Lecture 20: (Physical) Brownian Motion

Scribe: Neville E. Sanjana (and Martin Z. Bazant)

Department of Brain and Cognitive Sciences, MIT

April 21, 2005

1 Overview and simple models

When we talk about Brownian motion, we’re interested in the motion of a large particle in a gas
or liquid in equilibrium, which is roughly approximated by a random walk. You might imagine
something like this:

Figure 1: A large particle undergoing Brownian motion due to collision with smaller particles from a liquid
or gas.

Brownian motion is named for Robert Brown, who published a paper on his observations of
pollen particles.1 Prof. Bazant recommends looking at this web applet of a molecular dynamics
simulation of Brownian motion: http://www.phy.ntnu.edu.tw/java/gas2d/gas2d.html

1
Robert Brown’s original 1827 paper on his light microscope observations of pollen grains is available online (
http://sciweb.nybg.org/science2/pdfs/dws/Brownian.pdf). Interestingly, one of Brown’s first tasks was to make sure
that the motion continued even after plant death: “Having found motion in the particles of the pollen of all the living
plants which I had examined, I was led next to inquire whether this property continued after the death of the plant,
and for what length of time it was retained.” He goes on to show the existence of the motion in inorganic substances
and in organic substances after burning over a fire and thus concludes these “Active Molecules” are “readily obtained
from all bodies”.

1
M. Z. Bazant – 18.366 Random Walks and Diffusion – Lecture 20 2

Simplest models
1. Discrete RW with IID steps.
As mentioned in the first lecture, the simplest model of Brownian motion is a random walk
where the “steps” are random displacements, assumed to be IID random variables, between
nearly instantaneous collisions. The constant time step represents the mean time between
collisions, which is usually well defined near equilibrium, except in very small systems or very
low densities, where the mean free path is comparable to the system size. (We will discuss
this exotic regime of “low Knudsen number” gases later in the class.)

2. Continuous Wiener process.

This is the limit of infinitessimal displacements in the discrete random walk with IID steps,
first given a rigorous mathematical definition by MIT mathematician, Norbert Wiener, in the
1920s. We will discuss the Wiener process and its connection to discrete random walks later

in the class.

Mathematicians have come to call this formal construction “Brownian motion”, even though

it is only a crude approximation of the physical phenomenon of Brownian motion. Therefore,

the Wiener process is sometimes referred to as “mathematical Brownian motion”.

One curious feature of this limit, emphasized by Wiener, is that the particle has an infinite
velocity at all times2 It is easy to see this heuristically, assuming that the infinitessimal
displacement dx (stochastic differential) is a Gaussian random variable with mean zero and
variance 2D dt, where
√ D = limdt→0 �dx2 �/2 dt is the diffusivity. The typical displacement
� in
time dt has size 2D dt, which implies a divergent instantaneous velocity v = 2D/dt → ∞.

Both of these models have the attractive feature of mathematical simplicity, but each fails to describe
various important features of physical Brownian motion:

1. Inertia. Momentum is conserved after collisions, so a particle will recoil after a collision with
a bias in the previous direction of motion. This causes correlations in time, between successive
steps.

2. Ballistic motion. In a physical Brownian motion, there is in fact a well defined instantan-
teous velocity, which varies around some typical value. A more complete microscopic theory
of Brownian motion would account for the ballistic motion of a particle between collisions
(in a gas) or the gradual “turning of the steering wheel” when collision events are not well
separated in time (in a liquid).

3. Interactions between particles. Each particle, whether large or small, collides and inter-
acts with all of its neighbors, so it is not always possible to focus on just one independent
random walker interacting with a source of “noise”, although that is a reasonable first ap-
proximation.

Today, we will focus models addressing the first two problems.

NOTE: The material below is covered in more detail in lectures 9 –11 from 2003.
2
The trajectory of Brownian motion is thus an example of a curve which is continuous and yet nowhere differntiable.
This shows that stochastic calculus is markedly different from the more familiar continuous calculus.
M. Z. Bazant – 18.366 Random Walks and Diffusion – Lecture 20 3

2 Random walk with correlated steps

Consider unbiased (�Δx� = 0) identically distributed steps with �Δx� = σ 2 which are not indepen-
dent. In this case, if the correlations between steps are not too strong, the CLT still holds, but with
a modified diffusion coefficient.

N
�
XN = Δxn (1)
n=1
� � ��
2
XN = �Δxn · Δxm � (2)
n m
��
= Cn,m (3)
n m

Consider translation invariance in time, so that

Cn,m = C (|n − m|) Note: C(0) = σ 2 (4)

We have Cn,m symmetric with values decaying away from the diagonal. Since the correlations are
between times that are close together, then we expect the near diagonal entries to have greater
(magnitude) values than those further away from the diagonal. Of course, the diagonal terms C(0)
will all be σ 2 .
N
�
� 2
� 2
� �
XN
= Nσ + 2 N − n� C(n� ) where n� = n − m (5)
n� =1

In the continuum limit:

tn
N= , τ → 0, N →∞
τ
with t fixed
Δtn
= V (tn ) V = velocity
τ
Cn,m
= �V (tn ) · V (tm )�
τ2
σ2 (6)
Do =
2τ
�t
� 2
� � �
X(t) = 2Dt + V (0) · V (t� ) (t − t� )dt�
0
�∞
1 d � �
D ≡ lim X(t)2 = Do + �V (0) · V (t)�t>0 dt
2 dt
0

We have Do as the standard diffusion constant without any correlation and the additional
expectation to be taken t > 0. Note: It is common to introduce a delta function in �V (0) · V (t)�t>0 =
f (t) + δt>0 (t)Do . The first term f (t) is the memory of the past and the delta function is simply an
autocorrelation.
M. Z. Bazant – 18.366 Random Walks and Diffusion – Lecture 20 4

This is called the Green-Kubo relationship:

�∞
�V (0) · V (t)�t>0 dt (7)
0

It is useful in molecular dynamics simulations because you always know the velocity of every particle
at every time and thus can easily plug-in to get D. There exists a whole family of Green-Kubo
equations that relate transport coefficients (for mass, energy, momentum, etc.) to time integrals of
autocorrelation functions.
In the discrete case:
�
2 �

XN
D = lim (8)
N →∞ 2N τ

Note:
�
1. D
√ is finite (maybe =
� Do ) if C is short ranged (and not oscillating too fast). We get �X 2 � ∼

2Dt.
�

2. D → 0, C oscillates. We get �X 2 � ∝ tν with 0 ≤ ν < 12 . Subdiffusion.

�

3. D → ∞, C is long ranged. We get �X 2 � ∝ tν with 12 < ν ≤ 1. Superdiffusion.

3 Exponentially decaying correlations

Now, we will look at what happens when there is a direct correlation of p with the previous step
(lecture 9, 2003). When p > 0 this implies a tendency to keep going in the same direction after
a collision, sometimes called “persistence”, which approximates the effect of inertia in Brownian
motion. The “persistent random walk” can be traced back at least to 1921, in an early model of
G. I. Taylor for tracer motion in a turbulent fluid flow. (−1 < p < 1)
�

Δxn = pΔxn−1 + (1 − p2 )Δx�n (9)

This linear combination, illustrated in the figure on a circle, ensures that if the variables {Δx�n }N
n=1
are IID with zero mean and variance σ 2 , then the variables {Δxn }N n=1 are dependent, with correla-
tion p between successive steps and the same variance σ 2 . We have

C(1) = �Δxn · Δxn−1 � = pσ 2 (by construction) (10)

�
� � �

because Δx�n · Δx�m = σ 2 δn,m ( Δx�n = 0) (11)

C(2) = �Δxn · Δxn−2 � = p2 σ 2 (12)

m 2
C(m) = �Δxn · Δxn−m � = p σ (13)

−2m
⎫
C(m) = σ 2 e nc ⎪
⎬
−2 ⎪ (14)

nc = ⎭

logp
M. Z. Bazant – 18.366 Random Walks and Diffusion – Lecture 20
5

σn’

1-p 2

p σn-1

Figure 2: A geometrical construction for the “mixing” of a new random variable with the previous random
displacement to create a dependent random displacement, with correlation coefficient p.

� 2
� N
XN 2 � n
= 1+ p (N − n) (15)
N σ2 N
n=1
� �� �
geometric series

1+p 2 p(1 − pN )
= + (16)
1 − p N (1 − p)2

Long-time scaling:
D 1+p
= N →∞ (17)
Do 1−p
Recall from last lecture that p = 1/3 for a tetrahedral polymer chain.

Now, consider p → 1.

p = 1−� (18)
−2 2
nc = ∼ (19)
logp �
1+p 2
∼ = nc (20)
1−p �
Rescale with dimensionless variables

˜ = N,
N ˜ N = XN
X
nc nc σ

� � � �
˜2
X ˜ + 1 e−2N˜ − 1
∼ lim N (21)
N
�→0
� 2

N˜2 ˜
<< 1
N Ballistic motion at constant velocity
∼ (22)
N˜ ˜ >> 1
N Diffusion
M. Z. Bazant – 18.366 Random Walks and Diffusion – Lecture 20 6

With strong, exponentially decaying correlation (p = 1 − �), the correlated random walk is
equivalent at long times (N >> nc ) to an uncorrelated random walk with IID steps with size nc σ
and time step nc = tauc .

Historic moment3

Continuum limit
The persistent random walk in one dimension has an interesting continuum limit (see also lecture
10, 2003). Some definitions:

τc = τ /� , c = σ/τ (23)
p = 1−� (24)

Now, we will be taking the following limits

�→0 (25)
τ →0 (26)
σ→0 (27)

τc is fixed and c is fixed (ie. ballistic regime). We have P as the PDF for position under these
assumptions:

∂2P 1 ∂P 2
2∂ P
+ = c (28)
∂t2 τc ∂t ∂x2

∂2P 2
2∂ P
t << τc Wave equation ∼ c
∂t2 ∂x2 (29)
∂P ∂2P
t >> τc Diffusion equation ∼ D 2 D = σ 2 τc
∂t ∂x
The fundamental solution (Green function) for a localized initial condition at rest consists of two
delta function pulses moving in either direction (as in the wave equation), with decreasing amplitude,
to represent walkers undergoing ballistic motion before their first collision. Between these pulses is
a smooth distribution of walkers who have had at least one collision, which eventually tends to a
Gaussian shape (solving the diffusion equation) due to the long-time independence of steps which
implies the CLT. The variance takes the simple form
� � 2e2 � �
x (t)2 = 2 γt + e−γt − 1 (30)
γ
3
At this point in the lecture, Prof. Bazant received the important phone call that would, less than 12 hours later,
culminate in the birth of Isabella Rose Bazant. If there was ever a good reason to answer a cell phone in class, this
is it. Congratulations! (Note also that Prof. Bazant continued for an additional 15 minutes to actually derive the
continuum limit before going to the hospital. Impressive.)
M. Z. Bazant – 18.366 Random Walks and Diffusion – Lecture 20 7

showing a transition from ballistic to diffusive scaling at a time scale γ −1 . In the next lecture, we
will relate this parameter to the mass of the particle and the system temperature, by starting from
Newton’s laws of mechanics in a more complete stochastic theory of Brownian motion.
Note: a good reference for “physical Brownian motion” and continuous stochastic processes is
The Fokker-Planck Equation by Risken.