= ( 2 2 )
3
2
We note that this is an Euler equation, so we multiply by  :
8  
4
3
3
= [(  )2 ] = 3  2 


3
3
2
  2
1
4

4 ( )

or

4
3
3
(  )2 4 + 2 2
3
8
4

1
2

2 ( )


=0

Evidently, the quantity in [. . . ] is constant, we can evaluate it by taking z ,

:
4
3
3
3
(  )2 4 + 2 2 4
3
8
4
8
32
(  )2 = ( 2 2 )2
9
3
 = ( 2 2 ) ,  sign is physical root.
4 2
6

 z
d 
3

=
dz  =
2 ( 2 )2
4 2 0
0




+
1

= 3 ln + =
ln
2 4 2



(z)

3
z
4 2

3z
z/
2 2

ez/ 1
+ = ( )ez/ , = z/
, or
e
+1
(z) = tanh(z/2) , =

2 s

(t)1/2

Thus, the equilibrium density prole is a tanh

with a width that is of the same scale as the bulk correlation length for
a homogeneous system! We also note that since the shaded areas are equal,
N S = 0, so the Gibbs dividing surface indeed is at z = 0.
Next, we turn to the tension. Noting that 43 (  )2 = 38 ( 2 2 )2 for the
extremum prole,


8
8 pc 2
dz[  (z)]2 =
dx [tanh (x)]2
= pc

 
3
3 2

sec h4 (x)



4/3

or
=

2
16 pc
9

The scaling of this tension near Tc is thus

2
t , = 3/2

where we have used MF exponents. Thus vanishes continuously as T Tc !

This formula for is quite reasonably close, but not too close to the critical
point, where the gradient expansion can be justied (since  1 is then
1
). Very close to Tc , however, thermal uctuations of cause this
small c.f. LJ
formula to fail.
The proper critical scaling of can be obtained from a simple scaling argument, however. Since dimensionally, has units [] = energy
area we might guess:

kB T
2

t2v t1.26
7

This is indeed the proper scaling!

As a nal note, we point out that away from the critical region, more sophisticated DFT approaches, similar to ours but without the gradient expansion,
have proved quite accurate. These methods are also very useful in treating
liquid-solid transitions!

C. Capillary Waves
Fluctuations other than critical uctuations turn out to be important at all
nonzero temperatures for L-G and L-L interfaces. These are long-wavelength
undulations of the interface position, e.g., in a L-G system

Because of these waves, measurements of interfacial thickness always give

values larger than the intrinsic width, . Lets see how to analyze this interfacial
broadening.
Consider small displacements (x, y) of the Gibbs dividing surface of an
interface:

The extra surface energy associated with additional surface area of the distorted interface is:




1 + (x )2 + (y )2 1
E = dx dy


2 dx dy()2
Also opposing the height uctuations is the gravitational potential energy:


1
Eg = dx dy m(L G )g[(x, y)]2
2
 

If we superimpose these two, we get an eective Hamiltonian that can be used

to study the SM of capillary waves:


1
H[] =
dx dy{()2 + (mg) 2 }
2
8

We can connect to average properties via


Q = DeH[]

1
 (x, y) =
D[(x, y)]2 eH[]
Q
is the mean-squared height uctuation due to capillary waves.
so

Lecture 18
Recap:
Capillary wave uctuations (x, y)

H[] =

1
2

dy{||2 + (mg) 2 }

 
 
gravity
surface
tension

dx

mean-squared high t uids

1
 (r) = D[][(r)]2 eH[]
Q
2


Q=

D[] H[]

As we have done before, it is convenient to dene such functional integrals

by going to a Fourier representation. Here we use the 2-D representation:
1  ikr
e
k , r = (x, y)
(r) = 2
L
k

where, e.g.
2nx
, nx = 0, 1, 2, . . .
L
So we take our region of interface to be a square patch with area L2 and impose
PBCs. This leads to
 2
H[] =
{k + lambda2
c }k k
2L2
kx =

1/2



where c = mg
is the capillary length (typically 1 mm) and D

dk . In this representation, the interface broadening is given by:
k

w2  2 (r) =

1   i(k+q)r
e
k q
L4
q
k

However, due to the fact that H[] is diagonal in the Fourier representation:
k q = k,q
Thus,

L2 kB T
[k 2 + 2
c ]

1
+ 2
c
k
 2
kB T
1
1

d
k k2 +
2
L (2)2
c

k
kB T
1
2
W = (2)2 2
dk 2
, cuto
k + 2
c )
0




W 2=

kB T 1
L2

1 kB T
4

W =

k2

ln(1 +

1
2 2
2 ln(1+ c )
2 2
c )

As c 1 mm, 1/10
A 1/; c  1 W 2

1 kB T
2

ln(c ) , ln c 10

A couple of things to note:

1. The measured width of the interface is + W
2. The capillary wave contribution increases with T and with
1/2

3. As T Tc , W kBT
tv

10

Advanced Topics in Equilibrium Statistical

Mechanics

10. Electrolyte and Colloidal Solutions

A. Systems and Models
Thus far, we have restricted our attention to the equilibrium SM of simple
uids and polymers. However, there are many other systems of general interest
to chemical engineers:
colloidal
solutions:

1-1 electrolyte
solution:

polyelectrolyte
solution:
We model these systems a bit dierently, depending on the lengthscales
involved. In many cases, it is convenient and accurate to replace the solvent
by a continuum, so that we do not have the added computational expense
of describing the solvent structuring around the solute. Lets start with the
simplest case of a solution of neutral particles.
Imagine that we have a solution of non-charged, neutral molecules or colloids,
dissolved in a low molecular weight solvent:

If such systems are dilute in solute, then we might think about using a virial
expansion to address the thermodynamic properties. The eective interactions
between solute molecules would correspond to the potential of mean force w(r)
between two solute molecules separated by some distance r in the pure solvent.
Indeed there is a theory known as the McMillan-Mayer theory that makes a
rigorous connection between dilute gases and dilute nonelectrolyte solutions:
Gases

p
=+
Bn (T )n EOS
kB T
n=2
u(r)

pair potential in vacuum

Dilute Non-electrolyte solutions

=+
Bn (T, )n EOS
kB T
n=2

w(rj ): potential of mean force between two solutes at separation r in pure

solvent at chemical potential .
: solute concentration # density)
: osmotic pressure of the solution
The Bn are related to w in precisely the same way that Bn are related to u,
e.g

1
B2 (T, ) =
dr[ew(r,) 1]
2
Thus, dilute non-electrolyte solutions are no more dicult to deal with than
gases, provided we can make good models of w(r; ). For, large ( 1 m)
colloidal particles, it is sensible to neglect the granularity of the solvent and
model w by a simple hard sphere potential.
For sterically stabilized colloids with grafted polymers, we might want
to employ a softer repulsive potential with a longer range starting at 2R2 .
Finally, for molecular solutes, one might do a MC or MD simulation of 2 solutes
in the solvent to explore w(r; ).
Next lets consider a simple electrolyte solution, e.g. NaCl in watera socalled 1-1 electrolyte.

The crudest description of such a solution is the so-called primitive model:

2

 Treat solvent as a continuum with dielectric constant .

Let 1 be the cation (+) species and 2 be the anion() species. Model
their interactios as the sum of a hard sphere and coulomb potential: e.g.,

,
r < 12 (1 + 2 )
w12 (r, ) = q1 ,q2
1
r , r > 2 (1 + 2 )
Here, 1 and 2 are the cation and anion hard sphere diameters; q1 and
q2 are the charges (q1 = +e, q2 = e for NaCl).
This model clearly makes a drastic approximation for the short-ranged part of
wij (r, ). However, we shall work with it, since we shall see that the long-ranged
coulomb interaction is actually dominant at low concentrations.
Lets now consider the so-called restricted primitive model where 1 = 2 =
for simplicity. One might think that the McMillan-Mayer theory should apply,
so we could develop a virial expansion:
2

2

= 1 + 2 +
Bij
(T, )i j + . . .
kB T
i=1 j=1

(T, )
Bij

1
=
2

dr[eWij (rj ) 1]

where Bij
is a 2 2 matrix of second virial coecients. The problem with this
is that these virial coecients diverge for the coulomb potential:
qi qj

2
Bij
int
e r 1 ] =
dr r [ 


r1 for r



r for r

Clearly virial expansions fail for electrolyte solutions! A new approach is needed.
We shall see that the problem is that the osmotic pressure is not analytic in the
i for charged solutions. Instead,

3/2
1 + 2 + (i )
kB T
3/2

The Debye-H
uckel theory was developed to evaluate this (i ) electrostatic
term.

B. Debye-H
uckel Theory
There are two ways to deal with the divergences caused by electrostatic interactions:
3

 Mean-eld approximationsdescribed in Chapter 15 of McQuarrie

Summing innite classes of divergent diagramssee the Andersen handout.
We shall follow the second approach here, although both lead to the same limiting corrections to ideal gas behavior for dilute electrolytes.
Recall from the Andersen article that
Aex [A(T, V, N1 , N2 ) Aid (T, V, N1 , N2 )]
has the simple diagrammatic series:
sum of top distinct, irreducible
graphs with two or more -eld
Aex =
points, and at most one f -bond
between each pair of points.
=

Note that irreducible connected + no articulation vertices

Lecture 19
Debye-H
uckel Theory for electrolyte solution:
Aex =

where fij (r) = eWij (r) 1 , species 1, e.g., Na + species 2, e.g., Cl

Wij (r) = Wijd (r) +

hard sphere d =

qi qj r
r e

(r )
  

0 r<
=
1 r>

convergence factor 0+
4

In this expression, each fij bond represents:

fij (r) = eWij (r) 1
and we sum over all species indices when we evaluate diagrams. For example:
=


2 
2
1


dr1 dr2 fij (r12 )i j
2 i=1 j=1

Lets now write

Wij (r) =

qi qj r
e
Wijd (r) +
(r )
  
  
 r  

1, r >
hard sphere coulomb with a
=
0, r <
d=
convergence factor,
0+

By expanding the Coulomb potential:

d
d
fij = fij
+ [1 + fij
]

(ij )n /n!,

n=1

d
eWij 1 f d ( )
fij

ij

qi qj r
(r
r e

bond

)(

bond

It can be argued that this substitution in the diagrams leads to (see Andersen):
sum of top. distinct, irreducible
graphs that have 2 or more -eld
Aex = points, and at most one f d -bond
and any number of -bonds between
each pair of points
= Aex
d +
d
where Aex
d contains all the diagrams with only dashed f bonds. Notice that if
 bonds connect any two points, there is now a factor of 1
in the value of a
diagram!
Lets now look at some individual diagrams:

Aex
+
d 
 
V 3 2

(V 3 )
  
dominant term for
3  1
dilute

1
2

2

2 


i


dr2 ij (r12 )i j

dr1

= 21 V ( qi i )2


r>

But,

2

dr

1 r
e
r

qi i = 1 e2 = 0 which is a statement of electroneutrality. Thus

i=1

= 0. Now,

1
2
22! V

= 12 V

qi2 i

2 

r>

2 
qi i

dr

e2r


for 0

2 r2

drfd (r)

1 r
e
(r )
r





0
0

2 


1
1
= 23!
3V
qi3 i
dr 3 3 e3r
 r
r>
i




 
ln() , 0
0 for
11 electrolyte

3





2
1 3
dr12 dr23 13 r12 r113 r23
= 6 V
qi i


e(r12 +r13 +r23 )

(r12 )(r13 )(r23 )



3 for 0
By following this path, it can be shown that the largest diagrams for 0 at
each order in are the ring diagrams:
Aex
ring =

We can evaluate this series by using Fourier transforms; for 0, we can

ignore the -functions. Notice these diagrams all look like:
6

So lets group things this way, dening

4
2
1 r
2
e
qi i ; (r)


4r
i
Then

 1

2 2
Aex
dr12 (r12 22
(r12 )
ring = V ( )

1 2
+(1)3 32
dr3 (r13 )(r32 )


2 
)
+(1)4 (
dr3 dr4 (r13 )(r34 )(r42 ) + . . .
42

or

2 2
Aex
ring = V ( )

=V


(k)=

dr12 (r12 )C(r12 )


2 2 1 3

dk(k)
C(k)
( ) 2
1
ikr 1 r
= 2 +k2
dre
4r e

By the convolution theorem:

1)3 2
1
2
22 (k) + 32 [(k)]
4
(1)
2 2
3
+ 42 ( ) [(k)] + . . .
2
3
(1)4
2
2
+ (1)
= 12 (1)
2
3 +
4 ( )



(1)n

n
= 12 (2)
(2 )
2
n
n=2

C(k)=

So
Aex
ring =



2 )

2 +ln(1

V
2(2)3

2 + . . .

dk{2 (k)
+ ln(1 2 (k))}

Let q = k/, 0+

aex
ring
3
= 2(2)
3
V

3
12

dq{q 2 + ln(1 q 2 }

3/2

A few comments:
1. The sume of the ring diagrams is nite for 0 (coulomb potential), even
though the individual terms are divergent. Mayer (see Andersen) showed
that the entire Aex series can be renormalized, eectively replacing
bonds by C bonds and eliminating all long-ranged divergences.
2. All remaining diagrams in the renormalization series, as well as Aex
d /V ,
are (2 ) which are smaller than the ring sum (3/2 ). Thus the ring
sum gives the rst correction to IG behavior for 0. Note that the
nite size of the ions, , does not enter until (2 ).
7

In summary, we have derived the following Debye-H

uckel expression for
AV = A/V :




2
i ln i +
i (ln 3T,i 1)
AV =
12
i
i



ideal gas
+(2 )
where
2 =
and

1
2
8
I , I
q i ionic strength

2 i i

D 1 I 1/2 Debye screening length

We shall discuss the physical meaning of 0 = 1 shortly, but rst we note

that

A
= V
T,Ni

AV
= AV +
i
i
i
3

+ (2 )
=
i
24
i
as previously announced! The chemical potentials are:

A
q 2
i
= ln(i 3T,i ) i
Ni V,T
2
Recalling the denition of the activity coecient i :
i = k T ln(3T,i i i )
we nd a purely electrostatic contribution:
q 2

ln i = 2ki T

As shown in the attached gures, this limiting law works nicely at low concentrations, but typically breaks down around 0.01 moles/.
Our next task is to explain the signicance of the Debye length D = 1 .
We shall do this by studying the structure of an electrolyte solution, i.e. gij (r).
First, however, note that the electrostatic contribution to both AV and can
be written:
kB T
e
3
Ae
V kB T 3
D
This is reminiscent of our blob picture of critical phenomena and semi-dilute polymer solutions!

Our starting point for examining the structure of a dilute electrolyte solution
is the OZ equation:

2
hij (r12 ) = cij (r12 ) + k=1 k dr3 cik (r13 )hkj (r32 )
In the dilute limit of i 0,
0

cij (r) fij (r) = eWij (r) 1

where Wij0 (r) is the potential of mean force between two ions in the pure solvent.
Moreover, since ions are far apart in the dilute limit:
qi q j
cij (r) Wij0 (r) =
r
Similarly, we can write
hij (r) gij (r) 1 eWij (r) 1
Wij (r)
where Wij (r) is the potential of mean force between two ions when surrounded
by other ions at concentration {i , 2 }. With these approximations, the OZ
equation becomes

2


qi qk
qi qj
Wij (r12 ) =

k dr3
Wkj (r32 )
r12
r13
k=1

Dene Wij (r)

qi qj
 W (r)

W (r12 ) =

1
2

r12
4


dr3

1
W (r32 )
r13

This is a linear integral equatio that can be solved by Fourier transforms:

(k)=
W
(k)=
W
W (r)=

4
2 1
k2 k2 W (k)
4
k2 +2

1
(k)
dkeikr W
(2)3

er
r

q q

i j r/D
Wij (r) = r
e
Thus, we see that while the bare coulomb potential is long-ranged:

1
r
the potential of mean force between two ions is screened by the surrounding
ions:
1
W (r) = er/D
r
On lengthscales greater than D = 1 I 1/2 , two ions no longer feel each
other! The physical picture is:
W 0 (r) =

 Notice that D 30
A at = 0.01 mole/, which explains why 3
A is
not relevant.
Also note that the thermodynamics is evidently also controlled by these
3
clouds, which play the role of blobs, e Ae
V kB T /D !

Lecture 20
Recap: Debye-H
uckel Theory
For dilute electrolyte solutions:
3
2
Aex
)
V = 12 + (


8
1
2
qi2 i ionic strength
=  I , I2
i

D 1 I 1/2 Debye length

qi qj
bare coulomb inter.
Wij0 = r
qi qj r/D
potential of mean force
Wij (r) = r e

C. Application to Charged Colloids

A common strategy for preventing the aggregatio nof colloidal particles in solution (due to attractive VDW interactiosn) is to immobilize charges on their
surface. These charges are almost always derived from a neutral species that
dissociates, so counter ions are also present:

This situation can immediately be treated by using Debye-H

uckel theory,
provided:
1. The system is dilute in solute (colloid), i.e., s  1
2. The surface density of charge on the colloid is not too high so that the
local charge density in solution does not invalidate DH theory.
10

If these conditions are met, and the total charge on the colloid is +Q and the
counter ions are monovalent, then we can view the solution as a Q1 electrolyte:

Provided D , we can simply model the eective interactions between

two colloids as:
w(r) =

Q2 r/D
e
+
W S (r)
  
r  
short-ranged potential long-ranged
[hard-sphere (dia. )
electrostatic
and attractive VDW]
repulsion

We could then use W (r) as an eective pair potential and carry out MD or MC
simulatiosn to determine structure and thermo props of the solution. In such
a simulation, only the colloid positions are retainedthe water molecules and
counter ions enter only as a continuum through W (r)! Notice that the total
ionic strength enters D :




4
2
2
2
qi i
D = = 
i

Q2 + e2 Qe
(Q2 + eQ)

Q 1 electrolyte
No added salt

where is the colloid number density. It is common to add an extra free salt,
e.g. NaCl to increase the ionic strength and thereby decrease D the range of
W (r). Then:


2
D
I (Q2 + eQ) +
qS2 S
S

11