You are on page 1of 7

DEPARTMENT OF PHYSICS

INDIAN INSTITUTE OF TECHNOLOGY MADRAS

EP3120 Statistical Physics Problem Set 8 1.11.2023

The Langevin equation: Recall that we started with the simple Langevin equation
(LE) for a component of the velocity of a tagged colloidal particle (we will refer to it
as a Brownian particle) of mass m in a fluid,

Γ
v̇ + γv = ζ(t),
m
where γ is the friction constant, and ζ(t) is a stationary, Gaussian, Markov process
with zero mean and a δ-function autocorrelation (abbreviated as ‘Gaussian white
noise’ or GWN)
hζ(t)i = 0 and hζ(t) ζ(t0 )i = δ(t − t0 ).
The preservation of thermal equilibrium imposes a consistency condition relating the
strength of the fluctuating random force, Γ, to the strength of the dissipation in the
medium, γ, via the fluctuation-dissipation theorem:

Γ = 2mγkB T.

It follows that the driven random process v(t) is also a stationary, Gaussian, Markov
process with zero mean and an exponentially decaying autocorrelation function,
kB T −γ|t−t0 |
hv(t)i = 0, hv(t)v(t0 )i = e .
m
v(t) is called the Ornstein-Uhlenbeck (OU) process. Its conditional probabiliity den-
sity p(v, t|v0 ) satisfies the Fokker-Planck equation (FPE)

∂p ∂(vp) γkB T ∂ 2 p
=γ + .
∂t ∂v m ∂v 2
The initial condition
R∞ on p is of course p(v, 0|v0 ) = δ(v − v0 ), and the normalization
condition is −∞ p(v, t|v0 ) dv = 1 for all t > 0.

1. Show that, as t → ∞, p(v.t|v0 ) approaches a limiting distribution peq (v) that is


just the Maxwellian distrbution,
n m o1/2 n mv 2 o
peq (v) = exp − .
2πkB T 2kB T

2. Multiply both sides of the FPE by v and v 2 in turn, and integrate over v to
obtain ordinary differential equations for the conditional mean velocity v(t) and the
conditional mean square velocity v 2 (t) . Solve these with the appropriate initial and
boundary conditions to to obtain for t > 0 the expressions
kB T
v(t) = v0 (t)eγt and Var v(t) = 1 − e−2γt

m
1
(‘Condtional’ stands for averages taken for a given value of v0 .) Observe that the
conditional variance does not depend on the initial velocity v0 , altough the conditional
mean does so, oif course. Since it has already been stated (without proof) that the
driven process v(t) is also a Gaussian process, and since the latter is specified by its
mean and variance, we can write down the solution of the FPE in this instance as
n m o1/2 n m(v − v0 e−γt )2 o
p(v.t|v0 ) = exp − .
2πkB T (1 − e−2γt ) 2kB T (1 − e−2γt )

This is called the Ornstein-Uhlenbeck (OU) distribution. It is trivially checked that


lim p(v.t|v0 ) = peq (v).
t→∞

LE and FPE for a general diffusion process: The Langevin equation introduced
above is a special case of a much more general stochastic differential equaiton (SDE)
for what is known as a diffusion process. Thie latter is a random process ξ(t) driven
by a Gaussian white noise ζ(t) according to the SDE
˙ = f (ξ, t) + g(ξ, t)ζ(t),
ξ(t)

where f and g are specified functions of their arguments. One is then guaranteed that
ξ(t) is a Markov process that is, in general, neither stationary nor Gaussian, and is not
necessarily exponentially autocorrelated. (A necessary but not sufficient condition for
ξ to be a stationary process is that f and g have no explicit time dependence.) The two
terms on the right-hand side of the SDE above are called the frist and diffusionterms,
respectively. The crucial result is the one-to-0ne correspondence between this SDE
and the FPE for the conditional probability density p(ξ, t|ξ0 , t0 ), which reads

∂p ∂ 1 ∂2 2
= − (f p) + (g p).
∂t ∂ξ 2 ∂ξ 2
Thus, the LE and FPE for the √ velocity
pv(t) and its conditional density correspond
to ξ = v, f = −γv, and g = Γ/m = 2γkB T /m. When g is a constant, the noise
in the SDE is said to be additive; when it is a function of the random variable ξ, the
noise is said to be multiplicative.

Generalization to a multi-dimensional diffusion process: Both the SDE and


the FPE just discussed can be generalized to an n-dimensional process ξ(t) with com-
ponents ξj (t) (j = 1, 2, . . . , n), driven by a ν-component GWN ζ(t) with components
ζα (t) (α = 1, 2, . . . , ν) where 1 6 ν 6 n. The SDE is the matrix equation

ξ̇(t) = f (ξ, t) + g(ξ, t)ζ(t),

where ξ and f are (n × 1) column vectors, ζ is a (ν × 1) column vector, and g is a


(n × ν) natrix. The FPE for the conditional probability density of ξ corresponding
to this SDE is
∂p ∂  ∂2 
=− fj p + Djk p
∂t ∂ξj ∂ξj ∂ξk
where repeated indices are to be summed over, and D is the (n × n) matrix

D = 12 ggT .


2
3. Apply the foregoing to the case of (a single Cartesian component of) the position
x and velocity v of a Brownian particle, satisfying the coupled SDEs

Γ
ẋ = v, v̇ = −γv + ζ(t).
m
It is evident that n = 2 and ν = 1 in this case (the GWN has just one component,
and is an additive noise). Thus
     
x v √ 0
ξ= , f= , g=
v −γv Γ/m
in this simple instance. Use these identifications to obtain the following FPE for the
conditional probability density of the particle in phase space, p(x, v, t|x0 , v0 , 0):
∂p ∂p ∂ γkB T ∂ 2 p
= −v + γ (vp) + .
∂t ∂x ∂v m ∂v 2


Note that the pair x(t), v(t) forms a 2-dimensional Markov process, and that
v(t) by itself is a 1-dimensional Markov process. Its integral, the displacement x(t),
is not a Markov process in general. In the so-called high-friction approximation (or
diffusion regime), however, x(t) is a Markov process, as we shall see in the sequel.
Its conditional PDF then satisfies an FPE which is just the familiar diffusion equation.

4. Inclusion of an external force: Suppose the Brownian particle is also under the
influence of an external force F (x) = −U 0 (x) where the prime denotes the derivative
with respect to x. The SDEs for x and v are now given by

U 0 (x) Γ
ẋ = v, v̇ = −γv − + ζ(t).
m m
Hence obtain the FPE for the phase space PDF p(x, v, t|x0 , v0 , 0) of the Brownian
particle, namely,
∂p ∂p 0 ∂p ∂ γkB T ∂ 2 p
= −v + U (x) + γ (vp) + .
∂t ∂x ∂v ∂v m ∂v 2

5. Generalization to 3-dimensional motion: Next, consider the motion of a


Brownian particle in 3-dimensional space under the influence of a conservative force
F(r) = −∇r U (r). Write down the coupled SDEs for its position r and velocity v, and
hence obtain the FPE for its conditional PDF in phase space, PDF p(r, v, t|r0 , v0 , 0):
∂p γkB T 2
= −v · ∇r p + ∇r U (r) · ∇v p + γ∇v · (vp) + ∇v p.
∂t m
Here ∇r and ∇v denote, respectively, the gradient with respect to the components of
r and v.

3
Diffusion in a magnetic field: An interesting situation arises when a charged
Brownian particle is subjected to an applied magnetic field. For simplicity, let us
consider a constant, uniform magnetic field B = Bn directed along some unit vec-
tor n. The Lorentz force on the particle is q(v × B), which is a velocity-dependent
force. In the absence of any medium, any initial velocity vector v0 of the particle will
precess around the direction n of the magnetic field with a frequency ωc = qB/m
(the cyclotron frequency), the longitudinal component of v0 remaining unchanged.
In the presence of the medium, the initial velocity v0 may be expected to precess and
simultaneously get damped to zero magnitude on the average. Further, the trans-
verse components of the velocity will become correlated with each other owing to the
precession of the velocity vector, while the diffusion constant will be diminished in
the directions transverse to the magnetic field. Since the magnetic field does no work
on the charged particle, the equilibrium Maxwellian distribution of velocities must
remain unaltered, and the fluctuation-dissipation relation Γ = 2mγkB T continues to
hold good. All these assertions can be proved in a straightforward manner.

6. The Langevin equation for (each Cartesian component of) the velocity is now
given by √
Γ
v̇j = −γvj + ωc jkl vk nl + ζj (t) (j = 1, 2, 3).
m
It is very convenient to define a matrix R with elements Rjk ≡ jkl nl , so that the
LE becomes √
Γ
v̇j = −γvj + ωc Rjk vk + ζj (t) (j = 1, 2, 3).
m
Note that the white noise ζ(t) is also a 3-component vector, so that n = ν = 3 in this
case. Moreover, g is just a multiple of the unit matrix. Hence obtain the following
FPE for the conditional PDF of the velocity, p(v, t|v0 ):

∂p ∂ h  i γk T
B ∂ 2p
= γvj − ωc Rjk vk p + δjk .
∂t ∂vj m ∂vj ∂vk

The solution of this FPE turns out to be a modified Ornstein-Uhlembeck distribution


that incorporates the precession of the velocity vector. In order to understand the
nature of the solution, we must digress a bit to see how the precession comes about.

7. Precession due to the magnetic field: In the absence of the medium (i.e., of
both the noise and the frictional drag), the equation of motion iss given by

u̇j = ωc Rjk uk ,

where the velocity has been denoted by u in order to make it clear that this is the
velocity in the presence of the magnetic field alone. In matrix form, we have
du
= ωc R u,
dt
writing u as a column vector. Hence the formal solution is

u(t, v0 ) = eωc Rt v0

4
for any given initial velocity v0 . To find the required matrix exponential, show that
the matrix R2 has elements

(R2 )jk = nj nk − δjk ,

and that the matrix R3 has elements

(R3 )jk = −jkl nl = −Rjk .

Hence R3 = −R, which enables us to express eωc Rt explicitly as a linear combination


of the unit matrix I, R and R2 . Work out the algebra to show that

u(t, v0 ) = eωc Rt v0 = I + R sin ωc t + R2 (1 − cos ωc t) v0 .


 

Returning to the FPE for the velocity v, the solution for the conditional normalized
PDF p(v, t|v0 ) is again a Gaussian, but with the mean velocity given by

v(t) = u(t, v0 ) e−γt

and with the same variance as in the absence of B namely, (kB T /m)(1 − e−2γt ). This
is the modified Ornstein-Uhlenbeck distribution in the presence of a magnetic field.

8. The high-friction approximation√or diffusion regime: Returning to the


original Langevin equaiton mv̇ + mγv = Γζ(t), it has already been stated that the
displacement x(t) is not a Markov process in genral, but it becomes such a process
in the so-called high-friction approximation. The latter amounts to neglecting the
inertia term mv̇ in the equaiton of motion as compared to the dissipative or friction
term mγv. The LE immediately becomes

Γ
ẋ = ζ(t),

which is a particularly simple example of the SDE for a general diffusion
√ process,
with no drift term (f (x) = 0) and a constant diffusion term (g(x) = Γ/mγ). Hence
the FPE for the conditional PDF p(x, t|0, 0) is given by

∂p Γ ∂ 2p ∂ 2p
= = D ,
∂t 2m2 γ 2 ∂x2 ∂x2
which is just the basic diffusion equation in 1 spatial dimension, with the identifica-
tion of the diffusion constant as D = kB T /(mγ).

Use a Fourier transform with respect to x and a Laplace transform with respect to t
to show that the normalized fundamental solution to this diffusion equation, with the
initial condition p(x, 0|0, 0) = δ(x) and natural boundary conditions p(x, t|0, 0) → 0
as x → ±∞, is given by
1 2
p(x, 0|0, 0) = 1/2
e−x /(4Dt) .
(4πDt)

5
9. Hence show that the variance of x is equal ’to 2Dt, characteristic of diffusion.

10. Consider the diffusion equation for p(r, t|0, 0) in 3 dimensions,

∂p
= D∇2 p,
∂t
with the initial condition p(r, t|0, 0) = δ (3) (r) and natural boundary conditions. Write
own the normalized fundamental solution to the equation, and find the mean squared
displacement hr2 (t)i.

11. Diffusion in a potential: the Smoluchowski equation We have seen that


the Langevin equation for the velocity of a Brownian particle moving in 1 spatial
dimension in the presence of a potential U (x) is

U 0 (x) Γ
v̇ = −γv − + ζ(t).
m m
In the high-friction approximation, this becmes

U 0 (x) Γ
ẋ = − + ζ(t),
mγ mγ
which is an SDE for x that has the form of a diffusion process. Using the SDE↔FPE
correspondence, show that the associated FPE for the conditional PDF of x is given
by
∂p 1 ∂ 0 ∂ 2p
= [U (x) p] + D 2
∂t mγ ∂x ∂x
where D = kB T /(mγ) as usual. This is called the Smoluchowski equation for
difffusion in a potential.

12. Now consider a harmonically bound particle, for which U (x) = 21 mω 2 x2 . If the
oscillator is overdamped (high friction regime), the conditional PDF p(x, t|0, 0) f the
position x satisfies the FPE

∂p ω2 ∂ kB T ∂ 2 p
= (xp) + .
∂t γ ∂x mγ ∂x2
But this is precisely of the same form as the FPE for the conditional PDF of the
velocity of a freely diffusing Brownian particle! Hence write down the soluiion to the
FPE above. This explains why the OU process is also called the ‘oscillator process’.

13. Note that the FPE above shows that p(x, t|0, 0) tends to an equilibrium distri-
bution peq (x) as t → ∞.Write down this equilibrium distribution. Observe that the
mean squared displacement of the particle tends to an equilibrium value as t → ∞,
instead of diverging linearly with time, unlike the case of a freely diffusing particle.
In other words, there is no long-range diffusion of a harmonically bound Brownian
particle, owing to the confining potential.

6
14. Diffusion in the presence of a magnetic field: Let us return to the LE for
the velocity of a Brownian particle in a constant, uniform magnetic field, namely,

Γ
v̇j = −γvj + ωc Rjk vk + ζj (t) (j = 1, 2, 3).
m
In the high-friction approximation, this equation becomes

Γ
γvj − ωc Rjk vk = ζj (t)
m
or, in matrix form, √
Γ
(γI − ωc R)v = ζ(t).
m
But this is a Langevin equation for the position r(t) of the particle,

dr Γ
= (γI − ωc R)−1 ζ(t).
dt m
This SDE has no drift √ term, only a diffusion term: it is evident that, in this case,
n = ν = 3, f = 0, g = ( Γ/m)(γI − ωc R)−1 . The corresponding FPE will enable us
to identify the diffusion coefficients in the problem.

15. The matrix inverse (γI − ωc R)−1 must be found first. Since we know that
R3 = −R, it is clear that this matrix inverse must be a linear combination of I, R
and R2 . But there is a simpler way to find (γI − ωc R)−1 . The exponential matrix
eωc Rt has already been found as an explicit function of t. The Laplace transform of
this exponential is
Z ∞ Z ∞
dt eωc Rt−st
= dt e−(sI−ωc R)t = (sI − ωc R)−1 .
0 0

Substituting s by γ then yields the inverse sought. Take the Laplace transform of

eωc Rt = I + R sin ωc t + R2 (1 − cos ωc t)



to find the matrix g = ( Γ/m)(γI − ωc R)−1 , and hence calculate the elements of the
diffusion matrix D = 21 (ggT ) to get the reult

kB T n γ2 o
Djk = nj nk + 2 (δ jk − n n
j k ) .
mγ γ + ωc2

This expression shows precisely how the magnetic field curbs the diffusion transverse
to it, while leaving the longitudinal diffusion coefficient unaffected. It also illustrates
the interplay of the two time scales in the problem, γ −1 and ωc−1 . Taking B to be
directed along the z-axis, so that n1 = n2 = 0 and n3 = 1, we have

kB T kB T γ2
D33 = , while D11 = D22 = .
mγ mγ γ 2 + ωc2

You might also like