G25.

2651: Statisti al Me hani s

Notes for Le ture 1

Dening statisti al me hani s: Statisti al Me hani s provies the onne tion between mi ros opi motion of indi-
vidual atoms of matter and ma ros opi ally observable properties su h as temperature, pressure, entropy, free energy,
heat apa ity, hemi al potential, vis osity, spe tra, rea tion rates, et .
Why do we need Statisti al Me hani s:
1. Statisti al Me hani s provides the mi ros opi basis for thermodynami s, whi h, otherwise, is just a phenomeno-
logi al theory.
2. Mi ros opi basis allows al ulation of a wide variety of properties not dealt with in thermodynami s, su h
as stru tural properties, using distribution fun tions, and dynami al properties { spe tra, rate onstants, et .,
using time orrelation fun tions.
3. Be ause a statisti al me hani al formulation of a problem begins with a detailed mi ros opi des ription, mi-
ros opi traje tories an, in prin iple and in pra ti e, be generated providing a window into the mi ros opi
world. This window often provides a means of onne ting ertain ma ros opi properties with parti ular modes
of motion in the omplex dan e of the individual atoms that ompose a system, and this, in turn, allows for
interpretation of experimental data and an elu idation of the me hanisms of energy and mass transfer in a
system.

I. THE MICROSCOPIC LAWS OF MOTION

Consider a system of N lassi al parti les. The parti les a onned to a parti ular region of spa e by a \ ontainer"
of volume V . The parti les have a nite kineti energy and are therefore in onstant motion, driven by the for es
they exert on ea h other (and any external for es whi h may be present). At a given instant in time t, the Cartesian
positions of the parti les are r1 (t); :::; rN (t). The time evolution of the positions of the parti les is then given by
Newton's se ond law of motion:
mi ri = Fi (r1 ; :::; rN )
where F1 ; :::; FN are the for es on ea h of the N parti les due to all the other parti les in the system. The notation
ri = d2 ri =dt2 .
N Newton's equations of motion onstitute a set of 3N oupled se ond order dierential equations. In order to solve
these, it is ne essary to spe ify a set of appropriate initial onditions on the oordinates and their rst time derivaties,
fr1 (0); :::; rN (0); r_ 1 (0); :::; r_ N (0)g. Then, the solution of Newton's equations gives the omplete set of oordinates and
velo ities for all time t.

II. THE ENSEMBLE CONCEPT (HEURISTIC DEFINITION)

For a typi al ma ros opi system, the total number of parti les N  1023. Sin e an essentially innite amount
of pre ision is needed in order to spe ify the initial onditions (due to exponentially rapid growth of errors in this
spe i ation), the amount of information required to spe ify a traje tory is essentially innite. Even if we ontented
ourselves with quadrupole pre ision, however, the amount of memory needed to hold just one phase spa e point would
be about 128 bytes = 27  102 bytes for ea h number or 102  6  1023  1017 Gbytes. The largest omputers we
have today have perhaps 103 Gbytes of memory, so we are o by 14 orders of magnitude just to spe ify 1 point in
phase spa e.
Do we need all this detail? (Yes and No).

1
Yes { There are plenty of hemi ally interesting phenomena for whi h we really would like to know how individual
atoms are moving as a pro ess o urs. Experimental te hniques su h as ultrafast laser spe tros opy an resolve
short time s ale phenomena and, thus, obtain important insights into su h motions. From a theoreti al point of
view, although we annot follow 1023 parti les, there is some hope that we ould follow the motion of a system
ontaining 104 or 105 parti les, whi h might apture most of the features of true ma ros opi matter. Thus, by
solving Newton's equations of motion numeri ally on a omputer, we have a kind of window into the mi ros opi
world. This is the basis of what are known as mole ular dynami s al ulations.
No { Intuitively, we would expe t that if we were to follow the evolution of a large number of systems all des ribed by
the same set of for es but having starting from dierent initial onditions, these systems would have essentially
the same ma ros opi hara teristi s, e.g. the same temperature, pressure, et . even if the mi ros opi detailed
evolution of ea h system in time would be very dierent. This idea suggests that the mi ros opi details are
largely unimportant.
Sin e, from the point of view of ma ros opi properties, pre ise mi ros opi details are largely unimportant, we might
imagine employing a onstru t known as the ensemble on ept in whi h a large number of systems with dierent
mi ros opi hara teristi s but similar ma ros opi hara teristi s is used to \wash out" the mi ros opi details via
an averaging pro e ure. This is an idea developed by individuals su h as Gibbs, Maxwell, and Boltzmann.
Ensemble: Consider a large number of systems ea h des ribed by the same set of mi ros opi for es and sharing some
ommon ma ros opi property (e.g. the same total energy). Ea h system is assumed to evolve under the mi ros opi
laws of motion from a dierent initial ondition so that the time evolution of ea h system will be dierent from all
the others. Su h a olle tion of systems is alled an ensemble. The ensemble on ept then states that ma ros opi
observables an be al ulated by performing averages over the systems in the ensemble. For many properties, su h as
temperature and pressure, whi h are time-independent, the fa t that the systems are evolving in time will not ae t
their values, and we may perform averages at a parti ular instant in time. Thus, let A denote a ma ros opi property
and let a denote a mi ros opi fun tion that is used to ompute A. An example of A would be the temperature, and
a would be the kineti energy (a mi ros opi fun tion of velo ities). Then, A is obtained by al ulating the value of
a in ea h system of the ensemble and performing an average over all systems in the ensemble:
1X N
A=
N a =1
where N is the total number of members in the ensemble and a is the value of a in the th system.
The questions that naturally arise are:
1. How do we onstru t an ensemble?
2. How do we perform averages over an ensemble?
3. How many systems will an ensemble ontain?
4. How do we distinguish time-independent from time-dependent properties in the ensemble pi ture?
Answering these questions will be our main obje tive in this ourse.

III. THE LAGRANGIAN FORMULATION OF CLASSICAL MECHANICS

In order to begin to make a onne tion between the mi ros opi and ma ros opi worlds, we need to better understand
the mi ros opi world and the laws that govern it. We will begin pla ing Newton's laws of motion in a formal framework
whi h will be heavily used in our study of lassi al statisti al me hani s.
First, we begin by restri ting our dis ussion to systems for whi h the for es are purely onservative. Su h for es are
derivable from a potential energy fun tion U (r1 ; :::; rN ) by dierentiation:
U
Fi =
 ri

2
It is lear that su h for es annot ontain dissipative or fri tion terms. An important property of systems whose for es
are onservative is that they onserve the total energy
XN
E = K + U = 21 mi r_ 2i + U (r1 ; :::; rN )
i=1

To see this, simply dierentiate the energy with respe t to time:

dE X N X
N
U
= mi r_ i
ri +
r_
dt i=1
r i i=1 i

X
N X
N
= r_ i
Fi Fi
r_ i
i=1 i=1

=0
where, the se ond line, the fa ts that ri = Fi =mi (Newton's law) and Fi = U= ri ( onservative for e denition)
have been used. This is known as the law of onservation of energy.
For onservative systems, there is an elegant formulation of lassi al me hani s known as the Lagrangian formulation.
The Lagrangian fun tion, L, for a system is dened to be the dieren e between the kineti and potential energies
expressed as a fun tion of positions and velo ities. In order to make the nomen lature more ompa t, we shall
introdu e a shorthand for the omplete set of positions in an N -parti le system: r  r1 ; :::; rN and for the velo ities:
r_  r_ 1 ; :::; r_ N . Then, the Lagrangian is dened as follows:
XN
1 m r_ 2 U (r ; :::; r )
L(r; r_ ) = K U = i i 1 N
i=1
2

In terms of the Lagrangian, the lassi al equations of motion are given by the so alled Euler-Lagrange equation:
  L
d L
dt  r_ i  ri
=0
The equations that result from appli ation of the Euler-Lagrange equation to a parti ular Lagrangian are known as
the equations of motion. The solution of the equations of motion for a given initial ondition is known as a traje tory
of the system. The Euler-Lagrange equation results from what is known as an a tion prin iple. We shall defer
further dis ussion of the a tion prin iple until we study the Feynman path integral formulation of quantum statisti al
me hani s in terms of whi h the a tion prin iple emerges very naturally. For now, we a ept the Euler-Lagrange
equation as a denition.
The Euler-Lagrange formulation is ompletely equivalent to Newton's se ond law. In order to see this, note that
L
 r_ i
= mi r_ i
L U
r i
=  ri
= Fi
Therefore,
 
d L L
dt  r_ i  ri
= miri Fi = 0
whi h is just Newton's equation of motion.
An important property of the Lagrangian formulation is that it an be used to obtain the equations of motion of a
system in any set of oordinates, not just the standard Cartesian oordinates, via the Euler-Lagrange equation (see
problem set #1).

3
 IV. THE HAMILTONIAN FORMULATION OF CLASSICAL MECHANICS

The Lagrangian formulation of me hani s will be useful later when we study the Feynman path integral. For our
purposes now, the Lagrangian formulation is an important springboard from whi h to develop another useful formu-
lation of lassi al me hani s known as the Hamiltonian formulation. The Hamiltonian of a system is dened to be
the sum of the kineti and potential energies expressed as a fun tion of positions and their onjugate momenta. What
are onjugate momenta?
Re all from elementary physi s that momentum of a parti le, pi , is dened in terms of its velo ity r_ i by
pi = mi r_ i
In fa t, the more general denition of onjugate momentum, valid for any set of oordinates, is given in terms of the
Lagrangian:
L
pi =
 r_ i
Note that these two denitions are equivalent for Cartesian variables.
In terms of Cartesian momenta, the kineti energy is given by
XN
p2i
K=
i=1
2mi
Then, the Hamiltonian, whi h is dened to be the sum, K + U , expressed as a fun tion of positions and momenta,
will be given by
XN
p2i
H (p; r) =
i=1
2mi + U (r1 ; :::; rN ) = H (p; r)
where p  p1 ; :::; pN . In terms of the Hamiltonian, the equations of motion of a system are given by Hamilton's
equations:
H H
r_ i = p_ i =
 pi  ri
The solution of Hamilton's equations of motion will yield a traje tory in terms of positions and momenta as fun tions
of time. Again, Hamilton's equations an be easily shown to be equivalent to Newton's equations, and, like the
Lagrangian formulation, Hamilton's equations an be used to determine the equations of motion of a system in any
set of oordinates.
The Hamiltonian and Lagrangian formulations possess an interesting onne tion. The Hamiltonian an be dire tly
obtained from the Lagrangian by a transformation known as a Legendre transform. We will say more about Legendre
transforms in a later le ture. For now, note that the onne tion is given by
XN
H (p; r) = pi
r_ i L(r; r_ )
i=1
whi h, when the fa t that r_ i = pi =mi is used, be omes
XN
pi XN
1
 p 2
H (p; r) = pi
mi i
+ U (r1 ; :::; rN )
i=1
mi i=1
2 m i
XN 2
= 2pmi + U (r1 ; :::; rN )
i=1 i

Be ause a system des ribed by onservative for es onserves the total energy, it follows that Hamilton's equations
of motion onserve the total Hamiltonian. Hamilton's equations of motion onserve the Hamiltonian

4
 H (p(t); r(t)) = H (p(0); r(0)) = E
Proof: H = onst ) dH=dt = 0
N 
X 
dH H
dt
=
r_ + H
p_
 ri i  pi i
i=1
N 
X 
H H H H
=

 ri  pi


 pi  ri
=0
i=1

QED. This, then, provides another expression of the law of onservation of energy.

V. PHASE SPACE

We onstru t a artesian spa e in whi h ea h of the 6N oordinates and momenta is assigned to one of 6N mutually
orthogonal axes. Phase spa e is, therefore, a 6N dimensional spa e. A point in this spa e is spe ied by giving a
parti ular set of values for the 6N oordinates and momenta. Denote su h a point by
x = (p1 ; :::; pN ; r1 ; :::; rN )
x is a 6N dimensional ve tor. Thus, the time evolution or traje tory of a system as spe ied by Hamilton's equations
of motion, an be expressed by giving the phase spa e ve tor, x as a fun tion of time.
The law of onservation of energy, expressed as a ondition on the phase spa e ve tor:
H (x(t)) = onst = E
denes a 6N 1 dimensional hypersurfa e in phase spa e on whi h the traje tory must remain.

A. Classi al mi ros opi states or mi rostates and ensembles

A mi ros opi state or mi rostate of a lassi al system is a spe i ation of the omplete set of positions and momenta
of the system at any given time. In the language of phase spa e ve tors, it is a spe i ation of the omplete phase spa e
ve tor of a system at any instant in time. For a onservative system, any valid mi rostate must lie on the onstant
energy hypersurfa e, H (x) = E . Hen e, spe ifying a mi rostate of a lassi al system is equivalent to spe ifying a
point on the onstant energy hypersurfa e.
The on ept of lassi al mi rostates now allows us to give a more formal denition of an ensemble.
An ensemble is a olle tion of systems sharing one or more ma ros opi hara teristi s but ea h being in a unique
mi rostate. The omplete ensemble is spe ied by giving all systems or mi rostates onsistent with the ommon
ma ros opi hara teristi s of the ensemble.
The idea of ensemble averaging an also be expressed in terms of an average over all su h mi rostates (whi h omprise
the ensemble). A given ma ros opi property, A, and its mi ros opi fun tion a = a(x), whi h is a fun tion of the
positions and momenta of a system, i.e. the phase spa e ve tor, are related by
1 XN
A = haiensemble =
N a(x )
=1

where x is the mi rostate of the th member of the ensemble.

However, re all the original problem of determining the mi ros opi detailed motion of ea h individual parti le in a
system. In reality, measurements are made only on a single system and all the mi ros opi detailed motion is present.
However, what one observes is still an average, but it is an average over time of the detailed motion, an average that
also washes out the mi ros opi details. Thus, the time average and the ensemble average should be equivalent, i.e.

5
 1 Z T
A = haiensemble = lim dt a(x(t))
T !1 T 0
This statement is known as the ergodi hypothesis. A system that is ergodi is one for whi h, given an innite amount
of time, it will visit all possible mi ros opi states available to it (for Hamiltonian dynami s, this means it will visit
all points on the onstant energy hypersurfa e). No one has yet been able to prove that a parti ular system is truly
ergodi , hen e the above statement annot be more than a supposition. However, it states that if a system is ergodi ,
then the ensemble average of a property A(x) an be equated to a time average of the property over an ergodi
traje tory.

VI. PHASE SPACE DISTRIBUTION FUNCTIONS AND LIOUVILLE'S THEOREM

Given an ensemble with many members, ea h member having a dierent phase spa e ve tor x orresponding to a
dierent mi rostate, we need a way of des ribing how the phase spa e ve tors of the members in the ensemble will be
distributed in the phase spa e. That is, if we hoose to observe one parti ular member in the ensemble, what is the
probability that its phase spa e ve tor will be in a small volume dx around a point x in the phase spa e at time t.
This probability will be denoted
f (x; t)dx
where f (x; t) is known as the phase spa e probability density or phase spa e distribution fun tion. It's properties are
as follows:
Z f (x; t)  0
dxf (x; t) = Number of members in the ensemble

Liouville's Theorem: The total number of systems in the ensemble is a onstant. What restri tions does this pla e
on f (x; t)? For a given volume
in phase spa e, this ondition requires that the rate of de rease of the number of
systems from this region is equal to the ux of systems into the volume. Let n^ be the unit normal ve tor to the
surfa e of this region.

6
 n dS

Volume = Ω
FIG. 1.

The ux through the small surfa e area element, dS is just n^
x_ f (x; t)dS . Then the total ux out of volume is obtained
by integrating this over the entire surfa e that en loses
:
Z Z
dS n^
x_ f (x; t)) = rx
(_xf (x; t))

whi h follows from the divergen e theorem. rx is the 6N dimensional gradient on the phase spa e
    

rx =  p ; :::;  p ;  r ; :::;  r 1
1 N N

= (rp1 ; :::; rpN ; rr1 ; :::; rrN )

On the other hand, the rate of de rease in the number of systems out of the volume is
d
Z Z 
dxf (x; t) = dx f (x; t)
dt
t

Equating these two quantities gives

Z Z 
dxrx
(_xf (x; t)) = dx f (x; t)


t
But this result must hold for any arbitrary hoi e of the volume
, whi h we may also allow to shrink to 0 so that
the result holds lo ally, and we obtain the lo al result:

f (x; t) + rx
(_xf (x; t)) = 0
t

7
But
rx
(_xf (x; t)) = x_
rx f (x; t) + f (x; t)rx
x_
This equation resembles an equation for a \hydrodynami " ow in the phase spa e, with f (x; t) playing the role of a
density. The quantity rx
x_ , being the divergen e of a velo ity eld, is known as the phase spa e ompressibility, and
it does not, for a general dynami al system, vanish. Let us see what the phase spa e ompressibility for a Hamiltonian
system is:
X
N
rx
x_ = [rpi
p_ i + rri
r_ i ℄
i=1

However, by Hamilton's equations:

p_ i = rri H r_ i = rpi H
Thus, the ompressibility is given by
X
N
rx
x_ = [ rpi
rri H + rri
rpi H ℄ = 0
i=1

Thus, Hamiltonian systems are in ompressible in the phase spa e, and the equation for f (x; t) be omes
 df
f (x; t) + x_
rx f (x; t) = =0
t dt
whi h is Liouville's equation, and it implies that f (x; t) is a onserved quantity when x is identied as the phase
spa e ve tor of a parti ular Hamiltonian system. That is, f (xt ; t) will be onserved along a parti ular traje tory of
a Hamiltonian system. However, if we view x is a xed spatial label in the phase spa e, then the Liouville equation
spe ies how a phase spa e distribution fun tion f (x; t) evolves in time from an initial distribution f (x; 0).

8
 G25.2651: Statistical Mechanics

Notes for Lecture 2

I. THE LIOUVILLE OPERATOR AND THE POISSON BRACKET

From the last lecture, we saw that Liouville's equation could be cast in the form
@f + r
x_ f = 0
@t x

The Liouville equation is the foundation on which statistical mechanics rests. It will now be cast in a form that will be
suggestive of a more general structure that has a denite quantum analog (to be revisited when we treat the quantum
Liouville equation).
Dene an operator
iL = x_
rx
p
known as the Liouville operator (i = ?1 { the i is there as a matter of convention and has the eect of making L a
Hermitian operator). Then Liouville's equation can be written
@f + iLf = 0
@t
The Liouville operator also be expressed as
X  @H @ @H @ 
N

iL =
=1
i
@ p
@ r ? @ r
@ p  f:::; H g
i i i i

where fA; B g is known as the Poisson bracket between A(x) and B (x):
X  @A @B @A @B 
N

fA; B g = @r
@p ? @p
@r
i i i i
=1 i

Thus, the Liouville equation can be written as

@f + ff; H g = 0
@t
The Liouville equation is a partial dierential equation for the phase space probability distribution function. Thus,
it species a general class of functions f (x; t) that satisfy it. In order to obtain a specic solution requires more input
information, such as an initial condition on f , a boundary condition on f , and other control variables that characterize
the ensemble.

II. PRESERVATION OF PHASE SPACE VOLUME AND LIOUVILLE'S THEOREM

Consider a phase space volume element dx0 at t = 0, containing a small collection of initial conditions on a set of
trajectories. The trajectories evolve in time according to Hamilton's equations of motion, and at a time t later will
be located in a new volume element dx as shown in the gure below:
t

1
 q

dx t

dx0
p
FIG. 1.

How is dx0 related to dx ? To answer this, consider a trajectory starting from a phase space vector x0 in dx0 and
t

having a phase space vector x at time t in dx . Since the solution of Hamilton's equations depends on the choice of
t t

initial conditions, x depends on x0 :

t

x0 = (p1 (0); :::; p (0); r1 (0); :::; r (0))

N N

x = (p1 (t); :::; p (t); r1 (t); :::; r (t))

t N N

x = x (x10 ; :::; x60 )

i
t
i
t
N

Thus, the phase space vector components can be viewed as a coordinate transformation on the phase space from t = 0
to time t. The phase space volume element then transforms according to
dx = J (x ; x0 )dx0
t t

where J (x ; x0 ) is the Jacobian of the transformation:

t

(x1


x )
J (x ; x0 ) = @@ (x t
n
t
t 1



x )0
n
0

where n = 6N . The precise form of the Jacobian can be determined as will be demonstrated below.
The Jacobian is the determinant of a matrix M,
J (x ; x0 ) = det(M) = eTrlnM
t

whose matrix elements are

M = @x
i
t
ij
@ x0 j

Taking the time derivative of the Jacobian, we therefore have

2
 dJ = Tr M?1 dM  eTrlnM
dt dt
X X ?1 dM n n

=J M dt ij
ji

i=1 j =1

The matrices M?1 and dM=dt can be seen to be given by

M?1 = @ x0
i

@x
ij j
t

dM = @ x_
ji
j
t

dt @ x0 i

Substituting into the expression for dJ=dt gives

dJ = J X @ x0 @ x_
n i j
t

dt =1
@ x @ x0
i;j
j
t
i

X @ x0 @ x_ @ x
n i j k

=J t t

=1
@ x @ x @ x0
i;j;k
j
t
k
t
i

where the chain rule has been introduced for the derivative @ x_ =@ x0. The sum over i can now be performed:
j
t
i

X @ x0 @ x X ?1
n
X n n

= M M = M M?1 = 
i k
t

=1
i
@ x @ x0 =1
j
t
i
=1
i
ij ki

i
ki ij kj

Thus,
dJ = J X  @ x_ n j
t

dt =1
@x
j;k
jk k
t

X
= J @ x_ = J rx
x_
n j
t

j =1
x j
t

or
dJ = J r
x_
dt x

The initial condition on this dierential equation is J (0)  J (x0 ; x0 ) = 1. Moreover, for a Hamiltonian system
rx
x_ = 0. This says that dJ=dt = 0 and J (0) = 1. Thus, J (x ; x0 ) = 1. If this is true, then the phase space volume
t

element transforms according to

dx0 = dx t

which is another conservation law. This conservation law states that the phase space volume occupied by a collection
of systems evolving according to Hamilton's equations of motion will be preserved in time. This is one statement of
Liouville's theorem.
Combining this with the fact that df=dt = 0, we have a conservation law for the phase space probability:
f (x0 ; 0)dx0 = f (x ; t)dx t t

which is an equivalent statement of Liouville's theorem.

3
 III. LIOUVILLE'S THEOREM FOR NON-HAMILTONIAN SYSTEMS
The equations of motion of a system can be cast in the generic form
x_ =  (x)
where, for a Hamiltonian system, the vector function  would be

 (x) = ? @@H ; :::; ? @H ; @H ; :::; @H 
r1 @ r @ p1 @ p N N

and the incompressibility condition would be a condition on  :

rx
x_ = rx
 = 0
A non-Hamiltonian system, described by a general vector funciton  , will not, in general, satisfy the incompressibility
condition. That is:
r
x_ = r
 6= 0
x x

Non-Hamiltonian dynamical systems are often used to describe open systems, i.e., systems in contact with heat
reservoirs or mechanical pistons or particle reservoirs. They are also often used to describe driven systems or systems
in contact with external elds.
The fact that the compressibility does not vanish has interesting consequences for the structure of the phase space.
The Jacobian, which satises
dJ = J r
x_
dt x

will no longer be 1 for all time. Dening  = rx
x_ , the general solution for the Jacobian can be written as
Z t

J (x ; x0 ) = J (x0 ; x0 ) exp
t ds(x ) s
0

Note that J (x0 ; x0 ) = 1 as before. Also, note that  = d ln J=dt. Thus,  can be expressed as the total time derivative
of some function, which we will denote W , i.e.,  = W_ . Then, the Jacobian becomes
Z 
dsW_ (x )
t

J (x ; x0 ) = exp
t s
0

= exp (W (x ) ? W (x0 ))
t

Thus, the volume element in phase space now transforms according to

dx = exp (W (x ) ? W (x0 )) dx0
t t

which can be arranged to read as a conservation law:

e? W (xt )
dx = e?
t
W (x0 )
dx0
p exp(?W (x)). Introducing
Thus, we have a conservation law for a modied volume element, involvingpa \metric factor"
the suggestive notation pg = exp(?W (x)), the conservation law reads g(x dx = g(x0 dx0 . This is a generalized
t t

version of Liouville's theorem. Furthermore, a generalized Liouville equation for non-Hamiltonian systems can be
derived which incorporates this metric factor. The derivation is beyond the scope of this course, however, the result
is
@ (f pg) + rx
(_xf pg) = 0
We have called this equation, the generalized Liouville equation Finally, noting that pg satises the same equation as
J , i.e.,

4
 dpg = pg
dt
the presence of pg in the generalized Liouville equation can be eliminated, resulting in
@f + x_
r f = df = 0
@t x
dt
which is the ordinary Liouville equation from before. Thus, we have derived a modied version of Liouville's theorem
and have shown that it leads to a conservation law for f equivalent to the Hamiltonian case. This, then, supports the
generality of the Liouville equation for both Hamiltonian and non-Hamiltonian based ensembles, an important fact
considering that this equation is the foundation of statistical mechanics.

IV. EQUILIBRIUM ENSEMBLES

An equilibrium ensemble is one for which there is no explicit time-dependence in the phase space distribution
function, @f=@t = 0. In this case, Liouville's equation reduces to
ff; H g = 0
which implies that f (x) must be a pure function of the Hamiltonian
f (x) = F (H (x))
The specic form that F (H (x)) has depends on the specic details of the ensemble.
The integral over the phase space distribution function plays a special role in statistical mechanics:
Z
F = dxF (H (x)) (1)
It is known as the partition function and is equal to the number of members in the ensemble. That is, it is equal to
the number of microstates that all give rise to a given set of macroscopic observables. Thus, it is the quantity from
which all thermodynamic properties are derived.
If a measurement of a macroscopic observable A(x) is made, then the value obtained will be the ensemble average:
Z
hAi = 1 dxA(x)F (H (x))
F (2)
Eqs. (1) and (2) are the central results of ensemble theory, since they determine all thermodynamic and other
observable quantities.

A. Introduction to the Microcanonical Ensemble

The microcanonical ensemble is built upon the so called postulate of equal a priori probabilities:
Postulate of equal a priori probabilities: For an isolated macroscopic system in equilibrium, all microscopic
states corresponding to the same set of macroscopic observables are equally probable.

1. Basic denitions and thermodynamics

Consider a thought experiment in which N particles are placed in a container of volume V and allowed to evolve
according to Hamilton's equations of motion. The total energy E = H (x) is conserved. Moreover, the number of
particles N and volume V are considered to be xed. This constitutes a set of three thermodynamic variables N; V; E
that characterize the ensemble and can be varied to alter the conditions of the experiment.

5
 The evolution of this system in time generates a trajectory that samples the constant energy hypersurface H (x) = E .
All points on this surface correspond to the same set of macroscopic observables. Thus, by the postulate of equal
a priori probabilities, the corresponding ensemble, called the microcanonical ensemble, should have a distribution
function F (H (x)) that re ects the fact that all points on the constant energy hypersurface are equally probable. Such
a distribution function need only re ect the fact that energy is conserved and can be written as
F (H (x)) = (H (x) ? E )
where (x) is the Dirac delta function. The delta function has the property that
Z1
(x ? a)f (x)dx = f (a)
?1
for any function f (x).
Averaging over the microcanonical distribution function is equivalent to computing the time average in our thought
experiment. The microcanonical partition function
(N; V; E ) is given by
Z

(N; V; E ) = C N dx (H (x) ? E )
In Cartesian coordinates, this is equivalent to
Z Z

(N; V; E ) = C N dp n
d r(H (p; r) ? E )
N

D (V )

where C is a constant of proportionality. It is given by

N

C = N E!h03N
N

Here h is a constant with units Energy
Time, and E0 is a constant having units of energy. The extra factor of E0
is needed because the  function has units of inverse energy. Such a constant has no eect at all on any properties).
Thus,
(N; V; E ) is dimensionless. The origin of C is quantum mechanical in nature (h turns out to be Planck's
N

constant) and must be put into the classical expression by hand. Later, we will explore the eects of this constant on
thermodynamic properties of the ideal gas.
The microcanonical partition function function measures the number of microstates available to a system which
evolves on the constant energy hypersurface. Boltzmann identied this quantity as the entropy, S of the system,
which, for the microcanonical ensemble is a natural function of N , V and E :
S = S (N; V; E )
Thus, Boltzmann's relation between
(N; V; E ), the number of microstates and S (N; V; E ) is
S (N; V; E ) = k ln
(N; V; E )
where k is Boltzmann's constant 1=k = 315773.218 Kelvin/Hartree. The importance of Boltzmann's relation is that
it establishes a connection between the thermodynamic properties of a system and its microscopic details.
Recall the standard thermodynamic denition of entropy:
Z dQ
S= T
where an amount of heat dQ is assumed to be absorbed reversibly, i.e., along a thermodynamic path, by the system.
The rst law of thermodynamics states that the energy, E of the system is given by the sum of the heat absorbed by
the system and the work done on the system in a thermodynamic process:
E = Q+W
If the thermodynamic transformation of the system is carried reversibly, i.e., along a thermodynamic path, then the
rst law will be valid for the dierential change in energy, dE due to absorption of a dierential amount of heat, dQrev
and a dierential amount of work, dW done on the system:

6
 dE = dQ + dW
The work done on the system can be in the form of compression/expansion work at constant pressure, P , leading to
a change, dV in the volume and/or the insertion/deletion of particles from the system at constant chemical potential,
, leading to a change dN in the particle number. Thus, in general
dW = ?PdV + dN
(The above relation for the work is true only for a one-component
P system. If there are M types of particles present,
then the second term must be generalized according to =1  dN ). Then, using the fact that dQ = TdS , we have
M
k k k

dE = TdS ? PdV + dN

or
dS = dE P
T + T dV ? T dN


But since S = S (N; V; E ) is a natural function of N , V , and E , the dierential, dS is also given by
 @S   @S   @S 
dS = @E dE + @V dV + @N dN
N;V N;E E;V

Comparing these two expressions, we see that

 @S 
@E = T1
  N;V

@S = PT
@V
 @S  N;E

? @N = T
E;V

Finally, using Boltzmann's relation between the entropy S and the partition function
, we obtain a prescription for
obtaining the thermodynamic properties of the system starting from a microscopic, particle-based description of the
system:
 
1 = k @ ln

T @E
P = k @ ln


N;V

T @V
 = ?k @ ln


N;E

T @N V;E

Of course, the ultimate test of Boltzmann's relation between entropy and the partition function is that the above
relations correctly generate the known thermodynamic properties of a given system, e.g. the equation of state. We
will soon see several examples in which this is, indeed, the case.

7
 G25.2651: Statistical Mechanics

Notes for Lecture 3

I. MICROCANONICAL ENSEMBLE: CONDITIONS FOR THERMAL EQUILIBRIUM

Consider bringing two systems into thermal contact. By thermal contact, we mean that the systems can only
exchange heat. Thus, they do not exchange particles, and there is no potential coupling between the systems. In this
case, if system 1 has a phase space vector x1 and system 2 has a phase space vector x2 , then the total Hamiltonian
can be written as
H (x) = H1 (x1 ) + H2 (x2 )
Furthermore, let system 1 have N1 particles in a volume V1 and system 2 have N2 particles in a volume V2 . The total
particle number N and volume V are N = N1 + N2 and V = V1 + V2 . The entropy of each system is given by
S1 (N1 ; V1 ; E1 ) = k ln
1 (N1 ; V1 ; E1 )
S2 (N2 ; V2 ; E2 ) = k ln
2 (N2 ; V2 ; E2 )
The partition functions are given by
Z

1 (N1 ; V1 ; E1 ) = CN1 dx1 (H1 (x1 ) ? E1 )
Z

2 (N2 ; V2 ; E2 ) = CN2 dx2 (H2 (x2 ) ? E2 )
Z

(N; V; E ) = CN dx(H1 (x1 ) + H2 (x2 ) ? E ) 6=
1 (N1 ; V1 ; E1 )
2 (N2 ; V2 ; E2 )
However, it can be shown that the total partition function can be written as
Z E

(E ) = C 0
dE1
1 (E1 )
2 (E ? E1 )
0
where C is an overall constant independent of the energy. Note that the dependence of the partition functions on
0

the volume and particle number has been suppressed for clarity.
Now imagine expressing the integral over energies in the above expression as a Riemann sum:
P
X

(E ) = C
0

1 (E1(i) )
2 (E ? E1(i) )
i=1

where
is the small energy interval (which we will allow to go to 0) and P = E=
. The reason for writing the
integral this way is to make use of a powerful theorem on sums with large numbers of terms.
Consider a sum of the form
P
X
= ai
i=1

where ai > 0 for all ai . Let amax be the largest of all the ai 's. Clearly, then
P
X
amax  ai
i=1
XP
Pamax  ai
i=1

1
Thus, we have the inequality
amax    Pamax
or
ln amax  ln   ln amax + ln P
This gives upper and lower bounds on the value of ln . Now suppose that ln amax >> ln P . Then the above inequality
implies that
ln   ln amax
This would be the case, for example, if amax  eP . In this case, the value of the sum is given to a very good
approximation by the value of its maximum term.
WhyP
should this theorem apply to the sum expression for
(E )? Consider the case of a system of free particles
H = Ni=1 p2i =2mi, i.e., no potential. Then the expression for the partition function is
!

(E ) 
Z Z
d r d p
N N
N
X p2i ? E  V N
D (V ) i=1
2mi

since the particle integrations are restricted only the volume of the container. Thus, the terms in the sum vary
exponentially with N . But the number of terms in the sum P also varies like N since P = E=
and E  N , since E
is extensive. Thus, the terms in the sum under consideration obey the conditions for the application of the theorem.
Let the maximum term in the sum be characterized by energies E1 and E2 = E ? E1 . Then, according to the above
analysis,
 
S (E ) = k ln
(E ) = k ln
+ k ln
1 (E1 )
2 (E ? E1 ) + k ln P + k ln C
0

Since P = E=
, ln
+ ln P = ln
+ ln E ? ln
= ln E . But E  N , while ln

1  N . Since N >> ln N , the above
expression becomes, to a good approximation
 
S (E )  k ln
1 (E1 )
2 (E ? E1 ) + O(ln N ) + const
Thus, apart from constants, the entropy is approximately additive:
S (E ) = k ln
1 (E1 ) + k ln
2 (E2 )
= S1 (E1 ) + S2 (E2 ) + O(ln N ) + const
Finally, in order to compute the temperature of each system, we make a small variation in the energy E1 ; dE1 . But
since E1 + E2 = E , dE1 = ?dE2 . Also, this variation is made such that the total entropy S and energy E remain
constant. Thus, we obtain
0 = @S1 + @S2
@ E1 @ E1
0 = @S1 ? @S2
@ E1 @ E2
0= 1 ? 1
T1 T2
from which it is clear that T1 = T2 , the expected condition for thermal equilibrium.
It is important to point out that the entropy S (N; V; E ) dened via the microcanonical partition function is not the
only entropy that satises the properties of additivity and equality of temperatures at thermal equilibrium. Consider
an ensemble dened by the condition that the Hamiltonian, H (x) is less than a certain energy E . This is known as
the uniform ensemble and its partition function, denoted (N; V; E ) is dened by
Z Z
(N; V; E ) = C dx = C dx (E ? H (x))
H (x)<E

2
where (x) is the Heaviside step function. Clearly, it is related to the microcanonical partition function by
@ (N; V; E )

(N; V; E ) = @E
Although we will not prove it, the entropy S~(N; V; E ) dened from the uniform ensemble partition function via
S~(N; V; E ) = k ln (N; V; E )
is also approximately additive and will yield the condition T1 = T2 for two systems in thermal contact. In fact,
it diers from S (N; V; E ) by a constant of order ln N so that one can also dene the thermodynamics in terms of
S~(N; V; E ). In particular, the temperature is given by
!
1 = @ S~
 

T @E = k @@E
ln 
N;V N;V

II. REVERSIBLE LAWS OF MOTION AND THE ARROW OF TIME

Hamilton's equations of motion
q_i = @H
@pi ; p_i = ? @H
@q
i

are invariant under a reversal of the direction of time t ! ?t. Under such a transformation, the positions and
momenta transform according to
qi ! q i
pi = mi dq dqi
dt ! mi d(?t) = ?pi
i

Thus,
dqi @H d(?pi ) = ? @H
d(?t) = ? @pi ; d(?t) @qi
or
q_i = @H
@pi ; p_i = ? @H
@q
i

The form of the equations does not change! One of the implications of time reversal symmetry is as follows: Suppose
a system is evolved forward in time starting from some initial condition up to a maximum time t; at t, the evolution
is stopped, the sign of the velocity of each particle in the system is reversed, i.e., a time reversal transformation is
performed, and the system is allowed to evolve once again for another time interval of length t; the system will return
to its original starting point in phase space, i.e., the system will return to its initial condition. Now from the point
of view of mechanics and the microcanonical ensemble, the initial conditions (for the rst segment of the evolution)
and the conditions created by reversing the signs of the velocities for initiating the second segment are equally valid
and equally probably, both being points selected from the constant energy hypersurface. Therefore, from the point
of view of mechanics, without a priori knowledge of which segment is the forward evolving trajectory and which is
the time reversed trajectory, it should not be possible to say which is which. That is, if a movie of each trajectory
were to be made and shown to an ignorant observer, that observer should not be able to tell which is the forward-
evolving trajectory and which the time-reversed trajectory. Therefore, from the point of view of mechanics, which
obeys time-reversal symmetry, there is not preferred direction for the ow of time.
Yet our everyday experience tells us that there are plenty of situations in which a system seems to evolve in a
certain direction and not in the reverse direction, suggesting that there actually is a preferred direction in time. Some
common examples are a glass falling to the ground and smashing into tiny pieces or the sudden expansion of a gas into
a large box. These processes would always seem to occur in the same way and never in the reverse (the glass shards
never reassemble themselves and jump back onto the table forming an unbroken glass, and the gas particles never

3
suddenly all gather in one corner of the large box). This seeming inconsistency with the reversible laws of mechanics
is known as Loschmidt's paradox. Indeed, the second law of thermodynamics, itself, would seem to be at odds with
the reversibility of the laws of mechanics. That is, the observation that a system naturally evolves in such a way as
to increase its entropy cannot obviously be rationalized starting from the microscopic reversible laws of motion.
Note that a system being driven by an external agent or eld will not be in equilibrium with its surroundings and
can exhibit irreversible behavior as a result of the work being done on it. The falling glass is an example of such a
system. It is acted upon by gravity, which drives it uniformly toward a state of ever lower potential energy until the
glass smashes to the ground. Even though it is possible to write down a Hamiltonian for this system, the treatment
of the external eld is only approximate in that the true microscopic origin of the external eld is not taken into
account. This also brings up the important question of how one exactly denes non-equilibrium states in general,
and how do they evolve, a question which, to date, has no fully agreed upon answer and is an active area of research.
However, the expanding gas example does not involve an external driving force and still seems to exhibit irreversible
behavior. How can this observation be explained for such an isolated system?
One possible explanation was put forth by Boltzmann, who introduced the notion of molecular chaos. Under this
assumption, the momenta of two particles do not become correlated as the result of a collision. This is tantamount
to the assumption that microscopic information leading to a correlation between the particles is lost. This is not
inconsistent with microscopic reversibility from a probabilistic point of view, as the momenta of two particles before
a collision are certainly uncorrelated with each other. The assumption of molecular chaos allows one to prove the so
called Boltzmann H-theorem, a theorem that predicts an increase in entropy until equilibrium is reached. For more
details see Chapters 3 and 4 of Huang.
Boltzmann's assumption of molecular chaos remains unproven and may or may not be true. Another explanation
due to Poincare is based on his recurrence theorem. The Poincare recurrence theorem states that a system having
a nite amount of energy and conned to a nite spatial volume will, after a suciently long time, return to an
arbitrarily small neighborhood of its initial state.
Proof of the recurrence theorem (taken almost directly from Huang, pg. 91):
Let a state of a system be represented by a point x in phase space. As the system evolves in time, a point in
phase space traces out a trajectory that is uniquely determined by any given point on the trajectory (because of the
deterministic nature of classical mechanics). Let g0 be an arbitrary volume element in the phase space in the volume
!0 . After a time t all points in g0 will be in another volume element gt in a volume !t , which is uniquely determined
by the choice of g0. Assuming the system is Hamiltonian, then by Liouville's theorem:
!0 = !t
Let ?0 denote the subspace of phase space that is the union of all gt for 0  t < 1. Let its volume be
0 . Similarly,
let ? denote the subspace that is the union of all gt for   t < 1. Let its volume be
 . The numbers
0 and

are nite because, since the energy of the system is nite and the spatial volume occupied is nite, a representative
point is conned to a nite region in phase space. The denitions immediately imply that ?0 contains ? .
We may think of ?0 and ? in a dierent way. Imagine the region ?0 to be lled uniformly with representative
points. As time progresses, ?0 will evolve into some other regions that are uniquely determined. It is clear, from the
denitions, that after a time  , ?0 will become ? . Also, by Liouville's theorem:

0 =

Recall that ?0 contains all the future destinations of the points in g , which in turn is evolved from g0 after a time  .
It has been shown that ?0 has the same volume as ? since
0 =
 by Liouville's theorem. Therefore, ?0 and ?
must contain the same set of points (except possibly for a set of zero measure).
In particular, ? contains all of g0 (except possibly for a set of zero measure). But, by denition, all points in ?
are future destinations of the points in g0 . Therefore all points in g0 (except possibly for a set of zero measure) must
return to g0 after a suciently long time. Since g0 can be made arbitrarily small, Poincare's theorem follows.
Now consider an initial condition in which all the gas particles are initially in a corner of a large box. By Poincare's
theorem, the gas particles must eventually return to their initial state in the corner of the box. How long is this
recurrence time? In order to answer this, consider dividing the box up into M small cells of volume v. The total
number of microstates available to the gas varies with N like V N . The number of microstates corresponding to all the
particles occupying a single cell of volume v is vN . Thus, the probability of observing the system in this microstate is
approximately (v=V )N . Even if v = V=2, for N  1023 , the probability is vanishingly small. The Poincare recurrence
time, on the other hand, is proportional to the inverse of this probability or (V=v)N . Again, since N  1023, if
v = V=2, the required time is

4
 t  21023
which is many orders of magnitude longer than the current age of the universe. Thus, although the system will return
arbitrarily close to its initial state, the time required for this is unphysically long and will never be observed. Over
times relevant for observation, given similar such initial conditions, the system will essentially always evolve in the
same way, which is to expand and ll the box and essentially never to the opposite.

5
 G25.2651: Statistical Mechanics

Notes for Lecture 4

I. THE CLASSICAL VIRIAL THEOREM (MICROCANONICAL DERIVATION)

Consider a system with Hamiltonian H (x). Let xi and xj be specic components of the phase space vector. The
classical virial theorem states that
@H i = kT
hxi @x ij
j

where the average is taken with respect to a microcanonical ensemble.

To prove the theorem, start with the denition of the average:
@H i = C
hxi @x
Z
@H
j
(E ) dxxi @xj (E ? H (x))
where the fact that (x) = (?x) has been used. Also, the N and V dependence of the partition function have been
suppressed. Note that the above average can be written as
@H C @ Z

hxi @x i =
(E ) @E dx xi @x @H (E ? H (x))
j j
@
=
(CE ) @E @H
Z

dx xi @x
H (x)<E j

=
(CE ) @E
@ dx xi @ (H@x? E )
Z

H (x)<E j

However, writing
xi @ (H@x? E ) = @x@ [xi (H ? E )] ? ij (H ? E )
j j

allows the average to be expressed as

@H C @ dx @x@ [xi (H ? E )] + ij (E ? H (x))
Z  

hxi @x i =
(E ) @E
j H (x)<E j
C @
I Z 

=
(E ) @E xi (H ? E )dSj + ij dx (E ? H (x))
H =E H<E

The rst integral in the brackets is obtained by integrating the total derivative with respect to xj over the phase space
variable xj . This leaves an integral that must be performed over all other variables at the boundary of phase space
where H = E , as indicated by the surface element dSj . But the integrand involves the factor H ? E , so this integral
will vanish. This leaves:
@H i = Cij @
hxi @x
Z

dx(E ? H (x))
j
(E ) @E H (x)<E
Cij
=
(
Z

dx
E) H (x)<E

=
(ijE ) (E )

1
where (E ) is the partition function of the uniform ensemble. Recalling that
@ (E )

(E ) = @E
we have
@H i =  (E )
hxi @x ij @ (E )
j @E

= ij @ ln 1(E)
@E

= kij @1S~
@E

= kTij
which proves the theorem.
Example: xi = pi and i = j . The virial theorem says that
hpi @H
@p i = kT
i
2
h mpi i = kT
i
p 2
h 2mi i = 12 kT
i

Thus, at equilibrium, the kinetic energy of each particle must be kT=2. By summing both sides over all the particles,
we obtain a well know result
3
N 2 3N
X p
h 2m i = h 12 mi vi2 i = 23 NkT
i
X

i=1 i i=1

II. LEGENDRE TRANSFORMS

The microcanonical ensemble involved the thermodynamic variables N , V and E as its variables. However, it
is often convenient and desirable to work with other thermodynamic variables as the control variables. Legendre
transforms provide a means by which one can determine how the energy functions for dierent sets of thermodynamic
variables are related. The general theory is given below for functions of a single variable.
Consider a function f (x) and its derivative
df  g(x)
y = f 0 (x) = dx
The equation y = g(x) denes a variable transformation from x to y. Is there a unique description of the function
f (x) in terms of the variable y? That is, does there exist a function (y) that is equivalent to f (x)?
Given a point x0 , can one determine the value of the function f (x0 ) given only f 0 (x0 )? No, for the reason that the
function f (x0 ) + c for any constant c will have the same value of f 0 (x0 ) as shown in the gure below.

2
 f(x)

x0 x

b(x 0 )

FIG. 1.

However, the value f (x0 ) can be determined uniquely if we specify the slope of the line tangent to f at x0 , i.e.,
f 0 (x0 ) and the y-intercept, b(x0 ) of this line. Then, using the equation for the line, we have
f (x0 ) = x0 f 0 (x0 ) + b(x0 )
This relation must hold for any general x:
f (x) = xf 0 (x) + b(x)
Note that f 0 (x) is the variable y, and x = g?1(y), where g?1 is the functional inverse of g, i.e., g(g?1 (x)) = x. Solving
for b(x) = b(g?1 (y)) gives
b(g?1(y)) = f (g?1 (y)) ? yg?1(y)  (y)
where (y) is known as the Legendre transform of f (x). In shorthand notation, one writes
(y) = f (x) ? xy
however, it must be kept in mind that x is a function of y.

III. THE CANONICAL ENSEMBLE

A. Basic Thermodynamics
In the microcanonical ensemble, the entropy S is a natural function of N ,V and E , i.e., S = S (N; V; E ). This can be
inverted to give the energy as a function of N ,V , and S , i.e., E = E (N; V; S ). Consider using Legendre transformation
to change from S to T using the fact that

3
 T = @E
 

@S N;V

The Legendre transform E~ of E (N; V; S ) is

E~ (N; V; T ) = E (N; V; S (T )) ? S @E
@S
= E (N; V; S (T )) ? TS
The quantity E~ (N; V; T ) is called the Hemlholtz free energy and is given the symbol A(N; V; T ). It is the fundamental
energy in the canonical ensemble.
The dierential of A is
dA = @A @A @A
     

@T dT + @V dV + @N dN
N;V N;T T;V

However, from A = E ? TS , we have

dA = dE ? TdS ? SdT
From the rst law, dE is given by
dE = TdS ? PdV + dN
Thus,
dA = ?PdV ? SdT + dN
Comparing the two expressions, we see that the thermodynamic relations are
S = ? @A
 

@T N;V
@A
 

P = ? @V
N;T
@A
 

 = @N
V;T

B. The partition function

Consider two systems (1 and 2) in thermal contact such that
N2  N1
E2  E1
N = N1 + N2 ; E = E 1 + E2
dim(x1 )  dim(x2 )
and the total Hamiltonian is just H (x) = H1 (x1 ) + H2 (x2 )
Since system 2 is innitely large compared to system 1, it acts as an innite heat reservoir that keeps system 1 at a
constant temperature T without gaining or losing an appreciable amount of heat, itself. Thus, system 1 is maintained
at canonical conditions, N; V; T .
The full partition function
(N; V; E ) for the combined system is the microcanonical partition function
Z Z

(N; V; E ) = dx(H (x) ? E ) = dx1 dx2 (H1 (x1 ) + H2 (x2 ) ? E )
Now, we dene the distribution function, f (x1 ) of the phase space variables of system 1 as

4
 Z

f (x1 ) = dx2 (H1 (x1 ) + H2 (x2 ) ? E )

Taking the natural log of both sides, we have
Z

ln f (x1 ) = ln dx2 (H1 (x1 ) + H2 (x2 ) ? E )

Since E2  E1 , it follows that H2 (x2 )  H1 (x1 ), and we may expand the above expression about H1 = 0. To linear
order, the expression becomes

Z
@ Z

ln f (x1 ) = ln dx2 (H2 (x2 ) ? E ) + H1 (x1 ) @H (x ) ln dx2 (H1 (x1 ) + H2 (x2 ) ? E )




1 1 H1 (x1 )=0
Z
@ ln dx (H (x ) ? E )
= ln dx2 (H2 (x2 ) ? E ) ? H1 (x1 ) @E
Z

2 2 2

where, in the last line, the dierentiation with respect to H1 is replaced by dierentiation with respect to E . Note
that
ln dx (H ( ) ? E) = S2 (E )
Z

2 22 k
@ ln dx (H (x ) ? E ) = @ S2 (E ) = 1
Z

@E 2 2 2 @E k kT
where T is the common temperature of the two systems. Using these two facts, we obtain
ln f (x ) = S2 (E ) ? H1 (x1 )
1 k kT
f (x1 ) = eS2(E)=k e?H1 (x1 )=kT
Thus, the distribution function of the canonical ensemble is
f (x) / e?H (x)=kT
The prefactor exp(S2 (E )=k) is an irrelevant constant that can be disregarded as it will not aect any physical
properties.
The normalization of the distribution function is the integral:
Z

dxe?H (x)=kT  Q(N; V; T )

where Q(N; V; T ) is the canonical partition function. It is convenient to dene an inverse temperature  = 1=kT .
Q(N; V; T ) is the canonical partition function. As in the microcanonical case, we add in the ad hoc quantum corrections
to the classical result to give
Z

Q(N; V; T ) = N !h1 3N dxe?H (x)

The thermodynamic relations are thus,

Hemlholtz free energy:
A(N; V; T ) = ? 1 ln Q(N; V; T )

To see that this must be the denition of A(N; V; T ), recall the denition of A:
A = E ? TS = hH (x)i ? TS

5
But we saw that
S = ? @A
 

@T N;V

Substituting this in gives

A = hH (x)i ? T @A
@T
or, noting that
@A = @A @ = ? 1 @A
@T @ @T kT 2 @
it follows that
A = hH (x)i +  @A
@
This is a simple dierential equation that can be solved for A. We will show that the solution is
A = ? 1 ln Q( )

Note that
 @A 1 1 @Q
@ =  ln Q( ) ? Q @ = A ? hH (x)i
Substituting in gives, therefore
A = hH (x)i + A ? hH (x)i = A
so this form of A satises the dierential equation.
Other thermodynamics follow:
Average energy:
Z

E = hH (x)i = Q1 CN dxH (x)e?H (x)

@ ln Q(N; V; T )
= ? @

Pressure:
@A = kT @ ln Q@V
(N; V; T )
   

P = ? @V
N;T N;T

Entropy:
S = ? @A = ? @A @ = 1 @A
@T  @ @T kT 2 @
@
= k @ ?  ln Q(N; V; T ) = ?k @ @
ln Q + k ln Q

2 1

= kE + k ln Q = k ln Q + E
T

6
 Heat capacity at constant volume:
CV = @E = @E @ 2 @
 

@T N;V @ @T = k @ 2 ln Q(N; V; T )

C. Relation between canonical and microcanonical ensembles

We saw that the E (N; V; S ) and A(N; V; T ) could be related by a Legendre transformation. The partition functions

(N; V; E ) and Q(N; V; T ) can be related by a Laplace transform. Recall that the Laplace transform f~() of a function
f (x) is given by
Z 1
f~() = dxe?x f (x)
0
Let us compute the Laplace transform of
(N; V; E ) with respect to E :
Z 1 Z

~ (N; V; ) = CN dEe?E dx(H (x) ? E )
0
Using the -function to do the integral over E :
Z

~ (N; V; ) = CN dxe?H (x)
By identifying  =  , we see that the Laplace transform of the microcanonical partition function gives the canonical
partition function Q(N; V; T ).

D. Classical Virial Theorem (canonical ensemble derivation)

Again, let xi and xj be specic components of the phase space vector x = (p1 ; :::; p3N ; q1 ; :::; q3N ). Consider the
canonical average
@H i
hxi @x
j
given by
@H i = 1 C dxx @H e?H (x)
hxi @x
Z

j Q N i
@x j
@
Z  
1 1
= Q CN dxxi ?  @x e?H (x)
j
But
@xi
xi @x@ e?H (x) = @x@ xi e?H (x) ? e?H (x) @x
 

j j j
@  
= @x xi e?H (x) ? ij e?H (x)
j
Thus,
@H i = ? 1 C dx @ x e?H (x) + 1  C dxe?H (x)
hxi @x
Z   Z

j Q N Z @x j
i
Q ij N
1 C
= ? Q dx 0
Z

dx @ x e?H (x)  + kT

N j
@x i j
ij
Z
1

= dx0 xi e?H (x) 

xj =?1
+ kTij

Several cases exist for the surface term xi exp(?H (x)):

7
 1. xi = pi a momentum variable. Then, since H  p2i , exp(?H ) evaluated at pi = 1 clearly vanishes.
2. xi = qi and U ! 1 as qi ! 1, thus representing a bound system. Then, exp(?H ) also vanishes at qi = 1.

3. xi = qi and U ! 0 as qi ! 1, representing an unbound system. Then the exponential tends to 1 both at
qi = 1, hence the surface term vanishes.
4. xi = qi and the system is periodic, as in a solid. Then, the system will be represented by some supercell to which
periodic boundary conditions can be applied, and the coordinates will take on the same value at the boundaries.
Thus, H and exp(?H ) will take on the same value at the boundaries and the surface term will vanish.
5. xi = qi , and the particles experience elastic collisions with the walls of the container. Then there is an innite
potential at the walls so that U ?! 1 at the boundary and exp(?H ) ?! 0 at the bondary.
Thus, we have the result
@H i = kT
hxi @x ij
j

The above cases cover many but not all situations, in particular, the case of a system conned within a volume V
with re ecting boundaries. Then, surface contributions actually give rise to an observable pressure (to be discussed
in more detail in the next lecture).

8
 G25.2651: Statistical Mechanics

Notes for Lecture 5

I. TEMPERATURE AND PRESSURE ESTIMATORS

From the classical virial theorem
@H i = kT
hxi @x ij
j

we arrived at the equipartition theorem:

* +
N
X p2i = 3 NkT
i=1
2mi 2
where p1 ; :::; pN are the N Cartesian momenta of the N particles in a system. This says that the microscopic function
of the N momenta that corresponds to the temperature, a macroscopic observable of the system, is given by
N
K (p1 ; :::; pN ) =
X p2i
i=1
2mi
The ensemble average of K can be related directly to the temperature
2 hK (p ; :::; p )i = 2 hK (p ; :::; p )i
T = 3Nk N
1
3nR 1 3

K (p1 ; :::; pN ) is known as an estimator (a term taken over from the Monte Carlo literature) for the temperature.
An estimator is some function of the phase space coordinates, i.e., a function of microscopic states, whose ensemble
average gives rise to a physical observable.
An estimator for the pressure can be derived as well, starting from the basic thermodynamic relation:
   
P = ? @A@V = kT @ ln Q(N; V; T ) @V
N;T N;T

with
Z Z Z
Q(N; V; T ) = CN dx e?H (x) = CN dNp dN r e?H (p;r)
V
The volume dependence of the partition function is contained in the limits of integration, since the range of integration
for the coordinates is determined by the size of the physical container. For example, if the system is conned within
a cubic box of volume V = L3 , with L the length of a side, then the range of each q integration will be from 0 to L. If
a change of variables is made to si = qi =L, then the range of each s integration will be from 0 to 1. The coordinates
si are known as scaled coordinates. For containers of a more general shape, a more general transformation is
si = V ?1=3 ri
In order to preserve the phase space volume element, however, we need to ensure that the transformation is a canonical
one. Thus, the corresponding momentum transformation is
i = V 1=3 pi
With this coordinate/momentum transformation, the phase space volume element transforms as
dN p dN r = dN dN s

1
Thus, the volume element remains the same as required. With this transformation, the Hamiltonian becomes

H=
N
X p2i + U (r ; :::; r ) = XN
V ?2=3 i2 + U (V 1=3 s ; :::; V 1=3 s )
N N
i=1
2mi 1
i=1
2mi 1

and the canonical partition function becomes

( " #)

Q(N; V; T ) = CN
Z
dN 
Z
dN s exp ? V ?2=3 i2 + U (V 1=3 s ; :::; V 1=3 s )
N
X
N
i=1
2mi 1

Thus, the pressure can now be calculated by explicit dierentiation with respect to the volume, V :
P = kT 1 @Q
Q @V
Z Z " #
N N
= Q CN d  dN s 23 V ?5=3 2mi ? 3 V ?2=3 si
@U
kT X 2 X
N
@ (V 1=3 s )
e?H
i=1 i i=1 i
Z Z " #
N N
= kT  pi  @U ?H (p;r)
X 2 X
Q CN d p d r 3V i=1 mi ? 3V i=1 ri
@ ri e
N N

* +
N  2
1 X p
= 3V mi + r i
F i
i
i=1

= h? @H
@V i
Thus, the pressure estimator is
N  
(p1 ; :::; pN ; r1 ; :::; rN ) = (x) = 31V
X p2i + r
F (r)
i=1
2mi i i
and the pressure is given by
P = h(x)i
For periodic systems, such as solids and currently usedPmodels of liquids, an absolute Cartesian coordinate qi is
ill-dened. Thus, the virial part of the pressure estimator i qi Fi must be rewritten in a form appropriate for periodic
systems. This can be done by recognizing that the force Fi is obtained as a sum of contributions Fij , which is the
force on particle i due to particle j . Then, the classical virial becomes
N
X N
X X
ri
F i = ri
Fij
i=1 i=1 j 6=i
2 3

= 21 4
X X X X
ri
Fij + rj
Fji 5
i j 6=i j i6=j
2 3

= 21 4
X X X X
ri
Fij ? rj
Fij 5
i j 6=i j i6=j

= 21 (ri ? rj )
Fij = 12
X X
rij
Fij
i;j;i6=j i;j;i6=j

where rij is now a relative coordinate. rij must be computed consistent with periodic boundary conditions, i.e., the
relative coordinate is dened with respect to the closest periodic image of particle j with respect to particle i. This
gives rise to surface contributions, which lead to a nonzero pressure, as expected.

2
 II. ENERGY FLUCTUATIONS IN THE CANONICAL ENSEMBLE
In the canonical ensemble, the total energy is not conserved. (H (x) 6= const). What are the uctuations in the
energy? The energy uctuations are given by the root mean square deviation of the Hamiltonian from its average
hH i:
q p

E = h(H ? hH i)2 i = hH 2 i ? hH i2

@ ln Q(N; V; T )
hH i = ? @
Z
1
hH i = Q CN dxH 2 (x)e?H (x)
2

Z
1 @ 2 e?H (x)
= Q CN dx @ 2

=1 @ Q
2

Q @ 2
 2
@ 2
1
= @ 2 ln Q + Q2 @ @Q
 
@ 2 ln Q + 1 @Q 2
= @ 2 Q @
 2
@ 2
@
= @ 2 ln Q + @ ln Q
Therefore
@ ln Q
hH i ? hH i = @
2 2
2

But
@ 2 ln Q = kT 2C
V
@ 2
Thus,
p

E = kT 2CV
Therefore, the relative energy uctuation
E=E is given by
p

E = kT 2CV
E E
Now consider what happens when the system is taken to be very large. In fact, we will dene a formal limit called
the thermodynamic limit, in which N ?! 1 and V ?! 1 such that N=V remains constant.
Since CV and E are both extensive variables, CV  N and E  N ,

E  p1 ?! 0 as N ! 1
E N
But
E=E would be exactly 0 in the microcanonical ensemble. Thus, in the thermodynamic limit, the canonical and
microcanonical ensembles are equivalent, since the energy uctuations become vanishingly small.

3
 III. ISOTHERMAL-ISOBARIC ENSEMBLE
A. Basic Thermodynamics
The Helmholtz free energy A(N; V; T ) is a natural function of N , V and T . The isothermal-isobaric ensemble
is generated by transforming the volume V in favor of the pressure P so that the natural variables are N , P and
T (which are conditions under which many experiments are performed { \standard temperature and pressure," for
example).
Performing a Legendre transformation of the Helmholtz free energy
@A
A~(N; P; T ) = A(N; V (P ); T ) ? V (P ) @V
But
@A = ?P
@V
Thus,
A~(N; P; T ) = A(N; V (P ); T ) + PV  G(N; P; T )
where G(N; P; T ) is the Gibbs free energy.
The dierential of G is
     
@G
dG = @P dP + @G @G
dT + @N dN
N;T @T N;P P;T

But from G = A + PV , we have

dG = dA + PdV + V dP
but dA = ?SdT ? PdV + dN , thus
dG = ?SdT + V dP + dN
Equating the two expressions for dG, we see that
 
V = @P@G
N;T
 
S = ? @G@T N;P
 
 = @N@G
P;T

B. The partition function and relation to thermodynamics

In principle, we should derive the isothermal-isobaric partition function by coupling our system to an innite thermal
reservoir as was done for the canonical ensemble and also subject the system to the action of a movable piston under
the in uence of an external pressure P . In this case, both the temperature of the system and its pressure will be
controlled, and the energy and volume will uctuate accordingly.
However, we saw that the transformation from E to T between the microcanonical and canonical ensembles turned
into a Laplace transform relation between the partition functions. The same result holds for the transformation from
V to T . The relevant \energy" quantity to transform is the work done by the system against the external pressure P
in changing its volume from V = 0 to V , which will be PV . Thus, the isothermal-isobaric partition function can be
expressed in terms of the canonical partition function by the Laplace transform:

4
 Z 1

(N; P; T ) = V1 dV e?P V Q(N; V; T )
0 0

where V0 is a constant that has units of volume. Thus,

Z 1 Z

(N; P; T ) = V N1!h3N dV dxe?(H (x)+P V )
0 0

The Gibbs free energy is related to the partition function by

G(N; P; T ) = ? 1 ln
(N; P; T )

This can be shown in a manner similar to that used to prove the A = ?(1= ) ln Q. The dierential equation to start
with is
@G
G = A + PV = A + P @P
Other thermodynamic relations follow:
Volume:
 
V = ?kT @ ln
(@P
N; P; T )
N;T

Enthalpy:
@ ln
(N; P; T )
H = hH (x) + PV i = ? @

Heat capacity at constant pressure


 
CP = @@TH @ ln
(N; P; T )
2
= k 2 @
N;P

Entropy:
 
S = ? @G
@T N;P

= k ln
(N; P; T ) + H
T

The uctuations in the enthalpy
H are given, in analogy with the canonical ensemble, by
p

H = kT 2CP
so that
p

H = kT 2CP
H H
 H  1=pN which vanish in the thermodynamic limit.
so that, since CP and H are both extensive,
H=

5
 C. Pressure and work virial theorems
As noted earlier, the quantity ?@H=@V is a measure of the instantaneous value of the internal pressure Pint . Let
us look at the ensemble average of this quantity
Z 1 Z
1
hPint i = ?
CN dV e ? P V dx @H
@V e?H (x)
0
Z 1 Z
=
1 CN dV e?P V dxkT @V @ e?H (x)
0
Z 1
=
1 dV e?P V kT @V@ Q(N; V; T )
0

Doing the volume integration by parts gives

Z 1  
1
hPint i =
e
 ?P V 1
kTQ(N; V; T ) 0 ?


1 @
dV kT @V e ? P V Q(N; V; T )
0
Z 1
= P
1 dV e?P V Q(N; V; T ) = P
0

Thus,
hP i = P
int

This result is known as the pressure virial theorem. It illustrates that the average of the quantity ?@H=@V gives the
xed pressure P that denes the ensemble.
Another important result comes from considering the ensemble average hPint V i
Z 1 @ Q(N; V; T )
hP V i =
1
int dV e?P V kTV @V
0

Once again, integrating by parts with respect to the volume yields

1 Z 1 dV kT @ V e?P V Q(N; V; T )
 
hPint V i =
1 e?P V kTV Q(N; V; T ) 1
 
0
?
0 @V 
 Z 1 Z 1
1
=
?kT dV e?P V Q(V ) + P dV e?P V V Q(V )
0 0
= ?kT + P hV i
or
hP V i + kT = P hV i
int

This result is known as the work virial theorem. It expresses the fact that equipartitioning of energy also applies to
the volume degrees of freedom, since the volume is now a uctuating quantity.

6
 G25.2651: Statisti al Me hani s

Notes for Le ture 6

I. THE IDEAL GAS

The ideal gas is simplest thermodynami system that lends itself to analyti al solution in all the ensembles we have
studied thus far. It's importan e lies in the fa t that, for many \real" systems, the ideal gas behavior is the zeroth
order approximation to these systems. Thus, deviations from ideal behavior be ome manifestly lear. It is therefore
important that we establish what this ideal behaviour is. We will begin with the mi ro anoni al ensemble treatment.

A. Mi ro anoni al ensemble treatment

Consider a system of N parti les in a ubi box of volume V = L3 .

The parti les are assumed not to intera t with ea h other. Thus, the Hamiltonian in Cartesian oordinates may be
taken to be
N
X 2
pi
H=
i=1
2m
where we are assuming that all parti les are of the same type.
The mi ro anoni al partition fun tion is
Z !
N
X
E0 2
pi

(N; V; E ) = dN p dN r Æ E
N !h3N i=1
2m
Sin e the Hamiltonian is independent of the oordinates, the 3N oordinate integrations an be done straightforwardly.
The range of ea h one is 0 to L. Thus, these integrations give an overall fa tor L3N = V N :
Z !
N
X
E0 2
pi

(N; V; E ) = VN dN p Æ E
N !h3N i=1
2m
To do the momentum integrals, we rst hange variables to
0=
pi ppi
2m
dN p = (2m)3N=2 dN p0
Substitution into the partition fun tion gives
Z N !
E0 X

(N; V; E ) = V N (2m)3N=2 dN p0 Æ pi
02 E
N !h3N i=1

The 3N dimensional integral an now be seen to an integration over the surfa e of a sphere dened by the equation
N
X 02
pi =E
i=1

Therefore, it proves useful to transform to 3N dimensional spheri al oordinates, R; 1 ; :::; 3N 1 , where

1
 N
X 02
R2 = pi
i=1

Then
dN p0 = d3N 1
! dR R3N 1

where d3N 1 ! is the 3N 1 solid angle integral over the 3N 1 angles. The partition fun tion now be omes:
Z
E0


(N; V; E ) = V N (2m)3N=2 d3N 1
! dR R3N 1
Æ R2 E
N !h3N
At this point, we use an identity of Æ-fun tions:

1
Æ x2 a2 = [Æ(x a) + Æ(x + a)℄
2jaj
to write
E0
Z
1 h  p   p i

(N; V; E ) =
N !h3N
V N (2m)3N=2 d 3N 1
! dR R 3N 1
p Æ R E +Æ R+ E
Z
2 E
E0 1
= V N (2m)3N=2 E 3N=2 1 d3N 1 !
N !h 3N 2
We will also make use of the fa t that N is large, so that we may take 3N  1  3N . Using the general formula for
an n dimensional solid angle integral:
Z
2(n+1)=
2
dn ! = n+1
2

where (x) is the Gamma fun tion:

Z 1
(x) = dttx 1 e t
0

whi h satises
(n) = (n 1)!
 
1 (2n 1)!! 1=2
+n = 
2 2n
Thus, the solid angle integral is
Z
23N=2

d3N 1
! = 3N
2

and the partition fun tion nally be omes

 N
1 E0 V 1

(N; V; E ) = (2mE )3=2

N ! E h3 3N
2

The entropy S (N; V; E ) is given by

 
V E
S (N; V; E ) = k ln
(N; V; E ) = Nk ln (2mE )3=2 k ln N ! k ln (3N=2) k ln
h3 E0
Note that the term k ln(E=E0 )  ln N sin e E  N , whi h is negligibly small ompared to the term proportional to
N and ln N !. Thus, we an negle t it. Now, we an simplify ln N ! using Stirling's approximation
ln N !  N ln N N

2
whi h is valid for N very large. Also, note that
3N
(3N=2) = ( 1)!  (3N=2)!
2
so that
ln (3N=2)  (3N=2) ln(3N=2) (3N=2)
Substituting these approximations into the expression for the entropy, we obtain
"  3=2 #
V 4mE 3
S (N; V; E ) = k ln
(N; V; E ) = Nk ln 3 + Nk k ln N !
h 3N 2
We ould also simplify the ln N ! using Stirling's approximation, however, let us keep it as it is for now, sin e, as we
remember from our past treatment of the mi ro anoni al ensemble, this fa tor was in luded in the partition fun tion
in an ad ho manner, in order to a ount for the indistinguishability of the parti les. We will want to explore the
ee t of removing this term. Without it, the entropy is the purely lassi al entropy
"  3=2 #
V 4mE 3
S l (N; V; E ) = k ln
(N; V; E ) = Nk ln 3 + Nk
h 3N 2
Other thermodynami quantities an be easily obtained. For example, the temperature is
 
1  ln
3N
= =
kT E 2E
or
3
E = NkT
2
whi h is the result we obtained from our analysis of the lassi al virial theorem. The pressure is given by
 
 ln
N
P = kT =
V V
or
P V = NkT
whi h is the famous ideal gas law. This is a tually the equation of state of the ideal gas, as it expresses the pressure
as a fun tion of the volume and temperature. It an also be written as
P N
= =
kT V
where  is the onstant density of the gas.
One often expresses the equation of state graphi ally. For the ideal gas, if we plot P vs. V , for dierent values of
T , we obtain the following plot:

3
 T4>T3>T2>T1
10.0

8.0 T1
T2
T3
6.0
P T4

4.0

2.0

0.0
0.0 2.0 4.0 6.0 8.0
V

FIG. 1.

The dierent urves shown are alled the isotherms, sin e they represent P vs. V for xed temperature.
The heat apa ity of the ideal gas follows from the expression for the energy
 
E 3
CV = = Nk
T N;V 2
From our previous analysis of the virial theorem, we an on lude that ea h kineti mode ontributes k=2 to the heat
apa ity.

B. Relation to thermodynami entropy

In thermodynami s, the hange in entropy in a reversible pro ess whi h transforms the system from state 1 to state
2 is
Z 2
dQrev

S =
1 T
where dQrev is the heat absorbed in the pro ess. We an now ask if the entropy obtained starting from the mi ros opi
des ription agrees with the standard thermodynami denition. We will onsider two types of pro esses as des ribed
below:
I. Isothermal expansion/ ompression of the system from volume, V1 to V2 . In an isothermal pro ess,
the temperature, T , does not hange. Thus, the entropy relation an be integrated immediately to yield
Z
1 2

Qrev

S = dQrev =
T 1 T

where
Qrev is the heat absorbed as the state hanges from 1 to 2. Now, from the rst law of thermodynami s,
the hange in total internal energy of the system is

4

E =
Qrev +
Wrev

where
Wrev is the work done on the system. Sin e, for the ideal gas,
3
E = NkT
2
3

E = Nk
T
2
and
T = 0,
E = 0 and

Qrev =
Wrev

The expansion/ ompression of the system gives rise to a hange in pressure su h that dWrev = P (V )dV , where
P (V ) = NkT=V , is given by the equation of state (ideal gas law). Thus, the total work done on the system is
Z V2
NkT V

Wrev = dV = NkT ln 2
V1 V V1

Thus,
V2

Qrev = NkT ln
V1

and
V2

S = Nk ln
V1

If we now use the statisti al denition of entropy

"  3=2 #
V 4mE 3Nk
S = Nk ln 3 + k ln N !
h 3N 2

the hange in entropy is

"  3=2 #
V 4mE 3Nk

S = Nk ln 23 + k ln N !
h 3N 2
"  3=2 #
V 4mE 3Nk
Nk ln 13 + k ln N !
h 3N 2
= Nk ln( V2 ) Nk ln( V1 )

V2
= Nk ln
V1

where = (1=h3 )(4mE=3N )3=2. Thus, we see that the two agree exa tly.
II. Iso hori heating/ ooling from temperature T1 to T2 . In an iso hori pro ess, the volume remains
onstant. Hen e,

Wrev = 0

and, from the rst law,

5

Qrev =
E

dQrev = dE

However, for the ideal gas

3
E = NkT
2
3
dE = NkdT
2
3
E1 = NkT1
2
3
E2 = NkT2
2
E2 T2
=
E1 T1
Thus, the hange in entropy is
Z T2 T2 3 Z
dQrev dT 3 T

S = = Nk = Nk ln 2
T1 T T1 2 T 2 T1

From the statisti al denition:

"  3=2 #
V 4mE2 3Nk

S = Nk ln 3 + k ln N !
h 3N 2
"  3=2 #
V 4mE1 3Nk
Nk ln 3 + k ln N !
h 3N 2
   
= Nk ln aE23=2 Nk ln aE13=2
 3=2
E
= Nk ln 2
E1
3 E
= Nk ln 2
2 E1
3 T
= Nk ln 2
2 T1
whi h agrees exa tly with the thermodynami entropy hange.
These two examples illustrate that the statisti al approa h agrees exa tly with the standard thermodynami denition
of entropy.

C. Canoni al ensemble treatment

The anoni al partition fun tion for the ideal gas is mu h easier to evaluate than the mi ro anoni al partition
fun tion. Re all the expression for the anoni al partition fun tion Q(N; V; T ):
Z PN
1 p2i
Q(N; V; T ) = dN p dN r e  i=1 2m
N !h3N

6
Note that this an be expressed as
Z N
1
dpe  p =2m
2
Q(N; V; T ) = VN
N !h 3N

sin e the Hamiltonian is ompletely separable. Evaluating the Gaussian integral gives us the nal result immediately:
"  3=2 #N
V 1 2m
Q(N; V; T ) =
N ! h3 

The expressions for the energy


E= ln Q(N; V; T )

and pressure
 
 ln Q(N; V; T )
P = kT
V
gives rise to the results E = 3NkT=2 and P V = NkT just as for the mi ro anoni al ensemble. Note also that the
entropy S (N; V; T ) given by
E
S = k ln Q(N; V; T ) +
T
be omes
"  3=2 #
V 2m 3
S (N; V; T ) = Nk ln 3 + Nk k ln N !
h  2
whi h redu es to the mi ro anoni al expression exa tly if we use the fa t that  = 3N=2E :
"  3=2 #
V 4mE 3
S (N; V; E ) = k ln
(N; V; E ) = Nk ln 3 + Nk k ln N !
h 3N 2
Thus, the anoni al and mi ro anoni al ensembles gives rise to exa tly the same thermodynami s!! Let us now look
more arefully at the expression for the entropy.

D. The Gibbs paradox

Consider an ideal gas of N parti les in a ontainer with a volume V . A partition separates the ontainer into two
se tions with volumes V1 and V2 , respe tively, su h that V1 + V2 = V . Also, there are N1 parti les in the volume V1
and N2 parti les in the volume V2 . It is assumed that the number density is the same throughout the system
N1 N
= = 2
V1 V2

7
 N1 V1 N 2 V2

N1 + N2 =N
V 1 + V2 =V
FIG. 2.

If the partition is now removed, what should happen to the total entropy? Sin e the parti les are identi al, the
total entropy should not in rease as the partition is removed be ause the two states annot be dierentiated due to
the indistinguishability of the parti les. Let us analyze this thought experiment using the lassi al expression entropy
derived above (i.e., we leave o the ln N ! term).
The entropies S1 and S2 before the partition is removed are
3
S1  N1 k ln V1 + N1 k
2
3
S2  N2 k ln V2 + N2 k
2
and the total entropy is S = S1 + S2 .
After the partition is removed, the total entropy is
3
S  (N 1 + N2 )k ln(V1 + V2 ) + (N1 + N2 )k
2
Thus, the dieren e
S = Safter Sbefore is

S = (N1 + N2 )k ln(V1 + V2 ) N1 k ln V1 N2 k ln V2
= N1 k ln(V=V1 ) + N2 k ln(V=V2 ) > 0
This ontradi ts our predi ted result that
S = 0. Therefore, the lassi al expression must not be quite right.

Let us now restore the ln N !. Using the Stirling approximation ln N ! = N ln N N , the entropy an be written as
"  3=2 #
V 2m 5
S = Nk + Nk
Nh3  2
whi h is known as the Sa kur-Tetrode equation. Using this expression for the entropy, the dieren e now be omes

8
  
V1 + V 2

S = (N1 + N2 )k ln N1 k ln(V1 =N1 ) N2 k ln(V2 =N2 )
N1 + N2
= N1 k ln(V=V1 ) + N2 k ln(V=V2 ) N1 k ln(N=N1 ) N2 k ln(N=N2 )
   
V N1 V N2
= N1 k ln + N2 k ln
N V1 N V2
However, sin e the density  = N1 =V1 = N2 =V2 = N=V is onstant, the terms appearing in the log are all 1 and,
therefore, vanish. Hen e, the hange in entropy,
S = 0 as expe ted. Thus, it seems that the 1=N ! term is absolutely
ne essary to resolve the paradox. This means that only a orre t quantum me hani al treatment of the ideal gas gives
rise to a onsistent entropy.

E. Isothermal-isobari ensemble

Finally, let us look at the ideal gas in the isothermal-isobari ensemble. The partition fun tion is given by
1 1
Z

(N; P; T ) = dV e P V Q(N; V; T )
V0 0

2m
3N=2
1
Z 1
= dV e P V V N
 V0 N !h3N 0

We hange variables to
v = P V
so that

2m
3N=2
1 1
Z 1

(N; P; T ) = dve v v N
 V0 N !h (P )N +1
3N
0
"  3=2 #N
1 2m 1 1
=
h
3  P P V0
The enthalpy is given by
= 
H ln
(N; P; T )

whi h, for the ideal gas, be omes
= 3N 5N + 1
H kT + (N + 1)kT = kT
2 2
But re all that the enthalpy is H = E + P hV i, where E is the average energy. Sin e E = 3NkT=2, we have
P hV i = (N + 1)kT
This result an also be derived by onsidering the average volume, given by
 
 ln
N +1
hV i = kT = kT
P P
But re all the work virial theorem, whi h states that
P hV i = hPint V i + kT
where Pint is the instantaneous internal pressure obtained from Pint = H=V . Thus, the appropriate form of the
ideal gas law is
hP int V i = NkT

9
 G25.2651: Statistical Mechanics

Notes for Lecture 7

I. THE GRAND CANONICAL ENSEMBLE

In the grand canonical ensemble, the control variables are the chemical potential , the volume V and the tem-
perature T . The total particle number N is therefore allowed to uctuate. It is therefore related to the canonical
ensemble by a Legendre transformation with respect to the particle number N . Its utility lies in the fact that it closely
represents the conditions under which experiments are often performed and, as we shall see, it gives direct access to
the equation of state.

A. Thermodynamics
In the canonical ensemble, the Helmholtz free energy A(N; V; T ) is a natural function of N , V and T . As usual, we
perform a Legendre transformation to eliminate N in favor of  = @A=@N :
@A
 
A~(; V; T ) = A(N (); V; T ) ? N @N
V;T
= A(N (); V; T ) ? N
It turns out that the free energy A~(; V; T ) is the quantity ?PV . We shall derive this result below in the context of
the partition function. Thus,
?PV = A(N (); V; T ) ? N
To motivate the fact that PV is the proper free energy of the grand canonical ensemble from thermodynamic
considerations, we need to introduce a mathematical theorem, known as Euler's theorem:
Euler's Theorem: Let f (x ; :::; xN ) be a function such that
1

f (x ; :::; xN ) = n f (x ; :::; xN )

1 1

Then f is said to be a homogeneous function of degree n. For example, the function f (x) = 3x is a homogeneous
2

function of degree 2, f (x; y; z ) = xy + z is a homogeneous function of degree 3, however, f (x; y) = exy ? xy is not
2 3

a homogeneous function. Euler's Theorem states that, for a homogeneous function f ,

X N @f
nf (x ; :::; xN ) = xi @x
i
1

i =1

Proof: To prove Euler's theorem, simply dierentiate the the homogeneity condition with respect to lambda:
d d n
d f (x ; :::; xN ) = d  f (x ; :::; xN )
1 1

X N
xi @ (@f
x ) = nn? f (x ; :::; xN )
1

i
1

i
=1

Then, setting  = 1, we have

X N @f
xi @x = nf (x ; :::; xN )
i
1

i=1

which is exactly Euler's theorem.

1
 Now, in thermodynamics, extensive thermodynamic functions are homogeneous functions of degree 1. Thus, to see
how Euler's theorem applies in thermodynamics, consider the familiar example of the Gibbs free energy:
G = G(N; P; T )
The extensive dependence of G is on N , so, being a homogeneous function of degree 1, it should satisfy
G(N; P; T ) = G(N; P; T )
Applying Euler's theorem, we thus have
@G = N
G(N; P; T ) = N @N
or, for a multicomponent system,
X
G= j Nj
j
But, since
G = E ? TS + PV
it can be seen that G = N is consistent with the rst law of thermodynamics.
Now, for the Legendre transformed free energy in the grand canonical ensemble, the thermodynamics are
dA~ = dA ? dN ? Nd = ?PdV ? SdT ? Nd
But, since
A~ = A~(; V;
!T ) ! !
@ ~
A @ ~
A @ ~
A
dA~ = @ d + @V dV + @T dT
V;T ;T ;V
the thermodynamics will be given by
!
@ A~
N = ? @
!V;T
@A~
P = ? @V
! ;T
~
S = ? @@TA
V;
Since, A~ is a homogeneous function of degree 1, and its extensive argument is V , it should satisfy
A~(; V; T ) = A~(; V; T )
Thus, applying Euler's theorem,
@ A~ = ?PV
A~(; V; T ) = V @V
and since
A~ = A ? N = E ? TS ? N
the assignment A~ = ?PV is consistent with the rst law of thermodynamics. It is customary to work with PV ,
rather than ?PV , so PV is the natural free energy in the grand canonical ensemble, and, unlike the other ensembles,
it is not given a special name or symbol!

2
 B. Partition function

Consider two canonical systems, 1 and 2, with particle numbers N and N , volumes V and V and at temperature 1 2 1 2

T . The systems are in chemical contact, meaning that they can exchange particles. Furthermore, we assume that
N >> N and V >> V so that system 2 is a particle reservoir. The total particle number and volume are
2 1 2 1

V =V +V 1 2

N =N +N 1 2

The total Hamiltonian H (x; N ) is

H (x; N ) = H (x ; N ) + H (x ; N ) 1 1 1 2 2 2

If the systems could not exchange particles, then the canonical partition function for the whole system would be
Z
Q(N; V; T ) = N !h1 N dxe? H1 1 ;N1 H2 2 ;N2
3
( (x )+ (x ))

= N N!N! ! Q (N ; V ; T )Q (N ; V ; T )
1 2
1 1 1 2 2 2

where
Z
Q (N ; V ; T ) = N !h1 N1 dxe?H1 1 ;N1
1 1 1
(x )

Z
3
1

1
Q (N ; V ; T ) = N !h N2 dxe?H2 2 ;N2 (x )
2 2 2
3
2

However, N and N are not xed, therefore, in order to sum over all microstates, we need to sum over all values that
1 2

N can take on subject to the constraint N = N + N . Thus, we can write the canonical partition function for the
1 1 2

whole system as
N
X
Q(N; V; T ) = f (N ; N ) N N!N! ! Q (N ; V ; T )Q (N ; V ; T )
1
1 2
1 1 1 2 2 2

N1 =0

where f (N ; N ) is a function that weights each value of N for a given N . Thus,

1 2 1

f (0; N ) is the number of congurations with 0 particles in V and N particles in V . 1 2

f (1; N ) is the number of congurations with 1 particles in V and N ? 1 particles in V . 1 2

etc.
Determining the values of f (N ; N ) amounts to a problem of counting the number of ways we can put N identical
1

objects into 2 baskets. Thus,

f (0; N ) = 1
f (1; N ) = N = N !=1!(N ? 1)!
f (2; N ) = N (N ? 1)=2 = N !=2!(N ? 2)!
etc.
or generally,
f (N ; N ) = N !(NN?! N )! = N N!N! !
1
1 1 1 2

which is clearly a classical degeneracy factor. If we were doing a purely classical treatment of the grand canonical
ensemble, then this factor would appear in the sum for Q(N; V; T ), however, we always include the ad hoc quantum

3
correction 1=N ! in the expression for the canonical partition function, and we see that these quantum factors will
exactly cancel the classical degeneracy factor, leading to the following expression:
N
X
Q(N; V; T ) = Q (N ; V ; T )Q (N ; V ; T )
1 1 1 2 2 2

N1 =0

which expresses the fact that, in reality, the various congurations are not distinguishable from each other, and so
each one should count with equal weighting. Now, the distribution function (x) is given by
e?H ;N
(x; N ) = N hQ3N(N; V; T )
1 (x )
!

which is chosen so that

Z
dx(x; N ) = 1
However, recognizing that N  N , we can obtain the distribution for  (x ; N ) immediately, by integrating over the
2 1 1 1

phase space of system 2:

1 1 1 Z
 (x ; N ) = Q(N; V; T ) N !h N1 e ?H ;N ?H2 2 ;N2
N !h N2 dx e
1 1 1 (x ) (x )
1 1 1 2
3 3
1 2

where the 1=N !h N1 prefactor has been introduced so that

1
3

N Z
X
dx (x ; N ) = 1
1 1 1

N1 =0

and amounts to the usual ad hoc quantum correction factor that must be multiplied by the distribution function for
each ensemble to account for the identical nature of the particles. Thus, we see that the distribution function becomes
 (x ; N ) = QQ((NN;;V;V T; T) ) N !h1 N1 e?H1 1 ;N1
1 1 1
2 2 2

3
(x )

Recall that the Hemlholtz free energy is given by

A = ? 1 ln Q 
Thus,
Q(N; V; T ) = e?A N;V;T ( )

Q (N ; V ; T ) = e?A N2;V2 ;T = e?A N ?N1;V ?V1 ;T

2 2 2
( ) ( )

or
Q (N ; V ; T ) = e? A N ?N1;V ?V1 ;T ?A N;V;T
2 2 2 ( ( ) ( ))

Q(N; V; T )
But since N >> N and V >> V , we may expand:
1 1

@A N ? @A V +



A(N ? N ; V ? V ; T ) = A(N; V; T ) ? @N
1 1
@V 1 1

= A(N; V; T ) ? N + PV +


1 1

Therefore the distribution function becomes

 (x ; N ) = N !h1 N1 eN1 e?PV1 e?H1 1 ;N1
1 1 1
3
(x )

= N !h1 N1 ePV
1 eN1 e?H1 1 ;N1
1 3
(x )

4
Dropping the \1" subscript, we have
1
 1 
(x; N ) = ePV e N e?H ;N (x )

N !h N 3

We require that (x; N ) be normalized:

1 Z
X
dx(x; N ) = 1
"1 N #
1 eN Z dxe?H ;N
=0

1 X
ePV N N !h N 3
(x )
=1
=0

Now, we dene the grand canonical partition function

Z (; V; T ) =
1
X 1 eN Z dxe?H ;N (x )

N =0
N !h N
3

Then, the normalization condition clearly requires that

Z (; V; T ) = ePV
ln Z (; V; T ) = PV
kT
Therefore PV is the free energy of the grand canonical ensemble, and the entropy S (; V; T ) is given by
   
S (; V; T ) = @ (PV )
@T = k ln Z (; V; T ) ? k @ ln Z (; V; T ) @
;V ;V
We now introduce the fugacity  dened to be
 = e
Then, the grand canonical partition function can be written as
X1 Z X1
Z (; V; T ) = N !h1 N  N dxe?H ;N =  N Q(N; V; T )
3
(x )

N =0 N =0

which allows us to view the grand canonical partition function as a function of the thermodynamic variables  , V ,
and T .
Other thermodynamic quantities follow straightforwardly:
Energy:
X1 N Z
E = hH (x; N )i = Z1 N !h N dxH (x; N )e3
?H ;N (x )

N =0

= ? @ ln Z (; V; T ) @ ;V

Average particle number:

 
hN i = kT @ ln Z@
(; V; T )
V;T
This can also be expressed using the fugacity by noting that
@ @ @ @
@ = @ @ =  @
Thus,
hN i =  @@ ln Z (; V; T )

5
 C. Ideal gas
Recall the canonical partition function expression for the ideal gas:
"   #N
Q(N; V; T ) = N1 ! hV 2m = 3 2

3

Dene the thermal wavelength ( ) as
 =
( ) = 2h
2 1 2

m
which has a quantum mechanical meaning as the width of the free particle distribution function. Here it serves as a
useful parameter, since the canonical partition can be expressed as
1
 V N
Q(N; V; T ) = N !  3

The grand canonical partition function follows directly from Q(N; V; T ):

X1 1  V  N
Z (; V; T ) = = e V =3
N N! 
3
=0

Thus, the free energy is

PV = ln Z = V 
kT  3

In order to obtain the equation of state, we rst compute the average particle number hN i
hN i =  @@ ln Z = V  3

Thus, eliminating  in favor of hN i in the equation of state gives

PV = hN ikT
as expected. Similarly, the average energy is given by
 
E=? @ ln Z = 3V  @
@ = 3 hN ikT
@ V 4
2
where the fugacity has been eliminated in favor of the average particle number. Finally, the entropy
 @ ln Z (; V; T )  5
V  
S (; V; T ) = k ln Z (; V; T ) ? k = hN ik + hN ik ln hN i
3

@ ;V 2
which is the Sackur-Tetrode equation derived in the context of the canonical and microcanonical ensembles.

D. Particle number uctuations

In the grand canonical ensemble, the particle number N is not constant. It is, therefore, instructive to calculate
the uctuation in this quantity. As usual, this is dened to be
p

N = hN i ? hN i 2 2

Note that

6
 1 "1 #
@ @ 1 X N 1 X N 2

 @  @ ln Z (; V; T ) = Z N  Q(N; V; T ) ? Z
2
N Q(N; V; T ) 2
N =0 N =0

= hN i ? hN i
2 2

Thus,
(
N ) =  @
2 @  @ ln Z (; V; T ) = (kT ) @ ln Z (; V; T ) = kTV @ P
2
2 2

@ @ @
2 2

In order to calculate this derivative, it is useful to introduce the Helmholtz free energy per particle dened as follows:
a(v; T ) = 1 A(N; V; T )
N
where v = V=N = 1= is the volume per particle.
The chemical potential is dened by
@A = a(v; T ) + N @a @v
 = @N @v @N
@a
= a(v; T ) ? v @v
Similarly, the pressure is given by
@A = ?N @a @v = ? @a
P = ? @V @v @V @v
Also,
@ = ?v @ a 2

@v @v 2

Therefore,
 
@P = @P @v = @ a v @ a ? = 1
2 2 1

@ @v @ @v @v v 2 2

and
2
 
@ P = @ @P @v = 1 v @ a ? = ? 1 2 1

@ @v @ @ v @v
2
v @P=@v
2 2 3

But recall the denition of the isothermal compressibility:

T = ? V1 @V
@P = ? 1
v@p=@v
Thus,
@ P = 1 2

@ 2
v T 2

and
r

N = hN ikTT
v
and the relative uctuation is given by
r

N = 1 hN ikTT  p 1 ! 0 ashN i ! 1
N hN i v hN i
Therefore, in the thermodynamic limit, the particle number uctuations vanish, and the grand canonical ensemble is
equivalent to the canonical ensemble.

7
 G25.2651: Statistical Mechanics

Notes for Lecture 8

I. STRUCTURE AND DISTRIBUTION FUNCTIONS IN CLASSICAL LIQUIDS AND GASES

So far, we have developed the classical theory of ensembles and applied it to the ideal gas, for which there was no
potential of interaction between the particles: U = 0. We were able to derive all the thermodynamics for this sytem
using the various ensembles available to us, and, in particular, we could compute the equation of state. Now, we wish
to consider the general case that the potential is not zero. Of course, all of the interesting physics and chemistry of
real systems results from the specic interactions between the particles. Real systems can exhibit spatial structure,
undergo phase transitions, undergo chemical changes, exhibit interesting dynamics, basically, a wide variety of rich
behavior.
Consider the two snapshots below.
On the left is shown a conguration of an ideal gas, and on the right is shown a conguration of liquid argon.
Can you see any inherent structure in the snapshot of liquid argon? While it may not be readily apparent, there is
considerable structure in the liquid argon system that is clearly not present in the ideal gas. One way of quantifying
spatial structure is through the use of the radial distribution function g(r), which will be discussed in great detail
later. For now, it is sucient to know that g(r) is a measure of the probability that a particle will be located a
distance r from a another particle in the system. The gure below shows the function g(r) for the ideal gas and for
the liquid argon systems.
It can be seen that the radial distribution function for the ideal gas is completely featureless signifying that it is
equally likely to nd a particle at any distance r from a given particle. (Since the probability is uniform, g(r)  1=r2
for small r. This is the particularly normalization condition on g(r) that gives rise to uniform probability. Hence
its rapidly rising behavior for small r.) For the liquid argon system, g(r) exhibits several peaks, indicating that at
certain radial values, it is more likely to nd particles than at others. This is a result of the attractive nature of the
interaction at such distances. The plot of g(r) also shows that there is essentially zero probability of nding particles
at distances less than about 2.5A from each other. This is due to the presence of very strong repulsive forces at short
distances.
Sometimes, the structure can be readily seen in snapshots of congurations. Consider the following snapshot of a
system of water molecules:
The red spheres are oxygen atoms, the grey spheres are hydrogen atoms, the green lines are hydrogen bonds, and
the reddish-grey lines are covalent bonds. A good deal of structure can be seen in the form of a complex network of
hydrogen bonds. This high degree of structure is characteristic of water and gives rise to the ease which with water
can form stable, organized structures around other molecules. In water, one can ask several questions related to
structure. For example, what is the probability that an oxygen atom will be found at a distance r away from another
oxygen atoms? What is the probability that a hydrogen atom will be located at a distance r from an oxygen atom,
etc. The plot below shows the radial distribution functions corresponding to these two scenarios.
The peak in the O-O radial distribution function occurs at roughly 2.8 A which is the well known average hydrogen
bond length in water. Of course, one could also ask about structure from the point of view of a hydrogen atom and
obtain two other representations of structure in water.
One can also ask questions pertinent to chemical reactions. The following three snapshots depict a proton transfer
reaction in water upon solvation of an excess proton:
While the reaction dynamics is expected to be very complicated, involving the breaking of hydrogen bonds and other
uctuations away from the rst solvation shell, it might be possible to characterize certain features of the reaction
by a single, special coordinate that somehow describes the motion of the proton through the hydrogen bond. Such
a coordinate is called a reaction coordinate and can, in general, be a funtion of all of some subset of the cartesian
positions of the other atoms in the system:
qreac = q(r1 ; r2 ; :::; rN )
There are many things one can do with such a coordinate, such as study its dynamics or obtain its distribution and
hence its free energy prole. In the case of the latter, the free energy prole describes the \eective" potential at a
given temperature that the proton sees as it moves through the hydrogen bond. Computing this free energy prole,
one might obtain a plot that looks as follows:

1
 In this case, the free energy prole is obtained with and without quantum eects (which cannot be neglected for
protons). For the classical prole, the energy dierence between the top of the barrier and the wells is the eective
activation energy. Quantum mechanically, this activation energy is reduced due to zero-point motion (a concept we
will discuss in the context of quantum statistical mechanics).

A. General distribution functions and correlation functions

We begin by considering a general N -particle system with Hamiltonian
3N

H=
X p2i + U (r ; :::; r )
N
i=1
2m 1

For simplicity, we consider the case that all the particles are of the same type. Having established the equivalence
of the ensembles in the thermodynamic limit, we are free to choose the ensemble that is the most convenient on in
which to work. Thus, we choose to work in the canonical ensemble, for which the partition function is
Z P
Q(N; V; T ) = N !h3N 1 3N p 2
d p d re? i=1 2mi e?U (r1 ;:::;rN )
3N 3N

The 3N integrations over momentum variables can be done straighforwardly, giving

Z
Q(N; V; T ) = N !1 3N dr1


drN e?U (r1 ;:::;rN )
= NZ!N3N
p
where  = h2 =2m is the thermal wavelength and the quantity ZN is known as the congurational partition
function
Z
ZN = dr1


drN e?U (r1 ;:::;rN )
The quantity
e?U (r1 ;:::;rN ) dr


dr  P (N ) (r ; :::; r )dr


dr
N N N
ZN 1 1 1

represents the probability that particle 1 will be found in a volume element dr1 at the point r1 , particle 2 will be
found in a volume element dr2 at the point r2 ,..., particle N will be found in a volume element drN at the point rN .
To obtain the probability associated with some number n < N of the particles, irrespective of the locations of the
remaining n + 1; :::; N particles, we simply integrate this expression over the particles with indices n + 1; :::; N :
Z 
P (n) (r1 ; :::; rn )dr1


drn = Z1 drn+1


drN e?U (r1 ;:::;rN ) dr1


drn
N

The probability that any particle will be found in the volume element dr1 at the point r1 and any particle will be
found in the volume element dr2 at the point r2 ,...,any particle will be found in the volume element drn at the point
rn is dened to be

(n) (r1 ; :::; rn )dr1


drn = (N N?!n)! P (n)(r1 ; :::; rn )dr1


drn
which comes about since the rst particle can be chosen in N ways, the second chosen in N ? 1 ways, etc.
Consider the specal case of n = 1. Then, by the above formula,
Z
(1) (r1 ) = Z1 (N N?! 1)! dr2


drN e?U (r1 ;:::;rN )
N
Z
=Z N dr2


drN e?U (r1 ;:::;rN )
N

2
Thus, if we integrate over all r1 , we nd that
Z
1
V dr1 (1) (r1 ) = N
V =
Thus, (1) actually counts the number of particles likely to be found, on average, in the volume element dr1 at the
point r1 . Thus, integrating over the available volume, one nds, not surprisingly, all the particles in the system.

3
 G25.2651: Statistical Mechanics

Notes for Lecture 9

I. DISTRIBUTION FUNCTIONS IN CLASSICAL LIQUIDS AND GASES (CONT'D)

A. General correlation functions
A general correlation function can be dened in terms of the probability distribution function (n) (r1 ; :::; rn ) ac-
cording to
g(n) (r ; :::; r ) = 1 (n) (r ; :::; r )
1 n n 1 n
n
= Z NVn (NN !? n)! drn+1


drN e?U (r1 ;:::;rN )
Z

N
Another useful way to write the correlation function is
n
g(n)(r1 ; :::; rn ) = Z NVn (NN !? n)! dr01


dr0N e?U (r1 ;:::;rN ) (r1 ? r01 )


(rn ? r0n )
Z
0 0

N
V nN !
*
n +

= Z N n (N ? n)!
Y
0
(ri ? ri )
N i=1 r1 ;:::;rN
0 0

i.e., the general n-particle correlation function can be expressed as an ensemble average of the product of -functions,
with the integration being taken over the variables r01 ; :::; r0N .

B. The pair correlation function

Of particular importance is the case n = 2, or the correlation function g(2) (r1 ; r2 ) known as the pair correlation
function. The explicit expression for g(2)(r1 ; r2 ) is
g(2) (r1 ; r2 ) = N 2V(NN?! 2)! h(r1 ? r01 )(r2 ? r02 )i
2

= V (N ? 1) dr3


drN e?U (r1 ;:::;rN )

2
Z

NZN
= N2? 1) h(r1 ? r01 )(r2 ? r02 )ir1 ;:::;rN
N (
0 0

In general, for homogeneous systems in equilibrium, there are no special points in space, so that g(2) should depend
only on the relative position of the particles or the dierence r1 ? r2 . In this case, it proves useful to introduce the
change of variables
r = r 1 ? r2 ; R = 21 (r1 + r2 )
r1 = R + 21 r; r2 = R ? 12 r
Then, we obtain a new function g~(2) , a function of r and R:
g~(2)(r; R) = V (N ? 1) dr3


drN e?U (R+ 12 r;R? 12 r;r3 ;:::;rN )
2 Z

NZN  
N (
= 2N ? 1) 1 0
 
1 0
 R + 2 r ? r1  R ? 2 r ? r2


r1 ;:::;rN 0 0

1
In general, we are only interested in the dependence on r. Thus, we integrate this expression over R and obtain a
new correlation function g~(r) dened by
Z

g~(r) = 1 dRg~(2)(r; R)
V
V (
= NZN ? 1) Z

dRdr3


drN e?U (R+ 12 r;R? 21 r;r3 ;:::;rN )

N Z
(N
= Z ? 1) dRdr3


drN e?U (R+ 21 r;R? 12 r;r3 ;:::;rN )
N
For an isotropic system such as a liquid or gas, where there is no preferred direction in space, only the maginitude
or r, jrj  r is of relevance. Thus, we seek a choice of coordinates that involves r explicitly. The spherical-polar
coordinates of the vector r is the most natural choice. If r = (x; y; z ) then the spherical polar coordinates are
x = r sin  cos 
y = r sin  sin 
x = r cos 
dr = r2 sin drdd
where  and  are the polar and azimuthal angles, respectively. Also, note that
r = rn
where
n = (sin  cos ; sin  sin ; cos )
Thus, the function g(r) that depends only on the distance r between two particles is dened to be
Z
1
g(r) = 4 sinddg~(r)
? 1) Z sindddRdr


dr e?U (R+ 21 rn;R? 12 rn;r3 ;:::;rN )
= (4NZ 3 N
N 
? 1) (r ? r0 )
= (N4


rr0 r ; ; ;R ;r3 ;:::;rN

0 0 0 0 0 0

Integrating g(r) over the radial dependence, one nds that

Z 1
4 drr2 g(r) = N ? 1  N
0

The function g(r) is important for many reasons. It tells us about the structure of complex, isotropic systems, as
we will see below, it determines the thermodynamic quantities at the level of the pair potential approximation, and
it can be measured in neutron and X-ray diraction experiments. In such experiments, one observes the scattering of
neutrons or X-rays from a particular sample. If a detector is placed at an angle  from the wave-vector direction of
an incident beam of particles, then the intensity I () that one observes is proportional to the structure factor
* 2 +

I ()  N1
X 



eik
rm 

*
m +

= N1 eik
(rm ?rn)
X

m;n
 S (k)
where k is the vector dierence in the wave vector between the incident and scattered neutrons or X-rays (since
neutrons and X-rays are quantum mechanical particles, they must be represented by plane waves of the form exp(ik
r)).

2
By computing the ensemble average (see problem 4 of problem set #5), one nds that S (k) = S (k) and S (k) is given
by
Z 1

S (k) = 1 + 4 drr sin(kr)g(r)

k 0

Thus, if one can measure S (k), g(r) can be determined by Fourier transformation.

C. Thermodynamic quantities in terms of g(r)

In the canonical ensemble, the average energy is given by
@ ln Q(N; V;  )
E = ? @

ln Q(N; V;  ) = ln ZN ? 3N ln ( ) ? ln N !
Therefore,
@ ? 1 @ZN
E = 3N @ ZN @
Since
2 1=2
 = 2h
 

m
@ = 1 
@ 2
Thus,
Z

E = 32 NkT + Z1 dr1


drN U (r1 ; :::; rN )eU (r1 ;:::;rN )

N
= 23 NkT + hU i
In order to compute the average energy, therefore, one needs to be able to compute the average of the potential hU i.
In general, this is a nontrivial task, however, let us work out the average for the case of a pairwise-additive potential
of the form
U (r1 ; :::; rN ) = 21
X
u(jri ? rj j)  Upair (r1 ; :::; rN )
i;j;i6=j
i.e., U is a sum of terms that depend only the distance between two particles at a time. This form turns out to be an
excellent approximation in many cases. U therefore contains N (N ? 1) total terms, and hU i becomes
Z

hU i = 2Z1 dr1


drN u(jri ? rj j)e?Upair (r1 ;:::;rN )

X

N i;j;i6=j

= N (2NZ ? 1) dr1


drN u(jr1 ? r2 j)e?Upair (r1 ;:::;rN )

Z

N
The second line follows from the fact that all terms in the rst line are the exact same integral, just with the labels
changed. Thus,

3
 1 Z

hU i = 2 dr1 dr2 u(jr1 ? r2 j) Z N (N ?


1) Z

dr3


drN e?Upair (r1 ;:::;rN )



Z
N
= 21 dr1 dr2 u(jr1 ? r2 j)(2) (r1 ; r2 )
= 2NV 2
2
Z

dr1 dr2 u(jr1 ? r2 j)g(2) (r1 ; r2 )

Once again, we change variables to r = r1 ? r2 and R = (r1 + r2 )=2. Thus, we nd that

hU i = 2NV
2
Z

2
drdRu(r)~g(2) (r; R)
= 2NV 2
2
Z Z

dru(r) dRg~(2)(r; R)
N 2
Z

= 2V dru(r)~g (r)
Z 1
N
= 2V
2
dr4r2 u(r)g(r)
0

Therefore, the average energy becomes

1
E = 23 NkT + N2 4
Z

drr2 u(r)g(r)
0

Thus, we have an expression for E in terms of a simple integral over the pair potential form and the radial distribution
function. It also makes explicit the deviation from \ideal gas" behavior, where E = 3NkT=2.
By a similar procedure, we can develop an equation for the pressure P in terms of g(r). Recall that the pressure is
given by
P = 1 @ ln Q  @V
= Z1 @Z
@
N
N
The volume dependence can be made explicit by changing variables of integration in ZN to
si = V ?1=3 ri
Using these variables, ZN becomes
Z

ZN = V N ds1


dsN e?U (V 1=3 s1 ;:::;V 1=3 sN )

Carrying out the volume derivative gives
@ZN = N Z ? V N Z ds


ds 1 X N @U e?U (V 1=3 s1 ;:::;V 1=3 sN )
@V V N 1 N r i

3V i=1 @ ri
=N
Z
1 X r
F e?U (r1 ;:::;rN )
Z N +  dr 1


drN
V 3V i=1 i i
Thus,
1 @ZN = N + 
*
X N +

ZN @V V 3V ri
Fi
i=1
Let us consider, once again, a pair potential. We showed in an earlier lecture that

4
 X N X XN
ri
F i = ri
Fij
i=1 i=1 j =1;j 6=i
where Fij is the force on particle i due to particle j . By interchaning the i and j summations in the above expression,
we obtain
2 3
N
ri
Fi = 21 4
X X X
ri
Fij + rj
Fji 5
i=1 i;j;i6=j i;j;i6=j
However, by Newton's third law, the force on particle i due to particle j is equal and opposite to the force on particle
j due to particle i:
Fij = ?Fji
Thus,
2 3
N
ri
Fi = 21 4 rj
Fij 5 = 12 (ri ? rj )
Fij  21
X X X X X
ri
Fij ? rij
Fij
i=1 i;j;i6=j i;j;i6=j i;j;i6=j i;j;i6=j
The ensemble average of this quantity is
 XN
* + * +
 X r
F =  Z dr


dr X r
F e?Upair (r1 ;:::;rN )
3V i=1 ri
F i = 6V i;j;i6=j ij ij 6V ZN 1 N ij ij
i;j;i6=j
As before, all integrals are exactly the same, so that
 XN
*
N (
+
N ? 1) Z

3V i=1 ri
Fi = 6V ZN dr1


rN r12
F12 e?Upair (r1 ;:::;rN )

 Z

= 6V dr1 dr2 r12
F12


N (N ? 1) Z

dr3


drN e ? U pair (r1 ;:::;rN )



ZN
= 6V dr1 dr2 r12
F12 (2) (r1 ; r2 )
Z

= N
2
Z

6V 3 dr1 dr2 r12
F12 g (r1 ; r2 )

(2)

Then, for a pair potential, we have

F12 = ? @U pair 0 (r1 ? r2 ) 0 r12
@ r = ?u (jr1 ? r2 j) jr ? r j = ?u (r12 ) r
12 1 2 12

where u0 (r) = du=dr, and r12 = jr12 j. Substituting this into the ensemble average gives
 X N * +
N 2 Z dr dr u0 (r )r g(2) (r ; r )
r
3V i=1 i i
F = ? 6V 3 1 2 12 12 1 2

As in the case of the average energy, we change variables at this point to r = r1 ? r2 and R = (r1 + r2 )=2. This gives
 XN
*
N
+
2
Z
0
3V i=1 ri
Fi = ? 6V 3 drdRu (r)rg~ (r; R)
(2)

N 2
Z

= ? 6V 2 dru0 (r)rg~(r)
2 Z 1
N
= ? 6V 2 dr4r3 u0 (r)g(r)
0

5
Therefore, the pressure becomes
P =  ? 2 Z 1 dr4r3 u0 (r)g(r)
kT 6kT 0
which again gives a simple expression for the pressure in terms only of the derivative of the pair potential form and
the radial distribution function. It also shows explicitly how the equation of state diers from the that of the ideal
gas P=kT = .
From the denition of g(r) it can be seen that it depends on the density  and temperature T : g(r) = g(r; ; T ).
Note, however, that the equation of state, derived above, has the general form
P =  + B2
kT
which looks like the rst few terms in an expansion about ideal gas behavior. This suggests that it may be possible to
develop a general expansion in all powers of the density  about ideal gas behavior. Consider representing g(r; ; T )
as such a power series:
1
X
g(r; ; T ) = j gj (r; T )
j =0
Substituting this into the equation of state derived above, we obtain
P = +X1
kT Bj+2 (T )j+2
j =0
This is known as the virial equation of state, and the coecients Bj+2 (T ) are given by
Z 1
1
Bj+2 (T ) = ? 6kT dr 4r3 u0(r)gj (r; T )
0

are known as the virial coecients. The coecient B2 (T ) is of particular interest, as it gives the leading order deviation
from ideal gas behavior. It is known as the second virial coecient. In the low density limit, g(r; ; T )  g0 (r; T ) and
B2 (T ) is directly related to the radial distribution function.

6
 G25.2651: Statistical Mechanics

Notes for Lecture 10

I. DISTRIBUTION FUNCTIONS AND PERTURBATION THEORY

A. General formulation
Recall the expression for the congurational partition function:
Z
ZN = dr1


drN e?U (r1 ;:::;rN )
Suppose that the potential U can be written as a sum of two contributions
U (r1 ; :::; rN ) = U0 (r1 ; :::; rN ) + U1 (r1 ; :::; rN )
where U1 is, in some sense, small compared to U0 . An extra bonus can be had if the partition function for U0 can be
evaluated analytically.
Let
Z
ZN (0) = dr1


drN e?U0(r1 ;:::;rN )
Then, we may express ZN as
Z
ZN = ZZN (0) dr1


drN e?U0(r1 ;:::;rN ) e?U1 (r1 ;:::;rN )
(0)

N
= ZN (0) he?U1 (r1 ;:::;rN ) i0
where h


i0 means average with respect to U0 only. If U1 is small, then the average can be expanded in powers of U1 :
he?U1 i = 1 ?  hU i +  hU 2 i ?  hU 3 i +



2 3
0 1 0
2! 1 0
3! 1 0

1
X (? )k hU k i
=
k=0 k!
1 0

The free energy is given by

   
A(N; V; T ) = ? 1 ln NZ!N3N = ? 1 ln NZN!3N ? 1 lnhe?U1 i0
(0)

Separating A into two contributions, we have

A(N; V; T ) = A(0) (N; V; T ) + A(1) (N; V; T )
where A(0) is independent of U1 and is given by
 
A(0) (N; V; T ) = ? 1 ln NZN!3N
(0)

and
A(1) (N; V; T ) = ? 1 lnhe?U1 i
 0

1 (? )k
= ? 1 lnh
X
k
k=0 k! hU1 i0

1
We wish to develop an expansion for A(1) of the general form
1
X (? )k?1 !
A(1) = k
k=1 k!
where !kPare a set of expansion coecients that are determined by the condition that such an expansion be consistent
with lnh 1 k k
k=0 (? ) hU1 i0 =k!.
Using the fact that
1 k
(?1)k?1 xk
X
ln(1 + x) =
k=1
we have that
! !
1 X1 (? )k 1 X1 (? )k
?  ln k ! hU1k i0 = ?  ln 1 + k ! hU1k i0
k=0 k=0
!k
1 1(? )l hU l i
= ? 1 (?1)k?1 1
X X
 k=1 k l=1 l! 1 0
Equating this expansion to the proposed expansion for A(1) , we obtain
!k
1 1
(? )l hU l i 1
(?1)k?1 1 (? )k !kk!
X X X
k l=1 l! 1 0 =
k=1 k=1
This must be solved for each of the undetermined parameters !k , which can be done by equating like powers of  on
both sides of the equation. Thus, from the  1 term, we nd, from the right side:
Right Side : ? !
1!
1

and from the left side, the l = 1 and k = 1 term contributes:

Left Side : ?  hU1!1 i0
from which it can be easily seen that
!1 = hU1 i0
Likewise, from the  2 term,
Right Side : 2 !
2! 2
and from the left side, we see that the l = 1; k = 2 and l = 2; k = 1 terms contribute:
Left Side :  2 ?hU 2 i ? hU i2

2 1 0 1 0

Thus,
!2 = hU12 i0 hU1 i20
For  3 , the right sides gives:
Right Side : ? 3! !3
3

the left side contributes the l = 1; k = 3, k = 2; l = 2 and l = 3; k = 1 terms:

2
  2
Left Side : ? 6 hU13 i + (?1)2 31 (? hU1 i0 )3 ? 21 ? hU1 i0 + 21  2 hU12 i
3

Thus,
!3 = hU13 i0 + 2hU1i30 ? 3hU1i0 hU12 i0
Now, the free energy, up to the third order term is given by
A = A(0) + !1 ? 2 !2 + 6 !3



2

 
= ? 1 ln NZN!3N + hU1 i0 ? 2 hU12 i0 ? hU1 i20 + 6 hU13 i0 ? 3hU1 i0 hU12 i0 + 2hU1i30 +



(0) ?
2 ?

In order to evaluate hU1 i0 , suppose that U1 is given by a pair potential

U1 (r1 ; :::; rN ) = 12 u1 (jri ? rj j)
X

i6=j
Then,
Z
hU i = Z dr


drN 21 u (jri ? rj j)e?U0 r1 ;:::;rN
1 X
( )
1 0 1 1
N (0)
i6 j =
Z Z
= N (N ? 1) dr dr u (jr ? r j) dr


drN e?U0 r1 ;:::;rN ( )
2ZN (0) 1 2 1 1 2 3

Z
= 2NV
2
dr dr u (jr ? r j)g (r ; r )
2 1 2 1 1 2
(2)
0 1 2

Z 1
=V
2

2 0 4r u (r)g (r)dr

1 0
2

The free energy is therefore given by

  Z 1
1 ZN (0)
1
A(N; V; T ) = ?  ln N !3N + 2  V 2
4r2 u1 (r)g0 (r)dr ? 2 hU12 i0 ? hU1 i20



?

0

B. Derivation of the Van der Waals equation

As a specic example of the application of perturbation theory, we consider the Van der Waals equation of state.
Let U0 be given by a pair potential:
U0 (r1 ; :::; rN ) = 21 u0 (jri ? rj j)
X

i6=j
with

0
u0 (r) = 1 r>
r
This potential is known as the hard sphere potential. In the low-density limit, the radial distribution function can be
shown to be given correctly by g0 (r)  exp(?u0 (r)) or

g0 (r) = 10 r>
r
= (r ? )
u1 (r) is taken to be some arbitrary attractive potential, whose specic form is not particularly important. Then, the
full potential u(r) = u0 (r) + u1 (r) might look like:

3
 u(r)

r
σ

FIG. 1.

Now, the rst term in A(1) is

Z 1
A(1)  = 21 2 V 4r2 u1 (r)g0 (r)
0
Z 1
= 21 2 V 4r2 u1 (r)(r ? )
0
Z 1
= 22 V r2 u1 (r)dr  ?aN

where
Z 1
a = ?2 drr2 u1 (r) > 0

is a number that depends on  and the specic form of u1 (r).
Since the potential u0 (r) is a hard sphere potential, ZN (0) can be determined analytically. If  were 0, then u0
would describe an ideal gas and
ZN (0) = V N
However, because two particles may not approach each other closer than a distance  between their centers, there is
some excluded volume:

4
 d
d/2

d/2

FIG. 2.

If we consider two hard spheres at closest contact and draw the smallest imaginary sphere that contains both
particles, then we nd this latter sphere has a radius :

FIG. 3.

Hence the excluded volume for these two particles is

4 3
3
and hence the excluded volume per particle is just half of this:
2 3
3   b
Therefore Nb is the total excluded volume, and we nd that, in the low density limit, the partition function is given
approximately by
ZN (0) = (V ? Nb)N
Thus, the free energy is
 
A(N; V; T )  ? 1 ln (VN?!Nb)N ? aN 2
3N V

5
 If we now use this free energy to compute the pressure from

@A 
P = ? @V
N;T
we nd that
P = N ? aN 2
kT V ? Nb kTV 2
= 1 ? b ? a
2

kT
This is the well know Van der Waals equation of state. In the very low density limit, we may assume that b << 1,
hence
1
1 ? b  1 + b
Thus,
P   + 2 b ? a 
kT kT
from which we can approximate the second virial coecient:
a = 2 3 + 2 Z 1 drr2 u (r)
B2 (T )  b ? kT 3 kT  1

A plot of the isotherms of the Van der Waals equation of state is shown below:

6
 critical pt.

Critical isotherm

Volume discontinuity

FIG. 4.

The red and blue isotherms appear similar to those of an ideal gas, i.e., there is a monotonic decrease of pressure
with increasing volume. The black isotherm exhibits an unusual feature not present in any of the ideal-gas isotherms
- a small region where the curve is essentially horizontal ( at) with no curvature. At this point, there is no change
in pressure as the volume changes. Below this isotherm, the Van der Waals starts to exhibit unphysical behavior.
The green isotherm has a region where the pressure decreases with decreasing volume, behavior that is not expected
on physical grounds. What is observed experimentally, in fact, is that at a certain pressure, there is a dramatic,
discontinuous change in the volume. This dramatic jump in volume signies that a phase transition has occurred, in
this case, a change from a gaseous to a liquid state. The dotted line shows this jump in the volume. Thus, the small

7
 at neighborhood along the black isotherm becomes larger on isotherms below this one. The black isotherm just
represents a boundary between those isotherms along which no such phase transition occurs and those that exhibit
phase transitions in the form of discontinuous changes in the volume. For this reason, the black isotherm is called the
critical isotherm, and the point at which the isotherm is at and has zero curvature is called a critical point.
A critical point is a point at which
@P = 0; @2P = 0
@V @V 2
Using these two conditions, we can solve for the critical volume Vc and critical temperature Tc :
Vc = 3Nb 8a
kTc = 27 b
and the critical pressure is therefore
Pc = 27ab2
Using these values for the critical pressure, temperature and volume, we can show that the isothermal compress-
ibility, given by
 = 1 @V
T
V @P
diverges as the critical point is approached. To see this, note that

@P  = ? NkT + 2aN 2 = ? kT + 2a = 1  8a ? kT   (T ? T )
@V V =Vc 2N 2b2 27N 3b3 4Nb2 27Nb3 4Nb2 27b c

Thus,
T  (T ? Tc)?1
It is observed that at a critical point, T diverges, generally, as jT ? Tcj? . To determine the heat capacity, note that

Q(N; V; T ) = e?A(N;V;T ) = (VN?!Nb )N eaN 2=V

3N

so that
E = ? @@ ln Q(N; V; T )

= ? @@ N ln(V ? Nb) ? ln N ! ? 3N ln  + aN 2 
V
= 3N @
@ + aN 2

V
= 3N 2 + aN
2

V
= 32 NkT + aN
2

V
Then, since
 
CV = @E
@T V
it follows that
CV = 32 Nk  jT ? Tcj0

8
The heat capacity is observed to diverge as jT ? Tc j? . Exponents such as  and are known as critial exponents.
Finally, one other exponent we can easily determine is related to how the pressure depends on density near the
critical point. The exponent is called , and it is observed that
P  const + C ( ?  )
kT c
What does our theory predict for ? To determine  we expand the equation of state about the critical density and
temperature:
  
P = Pc + 1 @P  ( ?  ) + 1 @ 2 P  ( ?  )2 + 1 @ 3 P  ( ?  )3 +



kT kTc kTc @ =c c 2kT @2  c 6kTc @3 =c c
c =c
 " # 
Pc + 1
= kT @P @V  ( ?  ) + 1 @ 2 P  @V 2 + @P @ 2 V  ( ?  )2 + 1 @ 3 P  ( ?  )3 +



c 2kT @V 2 @ c c
c kTc @V @ =c c @V @2 =c 6kTc @3 =c
The second and third terms vanish by the conditions of the critical point. The third derivative term can be worked
out straightforwardly and does not vanish. Rather

1 @ 3 P  243 2
kTc @3 =c = 8 b 6= 0
Thus, we see that, by the above expansion,  = 3.
The behavior of these quantities near the critical temperature determine three critical exponents. To summarize
the results, the Van der Waals theory predicts that
=0 =1 =3
The determination of critical exponents such as these is an active area in statistical mechanical research. The reason
for this is that such exponents can be grouped into universality classes { groups of systems with exactly the same
sets of critical exponents. The existence of universality classes means that very dierent kinds of systems exhibit
essentially the same behavior at a critical point, a fact that makes the characterization of phase transitions via critical
exponents quite general. The values obtained above for , and  are known as the mean-eld exponents and
shows that the Van der Waals theory is really a mean eld theory. These exponents do not agree terribly well with
experimental values ( = 0:1; = 1:35,  = 4:2). However, the simplicty of mean-eld theory and its ability to give, at
least, qualitative results, makes it, nevertheless, useful. To illustrate universality classes, it can be shown that, within
mean eld theory, the Van der Waals gas/liquid and a magnetic system composed of spins at particular lattice sites,
which composes the so called Ising model, have exactly the same mean eld theory exponents, despite the completely
dierent nature of these two systems.
Another problem with the Van der Waals theory that is readily apparent is the fact that it predicts @P=@V > 0 for
certain values of the density . Such behavior is unstable. Possible ways of improving the approximations used are
the following:
1. Improve the approximation to g0 (r).
2. Choose a better zeroth order potential U0 (r1 ; :::; rN ).
3. Go to a higher order in perturbation theory.
Barker and Henderson have shown that going to second order in perturbation theory (see discussion in McQuarrie's
book, chapter 14), yields well converged results for a square well uid, compared to \exact" results from a molecular
dynamics simulation.

9
 G25.2651: Statistical Mechanics

Notes for Lecture 11

I. REACTION COORDINATES
Often simple chemical processes can be described in terms of a one-dimensional coordinate, which is not one of the
Cartesian coordinates in a system but a generalized coordinate, q. Such a coordinate is, in general, a function of the
Cartesian coordinates:
q = q(r1 ; :::; rN )
In some cases, a set of reaction coordinates is required.
Examples 1: Dissociation reactions. Consider a dissociation reaction of the form
AB ?! A + B
A reaction coordinate for such a process is the distance r between the two atoms in the diatomic. If r1 and r2 are the
Cartesian positions of atoms A and atom B , respectively, then
r = jr1 ? r2 j
What set of generalized coordinates contains r explicitly? Let us transform to center-of-mass and relative coordinate:
R = mAmr1 ++ m m B r2
A B
r = r 1 ? r2
Then, let r = (x; y; z ) be transformed to spherical polar coordinates
(x; y; z ) ?! (r; ; )
x = r sin  cos 
y = r sin  sin 
z = r cos 
Example 2: Proton transfer reactions. Consider proton transfer through a hydrogen bond according to:
A ? H


A * ) A


H ? A
which is illustrated schematically in the cartoon below:

H rp

d2
d1
s

O O
r1 r2
FIG. 1.

1
 The two heavy atoms are assumed to be of the same type (e.g. oxygen atoms). A reaction coordinate describing
this process is the dierence in the two distances d1 and d2 :
 = d1 ? d2 = jrp ? r1 j ? jrp ? r2 j
What generalized coordinate system contains  explicitly?
To see what coordinate system this is, consider transforming to center of mass and two relative coordinates as
follows:
R = mO(r21m+ r+2 )m+ mHrp
O H
r = r 1 ? r2
s = rp ? 21 (r1 + r2 )
Now the six degrees of freedom (rx ; ry ; rz ; sx; sy ; sz ) are transformed to 6 new degrees of freedom, which are the
spherical polar coordinates of r and the confocal elliptic coordinates for the position of the proton:
q
r = rx2 + ry2 + rz2
q
? 1
rx2 + ry2
 = tan rz
?
 = tan r1 ry
z
 = d1 + d2 = js + 21 rj + js + 21 rj
 = d1 + d2 = js + 12 rj ? js + 21 rj
 = tilt angle of plane containing three atoms
Then, the coordinate  is the reaction coordinate .

II. FREE ENERGY PROFILES

For a reaction coordinate q, we can dene a probability distribution function according to
P (q) = h(q(Zr1 ; :::; rN ) ? q)
= CQN dN pdN r e?H (p;r)(q(r1 ;:::;rN )?q)
Then, we dene the free energy prole, A(q), to be
A(q ) = ?kT ln P (q)
Apart from the normalization of P (q) by the partition function, Q, P (q ), itself is a kind of partition function corre-
sponding to a xed value of the reaction coordinate. Thus, dening the free energy prole in terms of the log of P (q)
is analogous to dening the global free energy A = ?kT ln Q.

A. Physical meaning of A(q)

Consider the free energy dierence between two values of the reaction coordinate q and q0 , which can be written as
Z q Z q
dA 1 dP
A(q) ? A(q0 ) = dq dq0 = ?kT dq0 P (
0
q0 q0 q0 ) dq0
The integrand can be written as

2
 R
1 dP = CQN dN pdN re?H (p;r) @@q (q(r1 ; :::; rN ) ? q0 )
0

P (q0 ) dq0 h(q ? q0 )i

Now, we introduce a coordinate transformation from Cartesian coordinates to a set of generalized coordinates that
contains q:
(r1 ; :::; rN ) ?! (u1 ; :::; un; q)
where n = 3N ? 1. In addition, a corresponding transformation is made on the conjugate momenta according to:
(p1 ; :::; pN ) ?! (pu1 ; :::; pun ; pq )
so that, in the measure, no overall Jacobian appears:
dN pdN r = dn pu dpq dn udq
Thus, we have
R
1 dP = dn pu dpq dn udqe?H (pu ;pq ;u;q) @@q (q ? q0 )
CN
Q 0

P (q0 ) dq0 h(q ? q0 )i

Next, the derivative with respect to q0 is changed to a derivative with respect to q:
R
1 dP = ? CQN dn pu dpq dn udqe?H (pu ;pq ;u;q) @q@ (q ? q0 )
P (q0 ) dq0 h(q ? q0 )i
Then, an integration by parts is performed, which yields
R h i
1 dP =
CN
Q dn pu dpq dn udq e?H (pu ;pq ;u;q) (q ? q0 )
@
@q
P (q0 ) dq0 h(q ? q0 )i
?CN R dn pu dpq dn udq @H e?H (pu ;pq ;u;q) (q ? q0 )
= Q @q
h(q ? q0 )i
h @H (q ? q0 )i
= ? @q h(q ? q0 )i
 ? h @H cond
@q iq 0

where the nal average is an ensemble average conditional upon the restriction that the reaction coordinate q is equal
to q0 . Generally,
h


icond = h(
h

()q(?q ? q0 )i
q 0
q )i
0
Thus, the free energy dierence becomes:
Z q
A(q) ? A(q0 ) = dq0 h @H cond
@q iq
0
q0

Note that the average in the above expression can also be performed with respect to Cartesian positions and momenta,
in which case the derivative can be carried out via the chain rule. If

H=
N
X p2i + U (r ; :::; r )
2m 1 N
i=1 i

then

3
  
N
@H = X @H @ pi @H @ ri
@q i=1 @ pi
@q + @ ri
@q
N  
=
X pi
@@qpi + @@Ur
@@qri
i=1
mi i
N 
=
X pi
@ pi ? F
@ ri 
i
i=1
mi @q @q
The quantity ?@H=@q is the generalized force on the generalized coordinate q. Thus, let the conditional average of
this force be Fq . Then,
0

Fq = ?h @H
0
cond
@q iq 0

and
Z q
A(q ) ? A(q0 ) = ? dq0 Fq 0

q0
Thus, the free energy dierence can be seen to be equal to the work performed against the averaged generalized force
in bringing the system from q0 to q, irrespective of the values of the other degrees of freedom. Such an integration is
called thermodynamic integration.
Another useful expression for the free energy derivative can be obtained by integrating out the momenta before
performing a transformation. We begin with
R
1 dP = CQN dN pdN re?H (p;r) @@q (q(r1 ; :::; rN ) ? q0 )
0

P (q0 ) dq0 h(q ? q0 )i

Now, noting that the -function condition is independent of momenta, we can integrate out the Cartesian momenta
to yield:
R
1 dP = QCN3N dN re?U (r1 ;:::;rN ) @@q (q(r1 ; :::; rN ) ? q0 )
0

P (q0 ) dq0 h(q ? q0 )i

Next, the coordinate transformation to generalized coordinates is performed:
(r1 ; :::; rN ) ?! (u1 ; :::; un; q)
associated with which there is a Jacobian given by
J (u1 ; :::; un ; q) = @@(u(r1; ;:::;
:::; rN )
1 un ; q )
Introducing the coordinate transformation, we obtain
R
1 dP = QCN3N dn udqJ (u1 ; :::; un ; q)e?U @@q (q ? q0 ) 0

P (q0 ) dq0
R
h(q ? q0 )i
CN
3N dn udqJ (u1 ; :::; un ; q)e?U @q@ (q ? q0 )
= ? Q h(q ? q0 )i
R h i
?U (q ? q0 )
Q3N d udq @q J (u1 ; :::; un ; q )e
CN n @
= h(q ? q0 )i i
R h
@ ? (U ?kT ln J )  (q ? q0 )
Q3N d udq @q e
CN n
= h(q ? q0 )i 
R 
?C3NN dn udq @U ? kT @ ln J e?(U ?kT ln J ) (q ? q0 )
Q @q @q
= h(q ? q0 )i
4
  
?C3NN R dn udqJ (u1 ; :::; un; q) @U ? kT @ @qln J e?U (q ? q0 )
Q @q
= h(q ? q0 )i
 
= ? h @U @ ln J cond
@q ? kT @q iq 0

or
 
dA @U @ ln J cond
dq0 = h @q ? kT @q iq 0

III. EXAMPLE OF A DISSOCIATION REACTION

Let us again look at the dissociation reaction
AB ?! A + B
for which the reaction coordinate r = jr1 ? r2 j is appropriate. The transformation to center of mass and relative
coordinate is:
R = mAr1 + mB r2 M
r = r 1 ? r2
The inverse of this transformation is
r1 = R + mM2 r
r2 = R ? mM1 r
Next, the vector r = (x; y; z ) is transformed to spherical polar coordinates according to
x = r sin  cos 
y = r sin  sin 
z = r cos 
and the derivative of the potential with respect to r can be easily worked out using the chain rule:
@U = @U
@ r
@r @ r @r
@U = @U m2 ? @U m1
@ r @ r1 M @ r 2 M
= ? M1 [m2 F1 ? m1 F2 ]
@ r =  x ; y ; z   n^  (sin  cos ; sin  sin ; cos )
@r r r r
@U = ? 1 [m F ? m F ]
n^
@r M 2 1 1 2
The Jacobian is clearly
J = r2 sin 
so that
@ ln J = 2
@r r
Thus, the free energy derivative is

5
 dA 1 2kT cond
dr0 = h? M [m2 F1 ? m1 F2 ]
n^ ? jr1 ? r2 j ir 0

which is expressed completely in terms of Cartesian quantities and can be calculated straightforwardly.
Once dA=dr0 is known, it can be integrated to obtain the full free energy prole A(r0 ). This is another use of
thermodynamic integration.
It is always interesting to see what A(r) looks like compared to the bare potential. Suppose the dissociation reaction
is governed by a pair potential describing the interaction of the dissociating molecule with a solvent:
X X
U (r1 ; :::; rN ) = u0 (jr1 ? r2 j) + [~u(jr1 ? rj j) + u~(jr2 ? rj j)] + v(jri ? rj j)
j 6=1;2 i;j;i6=j;i;j 6=1;2

The potential u0 (r) might look like:

u 0(r)

r
FIG. 2.

If, at a given temperature T , the solvent assists in the dissociation process, then we might expect A(r) to have a
lower dissociation threshold and perhaps a slightly longer eective minimum bond length:

6
 A(r)

u 0(r)

r
FIG. 3.

If, on the other hand, at a given temperature, T , the solvent hinders dissociation, by causing the molecule to bury
itself in a cavity, for example, then we might expect A(r) to appear as follows:

7
A(r)

u 0(r)

FIG. 4.

with a higher dissociation threshold energy and slightly shorter eective minimum bond length. Such curves will,
of course, be temperature dependent and depend on the specic nature of the interactions with the solvent.

8
 G25.2651: Statistical Mechanics

Notes for Lecture 12

I. THE FUNDAMENTAL POSTULATES OF QUANTUM MECHANICS

The fundamental postulates of quantum mechanics concern the following questions:
1. How is the physical state of a system described?
2. How are physical observables represented?
3. What are the results of measurements on quantum mechanical systems?
4. How does the physical state of a system evolve in time?
5. The uncertainty principle.

A. The physical state of a quantum system

The physical state of a quantum system is represented by a vector denoted
j (t)i
which is a column vector, whose components are probability amplitudes for dierent states in which the system might
be found if a measurement were made on it.
A probability amplitude  is a complex number, the square modulus of which gives the corresponding probability
P
P = jj2
The number of components of j (t)i is equal to the number of possible states in which the system might observed.
The space that contains j (t)i is called a Hilbert space H. The dimension of H is also equal to the number of states
in which the system might be observed. It could be nite or innite (countable or not).
j (t)i must be a unit vector. This means that the inner product:
h (t)j (t)i = 1
In the above, if the vector j (t)i, known as a Dirac \ket" vector, is given by the column
0 1
1
B C
j (t)i = B
C
2
B@ C

A


then the vector h (t)j, known as a Dirac \bra" vector, is given by
h (t)j = ( 1 2


)
so that the inner product becomes
X
h (t)j (t)i = j i j2 = 1
i
We can understand the meaning of this by noting that i , the components of the state vector, are probability
amplitudes, and j i j2 are the corresponding probabilities. The above condition then implies that the sum of all the
probabilities of being in the various possible states is 1, which we know must be true for probabilities.

1
 B. Physical Observables
Physical observables are represented by linear, hermitian operators that act on the vectors of the Hilbert space. If
A is such an operator, and ji is an arbitrary vector in the Hilbert space, then A might act on ji to produce a vector
j0 i, which we express as
Aji = j0 i
Since ji is representable as a column vector, A is representable as a matrix with components
A11 A12 A13


!
A = A21 A22 A23










The condition that A must be hermitian means that
Ay = A
or
Aij = Aji

C. Measurement
The result of a measurement of the observable A must yield one of the eigenvalues of A. Thus, we see why A is
required to be a hermitian operator: Hermitian operators have real eigenvalues. If we denote the set of eigenvalues of
A by fai g, then each of the eigenvalues ai satises an eigenvalue equation
Ajai i = ai jai i
where jai i is the corresponding eigenvector. Since the operator A is hermitian and ai is therefore real, we have also
the left eigenvalue equation
hai jA = hai jai
The probability amplitude that a measurement of A will yield the eigenvalue ai is obtained by taking the inner
product of the corresponding eigenvector jai i with the state vector j (t)i, hai j (t)i. Thus, the probability that the
value ai is obtained is given by
Pa = jhai j (t)ij2
i

Another useful and important property of hermitian operators is that their eigenvectors form a complete orthonormal
basis of the Hilbert space, when the eigenvalue spectrum is non-degenerate. That is, they are linearly independent,
span the space, satisfy the orthonormality condition
hai jaj i = ij
and thus any arbitrary vector ji can be expanded as a linear combination of these vectors:
X
ji = ci jai i
i
By multiplying both sides of this equation by haj j and using the orthonormality condition, it can be seen that the
expansion coecients are
ci = hai ji
The eigenvectors also satisfy a closure relation:

2
 X
I= jai ihai j
i
where I is the identity operator.
Averaging over many individual measurements of A gives rise to an average value or expectation value for the
observable A, which we denote hAi and is given by
hAi = h (t)jAj (t)i
That this is true can be seen by expanding the state vector j (t)i in the eigenvectors of A:
X
j (t)i = i (t)jai i
i
where i are the amplitudes for obtaining the eigenvalue ai upon measuring A, i.e., i = hai j (t)i. Introducing this
expansion into the expectation value expression gives
X
hAi(t) = i (t)j (t)hai jAjai i
i;j
X
= i (t)j ai (t)ij
i;j
X
= ai ji (t)j2
i
The interpretation of the above result is that the expectation value of A is the sum over possible outcomes of
a measurement of A weighted by the probability that each result is obtained. Since ji j2 = jhai j (t)ij2 is this
probability, the equivalence of the expressions can be seen.
Two observables are said to be compatible if AB = BA. If this is true, then the observables can be diagonalized
simultaneously to yield the same set of eigenvectors. To see this, consider the action of BA on an eigenvector jai i of
A. BAjai i = ai B jai i. But if this must equal AB jai i, then the only way this can be true is if B jai i yields a vector
proportional to jai i which means it must also be an eigenvector of B . The condition AB = BA can be expressed as
AB ? BA = 0
[A; B ] = 0
where, in the second line, the quantity [A; B ]  AB ? BA is know as the commutator between A and B . If [A; B ] = 0,
then A and B are said to commute with each other. That they can be simultaneously diagonalized implies that one
can simultaneously predict the observables A and B with the same measurement.
As we have seen, classical observables are functions of position x and momentum p (for a one-particle system).
Quantum analogs of classical observables are, therefore, functions of the operators X and P corresponding to position
and momentum. Like other observables X and P are linear hermitian operators. The corresponding eigenvalues x
and p and eigenvectors jxi and jpi satisfy the equations
X jxi = xjxi
P jpi = pjpi
which, in general, could constitute a continuous spectrum of eigenvalues and eigenvectors. The operators X and P
are not compatible. In accordance with the Heisenberg uncertainty principle (to be discussed below), the commutator
between X and P is given by
[X; P ] = ihI
and that the inner product between eigenvectors of X and P is
hxjpi = p 1 eipx=h
2h
Since, in general, the eigenvalues and eigenvectors of X and P form a continuous spectrum, we write the orthonormality
and closure relations for the eigenvectors as:

3
 hxjx0 i = Z(x ? x0 ) hpjp0 i = Z(p ? p0 )
ji = dxjxihxji ji = dpjpihpji
Z Z
I = dxjxihxj I = dpjpihpj
The probability that a measurement of the operator X will yield an eigenvalue x in a region dx about some point is
P (x; t)dx = jhxj (t)ij2 dx
The object hxj (t)i is best represented by a continuous function (x; t) often referred to as the wave function. It is a
representation of the inner product between eigenvectors of X with the state vector. To determine the action of the
operator X on the state vector in the basis set of the operator X , we compute
hxjX j (t)i = x (x; t)
The action of P on the state vector in the basis of the X operator is consequential of the incompatibility of X and P
and is given by
hxjP j (t)i = hi @x
@ (x; t)

Thus, in general, for any observable A(X; P ), its action on the state vector represented in the basis of X is
 h @ 
hxjA(X; P )j (t)i = A x; i @x (x; t)

D. Time evolution of the state vector

The time evolution of the state vector is prescribed by the Schrodinger equation
@ j (t)i = H j (t)i
ih @t
where H is the Hamiltonian operator. This equation can be solved, in principle, yielding
j (t)i = e?iHt=h j (0)i
where j (0)i is the initial state vector. The operator
U (t) = e?iHth
is the time evolution operator or quantum propagator. Let us introduce the eigenvalues and eigenvectors of the
Hamiltonian H that satisfy
H jEi i = Ei jEi i
The eigenvectors for an orthonormal basis on the Hilbert space and therefore, the state vector can be expanded in
them according to
X
j (t)i = ci (t)jEi i
i
where, of course, ci (t) = hEi j (t)i, which is the amplitude for obtaining the value Ei at time t if a measurement of
H is performed. Using this expansion, it is straightforward to show that the time evolution of the state vector can be
written as an expansion:

4
 j (t)i = e?iHth j (0)
Xi
=e ?iHt= h

jEi ihEi j (0)i
X i
= e?iEi t=h jEi ihEi j (0)i
i
Thus, we need to compute all the initial amplitudes for obtaining the dierent eigenvalues Ei of H , apply to each the
factor exp(?iEi t=h)jEi i and then sum over all the eigenstates to obtain the state vector at time t.
If the Hamiltonian is obtained from a classical Hamiltonian H (x; p), then, using the formula from the previous
section for the action of an arbitrary operator A(X; P ) on the state vector in the coordinate basis, we can recast the
Schrodiner equation as a partial dierential equation. By multiplying both sides of the Schrodinger equation by hxj,
we obtain
hxjH (X; P )j (t)i = ih @t@ hxj (t)i
 h @  @ (x; t)
H x; i @x (x; t) = ih @t
If the classical Hamiltonian takes the form
H (x; p) = 2pm + U (x)
2

then the Schrodinger equation becomes

 h2 @ 2 
? 2m @x2 + U (x) (x; t) = ih @t@ (x; t)
which is known as the Schrodinger wave equation or the time-dependent Schrodinger equation.
In a similar manner, the eigenvalue equation for H can be expressed as a dierential equation by projecting it into
the X basis:

 h @hxjH jEi i = Ei hxjEi i

H x; i @x i (x) = Ei i (x)
 h2 @ 2 
? 2m @x2 + U (x) i (x) = Ei i (x)

where i (x) = hxjEi i is an eigenfunction of the Hamiltonian.

E. The Heisenberg uncertainty principle

Because the operators X and P are not compatible, [X; P ] 6= 0, there is no measurement that can precisely
determine both X and P simultaneously. Hence, there must be an uncertainty relation between them that species
how uncertain we are about one quantity given a denite precision in the measurement of the other. Presumably, if
one can be determined with innite precision, then there will be an innite uncertainty in the other. Recall that we
had dened the uncertainty in a quantity by
p

A = hA2 i ? hAi2
Thus, for X and P , we have
p

x = hX 2 i ? hX i2
p

p = hP 2 i ? hP i2
These quantities can be expressed explicitly in terms of the wave function (x; t) using the fact that

5
 Z Z
hX i = h (t)jX j (t)i = dxh (t)jxihxjX j (t)i = dx  (x; t)x (x; t)
and
Z
hX i = h (t)jX j (t)i = (x; t)x2 (x; t)
2 2

Similarly,
Z Z
hP i = h (t)jP j (t)i = dxh (t)jxihxjP j (t)i = dx  (x; t) hi @x
@ (x; t)

and
Z  
@ 2 (x; t)
hP i = h (t)jP j (t)i =
2 2
dx  (x; t) ?h @x
2
2

Then, the Heisenberg uncertainty principle states that

>

x
p  h
which essentially states that the greater certainty with which a measurement of X or P can be made, the greater will
be the uncertainty in the other.

F. The Heisenberg picture

In all of the above, notice that we have formulated the postulates of quantum mechanics such that the state vector
j (t)i evolves in time but the operators corresponding to observables are taken to be stationary. This formulation of
quantum mechanics is known as the Schrodinger picture. However, there is another, completely equivalent, picture in
which the state vector remains stationary and the operators evolve in time. This picture is known as the Heisenberg
picture. This particular picture will prove particularly useful to us when we consider quantum time correlation
functions.
The Heisenberg picture species an evolution equation for any operator A, known as the Heisenberg equation. It
states that the time evolution of A is given by
dA = 1 [A; H ]
dt ih
While this evolution equation must be regarded as a postulate, it has a very immediate connection to classical
mechanics. Recall that any function of the phase space variables A(x; p) evolves according to
dA = fA; H g
dt
where f:::; :::g is the Poisson bracket. The suggestion is that in the classical limit (h small), the commutator goes over
to the Poisson bracket. The Heisenberg equation can be solved in principle giving
A(t) = eiHt=h Ae?iHt=h
= U y (t)AU (t)
where A is the corresponding operator in the Schrodinger picture. Thus, the expectation value of A at any time t is
computed from
hA(t)i = h jA(t)j i
where j i is the stationary state vector.
Let's look at the Heisenberg equations for the operators X and P . If H is given by
H = 2Pm + U (X )
2

6
then Heisenberg's equations for X and P are
dX = 1 [X; H ] = P
dt ih m
dP = 1 [P; H ] = ? @U
dt ih @X
Thus, Heisenberg's equations for the operators X and P are just Hamilton's equations cast in operator form. Despite
their innocent appearance, the solution of such equations, even for a one-particle system, is highly nontrivial and has
been the subject of a considerable amount of research in physics and mathematics.
Note that any operator that satises [A(t); H ] = 0 will not evolve in time. Such operators are known as constants of
the motion. The Heisenberg picture shows explicitly that such operators do not evolve in time. However, there is an
analog with the Schrodinger picture: Operators that commute with the Hamiltonian will have associated probabilities
for obtaining dierent eigenvalues that do not evolve in time. For example, consider the Hamiltonian, itself, which it
trivially a constant of the motion. According to the evolution equation of the state vector in the Schrodinger picture,
X
j (t)i = e?iE t=h jEi ihEi j (0)i
i

i
the amplitude for obtaining an energy eigenvalue Ej at time t upon measuring H will be
X
hEj j (t)i = e?iE t=h hEj jEi ihEi j (0)i
i

Xi
= e?iE t=h ij hEi j (0)i
i

i
= e?iE t=h hEj j (0)i
j

Thus, the squared modulus of both sides yields the probability for obtaining Ej , which is
jhEj j (t)ij2 = jhEj j (0)ij2
Thus, the probabilities do not evolve in time. Since any operator that commutes with H can be diagonalized simul-
taneously with H and will have the same set of eigenvectors, the above arguments will hold for any such operator.

7
 G25.2651: Statistical Mechanics

Notes for Lecture 13

I. PRINCIPLES OF QUANTUM STATISTICAL MECHANICS

The problem of quantum statistical mechanics is the quantum mechanical treatment of an N -particle system.
Suppose the corresponding N -particle classical system has Cartesian coordinates
q1 ; :::; q3N
and momenta
p1 ; :::; p3N
and Hamiltonian
3N
H=
X p2i + U (q ; :::; q )
3N
i=1 2mi
1

Then, as we have seen, the quantum mechanical problem consists of determining the state vector j (t)i from the
Schrodinger equation
@ j (t)i
H j (t)i = ih @t
Denoting the corresponding operators, Q1 ; :::; Q3N and P1 ; :::; P3N , we note that these operators satisfy the commu-
tation relations:
[Qi ; Qj ] = [Pi ; Pj ] = 0
[Qi ; Pj ] = ihIij
and the many-particle coordinate eigenstate jq1 :::q3N i is a tensor product of the individual eigenstate jq1 i; :::; jq3N i:
jq1 :::q3N i = jq1 i


jq3N i
The Schrodinger equation can be cast as a partial dierential equation by multiplying both sides by hq1 :::q3N j:
hq1 :::q3N jH j (t)i = ih @t@ hq1 :::q3N j (t)i
" #
3N
@ 2 + U (q ; :::; q ) (q ; :::; q ; t) = ih @ (q ; :::; q ; t)
? 2hm @q
X2
3N 3N 3N
i=1 i i
2 1 1
@t 1
where the many-particle wave function is (q1 ; ::::; q3N ; t) = hq1 :::q3N j (t)i. Similarly, the expectation value of an
operator A = A(Q1 ; :::; Q3N ; P1 ; :::; P3N ) is given by
Z  
hAi = dq1


dq3N  (q1 ; :::; q3N )A q1 ; :::; q3N ; hi @q@ ; :::; hi @q@ (q1 ; :::; q3N )
1 3N

1
 A. The density matrix and density operator
In general, the many-body wave function (q1 ; :::; q3N ; t) is far too large to calculate for a macroscopic system. If
we wish to represent it on a grid with just 10 points along each coordinate direction, then for N = 1023, we would
need
23
1010
total points, which is clearly enormous.
We wish, therefore, to use the concept of ensembles in order to express expectation values of observables hAi without
requiring direct computation of the wavefunction. Let us, therefore, introduce an ensemble of systems, with a total of
Z members, and each having a state vector j ()i,  = 1; :::; Z . Furthermore, introduce an orthonormal set of vectors
jk i (hk jj i = ij ) and expand the state vector for each member of the ensemble in this orthonormal set:
X
j () i = Ck() jk i
k
The expectation value of an observable, averaged over the ensemble of systems is given by the average of the expectation
value of the observable computed with respect to each member of the ensemble:
Z
hAi = Z1
X
h () jAj () i
=1
Substituting in the expansion for j () i, we obtain
hAi = Z1 Ck() Cl() hk jAjl i
X 

k;l
!
X Z
1X () ()
= Z Cl Ck hk jAjl i
k;l =1
Let us dene a matrix
Z
X
Cl() Ck()

lk =
=1
and a similar matrix
Z
~lk = Z1 Cl() Ck()
X 

=1
Thus, lk is a sum over the ensemble members of a product of expansion coecients, while ~lk is an average over the
ensemble of this product. Also, let Akl = hk jAjl i. Then, the expectation value can be written as follows:
hAi = 1  A = 1 (A) = 1 Tr(A) = Tr(~A)
X X
Z k;l lk kl Z k kk Z

where  and A represent the matrices with elements lk and Akl in the basis of vectors fjk ig. The matrix lk is
known as the density matrix. There is an abstract operator corresponding to this matrix that is basis-independent.
It can be seen that the operator
Z
X
= j () ih () j
=1
and similarly
Z
~ = Z1 j () ih () j
X

=1

2
have matrix elements lk when evaluated in the basis set of vectors fjk ig.
Z
X Z
X
hl j () ih () jk i = Cl() Ck() = lk

hl jjk i =
=1 =1
Note that  is a hermitian operator
y = 
so that its eigenvectors form a complete orthonormal set of vectors that span the Hilbert space. If wk and jwk i
represent the eigenvalues and eigenvectors of the operator ~, respectively, then several important properties they must
satisfy can be deduced.
Firstly, let A be the identity operator I . Then, since hI i = 1, it follows that
1 = 1 Tr() = Tr(~) = w
X
Z k
k
Thus, the eigenvalues of ~ must sum to 1. Next, let A be a projector onto an eigenstate of ~, A = jwk ihwk j  Pk .
Then
hPk i = Tr(~jwk ihwk j)
But, since ~ can be expressed as
X
~ = wk jwk ihwk j
k
and the trace, being basis set independent, can be therefore be evaluated in the basis of eigenvectors of ~, the
expectation value becomes
X X
hPk i = hwj j wi jwi ihwi jwk ihwk jwj i
j i
X
= wi ij ik kj
i;j
= wk
However,
Z
hPk i = Z1
X
h () jwk ihwk j () i
=1
1XZ
=Z jh () jwk ij2  0
=1
P
Thus, wk  0. Combining these two results, we see that, since k wk = 1 and wk  0, 0  wk  1, so that wk satisfy
the properties of probabilities.
With this in mind, we can develop a physical meaning for the density matrix. Let us now consider the expectation
value of a projector jai ihai j  Pai onto one of the eigenstates of the operator A. The expectation value of this operator
is given by
Z Z Z
hPai i = Z1 h () jPai j () i = Z1 h () jai ihai j ()i = Z1
X X X
jhai j ()ij2
=1 =1 =1
But jhai j () ij2  Pa(i) is just probability that a measurement of the operator A in the th member of the ensemble
will yield the result ai . Thus,
P
hPai i = Z1
X
Pa(i)
=1

3
or the expectation value of Pai is just the ensemble averaged probability of obtaining the value ai in each member of
the ensemble. However, note that the expectation value of Pai can also be written as
X X
hPai i = Tr(~Pai ) = Tr( wk jwk ihwk jai ihai j) = hwl jwk jwk ihwk jai ihai jwl i
k k;l
X
= wk kl hwk ai ihai wl i
k;l
X
= wk jhai jwk ij2
k
Equating the two expressions gives
1XZ X
()
Z hP ai i = wk jhai jwk ij2
=1 k
The interpretation of this equation is that the ensemble averaged probability of obtaining the value ai if A is measured
is equal to the probability of obtaining the value ai in a measurement of A if the state of the system under consideration
were the state jwk i, weighted by the average probability wk that the system in the ensemble is in that state. Therefore,
the density operator  (or ~) plays the same role in quantum systems that the phase space distribution function f (?)
plays in classical systems.

B. Time evolution of the density operator

The time evolution of the operator  can be predicted directly from the Schrodinger equation. Since (t) is given
by
Z
X
(t) = j () (t)ih () (t)j
=1
the time derivative is given by
Z
@ = X
  
@ j () (t)i h () (t)j + j ()(t)i @ h () (t)j


@t =1 @t @t
Z h
= i1h
X   i
H j () (t)i h () (t)j ? j () (t)h h () (t)jH
=1
= i1h (H ? H )
= i1h [H; ]
@ = 1 [H; ]
@t ih
where the second line follows from the fact that the Schrodinger equation for the bra state vector h () (t)j is
?ih @t@ h () (t)j = h () (t)jH
Note that the equation of motion for (t) diers from the usual Heisenberg equation by a minus sign! Since (t) is
constructed from state vectors, it is not an observable like other hermitian operators, so there is no reason to expect
that its time evolution will be the same. The general solution to its equation of motion is
(t) = e?iHt=h (0)eiHt=h
= U (t)(0)U y (t)
The equation of motion for (t) can be cast into a quantum Liouville equation by introducing an operator

4
 iL = i1h [:::; H ]
In term of iL, it can be seen that (t) satises
@ = ?iL
@t
(t) = e?iLt (0)
What kind of operator is iL? It acts on an operator and returns another operator. Thus, it is not an operator in the
ordinary sense, but is known as a superoperator or tetradic operator (see S. Mukamel, Principles of Nonlinear Optical
Spectroscopy, Oxford University Press, New York (1995)).
Dening the evolution equation for  this way, we have a perfect analogy between the density matrix and the state
vector. The two equations of motion are
@ i
@t j (t)i = ? h H j (t)i
@
@t (t) = ?iL(t)
We also have an analogy with the evolution of the classical phase space distribution f (?; t), which satises
@f = ?iLf
@t
with iL = f:::; H g being the classical Liouville operator. Again, we see that the limit of a commutator is the classical
Poisson bracket.

C. The quantum equilibrium ensembles

At equilibrium, the density operator does not evolve in time; thus, @=@t = 0. Thus, from the equation of motion, if
this holds, then [H; ] = 0, and (t) is a constant of the motion. This means that it can be simultaneously diagonalized
with the Hamiltonian and can be expressed as a pure function of the Hamiltonian
 = f (H )
Therefore, the eigenstates of , the vectors, we called jwk i are the eigenvectors jEi i of the Hamiltonian, and we can
write H and  as
X
H= Ei jEi ihEi j
i
X
= f (Ei )jEi ihEi j
i
The choice of the function f determines the ensemble.

1. The microcanonical ensemble

Although we will have practically no occasion to use the quantum microcanonical ensemble (we relied on it more
heavily in classical statistical mechanics), for completeness, we dene it here. The function f , for this ensemble, is
f (Ei )E = (Ei ? (E + E )) ? (Ei ? E )
where (x) is the Heaviside step function. This says that f (Ei )E is 1 if E < Ei < (E + E ) and 0 otherwise. The
partition function for the ensemble is Tr(), since the trace of  is the number of members in the ensemble:
X

(N; V; E ) = Tr() = [(Ei ? (E + E )) ? (Ei ? E )]
i

5
The thermodynamics that are derived from this partition function are exactly the same as they are in the classical
case:
S (N; V; E ) = ?k ln


(N; V;

E)
1 = ?k @ ln

T @E N;V
etc.

2. The canonical ensemble

In analogy to the classical canonical ensemble, the quantum canonical ensemble is dened by
 = e?H
f (Ei ) = e?Ei
Thus, the quantum canonical partition function is given by
X
Q(N; V; T ) = Tr(e?H ) = e?Ei
i
and the thermodynamics derived from it are the same as in the classical case:
A(N; V; T ) = ? 1 ln Q(N; V; T )

E (N; V; T ) = ? @@ ln Q(N; V; T )
@ ln Q(N; V; T )
P (N; V; T ) = 1 @V
etc. Note that the expectation value of an observable A is
hAi = 1 Tr(Ae?H )
Q
Evaluating the trace in the basis of eigenvectors of H (and of ), we obtain
hAi = 1 hE jAe?H jE i = 1 e?Ei hE jAjE i
X X
Q i i i Q i i i

The quantum canonical ensemble will be particularly useful to us in many things to come.

3. Isothermal-isobaric and grand canonical ensembles

Also useful are the isothermal-isobaric and grand canonical ensembles, which are dened just as they are for the
classical cases:
Z 1 Z 1

(N; P; T ) = dV e?PV Q(N; V; T ) = dV Tr(e?(H +PV ) )
0 0
1
X 1
X
Z (; V; T ) = eN Q(N; V; T ) = Tr(e?(H ?N ) )
N =0 N =0

6
 D. A simple example { the quantum harmonic oscillator
As a simple example of the trace procedure, let us consider the quantum harmonic oscillator. The Hamiltonian is
given by
H = 2Pm + 12 m!2 X 2
2

and the eigenvalues of H are

 
En = n + 21 h!; n = 0; 1; 2; :::
Thus, the canonical partition function is
1
X 1 ?
X
e?h ! n


Q( ) = e?(n+1=2)h! = e?h !=2
n=0 n=0
This is a geometric series, which can be summed analytically, giving
?h !=2
Q( ) = 1e? e?h ! = eh !=2 ?1e?h !=2 = 21 csch( h!=2)
The thermodynamics derived from it as as follows:
1. Free energy:
The free energy is
A = ? 1 ln Q( ) = h2! + 1 ln 1 ? e?h !
?

2. Average energy:
The average energy E = hH i is
 
@ ln Q( ) = h! + h!e?h! = 1 + hni h!
E = ? @ 2 1 ? e?h ! 2
3. Entropy
The entropy is given by
?h !
S = k ln Q( ) + ET = ?k ln 1 ? e?h ! + hT! 1 ?e e?h !
?

Now consider the classical expressions. Recall that the partition function is given by
 
Z p2 1  1=2  1=2
Q( ) = h1 dpdxe
? 2m + 2 m!2 x2 1
=h  2m 2 = 2 = 1
m! 2 !h  h!
Thus, the classical free energy is
Acl = 1 ln( h!)
In the classical limit, we may take h to be small. Thus, the quantum expression for A becomes, approximately, in
this limit:
AQ ?! h2! + 1 ln( h!)
and we see that
AQ ? Acl ?! h2!
The residual h !=2 (which truly vanishes when h ! 0) is known as the quantum zero point energy. It is a pure
quantum eect and is present because the lowest energy quantum mechanically is not E = 0 but the ground state
energy E = h !=2.

7
 G25.2651: Statistical Mechanics

Notes for Lecture 14

I. DERIVATION OF THE DISCRETIZED PATH INTEGRAL

We begin our discussion of the Feynman path integral with the canonical ensemble. The expressions for the partition
function and expectation value of an observable A are, respectively
Q(N; V; T ) = Tr(e?H )
hAi = Q1 Tr(Ae?H )
It is clear that we need to be able to evaluate traces of the type appearing in these expressions. We have already
derived expressions for these in the basis of eigenvectors of H . However, since the trace is basis independent, let
us explore carrying out these traces in the coordinate basis. We will begin with the partition function and treat
expectation values later.
Consider the ensemble of a one-particle system. The partition function evaluated as a trace in the coordinate basis
is
Z

Q( ) = dxhxje?H jxi
We see that the trace involves the diagonal density matrix element hxje?H jxi. Let us solve the more general problem
of any density matrix element hxje?H jx0 i.
If the Hamiltonian takes the form
H = 2Pm + U (X )  K + U
2

then we cannot evaluate the operator exp(?H ) explicitly because the operators for kinetic (T ) and potential energies
(U ) do not commute with each other, being, respectively, functions of momentum and position, i.e.,
[K; U ] 6= 0
In this instance, we will make use of the Trotter theorem, which states that given two operators A and B , such that
[A; B ] 6= 0, then for any number ,
B=2P eA=P eB=2P P
h i
e(A+B) = Plim
!1
e
Thus, for the Boltzmann operator,
?U=2P e?K=P e?U=2P P
h i
e?(K +U ) = Plim
!1
e
and the partition function becomes
?U=2P e?K=P e?U=2P P jxi
Z h i
Q( ) = Plim
!1
dxhxj e
Dene the operator in brackets to be
:

= e?U=2P e?K=P e?U=2P
Then,
Z

Q( ) = Plim
!1
dxhxj
P jxi

1
In between each of the P factors of
, the coordinate space identity operator
Z

I= dxjxihxj
is inserted. Since there are P factors, there will be P ? 1 such insertions. the integration variables will be labeled
x2 ; :::; xP . Thus, the expression for the matrix element becomes
Z

hxj
jx0 i = dx2


dxP hxj
jx2 ihx2 j
jx3 ihx3 j


jxP ihxP j
jx0 i
Z P
Y
= dx2


dxP hxi j
jxi+1 ijx1 =x;xP +1 =x 0

i=1
The next step clearly involves evaluating the matrix elementx
hxi j
jxi+1 i = hxi je?U (X )=2P e?P 2 =2mP e?U (X )=2P jxi+1 i
Note that in the above expression, the operators involving the potential U (X ) act on their eigenvectors and can thus
be replaced by the corresponding eigenvalues:
hxi j
jxi+1 i = e?(U (xi )+U (xi+1 )=2 hxi je?P 2 =2mP jxi+1 i
In order to evaluate the remaining matrix element, we introduce the momentum space identity operator
Z

I= dpjpihpj
Letting K = P 2 =2m, the matrix remaining matrix element becomes
Z

hxi je?K=P jxi+1 i = dphxi jpihpje?P 2 =2mP jxi+1 i

Z

= dphxi jpihpjxi+1 ie?p2 =2mP

Using the fact that
hxjpi = p 1 eipx=h
2h
it follows that
Z

hxi je?K=P jxi+1 i = 21h dpeip(xi ?xi+1 )=h e?p2 =2mP

The remaining integral over p can be performed by completing the square, leading to the result
mP 1=2 exp ? mP (x ? x )2
   

hxi je?K=P jxi+1 i =

2 h2 2 h2 i+1 i
Collecting the pieces together, and introducing the P ! 1 limit, we have for the density matrix
mP
  P=2 Z "
P 
mP (x ? x )2 +  (U (x ) + U (x ))
#

hxje?H jx0 i = Plim
X
dx2


dxP exp ? 
!1 2 h2 2 h2 i+1 i 2P i i+1 

i=1 x1 =x;xP +1 =x 0

The partition function is obtained by setting x = x0 , which is equivalent to setting x1 = xP +1 and integrating over x,
or equivalently x1 . Thus, the expression for Q( ) becomes
mP P=2 Z dx


dx exp ? X
 P 1 "   #

Q( ) = Plim m! 2 (x ? x )2 + 1 U (x ) 
1 P P i+1 i P i
!1 2 h2 i=1 2 
xP +1 =x1

2
where we have introduced a \frequency"
p
!P =  Ph
When expressed in this way, the partition function, for a nite value of P , is isomorphic to a classical conguration
integral for a P -particle system, that is a cyclic chain of particles, with harmonic nearest neighbor interactions and
interacting with an external potential U (x)=P . That is, the partition function becomes
Z

Q( )  dx1


dxP e?Ueff (x1 ;:::;xP )

where
P
X

1 m!2 (x ? x )2 + 1 U (x )


Ue (x1 ; :::; xP ) = 2 P i+1 i P i

i=1
Thus, for nite (if large) P the partition function in the discretized path integral representation can be treated as any
ordinary classical conguration integral. Consider the integrand of Q( ) in the limit that all P points on the cyclic
chain are at the same locationPP
x. Then the harmonic nearest neighbor coupling (which is due to the quantum kinetic
energy) vanishes and (1=P ) i=1 U (xi ) ! U (x), and the integrand becomes
e?U (x)
which is just the true classical canonical position space distribution function. Therefore, the greater the spatial spread
in the cyclic chain, the more \quantum" the system is, since this indicates a greater contribution from the quantum
kinetic energy. The spatially localized it is, the more the system behaves like a classical system.
It remains formally to take the limit that P ! 1. There we will see an elegant formulation for the density matrix
and partition function emerges.

3
 G25.2651: Statistical Mechanics

Notes for Lecture 15

I. THE FUNCTIONAL INTEGRAL REPRESENTATION OF THE PATH INTEGRAL

A. The continuous limit
In taking the limit P ! 1, it will prove useful to dene a parameter
" =  h P
so that P ! 1 implies " ! 0. In terms of ", the partition function becomes
 m P=2 Z " X  xi+1 ? xi 2 !#
+ U (xi ) 
P
Q( ) = P !1
lim;"!0 2"h "
dx1


dxP exp ? h m
i=1
2 " xP +1 =x1

We can think of the points x1 ; :::; xP as specic points of a continuous functions x( ), where
xk = x( = (k ? 1)")
such that x(0) = x( = P") = x( =  h ):
x

x4
x2

x P+1
x1

τ
0 2 ε 4ε Pε
FIG. 1.

1
 Note that
  
xk+1 ? xk = lim x(k") ? x((k ? 1)") = dx

lim
"!0 " "!0 " d
and that the limit
"  2 #
lim "X P
m xi+1 ? xi + U (xi )
h i=1 2
P !1;"!0  "
is just a Riemann sum representation of the continuous integral
"  2 #
1 j Z h d m dx + U (x( ))
h 0 2 d
Finally, the measure
 m P=2
lim
P !1;"!0 2"
dx1


dxP
h2
represents an integral overa all values that the function x( ) can take on between  = 0 and  =  h such that
x(0) = x( h). We write this symbolically as Dx(
). Therefore, the P ! 1 limit of the partition function can be
written as
Z " Z h 
Z x(h)=x #
Q( ) = dx 1
Dx(
) exp ? h m
d 2 x_ + U (x( ))
2

I " Z h 
x(0)=x 0
#
= 1
Dx(
) exp ? hm 2 d 2 x_ + U (x( ))
0

The above expression in known as a functional integral. It says that we must integrate over all functions (i.e., all
values that an arbitrary function x( ) may take on) between the values  = 0 and  =  h. It must really be viewed
as the limit of the discretized integral introduced in the last lecture. The integral is also referred to as a path integral
because it implies an integration over all paths that a particle might take between  = 0 and  =  h such that
x(0) = x( h , where the paths are paramterized by the variable  (which is not time!). The second line in the above
expression, which is equivalent to the rst, indicates that the integration is taken over all paths that begin and end
at the same point, plus a nal integration over that point.
The above expression makes it clear how to represent a general density matrix element hxj exp(?H )jx0 i:
Z x(h)=x " Z h  #
1 m
0

hxje?H jx0 i = Dx(
) exp ? h d 2 x_ + U (x( ))

2
x(0)=x 0

which indicates that we must integrate over all functions x( ) that begin at x at  = 0 and end at x0 at  =  h:

2
 x

x’

τ
0 βh
FIG. 2.

Similarly, diagonal elements of the density matrix, used to compute the partition function, are calculated by
integrating over all periodic paths that satisfy x(0) = x( h) = x:

3
 x

x’

τ
0 βh
FIG. 3.

Note that if we let  = it=h, then the density matrix becomes

(x; x0 ; it=h) = hxje?iHt=h jx0 i = U (x; x0 ; t)
which are the coordinate space matrix elements of the quantum time evolution operator. If we make a change of
variables  = is in the path integral expression for the density matrix, we nd that the quantum propagator can also
be expressed as a path integral:
Z x(t)=x 0  i Z t m 
U (x; x0 ; t) = hxje?iHt=h jx0 i = Dx(
) exp h ds 2 x_ (s) ? U (x(s))
x(0)=x 0

Such a variable transformation is known as a Wick rotation. This nomenclature comes about by viewing time as a
complex quantity. The propagator involves real time, while the density matrix involves a transformation t = ?i h to
the imaginary time axis. It is because of this that the density matrix is sometimes referred to as an imaginary time
path integral.

B. Dominant paths in the propagator and density matrix

Let us rst consider the real time quantum propagator. The quantity appearing in the exponential is an integral of
1 mx_ 2 ? U (x)  L(x; x_ )
2
which is known as the Lagrangian in classical mechanics. We can ask, which paths will contribute most to the integral
Z t hm i Zt
ds 2 x_ (s) ? U (x(s)) = dsL(x(s); x_ (s)) = S [x]
2
0 0

4
known as the action integral. Since we are integrating over a complex exponential exp(iS=h), which is oscillatory,
those paths away from which small deviations cause no change in S (at least to rst order) will give rise to the
dominant contribution. Other paths that cause exp(iS=h) to oscillate rapidly as we change from one path to another
will give rise to phase decoherence and will ultimately cancel when integrated over. Thus, we consider two paths x(s)
and a nearby one constructed from it x(s) + x(s) and demand that the change in S between these paths be 0
S [x + x] ? S [x] = 0
Note that, since x(0) = x and x(t) = x0 , x(0) = x(t) = 0, since all paths must begin at x and end at x0 . The change
in S is
Zt Zt
S = S [x + x] ? S [x] = dsL(x + x; x_ + x_ ) ? dsL(x; x_ )
0 0

Expanding the rst term to rst order in x, we obtain

Zt  @L @L  Zt Z t  @L @L 
S = ds L(x; x_ ) + @ x_ x_ + @x x ? L(x; x_ ) = ds @ x_ x_ + @x x
0 0 0

The term proportional to x_ can be handled by an integration by parts:

Z t @L Z t @L d @L
t Z t d @L
ds @ x_ x_ = ds @ x_ dt x = @ x_ x ? ds dt @ x_ x
0 0 0 0

because x vanishes at 0 and t, the surface term is 0, leaving us with

Z t  d @L @L 
S = ds ? + x = 0
dt @ x_ @x
0

Since the variation itself is arbitrary, the only way the integral can vanish, in general, is if the term in brackets
vanishes:
d @L @L
dt @ x_ ? @x = 0
_ 2 ? U (x), they give
This is known as the Euler-Lagrange equation in classical mechanics. For the case that L = mx=
d @U
dt (mx_ ) + @x = 0
mx = ? @U
@x
which is just Newton's equation of motion, subject to the conditions that x(0) = x, x(t) = x0 . Thus, the classical
path and those near it contribute the most to the path integral.
The classical path condition was derived by requiring that S = 0 to rst order. This is known as an action
stationarity principle. However, it turns out that there is also a principle of least action, which states that the
classical path minimizes the action as well. This is an important consideration when deriving the dominant paths for
the density matrix, which takes the form
Z x(h=x " Z h  m  #
Dx(
) exp ? h1
0

(x; x0 ;  ) = d 2 x_ ( ) + U (x( ))
x(0)=x 0

The action appearing in this expression is

Z h h m i Z h
SE [x] = d 2 x_ + U (x( )) =
2
dH (x; x_ )
0 0

which is known as the Euclidean action and is just the integral over a path of the total energy or Euclidean Lagrangian
H (x; x_ ). Here, we see that a minimum action principle is needed, since the smallest values of SE will contribute most
to the integral. Again, we require that to rst order SE [x + x] ? SE [x] = 0. Applying the same logic as before, we
obtain the condition

5
 d @H @H
d @ x_ ? @x = 0
@ U (x)
mx = @x
which is just Newton's equation of motion on the inverted potential surface ?U (x), subject to the conditions x(0) = x,
x( h ) = x0 . For the partition function Q( ), the same equation of motion must be solved, but subject to the conditions
that x(0) = x( h ), i.e., periodic paths.

II. DOING THE PATH INTEGRAL: THE FREE PARTICLE

The density matrix for the free particle

H = 2Pm
2

will be calculated by doing the discrete path integral explicitly and taking the limit P ! 1 at the end.
The density matrix expression is
 P=2 Z " X #
(x; x0 ;  ) = Plim mP P
dx2


dxP exp ? mP2 (xi+1 ? xi )2 
!1 2 h2 2 h i=1 x1=x;xP +1=x 0

Let us make a change of variables to

u1 = x 1
uk = xk ? x~k

x~k = (k ? 1)xkk+1 + x1
The inverse of this transformation can be worked out explicitly, giving
x1 = u 1
X k?1
P +1
P ? k + 1u
xk = l?1 u l+
P 1
l=1

The Jacobian of the transformation is simply

0 1 ?1=2 0 0


1
BB 0 1 ?2=3 0


C
J = det B @ 00 00 1 ?3=4


C C=1
0 1


A









Let us see what the eect of this transformation is for the case P = 3. For P = 3, one must evaluate
(x1 ? x2 )2 + (x2 ? x3 )2 + (x3 ? x4 )2 = (x ? x2 )2 + (x2 ? x3 )2 + (x3 ? x0 )2
According to the inverse formula,
x1 = u1
x2 = u2 + 21 u3 + 13 x0 + 32 x
x3 = u3 + 32 x0 + 31 x
Thus, the sum of squares becomes

6
 (x ? x2 )2 + (x2 ? x3 )2 + (x3 ? x0 )2 = 2u22 + 23 u23 + 31 (x ? x0 )2
= 2 ?2 1 u22 + 3 ?3 1 u23 + 31 (x ? x0 )2
From this simple exmple, the general formula can be deduced:
X
P X
P
k 2 1 02
(xi+1 ? xi )2 = k ? 1 uk + P (x ? x )
i=1 k=2

Thus, substituting this transformation into the integral gives

 m 1=2 Y
P   " P #  
(x; x0 ;  ) = mk P 1=2 Z du


du exp ? X mk P u2 exp ? m (x ? x0 )2
P
2 h2 2 h2
k=2
2
2 h2 k k=2 2 h2
where
mk = k ?k 1 m
and the overall prefactor has been written as
 mP P=2  m 1=2 Y
P 
mkP
1=2
=
2 h2 2 h2 k=2 2 h2
Now each of the integrals over the u variables can be integrated over independently, yielding the nal result
 m 1=2  m 
(x; x0 ;  ) = exp ? (x ? x 0 )2
2 h2 2 h2
In order to make connection with classical statistical mechanics, we note that the prefactor is just 1=, where 
 1=2  h2 1=2
 = 2
m
h2 =2m
is the kinetic prefactor that showed up also in the classical free particle case. In terms of , the free particle density
matrix can be written as
(x; x0 ;  ) = 1 e?(x?x )2 =2

0

Thus, we see that  represents the spatial width of a free particle at nite temperature, and is called the \thermal de
Broglie wavelength."

7
 G25.2651: Statistical Mechanics

Notes for Lecture 16

I. THE HARMONIC OSCILLATOR { EXPANSION ABOUT THE CLASSICAL PATH

It will be shown how to compute the density matrix for the harmonic oscillator:
2
H = 2Pm + 12 m!2 X 2
using the functional integral representation. The density matrix is given by
" #
 )=x0
Z x(h
1  1
Z h
1
0
(x; x ;  ) = Dx( ) exp ? h 2 2
d 2 mx_ + 2 m! x 2
x(0)=x 0
As we saw in the last lecture, paths in the vicinity of the classical path on the inverted potential give rise to the
dominant contribution to the functional integral. Thus, it proves useful to expand the path x( ) about the classical
path. We introduce a change of path variables from x( ) to y( ), where
x( ) = xcl ( ) + y( )
where xcl ( ) satises
mxcl = m!2xcl
subject to the conditions
xcl(0) = x; xcl ( h) = x0
so that y(0) = y( h) = 0.
Substituting this change of variables into the action integral yields
Z h 1 1

S= 2
d 2 mx_ + 2 m! x 2 2
0
Z h 1 
= d 2 m(x_ cl + y_ )2 + 12 m!2 (xcl + y)2
0
Z h 1 1
 Z h  1 1

= 2 2
d 2 mx_ cl + 2 m! xcl +2 2 2
d 2 my_ + 2 m! y 2
0 0
Z h  
+ d mx_ cl y_ + m!2 xcl y
0
An integration by parts makes the cross terms vanish:
Z h   
Z h  
2
d mx_ cly_ + m! xcl y = mx_ cl yj0 +  h d ?mxcl + m!2xcl y = 0
0 0
where the surface term vanishes becuase y(0) = y( h ) = 0 and the second term vanishes because xcl satises the
classical equation of motion.
The rst term in the expression for S is the classical action, which we have seen is given by
Z h  1 
m! h(x2 + x0 2 )cosh( h !) ? 2xx0 i
d 2 mx_ 2cl + 12 m!2 x2cl = 2sinh(  h !)
0
Therefore, the density matrix for the harmonic oscillator becomes

1
  
(x; x0 ;  ) = I [y] exp m! (x2 + x0 2 )cosh( h!) ? 2xx0 
? 2sinh( h!)
where I [y] is the path integral
" #
y(h )=0  m
2 + m!2 y2
Z Z h 
I [y] = Dy( ) exp ? h1 2 y
_ 2
y(0)=0 0
Note that I [y] does not depend on the points x and x0 and therefore can only contribute an overall (temperature
dependent) constant to the density matrix. This will aect the thermodynamics but not any averages of physical
observables. Nevertheless, it is important to see how such a path integral is done.
In order to compute I [y] we note that it is a functional integral over functions y( ) that vanish at  = 0 and  =  h.
Thus, they are a special class of periodic functions and can be expanded in a Fourier sine series:
1
X
y( ) = cn sin(!n  )
n=1
where
!n = n
 h
Thus, we wish to change from an integral over the functions y( ) to an integral over the Fourier expansion coecients
cn . The two integrations should be equivalent, as the coecients uniquely determine the functions y( ). Note that
1
X
y_ ( ) = !n cn cos(!n  )
n=1
Thus, terms in the action are:
Z h 1 m X1 X 1 
Z h
2
d m y_ = 2 cn cn0 !n !n0 d cos(!n  ) cos(!n0  )
0 n=1 n0 =1 0
Since the cosines are orthogonal between  = 0 and  =  h, the integral becomes
Z h 1 m X1 Z h m X 1  1 1
Z h 
m h
 X1
2
d m y_ = 2 cn !n 2 2 2
d cos (!n  ) = 2 cn !n 2 2 d 2 + 2 cos(2!n ) = 4 c2n !n2
0 n=1 0 n=1 0 n=1
Similarly,
Z h 1 1
m! 2 y2 = m h !2 X c2
2 4 n
0 n=1
The measure becomes
1
Y dcn
Dy(t) ! p
2
n=1 4=m!n
which, is not an equivalent measure (since it is not derived from a determination of the Jacobian), but is chosen
q give the correct free-particle (! = 0) limit, which can ultimately be corrected by attaching an overall factor of
to
m=2 h 2 .
With this change of variables, I [y] becomes
Y1 Z1 dcn
 
m (!2 + !2 )c2 = Y1  !2 1=2
I [y] = p exp ? n
4=m! 2 4 n n ! 2 + !n2
n=1 ?1 n n=1
The innite product can be written as

2
 1 1 " #?1
2 n2 = 2 h2 = Y  2 h2 !2
  
Y
2 2 2 2 2 1 + 2 n2
n=1 ! +  n = h n=1
the product in the square brackets is just the innite product formula for sinh( h!)=( h!), so that I [y] is just
s
 h !
I [y] = sinh(  h!)
q
Finally, attaching the free-particle factor m=2 h2 , the harmonic oscillator density matrix becomes:
r  
(x; x0 ;  ) = m! exp ? m! (x2 + x0 2 )cosh( h!) ? 2xx0 
2hsinh( h!) 2sinh( h!)
Notice that in the free-particle limit (! ! 0), sinh( h!)   h! and cosh( h!)  1, so that
r  
(x; x0 ;  ) ! m exp ? m (x ? x0 )2
2 h2 2 h2
which is the expected free-particle density matrix.

II. THE STATIONARY PHASE APPROXIMATION

Consider the simple integral:
Z 1
I = lim
!1
dx e?f (x)
?1
Assume f (x) has a global minimum at x = x0 , such that f 0 (x0 ) = 0. If this minimum is well separated from other
minima of f (x) and the value of f (x) at the global minimum is signicantly lower than it is at other minima, then
the dominant contributions to the above integral, as  ! 1 will come from the integration region around x0 . Thus,
we may expand f (x) about this point:
f (x) = f (x0 ) + f 0(x0 )(x ? x0 ) + 21 f 00 (x0 )(x ? x0 )2 +



Since f 0 (x0 ) = 0, this becomes:
f (x)  f (x0 ) + 21 f 00 (x0 )(x ? x0 )2
Inserting the expansion into the expression for I gives
Z 1
I = lim
!1
e ?f (x 0 ) dx e? 2 f 00 (x0 )(x?x0)2
?1
 1=2
= lim 2 e?f (x0 )
!1 f 00 (x0 )
Corrections can be obtained by further expansion of higher order terms. For example, consider the expansion of f (x)
up to fourth order:
f (x)  f (x0 ) + 21 f 00 (x0 )(x ? x0 )2 + 16 f 000 (x0 )(x ? x0 )3 + 24 1 f (iv) (x )(x ? x )4
0 0
Substituting this into the integrand and further expanding the exponential would give, as the lowest order nonvanishing
correction:
Z 1  
I = lim e ?f (x0 ) dx e
? f 00 (x0 )(x?x0 )2
1 ?  f (iv) (x )(x ? x )4
!1
2
24 0 0
?1

3
This approximation is known as the stationary phase or saddle point approximation. The former may seem a little
out-of-place, since there is no phase in the problem, but that is because we formulated it in such a way as to anticipate
its application to the path integral. But this is only if  is taken to be a real instead of an imaginary quantity.
The application to the path integral follows via a similar argument. Consider the path integral expression for the
density matrix:
 )=x0
Z x(h
(x; x0 ;  ) = D[x]e?SE [x]=h
x(0)=x
We showed that the classical path satisfying

mxcl = @U


@x x=xcl
x(0) = x x( h) = x0
is a stationary point of the Euclidean action SE [x], i.e., SE[xcl ] = 0. Thus, we can develop a stationary phase or
saddle point approximation for the density matrix by introducing an expansion about the classical path according to
X
x( ) = xcl ( ) + y( ) = xcl ( ) + cn n ( )
n
where the correction y( ), satisfying y(0) = y( h) = 0 has been expanded in a complete set of orthonormal functions
fn ( )g, which are orthonormal on the interval [0;  h] andsatisfy n (0) = n ( h) = 0 as well as the orthogonality
condition:
Z h
d n ( )m ( ) = mn
0
Setting all the expansion coecients to 0 recovers the classical path. Thus, we may expand the action S [x] (the \E"
subscript will henceforth be dropped from this discussion) with respect to the expansion coecients:
 
X @S 
S [x] = S [xcl ] + @c  c + 1 X @ 2 S  cj ck +



j 2
j fcg=0 j;k @cj ck fcg=0

j
Since
Z h  
S [x] = d 12 mx_ 2 + U (x( ))
0
the expansion can be worked out straightforwardly by substitution and subsequent dierentiation:
2 !2 3

Z h
1 X X
S [x] = d 4 2 m x_ cl + cn _ n + U (xcl + cn n )5
0 n n
" ! #
@S = Z h d m(x_ + X c _ )_ + U 0 x + X c  
@cj 0 cl n n j cl n n j
n n

@S   h
Z h i
= d m x
_ cl _ j + U 0 (xcl )j

@cj fcg=0 0

Z h 
= mx_ cl j j0h + d [?mxcl + U 0 (xcl )] j
0
=0
" ! #
@ 2 S = Z h m_ _ + U 00 x + X c  ]  
@cj @ck 0 j k cl n n j k
n

4
 
@ 2 S   h
Z h i
= d m _
 j _
 k + U 00 (xcl )j k
@cj @ck fcg=0 0
Z h h i
= d ?mj k + U 00 (xcl ( ))j k
0
Z h 2  
= dj ( ) ?m dd 2 + U 00 (xcl ( )) k ( )
0
where the fourth and eighth lines are obtained from an integration by parts. Let us write the integral in the last line
in the suggestive form:

@ 2 S  = h j ? m @ 2 + U 00 (x ( ))j i =

@cj @ck fcg=0 j @ 2 cl k jk

which emphasizes the fact that we have matrix elements of the operator ?md2=d 2 + U 00 (xcl ( )) with respect to the
basis functions. Thus, the expansion for S can be written as
S [x] = S [xcl ] + 12 cj
jk ck +



X

j;k
and the density matrix becomes
Z Y P
(x; x0 ;  ) = N pdcj e?Scl (x;x0;) e? 2 j;k cj
jk ck =h

1

j 2h
where Scl(x; x0 ;  ) = S [xcl ]. N is an overall normalization constant. p The integral over the coecients becomes a
generalized Gaussian integral, which brings down a factor of 1= det
:
(x; x0 ;  ) = N e?Scl (x;x0 ;) p 1
det

0
= N e?Scl (x;x ;) q ? 1


det ?m dd 2 + U 00 (xcl ( ))
2

where the last line is the abstract representation of the determinant. The determinant is called the Van Vleck-Pauli-
Morette determinant.
If we choose the basis functions n ( ) to be eigenfunctions of the operator appearing in the above expression, so
that they satisfy

d2 00

?m d 2 + U (xcl ( )) n ( ) = n n ( )
Then,

jk = j jk = j (x; x0 ;  )jk
and the determinant can be expressed as a product of the eigenvalues. Thus,
(x; x0 ;  ) = N e?Scl (x;x0;) p 1 0
Y

j j (x; x ;  )
The product must exclude any 0-eigenvalues.
Incidentally, by performing a Wick rotation back to real time according to  = ?it=h, the saddle point or stationary
phase approximation to the real-time propagator can be derived. The derivation is somewhat tedious and will not be
given in detail here, but the result is
U (x; x0 ; t) = e hi Scl (x;x0;t) q ? 1 ?i=2
d
e
det ?m dt2 ? U 00 (xcl (t))
2

5
where xcl (t) satises

mxcl = ? @U


@x  x=xcl
xcl (0) = x xcl (t) = x0
and  is an integer that increases by 1 each time the determinant vanishes along the classical path.  is called the
Maslov index. It is important to note that because the classical paths satisfy an endpoint problem, rather than an
initial value problem, there can be more than one solution. In this case, one must sum the result over classical paths:
U (x; x0 ; t) =
X
e hi Scl (x;x0;t)?i=2 q ? 1


classical paths det ?m dtd 2 ? U 00 (xcl (t))
2

with a similar sum for the density matrix.

6
 G25.2651: Statistical Mechanics

Notes for Lecture 17

I. EXPECTATION VALUES OF OBSERVABLES

Recall the basic formula for the expectation value of an observable A:
hAi = Q(1 ) Tr(Ae?H )
Two important cases pertaining to the evaluation of the trace in the coordinate basis for expectation values will be
considered below:

A. Case 1: Functions only of position

If A = A(X ), i.e., a function of the operator X only, then the trace can be easily evaluated in the coordinate basis:
Z
hAi = 1 dxhxjA(X )e?H jxi
Q
Since A(X ) acts to the left on one of its eigenstates, we have
Z
hAi = 1 dxA(x)hxje?H jxi
Q
which only involves a diagonal element of the density matrix. This can, therefore, be written as a path integral:
  "  #
hAi = Q1 Plim mP P=2 Z dx


dx A(x ) exp ? X P
1 m!2 (x ? x )2 + 1 U (x )
P
!1 2 h2 1 1
i=1
2 P i+1 i P i
However, since all points x1 ; ::; xP are equivalent, due to the fact that they are all integrated over, we can make P
equivalent cyclic renaming of the coordinates x1 ! x2 , x2 ! x3 , etc. and generate P equivalent integrals. In each,
the function A(x1 ) or A(x2 ), etc. will appear. If we sum these P equivalent integrals and divide by P , we get an
expression:
 P=2 Z P
"
P   #
mP X X
hAi = Q1 Plim dx1


dxP P1 A(xi ) exp ? 1 m!2 (x ? x )2 + 1 U (x )
2 P i+1 i
P i
!1 2 h2 i=1 i=1
This allows us to dene an estimator for the observable A. Recall that an estimator is a function of the P variables
x1 ; :::; xP whose average over the ensemble yields the expectation value of A:
XP
aP (x1 ; :::; xP ) = P1 A(xi )
i=1
Then
hAi = Plim ha i
!1 p x ;:::;x
1 P

where the average on the right is taken over many congurations of the P variables x1 ; :::; xP (we will discuss, in the
nex lecture, a way to generate these congurations).
The limit P ! 1 can be taken in the same way that we did in the previous lecture, yielding a functional integral
expression for the expectation value:
I " Z # " Z h  #
hAi = Q Dx( ) 1h
1 1 d 21 mx_ 2 + U (x( ))
h

dA(x( )) exp ? h
0 0

1
 B. Case 1: Functions only of momentum
Suppose that A = A(P ), i.e., a function of the momentum operator. Then, the trace can still be evaluated in the
coordinate basis:
Z
hAi = 1 dxhxjA(P )e?H jxi
Q
R
However, A(P ) acting to the left does not act on an eigenvector. Let us insert a coordinate space identity I = dxjxihxj
between A and exp(?H ):
Z
hAi = 1 dxdx0 hxjA(P )jx0 ihxje?H jxi
Q
Now, we see that the expectation value can be obtained by evaluating all the coordinate space matrix elements of the
operator and all the coordinate space matrix elements of the density matrix.
A particularly useful form for the expectation value can be obtained if a momentum space identity is inserted:
Z
hAi = 1 dxdx0 dphxjA(P )jpihpjx0 ihx0 je?H jxi
Q
Now, we see that A(P ) acts on an eigenstate (at the price of introducing another integral). Thus, we have
Z Z
hAi = 1 dpA(p) dxdx0 hxjpihpjx0 ihx0 je?H jxi
Q
Using the fact that hxjpi = (1=2h) exp(ipx=h), we nd that
Z Z
hAi = 21hQ dpA(p) dxdx0 eip(x?x )=h hx0 je?H jxi
0

In the above expression, we introduce the change of variables

0
r = x +2 x s = x ? x0
Then
Z Z
hAi = 2h Q dpA(p) dr dseips=h hr ? 2s je?H jr + 2s i
1

Dene a distribution function
Z
W (r; p) = 2h dseips=h hr ? 2s je?H jr + 2s i
1

Then, the expectation value can be written as

Z
hAi = 1 drdpA(p) (r; p)
Q W

which looks just like a classical phase space average using the \phase space" distribution function W (r; p). The
distribution function W (r; p) is known as the Wigner density matrix and it has many interesting features. For one
thing, its classical limit is
  
W (r; p) = exp ? 2pm + U (r)
2

which is the true classical phase space distribution function. There are various examples, in which the exact Wigner
distribution function is the classical phase space distribution function, in particularly for quadratic Hamiltonians.
Despite its compelling appearance, the evaluation of expectation values of functions of momentum are considerably
more dicult than functions of position, due to the fact that the entire density matrix is required. However, there are
a few quantities of interest, that are functions of momentum, that can be evaluated without resorting to the entire
density matrix. These are thermodynamic quantities which will be discussed in the next section.

2
 II. THERMODYNAMICS FROM PATH INTEGRALS
Although general functions of momentum are dicult (though not intractable) to evaluate by path integration, cer-
tain functions of momentum (and position) can be evaluated straightforwardly. These are thermodynamic quantities
such as the energy and pressure, given respectively by
@ ln Q(; V )
E = ? @
@ ln Q(; V )
P = 1 @V
We shall derive estimators for these two quantities directly from the path integral expression for the partition function.
However, let us work with the partition function for an ensemble of 1-particle systems in three dimensions, which is
given by
 3P=2 Z " P  #
mP X 1 1
Q(; V ) = Plim
!1 2 h2
dr1


drP exp ? 2 m!P (ri+1 ? ri ) + P U (ri )
2 2

i=1

Using the above thermodynamic relation, the energy becomes

E = ? 1 @Q Q @
 3P=2 Z " P  #
mP X
= Q1 Plim dr1


drP exp ? 1 1
2 m!P (ri+1 ? ri ) + P U (ri )
2 2
!1 2 h2 i=1
" P P #
X 1X
 32P ? 1
2 m!P2 (ri+1 ? ri )2 + P U (ri )
i=1 i=1

= Plim h" (r ; :::; rP )i

!1 P 1
where
XP XP
3P 1 2 2 1
"P (r1 ; :::; rP ) = 2 ? 2 m!P (ri+1 ? ri ) + P U (ri )
i=1 i=1

is the thermodynamic estimator for the total energy.

Similarly, an estimator for the internal pressure can be derived using P = kT@ ln Q=@V . As we have done in the
past for classical systems, the volume dependence can be made explicity by introducing the change of variables:
rk = V 1=3 sk
In terms of the scaled variables sk , the partition function expression reads:
 3P=2 Z " P  #
mP X 1
Q(; V ) = Plim
!1 2 h2
V P ds1


dsP exp ? 2 m!P V
2 2=3
(si+1 ? si ) + P1 U (V 1=3 si )
2

i=1

Evaluating the derivative with respect to volume gives the internal pressure:
P = 1 @Q
Q @V
 3P=2 Z " P  #
mP X
= Q1 Plim V P ds1


dsP exp ? 1
2 m!P V
2 2=3
(si+1 ? si ) + P1 U (V 1=3 si )
2
!1 2 h2 i=1
" P P #
P 1 2 ?1=3
X 1 X @U
1 V ?2=3 s
 V ? 3 m!P V (si+1 ? si ) ? P
2
@ (V 1=3 s ) 3 i
i=1 i=1 i

3
  3P=2 Z " P  #
mP X
= Q1 Plim dr1


drP exp ? 1 1
2 m!P (ri+1 ? ri ) + P U (ri )
2 2
!1 2 h2 i=1
" P P #
P ? 1 m!2 (r ? r )2 ? X 1 X @U
 V 3V P i=1 i+1 i 3V P @ ri
si
i=1

= Plim hp (r ; :::; rP )i
!1 P 1
where
P  
P ? 1 X
pP (r1 ; :::; rP ) = V m! 2
(r ? r ) 2
+ 1 r
@U
3V i=1 P i+1 i
P i @ ri
is the thermodynamic estimator for the pressure. Clearly, both the energy and pressure will be functions of the particle
momenta, however, because they are related to the partition function by thermodynamic dierentiation, estimators
can be derived for them that do not require the o-diagonal elements of the density matrix.

III. PATH INTEGRAL MOLECULAR DYNAMICS (OPTIONAL READING)

Consider once again the path integral expression for the one-dimensional canonical partition function (for a nite
but large value of P ):
 P=2 Z " P  #
X
Q( ) = mP 2 dx1


dxP exp ? 1 1
2 m!P (xi+1 ? xi ) + P U (xi )
2 2
(1)
2 h i=1

(the condition xP +1 = x1 is understood). Recall that, according to the classical isomorphism, the path integral
expression for the canonical partition function is isomorphic to the classical conguration integral for a certain P -
particle system. We can carry this analogy one step further by introducing into the above expression a set of P
momentum integrations:
Z " P  #
X p2i + 1 m!2 (x ? x )2 + 1 U (x )
Q( ) = dp1


dpP dx1


dxP exp ? 2m0 2 P i+1 i P i (2)
i=1

Note that these momentum integrations are completely uncoupled from the position integrations, and if we were to
carry out these momentum integrations, we would reproduce Eq. (1) apart from trivial constants. Written in the form
Eq. (2), however, the path integral looks exactly like a phase space integral for a P -particle system. We know from
our work in classical statistical mechanics that dynamical equations of motion can be constructed that will generate
this partition function. In principle, one would start with the classical Hamiltonian
P 
X 
H= p2i + 1 m!2 (x ? x )2 + 1 U (x )
i=1
2m0 2 P i+1 i P i
derive the corresponding classical equations of motion and then couple in thermostats. Such an approach has certainly
been attempted with only limited success. The diculty with this straightforward approach is that the more \quan-
tum" a system is, the large the paramester P must be chosen in order to converge the path integral. However, if P
is large, the above Hamiltonian describes a system with extremely sti nearest-neighbor harmonic bonds interacting
with a very weak potential U=P . It is, therefore, almost impossible for the system to deviate far harmonic oscillator
solutions and explore the entire available phase space. The use of thermostats can help this problem, however, it is
also exacerbated by the fact that all the harmonic interactions are coupled, leading to a wide variety of time scales
associated with the motion of each variable in the Hamiltonian. In order to separate out all these time scales, one
must somehow diagonalize this harmonic interaction. One way to do this is to use normal mode variables, and this is

4
a perfectly valid approach. However, we will explore another, simpler approach here. It involves the use of a variable
transformation of the formed used in previous lectures to do the path integral for the free-particle density matrix.
Consider a change of variables:
u1 = x 1
uk = xk ? x~k k = 2; :::; P
where
x~ = (k ? 1)xk+1 + x1
i
k
The inverse of this transformation can be worked out in closed form:
x1 = u1
xk = u1 + ((kl ?? 1)
XP
1) u
l
l=k
and can also be expressed as a recursive inverse:
x1 = u 1
xk = uk + k ?k 1 xk+1 + k1 x1
The term k = P here can be used to start the recursion. We have already seen that this transformation diagonalized
the harmonic interaction. Thus, substituting the transformation into the path integral gives:
Z " P  #
X p2i + 1 m !2 u2 + 1 U (x (u ; :::; u ))
Q( ) = dp1


dpP du1


duP exp ? 2m0i 2 i P i P i 1 P
i=1
The parameters mi are given by
m1 = 0
mi = i ?i 1 m
Note also that the momentum integrations have been changed slightly to involve a set of parameters m0i . Introducing
these parameters, again, only changes the partition function by trivial constant factors. How these should be chosen
will become clear later in the discussion. The notation xi (u1 ; :::; uP ) indicates that each variable xi is a generally a
function of all the new variables u1 ; :::; uP .
A dynamics scheme can now be derived using as an eective Hamiltonian:
P 
X 
H= p2i + 1 m !2 u2 + 1 U (x (u ; :::; u ))
i=1
2m0i 2 i P i P i 1 P

which, when coupled to thermostats, yields a set of equations of motion

m0 u = ?m !2 u ? 1 @U ? _ u_
i i i P i
P @ui i i

Qi = mi u_ 2i ? 1 (3)
These equations have a conserved energy (which is not a Hamiltonian):
P 
X 
0
H = 1 0 1 1 1 1
mi u_ i + mi !P ui + U (xi (u1 ; :::; uP )) + Q_i + i
2 2 2 2
i=1
2 2 P 2 
Notice that each variable is given its own thermostat. This is done to produce maximum ergodicity in the trajectories.
In fact, in practice, the chain thermostats you have used in the computer labs are employed. Notice also that the

5
time scale of each variable is now clear. It is just determined by the parameters fmi g. Since the object of using such
dynamical equations is not to produce real dynamics but to sample the phase space, we would really like each variable
to move on the same time scale, so that there are no slow beads trailing behind the fast ones. This eect can be
produced by choosing each parameter m0i to be proportional to mi : m0i = cmi . Finally, the forces on the u variables
can be determined easily from the chain rule and the recursive inverse given above. The result is
P
1 @U = 1 X @U
P @u1 P i=1 @xi
 
1 @U = 1 (k ? 2) @U + @U
P @ui P (k ? 1) @uk?1 @xk
where the rst (i = 1) of these expressions starts the recursion in the second equation.
Later on, when we discuss applications of path integrals, we will see why a formulation such as this for evaluating
path integrals is advantageous.

IV. PATH INTEGRALS FOR N -PARTICLE SYSTEMS

If particle spin statistics must be treated in a given problem, the formulation of the path integral is more complicated,
and we will not treat this subject here. The extension of path integrals to N -particle systems in which spin statistics
can safely be ignored, however, is straightforward, and we will give the expressions below.
The partition function for an N -particle system in the canonical ensemble without spin statistics can be formulated
essentially by analogy to the one-particle case. The partition function that one obtains is
"N  3P=2 Z # " P N
!#
Y mI P X X 1
Q(N; V; T ) = Plim drI


drI
(1) (P )
exp ? 2 mI !P (rI
2 (i+1)
? rI ) + P1 U (r(1i) ; :::; r(Ni) )
(i) 2
!1 I =1 2 h2 i=1 I =1

Thus, it can be seen that the N -particle potential must be evaluated for each imaginary time discretization, however,
there is no coupling between separate imaginary time slices due arising from the potential. Thus, interactions occur
only between particles in the same time slice. From a computational point of view, this is advantageous, as it allows
for easily parallelization over imaginary time slices.
The corresponding energy and pressure estimators for the N -particle path integral are given by
P X
X N  2 P
X
P (fr(1) ; :::; r(P ) ) = 3NP
2 ?
1
2 mI !P
2 (i)
rI ? r(Ii+1) + P1 U (r(1i) ; :::; r(Ni) )
i=1 I =1 i=1
N 
P X
X  2 
pP (fr(1) ; :::; r(P ) ) = NP ?
V 3V
1 mI !P2 (i)
rI ? r(Ii+1) + P1 r(Ii)
rr U (i)
I
i=1 I =1

6
 G25.2651: Statisti al Me hani s

Notes for Le ture 18

I. INTRODUCTION TO SPIN

The path integral formulation of quantum statisti al me hani s is parti ularly useful for situations in whi h parti le
spin statisti s an be largely ignored. In the quantum ideal gases, we have a situation in whi h the spin statisti s
determine all of the interesting behavior! The fully quantum treatment of the ideal gas will be the subje t of the next
several le tures.
The spin degree of freedom of a parti le is a purely quantum me hani al aspe t (with no lassi al analog). In
quantum me hani s, spin is analogous to an angular momentum. It is des ribed by a Hermitian operator S =
(Sx ; Sy ; Sz ), where the omponents satisfy angular momentum type ommutation relations:
[Sx ; Sy ℄ = ih Sz [Sy ; Sz ℄ = ih Sx [Sz ; Sx ℄ = ih Sy
The spin operators for a spin-s parti le are represented by (2s + 1)  (2s + 1) matri es (whi h dene dierent
representations of the group SU(2)). For example, for a spin-1/2 parti le, su h as an ele tron, the three spin operators
are
     
h 0 1 h 0 i h 1 0
Sx = Sy = Sz =
2 1 0 2 i 0 2 0 1
whi h an be shown to satisfy the above ommutation relations. Sin e the three omponents of spin to do not ommute,
we hoose, by onvention, to work in a basis in whi h Sz is diagonal. Thus, there will be (2s + 1) eigenvalues given
by sh ; :::; sh . In the example of the spin-1/2 parti le, we see that the allowed spin eigenvalues (denoted m) are
m = hbar=2 and m = h =2. The orresponding eigenstates are just
   
jm = h=2i  j1=2 i = 10 jm = h=2i  j 1=2 i = 01
whi h are denoted the \spin-up" and \spin-down" states, respe tively. Note that the operator S 2 = Sx2 + Sy2 + Sz2 is
also diagonal so that the spin-up and spin-down eigenstates of Sz are also eigenstate of S 2 , both having the eigenvalue
s(s + 1)h . Thus, given a Hamiltonian H for a system, if H is independent of spin, then the eigenstates of H must
2

also be eigenstates of S 2 and Sz sin e all three an be simultaneously diagonalized.

What happens in quantum me hani s when we have systems of identi al parti les of a given type of spin? Consider
the simple example of a system of two identi al spin-1/2 parti les. Suppose we perform a measurement whi h is able
to determine that one of the parti les has an Sz eigenvalue of ma h and the other mb h su h that ma 6= mb . Is the
state ve tor of the total system just after this measurement
jma ; mb i or jmb ; ma i
where, in the rst state, parti les 1 and 2 have Sz eigenvalues ma h and mb h , respe tively, and, in the se ond state, it
is the reverse of this? The answer is that neither state is the orre t state ve tor sin e the measurement is not able
to assign the parti ular spin states of ea h parti le. In fa t, the two state jma ; mb i and jmb ; ma i are not physi ally
equivalent states. Two states j i and j 0 i an only be physi all equivalent if there is a omplex number  su h that
j i = j 0 i
and there is no su h number onne ting jma ; mb i and jmb ; ma i. However, it is possible to onstru t a new state ve tor
j (ma ; mb )i su h that j (mb ; ma )i is physi ally equivalent to j (ma ; mb )i. Let
j (ma ; mb )i = C jma ; mb i + C 0 jmb ; ma i
If we require that
j (ma ; mb )i = j (mb ; ma )i
1
then
C jma ; mb i + C 0 jmb ; ma i =  (C jmb ; ma i + C 0 jma ; mb i)
from whi h we see that
C = C 0 C 0 = C
or
C 0 = 2 C 0
from whi h  = 1 and C = C 0 . This gives us two possible physi al states of the system
j S (ma ; mb )i / jma ; mb i + jmb ; ma i
j A (ma ; mb )i / jma ; mb i jmb ; ma i
whi h are symmetri and antisymmetri , respe tively, with respe t to an ex hange of the parti le spin eigenvalues.
The analog in ordinary one-dimensional quantum me hani s would be the ase of two identi al parti les moving along
the x axis. If a measurement performed on the system determined that a parti le was at position x = a and the other
was at x = b, then the state of the system after the measurement would be one of the two following possibilities:
j S (a; b)i / ja bi + jb ai
j A(a; b)i / ja bi jb ai
The standard postulates of quantum me hani s now need to supplemented by an additional postulate that allows us
to determine whi h of the two possible physi al states a system will assume. The new postulate states the following:
In nature, parti les are of two possible types { those that are always found in symmetri (S) states and those that
are always found in antisymmetri (A) states. The former of these are known as bosons and the latter are known as
fermions. Moreover, fermions possess only half-integer spin, s=1/2,3/2,5/2,..., while bosons possess only integer spin,
s=0,1,2,....
Suppose a system is omposed of N identi al fermions or bosons with oordinate labels r1 ; :::; rN and spin labels
s1 ; :::; sN . Let us dene, for ea h parti le, a ombined lable xi  ri ; si . Then, for a given permutation P (1); :::; P (N )
of the parti le indi es 1,..,N , the wave fun tion will be totally symmetri if the parti les are bosons:
B (x1 ; :::; xN ) = B (xP (1) ; ::::; xP (N ) )
For fermions, as a result of the Pauli ex lusion prin iple, the wave fun tion is antisymmetri with respe t to an
ex hange of any two parti les in the systems. Therefore, in reating the given permutation, the wave fun tion will
pi k up a fa tor of -1 for ea h ex hange of two parti les that is performed:
F (x1 ; :::; xN ) = ( 1)Nex F(xP (1) ; ::::; xP (N ) )
where Nex is the total number of ex hanges of two parti les required in order to a hieve the permutation P (1); :::; P (N ).
An N -parti le bosoni or fermioni state an be reated from a state (x1 ; :::; xN ) whi h is not properly symmetrized
but whi h, nevertheless, is an eigenfun tion of the Hamiltonian
H  = E
Noting that there will be N ! possible permutations of the N parti le labels in an N -parti le state, the bosoni state
B (x1 ; :::; xN ) is reated from (x1 ; :::; xN ) a ording to
N!
B (x1 ; :::; xN ) = N1 ! P (x1 ; :::; xN )
X

=1
where Palpha reates 1 of the N ! possible permutations of the indi es. The fermioni state is reated from
N!
F (x1 ; :::; xN ) = N1 ! ( 1)Nex () P (x1 ; :::; xN )
X

=1
where Nex() is the number of ex hanges needed to reate permutation .
This simple dieren e in the symmetry of the wavefun tion leads to stark ontrasts in the properties of fermoni
and bosoni systems. With these quantum me hani al rules in mind, let us work out what these properties are.

2
 II. SOLUTION OF THE N -PARTICLE EIGENVALUE PROBLEM
The Hamiltonian for an ideal gas of N parti les is
N
H=
X P2i
i=1 2m
The eigenvalue problem for the Hamiltonian is in the form of the time-independent S hrodinger equation for the
(unsymmetrized) eigenfun tions
N
h 2 X
2m i=1 ri (x1 ; :::; xN ) = E (x1 ; :::; xN )
2

First, we noti e that the equation is ompletely separable in the N parti le oordinate/spin labels x1 ; :::; xN ,
meaning that the Hamiltonian is of the form
N
X
H= hi
i=1
P2
hi = i
2m
Note, further, that H is independent of spin, hen e, the eigenfun tions must also be eigenfun tions of S 2 and Sz .
Therefore, the solution an be written as
N
Y
1 m1 ;:::; N mN
(x1 ; :::; xN ) =  m (xi )
i i

i=1
where  m (xi ) is a single parti le wave fun tion hara terized by a set of spatial quantum numbers i and Sz
eigenvalues mi . The spatial quantum numbers i are hose to hara terized the spatial part of the eigenfun tions
i i

in terms of appropriately hosen observables that ommute with the Hamiltonian. Note that ea h single-parti le
fun tion  m (xi ) an be further de omposed into a produ t of a spatial fun tion  (ri ) and a spin eigenfun tion
m (si ), where
i i i

m (s) = hsjm i = Æms

Substituting this ansatz in to the wave equation yields a single-parti le wave equation for ea h single parti le fun tion:
h 2 2
2m ri  (ri ) = "  (ri )
i i i

Here, " is a single parti le eigenvalue, and the N -parti le eigenvalue is, therefore, given by
i

N
X
E1 ;:::; =N
" i

i=1
We will solve the single-parti le wave equation in a ubi box of side L for single parti le wave fun tions that satisfy
periodi boundary onditions:
 (xi ; yi ; zi ) =  (xi + L; yi ; zi ) =  (xi ; yi + L; zi ) =  (xi ; yi ; zi + L)
i i i i

Note that the momentum operator Pi ommutes with the orresponding single-parti le Hamiltonian
[Pi ; hi ℄ = 0
This means that the the momentum eigenvalue pi is a good number for hara terizing the single parti le states
 = p . In fa t, the solutions of the single-parti le wave equation are of the form
i i

3
 = Ceip
r =h pi (ri ) i i

provided that the single parti le eigenvalues are given by

p2
"p = i
2m i

R
The onstant C is an overall normalization onstant on the single-parti le states to ensure that d3 ri j i (ri )j2 = 1.
Now, we apply the periodi boundary ondition. Consider the boundary ondition in the x-dire tion. The ondition
p (xi ; yi ; zi ) = p (xi + L; yi ; zi )
i i

leads to
eip x =h = eip (x +L)=h xi i xi i

or
   
1= eipxi L=h = os pxh L + i sin pxh L
i i

whi h will be satised if

px L
h
= 2nx i
i

where nx is an integer, 0; 1; 2,.... Thus, the momentum px an take on only dis rete values, i.e., it is quantized,
and is given by
i i

px =
2h n
i
L x i

Applying the boundary onditions in y and z leads to the onditions

p =
2h n
yi
L y i

pz =
2h n
L z
i i

Thus, the momentum ve tor pi an be written generally as

p = 2h n i
L i
where ni is a ve tor of integers ni = (nx ; ny ; nz ). This ve tor of integers an be used in pla e of pi to hara terize
the single-parti le eigenvalues and wave fun tions. The single-parti le energy eigenvalues will be given by
i i i

"n = i =
p2 22 h 2 jn j2
2m mL2 i i

and the single-parti le eigenfun tions are given by

2in
r =L
n (ri ) = Ce i
i i

Finally, the normalization onstant C is determined by the ondition

Z Z L Z L Z L
d3 ri j n (ri )j = jC j
i
2 2
dxi dyi dzi e 2in
r =Le2in
r =L = 1
i i i i

0 0 0
Z L Z L Z L
jC j2 dxi dyi dzi = jC j2 L3 = 1
0 0 0

C= p1
V

4
Therefore, the omplete solution for the single-parti le eigenvalues and eigenfun tions is
hx jn m i =  (x ) = p1 e2in
r =L  (s )
i i i n i mi i i i
mi i
V
"n =
22 h2 2
jn j
mL2 i
i

and the total energy eigenvalues are given by

N
En1 ;:::;n =
X 22 h2 jn j2
i
N

i=1 mL2
Another way to formulate the solution of the eigenvalue problem is to onsider the single parti le eigenvalue and
eigenfun tion for a given ve tor of integers n:
(r) = p1 e2in
r=L
n
V
"n =
22 h2 jnj2
mL2
and ask how many parti les in the N -parti le system o upy this state. Let this number be fnm. fnm is alled an
o upation number and it tells just how many parti les o upy the state hara terized by a ve tor of integers n. Sin e
there are an innite number of possible hoi es for n, there is an innite number of o upation numbers. However,
they must satisfy the obvious restri tion
XX
fnm = N
m n
where
X 1
X 1
X 1
X

n nx = 1n y = 1n z = 1
and
X s
X

m m= s
runs over the (2s +1) possible values of m for a spin-s parti le. These o upation numbers an be used to hara terize
the total energy eigenvalues of the system. The total energy eigenvalue will be given by
XX
Effnm g = "n fnm
m n

III. AN IDEAL GAS OF DISTINGUISHABLE QUANTUM PARTICLES

As an illustration of the use of o upation numbers in the evaluation of the quantum partition fun tion, let us
onsider the simple ase of Boltzmann statisti s (ignoring spin statisti s or treating the parti les as distinguishable).
The anoni al partition fun tion Q(N; V; T ) an be expressed as a sum over the quantum numbers n1 ; :::; nN for ea h
parti le:
XX X
Q(N; V; T ) =


e En1 n ;:::; N

n1 n2 nN
XX X
=


e "n1 e "n2


e "n N

n1 n2 nN
! ! !
X X X
= e "n1 e "n2


e "nN
n1 n2 nN
!N
X
= e "nN
n

5
In terms of o upation numbers, the partition is
X P
Q(N; V; T ) = g(ff g)e  n "n fn
ff g
where g(ff g) is a fa tor that tells how many dierent physi al states an be represented by a given set of o upation
numbers ff g. For Boltzmann parti les, ex hanging the momentum labels of two parti les leads to a dierent physi al
state but leaves the o upation numbers un hanged. In fa t the ounting problem is merely one of determining how
many dierent ways an N parti les be pla ed in the dierent physi al states. This is just
N!
g(ff g) = Q
n fn !
For example, if there are just two states, then the o upation numbers must be N1 and N2 where N1 + N2 = N . The
above formula gives
N! N!
g(N1; N2 ) = =
N1 !N2 ! N1 !(N N1 )!
whi h is the expe ted binomial oeÆ ient.
The partition fun tion therefore be omes
X N! Y
Q(N; V; T ) = Q e fn"n
ff g n fn ! n
whi h is just the multinomial expansion for
!N
X
Q(N; V; T ) = e "n
n
Again, if there were two states, then the partition fun tion would be
"1 "2 )N
X N!
(e +e = e N1"1 e N2 "2
N1 ;N2 ;N1 +N2 =N N1 !N 2 !
using the binomial theorem.
Therefore, we just need to be able to evaluate the sum
X X
 2 jnj2 =mL2
e "n = e 2 2  h
n n
But we are interested in the thermodynami limit, where L ! 1. In this limit, the spa ing between the single-parti le
energy levels be omes quite small, and the dis rete sum over n an, to a very good approximation, be repla ed by an
integral over a ontinuous variable:
X Z
e 2 h jnj =mL = d3 ne 2 h jnj =mL
2 2 2 2 2 2 2 2

n
Sin e the single-parti le eigenvalues only depend on the magnitude of n, this be omes
Z 1 m
 3=2  
4 dnn e
2  2 jnj2 =mL2
2 2  h
=V = V3
0 2 h2
where  is the thermal deBroglie wavelength.
Hen e,
 
V N
Q(N; V; T ) = 3

whi h is just the lassi al result. Therefore, we see that an ideal gas of distinguishable parti les, even when treated
fully quantum me hani ally, will have pre isely the same properties as a lassi al ideal gas. Clearly, all of the quantum
ee ts are ontained in the parti le spin statisti s. In the next few le tures we will see just how profound an ee t
the spin statisti s an have on the equilibrium properties.

6
 IV. GENERAL FORMULATION FOR FERMIONS AND BOSONS

For systems of identi al femions and identi al bosons, an ex hange of parti les does not hange the physi al state.
Therefore the fa tor g(ffnmg) is just 1 for both of kinds of systems. Moreover, the o upation number of a state
hara terized by n for a system of identi al bosons an be any number between 0 and N :
fnm = 0; 1; 2; :::; N
For fermions, the Pauli ex lusion prin iple forbids two identi al parti les from o upying the same quantum state.
This restri ts the o upation numbers to be either 0 or 1:
fnm = 0; 1
Given these possibilities for the o upation numbers, the anoni al partition fun tion an be formulated:
X P P X YY
Q(N; V; T ) = e  m n fnm "n = e fn"n
ffnm g ffnm g n m

Note that the sum over o upation numbers must be performed subje t to the restri tion
XX
fnm = N
m n
a ondition that makes the evaluation of Q(N; V; T ) extremely diÆ ult. Therefore, it seems that the anoni al
ensemble is not the best hoi e for arrying out the al ulation. No worry, there are other ensembles from whi h to
hoose, and of these, it turns out that the grand anoni al ensemble is signi antly easier to work with. Re all that
in the grand anoni al ensemble, , V and T are the ontrol variables and the partition fun tion is given by
1
X
Z (; V; T ) =  N Q(N; V; T )
N =0
1
X X YY
= eN e fnm"n
N =0 ffnm g m n
P P
Note that the inner sum over o upation numbers is still subje t to the restri tion m n fnm = N . However, there
is a nal sum over all possible values that N , the number that restri ts the sum over o upation numbers, an take on.
Therefore, if we let the sum over o upation numbers be unrestri ted, then they ould sum to any value they liked.
This would be equivalent to performing an unrestri ted sum over o upation numbers without performing the nal
sum over N , sin e in the ourse of summing, unrestri ted, over o upation numbers, we would obtain every possible
value of N as required by the nal sum over N . This is the main advantage of using this ensemble for bosoni and
fermoni systems. Thus, the grand anoni al partition fun tion be omes
X YY
Z (; V; T ) = e( "n )fn
ffnm g m n
Note also that the sum of produ ts is just
0 10 10 1
XXX X X X



e( "1 )f1 e ( "2 )f2 e ( "3 )f3


=  e( "1 )f1 A  e( "1 )f2 A  e( "1 )f3 A



f1 f2 f3 f1 f2 f3
YY X
= e  ( "n )fnm
m n ffnm g

For bosons, ea h individual sum is just the sum of a geometri series. Hen e,
Z (; V; T ) =
YY 1
m n 1 e
 ( "n )

7
whereas, for fermions, ea h individual sum ontains only two terms orresponding to fn = 0 and fn = 1. Thus, for
fermions:
YY 
Z (; V; T ) = 1 + e( "n)
m n
Note that the summands are independent of the quantum number m so that we may perform the produ t over m
values trivially with the result
" #g
Z (; V; T ) =
Y 1
n 1 e( "n )

for bosons and

" #
Y  g
Z (; V; T ) = 1 + e( "n )
n
for fermions, where g = (2s + 1) is the number of eigenstates of Sz (also known as the spin degenera y).
At this point, let us re all the pro edure for al ulating the equation of state in the grand anoni al ensemble. The
free energy in this ensemble is P V=kT given by
PV
kT
= ln Z (; V; T )
and the average parti le number is given by
hN i =   ln Z (; V; T )
The fuga ity  must be eliminated in favor of hN i using the se ond equation and substituted into the rst equation
to yield the equation of state. Re all that, for the lassi al ideal gas,
Z (; V; T ) = eV =3
PV V 
kT
= 3
hN i =   ln Z = V3
Eliminating  in favor hN i is trivial in this ase, leading to the lassi al ideal gas equation
P V = hN ikT
For the ideal gas of identi al fermions, the equations one must solve are
" #g
PV Y
X
"n

kT
= ln Z (; V; T ) = ln 1 + e "n =g ln 1 + e
n n
 X e "n
hN i =   ln Z = g 1 + e "n
n
and for bosons, they are
" #g
PV
= ln Z (; V; T ) = ln
Y 1 = g
X
ln 1 e "n

kT n 1 e "n
n
"n
hN i =   ln Z = g 1 ee "n
X

n
It is not diÆ ult to see that the problem of solving for  in terms of hN i is highly non-trivial for both systems. The
next two le tures will be devoted to just this problem and exploring the ri h behavior that the quantum ideal gases
exhibit.

8
 G25.2651: Statisti al Me hani s

Notes for Le ture 19

I. THE FERMION QUANTUM IDEAL GAS: INTRODUCTION

For an ideal gas of fermions, we had shown that the problem of determining the equation of state was one of solving
two equations
PV X

=g ln 1 + e "n
kT n
X e "n
hN i = g "n
n 1 + e
where the se ond of these must be solved for  in terms of  and substituted into the rst to obtain P as a fun tion
of .
As we did in the Boltzmann ase, let us onsider the thermodynami limit L ! 1 so that the spa ing between
energy levels be omes small. Then the sums an be repla ed by integrals over the ontinuous variable n. For the
pressure, this repla ement give rise to
Z
PV

=g d3 n ln 1 + e "n
kT
Z  
d3 n ln 1 + e 2 h jnj =mL
2 2 2 2
=g
Z 1  
dnn2 ln 1 + e 2 h jnj =mL
2 2 2 2
= 4g
0

Change variables to
s
2 2 h2
x= n
mL2
Then,
PV m
 3=2 1   Z 
= 4gV dxx 2
ln 1 + e x2
kT 2 2  h2 0
Z 1  
4V g
=p 3
2
dxx2 ln 1 + e x
 0
The remaining integral an be evaluated by expanding the log in a power series and integrating the series term by
term:
  1
X ( 1)l+1  l
ln 1 + e x2 = e lx2
l=1
l

PV 1 ( 1)l+1  l
4V g X
Z 1
=p 3
2
dxx2 e lx
kT  l=1 l 0

Vg
1
X
( 1)l+1  l
=
3 l=1 l5=2

By the same te hnique, the average parti le number hN i an be shown to be equal to

1
 1
X ( 1)l+1  l
hN i = V3g l3=2
l=1

Multipling both over these equations on both sides by 3 =V gives

P 3 X (
1 1)l+1  l
=
gkT l=1 l5=2
3 X (
1 1)l+1  l
=
g l=1
l3=2

Although exa t solution of these equations analyti ally is intra table, we will onsider their solutions in two interesting
limits: The high temperature, low density limit and its ounterpart, the low temperature, high density limit.

II. THE HIGH TEMPERATURE, LOW DENSITY LIMIT

Sin e  =  (), in the low density limit, the fuga ity an be expanded in the form
 = a 1  + a 2 2 + a 3 3 +



Writing out the rst few terms in the pressure and density equations, we have
P 3 2 3 4
= + +



gkT 25=2 35=2 45=2
3 2 3 4
= + +



g 23=2 33=2 43=2
Substituting the expansion for  () into the density equation gives
3 1 1
= (a1  + a2 2 + a3 3 +


) (a  + a2 2 + a3 3 +


)2 + 3=2 (a1  + a2 2 + a3 3 +


)3 +



g 23=2 1 3
This equation an now be solved perturbatively, equating like powers of  on both sides. For example, working only
to rst order in , yields:
3 3 3

g
= a1  ) a1 =
g
)   g
When this is substituted into the pressure equation, and only rst order terms in the density are kept, we nd
P 3
=
3
) P
==
hN i
gkT g kT V
whi h is just the lassi al ideal gas equation. Working, now, to se ond order in , we have, from the density equation
3  3  1 6 2
= + a 2 2
g g 23=2g2
or
6
a2 =
23=2 g 2
Thus,
3 6
  g + 23=2 g2 2
and the equation of state be omes

2
 P 3
=  + 5=2 2
kT 2 g
From this, we an read o the se ond virial oeÆ ient
3 3
B2 (T ) =
25=2 g
 0:1768 g >0

It is parti ularly interesting to note that there is a nonzero se ond virial oeÆ ient in spite of the fa t that there are
no intera tions among the parti les. The impli ation is that there is an \ee tive" intera tion among the parti les
as a result of the fermioni spin statisti s. Moreover, this ee tive intera tion is su h that is tends to in rease the
pressure above the lassi al ideal gas result (B2 (T ) > 0). Thus, the ee tive intera tion is repulsive in nature. This is
a onsequen e of the Pauli ex lusion prin iple: The parti le energies must be distributed among the available levels
in su h a way that no two parti les an o upy the same quantum state, thus giving rise to an \ee tive" repulsion
between them.
If we look at the third order orre tion to the pressure, we nd that
 
1 1 9
a3 =
4 33=2 g 3
3   9
  6 2 1 1 
= + 3=2 2  + 3
g 2 g 4 33=2 g 3
3 6  
P   1 2
=  + 5=2 2 + 2 3
kT 2 g g 8 35=2
so that B3 (T ) < 0. Thus, one must go out to third order in the density expansion to nd a ontribution that tends
to de rease the pressure.

III. THE HIGH DENSITY, LOW TEMPERATURE LIMIT

Re all that the density equation ould be expressed as an integral

Z 1 x2 dx
3 = p4g  1 e x2 +1
0

whi h lead to an expansion in powers of  . It is also possible to develop an expansion in powers of ln   mu=kT .
This is a omplished by letting

 = ln  =
kT
and developing an expansion p in powers of p. In order to see how this is done, onsider making a hange of variables
in the integral y = x2 , x = y , dx = dy=(2 y ). Then
Z 1 pydy
3 = p2g ey  + 1
0

Integrate by parts using

1 1
u= y  du = ey  dy
e +1 (ey + 1)2
2
dv = y 1=2 dy v = y 3=2
3
so that
4g
Z 1 y3=2 ey  dy
3 = p
3  (ey  + 1)2
0

3
If we now expand y 3=2 about y =  :
3 3
y 3=2 =  3=2 +  1=2 (y  ) +  1=2 (y  )2 +



2 8
substitute this expansion into the integral and perform the resulting integrals over y , we nd
 
4g 2
 = p (ln  )3=2 + (ln  )
3 1=2
+


+ O(1= )
3  8
where the fa t that =kT  1 has been used owing to the low temperature. Sin e we are in the high density limit,
 () is expe ted to be large as well so that the series, whose error goes as powers of 1= will onverge. As T ! 0,
 ! 1 and only one term in the above expansion survives:
  
2 h2  3=2
 =  3
mkT
 34pg (ln  )3=2 = 34pg kT
Solving for  gives
 2=3
 2 6 2 
h
=
2m g
 0 = "F
whi h is independent of T . The spe ial value of the hemi al potential 0 = (T = 0) is known as the Fermi energy .
To see what its physi al meaning is, onsider the expression for the average number of parti les:
XX e "n
hN i = 1 + e "n
m n

However, re all that

XX
fnm = N
m n
for a spe i number of parti les. Averaging both sides gives
XX
hN i = hfnm i
m n
Comparing these two expressions, we see that the average o upation number of a given state with quantum number
n and m is
 ("n ) 1
hfnm i = 1 +e e  ("n ) =
1 + e ("n )
As T ! 0,  ! 1, and e("n 0 ) ! 1 if "n > 0 , and e("n 0 ) ! 0 if "n < 0 . Thus, at T = 0, we have the result

hfnm i = 0 "n > "F
= ("F "n )
1 "n < "F
A plot of the o upation average o upation number vs. "n at T = 0 is shown in the plot below:

4
 FIG. 1.

Thus, at T = 0, the parti les will exa tly ll up all of the energy levels up to an energy value "F above whi h no
energy levels will be o upied. As T is in reased, the probability of an ex itation above the Fermi energy be omes
nonzero, and the average o upation (shown for several dierent values of  ) appears as follows:

5
 FIG. 2.

Thus, there is a nite probability that some of the levels just above the Fermi energy will be ome o upied as T is
raised slightly above T = 0. At T = 0, the highest o upied energy eigenvalue must satisfy
"n = " F

6
 2 2 h2 2
mL2
jnj = "F
2 2 h2 2
(n + n2y + n2z ) = "F
mL2 x
This denes a spheri al surfa e in n spa e, whi h is known as the Fermi Surfa e. Note that the Fermi surfa e is only
a sphere for the ideal gas. For systems in whi h intera tions are in luded, the Fermi surfa e an be a mu h more
ompli ated surfa e, and studying the properties of this surfa e is a task that o upies the time of many a solid-state
physi ist.

A. Zero-temperature thermodynami s

In order to derive an expression for the average parti le number, re all that
XX XX X
hN i = hfnm i = ("F "n ) = g ("F "n )
m n m n n
In the thermodynami limit, we may take the sum over to an integration:
Z 1
hN i = 4g dnn2 ("F "n )
0

But
2 2 h2 2
"n = n
mL2
Therefore, it proves useful to hange variables of integration from n to "n , using the above relation:
 1=2
mL2
n= "1n=2
2 2 h2
 1=2
1 mL2
dn = "n 1=2
2 2 2 h2
Thus,
Z 1
hN i = 4g dnn2 ("F "n )
0
mL2 3=2 1
  Z
= 2 d"n "1n=2 ("F "n )
2 2 h2 0
 3=2 Z "F
mL2
= 2g d""1=2
2 2 h2 0
 
4g m 3=2
hN i = 3 V "F 3=2
2 h2
In order to derive an expression for the average energy, re all that the energy eigenvalues were given by
XX
Effn g = fnm "n
m n
Therefore, the average energy is given by
XX
hH i = E = hfnm i"n
m n
At T = 0, this be omes

7
 X
E=g ("F "n )"n
Z
n
!g d3 n("F "n )"n
Z 1
= 4g dn n2 ("F "n )"n
0

If the same hange of variables is made, one nds that

Z 1 
1 mL2 3=2 3=2

E = 4g d"n "n ("F "n )
0 2 2 2 h2
  Z "F
m 3=2
= 2g V d"n "n 3=2
2 2 h2 0
 3=2
4g m
= "F 5=2
5 2 2 h2
Thus, the average energy an be seen to be related to the average parti le number by
3
E=
5
hN i"F
whi h is learly not 0 (as it would be lassi ally).
Note that the pressure an be obtained simply in the following way: Re ognize that
PV
1
V g X ( 1)l+1  l
= 3 = ln Z (; V; T )
kT  l=1 l5=2

The energy is given by

 

E= ln Z (; V; T )
 ;V

Thus,
3 Vg X1 ( 1)l+1  l
E=
2 3 l=1 l5=2

Comparing these two equations for the energy and pressure shows that
3 2E
E = PV
2
) P =
3V
Note, that just like the energy, the pressure at T = 0 is not zero. The T = 0 values of both the energy and pressure
are:
3
E=
5
hN i"F
2 hN i
P= "
5 V F
These are referred to as the zero-point energy and pressure and are purely quantum me hani al in nature. The fa t
that the pressure does not vanish at T = 0 is again a onsequen e of the Pauli ex lusion prin iple and the ee tive
repulsive intera tion that also showed up in the low density, high temperature limit. Using the expansion for 3 , we
an derive the thermodynami s in this limit.

8
 B. Thermodynami s at low temperature

Finite temperature thermal orre tions an be obtained by starting with the expansion derived earlier: Note that
 
4g 2
3 = p
3 
(ln  )3=2 + (ln  ) 1=2 +



8
  
4g  3=2  2    1=2
= p + +



3  kT 8 kT
"   !#
4g  
 3=2  2 kT 2
= p 1+ +



3  kT 8 

The term proportional to T 2 is a small thermal orre tion to the T = 0 limit. As su h, it is small and we an repla e
the  appearing there with 0 = "F to the same order in T :
"  2 !#
4g   3=2 2 kT
3 = p 1+ +



3  kT 8 "F

Solving this, now, for  (whi h is equivalent to solving for  ) gives


33 
p 2=3
1
  kT  2 2=3
4g 
1 + 8 kT
2
"F
"  2 #
2 kT
 "F 1
12 "F
+

where the se ond line is obtained by expanding 1=(1 + x)2=3 about x = 0.

In order to obtain the thermal orre tions, one must expand the average o upation number formula about the
0 = "F value using the expansion obtained above for  and the do the integrals. The result is simply
"  2 #
3 5 kT
E = N"F 1 +  2 +



5 12 "F

The thermal orre tion is ne essary in order to obtain the heat apa ity at onstant volume, whi h is given by
 
E
CV =
T V
Using the above expression for the energy, one nds
CV  2 kT
=
hN ik 2"F
From the thermally- orre ted expression for the energy, the pressure an be obtained immediately:
"  2 #
2 5 kT
P = "F 1 +  2 +



5 12 "F

9
 G25.2651: Statisti al Me hani s

Notes for Le ture 20

I. THE IDEAL BOSON GAS: INTRODUCTION

For the bosoni ideal gas, one must solve the equations
PV X

= g ln 1 e "n
kT n
X e "n
hN i = g 1 e "n
n
in order to obtain the equation of state. Examination of these equations, however, shows an immediate problem: The
term n = (0; 0; 0) is divergent both for the pressure and the average parti le number. These terms need to be treated
arefully, and so we split them o from the rest of the sum, giving:
PV X 0

= g ln 1 e "n g ln(1 )
kT n
X0 "n
hN i = g 1 ee "n + g 1  
n
where 0 means that the n = (0; 0; 0) term is ex luded. With these divergent terms split o, the thermodynami
P

limit an be taken and the remaining sums onverted to integrals as was done in the fermion ase. Thus, for the
pressure, we nd
Z
PV

= g dn ln 1 e "n g ln(1 )
kT
Z 1  
e 2 h jnj =mL
2 2 2 2
= 4g dn n2 ln 1 g ln(1 )
0
Z 1
= p4
Vg
3
dx x2 ln(1
2
e x ) g ln(1 )
0

where the hange of variables

s
2 2  h2
x= n
mL2
has been made. Using the expansion
2
1 l
X 2
ln(1 e x ) = e lx
l=1
l

the pressure equation be omes

P 3 X  l
1 3
= ln(1 )
gkT l=1
l5=2 V

and by a similar pro edure, the average parti le number be omes

3 X  l
1 3 
= +
g l=1
l3=2 V 1 

1
In this equation, the term that has been split o represents the average o upation of the ground (n = (0; 0; 0)) energy
state:
hf0 i = 1  
Sin e hf0 i must be greater than or equal to 0, it an be seen that there are restri tions on the allowed values of  .
Firstly, sin e  = exp(),  must be a positive number. However, in order that the average o upation of the ground
state be positive,
0<1
from whi h it follows that
<0
The fa t that as  ! 1 auses hf0 i to diverge will have interesting onsequen es to be dis ussed below. However, let
us rst onsider the low density limit with  << 1.

II. LOW DENSITY, SMALL  LIMIT

In a manner ompletely analogous to what was done for the fermion ase, the low density limit an be treated
by perturbation theory. Note that if  is not lose to 1, then the divergent terms, whi h have a 3 =V prefa tor
a ompanying them, will vanish in the thermodynami limit. Thus, for the pro eeding analysis, these terms an be
negle ted.
As before, we assume the fuga ity an be expanded as
 = a 1  + a 2 2 + a 3 3 +



Then the equation for the density be omes
3 1 1
= (a1  + a2 2 + a3 3 +


) (a1  + a2 2 + a3 3 +


)2 + (a1  + a2 2 + a3 3 +


)3 +



g 23=2 33=2
By equating like powers of  on both sides, the oeÆ ients a1 ; a2 ; a3 ; ::: an be determined as they were for the fermion
gas. Working to rst order in  gives
3 3
a1 =
g
  g
and the equation of state is
P
=
kT
whi h is just the lassi al ideal gas equation. To se ond order, however, we nd
6 3  6
a2 = = 2
23=2 g 2 g 23=2 g 2
and the se ond order equation of state be omes
P 3
= 2
kT 25=2 g
The se ond virial oeÆ ient an be read o and is given by
1 3 0:1768 3
B2 (T ) =  =  <0
25=2 g g
Interestingly, in ontrast to the fermioni system, the pressure is a tually de reased from its lassi al value as a result
of bosoni spin statisti s. Thus, it appears that there is an \ee tive attra tion" between the parti les. This fa t is
not entirely unexpe ted, given that any number of bosons an o upy the same quantum state.

2
 III. THE HIGH DENSITY,  ! 1 LIMIT

The  ! 1 limit is the limit of maximum hemi al potential, whi h is expe ted at high density. However, sin e
 < 0, maximum hemi al potential will be the limit  ! 0. In this limit, the full problem, in luding the divergent
terms, must be solved:
1
P 3 X  l 3
= ln(1  )
gkT l=1 l5=2 V
3 X 1 l 3 
= +
g l=1
l3=2 V 1 

We will need to refer to these two sums often in this se tion, so let us dene them to be
1
X l
g3=2 ( ) =
l=1
l3=2
X1 l
g5=2 ( ) =
l=1
l5=2

Thus, the problem be omes one of solving

P 3 3
= g5=2 ( ) ln(1  )
gkT V
3 3 
= g3=2 ( ) +
g V 1 
We examine, rst the density equation. The se ond term will diverge at  = 1. It is instru tive to ask what is the
behavior of the rst term g3=2 ( ) at  = 1. In fa t g3=2 (1) is nothing but a Riemann zeta-fun tion:
1
X 1
g3=2 (1) = = R(3=2)
l=1
l3=2

In general, a Riemann zeta-fun tion R(n) is given by

1
X 1
R(n) =
l=1
ln

and the values of this fun tion are given in many standard math tables. The parti les value of R(3=2) is approximately
2.612... Moreover, from the form of g3=2 ( ), it is lear that, sin e  < 1, g3=2 (1) is the maximum value of g3=2 ( ). A
plot of g3=2 ( ) is given below:

3
 FIG. 1.

The gure also indi ates that the derivative g30 =2 ( ) diverges at  = 1 despite the fa t that the value of the fun tion
is nite. Note that, sin e  < 1
g3=2 ( ) < g5=2 ( )
It is possible to solve the density equation for  by noting that unless  is very lose to 1, the divergent term will
still vanish in the thermodynami limit as a result of its 3 =V prefa tor. How lose to 1 must it be for this term to
dominate? It an only be dierent from 1 by an amount on the order of 1=V . Thus, let us take  to be of the form
a
=1
V

4
where a is a positive onstant. Substituting this ansatz into the equation for the density gives
3 3 1 a=V
= g3=2 (1 a=V ) +
g V a=V
Sin e g3=2 ( ) does not hange its value mu h if  is displa ed just a little from 1, we an repla e the rst term by
R(3=2). Then,
3 3 1
g
 g3=2 (1) + V a=V
a=V
an be solved for a to yield
3
a = 3
g R(3=2)

where we have negle ted a term 3 =V , whi h vanishes in the thermodynami limit. Sin e a must be positive, this
solution is only valid for 3 =g > R(3=2). For 3 =g < R(3=2),  will be dierent from 1 by more than an amount
1=V so in this regime, the =(1  ) term an be negle ted, leaving the problem of solving 3 =g = g3=2 ( ). Therefore,
the solution for  an be expressed as
8
3 =V 3 > R(3=2)
< 1 3 R(3=2) g
= g
:
root of g3=2 ( ) = 3 3 < R(3=2)
g g

whi h, in the thermodynami limit, be omes

(
1 3 > R(3=2)
= g
root of g3=2 ( ) = 3 3 < R(3=2)
g g

A plot of  vs. v=3 = V=hN i3 is shown below:

5
 ζ

1/2.612 3
V/<N> λ
FIG. 2.

Clearly, point R(3=2) is spe ial, as  undergoes a transition there to a onstant value of 1.
Re all that the o upation of the ground state is

hf0 i = 1  
Thus, for  = 1 a=V , this be omes
V 3
hf0 i  Va = (
3 g
R(3=2))

for 3 =g > R(3=2). At 3 =g = R(3=2) the o upation of the ground state be omes 0. To what temperature does
this orrespond? We an nd this out by solving
3
= R(3=2)
g
 3=2
 2 h2
= R(3=2)
g mkT0
 2=3
 2 h2
kT0 =
gR(3=2) m
so that for temperatures less than T0 the o upation of the ground state be omes
 
hf0 i = Vg 1 g
3
R(3=2)

6
  
=
hNi
1
g
R(3=2)
g 3
"  3=2  3=2 #
=
hN i 1
gR(3=2) mkT kT0
g  2 h2 kT0
"  3=2 #
hN i 1
T
g T0
"  3=2 #
hf0 i = 1 1
T
hN i g T0

Thus, at T = 0

hf0 i = hNg i
whi h is equivalent to

hfn=(0;0;0);mi = hNg i
If we sum both sides over m, this gives
X
hfn=(0;0;0);mi =
X hN i
m m g

hf0 i = hN i
where hf0 i indi ates that the spin degenera y has been summed over. For T > T0 , 3 =g < R(3=2) and  is not
within 1=V of 1. This means that =(1  ) is nite and
hf0 i = 1  ! 0
hN i hN i 1 
as hN i ! 1. Therefore, we have, for the o upation of the ground state:
hf0 i =  1 (T=T0 )3=2 T < T0
hN i 0 T > T0
whi h is shown in the gure below:

7
 FIG. 3.

The o upation of the ground state undergoes a transition from a nite value to 0 at T = T0 and for all higher
temperatures, remains 0. Now, hf0 i=hN i represents the probability that a parti le will be found in the ground state.
It also represents the fra tion of the total number of parti les that will be found in the ground state. For T << T0 ,
this number is very lose to 1, and at T = 0, it be omes exa tly 1, implying that at T = 0 all parti les will be found
in the ground state. This is a phenomenon known as Bose-Einstein ondensation. The o upation number of the
ground state as a fun tion of temperature is shown in the plot below:
Note that there is also a riti al density orresponding to this temperature. This will be given by the solution of
3
= R(3=2)
g
whi h an be solved to yield

8
  3=2
gR(3=2) mkT0
= = gR(3=2)  0
3
2 h2
and the o upation number, expressed in terms of the density is
hf0 i =  1 (0 =)  > 0
hN i 0  < 0
The term in the pressure equation
3
ln(1 )
V
be omes, for  very lose to 1
3
V
ln(V=a)  lnVV
whi h learly vanishes in the thermodynami limit, sin e V  hN i. This allows to dedu e the equation of state as
(
g5=2 (1)
P 3  > 0
= g5=2 ( )
gkT 3  < 0
where  in the above equation omes from the a tual solution of 3 =g = g3=2 ( ). What is parti ularly interesting to
note about the equation of state is that the pressure is independent of the density for  > 0 . Isotherms of the ideal
Bose gas are shown below:

Transition line

kT R( 5 )
3 2
λ

v=V/<N>
v0
FIG. 4.

9
 Here, v0 orresponds to the riti al density 0 . As a fun tion of temperature, we see that P  T 5=2 , whi h is quite
dierent from the lassi al ideal gas. This is also in ontrast to the fermion ideal gas, where as T ! 0 the pressure
remains nite. For the Boson gas, as T ! 0 the pressure vanishes, in keeping with the notion of an \ee tive"
attra tion between the parti les.
Other thermodynami quantities an be determined in a similar manner. The energy an be obtained from E =
3P V=2 straightforwardly:
 3 kT V
E= 2 3 g5=2 (1)  > 0 ; T < T 0
3 kT V
2 3 g5=2 ( )  < 0 ; T > T0
and the heat apa ity at onstant volume from
 
E
CV =
T V
whi h gives
(
15 g5=2 (1) T < T0
CV 4 3
hN iK = 15 g5=2 ( )
4 3
9 g3=2 ( )
4 g1=2 ( ) T > T0

A plot of the heat apa ity exhibits a usp at T = T0 :

CV
Nk

3/2

3/2
~T

T
T0
FIG. 5.

Experiments arried out on liquid He4 , whi h has been observed to undergo Bose-Einstein ondensation at around
T =2.18 K, have measured an a tual dis ontinuity in the heat apa ity at the transition temperature, suggesting that
Bose-Einstein ondensation is a phase transition known as the  transition. The experimental heat apa ity is shown
roughly below:

10
 CV
Nk

T
1 2.18K
FIG. 6.

By ontrast, the ideal Bose gas undergoes a rst order phase transition. However, using the mass and density of
liquid He4 in the expression for T0 given above, one would predi t that T0 is about 3.14 K, whi h is not far o the
experimental transition temperature of 2.18 K for real liquid helium.
For ompleteness, other thermodynami properties of the ideal Bose gas are given as follows: The entropy is
 5 1
S 2 3 g5=2 (1) T < T0
hN iK = 5 1
2 3 g5=2 ( ) ln  T > T0
The Gibbs free energy is given by

G 0 T < T0
hN iK = ln  T > T0
It is lear from the analysis of this and the fermion ideal gas that quantum statisti s give rise to an enormously ri h
behavior, even when there are no parti le intera tions!

11
 G25.2651: Statistical Mechanics

Notes for Lecture 21

I. CLASSICAL LINEAR RESPONSE THEORY

Consider Hamilton's equations in the form
q_i = @H
@pi
p_i = ? @H
@qi
We noted early in the course that an ensemble of systems evolving according to these equations of motion would
generate an equilibrium ensemble (in this case, microcanonical). Recall that the phase space distribution function
f (?; t) satised a Liouville equation:
@f + iLf = 0
@t
where iL = f:::; H g. We noted that if @f=@t = 0, then f = f (H ) is a pure function of the Hamiltonian which dened
the general class of distribution functions valid for equilibrium ensembles.
What does it mean, however, if @f=@t 6= 0? To answer this, consider the problem of a simple harmonic oscillator.
In an equilibrium ensemble of simple harmonic oscillators at temperature T , the members of the ensemble will
undergo oscillatory motion about the potential minimum, with the amplitude of this motion determined by the
temperature. Now, however, consider driving each oscillator with a time-dependent driving force F (t). Depending on
how complicated the forcing function F (t) is, the motion of each member of the ensemble will, no longer, be simple
oscillatory motion about the potential minimum, but could be a very complex kind of motion that explores large
regions of the potential energy surface. In other words, the ensemble of harmonic oscillators has been driven away
from equilibrium by the time-dependent force F (t). Because of this nonequilibrium behavior of the ensemble, averages
over the ensemble could become time-dependent quantities rather than static quantities. Indeed, the distribution
function f (?; t), itself, could be time-dependent. This can most easily be seen by considering the equation of motion
for a forced oscillator
mx = ?m!2x + F (t)
The solution now depends on the entire history of the forcing function F (t), which can introduce explicit time-
dependence into the ensemble distribution function.

A. Generalized equations of motion

The most general way a system can be driven away from equilibrium by a forcing function Fe (t) is according to the
equations of motion:
q_i = @H
@pi + Ci (?)Fe (t)
p_i = ? @H
@pi + Di (?)Fe (t)
where the 3N functions Ci and Di are required to satisfy the incompressibility condition
3N 
X 
@Ci + @Di = 0
i=1 @qi @pi

1
in order to insure that the Liouville equation for f (?; t) is still valid. These equations of motion will give rise to a
distribution function f (?; t) satisfying
@f + iLf = 0
@t
R
with @f=@t 6= 0. (We assume that f is normalized so that d?f (?; t) = 1.)
What does the Liouville equation say about the nature of f (?; t) in the limit that Ci and Di are small, so that
the displacement away from equilibrium is, itself, small? To examine this question, we propose to solve the Liouville
equation perturbatively. Thus, let us assume a solution of the form
f (?; t) = f0(H (?)) +
f (?; t)
Note, also, that the equations of motion ?_ take a perturbative form
?_ (t) = ?_ 0 +
?_ (t)
and as a result, the Liouville operator contains two pieces:
iL = ?_
r? = ?_ 0
r? +
?_
r? = iL0 + i
L
where iL0 = f:::; H g and f0 (H ) is assumed to satisfy
iL0f0 (H (?)) = 0
?_ 0 means the Hamiltonian part of the equations of motion
q_i = @H
@pi
p_i = ? @H
@qi
For an observable A(?), the ensemble average of A is a time-dependent quantity:
Z
hA(t)i = d?A(?)f (?; t)
which, when the assumed form for f (?; t) is substituted in, gives
Z Z Z
hA(t)i = d?A(?)f (?) + d?A(?)
f (?; t) = hAi + d?A(?)
f (?; t)
0 0

where h
i0 means average with respect to f0 (?).

B. Perturbative solution of the Liouville equation

Substituting the perturbative form for f (?; t) into the Liouville equation, one obtains
@
@t (f0 (?) +
f (?; t)) + (iL0 + i
L(t))(f0 (?) +
f (?; t)) = 0
Recall @f0 =@t = 0. Thus, working to linear order in small quantities, one obtains the following equation for
f (?; t):
@ 
@t + iL0
f (?; t) = ?i
Lf0(?)
which is just a rst-order inhomogeneous dierential equation. This can easily be solved using an integrating factor,
and one obtains the result
Zt

f (?; t) = ? dse?iL0 (t?s) i
L(s)f0(?)
0

2
Note that
i
Lf0(?) = iLf0(?) ? iL0f0 (?) = iLf0(?) = ?_
r? f0 (?)
But, using the chain rule, we have
?_
r? f0 (?) = ?_
@f0 @H
@H @ ?
@f XN  @H @H

p_ i

= @H @ pi + q_ i @ qi
0

i=1
@f XN p p 
= @H 0 i i

(Fi + Di Fe (t)) ? Fi
m + Ci Fe (t)
i=1 mi i
XN  
= @f 0
@H i=1 i D (? )
pi ? C (?)
F F (t)
mi i i e

where Fi (q1 ; :::; qN ) is the force on the ith particle. Dene

N 
X pi

j (?) = Ci (?)
Fi ? Di (?)
m
i=1 i
called the dissipative ux. Then
@f
?_
r? f0 (?) = ? 0 j (?)Fe (t)
@H
Now, suppose f0 (?) is a canonical distribution function
f0 (H (?)) = Q(N;1V; T ) e?H (?)
then
@f0 = ?f (H )
@H 0

so that
?_
r? f0 (?) = f0 (?)j (?)Fe (t)
Thus, the solution for
f (?; t) is
Zt

f (?; t) = ? dse?iL0 (t?s) f0 (?)j (?)Fe (s)
0

The ensemble average of the observable A(?) now becomes

Z Zt
hA(t)i = hAi ?  d?A(?)
0 dse?iL0 (t?s) f0 (?)j (?)Fe (s)
0
Zt Z
= hAi0 ?  ds d?A(?)e?iL0 (t?s) f0 (?)j (?)Fe (s)
0
Zt Z
= hAi0 ?  ds d?f0 (?)A(?)e?iL0 (t?s) j (?)Fe (s)
0

Recall that the classical propagator is exp(iLt). Thus the operator appearing in the above expression is a classical
propagator of the unperturbed system for propagating backwards in time to ?(t ? s). An observable A(?) evolves in
time according to

3
 dA = iLA
dt
A(t) = eiLtA(0)
A(?t) = e?iLtA(0)
Now, if we take the complex conjugate of both sides, we nd
A (t) = A (0)e?iLt
where now the operator acts to the left on A (0). However, since observables are real, we have
A(t) = A(0)e?iLt
which implies that forward evolution in time can be achieved by acting to the left on an observable with the time
reversed classical propagator. Thus, the ensemble average of A becomes
Zt Z
hA(t)i = hAi ? 
0 dsFe (s) d?0 f0 (?0 )A(?t?s (?0 ))j (?0 )
0
Zt
= hAi0 ?  dsFe (s)hj (0)A(t ? s)i0
0

where the quantity on the last line is an object we have not encountered yet before. It is known as an equilibrium
time correlation function. An equilibrium time correlation function is an ensemble average over the unperturbed
(canonical) ensemble of the product of the dissipative ux at t = 0 with an observable A evolved to a time t ? s.
Several things are worth noting:
1. The nonequilibrium average hA(t)i, in the linear response regime, can be expressed solely in terms of equilibrium
averages.
2. The propagator used to evolve A(?) to A(?; t ? s) is the operator exp(iL0 (t ? s)), which is the propagator for
the unperturbed, Hamiltonian dynamics with Ci = Di = 0. That is, it is just the dynamics determined by H .
3. Since A(?; t ? s) = A(?(t ? s)) is a function of the phase spaceRvariables evolved to a time t ? s, we must now
specify over which set of phase space variables the integration d? is taken. The choice is actually arbitrary,
and for convenience, we choose the initial conditions. Since ?(t) is a function of the initial conditions ?(0), we
can write the time correlation function as
Z
hj (0)A(t ? s)i = 1 d? e?H (?0 ) j (? )A(? (? ))
0
Q 0 0 t?s 0

C. General properties of time correlation functions

Dene a time correlation function between two quantities A(?) and B (?) by
CAB (t) = hZA(0)B (t)i
= d?f (?)A(?)eiLt B (?)
The following properties follow immediately from the above denition:
hA(0)B (t)i = hA(?t)B (0)i
1.
CAB (0) = hA(?)B (?)i
2. Thus, if A = B , then

4
 CAA (t) = hA(0)A(t)i
known as the autocorrelation function of A, and
CAA (0) = hA2 i
If we dene A = A ? hAi, then
CAA (0) = h(A)2 i = h(A ? hAi)2 i = hA2 i ? hAi2
which just measures the uctuations in the quantity A.
3. A time correlation function may be evaluated as a time average, assuming the system is ergodic. In this case,
the phase space average may be equated to a time average, and we have
1 Z T ?t
CAB (t) = Tlim
!1 T ? t
dsA(?(s))B (?(t + s))
0

which is valid for t << T . In molecular dynamics simulations, where the phase space trajectory is determined
at discrete time steps, the integral is expressed as a sum
?k
NX
CAB (k
t) = N 1? k A(?k )B (?k+j ) k = 0; 1; 2; :::; Nc
j=1

where N is the total number of time steps,
t is the time step and Nc << N .
4. Onsager regression hypothesis: In the long time limit, A and B eventually become uncorrelated from each other
so that the time correlation function becomes
CAB (t) = hA(0)B (t)i ! hAihB i
For the autocorrelation function of A, this becomes
CAA (t) ! hAi2
Thus, CAA (t) decays from hA2 i at t = 0 to hAi2 as t ! 1.
An example of a signal and its time correlation function appears in the gure below. In this case, the signal is the
magnitude of the velocity along the bond of a diatomic molecule interacting with a Lennard-Jones bath. Its time
correlation function is shown beneath the signal:

5
 FIG. 1.

Over time, it can be seen that the property being autocorrelated eventually becomes uncorrelated with itself.

6
 II. TIME CORRELATION FUNCTIONS AND TRANSPORT COEFFICIENTS
A. The shear viscosity
The shear viscosity of a system measures is resistance to ow. A simple ow eld can be established in a system by
placing it between two plates and then pulling the plates apart in opposite directions. Such a force is called a shear
force, and the rate at which the plates are pulled apart is the shear rate. A set of microscopic equations of motion
for generating shear ow is
q_ = pi + y x^
i i
mi
p_ i = Fi ? py x^
i

where is a parameter known as the shear rate. These equations have the conserved quantity
N
X
H0 = (pi + mi yi x^ )2 + U (q1 ; ::; qN )
i=1
The physical picture of this dynamical system corresponds to the presence of a velocity ow eld v(y) = yx^ shown
in the gure.
The ow eld points in the x^ direction and increases with increasing y-value. Thus, layers of a uid, for example,
will slow past each other, creating an anisotropy in the system. From the conserved quantity, one can see that the
momentum of a particle is the value of pi plus the contribution from the eld evaluated at the position of the particle
pi ! pi + mi v(yi )

7
 FIG. 2.

Such an applied external shearing force will create an asymmetry in the internal pressure. In order to describe
this asymmetry, we need an analog of the internal pressure that contains a dependence on specic spatial directions.
Such a quantity is known as the pressure tensor and can be dened analogously to the isotropic pressure P that we
encountered earlier in the course. Recall that an estimator for the pressure was
1 XN  p2 
p = 3V i + qi
Fi
i=1 mi
and P = hpi in equilibrium. Here, V is the volume of the system. By analogy, one can write down an estimator for
the pressure tensor p :
XN  (p
e^ )(p
e^ ) 
p = V1 i  i  + (q
e^ )(F
e^ )
mi i  i 
i=1
and
P = hp i
where e^ is a unit vector in the  direction,  = x; y; z . This (nine-component) pressure tensor gives information
about spatial anisotropies in the system that give rise to o-diagonal pressure tensor components. The isotropic
pressure can be recovered from

8
 X
P = 13 P

which is just 1/3 of the trace of the pressure tensor. While most systems have diagonal pressure tensors due to
spatial isotropy, the application of a shear force according to the above scheme gives rise to a nonzero value for the
xy component of the pressure tensor Pxy . In fact, Pxy is related to the velocity ow eld by a relation of the form
Pxy = ? @v x
@y = ?
where the coecient  is known as the shear viscosity and is an example of a transport coecient. Solving for  we
nd
 = ? Pxy = ? lim hpxy (t)i
t!1
where hpxy (t)i is the nonequilibrium average of the pressure tensor estimator using the above dynamical equations of
motion.
Let us apply the linear response formula to the calculation of the nonequilibrium average of the xy component of
the pressure tensor. We make the following identications:
Fe (t) = 1 Ci (?) = yi x^ Di (?) = ? py x^
i

Thus, the dissipative ux j (?) becomes

N 
X pi

j (?) = Ci
F i ? D i
m
i=1 i
XN  
= i
x^
yi (Fi
x^ ) + py pm
i=1 ii

XN  (p
y^ )(p
x^ ) 
= i i + (qi
y^ )(Fi
x^ )
i=1 mi
= V pxy
According to the linear response formula,
Zt
hpxy (t)i = hpxy i ?  V
0 dshpxy (0)pxy (t ? s)i0
0

so that the shear viscosity becomes

 hp i Zt 
xy
!1 ? + V 0 dshpxy (0)pxy (t)i0
 = tlim 0

Recall that h


i0 means average of a canonical distribution with = 0. It is straightforward to show that hpxy i0 = 0
for an equilibrium canonical distribution function. Finally, taking the limit that t ! 1 in the above expression gives
the result
V Z 1 dthp (0)p (t)i
 = kT xy xy 0
0

which is a relation between a transport coecient, in this case, the shear viscosity coecient, and the integral of
an equilibrium time correlation function. Relations of this type are known as Green-Kubo relations. Thus, we have
expressed a new kind of thermodynamic quantity to an equilibrium time correlation function, which, in this case, is
an autocorrelation function of the xy component of the pressure tensor.

9
 B. The diusion constant
The diusive ow of particles can be studied by applying a constant force f to a system using the microscopic
equations of motion
q_ = pi i mi
p_ i = Fi (q1 ; ::; qN ) + f x^
which have the conserved energy
XN p2 N
H0 = i + U (q1 ; :::; qN ) ? f X xi
i=1 2mi i=1
Since the force is applied in the x^ direction, there will be a net ow of particles in this direction, i.e., a current Jx .
Since this current is a thermodynamic quantity, there is an estimator for it:
N
X
ux = ? x_ i
i=1
and Jx = hux i. The constant force can be considered as arising from a potential eld
(x) = ?xf
The potential gradient @=@x will give rise to a concentration gradient @c=@x which is opposite to the potential
gradient and related to it by
@c = ? 1 @
@x
kT @x
However, Fick's law tells how to relate the particle current Jx to the concentration gradient
Jx = ?D @x@c = D @ = ? D f
kT @x kT
where D is the diusion constant. Solving for D gives
D = ?kT Jfx = ?kT tlim hux (t)i
!1 f
Let us apply the linear response formula again to the above nonequilibrium average. Again, we make the identication:
Fe (t) = 1 Di = f x^ Ci = 0
Thus,
Zt N ! N !
X X
hux (t)i = hux i ? 
0 dsf h x_ i (0) x_ i (t ? s) i0
0 i=1 i=1
Zt X
= hux i0 ? f ds hx_ i (0)x_ j (t ? s)i0
0 i;j
In equilibrium, it can be shown that there are no cross correlations between dierent particles. Consider the initial
value of the correlation function. From the virial theorem, we have
hx_ i x_ j i = ij hx_ i i
0
2
0

which vanishes for i 6= j . In general,

hx_ i (0)x_ j (t)i = ij hx_ i (0)x_ i (t ? s)i
0 0

10
Thus,
Zt XN
hux (t)i = hux i ? f
0 ds x_ i (0)x_ i (t ? s)i0
0 i=1
In equilibrium, hux i0 = 0 being linear in the velocities (hence momenta). Thus, the diusion constant is given by,
when the limit t ! 1 is taken,
Z 1X
N
D= hx_ i (0)x_ i (t)i 0
0 i=1
However, since no spatial direction is preferred, we could also choose to apply the external force in the y or z directions
and average the result over the these three. This would give a diusion constant
Z1 XN
D = 31 dt hq_ i (0)
q_ i (t)i0
0 i=1
The quantity
N
X
hq_ i (0)
q_ i (t)i 0
i=1
is known as the velocity autocorrelation function, a quantity we will encounter again in other contexts.

11
 G25.2651: Statisti al Me hani s

Notes for Le ture 22

I. QUANTUM TIME CORRELATION FUNCTIONS AND SPECTRA

A. The Hamiltonian

Consider a quantum system with a Hamiltonian H0 . Suppose this system is subje t to an external driving for e
Fe (t) su h that the full Hamiltonian takes the form
H = H0 BFe (t) = H0 + H 0
where B is an operator through whi h this oupling o urs. This is the situation, for example, when the infrared
spe trum is measured experimentally { the external for e Fe (t) is identied with an ele tri eld E (t) and B is
identied with the ele tri dipole moment operator. If the eld Fe (t) is inhomogeneous, then H takes the more
general form
Z X
H = H0 d3 x B (x)Fe (x; t) = H0 Bk Fe;k (t)
k
where the sum is taken over Fourier modes. Often, B is an operator su h that, if Fe (t) = 0, then
H

hB i = TrTr Be
(e H )
Suppose we take Fe (t) to be a mono hromati eld of the form
Fe (t) = F! ei!t
Generally, the external eld an indu e transitions between eigenstates of H0 in the system. Consider su h a transition
between an initial state jii and a nal state jf i, with energies Ei and Ef , respe tively:
H0 jii = Ei jii
H0 jf i = Ef jf i
(see gure below).

FIG. 1.

1
This transition an only o ur if
Ef = Ei + h!

B. The transition rate

In the next le ture, we will solve the quantum Liouville equation


ih = [H; ℄
t
perturbatively and derive quantum linear response theory. However, the transition rate an a tually be determined
dire tly within perturbation theory using the Fermi Golden Rule approximation, whi h states that the probability of
a transition's o uring per unit time, Ri!f , is given by
2 2
Ri!f (! ) = jhf jH 0 jiij2 Æ (Ef Ei h! ) = jF! j2 jhf jB jiij2 Æ (Ef Ei h! )

h h
The Æ-fun tion expresses the fa t that energy is onserved. This des ribes the rate of transitions between spe i
states jii and jf i. The transition rate between any initial and nal states an be obtained by summing over both i
and f and weighting the sum by the probability that the system is found in the initial state jii:
X
P (! ) = Ri!f (! )wi
i;f

where wi is an eigenvalue of the density matrix, whi h we will take to be the anoni al density matrix:
e Ei
wi =
Tr (e H )
Using the expression for Ri!f (!), we nd
2 X
P (! ) =

h
jF ! j2 wi jhijB jf ij2 Æ (Ef Ei h!)
i;f

Note that
2 X
P ( !) =

h
jF! j2 wi jhijB jf ij2 Æ (Ef Ei + h! )
i;f

This quantity orresponds to a time-reversed analog of the absorption pro ess. Thus, it des ribes an emission event
jii ! jf i with Ef = Ei h !, i.e., emission of a photon with energy h !. If an also be expressed as a pro ess jf i ! jii
by re ognizing that
e Ef e  (Ei h !)
wf = =
Tr (e H ) Tr (e H )
or
wf = eh ! wi ) wi = e h ! wf
Therefore
2 X
P ( !) = jF j2 e
 !
h
!
h wf jhijB jf ij2 Æ (Ef Ei + h! )
i;f

If we now inter hange the summation indi es, we nd

2
 2 X
P ( !) = jF j2 e
 !
h
!
h wi jhijB jf ij2 Æ (Ei Ef + h! )
i;f
2 X
=

h
jF! j2 wi jhijB jf ij2 e  ! Æ (E
h
i Ef h!)
i;f
where the fa t that Æ(x) = Æ( x) has been used. Comparing this expression for P ( !) to that for P (!), we nd
P ( ! ) = e h ! P (! )
whi h is the equation of detailed balan e. We see from it that the probability of emission is less than that for absorption.
The reason for this is that it is less likely to nd the system in an ex ited state jf i initially, when it is in onta t
with a heat bath and hen e thermally equilibrated. However, we must remember that the mi ros opi laws of motion
(Newton's equations for lassi al systems and the S hrodinger equation for quantum systems) are reversible. This
means that
Ri!f (! ) = Rf !i ( ! )
The on lusion is that, sin e P (!) > P ( !), reversibility is lost when the system is pla ed in onta t with a heat
bath, i.e., the system is being driven irreversibly in time.
Dene
X
C> (! ) = wi jhijB jf ij2 Æ (Ef Ei h
 !)
i;f
X
C< (! ) = wi jhijB jf ij2 Æ (Ef Ei + h! )
i;f
then
C< (! ) = e h ! C> (! )
Now using the fa t that the Æ-fun tion an be written as
1 1
Z
Æ (E ) = dte iEt
2 1
C> (! ) be omes
Z 1
1 X
C> (! ) = dt w jhijB jf ij2 e i(Ef Ei h !)t=h
2h 1 i;f i
1 1
Z X
= dt ei!t wi jhijB jf ij2 e i(Ef Ei )t=h
2h 1 i;f
Z 1
1 X
= dt ei!t wi hijB jf ihf jB jiie iEf t=h eiEi t=h
2h 1 i;f
Z 1
1 X
= dt ei!t wi hijeiH0 t=h Be iH0 t=h jf ihf jB jii
2h 1 i;f
Re all that the evolution of an operator in the Heisenberg pi ture is given by
B (t) = eiH0 t=h Be iH0 t=h
if the evolution is determined solely by H0 . Thus, the expression for C> (!) be omes
1 1
Z X
C> (! ) = dt ei!t wi hijB (t)jf ihf jB jii
2h 1 i;f
1 1
Z X
= dt ei!t wi hijB (t)B (0)jii
2h 1 i
1 1 i!t
Z
= e Tr [B (t)B (0)℄
2h 1
Z 1
1
= ei!t hB (t)B (0)i
2h 1

3
whi h involves the quantum auto orrelation fun tion hB (t)B (0)i.
In general, a quantum time orrelation fun tion in the anoni al ensemble is dened by
 H 
Tr A(t)B (0)e
CAB (t) =
Tr [e H ℄
In a similar manner, we an show that
1
Z 1
C< (! ) = dt ei!t hB (0)B (t)i 6= C> (! )
2h 1
sin e
[B (0); B (t)℄ 6= 0
in general. Also, the produ t B (0)B (t) is not Hermitian. However, a hermitian ombination o urs if we onsider the
energy dieren e between absorption and emission. The energy absorbed per unit of time by the system is P (!)h!,
while the emitted into the bath by the system per unit of time is P ( !)h!. The energy dieren e Q(!) is just
Q(! ) = [P (! ) P ( ! )℄h!
= P (!)[1 e h ! ℄h!
= 2!jF! j2 C> (!)[1 e h ! ℄
But sin e
C< (! ) = e h ! C> (! )
it follows that


C> (! ) + C< (! ) = 1 + e h ! C> (! )
or
C> (! ) + C< (! )
C> (! ) =
1 + e h !
Note, however, that
1
Z 1
C> (! ) + C< (! ) = dtei!t hB (t)B (0) + B (0)B (t)i
2h 1
1 1
= dt ei!t h [B (0); B (t)℄+ i

h 2
where [:::; :::℄+ is known as the anti ommutator: [A; B ℄+ = AB + BA. The anti ommutator between two operators
is, itself, hermitian. Therefore, the energy dieren e is
1 e h ! 1
Z
2!
Q(! ) = jF! j 2 dt ei!t h[B (0); B (t)℄+ i

h 1 + e h ! 1
Z 1
2!
= jF! j tanh( h !=2)
2 dt ei!t h[B (0); B (t)℄+ i

h 1
The quantity h[B (0); B (t)℄+ i is the symmetrized quantum auto orrelation fun tion. The lassi al limit is now manifest
(tanh( h!=2) !  h!=2):
Z 1
Q(! ) ! jF! j2 dt ei!t hB (0)B (t)i
1
The lassi ally, the energy spe trum Q(!) is dire tly related to the Fourier transform of a time orrelation fun tion.

4
 C. Examples

Dene
1 1
Z
1
G(! ) = dtei!t h [B (0); B (t)℄+ i
2 1 2
whi h is just the frequen y spe trum orresponding to the auto orrelation fun tion of B . For dierent hoi es of B ,
G(! ) orresponds to dierent experimental measurements.
Consider the example of a mole ule with a transition dipole moment ve tor . If an ele tri eld E(t) is applied,
then the Hamiltonian H 0 be omes
H 0 = 
E(t)
If we take E(t) = E (t)^z, then
H 0 = z E (t)
Identifying B = z , the spe trum be omes
1 1
Z
1
G(! ) = dt ei!t h [z (0); z (t)℄+ i
2 1 2
or for a general ele tri eld, the result be omes
1 1
Z
1
G(! ) = dt ei!t h ((0)
(t) + (t)
(0))i
2 1 2
These spe tra are the infrared spe tra.
As another example, onsider a blo k of material pla ed in a magneti eld H(t) in the z dire tion. The spin Sz of
ea h parti le will ouple to the magneti eld giving a Hamiltonian H 0
N
X
H0 = Si;z H(t)
i=1
The net magnetization reated by the eld mz is given by
N
1 X
mz = S
N i=1 i;z
so that
H 0 = Nmz H(t)
Identify B = mz (the extra fa tor of N just expresses the fa t that H 0 is extensive). Then the spe trum is
1 1
Z
1
G(! ) = dt ei!t h [mz (0); mz (t)℄+ i
2 1 2
whi h is just the NMR spe trum. In general for ea h orrelation fun tion there is a orresponding experiment that
measures its frequen y spe trum.
To see what some spe i lineshapes look like, onsider as an ansatz a pure exponential de ay for the orrelation
fun tion CBB (t):
CBB (t) = hB 2 ie jtj
The spe trum orresponding to this time orrelation fun tion is
1 1
Z
G(! ) = dtei!t CBB (t)
2 1
and doing the integral gives
G(! ) =
hB 2 i
 !2 + 2
whi h is shown in the gure below:

5
 FIG. 2.

We see that the lineshape is a Lorentzian with a width .

As a further example, suppose CBB (t) is a de aying os illatory fun tion:
CBB (t) = hB 2 ie jtj os !0 t
whi h des ribes well the behavior of a harmoni diatomi oupled to a bath. The spe trum an be shown to be
 
G(! ) =
hB 2 i 2 + !2 + !2
0
 ( 2 + (! !0 )2 ) ( 2 + (! + !0 )2 )
p
whi h ontains two peaks at ! =  !02 2 as shown in the gure below:

6
FIG. 3.

7
 G25.2651: Statisti al Me hani s

Notes for Le ture 23

I. QUANTUM LINEAR RESPONSE THEORY

Consider again the Hamiltonian for a system oupled to a time-dependent eld

H = H0 BFe (t)
We wish to solve the quantum Liouville equation

ih = [H; ℄
t
in the linear regime where Fe (t) is small.

A. Perturbative solution of the Liouville equation

As in the lassi al ase, we assume a solution of the form

(t) = 0 (H0 ) +
(t)
where
[H0; 0℄ = 0 ) t0 = 0
and we will assume
e H0
0 (H0 ) =
Q(N; V; T )
Substituting into the Liouville equation and working to rst order in small quantities, we nd

 1 [H0;
℄ 1 [B; 0 ℄Fe (t)
t
= 
ih ih
whi h is a rst order inhomogeneous equation that an be solved by using an integrating fa tor:

(t) = i1h
Z t
ds e iH0 (t s)=h [B; 0 ℄eiH0 (t s)=h Fe (s)
1
(Note that we have hosen the origin in time to be t = 1, whi h is an arbitrary hoi e.)
For an observable A, the expe tation value is
hA(t)i = Tr(A) = hAi0 + Tr(
(t)A)
when the solution for
 is substituted in, this be omes
hA(t)i = hAi0 ih 1 Z t h i
ds Tr Ae iH0 (t s)=hbar [B; 0 ℄eiH0 (t s)=h Fe (s)
1
= hAi0 ih 1 Z t h i
ds Tr eiH0 (t s)=h Ae iH0 (t s)=hbar [B; 0 ℄ Fe (s)
1
1
= hAi0 ih ds Tr [A(t s)[B; 0℄℄ Fe (s)
1
where the y li property of the tra e has been used and the Heisenberg evolution for A has been substituted in.
Expanding the ommutator gives
hA(t)i = hAi0 i1h
Z t
ds Tr[A(t s)B0 A(t s)0 B ℄ Fe (s)
1
= hAi0 i1h
Z t
ds Tr[0 (A(t s)B BA(t s))℄ Fe (s)
1
= hAi0 ih1 Z t
ds Fe (s)h[A(t s); B (0)℄0 i
1
where the y li property of the tra e has been used again. Dene a fun tion
AB (t) = hi h[A(t); B(0)℄i0
alled the after ee t fun tion. It is essentially the antisymmetri quantum time orrelation fun tion, whi h involves
the ommutator between A(t) and B(0). Then the linear response result an be written as
Z t
hA(t)i = hAi0 + dsFe (s)AB (t s)
1
whi h is the starting point for the theory of quantum transport oeÆ ients. If we hoose to measure the operator B,
then we nd
Z t
hB (t)i = hB i0 + ds Fe (s)BB (t s)
1

B. Relation to spe tra

Suppose that Fe (t) is a mono hromati eld

Fe (t) = F! ei!t et
where the parameter  insures that eld goes to 0 at t = 1. We will take  ! 0+ at the end of the al ulation. The
expe tation value of B then be omes
Z t
hB (t)i = hB i0 + ds BB (t s)F! ei!s es
1 Z 1
= hBi0 + F! e (i! +)t
d BB ( )e i(! i)
0

where the hange of integration variables  = t s has been made.

Dene a frequen y-dependent sus eptibility by
Z 1
BB (! i) = d BB ( )e i(! i)
0

then
hB (t)i = hB i0 + F! ei!t etBB (! i)
If we let z = ! i, then we see immediately that
Z 1
BB (z ) = d BB ( )e iz
0

i.e., the sus eptibility is just the Lapla e transform of the after ee t fun tion or the time orrelation fun tion.
2
 Re all that
AB (t) = hi h[A(t); B(0)℄i0 = hi h[eiH0 t=h Ae iH0 th

;B ℄i0
Under time reversal, we have
h i
AB ( t) = hi h e iH0 t=h AeiH0 t=h ; B i0
 
= hi h e iH0 t=h AeiH0 t=h B Be iH0 t=h AeiH0 t=h i0
 
= hi h AeiH0 t=h Be iH0 t=h eiH0 t=h Be iH0 t=h A i0
= hi h(AB(t) B(t)A)i0
= hi h[B(t); A℄i = BA(t)
Thus,
AB ( t) = BA(t)
and if A = B, then
BB ( t) = BB (t)
Therefore
Z 1
BB (! ) = lim dt e i(! it) BB (t)
!0+ 0
Z 1
= !lim0+ dt e t [BB (t) os !t iBB (t) sin !t℄
0
= Re(BB (!)) iIm(BB (!))
From the properties of BB (t) it follows that
Re(BB (!) = Re(BB ( !)
Im(BB (!) = Im(BB ( !)
so that Im(BB (!)) is positive for ! > 0 and negative for ! < 0. It is a straightforward matter, now, to show that
the energy dieren e Q(!) derived in the le ture from the Fermi golden rule is related to the sus eptibility by
Q(! ) = 2! jF! j2 Im(BB (! ))

C. Kubo transform expression for the time orrelation fun tion (Optional)

We shall derive the following expression for the quantum time orrelation fun tion
Z 
AB (t) = d hB_ ( ih)A(t)i0
0

known as a Kubo transform relation. Sin e B_ is given by the Heisenberg equation:

1
B_ = [B; H0 ℄

ih
it follows that

3
 B_ (t) =
1 eiH0 t=h [H ; B(0)℄e iH0 t=h


ih 0

Evaluating the expression at t = ih  gives

1
B_ ( ih) = eH0 [B (0); H0 ℄e H0

ih
Thus,
 
AB (t) = dheH0 i1h [B(0); H0 ℄ e H0 A(t)i0
Z 

By performing the tra e in the basis of eigenve tors of H0, we obtain

 
AB (t) = Q1 d hnjeH0 i1h [B(0); H0 ℄e H0 A(t)jnie En
Z  X

0 n
 
= Q1 d hnjeH0 i1h [B(0); H0 ℄e H0 jmihmjA(t)jnie En
Z  X

0 m;n

= Q1 d eEn e Em i1h hnj[B(0); H0℄jmihmjA(t)jnie En

Z  X

0 m;n

= Q e En e (En Em) 1 i1h hnj[B(0); H0 ℄jmihmjA(t)jnie En

1 X  (En Em )

m;n

But
hnj[B (0); H0 ℄jmi = hnjB (0)H0 H0 B (0)jmi = (Em En )hnjB (0)jmi
Therefore,
AB (t) = ih1Q e En e Em
hnjB(0)jmihmjA(t)jni
X

m;n
" #
1 X
= ih Q e m hmjA(t)jnihnjB(0)jmi
E
X
e E n hnjB (0)jmihmjA(t)jni
m;n m;n

= hi h[A(t); B(0)℄i0

whi h proves the relation. The lassi al limit an be dedu ed easily from the Kubo transform relation:
AB (t) ! hB_ (0)A(t)i0
Note further, by using the yli properties of the tra e, that
hB_ ( ih)B (t)i0 = d hB ( ih)B (t)i0
dt

D. The Onsager u tuation regression theorem (Optional)

Suppose that Fe (t) is of the form

Fe (t) = F0 et ( t)
whi h adiabati ally indu es a u tuation in the system for t < 0 and the lets the system evolve in time a ording
to the unperturbed Hamiltonian for t > 0. How will the indu ed u tuation evolve in time? Combining the kubo
transform relation with the linear response result for hB(t)i, we nd that
4
 Z 0 Z 
hB (t)i = dses dhB_ ( ih)B (t s)i0
1 0
Z  Z 1 d
= et d due u hB ( ih)B (u)i0
du
0 t
where the hange of variables u = t s has been made. Taking the limit  ! 0, and performing the integral over u,
we nd
Z 
hB (t)i = d [hB ( ih)B (1)i0 hB ( ih )B (t)i0 ℄
0

Sin e we assumed that hBi0 = 0, we have hB( ih)B(1)i0 = hB( ih)i0 hB(1)i0 = 0. Thus, dividing by hB(0)i,
we nd
R
hB (t)i = 0 dB ( ih)B (t)i0 ! hB (0)B (t)i0
hB (0)i R0 dB ( ih)B (0)i0 h !0 hB (0)2 i0
Thus at long times in the lassi al limit, the u tuations de ay to 0, indi ting a omplete regression or suppression of
the indu ed u tuation:
hB (t)i ! 0
hB (0)i

5
 G25.2651: Statistical Mechanics

Notes for Lecture 24

I. THE HARMONIC BATH HAMILTONIAN

In the theory of chemical reactions, it is often possible to isolate a small number or even a single degree of freedom in
the system that can be used to characterize the reaction. This degree of freedom is coupled to other degrees of freedom
(for example, reactions often take place in solution). Isomerization or dissociation of a diatomic molecule in solution
is an excellent example of this type of system. The degree of freedom of paramount interest is the distance between
the two atoms of the molecule { this is the degree of freedom whose detailed dynamics we would like to elucidate.
The dynamics of the \bath" or environment to which is couples is less interesting, but still must be accounted for in
some manner. A model that has maintained a certain level of both popularity and success is the so called \harmonic
bath" model, in which the environment to which the special degree(s) of freedom couple is replaced by an eective
set of harmonic oscillators. We will examine this model for the case of a single degree of freedom of interest, which
we will designate q. For the case of the isomerizing or dissociating diatomic, q could be the coordinate r ? hri, where
r is the distance between the atoms. The particular denition of q ensures that hqi = 0. The degree of freedom q is
assumed to couple to the bath linearly, giving a Hamiltonian of the form
"  2 #
2 X p2
H = 2pm + (q) +  + 1 m ! 2 x + g q
  
 2m 2 m !2
where the index  runs over all the bath degrees of freedom, ! are the harmonic bath frequencies, m are the
harmonic bath masses, and g are the coupling constants between the bath and the coordinate q. p is a momentum
conjugate to q, and m is the mass associated with this degree of freedom (e.g., the reduced mass  in the case of a
diatomic). The coordinate q is assumed to be subject to a potential (q) as well (e.g., an internal bond potential).
The form of the coupling between the system (q) and the bath (x ) is known as bilinear.
Below, using a completely classical treatment of this Hamiltonian, we will derive an equation for the detailed
dynamics of q alone. This equation is known as the generalized Langevin equation (GLE).

II. DERIVATION OF THE GLE

The GLE can be derived from the harmonic bath Hamiltonian by simply solving Hamilton's equations of motion,
which take the form
q_ = mp
X g2
p_ = ? @
X
?
@q  g x ?
m
 q
!2
 
x_  = mp

p_ = ?m !2 x ? g q
This set of equations can also be written as second order dierential equation:
X g2
mq = ? @
X
@q ? g  x ?

m !2 q
   
m x = ?m   ? g q
!2 x
In order to derive an equation for q, we solve explicitly for the dynamics of the bath variables and then substitute into
the equation for q. The equation for x is a second order inhomogeneous dierential equation, which can be solved
by Laplace transforms. We simply take the Laplace transform of both sides. Denote the Laplace transforms of q and
x as

1
 Z 1
q~(s) = dt e?st q(t)
Z0 1
x~ = dt e?st x (t)
0
and recognizing that
Z 1
dt e?st x (t) = s2 x~ (s) ? sx (0) ? x_  (0)
0
we obtain the following equation for x~ (s):
(s2 + !2 )~x (s) = sx (0) + x_  (0) ? mg q~(s)

or
x~ (s) = s2 +s !2 x (0) + s2 +1 !2 x_  (0) ? mg s2q~+(s!) 2
   
x (t) can be obtained by inverse Laplace transformation, which is equivalent to a contour integral in the complex
s-plane around a contour that encloses all the poles of the integrand. This contour is known as the Bromwich contour.
To see how this works, consider the rst term in the above expression. The inverse Laplace transform is
1 I ds sest = 1 I ds sest
2i s2 + !2 2i (s + i! )(s ? i! )
The integrand has two poles on the imaginary s-axis at i!. Integration over the contour that encloses these poles
picks up both residues from these poles. Since the poles are simple poles, then, from the residue theorem:
  
1 I ds sest 1 i! ei! t ?i! e?i! t
2i (s + i! )(s ? i! ) = 2i 2i 2i! + ?2i! = cos ! t
By the same method, the second term will give (sin ! t)=! . The last term is the inverse Laplace transform of a
product of q~(s) and 1=(s2 + !2 ). From the convolution theorem of Laplace transforms, the Laplace transform of a
convolution gives the product of Laplace transforms:
Z 1 Z t
dt e?st d f ( )g(t ?  ) = f~(s)~g(s)
0 0
Thus, the last term will be the convolution of q(t) with (sin ! t)=!. Putting these results together, gives, as the
solution for x (t):
Z t
x (t) = x (0) cos ! t + x
_  (0)
sin ! t ? g 
dq( ) sin ! (t ?  )
    
! m ! 0
The convolution term can be expressed in terms of q_ rather than q by integrating it by parts:
g Z t d q( ) sin ! (t ?  ) = g [q(t) ? q(0) cos ! t] ? g Z t d q_( ) cos ! (t ?  )
  
m ! 0 m !2 m !2 0
The reasons for preferring this form will be made clear shortly. The bath variables can now be seen to evolve according
to
Z t
x (t) = x (0) cos ! t + x_  (0) sin ! t + g
  
! 
m !2 0 d q_( ) cos ! (t ?  ) ? g [q(t) ? q(0) cos ! t]

m !2 

Substituting this into the equation of motion for q, we nd

 
@ p g X g2 Z t 2
 (0)
d q_( ) cos ! (t? )+ mg!2 q(t)?
X  X
mq = ? @q ? g x (0) cos ! t + m ! sin ! t + m !2 q(0) cos ! t ? m !2
        0   
We now introduce the following notation for the sums over bath modes appearing in this equation:

2
 1. Dene a dynamic friction kernel
 (t) =
X g
 m !2 cos ! t
2. Dene a random force
  
x (0) + mg!2 q(0) cos ! t + m
p (0) sin ! t
X
R(t) = ? g 
    !

Using these denitions, the equation of motion for q reads

Z t
@
mq = ? @q ? d q_( ) (t ?  ) + R(t) (1)
0
Eq. (1) is known as the generalized Langevin equation. Note that it takes the form of a one-dimensional
Rt
particle subject
to a potential (q), driven by a forcing function R(t) and with a nonlocal (in time) damping term ? 0 d q_( ) (t ?  ),
which depends, in general, on the entire history of the evolution of q. The GLE is often taken as a phenomenological
equation of motion for a coordinate q coupled to a general bath. In this spirit, it is worth taking a moment to discuss
the physical meaning of the terms appearing in the equation.

III. PROPERTIES OF THE GLE

Below we discuss the physical meaning of the terms appearing the GLE

A. The random force term

Within the context of a harmonic bath, the term \random force" is something of a misnomer, since R(t) is completely
deterministic and not random at all!!! We will return to this point momentarily, however, let us examine particular
features of R(t) from its explicit expression from the harmonic bath dynamics. Note, rst of all, that it does not
depend on the dynamics of the system coordinate q (except for the appearance of q(0)). In this sense, it is independent
or \orthogonal" to q within a phase space picture. From the explicit form of R(t), it is straightforward to see that
the correlation function
hq_(0)R(t)i = 0
i.e., the correlation function of the system velocity q_ with the random force is 0. This can be seen by substituting
in the expression for R(t) and integrating over initial conditions with a canonical distribution weighting. For certain
potentials (q) that are even in q (such as a harmonic oscillator), one can also show that
hq(0)R(t)i = 0
Thus, R(t) is completely uncorrelated from both q and q_, which is a property we might expect from a truly random
process. In fact, R(t) is determined by the detailed dynamics of the bath. However, we are not particularly interested
or able to follow these detailed dynamics for a large number of bath degrees of freedom. Thus, we could just as well
model R(t) by a completely random process (satisfying certain desirable features that are characteristic of a more
general bath), and, in fact, this is often done. One could, for example, postulate that R(t) act over a maximum time
tmax at discrete points in time k
t, giving N = tmax=
t values of Rk = R(k
t), and assume that Rk takes the form
of a gaussian random process:
N h
X i
Rk = aj e2ijk=N + bj e?2ijk=N
j =1
where the coecients faj g and fbj g are chosen at random from a gaussian distribution function. This might be
expected to be suitable for a bath of high density, where strong collisions between the system and a bath particle are
essentially nonexistent, but where the system only sees feels the relatively \soft" uctuations of the less mobile bath.
For a low density bath, one might try modeling R(t) as a Poisson process of very strong collisions.
Whatever model is chosen for R(t), if it is a truly random process that can only act at discrete points in time, then
the GLE takes the form of a stochastic (based on random numbers) integro-dierential equation. There is a whole
body of mathematics devoted to the properties of such equations, where heavy use of an It^o calculus is made.

3
 B. The dynamic friction kernel
The convolution integral term
Z t
d q_( ) (t ?  )
0
is called the memory integral because it depends, in general, on the entire history of the evolution of q. Physically
it expresses the fact that the bath requires a nite time to respond to any uctuation in the motion of the system
(q). This, in turn, aects how the bath acts back on the system. Thus, the force that the bath exerts on the system
presently depends on what the system coordinate q did in the past. However, we have seen previously the regression
of uctuations (their decay to 0) over time. Thus, we expect that what the system did very far in the past will no
longer the force it feels presently, i.e., that the lower limit of the memory integral (which is rigorously 0) could be
replaced by t ? tmem, where tmem is the maximum time over which memory of what the system coordinate did in the
past is important. This can be interpreted as a indicating a certain decay time for the friction kernel  (t). In fact,
 (t) often does decay to 0 in a relatively short time. Often this decay takes the form of a rapid initial decay followed
by a slow nal decay, as shown in the gure below:
Consider the extreme case that the bath is capable of responding innitely quickly to changes in the system coordinate
q. This would be the case, for example, if there were a large mass disparity between the system and the bath
(m >> m ). Then, the bath retains no memory of what the system did in the past, and we could take  (t) to be a
-function in time:
 (t) = 20 (t)
Then
Z t Z t Z t
d q_( ) (t ?  ) = d q_(t ?  ) ( ) = 20 d ( )q_(t ?  ) = 0 q_(t)
0 0 0
and the GLE becomes
mq = ? @
@q ? 0 q_ + R(t)
This simpler equation of motion is known as the Langevin equation and it is clearly a special case of the more
generalized equation of motion. It is often invoked to describe brownian motion where clearly such a mass disparity
is present. The constant 0 is known as the static friction and is given by
Z 1
0 = dt  (t)
0
In fact, this is a general relation for determining the static friction constant.
The other extreme is a very sluggish bath that responds slowly to changes in the system coordinate. In this case,
we may take  (t) to be a constant    (0), at least, for times short compared to the response time of the bath. Then,
the memory integral becomes
Z t
d q_( ) (t ?  )   (q(t) ? q(0))
0
and the GLE becomes
 
@ (q) + 1  (q ? q )2 + R(t)
mq = ? @q 2 0

where the friction term now manifests itself as an extra harmonic term added to the potential. Such a term has
the eect of trapping the system in certain regions of conguration space, an eect known as dynamic caging. An
example of this is a dilute mixture of small, light particles in a bath of heavy, large particles. The light particles can
get trapped in regions of space where many bath particles are in a sort of spatial \cage." Only the rare uctuations
in the bath that open up larger holes in conguration space allow the light particles to escape the cage, occasionally,
after which, they often get trapped again in a new cage for a similar time interval.

4
 C. Relation between the dynamic friction kernel and the random force
From the denitions of R(t) and  (t), it is straightforward to show that there is a relation between them of the
form
hR(0)R(t)i = kT (t)
This relation is known as the second uctuation dissipation theorem. The fact that it involves a simple autocorrelation
function of the random force is particular to the harmonic bath model. We will see later that a more general form of
this relation exists, valid for a general bath. This relation must be kept in mind when introducing models for R(t)
and  (t). In eect, it acts as a constraint on the possible ways in which one can model the random force and friction
kernel.

IV. MORI-ZWANZIG THEORY: A MORE GENERAL DERIVATION OF THE GLE

A derivation of the GLE valid for a general bath can be worked out. The details of the derivation are given in the
book by Berne and Pecora called Dynamic Light Scattering. The system coordinate q and its conjugate momentum
p are introduced as a column vector:
 
A = pq
and, in addition, one introduces statistical projection operators P and Q that project onto subspaces in phase space
parallel and orthogonal to A. These operators take the form
P = h:::AT ihAAT i?1
Q=I ?P
These operators are Hermitian and satisfy the property of idempotency:
P2 = P
Q2 = Q
Also, note that
PA = A
QA = 0
The time evolution of A is given by application of the classical propagator:
A(t) = eiLtA(0)
Note that the evolution of A is unitary, i.e., it preserves the norm of A:
jA(t)j2 = jA(0)j2
Dierentiating both sides of the time evolution equation for A gives:
dA = eiLt iLA(0)
dt
Then, an identity operator is inserted in the above expression in the form I = P + Q:
dA = eiLt (P + Q)iLA(0) = eiLtPiLA(0) + eiLtQiLA(0)
dt
The rst term in this expression denes a frequency matrix acting on A:

5
 eiLt PiLA(0) = eiLt hiLAAT ihAAT i?1 A
= hiLAAT ihAAT i?1 eiLt A

= hiLAAT ihAAT i?1 A(t)

 i
A(t)
where

= hLAAT ihAAT i?1
In order to evaluate the second term, another identity operator is inserted directly into the propagator:
eiLt = ei(P +Q)Lt
Consider the dierence between the two propagators:
eiLt ? eiQLt
If this dierence is Laplace transformed, it becomes
(s ? iL)?1 ? (s ? iQL)?1
which can be simplied via the general operator identity:
A?1 ? B?1 = A?1 (B ? A)B?1
Letting
A = (s ? iL)
B = (s ? iQL)
we have
(s ? iL)?1 ? (s ? iQL)?1 = (s ? iL)?1 (s ? iQL ? s + iL)(s ? iQL)?1
= (s ? iL)?1 iPL(s ? iQL)?1
or
(s ? iL)?1 = (s ? iQL)?1 + (s ? iL)?1(s ? iQL ? s + iL)(s ? iQL)?1
Now, inverse Laplace transforming both sides gives
Z t
eiLt = eiQLt + d eiL(t? )iPLeiQL
0
Thus, multiplying fromthe right by QiLA gives
Z t
eiLt QiLA = eiQLt QiLA + d eiL(t? )iPLeiQL QiLA
0
Dene a vector
F(t) = eiQLt QiLA(0)
so that
Z t
eiLt QiLA = F(t) + d hiLF( )AT ihAAT i?1 A(t ?  )
0

6
Because F(t) is completely orthogonal to A(t), it is straightforward to show that
QF(t) = F(t)
Then,
hiLF( )AT ihAAT i?1 A = hiLQF( )AT ihAAT i?1 A
= ?hQF( )(iLA)T ihAAT i?1 A

= ?hQ2 F( )(iLA)T ihAAT i?1 A

= ?hQF( )(QiLA)T ihAAT i?1 A

= ?hF( )FT (0)ihAAT i?1 A

Thus,
Z t
eiLt QiLA = F(t) ? d hF( )FT (0)ihAAT i?1 A(t ?  )
0
Finally, we dene a memory kernel matrix:
K(t) = hF( )FT (0)ihAAT i?1
Then, combining all results, we nd, for dA=dt:
dA = i
(t)A ? Z t d K( )A(t ?  ) + F(t)
dt 0
which equivalent to a generalized Langevin equation for a particle subject to a harmonic potential, but coupled to a
general bath. For most systems, the quantities appearing in this form of the generalized Langevin equation are
 
0
i
= ?m!2 0 1=m
 
K(t) = 00  (t)0=m
 
F(t) = R0(t)
It is easy to derive these expressions for the case of the harmonic bath Hamiltonian when (q) = m!2 q2 =2.
For the case of a harmonic bath Hamiltonian, we had shown that the friction kernel was related to the random
force by the uctuation dissipation theorem:
hR(0)R(t)i = hR(0)eiLt R(0)i = kT (t)
For a general bath, the relation is not as simple, owing to the fact that F(t) is evolved using a modied propagator
exp(iQLt). Thus, the more general form of the uctuation dissipation theorem is
hR(0)eiQLt R(0)i = kT (t)
so that the dynamics of R(t) is prescribed by the propagator exp(iQLt). This more general relation illustrates the
diculty of dening a friction kernel for a general bath. However, for the special case of a sti harmonic diatomic
molecule interacting with a bath for which all the modes are soft compared to the frequency of the diatomic, a very
useful approximation results. One can show that
hR(0)eiQLt R(0)i  hR(0)eiLcons t R(0)i
where iLcons is the Liouville operator for a system in which the diatomic is held rigidly xed at some particular bond
length (i.e., a constrained dynamics). Since the friction kernel is not sensitive to the details of the internal potential of
the diatomic, this approximation can also be used for diatomics with sti, anharmonic potentials. This approximation
is referred to as the rigid bond approximation (see Berne, et al, J. Chem. Phys. 93, 5084 (1990)).

7
 V. EXAMPLE: VIBRATIONAL DEPHASING AND ENERGY RELAXATION
Recall that the Fourier transform of a time correlation function can be related to some kind of frequency spectrum.
For example, the Fourier transform of the velocity autocorrelation function of a particular degree of freedom q of
interest
Cvv (t) = hq_(0) q_(t)i
hq_ i
2

where v = q_, gives the relevant frequencies contributing to the dynamics of q, but does not give amplitudes. This
\frequency" spectrum I (!) is simply given by
Z 1
I (!) = dt ei!t Cvv (t)
0
That is, we take the Laplace transform of Cvv (t) using s = ?i!. Since Cvv (t) carries information about the relevant
frequencies of the system, the decay of Cvv (t) in time is a measure of how strongly coupled the motion of q is to the
rest of the bath, i.e., how much of an overlap there is between the relevant frequencies of the bath and those of q.
The more of an overlap there is, the more mixing there will be between the system and the bath, and hence, the more
rapidly the motion of the system will become vibrationally \out of phase" or decorrelated with itself. Thus, the decay
time of Cvv (t), which is denoted T2 is called the vibrational dephasing time.
Another measure of the strength of the coupling between the system and the bath is the time required for the
system to dissipate energy into the bath when it is excited away from equilibrium. This time can be obtained by
studying the decay of the energy autocorrelation function:
C = h"(0)"(t)i
""
h"2 i
where "(t) is dened to be
"(t) = 21 mq_2 + (q) ? kT
The decay time of this correlation function is denoted T1.
The question then becomes: what are these characteristic decay times and how are they related? To answer this,
we will take a phenomenological approach. We will assume the validity of the GLE for q:
Z t
mq = ? @
@q ? d q_( ) (t ?  ) + R(t)
0
and use it to calculate T1 and T2 .
Suppose the potential (q) is harmonic and takes the form
(q) = 12 m!2 q2
Substituting into the GLE and dividing through by m gives
Z t
q = ?!2q ? d q_(t ?  ) ( ) + f (t)
0
where
(t) = m
(t) f (t) = Rm(t)
An equation of motion for Cvv (t) can be obtained directly by multiplying both sides of the GLE by q_(0) and averaging
over a canonical ensemble:
Z t
hq_(0)q(t)i = ?!2hq_(0)q(t)i ? d hq_(0)q_(t ?  )i ( ) + hq_(0)f (t)i
0

8
Recall that
hq_(0)f (t)i = m1 hq_(0)R(t)i = 0
and note that
hq_(0)q(t)i = dtd hq_(0)q_(t)i = dCdtvv
also
Z t
d hq_(0)q_( )i = hq_(0)q(t)i ? hq_(0)q(0)i = hq_(0)q(t)i
0
Thus,
Z t
hq_(0)q(t)i = d Cvv ( )
0
Combining these results gives an equation for Cvv (t)
d C (t) = ? Z t d ?!2 + (t ?  )
C ( )
dt vv 0
vv
Z t
d
dt Cvv (t) = ? d K (t ?  )Cvv ( )
0
which is known as the memory function equation and the kernel K (t) is known as the memory function or memory
kernel. This type of integro-dierential equation is called a Volterra equation and it can be solved by Laplace
transforms.
Taking the Laplace transform of both sides gives
sC~vv (s) ? Cvv (0) = ?C~vv (s)K~ (s)
However, it is clear that Cvv (0) = 1 and also
2
K~ (s) = !s + ~(s)
Thus, it follows that
 2 
sC~vv (s) ? 1 = !s + ~(s) C~vv (s)
C~vv (s) = s2 + s ~s(s) + !2
In order to perform the inverse Laplace transform, we need the poles of the integrand, which will be determined by
the solutions of
s2 + s ~(s) + !2 = 0
which we could solve directly if we knew the explicit form of ~(s).
However, if ! is suciently larger than ~(0), then it is possible to develop a perturbation solution to this equation.
Let us assume the solutions for s can be written as
s = s0 + s1 + s2 +



Substituting in this ansatz gives
(s0 + s1 + s2 +


)2 + (s0 + s1 + s2 +


)~ (s0 + s1 + s2 +


) + !2 = 0
Since we are assuming ~ is small, then to lowest order, we have

9
 s20 + !2 = 0
so that s0 = !. The rst order equation then becomes
2s0 s1 + s0 ~(s0 ) = 0
or
s1 = ? ~(2s0 ) = ? ~(2i!)
Note, however, that
Z 1
~(i!) = dt (t)ei!t
0
Z 1
= dt [ (t) cos !t  i (t) sin !t]
0
 0 (!)  i 00(!)
Thus, stopping the rst order result, the poles of the integrand occur at
0 0
s  i (! + 00 (!)) ? (2!)  i
? (2!)
Dene
0
s+ = i
? (2!)
0
s? = ?i
? (2!)
Then
C~vv (s)  (s ? s )(s s ? s )
+ ?
and Cvv (t) is then given by the contour integral
I
Cvv (t) = 21i sest ds
(s ? s+ )(s ? s? )
Taking the residue at each pole, we nd
Cvv (t) = (ss+?e s ) + (ss??e s )
s+ t s? t
+ ? ? +
which can be simplied to give
 0 
Cvv (t) = e? 0(!)t=2 cos
t ? 2

(!) sin
t

Thus, we see that the GLE predicts Cvv (t) oscillates with a frequency
and decays exponentially. From the expo-
nential decay, we can directly read o the time T2 :
1 = 0 (!) =  0 (!)
T2 2 2m
That is, the value of the real part of the Fourier (Laplace) transform of the friction kernel evaluated at the renormalized
frequency divided by 2m gives the vibrational dephasing time! By a similar scheme, one can easily show that the
position autocorrelation function Cqq (t) = hq(0)q(t)i decays with the same dephasing time. It's explicit form is
 0 (!) 
Cqq (t) = e? 0(!)t=2 cos
t + 2
sin
t

10
 The energy autocorrelation function C"" (t) can be expressed in terms of the more primitive correlation functions
Cqq (t) and Cvv (t). It is a straightforward, although extremely tedious, matter to show that the relation, valid for the
harmonic potential of mean force, is
C"" (t) = 21 Cvv
2 (t) + 1 C 2 (t) + 1 C_ 2 (t)
2 qq !2 qq
Substituting in the expressions for Cqq (t) and Cvv (t) gives
C"" (t) = e? 0(!)t  (oscillatory functions of t)
so that the decay time T1 can be seen to be
1 = 0 (!) =  0 (!)
T1 m
and therefore, the relation between T1 and T2 can be seen immediately to be
1 1
T2 = 2T1
The incredible fact is that this result is also true quantum mechanically. That is, by doing a simple, purely classical
treatment of the problem, we obtained a result that turns out to be the correct quantum mechanical result!
Just how big are these times? If ! is very large compared to any typical frequency relevant to the bath, then the
friction kernel evaluated at this frequency will be extremely small, giving rise to a long decay time. This result is
expect, since, if ! is large compared to the bath, there are very few ways in which the system can dissipate energy into
the bath. The situation changes dramatically, however, if a small amount of anharmonicity is added to the potential of
mean force. The gure below illustrates the point for a harmonic diatomic molecule interacting with a Lennard-Jones
bath. The top gure shows the velocity autocorrelation function for an oscillator whose frequency is approximately 3
times the characteristic frequency of the bath, while the bottom one shows the velocity autocorrelation function for
the case that the frequency disparity is a factor of 6.

11
FIG. 1.

12
 G25.2651: Statistical Mechanics

Notes for Lecture 25

I. OVERVIEW OF CRITICAL PHENOMENA

Consider the phase diagram of a typical substance:

Melting curve

Solid Liquid Supercritical

fluid region

Pc . Critical point
Triple point . Gas
Boiling curve

Sublimation curve Tc T
FIG. 1.

The boundary lines between phases are called the coexistence lines. Crossing a coexistence line leads to a rst order
phase transition, which is characterized by a discontinuous change in some thermodynamic quantity such as volume,
enthalpy, magnetization.
Notice that, while the melting curve, in principle, can be extended to innity, the gas-liquid/boiling curve terminates
at a point, beyond which the two phases cannot be distinguished. This point is called a critical point and beyond
the critical point, the system is in a supercritical uid state. The temperature at which this point occurs is called the
critical temperature Tc. At a critical point, thermodynamic quantities such as given above remain continuous, but
derivatives of these functions may diverge. Examples of such divergent quantities are the heat capacity, isothermal
compressibility, magnetic susceptibility, etc. These divergences occur in a universal way for large classes of systems
(such systems are said to belong to the same universality class, a concept to be dened more precisely in the next
lecture).
Recall that the equation of state of a system can be expressed in the form
P = g(; T )
i.e., pressure is a function of density an temperature. A set of isotherms of the equation of state for a typical substance
might appear as follows:

1
 P T > Tc

T = Tc

T < Tc

Inflection pt.
Pc .

Linear regime ρG ρ ρ ρ
c L
FIG. 2.

The dashed curve corresponds to the gas-liquid coexistence curve. Below the critical isotherm, the gas-liquid
coexistence curve describes how large a discontinuous change in the density occurs during rst-order gas-liquid phase
transition. At the in ection point, which corresponds to the critical point, the discontinuity goes to 0.
As noted above, the divergences in thermodynamic derivative quantities occur in the same way for systems belonging
to the same universality class. These divergences behave as power laws, and hence can be described by the exponents
in the power laws. These exponents are known as the critical exponents. Thus, the critical exponents will be the
same for all systems belonging to the same universality class. The critical exponents are dened as follows for the
gas-liquid critical point:
1. The heat capacity at constant volume dened by
CV = @E = ?T @@TA2
   
2

@T V V
diverges with temperature as the critical temperature is approached according to
CV  jT ? Tc j?

2. The isothermal compressibility dened by

T = ? V1 @V @
   

@P = 1 @T
T T
diverges with temperature as the critical temperature is approached according to
T  jT ? Tc j?

2
 3. On the critical isotherm, the shape of the curve near the in ection point at (c ; Pc ) is given by
P ? Pc  j ? c j sign( ? c ) >0

4. The shape of the coexistence curve in the -T plane near the critical point for T < Tc is given by
L ? G  (Tc ? T )
These are the primary critical exponents.

A. Review of the Van der Waals theory

Recall that the Van der Waals equation of state was derived earlier by perturbation theory. The unperturbed
Hamiltonian describes a system of hard spheres and is given by
N 2
X X
H0 = pi
+ u0 (jri ? rj j)
i=1 2m i<j
and a perturbation of the form
X
H1 = u1 (jri ? rj j)
i<j
where u0 (r) is the hard-sphere potential given by

0
u0 (r) = 1 r>
r
and u1 (r) is an arbitrary attractive potential. In the low density limit, we had
g(r)  e?u0 (r) = (r ? )
and the free energy was determined to be
A(N; V; T )  ? 1 ln (VN?!Nb)N ? aN 2
 

3N V
with
b = 23 3
Z 1
a = ?2 dr r2 u1 (r)

The equation of state takes the form
P = 1kT
? b ? a
2

The critical point is dened by the conditions:

@P = 0 @2P = 0
@V @V 2
which leads to the following values for the critical pressure, temperature and density:
Pc = 27ab2 Tc = 27 8a  = 1
b c
3b
The critical exponents predicted by the theory are as follows:

3
1. The internal energy is given by
E = 32 NkT + aN
2

V
from which it can be seen that the heat capacity is
CV = @E
 

= 3 Nk  jT ? T j0
c
@T V 2
)  = 0.
2. The isothermal compressibility can be expressed as
1
T = ? V (@P=@V )
and 
@P  8a
 
1
@V V =Vc = 4Nb2 27b ? kT  (T ? Tc )
so that
T  (T ? Tc)?1
) =1
3. By Taylor expanding the equation of state about the critical pressure and density, it is easy to show that
1 (P ? P )  const + ( ?  )3 +



c c
kT
)=3
4. The exponent  can be computed using the Maxwell construction (see problem set 11), which attempts to x the
fact that the Van der Waals equation has an unphysical region where @P=@V > 0. The Maxwell construction is
illustrated below:
P

Van der Waals eqn

a2
Tie line
a1

VL VG V
FIG. 3.

4
 When the Maxwell construction is carried out, it can be shown that
L ? G  (Tc ? T )1=2
)  = 1=2.
The following table compares the Van der Waals exponents to the experimental critical exponents:

α β γ δ

VdW 0 1/2 1 3

Exp. 0.1 0.34 1.35 4.2

FIG. 4.

Thus, one sees that the Van der Waals theory is only qualitatively correct, but not quantitatively. It is an example
of a mean eld theory.

II. MAGNETIC SYSTEMS AND THE ISING MODEL

Imagine a cubic lattice in which particles carrying spin S are placed on the lattice sites with jSj = h=2 as shown
below:

FIG. 5.

5
 Such a model describes ferromagnetic materials, which can be magnetized by applying an external magnetic eld
h. In the absence of a eld, the unperturbed Hamiltonian takes the form
H0 = ? 12 i
Jij
j
X

i;j
where J is a tensor and  is a spin vector such that jj = 1. Quantum mechanically,  would be the vector of Pauli
matrices. In general, the spins can point in any spatial direction, a fact that makes the problem dicult to solve.
A simplication introduced by Ising was to allow the spins to point in only one of two possible directions, up or
down, e.g., along the z -axis only. In addition, the summation is restricted to nearest neighbor interactions only. In
this model, the Hamiltonian becomes
H0 = ? 21 Jij i j
X

hi;ji
where hi; j iindicates restriction of the sum to nearest neighbor pairs only. The variables i now can take on the values
1 only. The couplings Jij are the spin-spin \J " couplings.
In the presence of a magnetic eld, the full Hamiltonian becomes
N
H = ? 21
X X
Jij i j ? h i
hi;ji i=1
which describes a uniaxial ferromagnetic system in a magnetic eld. The parameters T and h are experimental control
parameters.
Dene the magnetization per spin as
N
m = N1 i
X

i=1
Then the phase diagram looks like:
h

Spin up phase
T
Spin down phase
Tc

FIG. 6.

6
 where the colored lines indicate a nonzero magnetization at h = 0 below a critical temperature T . The persistence
of a nonzero magnetization in ferromagnetic systems at h = 0 below T = Tc indicates a transition from a disordered
to an ordered phase. In the latter, the spins are aligned in the direction of the applied eld before it is switched o.
If h ?! 0+ , then the spins will point in one direction and if h ?! 0? , it will be in the opposite direction. A plot of
the isotherms of m vs. h yields:
m

T < TC

T = TC

T>TC

FIG. 7.

Notice an in ection point along the isotherm T = Tc, at h = 0, where @m=@h ?! 1.

The thermodynamics of the magnetic system can be dened in analogy with the liquid-gas system. The analogy is
shown in the table below:
Gas-Liquid Magnetic

P h
V ?M = ?Nm

T = ? V1 @V
@P  = @m
@h
A = A(N; V; T ) A = A(N; M; T )
G = A + PV (P ) G = A ? hM (h)
@A
P = ? @V @A
h = @M

V = @P@G M = ? @G
@h
CV = ?T @@TA2 CM = ?T @@TA2
   
2 2

V M

7
 CP = ?T @@TG2 Ch = ?T @@TG2
   
2 2

P h
where  is the magnetic susceptibility. The magnetic exponents are then given by
Ch  jT ? Tcj? h = 0 limit
  jT ? Tcj? h = 0 limit
h  jmj sign(m) at T = Tc
m  (Tc ? T ) T < Tc

8
 G25.2651: Statistical Mechanics

Notes for Lecture 26

I. MEAN FIELD THEORY CALCULATION OF MAGNETIC EXPONENTS

The calculation of critical exponents is nontrivial, even for simple models such as the Ising model. Here, we will
introduce an approximate technique known as mean eld theory. The approximation that is made in the mean eld
theory (MFT) is that uctuations can be neglected. Clearly, this is a severe approximation, the consequences of which
we will see in the nal results.
Consider the Hamiltonian for the Ising model:
H = 21 Jij i j + h i
X X

hi;ji i
The partition function is given by
h P P i
XX X  21 hi;ji i j +h i i

(N; h; T ) =


e
1 2 N
Notice that we have written the partition function as an isothermal-isomagnetic partition function in analogy with
the isothermal-isobaric ensemble. (Most books use the notation Q for the partition function and A for the free energy,
which is misleading). This sum is nontrivial to carry out.
In the MFT approximation, one introduces the magnetization
N
m = N1 h i i
X

i=1
explicitly into the partition function by using the identity
i j = (i ? m + m)(j ? m + m)
= m2 + m(i ? m) + m(j ? m) + (i ? m)(j ? m)
The last term is quadratic in the spins and is of the form (i ? hi)(j ? hi), the average of which measures the spin
uctuations. Thus, this term is neglected in the MFT. If this term is dropped, then the spin-spin interaction term in
the Hamiltonian becomes:
1 XJ    1 XJ m2 + m( +  ) ? 2m2
2 hi;ji ij i j 2 hi;ji ij i j

= 12 Jij ?m2 + m(i + j )

X  

hi;ji
P
We will restrict ourselves to isotropic magnetic systems, for which j Jij is independent of i (all sites are equivalent).
P
Dene j Jij  Jz ~ , where z is the number of nearest neighbors of each spin. This number will depend on the number
of spatial dimensions. Since this dependence on spatial dimension is a trivial one, we can absorb the z factor into the
coupling constant and redene J = Jz ~ . Then,
1 X J = 1 NJ
2 i 2
where N is the total number of spins. Finally,
1 X J m( +  ) = Jm X 
2 hi;ji ij i j i
i

1
and the Hamiltonian now takes the form
1 X J   + h X  ?! ? 1 NJm2 + (Jm + h) X 
2 hi;ji ij i j i
i 2 i
i

and the partition function becomes

2 X X P

(N; h; T ) = e?NJm =2


e(Jm+h) i i
1 N
!N
2 X
= e?NJm e(Jm+h)
=1
2 =2
=e ?NJm [2cosh (Jm + h)]N
The free energy per spin g(h; T ) = G(N; h; T )=N is then given by
g(h; T ) = ? 1 ln
(N; h; T )
N
= 12 Jm2 ? 1 ln [2cosh (Jm + h)]
The magnetization per spin can be computed from
 
m = ? @h @g
h
= tanh (Jm + h)
Allowing h ?! 0, one nds a transcendental equation for m
m = tanh(mJ )
which can be solved graphically as shown below:

2
 tanh(βJm)
f(m)=m

− m0
m
m0

FIG. 1.

Note that for J > 1, there are three solutions. One is at m = 0 and the other two are at nite values of m, which
we will call m0. For J < 1, there is only one solution at m = 0. Thus, for J > 1, MFT predicts a nonzero
magnetization at h = 0. The three solutions coalesce onto a single solution at J = 1. The condition J = 1 thus
denes a critical temperature below which (J > 1) there is a nite magnetization at h = 0. The condition J = 1
denes the critical temperature, which leads to
kTc = J
To see the physical meaning of the various cases, consider expanding the free energy about m = 0 at zero-eld. The
expansion gives
g(0; m) = const + J (1 ? J )m2 + cm4
where c is a (possibly temperature dependent) constant with c > 0. For J > 1, the sign of the quadratic term is
negative and the free energy as a function of m looks like:

3
 g(0,m)

m
−m 0 m0
FIG. 2.

Thus, there are two stable minima at m0 , corresponding to the two possible states of magnetization. Since a large
portion of the spins will be aligned below the critical temperature, the magnetic phase is called an ordered phase. For
J > 1, the sign of the quadratic term is positive and the free energy plot looks like:

4
 g(0,m)

m
FIG. 3.

i.e., a single minimum function at m = 0, indicating no net magnetization above the critical temperature at h = 0.
The exponent  can be obtained directly from this expression for the free energy. For T < Tc, the value of the
magnetization is given by

@g 
@m m=m0 = 0
which gives
2J (T ? T )m + 4cm3 = 0
T c 0 0

1=2
m0  (Tc ? T )
Thus,  = 1=2.
From the equation for the magnetization at nonzero eld, the exponent  is obtained as follows:
m = tanh (Jm + h)
 (Jm + h) = tanh?1 m
 
h  kT m + m3 +


? Jm
3

= mk(T ? T ) + kT m3
c 3
where the second line is obtained by expanding the inverse hyperbolic tangent about m = 0. At the critical temper-
ature, this becomes
h  m3
so that  = 3.
For the exponent , we need to compute the heat capacity at zero-eld, which is either Ch or Cm . In either case,
we have, for T > Tc, where m = 0,

5
 G = ?NkT ln 2
so
 
Ch = ?T @@TG2 = 0
2

from which is it clear that  = 0. For T < Tc, Ch approaches a dierent value as T ! Tc , however, the dependence
on jT ? Tcj is the same, so that  = 0 is still obtained.
Finally, the susceptibility, which is given by
 = @m = 1
@h @h=@m
but, near m = 0,
h = mk(T ? Tc ) + kT
3m
3

@h = k(T ? T ) + kTm2
@m c
As the critical temperature is approached, m ! 0 and
  jT ? Tcj?1
which implies = 1.
The MFT exponents for the Ising model are, therefore
=0  = 1=2 =1 =3
which are exactly the same exponents that the Van der Waals theory predict for the uid system. The fact that two
(or more) dissimilar systems have the same set of critical exponents (at least at the MFT level) is a consequence of a
more general phenomenon known as universality, which was alluded to in the previous lecture.
Systems belonging to the same universality class will exhibit the same behavior about their critical points, as
manifested by their having the same set of critical exponents.
A universality class is characterized by two parameters:
1. The spatial dimension d.
2. The dimension, n, of the order parameter.
An order parameter is dened as follows:
Suppose the Hamiltonian H0 of a system is invariant under all the transformations of a group G . If two phases
can be distinguished by the appearance of a thermodynamic average hi, which is not invariant under G , then
hi is an order parameter for the system.
The Ising system, for which H0 is given by
H0 = ? 21
X
Jij i j
hi;ji
is invariant under the group Z2 , which is the group that contains only two elements, an identity element and a spin
re ection transformation: Z2 = 1; ?1. Thus, under Z2 , the spins transform as
i ! i i ! ?i
From the form of H0 is can be seen that H0 ! H0 under both transformations of Z2 , so that it is invariant under Z2 .
However, the magnetization
m = N1 h = 1N i i
X

6
is not invariant under a spin re ection for T < Tc, when the system is magnetized. In a completely ordered state,
with all spins aligned, under a spin re ection m ! ?m. Thus, m is an order parameter for the Ising model, and,
since it is a scalar quantity, its dimension is 1.
Thus, the Ising model denes a universality class known as the Ising universality class, characterized by d = 3,
n = 1 in three dimensions. Note that the uid system, which has the same MFT critical exponents as the Ising
system, belongs to the same universality class. The order parameter for this system, by the analogy table dened
in the last lecture, is the volume dierence between the gas an liquid phases, VL ? VG , or equivalently, the density
dierence, L ? G . Although the solid phase is the truly ordered phase, while the gas phase is disordered, the liquid
phase is somewhere in between, i.e., it is a partially ordered phase. The Hamiltonian of a uid is invariant under
rotations of the coordinate system. Ordered and partially ordered phases break this symmetry. Note also that a true
magnetic system, in which the spins can point in any spatial direction, need an order parameter that is the vector
generalization of the magnetization:
N
m = N1 h i i
X

i=1
Since the dimension of the vector magnetization is 3, the true magnetic system belongs to the d = 3, n = 3 universality
class.

II. EXACT SOLUTIONS OF THE ISING MODEL IN 1 AND 2 DIMENSIONS

Exact solutions of the Ising model are possible in 1 and 2 dimensions and can be used to calculate the exact critical
exponents for the two corresponding universality classes.
In one dimension, the Ising Hamiltonian becomes:
N
X N
X
H=? Ji;i+1 i i+1 ? h i
i=1 i=1
which corresponds to N spins on a line. We will impose periodic boundary conditions on the spins so that N +1 = 1 .
Thus, the topology of the spin space is that of a circle. Finally, let all sites be equivalent, so that Ji;i+1  J . Then,
N
X N
X
H = ?J i i+1 ? h i
i=1 i=1
The partition function is then
 P 
X X 1 PN

(N; h; T ) =


e J i i i+1 + 2 h i=1 (i +i+1 )
1 N
In order to carry out the spin sum, let us dene a matrix P with matrix elements:
hjPj0 i = e[J0 +h(+0 )=2]
h1jPj1i = e(J +h)
h?1jPj ? 1i = e(J ?h)
h1jPj ? 1i = h?1jPj1i = e?J
Thus, the matrix P becomes is a 22 matrix given by
 
 (J +h) ?J
P = e e?J ee(J ?h)
so that the partition function becomes
X X

(N; h; T ) =


h1 jPj2 ih2 jPj3 i


hN ?1 jPjN ihN jPj1 i
1 N
X
= h1 jPN j1 i
1
?

= Tr PN

7
A simple way to carry out the trace is diagonalize the matrix, P. From
det(P ? I) = 0
the eigenvalues can be seen to be
 q 
 = eJ cosh(h)  sinh2 (h) + e?4J  
where + corresponds to the choice of + in the eigenvalue expression, etc.
The trace of the PN is then
?

Tr PN = N+ + N?
We will be interested in the thermodynamic limit. Note that + > ? for any h, so that as N ! 1, N+ dominates
over N? . Thus, in this limit, the partition function has the single term:

(N; h; T ) ?! N+
Thus, the free energy per spin becomes
g(h; T ) = ?kT ln +  
q
= ?J ? kT ln cosh(h) + sinh (h) + e
2 ?4J

and the magnetization becomes

 
m = @h@g
@ ln +
= ? @ (h)
sinh(h) + psinh(h)cosh(h)
sinh2 (h)+e?4J
= q
cosh(h) + sinh2 (h) + e?4J
which, as h ! 0, since cosh(h) ! 1 and sinh(h) ! 0, itself vanishes. Thus, there is no magnetization at any nite
temperature in one dimension, hence no nontrivial critical point.
While the one-dimensional Ising model is a relatively simple problem to solve, the two-dimensional Ising model
is highly nontrivial. It was only the pure mathematical genius of Lars Onsager that was able to nd an analytical
solution to the two-dimensional Ising model. This, then, gives an exact set of critical exponents for the d = 2, n = 1
universality class. To date, the three-dimensional Ising model remains unsolved.
Here, the Onsager solution will be outlined only and the results stated. Consider a two-dimension spin-lattice as
shown below:

8
 n+1
. . . . .
.
. . . . . .
.
. . . . . .
2 . . . . .
1 . . . . .
1 2 ...... n+1
FIG. 4.

The Hamiltonian can be written as

X X
H = ?J (i;j i+1;j + i;j+1 i;j ) ? h i;j
i;j i;j
where the spins are now indexed by two indices corresponding to a point on the 2-dimensional lattice. Introduce a
shorthand notation for H :
n
X
H= [E (j ; j+1 ) + E (j )]
j =1
where
n
X
E (j ; k )  ? ij ik
i=1
Xn X
E (j )  ?J ij i+1;j ? h j
i=1 i;j
and j is dened to be a set of spins in a particular column:
j  f1j ; :::; nj g
Then, dene a transfer matrix P, with matrix elements:

9
 hj jPjk i = e? E j ;k
[ ( E j )]
)+ (

which is a 2n  2n matrix. The partition function will be given by

= Tr (Pn )
and, like, in the one-dimensional case, the largest eigenvalue of P is sought. This is the nontrivial problem that is
worked out in 20 pages in Huang's book.
In the thermodynamic limit, the nal result at zero eld is:
Z   
g(T ) = ?kT ln [2cosh(2J )] ? kT d ln 1 1 + q1 ? K 2 sin2 
2 0 2
where
2
K = cosh(2J )coth(2 J )
The energy per spin is
Z 
"(T ) = ?2J tanh(2J ) + 2K dK d sin2 
d 0
(1 +
)
where
q

= 1 ? K 2 sin2 
The magnetization, then, becomes
n o1=8
m = 1 ? [sinh(2J )]?4
for T < Tc and 0 for T > Tc , indicating the presence of an order-disorder phase transition at zero eld. The condition
for determining the critical temperature at which this phase transition occurs turns out to be
2tanh2 (2J ) = 1
kTc  2:269185J
Near T = Tc, the heat capacity per spin is given by
 2      
C (t) = 2 2J ? ln 1 ? TT + ln kT c ? 1 +  
k  kTc c 2J 4
Thus, the heat capacity can be seen to diverge logarithmically as T ! Tc.
The critical exponents computed from the Onsager solution are
 = 0 (log divergence)
 = 81
= 74
 = 15
which are a set of exact exponents for the d = 2, n = 1 universality class.

10
 G25.2651: Statistical Mechanics

Notes for Lecture 27

I. THE EXPONENTS  AND 

Consider a spin-spin correlation function at zero eld of the form
hi j i = 1


i j e?H
X X
Q 1 N

If i and j occupy lattice sites at positions ri and rj , respectively, then at large spatial separation, with r = jri ? rj j,
the correlation function depends only r and decays exponentially according to
?r=
G(r)  hi j i ? hi ihj i  red?2+
for T < Tc. The quantity  is called the correlation length. Since, as a critical point is approached from above, long
range order sets in, we expect  to diverge as T ! Tc+ . The divergence is characterized by an exponent  such that
  jT ? Tcj?
At T = Tc, the exponential dependence of G(r) becomes 1, and G(r) decays in a manner expected for a system
with long range order, i.e., as some small inverse power of r. The exponent  appearing in the expression for G(r)
characterizes this decay at T = Tc.
The exponents,  and  cannot be determined from MFT, as MFT neglects all correlations. In order to calculate
these exponents, a theory is needed that restores uctuations at some level. One such theory is the so called Landau-
Ginzberg theory. Although we will not discuss this theory in great detail, it is worth giving a brief introduction to
it.

II. INTRODUCTION TO LANDAU-GINZBERG THEORY

The Landau-Ginzberg (LG) theory is a phenomenological theory meant to be used only near the critical point.
Thus, it is formulated as a macroscopic theory. The basic idea of LG theory is to introduce a spin density eld
variable S (x) dened by
N
X
S (x) = i (x ? xi )
i=1

Then the total magnetization is given by

Z
M= dd xS (x)
Since the free energy at zero eld is A = A(N; M; T ), there should be a corresponding free energy density A(T; S (x)) =
a(x) such that
Z
A(N; M; T ) = dd xa(x)
It is assumed that a(x) can be represented as a power series according to
a(x) = a0 + a1 (T )S 2(x) + a2 (T )S 4 (x) +

1
such that a1 and a2 are both positive for T > Tc and a1 vanishes at T = Tc. These conditions are analogous to those
exhibited by the total free energy in MFT above and near the critical point (see previous lecture). By symmetry, all
odd terms vanish and are, therefore, not explicitly included.
In the presence of a magnetic eld h(x), we have a Gibbs free energy density given by
g(x) = a(x) ? h(x)S (x) = a0 + a1 (T )S 2(x) + a2 (T )S 4(x) ? h(x)S (x)
In addition, a term a3 (rS )2 with a3 > 0 is added to the free energy in order to damp out local uctuations. If there
are signicant local uctuations, then (rS )2 becomes large, so these eld congurations contribute negligibly to the
partition function, which is dened by
Z  Z 
Z = D[S ] exp ? dd xg(x)
Z  Z 
D[S ] exp ? dd x a + a (T )S (x) + a (T )S (x) + a (rS ) ? h(x)S (x)
?

= 0 1
2
2
4
3
2

Thus, the LG theory is a eld theory. The form of this eld theory is well known in quantum eld theory, and is known
as a d-dimensional scalar Klein-Gordon theory. In terms of Z , a correlation function at zero eld can be dened by

2 ln Z 
hS (x)S (x0 )i = 1
 2 h(x)h(x0 ) h=0
Thus, by studying the behavior of the correlation function, the exponents  and  can be determined.
For the choice a2 = 0, the theory can be solved exactly analytically. This is known as the Gaussian model, for
which
Z  Z 
Z = D[S ] exp ? dd x a0 + a1 (T )S 2(x) + a3 (rS )2 ? h(x)S (x)
?

which leads to values of  and  of 0 and 1/2, respectively. These values are known as the \classical" exponents. They
are independent of the number of spatial dimensions d. Dependence on d comes in at higher orders in S . Comparing
these to the exact exponents from the Onsager solution, which gives  = 1=4 and  = 1, it can be seen that the
classical exponents are only qualitatively correct. Going to higher orders in the theory leads to improved results,
although the theory cannot be solved exactly analytically. It can either be solved numerically using path integral
Monte Carlo or Molecular Dynamics or analytically perturbatively using Feynman diagrams.

III. RENORMALIZATION GROUP AND THE SCALING HYPOTHESIS

A. General formulation
The renormalization group (RG) has little to do with \group theory" as it is meant mathematically. Also, there is
no uniqueness to the renormalization group, so the use of \the" in this context is misleading. Rather, the RG is an
idea that exploits the physics of systems near their critical point which leads to a procedure for nding the critical
point. It also oers an explanation of universality, perhaps the closest thing there is to a proof of this concept. Finally,
through the scaling hypothesis, it generates relations, called scaling relations satised by the critical exponents. It
does not allow actual determination of specic exponents. However, given a numerical calculation or some other
method of determining a small subset of exponents, the scaling relations can be used to determine the remaining
exponents.
In order to see how the RG works, we will consider a specic example. Consider a square spin lattice:

2
 . . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
FIG. 1.

which has been separated into 33 blocks as shown. Consider dening a new spin lattice from the old by application
of a coarse-graining procedure in which each 33 block is replaced by a single spin. The new spin is an up spin if the
majority of spins in the block is up and down is the majority point down. The new lattice is shown below:

3
 . .

. .
FIG. 2.

Such a transformation is called a block spin transformation. Near a critical point, the system will exhibit long-range
ordering, hence the coarse-graining procedure should yield a new spin lattice that is statistically equivalent to the old
spin lattice. If this is so, then the spin lattice is said to possess scale invariance.
What will be the Hamiltonian for the new spin lattice? To answer this, consider the partition function at h = 0 for
the old spin lattice using the Ising Hamiltonian as the starting point:
X X
Q=
e?H0 (1 ;:::;N )  Tr e?H0 (1 ;:::;N )
1 N

The block spin transformation can be expressed by dening, for each block, a transformation function:
 P9
T (0 ; 1 ; :::; 9 ) = 01  >0 0 i=1 i
otherwise
When inserted into the expression for the partition function, T acts to project out those congurations that are
consistent with the block spin transformation, leaving a function of only the new spin variables f10 ; :::; N0 g, in terms
0

of which a new partition function can be dened. To see how this works, let the new Hamiltonian be dened through:
" #
Y
e?H0 (f g) = Tr
0 0
T (0 ; 1 ; :::; 9 ) e?H0 (fg)
blocks

That this is a consistent denition follows from the fact that

X
T (0 ; 1 ; :::; 9 ) = 1
 0

Thus, tracing both sides of the projected partition function expression over 0 yields:
Tr e?H0 (f g) = Tr e?H0 (fg)
0 0
0

which states that the partition function is preserved by the block spin transformation, hence the physical properties
are also preserved.
Transformations of this type should be chosen so as to preserve the functional form of the Hamiltonian, for if this is
done, then the transformation can be iterated in exactly the same way for each new lattice produced by the previous
iteration. The importance of being able to iterate the procedure is that, in a truly ordered state formed at a critical
point, the iteration transformation will produce exactly the same lattice as the previous iteration, thus signifying
the existence of a critical point. If the functional form of the Hamiltonian is preserved, then only its parameters are
aected by the transformation, so we can think of the transformation as acting on these parameters. If the original
Hamiltonian contains parameters K1 ; K2 ; :::;  K (e.g., the J coupling in Ising model), then the transformation yields
a Hamiltonian with a new set of parameters K0 = (K10 ; K20 ; :::), such that the new parameters are functions of the old
parameters

4
 K0 = R(K)
The vector function R characterizes the transformation. These equations are called the renormalization group equa-
tions or renormalization group transformations. By iterating the RG equations, it is possible to determine if a system
has an ordered phase or not and for what values of the parameters such a phase will occur.

B. Example: The one-dimensional Ising model

For the one-dimensional Ising model:
N
X
H0 = ?J i i+1
i=1

Dene:
H = H
0

K = J
so that
N
X
H = ?K
0 i i+1
i=1

and the partition function becomes

Q = Tr e?H0
We will consider a simple block spin transformation as illustrated below:

2 3
1

σ1 σ2 σ3 σ1 σ2 σ3 σ1 σ2
. . . . . . . .
σ1 σ2 σ3 σ4 σ5 σ6 σ7 σ8
FIG. 3.

The gure shows the one-dimension spin lattice numbered in two dierent ways { one a straight numbering and
one using blocks of three spins, with spins in each block numbered 1-3. The block spin transformation to be employed
here is that the spin of a block will be determined by the value of the spin in the center of the block. Thus, for
block 1, it is the value of spin 2, for block 2, it is the value of spin 5, etc. This rather undemocratic choice should be
reasonable at low temperature, where local ordering is expected, and spins close to the center spin would be expected
to be aligned with it, anyway. The transformation function, T for this case is
T (0 ; 1 ; 2 ; 3 ) =  2
0

The new lattice will look like:

5
 . . .
σ1’ σ2 ’ σ 3’
FIG. 4.

with 10 = 2 , 20 = 5 , etc.

The new Hamiltonian is computed from
XXX X?

e?cHz (f g) =
0 0



1 2 2 5


eK1 2 eK2 3 eK3 4 eK4 5



0 0

1 2 3 N
XXXX
=


eK1 1 eK1 3 eK3 4 eK4 2



0 0 0

1 3 4 6

The idea is then to nd a K 0 such that when the sum over 3 and 4 are performed, the new interaction between 10
and 20 is of the form exp(K 0 10 20 ), which preserves the functional form of the old Hamiltonian. The sum over 3 and
4 is
XX
eK1 3 eK3 4 eK4 2
0 0

3 4

Note that 3 4 = 1. Then, since

e = cosh + sinh = cosh [1 + tanh]
e? = cosh ? sinh = cosh [1 ? tanh]
we can express exp(K3 4 ) as
eK3 4 = coshK [1 + 3 4 tanhK ]
Letting x = tanhK , the product of the three exponentials becomes:
eK1 3 eK3 4 eK4 2 = cosh3 K (1 + 10 3 x)(1 + 3 4 x)(1 + 4 20 x)
0 0

= cosh3 K (1 + 10 3 x + 3 4 x + 4 20 x + 10 32 4 x2 + 10 3 4 20 x2 + 3 42 20 x2 + 10 32 42 20 x3 )
When summed over 3 and 4 , most terms in the above expression will cancel, yielding the following expression:
XX  
eK1 3 eK3 4 eK4 2 = 2cosh3 K 1 + 10 20 x3  coshK 0 [1 + 10 20 x0 ]
0 0

3 4

where the last expression puts the interaction into the original form with a new coupling constant K 0. One of the
possible choices for the new coupling constant is
tanhK 0 = tanh3 K 
K 0 = tanh?1 tanh3 K
This, then, is the RG equation for this particular block spin transformation. With this identication of K 0, the new
Hamiltonian can be shown to be
X N0
H0 (f0 g) = N 0 g(K ) ? K 0 0 0
0 i i+1
i=1

where the spin-independent function g(K ) is given by

6
  
g(K ) = ? 13 ln cosh
3
K ? 2 ln 2
coshK 0 3
Thus, apart from the additional term, the new Hamiltonian is exactly the same functional form as the original
Hamiltonian but with a dierent set of spin variables and a dierent coupling constant.
The transformation can now be applied to the new Hamiltonian, yielding the same relation between the new and
old coupling constants. This is equivalent to iterating the RG equation. Since the coupling constant K depends on
temperature through K = J=kT , the purpose of the iteration would be to nd out if, for some value of K , there is an
ordered phase. In an ordered phase, the transformed lattice would be exactly the same as the old lattice, and hence
the same coupling constant and Hamiltonian would result. Such points are called xed points of the RG equations
and are generally given by the condition
K = R(K)
The xed points correspond to critical points. For the one-dimensional Ising model, the xed point condition is
 
K = tanh?1 tanh3 K
or, in terms of x = tanhK ,
x = x3
Since K is restricted to K  0, the only solutions to this equation are x = 0 and x = 1, which are the xed points of
the RG equation.
To see what these solutions mean, consider the RG equation away from the xed point:
x0 = x 3
Since K = J=kT , at high T , K ! 0 and x = tanhK ! 0+ . At low temperature, K ! 1 and x ! 1?. Viewing the
RG equation as an iteration or recursion of the form
xn+1 = x3n
if we start at x0 = 1, each successive iteration will yield 1. However, for any value of x less than 1, the iteration
eventually goes to 0 in some nite (though perhaps large) number of iterations of the RG equation. This can be
illustrated pictorially as shown below:

Unstable Stable
. .
x=1 x=0
K= K=0
FIG. 5.

The iteration of the RG equation produces an RG ow through coupling constant space. The xed point at x = 1 is
called an unstable xed point because any perturbation away from it, if iterated through the RG equation, ows away
from this point to the other xed point, which is called a stable xed point. As the stable xed point is approached,
the coupling constant gets smaller and smaller until, at the xed point, it is 0. The absence of a xed point for any
nite, nonzero value of temperature tells us that there can be no ordered phase in one dimension, hence no critical
point in one dimension. If there were a critical point at a temperature Tc , then, at that point, long range order would
set it, and there would be a xed point of the RG equation at Kc = J=kTc. Note, however, that at T = 0, K = 1,
there is perfect ordering in one dimension. Although this is physically meaningless, it suggests that ordered phases
and critical points will be associated with the unstable xed points of the RG equations.

7
 Another way to see what the T = 0 unstable xed point means is to study the correlation length. The correlation
is a quantity that has units of length. However, if we choose to measure it in units of the lattice spacing, then the
correlation length will be a number that can only depend on the coupling constant K or x = tanhK :
 =  (x)
Under an RG transformation, the lattice spacing increases by a factor of 3 as a result of coarse graining. Thus, the
correlation length, in units of the lattice spacing, must decrease by a factor of 3 in order for the same physical distance
to be maintained:
 (x0 ) = 31  (x)
In general, for block containing b spins, the correlation length transforms as
 (x0 ) = 1  (x)
b
A function satisfying this equation is
 (x)  ln1x
Since, for arbitrary b, the RG equation is
x0 = x b
we have
 (x0 ) =  (xb )
 ln1xb
= b ln1 x
= 1  (x)
b
so that
1
 (K )  ln tanhK ?! 1 asT ?! 0
so that at T = 0 the correlation length becomes innite, indicating an ordered phase.
Note, also, that at very low T , where K is large, motion toward the stable xed point is initially extremely slow.
To see this, rewrite the RG equation as
tanhK 0 = tanh3 K
= tanhK tanh

2
K 
cosh(2 K )
= tanhK cosh(2K ) + 1 ? 1

Notice that the term in brackets is extremely close to 1 if K is large. To leading order, we can say
K0  K
Since the interactions between blocks are predominantly mediated by interactions between boundary spins, which, for
one dimension, involves a single spin pair between blocks, we expect that a block spin transformation in 1 dimension
yields a coupling constant of the same order as the original coupling constant when T is low enough that there is
alignment between the blocks and the new lattice is similar to the original lattice. This is re ected in the above
statement.

8
 G25.2651: Statistical Mechanics

Notes for Lecture 28

I. FIXED POINTS OF THE RG EQUATIONS IN GREATER THAN ONE DIMENSION

In the last lecture, we noted that interactions between block of spins in a spin lattice are mediated by boundary
spins. In one dimension, where there is only a single pair between blocks, the block spin transformation yields a
coupling constant that is approximately equal to the old coupling constant at very low temperature, i.e., K 0  K .
Let us now explore the implications of this fact in higher dimensions.
Consider a two-dimensional spin lattice as we did in the previous lecture. Now interactions between blocks can be
mediated by more than a single pair of spin interactions. For the case of 33 blocks, there will be 3 boundary spin
pairs mediating the interaction between two blocks:

. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
FIG. 1.

1
 Since 3 boundary spin pair interactions mediate the block-block interaction, the result of a block spin transformation
should yield, at low T , a coupling constant K 0 roughly three times as large as the original coupling constant, K :
K 0  3K
In a three-dimensional lattice, using 3  3  3 blocks, there would be 32 = 9 spin pairs. Generally, in d dimensions
using blocks containing bd spins, the RG equation at low T should behave as
K 0  bd?1K T ! 0; K ! 1
The number b is called the length scaling factor. The above RG equation implies that for d > 1, K 0 > K for low T .
Thus, iteration of the RG equation at low temperature should now ow toward K = 1 and the xed point at T = 0
is now a stable xed point. However, we know that at high temperature, the system must be in a paramagnetic state,
so the xed point at T = 1 must remain a stable xed point. These two facts suggest that, for d > 1, between T = 0
and T = 1, there must be another xed point, which will be an unstable xed point. To the extent that an RG ow
in more than one dimension can be considered a one-dimensional ow, the ow diagram would look like:

Unstable Stable
Stable
. . .
x=1 K=K c x=0
K= K=0
T=T c
FIG. 2.

Any perturbation to the left of the unstable xed point iterates to T = 0, K = 1 and any perturbation to the
right iterates to T = 1 and K = 0. The unstable xed point corresponds to a nite, nonzero value of K = Kc and a
temperature Tc , and corresponds to a critical point.
To see that this is so, consider the correlation length evolution of the correlation length under the RG ow. Recall
that for a length scaling factor b, the correlation length transform as
 (K 0 ) = 1  (K )
b
 (K ) = b (K 0 )
Suppose we start at a point K near Kc and require n(K ) iterations of the RG equations to reach a value K0 between
K = 0 and K = Kc under the RG ow:

Unstable Stable
Stable
. . . .
x=1 K=K c K0 x=0
K= K=0
T=T c
FIG. 3.

2
 If 0 is the correlation length at K = K0, which can expect to be a nite number of order 1, then, by the above
transformation rule for the correlation length, we have
 (K ) = 0 bn(K )
Now, recall that, as the starting point, K is chosen closer and closer to Kc, the number of iterations needed to reach
K0 gets larger and larger. (Recall that near an unstable xed point, the initial change in the coupling constant is
small as the iteration proceeds). Of course, if K = Kc initially, than an innite number of iterations is needed. This
tells us that as K approaches Kc, the correlation length becomes innite, which is what is expected for an ordered
phase. Thus, the new unstable xed point must correspond to a critical point.
In fact, we can calculate the exponent  knowing the behavior of the RG equation near the unstable xed point.
Since this is a xed point, Kc satises, quite generally,
Kc = R(Kc)
Near the xed point, we can expand the RG equation, giving
K 0  R(Kc) + (K ? Kc)R0 (Kc ) +



Dene an exponent y by
0
y = ln Rln(bKc )
so that
K 0  Kc + by (K ? Kc )
Near the critical point,  diverges according to
?  ?  ? K ?
 K1 ? K1    K ?K Kc    K K
     

  jT ? Tcj? c
  


c c 
Thus,
  jK ? Kcj?
but
 (K ) = b (K 0 )
which implies
jK ? Kc j?  bjK 0 ? Kc j?
= bjby (K ? Kc)j?
which is only possible if
 = y1
This result illustrates a more general one, namely, that critical exponents are related to derivatives of the RG trans-
formation.

II. GENERAL LINEARIZED RG THEORY

The above discussion illustrates the power of the linearized RG equations. We now generalize this approach to a
general Hamiltonian H0 with parameters K1; K2 ; :::;  K. The RG equation
K0 = R(K)
3
can be linearized about an unstable xed point at bK  according to
Ka0 ? Ka  Tab (Kb ? Kb)
X

b
where


Tab = @R a 
@Kb K=K
The matrix T need not be a symmetric matrix. Given this, we dene a left eigenvalue equation for T according to
X
ia Tab = i ib
a
where the eigenvalues fg can be assumed to be real (although it cannot be proved). Finally, dene a scaling variable,
ui by
ia (Ka ? Ka )
X
ui =
a
They are called scaling variables because they transform multiplicatively near a xed point under the linearized RG
ow:
u0i = ia (Ka0 ? Ka )
X

a
ia Tab (Kb ? Kb)
XX
=
a b
i ib (Kb ? Kb )
X
=
b
= i ui
Since ui scales with i , it will increase if i > 1 and will decrease if i < 1. Redening the eigenvalues i according
to
i = b y i

we see that
u0i = by ui i

By convention, the quantities fyig are called the RG eigenvalues. These will soon be shown to determine the scaling
relations among the critical exponents.
For the RG eigenvalues, three cases can be identied:
1. yi > 0. The scaling variable ui is called a relevant variable. Repeated RG transformations will drive it away
from its xed point value, ui = 0.
2. yi < 0. The scaling variable ui is called an irrelevant variable. Repeated RG transformations will drive it toward
0.
3. yi = 0. The scaling variable ui is called a marginal variable. We cannot tell from the linearized RG equations
if ui will iterate toward or away from the xed point.
Typically, scaling variables are either relevant or irrelevant. Marginality is rare and will not be considered here. The
number of relevant scaling variables corresponds to the number of experimentally `tunable' parameters or `knobs'
(such as T and h in the magnetic system, or P and T in a uid system; in the case of the former, the relevant
variables are called the thermal and magnetic scaling variables, respectively).

4
 III. UNDERSTANDING UNIVERSALITY FROM THE LINEARIZED RG THEORY
In the linearized RG theory, at a xed point, all scaling variables are 0, whether relevant, irrelevant or marginal.
Consider the case where there are no marginal scaling variables. Recall, moreover, that irrelevant scaling variables
will iterate to 0 under repeated RG transformations, starting from a point near the unstable xed point, while the
relevant variables will be driven away from 0. These facts provide us with a formal procedure for locating the xed
point:
i. Start with the space spanned by the full set of eigenvectors of T.
ii. Project out the relevant subspace by setting all the relevant scaling variables to 0 by hand.
iii. The remaining subspace spanned by the irrelevant eigenvectors of T denes a hypersurface in the full coupling
constant space. This is called the critical hypersurface.
iv. Any point on the critical hypersurface belongs to the irrelevant subspace and will iterate to 0 under successive
RG transformations. This will dene a trajectory on the hypersurface that leads to the xed point as illustrated
below:

K 3

Fixed point

. K2

K1
FIG. 4.

5
 This xed point is called the critical xed point. Note that it is stable with respect to irrelevant scaling variables
and unstable with respect to relevant scaling variables.
What is the importance of the critical xed point? Consider a simple model in which there is one relevant and one
irrelevant scaling variable, u1 and u2 , with corresponding couplings K1 and K2 . In an Ising-type model, K1 might
represent the reduced nearest neighbor coupling and K2 might represent a next nearest neighbor coupling. Relevant
variables also include experimentally tunable parameters such as temperature and magnetic eld. The reason u1 is
relevant and u2 is irrelevant is that there must be at least nearest neighbor coupling for the existence of a critical
point and ordered phase at h = 0 but this can happen whether or not there is a next nearest neighbor coupling. Thus,
the condition
u1 (K1 ; K2) = 0
denes the critical surface, in this case, a one-dimensional curve in the K1 -K2 plane as illustrated below:
K 2

K1
FIG. 5.

Here, the blue curve represents the critical \surface" (curve), and the point where the arrows meet is the critical
xed point. The full coupling constant space represents the space of all physical systems containing nearest neighbor
and next nearest neighbor couplings. If we wish to consider the subset of systems with no next nearest neighbor
coupling, i.e., K2 = 0, the point at which the line K2 = 0 intersects the critical surface (curve) denes the critical
value, K1c and corresponding critical temperature and will be an unstable xed point of an RG transformation with
K2 = 0. Similarly, if we consider a model for which K2 6= 0, but having a xed nite value, then the point at which
this line intersects the critical surface (curve) will give the critical value of K1 for that model. For any of these
models, K1c lies on the critical surface and will, under the full RG transformation iterate toward the critical xed
point. This, then, denes a universality class: All models characterized by the same critical xed point belong to
the same universality class and will share the same critical properties. This need not include only magnetic systems.
Indeed, a uid system, near its critical point can be characterized by a so called lattice gas model. This model is
similar to the Ising model, except that the spin variables are replaced by site occupance variables ni , which can take
on values 0 or 1, depending on whether a given lattice site is occupied by a particle or not. The grand canonical
partition function is

6
 P P
Z=
X




X
 i
ni e2 hi;j i
Jij ni nj
n1 =0;1 nN =0;1
and hence belongs to the same universality class as the Ising model. The critical surface, then, contains all physical
models that share the same universality properties and have the same critical xed point.
In order to see how the relevant scaling variables lead to the scaling relations among the critical exponents, we next
need to introduce the scaling hypothesis.

IV. THE SCALING HYPOTHESIS

Recall that the RG transformation preserves the partition function:
Tr0 e?H0 (f g;K ) = Tr e?H0 (fg;K )
0 0 0

For the one-dimensional Ising model, we found that the block spin transformation lead to a transformed Hamiltonian
of the form
N 0

H0 (f0 g; K 0) = N 0 g(K ) ? K 0 i0 i0+1

X
0
i=1
Thus, H0 (f0 g; K 0) contains a term that is of the same functional form as H0 (fg; K ) plus an analytic function of K .
Dening the reduced free energy per spin from the spin trace as f (K ), the equality of the partition functions allows
us to write generally:
e?Nf (fK g) = Tr e?H0 (fg;fK g) = e?N 0 g(fK g) Tr e?H0 (f0 g;fK 0 g) = e?N 0 g(fK g) e?N 0 f (fK 0 g)
which implies that
0
f (fK g) = g(fK g) + NN f (fK 0g)
If b is the size of the spin block, then
N 0 = b?dN
from which
f (fK g) = g(fK g) + b?df (fK 0g)
Now, g(fK g) is an analytic function and therefore plays no role in determining critical exponents, since it does not
lead to divergences. Only the so called singular part of the free energy is important for this. Thus, the singular part
of the free energy fs (fK g) can be seen to satisfy a scaling relation:
fs (fK g) = b?d fs (fK 0g)
This is the basic scaling relation for the free energy. From this simple equation, the scaling relations for the critical
exponents can be derived.
To see how this works, consider the one-dimensional Ising model again with h 6= 0. The free energy depends on
the scaling variables through the dependence on the couplings K . For h = 0, we saw that there was a single relevant
scaling variable corresponding to the nearest neighbor coupling K = J=kT . This variable is temperature dependent
and is called a thermal scaling variable ut . For h 6= 0 there must also be a magnetic scaling variable, uh. These will
transform under the linearized RG as
u0t = by ut t

u0h = by uh h

where yt and yh are the relevant RG eigenvalues. Therefore, the scaling relation for the free energy becomes
fs (ut ; uh) = b?dfs (u0t ; u0h ) = b?dfs (by ut ; by uh) t h

7
Now, after n iterations of the RG equations, the free energy becomes
fs (ut ; uh) = b?nd fs (bny ut ; bny uh) t h

Recall that relevant scaling variables are driven away from the critical xed point. Thus, let us choose an n small
enough that the linear approximation is still valid. In order to determine n, we only need to consider one of the scaling
variables, so let it be ut . Thus, let ut0 be an arbitrary value of ut obtained after n iterations of the RG equation, such
that ut0 is still close enough to the xed point that the linearized RG theory is valid. Then,
ut0 = bny ut t

or
  1=y  

n = y1 logb  uut0  = logb  uut0 

t
   

t t t
and
 d=y
fs (ut ; uh ) =  uut  fs (ut0; uh jut =ut0j?y
t
 
h =yt )
t0
Now, let
t = T ?T Tc
We know that fs must depend on the physical variables t and h. In the linearized theory, the scaling variables ut and
uh will be related linearly to the physical variables:
ut t
ut0 = t0
uh = hh
0

Here, t0 and h0 are nonuniversal proportionality constants, and we see that ut ! 0 when t ! 0 and the same for uh.
Then, we have
ut0 ; jt=th=h
 

fs (t; h) = jt=t0jd=yt fs 0

0j
y =y h t

The left side of this equation does not depend on the nonuniversal constant ut0, hence the right side cannot. This
means that the function on the right side depends on a single argument. Thus, we rewrite the free energy equation as
h=h0
 

fs (t; h) = jt=t0 jd=yt  jt=t0 jy =y

h t

The function  is called a scaling function. Note that the dependence on the system-particular variables t0 and h0
is trivial, as this only comes in as a scale factor in t and h. Such a scaling relation is a universal scaling relation,
in that it will be the same for all systems in the same universality class. From the above scaling relation come all
thermodynamic quantities. Note that the scaling relation depends on only two exponents yt and yh. This suggests
that there can only be two independent critical exponents among the six, ; ; ; ;  , and . There must, therefore,
be four relations relating some of the critical exponents to others. These are the scaling relations. To derive these,
we use the scaling relation for the free energy to derive the thermodynamic functions:
1. Heat Capacity:


Ch  @@tf2   jtjd=y ?2
2 
t

h=0

but

8
 Ch  jtj?
Thus,
 = 2 ? yd
t

2. Magnetization:
m = @f




 jt=t0 jd=y  jtj(d?y t

h) =yt
@h h=0 jt=t0 jy =y h t

but
m  jtj
Thus,
 = d ?y yh
t

3. Magnetic susceptibility:
 = @m  @ 2 f  jtj(d?2y
@h @h2
h)=yt

but
  jtj?
Thus,
= 2yhy? d
t

4. Magnetization vs. magnetic eld:

m = @f h=h0
 
(d?y =yt 0
@h = jt=t0 j
h)

jt=t0 jy =y
h t

but m should remain nite as t ! 0, which means that 0 (x) must behave as xd=y ?1 as x ! 1, since then
h

m  jt=t0 j(d?y h)=yt (h=h0)(d?y=yt h)

jt=t0jyh (d?yh)=(yt yh )
(d?yh )=yh
 (h=h 0 )
But
m  h1=
Thus,
 = d ?yhy
h

9
 From these, it is straightforward to show that the four exponents ,  , , and  are related by
 + 2 + = 2
 +  (1 + ) = 2
These are examples of scaling relations among the critical exponents.
The other two scaling relations involve  and  and are derived from the scaling relation satised by the correlation
function:
G(r) = b?2(d?y ) G(r=b; by t)
h t

r
 

G(r) = jt=t0 j2(d?y =yt

h)

jt=t j? =y 1 t
0

This leads to the relations

 = y1
t
 = d + 2 ? 2yh
and the scaling relations:
 = 2 ? d
=  (2 ? )

10