Chapter 4

Calculations of Time Correlation

Functions

In the last chapter, we explored the low-frequency, long-wavelength behavior of a system that is
disturbed from equilibrium. In the first section of this chapter, we study the opposite limit, and
describe how a system behaves at very short times and high frequencies. The study of systems in
this limit is referred to as Molecular dynamics.

We are ultimately interested in developing a set of expressions that describe a system at all times
and frequencies. In section 2, we will introduce the projection operator and use it to derive the
Generalized Langevin Equation. The projection operator allows us to study only the portion of the
system we are interested in, and treat the rest as a statistical bath. In section 3, we will use the
GLE to derive the viscoelastic model for transverse current. Finally, in section 4, we will introduce
mode-coupling theory and discuss its ability to predict long-time tails in velocity correlation func-
tions.

For further information on the subjects covered in this chapter, please consult books by Hansen
and McDonald[1], McQuarrie[2], Boon and Yip[3], and Berne [4].

4.1 Short-Time Behavior

4.1.1 Moment Expansion
In chapter 2, we introduced the concept of the time-correlation function. The correlation function
for an operator A(t) is given by

C(t) = ⟨A(t)A(0)⟩ = T rA(t)A(0)ρeq (4.1)

Where the equilibrium density matrix is given by

e−βH
ρeq =
Q

Here H and Q are the Hamiltonian and Partition function for the system. The time evolution of
A is given by
A(t) = eiLt A(0)

1
Chapter 4. Calculations of Time Correlation Functions 2

or
Ȧ(t) = iLA(0)
Here, L is an operator which describes the time evolution of an operator. For quantum mechanical
systems, L is defined as the Liouville operator
1
iL = [. . . , H]
i}
And for classical systems it is defined as the Poisson operator

iL = {. . . , H}

The evolution operator L is Hermitian, L+ = L. This operator will be discussed in much more
detail in section 4.2.

The value of a correlation function in the short time limit t → 0 can be approximated using a
moment expansion. As shown in Eq.(4.1), the correlation function of a quantity A(t) is given by

C(t) = ⟨A(t)A(0)⟩

This quantity C(t) can be written as a Taylor expansion

∞ n
∑ t
C (t) = C (n) (0)
n!
n=0

This formula can be simplified by noting that all correlation functions are even in time. As a result,
any odd-valued derivative of C(t) will be zero when evaluated at t = 0. Therefore, all of the odd
terms of this expansion can be dropped
∞
∑ t2n (2n)
C (t) = C (0)
(2n)!
n=0

The derivative of a correlation function can be written as

⟨ ⟩
C (2n) (t) = (−1)n A(2n) (t)A(0)

Using this expression, the Taylor expansion can be written in terms of the function A(t)

t2n ⟨ (2n) ⟩
∞
∑
C(t) = (−1)n A (0)A(0)
(2n)!
n=0
⟨ ⟩ ⟨ ⟩
This expression can be further simplified using the definition A(2n) (0)|A(0) = − A(n) (0)|A(n) (0) ,
where the notation ⟨A|B⟩ = ⟨AB + ⟩ = T rAB + ρeq

t2n ⟨ (n) (n) ⟩

∞
∑
C (t) = A A (4.2)
(2n)!
n=0

In this expression, we are only concerned with the value of A(t) at time 0, and so the explicit
time dependence has been dropped. This expression could also be obtained by performing a Taylor
expansion on A(t) and substituting it into Eq.(4.1).

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 3

We can use the Fourier transform of C(t) to find a general expression for C(2n) . Since
∫ ∞
1
C (2n)
= C̃(ω)e−iωt dt
2π −∞

the time derivatives can be evaluated easily as

( )(2n) ∫ ∞
(2n) (n) ∂ 1 ⟨ ⟩
C = (−1) C(t) = C̃(ω)ω 2n dt = ω 2n (4.3)
∂t 2π −∞

In the next sections, this method is applied to velocity correlation functions and self-scattering
functions.

4.1.2 Velocity Correlation Function and Self-Scattering Functions

1. The Velocity Correlation Function The velocity correlation function for the z -directed
motion of a particle is defined as
1
C(t) = ⟨⃗v (t)⃗v (0)⟩ = ⟨ż(t)ż(0)⟩
3
This expression can be evaluated using Eq.(4.2). At short times, the value of C(t) can be reasonably
approximated by taking the first two moments of the expansion

1 t2 1 ⟨ ˙ ˙ ⟩
C(t) = ⟨⃗v |⃗v ⟩ − ⃗v |⃗v + · · ·
3 23
The first moment is simply the the average thermal velocity in the z -direction
1 1
⟨⃗v |⃗v ⟩ = = vo2
3 βm

where β = (kB T )−1 . The second moment can be evaluated using Newton’s equation, F = ma.
Since a = v̇ and F = −∇U , where U is the potential energy, v̇ = m F
= − ∇U
m . Therefore, the second
moment is given by
1 ⟨ ˙ ˙ ⟩ 1 ⟨∇U |∇U ⟩
⃗v |⃗v =
3 3 m2
To evaluate this expression, write it in its explicit form
∫
1 ⟨∇U |∇U ⟩ 1
= dz ∂z U ∂z U e−βU
3 m2 3m2

Note that ∂z U e−βU = − β1 ∂z e−βU . This allows us to combine terms in the integral to get the
expression: ∫
1 1
2
dz∂z U ( ∂z e−βU )
3m β
Now, carry out a partial integration to get the expression:
∫ ⟨ 2 ⟩
1 2 −βU 1 ∂z U
2
dz ∂z U e =
3βm 3βm m

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 4

Note that we have proven a general property here. For any operator A
∫
1
⟨∇U A⟩ = drA∇U e−βU = kB T ⟨∇A⟩ (4.4)
Q
⟨ 2 ⟩
We have shown that the second term in the expansion of C(t) is proportional to ∂m zU
, the
curvature of the potential averaged with the Boltzmann weight. This term is called the Einstein
frequency Ω2o . It is the average collision frequency of the particles in the system. For the specific
case of a harmonic potential this is simply the frequency ω 2 . However, it can be defined for many
types of systems. Simply find the collision frequency for each pair of particles in the system and
sum over all pairs. For the velocity correlation function, this can be expressed as
∫
1 ⟨ 2 ⟩ ρ
2
Ωo = ∇ U = drg(r)∇2 ϕ (4.5)
3m 3m

where ϕ is the pairwise potential between each set of two particles.

Finally, we can write second moment the expansion of C(t) as

( )
t2 2
C(t) ≃ vo2 1 − Ω0
2

2. Self-Intermediate Scattering Function The moment expansion method of estimating the

short time behavior of correlation functions can also be applied to self-scattering functions. In
chapter 3, we introduced the self-density of a particle i as

ns (R, t) = δ(R − r(t))

Which has the Fourier transform

⃗ r(t)
⃗ t) = e−ik⃗
ns (k,
The self-intermediate scattering function is defined as
⟨ ⟩ ⟨ ⟩ ⟨ ⟩
⃗ 0) = e−ik⃗⃗ r(t) −ik⃗
⃗ ⃗
⃗ t) = ns (k,
Fs (k, ⃗ t)|ns (k, |e r(0) = e−ik(⃗r(t)−⃗r(0))

We can apply Eq.(4.2) to estimate the short time behavior of this function. The zero-th moment
term is trivial to evaluate: ⟨ ⟩
⃗ r(0)−r(0))
⃗ 0) = e−ik(⃗ ⃗
C0 = Fs (k, =1

The second order term is given by

⟨ ⟩ ⟨ ⟩
⃗ r˙ (0)e−ik⃗ ⃗r(0) | − ik⃗
C2 = ω 2 = ⟨ṅs |ṅs ⟩ = −ik⃗
⃗ r(0)
⃗ r˙ (0)e−ik⃗

This can be simplified to ⟨ ⟩

⃗ r(0)−⃗
⃗ v (0))2 e−ik(⃗ r(0))
C2 = (k⃗ = k 2 vo2

We can defined ωo = kvo , which gives second moment of the correlation function

C2 = ωo2

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 5

The fourth moment of this correlation function is given by

⟨ ⟩
⟨ 4⟩ d ⃗˙ −i⃗k⃗ d ⃗˙ ⃗ r(0)
−ik⃗
C4 = ω = ⟨n̈s |n̈s ⟩ = −i (k⃗r(0)e r(0)
)| − i (k⃗r(0)e )
dt dt
Evaluate these derivatives using the product rule and multiply out the terms. The resulting equation
will have four terms, two of which cancel out. The remaining two terms are
⟨ ⟩
⃗ v˙ )2
⃗ v )4 + (k⃗
C4 = (k⃗

The first term is simply 3ωo4 . The second term can be evaluated by following a similar method to
the one we used to calculate the second moment of the velocity correlation function in the previous
section. As we demonstrated in that problem, The derivative of the velocity ⃗v is equivalent to the
derivative of the potential divided by the mass. Therefore, this term can be written as
⟨ ⟩
⃗ v˙ )2 = 1 k 2 ⟨∇z V ∇z V ⟩
(k⃗
m2
Using Eq.(4.4), we can rewrite this term as
1 ⟨ ⟩
2
kB T k 2 ∇2z V
m
Finally, by doing some rearranging and using vo2 = kBmT , we find that this term can be written as
⟨ ⟩ ⟨ 2 ⟩
kB T 2 ∇2z V 2 2 ∇z V
k = k vo = ωo2 Ω2o
m m m

Where Ω2o is the Einstein frequency, as defined in the previous section. Therefore, the short time
⃗ t),
expansion of Fs (k,
⟨ ⟩ 2 ⟨ ⟩ 4
⃗ t) = 1 − ω 2 t + ω 4 t − · · ·
Fs (k,
2! 4!
can be evaluated to ( ) t2 ( ) t4
Fs ⃗k, t = 1 − ωo2 + 3ωo4 + ωo2 Ω2o − ··· (4.6)
2! 4!

3. Free-Particle Limit (Ideal Fluid) We can use the short time expansion of the self-
intermediate scattering function to find an expression for Fs (k, ⃗ t) in the free-particle limit. In
the free-particle limit, we assume that the particles behave as an ideal gas; that is, there is no at-
traction or repulsion between the particles, and their interaction potential is zero ϕ(⃗r) = 0. Recall
that the Einstein frequency can be written as (Eq.(4.5))
∫
ρ
2
Ωo = d⃗rg(⃗r)∇2 ϕ(⃗r)
3m
Therefore, if the interaction potential is zero, the Einstein frequency will also be zero. Our expansion
⃗ t) becomes
for Fs (k,
2 4
⃗ t) = 1 − ω 2 t + ω 4 t − · · ·
Fs (k, o o
2! 8
This is simply the short time expansion for the function
⃗ t) = e− 21 ωo2 t2
Fs (k,

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 6

For free particles the self intermediate scattering function takes on a Gaussian form.

Only ideal systems can by truly described with the free-particle model. However, there are many
real systems that also show this limiting behaviour. Using these results, we can find the condition
for a system that will allow us to ignore the effects of molecular collisions. From Eq.(4.6), we can
see that the scattering function will take on a Gaussian form when Ω2o = 0 (the ideal case) or when
ωo2 Ω2o is sufficiently smaller than 3ωo4 that it can be ignored. Therefore, the condition for ignoring
collisions can be written as
Ω2o ≪ 3ωo2
Using the definitions of ωo2 and vo2 and rearranging, we find

Ωo
k≫√
3kB T
m

Now, define the parameter l as √

3kB T
l= Ωo
m
√
3kB T
This term gives the average thermal velocity, m , divided by the average collision frequency
Ωo . Therefore, it can be interpreted as the mean free path of the particles, or the average distance
a particle can travel before experiencing a collision. With the definition of l in hand, we can rewrite

1
k≫
l
or
λ≪l (4.7)
This indicates that a system can be treated in the free-particle limit when the wavelength, or
spatial range, that it used to investigate the system is less than the mean free path travelled by
the particles. For further discussion of self-intermediate scattering functions, please see Dynamics
of the Liquid State by Umberto Balucani [5].

4.1.3 Collective Properties

1. Density Fluctuations We can extend our previous discussion of the self-density function
ns (⃗r, t) by considering the density function ρ, which is simply a sum of self-density functions
∑
ρ(⃗r, t) = δ(⃗r − ⃗ri (t))
i

We define the density fluctuation as

δρ(⃗r, t) = ρ(⃗r, t) − ⟨ρ⟩

The Fourier transform of the density fluctuation is given by

∑ ⃗
ρk (t) = e−ik⃗ri (t) − (2π)3 δ(k)ρ
⃗ o
i

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 7

where ρo = ⟨ρ⟩. Then, we define the intermediate scattering function as the correlation function of
ρk (t)
⟨ ⟩
⃗ t) = 1 ⟨ρk (t)|ρk (0)⟩ = 1 ρ(k,
F (k, ⃗ t)|ρ(−k,
⃗ 0)
N N
Once again, we can find an expression for the short time behavior of F (k,⃗ t) using the moment
expansion in equation Eq.(4.2). We can find the zeroth moment of F (k, ⃗ t) by substituting in the
definition of ρk (t) and solving at time t = 0.
⟨ ⟩
∑ ⃗ rj (0) ∑ −ik⃗
⃗ 0) = 1
C0 = F (k, e ik⃗
e
⃗ ri (0)
N
j i

Note that when we consider the correlation of a particle with itself (that is, when i = j), the terms
in the exponentials will cancel, giving a value of 1. Summing over all N particles gives a value of
N. Therefore, we can write the zeroth moment as
⟨ ⟩
1 ∑ −ik⃗ ⃗ rij ⃗ o
C0 = 1 + e (2π)3 δ(k)ρ
N
̸ j
i=

where ⃗rij = ⃗ri (0) − ⃗rj (0). In Chapter 3, we defined the second term as ρo g(⃗r), where g(⃗r), is the
pair distribution function. The zeroth moment becomes
⃗ o = 1 + ρo h
C0 = 1 + ρo g (⃗r) − (2π)3 δ(k)ρ ˜ = S(k)
⃗

⃗ is the static structure factor. From thermodynamics, we know that

where S(k)

S(0) = 1 + ρo kB T χT 6 1

where χT is the isothermal compressibility,

1 ∂ρ
χT =
ρo ∂t
The pairwise correlation functions arises from the real space Van Hove Correlation function
⟨ ⟩
1 ∑
G(⃗r, t) = ⃗r(0) − ⃗rij (0) − ρo = ⟨δρ(⃗r, t)ρ(⃗r, 0)⟩
N
i,j

At time t = 0, the Van Hove function becomes

G(⃗r, 0) = δ(⃗r) + ρo h(⃗r)

where g = 1 + h

2. The Short time expansion In the previous section, we demonstrated that the zeroth
moment C0 of the short-time expansion of the intermediate scattering function is given by the
⃗ Therefore, we can write
static structure factor S(k).

⟨ ⟩ 2 ⟨ ⟩ 4
⃗ − ω2 t + ω4 t − · · ·
⃗ t) = S(k)
F (k,
2! 4!

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 8

To evaluate the second and fourth moments, we will consider the interactions of each particle with
itself (the self-part, i = j) separately from the interactions of each particle with other particles (the
distinct part, i ̸= j). To evaluate the self-part, use the results from section 2:

∑
N
ṅk = ⃗ vi )e−i⃗k⃗ri (t)
−i(k⃗
i=1

N [
∑ ]
⃗ ri (t)
⃗ v˙ i ) e−ik⃗
⃗ vi )2 − i(k⃗
n̈k = −(k⃗
i=1
Then we can evaluate the second moment of the self-part as
⟨ ⟩ 1 ⟨ ⟩
C2 = ω 2 = ⃗ v˙ 2 = ω 2
⟨ṅk |ṅk ⟩ = (k⃗) o
N
This gives the entire value of the second moment because the i = ̸ j terms do not contribute. The
fourth moment is given by
⟨ ⟩ 1 ∑ ⟨[ ⃗ 2 ⃗ 2 ⃗ v˙ i )(k⃗ ⃗ v˙ j )(k⃗
⃗ vj )2 − i(k⃗ ⃗ v˙ i )(k⃗
⃗ vk )2 + (k⃗
]
⃗ v˙ j ) e−i⃗k⃗rij
⟩
C4 = ω 4 = (k⃗vi ) (k⃗vj ) + i(k⃗
N
i,j

Both the self-part and the distinct part contribute to the fourth moment.
 
⟨ ⟩ 1  ∑ ∑
C4 = ω 4 = ⟨· · · · · · ⟩ + ⟨· · · · · · ⟩
N
i=j i̸=j

When i = j, the middle two terms of the fourth moment cancel out and the exponential becomes
1. Therefore, the self-part of the fourth moment is given by
1 ⟨ ⃗ 4⟩ 1 ⟨ ⃗ ˙ 2⟩
(k⃗v ) + (k⃗v ) = 3ωo4 + ωo2 Ω2o
N N
We can evaluate each of the terms of the distinct part of the fourth moment separately. The first
term is given by
1 ∑ ⟨ ⃗ 2 ⃗ 2 −i⃗k⃗rij ⟩ 1 ∑ ⟨ −ik⃗
⃗
⟩
= (k⃗vi ) (k⃗vj ) e = (k 2 vo2 )2 e rij = ωo4 g̃ρo
N N
i̸=j i̸=j

The second and third term can be combined to give

1 ∑ ⟨[ ⃗ ˙ ⃗ 2 ]
⃗ rij
⃗ vk )2 e−ik⃗
⃗ v˙ j )(k⃗
⟩
i(k⃗vi )(k⃗vj ) − i(k⃗
N
i̸=j

1 ∑⟨ [ ]
⃗
⟩
= (kvo )2 ⃗k⃗v˙ i − ⃗k⃗v˙ j e−ik⃗rij
N
̸ j
i=

1 ∑ ⟨ −ik⃗
⃗
⟩
= −2ωo4 e rij = −2ωo4 g̃ρo
N
̸ j
i=

And the fourth term gives

1 ∑ ⟨ ⃗ ˙ ⃗ ˙ −ik⃗ ⃗
⟩
(k⃗vi )(k⃗vj )e rij
N
i̸=j

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 9

k2 1 ∑ ⟨ ⃗ rij
−ik⃗
⟩
= ∇z i U ∇ zj U e
m2 N
i̸=j

Using Eq.(4.4), we can write this as

⟨( ) ⟩
k2 1 ∑ 1 ∂ ∂ k2 −i⃗k⃗
− U + e rij
m2 N β ∂zi ∂zj β2
i̸=j

= −ωo2 Ω2L + ωo4 g̃ρo

Then the distinct part of the fourth moment is given by

ωo4 g̃ρo − 2ωo4 g̃ρo + ωo4 g̃ρo − ωo2 Ω2L = −ωo2 ΩL

2

where ∫
1 ⟨ 2 −i⃗k⃗zij ⟩ ρo ⃗
Ω2L = ∂ ϕe = d⃗re−ik⃗z ∂z2 ϕg(⃗r)
m z m
Therefore, the fourth moment of the intermediate scattering function is given by
⟨ ⟩
C4 = ω 4 = 3ωo4 + ωo2 Ω2o − ωo2 Ω2L

3. Comparison to Self-intermediate Scattering Function With our results from the pre-
vious sections, we can write the short-time expansion of the intermediate scattering function as
2 [ ] 4
⃗ ) − ω 2 t + 3ω 2 + Ω2 − Ω2 ω 2 t − · · ·
⃗ t) = S(k
F (k, o o o L o
2! 4!
2
⃗
o 2! as the initial decay term and define the frequency ωL = 3ωo +Ωo −ΩL .
We can interpret S(k)−ω 2t 2 2 2 2

For comparison, the self-intermediate scattering function is given by

2 [ ] 4
⃗ t) = 1 − ω 2 t + 3ω 2 + Ω2 ω 2 t − · · ·
Fs (k, o o o o
2! 4!
How do these compare in the long wavelength limit k → 0? In the short-time limit, the scattering
functions will be largely determined by the first terms in the expansions. We see that as
⃗ = S(0) ≤ 1
lim S(k)
k→0

⃗ t) decays slower than the self-

Therefore, in this limit, the intermediate scattering function F (k,
⃗ t).
intermediate scattering function Fs (k,

Transverse and Longitudinal Current Transverse and longitudinal current were introduced
in chapter 3, where the Navier-Stokes equation was used to predict their rate of dissipation. Here,
we will apply the short-time expansion to the current correlation functions to define the transverse
and longitudinal speeds of sound and find their behavior in the free particle limit.

To review, the current is defined as

∑ ⃗
J⃗k (t) = ⃗vi (t)e−ik⃗ri
i

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 10

Longitudinal current exists when the direction of motion of the particles (the velocity) is parallel
with the direction of propagation of the waves. For waves propagating in the z-direction, the
longitudinal current is given by ∑
J⃗L (k, t) = ⃗żi (t)e−i⃗k⃗zi
i
Transverse current exists when the direction of motion of the particles is perpendicular to the
direction of propagation of the waves. For waves propagating in the z-direction, the transverse
current is given by ∑ ⃗ zi
J⃗T (k, t) = ⃗ẋi (t)e−ik⃗
i
The longitudinal current correlation function is given by
1 ⟨⃗ ⃗ ⃗ t)
⟩
CL = JL (k, t)|J⃗L (k,
N
And the transverse current correlation function is given by
1 ⟨⃗ ⃗ ⃗ ⃗
⟩
CT = JT (k, t)|JT (k, t)
N
Using Eq.(4.2), we can write the short time expansion of each of these functions as
2
⃗ t) = v 2 (1 − ω 2 t ) + · · ·
CL (k, o L
2
t 2
CT (⃗k, t) = vo2 (1 − ωT2 ) + · · ·
2
In the long wavelength limit the transverse and longitudinal frequencies ωT and ωL are related to
the transverse and longitudinal speeds of sound by
ω 2L = k 2 c2L
T T

And the transverse and longitudinal speeds of sound are given by

∫
ρo
c2L = 3vo2 + d⃗rg(⃗r)∂z2 ϕ⃗z
2m
∫
2 2 ρo
cT = vo + d⃗rg(⃗r)∂x2 ϕ⃗z
2m
Therefore, in the long wavelength limit,
ωL2 v 3ωT2

5. Free-Particle Limit In the free-particle limit, the forces between particles can be ignored.
The longitudinal and transverse current correlation functions are then given by
⟨ ⟩
⃗ t) = v 2 e−i⃗k⃗vz t = v 2 (1 − ω 2 t2 )e− 12 (ωo t)2
CL (k, z o o
⟨ ⟩
⃗
CT (⃗k, t) = vx2 e−ik⃗vz t = vo2 e− 2 (ωo t)
1 2

⃗ ω) is a Gaussian
We can see that the Fourier transform of the transverse correlation function C˜T (k,
while ˜ ⃗
√ the Fourier transform of the longitudinal correlation function CL (k, ω) has poles at ω =
± 2ωo

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 11

4.2 Projection Operator Method

In chapter 3, we explored the long time behavior of time correlation functions, and in the previous
section, we explored their short-time behavior. However, we are ultimately interested in finding
expressions for the time correlation functions that apply at all time scales.

There are many different ways to approach this problem. In principle, we could simply calculate
the position and velocity of each particle in the system at all times. Then, any other phase-space
variable A(t) could easily be determined. However, this is clearly not viable for macroscopic sys-
tems, which contain on the order of 1023 individual particles.

Another approach, and the one we will explore here, is to consider only that part of the sys-
tem that we care about and treat all the rest as a statistical bath. This can be accomplished using
projection operator techniques. To understand this method, it is useful to consider an analogy to
a three dimensional vector space. Any three dimensional vector can be projected onto a vector of
interest to find its component in that direction. Similarly, we can project the position and velocity
information for the entire system onto A(t), and treat the rest as a statistical bath.

To illustrate this idea, let A(t) be the velocity of a Brownian particle. We could always calcu-
late A(t) by determining the positions and velocities of all the particles in the system. However,
this would be very time consuming and generate much unnecessary information. Instead, we can
project the system velocity onto the velocity of the Brownian particle and treat the rest of the
system as a bath.

We have already solved this problem for one specific case: in chapter 1, we used the Langevin
equation to describe the evolution of the velocity of a particle under the influence of friction and a
random force. In this section, we will use the projection operator technique to derive the General-
ized Langevin Equation. However, first we need to define some terms.

4.2.1 Definitions
1. The Projection Operator Given the column vectors A and B, the projection of B onto A
is given by the expression
⟨B|A⟩
PA B = A (4.8)
⟨A|A⟩
By definition,
PA2 = PA (4.9)
For an equilibrium system, the operator product is

⟨B|A⟩ = T rBA+ ρeq (4.10)

or, in the phase space representation

∫
⟨B|A⟩ = dΓB(Γ)A+ (Γ)ρeq (4.11)

Similarly, we can define the orthogonal operator Q = P − 1, which projects onto a subspace that
is orthogonal to A.

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 12

2. Operator Identity If a and b are operators, the following are identities

1 1 1 1
= + b (4.12)
s−a−b s−a s−a−b s+a
∫ t
(a+b)t at
e =e + e(a+b)(t−τ ) beaτ dτ (4.13)
0

3. The Liouville Operator The time evolution of an operator A in a system with the Hamil-
tonian H is found using the Liouville operator L
dA
= iLA (4.14)
dt
The Liouville operator is a special form of operator called a “superoperator” because it acts upon
other operators rather than on functions. In quantum mechanics, the Liouville operator for a
system with the Hamiltonian H is defined as
1
iLA ≡ [A, H] (4.15)
i}
where [. . . , . . .] indicates the commutator. In the classical limit as } → 0, this becomes

iLA ≡ i {A, H} (4.16)

where {. . . , . . .} is the Poisson bracket. One important property of L is that it is Hermitian. This
property is demonstrated in the following proof
( )
⟨LA|B⟩ = T r [A, H] B + ρ (4.17)
( )
= T r AHB + ρ − HAB + ρ
( )
= T r AHB + ρ − AB + Hρ
[ ]
= T rA H, B + ρ
= T rA [H, B]+ ρ
= ⟨A|HB⟩

4.2.2 The Generalized Langevin Equation

The Liouville equation
d
A(t) = iLA(t) (4.18)
dt
has the formal solution
A(t) = eiLt A(0) (4.19)
From this equation it is clear that the function eiLt acts as a time propagator of A from an initial
value A(0). However, it is not very helpful in this form. We will use the projection operator to
rewrite this equation in a more useful form. To simplify the notation, A(t) will be written as A
from now on. Start by writing the new equation of motion for A(t)
dA
= iLeiLt A
dt

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 13

Insert the identity, I = (P + Q)

dA
= eiLt (P + Q) iLA = eiLt P iLA + eiLt QiLA
dt
Begin by evaluating the first term. Using the definition of the projection operator, we can rewrite
this as
⟨iLA|A⟩
eiLt P iLA = eiLt A (4.20)
⟨A|A⟩
⟨LA|A⟩ iLt
=i e A
⟨A|A⟩
= iΩA(t)
where Ω is called the frequency matrix and is defined as

⟨LA|A⟩
Ω=
⟨A|A⟩

To evaluate the second term, we will need to rewrite the time propagator in terms of P and Q.
Start by inserting the identity, and then rewrite the expression using the operator identity defined
in section A.2, with a = −iQL, b = −iP L, and (a + b) = −iL

eiLt = ei(P+Q)Lt (4.21)

∫ t
= e iQLt
+ eiL(t−τ ) iP LeiQLt dτ (4.22)
0

Now, apply this expansion to iQLA

∫ t
e iLt
iQLA = e iQLt
iQLA + eiL(t−τ ) iP LeiQLt (iQLA)dτ
0

To understand this expression, start by examining the first term. The operator Q projects the
system into the solvent degrees of freedom, which are orthogonal to A. However, we are primarily
interested in describing only the propagation in the A direction. Therefore, this term gives the
random force or noise in the system, which we will denote R(t)

R(t) = eiQLt iQLA

where R(0) = iQLA and eiQLt describes the time propagation of R(t). The second term in this
expression describes the friction in the system. One interesting thing to note is that the expression
for R(t) appears in this term, indicating that the friction and noise of the system are related. This
relation is called the fluctuation-dissipation theorem, and will be given more explicitly later. Using
the definition of R(t), we can rewrite the second term in the expression as
∫ t ∫ t
e iL(t−τ )
iP Le iQLt
(iQLA)dτ = eiL(t−τ ) iP LR(t)dτ
0 0

Then, use the definition of the projection operator P to write

∫ t
⟨iLR(t)|A⟩
eiL(t−τ ) Adτ
0 ⟨A|A⟩

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 14

Since the noise term R(t) is already projected into the orthogonal space, we can always operate on
it with Q without changing its value (recall that for any projection operator P, P 2 = P)
∫ t
⟨iLQR(t)|A⟩
eiL(t−τ ) Adτ
0 ⟨A|A⟩
Then, use the fact that Q and L are both Hermitian operators to rearrange the expression
∫ t
⟨R(t)|iQLA⟩
− eiL(t−τ ) Adτ
0 ⟨A|A⟩

Finally, use the definition R(0) = iQLA and eiL(t−τ ) A = A(t − τ ) to write the expression as
∫ t
⟨R(t)|R(0)⟩
− A(t − τ )dτ
0 ⟨A|A⟩

Define the memory kernel κ(t) as

⟨R(t)|R(0)⟩
κ(t) =
⟨A|A⟩
This term gives the final fluctuation-dissipation theorem. The second term can then be written as
simply ∫ t
− κ(t)A(t − τ )dτ
0
With all of this in hand, we can finally write out the full Generalized Langevin equation
∫ t
dA
= iΩA(t) − κ(t)A(t − τ )dτ + R(t) (4.23)
dt 0

where the frequency matrix is

⟨LA|A⟩
Ω= (4.24)
⟨A|A⟩
the random force is
R(t) = eiQLt iQLA (4.25)
and the memory kernel, which defines the fluctuation-dissipation theorem, is
⟨R(t)|R(0)⟩
κ(t) = (4.26)
⟨A|A⟩
Let’s take a closer look at the frequency matrix and the memory kernel. For one-dimensional prob-
lems, the frequency matrix will evaluate to zero. To understand why, remember that iLA = dA dt .
⟨ dA ⟩
This allows us to rewrite the numerator of the frequency matrix as dt |A , which is simply the
derivative of the correlation function C(t) = ⟨A(t)|A(0)⟩, evaluated at zero. Since all correlation
functions are even in time, the derivative at zero must equal zero. This rule will apply to all of the
problems that we address in this section.

As stated earlier, the definition of the memory kernel relates the fluctuation, or noise in the system,
to the dissipation of A. The fluctuation term, ⟨R(t)|R(0)⟩ ⟨A|A⟩−1 , will be zero when the noise in
the system is zero. This indicates that in an isolated system with no noise, A will quickly decay to
zero.

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 15

4.2.3 Applications of the GLE

1. GLE for Brownian Motion In chapter 1, we used the Langevin equation to explore the
motion of a Brownian particle. Here, we will perform the same analysis using the Generalized
Langevin equation. Recall that Brownian motion describes the discrete and random motion that is
observed when a large particle is immersed in a fluid of smaller particles. We want to use the GLE
to describe the velocity of the large particle without having to solve for the motion of the entire bath.

Begin by writing the GLE for the the velocity of the particle. For this system, the frequency
matrix Ω is zero, so the full GLE is given by
∫ t
dv f (t)
=− γ(t − τ )v(τ )dτ +
dt 0 m
f (t)
where γ(t) represents the memory kernel and m = R(t) represents the random force. For this
system, the memory kernel is given by

⟨f (t)|f (0)⟩
γ(t) =
m2 ⟨v|v⟩
⟨ ⟩
The normalization factor ⟨v|v⟩−1 is simply the average value of the squared velocity, v 2 = vo2 =
kB T
m . Therefore we can write this as

β
γ(t) = ⟨f (t)|f (0)⟩
m
where β = kB T −1 . The friction coefficient for the system is given by ξ(t) = mγ(t). Using this, we
can write the fluctuation-dissipation relation

ξ(t) = β ⟨f (t)|f (0)⟩

We can use the GLE to find the velocity autocorrelation function C(t) = ⟨v(t)v(0)⟩ for the Brownian
particle. Begin by multiplying the GLE through by v(0) and taking the thermal average.
∫ t
dv f (t)
=− γ(t − τ )v(τ )dτ +
dt 0 m
∫ t
dv f (t)
v(0) = − γ(t − τ )v(τ )v(0)dτ + v(0)
dt 0 m
⟨ ⟩ ∫ t
dv v(0)
v(0) = − γ(t − τ ) ⟨v(τ )v(0)⟩ dτ + ⟨f (t)⟩
dt 0 m
∫ t
dC(t)
=− γ(t − τ )C (τ )dτ
dt 0

Here we have used the fact that the thermal average over the random force ⟨f (t)⟩ = 0. This
gives us the equation of motion for C(t), which can be solved using Laplace transformation. The
Laplace transform of this equation gives

ˆ
sC(s) ˆ
+ γ̂(s)C(s) = C(0)

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 16

Using C(0) = ⟨v(0)v(0)⟩ = vo2 and rearranging, we get the general Laplace transformed solution for
C(t)
ˆ = v02
C(s) (4.27)
s + γ̂(s)
which can be solved for specified values of γ(t).

The Laplace transformed solution for C(t) can be used to find an equation for the diffusion constant
D. The Green-Kubo relation defines the diffusion constant as
∫ ∞
D= ˆ = 0)
C(t)dτ = C(s
0

Using the solution that we derived above

v02 kB T
D= = (4.28)
γ̂(0) mγ̂(0)

This is a generalized form of Einstein’s relation, which we derived in chapter 1 for the Brownian
particle.

The Brownian particle experiences white-noise, which can be modelled by making the memory
function a delta function γ(t) = γo δ(t). Then the GLE simplifies to
∫ t
dv f (t)
= − γo δ(t − τ )v(τ )dτ + (4.29)
dt 0 m
f (t)
= −γo v(t) + (4.30)
m
which has the formal solution (chapter 1)
∫ t
−γt 1
v(t) = v(0)e + e−γ(t−τ ) f (t)dτ (4.31)
m 0

and the correlation function C(t) = C(0)e−γt . Using the white-noise memory function, we can also
reproduce Einstein’s relation from chapter 1. The Laplace transform of γ(t) = γo δ(t) is γ̂(s) = γo .
Substituting this for γ̂(0) in Eq.(4.28) and using the friction coefficient ξ(t) = mγo gives the familiar
Einstein relation
kB T kB T
D= = (4.32)
mγ̂(0) ξ

2. Exponential-decay memory In addition to the delta function memory kernel, which gives
the dynamics of a Brownian particle, we can also consider a case in which the friction has the same
overall strength γo but varies with time. We can model this with the exponential decay memory
kernel
γ(t) = γo αe−αt
This memory kernel has the special property that no matter the value of α, the integral of the
function will always equal γo . In the limit as α → ∞, this function approaches γo δ(t+ ).

The correlation function for this memory kernel is relatively easy to find because the Laplace

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 17

transform of an exponential decay function is well defined. For the exponential decay γ(t) defined
above, the Laplace transform is
γo α
γ̂(t) =
s+α
Therefore, to solve the correlation function, we only need to find the value of γo . This can be
estimated using the definition of the memory kernel

⟨iQLv|iQLv⟩
γ(0) =
m2 ⟨v 2 ⟩

−1 ∂U
Here, iLv = dv F
dt is simply the acceleration. Using Newton’s law, we can write iLv = m = m ∂x ,
which is the gradient of the potential, or the non-random component of the force. Putting every-
thing together, we find that the memory kernel evaluated at zero is
⟨ 2 ⟩
∂ U
γ(0) = x ≡ Ω2o
m

This is the average curvature of the potential. For a harmonic oscillator, this is simply the average
frequency.

We can now use the Laplace transform of the exponential decay memory kernel to find the corre-
lation function.
vo2 s+α
Ĉ(0) = Ω2o
= vo2 2
s + s+α s + sα + Ω2o

This is relatively easy to solve because it is quadratic. To generate the solutions, find the eigenvalues
by solving the quadratic equation s2 + sα + Ω2o [1]. This gives the results
√
α α − 4Ω2o
λ± = − ± (4.33)
2 4
1 ( )
2
C(t) = vo e−λ+ t − e−λ− t (4.34)
λ+ + λ−

Some interesting results arise from this solution. We can see that if α < 2Ω, then λ± are complex
numbers and C(t) becomes oscillatory

1
C(t) = vo2 (cos ∆t)
λ+ + λ−

We can examine these results for different relations between α and Ωo .

1. Solids When α ≪ Ω, the decay time is much longer than one oscillation period. The
correlation function shows persistant oscillations at many frequencies. There is virtually no damping
and the decay occurs primarily through dephasing (see Figure 4.1).
Physically, this represents a solid. In a solid, each individual particle is locked into position by
strong bonding between itself and its neighbors. If it is disturbed from equilibrium, it can only
vibrate within the small area allowed by these bonds.

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 18

C(t)

0
time

Figure 4.1: Velocity correlation function of a solid

2. Liquids When αlΩ, the decay time is longer than one oscillation period. The correlation
function shows one or two oscillations which are quickly damped out and a long time decay tail
(Figure 4.2).
Physically, this represents a liquid. At short times, a molecule in a liquid is “trapped” within
a solvation shell formed by weak intermolecular bonds. When it is disturbed from equilibrium,
it will initially vibrate within this shell. However, at longer times, this vibration will cause a
rearrangement of the solvation shell, allowing the molecule to travel away from its initial position.
This damps out the oscillations.

3. Gases When α Ω, the decay time is shorter than one oscillation period. The correlation
function decays completely before undergoing an oscillation (Figure 4.3).
Physically, this represents a gas. In a gas, the molecules are not confined by intermolecular
bonding, and the correlation function will decay without any oscillation.

3. Generalized Diffusion Constant We can use the GLE to derive the Green-Kubo relation
for the generalized diffusion constant. Following a similar procedure as that used for the velocity
correlation function, we can show that the equation of motion for the intermediate scattering
function (which we have discussed in depth in Chapter 3 and in section IC of chapter 4) is given

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 19

C(t)

0
time

Figure 4.2: Velocity correlation function of a liquid

by ∫ t
⃗ t) = −
F˙ (k, ⃗ t)F (k,
κ(k, ⃗ t − τ )dτ
0
where the memory kernel has its usual definition

⟨R(t)|R(0)⟩
κ(t) =
⟨ρk |ρk ⟩

Using the definition of the noise term for this system

R(t) = eiQLt iQLρk

We can rewrite the memory kernel

⟨ iQLt ⟩
e iQLρk |iQLρk
κ(t) =
⟨ρk |ρk ⟩
⟨ ⟩
= ρ̇k |eiQLt |ρ̇k
By evaluating ρ̇k (section IC) we can write the memory kernel as
⟨ ⟩
⃗ ⃗
κ(t) = k 2 ⃗v e−ik⃗r |eiQLt |⃗v e−ik⃗r

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 20

C(t)

0
0
time

Figure 4.3: Velocity correlation function of a gas

However, this term is simply the diffusion constant D(k,⃗ t)multiplied by k 2 . Therefore, the equation
of motion for the intermediate scattering function can be written as
∫ t
⃗
F (k, t) = −k
˙ 2 ⃗ τ )F (k,
D(k, ⃗ t − τ )dτ
0

In the long time limit t → ∞ and the long wavelength limit k → 0, we find the Green-Kubo relation
∫ ∞ ∫ ∞
D= D(0, t)dt = C(t)dt
0 0

4.3 Viscoelastic Model

4.3.1 Introduction
The Generalized Langevin Equation and Mode-Coupling theory are subsets of molecular hydrody-
namics, the theory that was developed to bridge the gap between hydrodynamics and molecular
dynamics. Hydrodynamics, which we discussed in chapter 3, describes the macroscopic, long time
behavior of systems in the limit as t → ∞ and k → 0. It uses the transport coefficients D, λ,
η, and ηB to predict long time fluctuations. Molecular dynamics, which we discussed in section I
of chapter 4, describes the microscopic, short time behavior of systems in the limit as t → 0 and
k → ∞. In this limit, systems behave as static liquid structures, and their dynamics are largely

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 21

determined by the pairwise interaction potential.

In this section, we will use the GLE to derive the viscoelastic model for transverse current. By
taking the appropriate limits, we can show that the results of the viscoelastic model are consistent
with those of hydrodynamics and molecular dynamics, and that this model provides a successful
bridge between the two limits.

4.3.2 Phenomenological Viscosity

Consider a constant shear force applied to a viscous liquid. At long times, the shear stress σxz in
the liquid is related to the rate of strain ∂z ⃗vx + ∂x⃗vz by
σxz = −η (∂z ⃗vx + ∂x⃗vz )
Liquids behaving in this fashion do not support shear waves. However, if the force is applied
instantaneously, the system does not have the time to relax like a liquid. Instead, it behaves like
an elastic solid. The stress is now proportional to the strain rather than the rate of strain. The
short term response is
σxz = −G (∂z ⃗x + ∂x⃗z)
where G is the modulus of rigidit√
y. When the liquid is behaving like a solid, it supports shear waves
G
propagating at a speed of vs = ρm .

To determine the time scale on which the liquid behaves like an elastic solid, define the constant
η
τM =
G
This is the Maxwell relaxation time. For the timescales t when
t
≪1
τM
the system behaves like an elastic solid. For the timescales when
t
≫1
τM
the system behaves like a viscous liquid.

4.3.3 Viscoelastic Approximation

1. To interpolate between the two extremes, we can write
( ) ( )
1 1 ∂ ∂ ∂
+ σxz = − ⃗vx + ⃗vz
η G ∂t ∂z ∂x
The Laplace transform of this equation yields
G G
η̂(s) = G
= 1
s+ η s + τM
In the steady-state limit, as s → 0
lim η̂ = η
s→0
and in the high-frequency limit, as s → ∞
G
lim η̂ =
s→∞ s

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 22

2. Transverse Current Correlation Function We will use the transverse current correlation
function to demonstrate the viscoelastic approximation. In Section I, we defined the transverse
current as
∑
N
⃗ zj
⃗
JT (k, t) = ⃗ẋj eik⃗
j=1

and the transverse current correlation function as

⟨ ⟩
⃗ t) = 1 JT (k,
CT (k, ⃗ t)|JT (k,
⃗ 0)
N
We have studied the transverse current in both the hydrodynamic limit (k → 0, ω → 0) and the
short-time expansion limit (k → ∞, ω → ∞). In chapter 3, we used the Navier Stokes equation to
find an equation of motion for the transverse correlation function in the hydrodynamic limit
⃗ t) = −νk 2 CT (k,
C˙ T (k, ⃗ t) (4.35)

This has the solution

⃗ 0)e−νk
⃗ t) = CT (k,
CT (k,
2t
= vo2 e−νk
2t

η
where ν = mρ is the shear viscosity. Therefore, in the hydrodynamic limit, transverse current
fluctuations decay exponentially with a rate determined by the shear viscosity ν.

In section I of this chapter, we used the short-time expansion approximation to show that in
the k → ∞, ω → ∞ limit, the transverse current correlation function can be written as
2
⃗ t) = v 2 (1 − ω 2 t ) + · · ·
CT (k, o T
2
where the transverse frequency ωT is related to the transverse speed of sound cT (k) by

ωT2 = k 2 c2T (k)

And the transverse speed of sound is given by

∫
2 2 ρ [1 − eikz ]
cT (k) = vo + g(⃗r)∂x2 U (⃗r) d⃗r
m k2
where g(⃗r) is the pairwise correlation function and U (⃗r) is the pairwise interaction potential. This
frequency term can also be written as [3]
(kvo )2
ωT2 = G∞ (k)
nM
where G∞ (k) is the shear modulus. This indicates that at short times and wavelengths, the dissipa-
tion effects are diminished and transverse current fluctuations can propagate through the material
with speed cT (k).

Using the Generalized Langevin equation, we can generate a model for transverse current fluc-
tuations that replicates the results of hydrodynamics and the short-time expansion when the ap-
propriate limits are taken. Begin by writing the GLE for transverse current. Since the frequency
matrix is zero, the GLE is written
∫ t
JT (k, t) = −
˙ κ(t − τ )JT (k, τ )dτ + f (t)
0

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 23

where κ(t) is the memory function and f (t) is the noise term. Multiplying through by JT (k, 0) and
taking the average gives us the equation of motion for the transverse current correlation function
∫ t
ĊT (k, t) = − κ(t − τ )CT (k, τ )dτ
0
Take a closer look at the memory kernel
⟨ iQLt ⟩
⟨R(t)|R(0)⟩ e iQLJT (k, 0)|iQLJT (k, 0)
κ(t) = =
⟨A|A⟩ ⟨JT (k, 0)|JT (k, 0)⟩
βm
The normalization factor is simply ρ . By writing the projection operator Q as 1 − P and elimi-
nating terms, this can be written
βm ⟨ i(1−P)Lt ⟩
κ(t) = e iLJT (k, 0)|iLJT (k, 0)
ρ
βm ⟨ ˙ ⟩
= JT (k, t)|ei(1−P)Lt |J˙T (k, t)
ρ
The equation of motion for the transverse current can be written as [5]

⃗ t) = i k σ zx (k,
J˙T (k, ⃗ t)
m
⃗ t) is the zx-component of the microscopic stress tensor
where σ zx (k,
 
∑ 1 ∑ zij xij 
⃗) =
σ zx (k mvi,z vi,x − P k (rij ) eikzi
 2 r2 
i j̸=i

Then the memory kernel becomes

β ⟨ zx ⃗ ⟩
⃗ t) = k 2 ν(k, t)
κ(t) = k 2 σ (k, t)|ei(1−P)Lt |σ zx (k,
ρm
where ν(k, t) is defined as
β ⟨ zx ⃗ ⃗ t)
⟩
ν(k, t) = σ (k, t)ei(1−P)Lt σ zx (k,
ρm
This demonstrates that the memory kernel is proportional to k 2 . Then the transverse current
correlation function can be written
∫ t
ĊT (k, t) = −k 2
ν(k, t − τ )CT (k, τ )dτ
0
The memory kernel is the key element that links the two limits. In general, the presence of the
propagator eiQLt makes it very difficult to evaluate ν(k, t) explicitly. However, the presence of Q
indicates that we can separate out fast and slow motions and use this to construct a form for ν(k, t)
that will bridge the short and long time limits. To find this form, the viscoelastic model starts my
assuming that the memory kernel has an exponential form:
[ ]
−t
ν(k, t) = ν(k, 0) exp (4.36)
τM (k)
where τM (k) is the Maxwell relaxation time, discussed above. Before using this function, it is
necessary to specify the values of the two parameters, ν(k, 0) and τM (k). These can be found by
taking the short and long time limits of the GLE and comparing them to the short-time expansion
and hydrodynamic results, respectively.

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 24

3. The Short Time Limit The value of ν(k, t) at short times can be obtained by comparing
the GLE at time t = 0 to the short time expansion of the transverse correlation function. To find
the GLE at time t = 0, take its time derivative
∫ t
2 d
C̈T (k, t) = −k ν(k, t − τ )CT (k , τ )dτ
dt 0
C̈T (k, t) = −k 2 ν(k, 0)CT (k, t)
C̈T (k, 0) = −k 2 ν(k, 0)CT (k, 0)
C̈T (k, 0) = −k 2 ν(k, 0)vo2 (4.37)
The first two terms of the short time expansion of the correlation function are
2
⃗ t) = v 2 (1 − ω 2 t ) + · · ·
CT (k, o T
2
The second derivative of this expansion gives
C̈T (k, 0) = −k 2 c2t (k)vo2 (4.38)
Comparison of equations (4.38) and (4.35) shows that
G(k )
ν(k, 0) = c2t (k) =
ρm
Further, we see that in this limit the material supports propagating waves. The form of the waves
can be found by solving the differential equation Eq.(4.38), and is given by
CT (k, t) = vo2 cos(ωt t − kx)
√
where ω = kct (k) and the speed of the waves are ct = G(k) ρm .

4. The Hydrodynamic Limit The value of ν(k, t) at long times can be obtained by comparing
the hydrodynamic equation to the long time limit of the GLE for CT (k, t):
∫ t
ĊT (k, t) = −k 2
ν(k, t − τ )CT (k, τ )dτ
0
To take the long time limit of this equation, note that the memory function will generally be
characterized by some relaxation time τκ . When the time t is much greater than this relaxation
time, the major contribution to the integral will come when t ∼ τ . Therefore, we can approximate
CT (k, τ ) ∼ CT (k, t). With this approximation, the correlation function can be taken out of the
integral in the GLE: ∫ ∞
ĊT (k, t) = −k 2 CT (k, t) ν(k, t − τ )dτ
0
where the integration limit has been extended to ∞ to indicate that we are taking the long time
limit.

This result should be identical to the hydrodynamic solution in the long time and long wave-
length limit. By taking the long wavelength limit (k → 0) and comparing to the hydrodynamic
result (Eq.(4.35)), we see that this only holds when:
∫ ∞
η
ν(k, t − τ )dτ = ν =
0 ρm

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 25

5. The Viscoelastic Solution We now have the information we need to construct the explicit
form of the viscoelastic memory kernel.
[ ]
−t
ν(k, t) = ν(k, 0)
τM (k)
G(k)
From the short time limit, we found that ν(k, 0) = ρm , which allows us to write
[ ]
G(k) −t
ν(k, t) = exp
ρm τM (k)

From the long time limit, we know that

∫ ∞
η
ν(k, t − τ )dτ =
0 ρm

Now, plug in the exponential memory kernel for k = 0

∫ ∞ [ ]
G(0) −t η
exp dτ =
0 ρm τM (0) ρm

The elastic modulus has no time dependence, so it can be taken out of the integral
∫ [ ]
G(0) ∞ −t η
exp dτ =
ρm 0 τM (0) ρm

Finally, evaluate the integral to find the Maxwell relaxation time at k = 0. It is reasonable
to assume that the Maxwell relaxation time remains constant over all k values. Therefore, the
Maxwell relaxation time can be written as the ratio of the shear viscosity coefficient of the liquid
to the modulus of rigidity of the elastic solid at k = 0.
η
τM =
G(0)

When τM is small compared to the time t, the viscosity term dominates and the system will behave
as a viscous liquid. However, when τM is large compared to the time t, the system does not have
time to respond to a stimulus as a viscous liquid. The modulus of rigidity dominates, and the
material will behave as an elastic solid, supporting propagating shear waves.

Finally, we can use the Maxwell relaxation time to write the explicit form of the viscoelastic
memory kernel. [ ]
G(k) −ηt
ν(k, t) = exp
ρm G(0)
With this memory kernel in hand, we can now go on to find an explicit solution to the transverse
current correlation function.

To find the equation for the viscoelastic wave, we first find the Laplace transform of the trans-
verse current correlation function
C(k, t = 0) vo2
ĈT (k, s) = =
s + k 2 ν̂(k, s) s + k 2 ν̂(k, s)

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 26

Now, solve this equation using the exponential memory kernel ν(k, t) = G(k)
ρm exp τ−t
M
. The Laplace
transform of an exponential function is well defined
1
L[exp [−αtu(t)]] =
s+α
Therefore, the Laplace transform of the viscoelastic memory kernel is
ν(k, 0)
ν̂(k, s) = −1
s + τM
Plug this into the the Laplace transform of the transverse current correlation function

vo2
ĈT (k, s) = ( )
ν(k,0)
s + k2 s+τM−1

−1
vo2 (s + τM )
CˆT (k, s) = −1
s(s + τM ) + k 2 ν(k, 0)
Since function is quadratic, it is relatively easy to find the reverse Laplace transform, using the
same method as that presented in section 4.2.C.2, or reference [1].
1 ( )
CT (k, t) = vo2 e−λ+ t − e−λ− t
λ+ + λ−
−1
where the eigenvalues λ± are given by the solutions to the quadratic equation s2 + τM s + k 2 ν(k, 0):
√
−1 −2
−τM τM − 4k 2 ν(k, 0)
λ± = − ±
2 4
Complex eigensolutions exist if
1 2
2 < 4k ν(k, 0)
τM
Recall that ν(k, 0) = c2t (k). Then we can rewrite the above inequality in terms of the wavenumber
1
k>
2τM ct (k)

Define the critical wavenumber, kc = (2τM ct (k))−1 . For more information on the viscoelastic
approximation and its application to transverse current, please see Chapter 6 of Molecular Hydro-
dynamics by Jean-Pierre Boon and Sidney Yip [3] and chapter 3 and chapter 6 of Dynamics of the
Liquid State by Umberto Balucani [5].

4.4 Long-time Tails and Mode-coupling Theory

4.4.1 The Problem
f
dt = −γv + m , we found that the velocity correlation function should
From the Langevin equation dv
take the form of an exponential decay (see Figure 4.4)

C(t) = ⟨v(t)v(0)⟩ = vo2 e−γt

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 27

Figure 4.4: Uncorrelated collision of two particles in a fluid

where the memory kernel is related to the diffusion constant through Einstein’s relation
1
γ=
mDβ
In 1967, Alder and Wainwright used computer simulations to calculate the velocity correlation
function of hard-sphere gases [6]. They found that at long times, C(t) exhibits a power-law decay
rather than an exponential decay. That is, C(t) decays according to t− 2
3

1
lim C(t) ∝ 3
t→∞ t2
This is the famous long-time tail problem in kinetic theory.

4.4.2 Memory Effects

The fundamental assumption underlying the exponential decay model of the velocity correlation
function is that collisions between particles in a dilute hard-sphere gas are independent. This means
that after each collision, a particle will lose memory of its original velocity until its motion has be-
come completely randomized. This assumption leads to the exponential decay C(t) = vo2 e−t/τcoll ,
where τcoll = γ1 is the average collision time.

However, it is also possible that collisions are not completely independent but instead are cor-
related. A correlation would occur if, for example, two particles collide and then collide again
after undergoing some number of other collisions (see Figures 4.5 and 4.6). This implies that there
is a long term memory in the system which would lead to a decay that is slower than an exponential.

To estimate the form of this decay, we can consider the probability that, following a collision,
a particle remains at or returns to its initial position after a time t, P (r = 0, t). To make a
rough estimate of this probability, imagine that at any moment in time we can draw a “probability
sphere” for the particle. The probability of finding the particle inside the sphere is constant, and
the probability of finding the particle outside the sphere is zero. At time t = 0, the particle has not
had time to travel away from its initial position. Therefore, P (r = 0, t) = δ(r). As t increases, the
particle begins to diffuse away from its initial position. The radius of the sphere increases linearly

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 28

Figure 4.5: A test particle (blue star) collides with one particle (red circle), altering that particle’s
trajectory such that it collides with a third particle (yellow square). The subsequent collision of
the third particle with the test particle is correlated with the initial collision.

Figure 4.6: Vortex rings in the fluid around a hard sphere particle can contribute to the long-time
tail of the velocity autocorrelation function [6].

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 29

with time according to R = 2D2 t, where D is the diffusion constant.

To estimate P (r = 0, t) from probability sphere, note that the probability of finding the parti-
cle at any point in space must be unity:
∫
P (r, t)dr = 1

We only need to integrate over the volume of the sphere, since the probability of finding the particle
anywhere else is zero. Within the sphere, the probability is a constant, P (r, t) = P (t). Therefore,
this integral simplifies to ∫
1= P (t)dr ∝ V P (t)
sphere
d
The volume of the sphere goes as Rd = t 2 , where d is the spatial dimension. Therefore, the
probability goes as
1 1 1
P (r = 0, t) = ∝ d ∝ d
V R t2
For a three dimensional system, we get a result consistent with Alder and Wainwright’s predictions
1
P (r = 0, t) ∝ 3
t2
This shows that memory effects may be the source of the power law decay.

We can construct a simple model of the behavior of a system for which memory effects are impor-
tant. In this model, a particle with a velocity vo creates a velocity field through its interactions
with other particles. This velocity field can in turn influence the long time behavior of the particle.

We can start by getting a rough estimate of this velocity field from the transverse current
JT (k, t) = JT (k, 0)e−νk
2t

η
where ν = mρ is the diffusion constant, η is the shear viscosity, and JT (k, 0) ∝ vo . The transverse
current is written in k -space. By transforming this into real space, we obtain an expression for the
dissipation of the velocity field over time and space
vo −r 2
v(r, t) ∝ (√ )3 e 2νt
4πνt
The velocity field dissipates due to friction. At short times, the decay has a Gaussian form.
However, at long times the decay is dominated by the prefactor, which goes as τ − 2 .
3

4.4.3 Hydrodynamic Model

A simple way of deriving the above result would be to evaluate the velocity correlation function
C(t) = ⟨v(t)v(0)⟩
Using the hydrodynamic model, we can find this correlation function by taking the equilibrium
average of the non-equilibrium average thermal velocity
∫ ∫
1
C(t) = ⟨v(t)v(0)⟩ = d⃗ro d⃗vo fB vo ⟨v(t)⟩n.e.
V

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 30

where fB is the Boltzmann distribution and ⟨v(t)⟩n.e. is a non-equilibrium velocity field:

⟨vs (t)δ (vs − vo ) δ (rs − ro )⟩

⟨v(t)⟩n.e. =
⟨δ (vs − vo ) δ (vs − vo ) δ (rs − ro )⟩

Here, s describes the tagged particle.

The non-equilibrium velocity field can be represented as a coupling of the linear modes ρs and
vs .
∫
⟨v(t)⟩n.e. = ρs (r, t) vs (r, t) dr

∫
1
= ρs (r, t) J (r, t) dr
ρo
∫
ρ−1
o
= ρs (k, t) J (k, t) dk
(2π)3
We can solve this using the solutions of the hydrodynamic modes:

ρs (k, t) = ρs (k, 0) e−k

2 Dt

J⊥ (k, t) = J⊥ (k, 0) e−k

2 νt

JL (k, t) = JL (k, 0) e−k

2 Γ t±ic kt
2 s

For transverse modes,

∫
1 2
ρs (k, 0) e−k Jx (k, 0) e−k
2 Dt 2 νt
vo ⟨v(t)⟩n.e. = 3 vxs (0)dk
(2π) ρo 3
∫
1 2
ρs (k, 0) Jx (k, 0) vxs (0)e−k
2 (D+ν)t
= 3 dk
(2π) ρo 3
Then, take the equilibrium average

⟨ρs (k, 0) Jx (k, 0) vxs (0)⟩ = vo2

Finally,
∫
2 2 1
d⃗ke−k (D+ν)t
2
C(t) = v
3ρo o (2π)3

vo2 1
=
12ρo [πt (D + ν)] 32

1
∝ 3
t2

5.72, Spring 2008 J. Cao

Chapter 4. Calculations of Time Correlation Functions 31

4.4.4 Mode-Coupling Theory

As shown above, the correlation of a given dynamic quantity decays predominantly into pairs of
hydrodynamic modes with conserved variables. Mode-Coupling Theory is the formalism that cal-
culates their coupling.

From the discussion about, the velocity of the tagged particle is coupled to a bilinear mode,
A = ρs (k)Jx∗ (k). then ⟨ ⟩ ⟨ ⟩
C(t) = vx |eiLt |vx ⇒ vx |P eiLt P |vx
where P is the projection operator associated with A. By expanding the projection operator
⟨ ⟩
C(t) = ⟨vs |A⟩ ⟨A|A⟩−1 A|eiLt |A ⟨A|A⟩−1 ⟨A|vs ⟩
Now, use the linear hydrodynamic modes to evaluate the correlation function.
ρs = e−ikrs
∑
Jx = vxi e−ikri
i
Then ∑⟨ ⟩
⟨vs |A⟩ = vxs |e−ikrs vxi eikri = vo2
i
and
⟨A|A⟩ = N vo2
so that
1
⟨vs |A⟩ ⟨A|A⟩−1 =
N
Therefore,
1 ∑∑⟨ ⟩
C(t) = 2
A(k)|eiLt |A(k ′ )
N ′ k k
1 ∑∑⟨ ( )⟩ ⟨ ( )⟩
≈ 2 ρs (k, t) |ρs k ′ , 0 Jx (k, t) |Jx k ′ , 0
N
k k′
1 ∑
= Fs (k, t) Ct (k, t)
N
k
∫ ( ) ( )
V 1 ⃗ s ⃗k, t Ct ⃗k, t
= dkF
N (2π)3
Now, ( )
Fs ⃗k, t = e−k Dt
2

and ( )
Ct ⃗k, t = vo2 e−k νt
2

Then ∫
1 vo2 ⃗ −k2 (D+ν)t
C(t) = dke
ρ (2π)3
By incorporating the three-spatial components of J and v, we have
[ ]3
1 2 1 2
C(t) = vo
12ρ π (D + ν) t

5.72, Spring 2008 J. Cao

References

[1] Jean Pierre Hansen and Ian R. McDonald. Theory of Simple Liquids. Burlington, MA: Elsevier
Academic Press, 2006.

[2] Donald A. McQuarrie. Statistical Mechanics. Sausalito: Univerity Science Books, 2000.

[3] Jean-Pierre Boon and Sidney Yip. Molecular Hydrodynamics. New York: McGraw-Hill, 1980.

[4] Bruce J. Berne and Robert Pecora. Dynamic Light Scattering: with Applications to chemistry,
biology, and physics. New York: Wiley, 1976.

[5] Umberto Balucani. Dynamics of the Liquid State. New York: Oxford University Press, 1994.

[6] B. J. Alder and T. E. Wainwright. Phys. Rev., A1, 1970.

32
MIT OpenCourseWare
http://ocw.mit.edu

5.72 Statistical Mechanics

Spring 2012

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.