Physics 127b: Statistical Mechanics

Brownian Motion
Brownian motion is the motion of a particle due to the buffeting by the molecules in a gas or liquid.
The particle must be small enough that the effects of the discrete nature of matter are apparent, but
large compared to the molecular scale (pollen in the early experiments, various plastic beads these
days). It is a convenient example to display the residual effects of molecular noise on macroscopic
degrees of freedom. I will use this example to investigate the type of physics encountered, and the
tools used to treat the uctuations.
RandomWalk
The rst observation of Brownian motion is that the particle under the microscope appears to
perform a random walk, and it is rst useful to study this aspect in its simplest form.
Lets consider rst a one dimensional random walk, consisting of n jumps of l along the x axis.
We take n to be even (the odd case is essentially the same, but differs in minor details). For the
particle after n jumps to be at x = ml there must have been
1
2
(nm) forward jumps, and
1
2
(nm)
backwards jumps (in any order), and m must be even. The probability of arriving at x = ml is
therefore
p
n
(m) =
n!
_
1
2
(n m)
_
!
_
1
2
(n m)
_
!
. (1)
For large m, n Stirlings approximation n! . (2n)
1/2
(n/e)
n
gives
p
n
(m) =
2

2n
e
m
2
/2n
. (2)
This is a Gaussian probability centered around m = 0 (the most probable and mean position is the
origin) and the mean square displacement
_
m
2
_
= n, or
_
x
2
_
= nl
2
. (3)
For large n the discreteness of the displacements is unimportant compared to the root mean square
distance of the walk. Transforming to a continuous variable x and a probability density p(x, t )
using p
n
(m) = p(x)2l (since the interval between the discrete results is dx = 2l) and introducing
time supposing there are n jumps in time t
p(x, t ) =
1

4Dt
exp
_

x
2
4Dt
_
(4)
where we have written
nl
2
/2t = D. (5)
We recognize that this is the expression for diffusion, with p satisfying
p
t
= D

2
p
x
2
, p(x, t = 0) = (x) (6)
1
with the diffusion constant D. In terms of D
_
x
2
_
= 2Dt. (7)
These results are readily extended to 3 dimensions, since we can consider a walk with steps
(l, l, l) for example, so that the walk is the product of walks in each dimension. The mean
square distance gone after n walks is again
_
r
2
_
= nL
2
with L =

3l the length of each step. The
probability distribution p( x, t ) satises the 3d diffusion equation
p
t
=
1
(4Dt )
3/2
exp
_

r
2
4Dt
_
(8)
with r
2
= x
2
y
2
z
2
. This equation is simply the product of three 1d diffusion equations with
D = nl
2
/2t as before. The means square distance is
_
r
2
_
= 6Dt. (9)
(The results in 2d can similarly be constructed.)
The fact that the mean displacement is zero, and the mean square displacement grows linearly in
time can be derived by very simple arguments. Lets consider the two dimensional case of a random
walk consisting of n vectors of length s but with arbitrary angles
i
taken froma uniformprobability
distribution. The total displacement in the x direction is
X =

i
s cos
i
. (10)
Clearly X) = 0 since cos
i
is equally likely to be positive or negative. On the other hand
_
X
2
_
=
__

i
s cos
i
_
2
_
= s
2
_

i
cos
i

j
cos
j
_
(11)
= s
2
_

i
(cos
i
)
2
_
= ns
2
/2 (12)
where we have used the fact that
_

i,j,=i
cos
i
cos
j
_
= 0 since again each cos
i
is equally likely
to be positive or negative. Thus the mean square distance is
_
R
2
_
=
_
X
2
Y
2
_
= ns
2
. (13)
This specic result is useful in adding complex numbers with randomphases: the average amplitude
is zero, and the mean square magnitude (the intensity) scales linearly with the number of vectors.
Some general nomenclature
The position x(t ) in a one dimensional random walk forms a one dimensional random processin
general a scalar function y(t ) for which the future data is not determined uniquely by the known
initial data.
2
The random process is in general characterized by probability distributions p
1
, p
2
. . . such that
p
n
(y
1
, t
1
; y
2
, t
2
. . . ; y
n
, t
n
)dy
1
dy
2
. . . dy
n
(14)
is the probability that a single process drawn from the ensemble of processes will take on a value
between y
1
and y
1
dy
1
at t
1
etc. The different p
n
are related by
_

p
n
dy
j
p
n1
. (15)
Ensemble averages are determined by integrating over the appropriate distribution, e.g. for the mean
and the two point correlation function
y(t
1
)) =
_

y
1
p
1
(y
1
, t
1
)dy
1
, (16)
y(t
1
)y(t
2
)) =
_

y
1
y
2
p
2
(y
1
, t
1
; y
2
, t
2
)dy
1
dy
2
. (17)
Higher order correlation functions require the knowledge of higher order distribution functions. In
the random walk we have just looked at p
1
.
A stationary random process is one for which the p
n
depend only on time differences, or
p
n
(y
1
, t
1
; y
2
, t
2
; . . . ; y
n
, t
n
) = p
n
(y
1
, t
1
; y
2
, t
2
. . . ; y
n
, t
n
). (18)
I have chosen to formulate the random walk as starting a particle from a particular position at time
t = 0, so that x(t ) is not stationary. Alternatively we could have considered a stationary process
(e.g. the eld of vision of a microscope with many Brownian particles) and then calculated the
conditional probability P
2
(x
1
, t
1
[x
2
, t
2
) which is the probability of the particle being at x
2
at time t
2
given that it was at x
1
at time t
1
. Then P
2
(0, 0[x, t ) takes the diffusive form that we have calculated
and the p
n
all just depend on the time difference (p
1
(x, t ) is just constant, for example).
Means and correlation functions are dened with respect to the ensemble average. For a stationary
random process we usually assume ergodicity, and replace the ensemble average by a time average,
e.g.
y) = y(t ) = lim
T
1
T
_
T/2
T/2
y(t )dt. (19)
The probability distribution for the random walk is a Gaussian function. A Gaussian process in
general is one in which all the probability distributions are Gaussian
p
n
(y
1
, t
1
; y
2
, t
2
. . . ; y
n
, t
n
) = Aexp
_
_

j=1
n

k=1

jk
(y
j
y))(y
k
y))
_
_
(20)
where y) is the mean of y,
jk
is a positive denite matrix and A is a normalization constant. For
a stationary process y) is time independent and and A depend only on time differences.
Gaussian processes are important in physics because of the central limit theorem: if
Y =
1
N

y
i
(21)
3
with y
i
a randomprocess or variable with arbitrary distribution but with nite mean y) and variance

2
y
then for N large Y is a Gaussian process or variable
p(Y) =
1
_
2
2
Y
exp
_

(Y Y))
2
2
2
Y
_
(22)
with Y) = y) and
Y
=
y
/

N. The central limit theorem is why the Gaussian distribution

y(t )e
i2f t
dt, (23a)
y(t ) =
_

y(f )e
i2f t
df. (23b)
The correctness of the inverse is shown from the result
_

e
i2xy
dx = lim
x
sin 2xy
y
= (y). (24)
For a real function y(t ) we have y

(f ) = y(f ).
For a stationary random process the integral dening y(f ) diverges, so we instead dene the
auxiliary process
y
T
(t ) =
_
y(t ) T/2 < t < T/2
0 otherwise
(25)
and then use the nite y
T
(f ).
Parsevals theorem tells us
lim
T
1
T
_
T/2
T/2
[y(t )]
2
= lim
T
2
T
_

0
[ y
T
(f )[
2
df. (26)
Here and elsewhere we use y

(f ) = y(f ) to restrict the frequency domain to positive values.

_
T/2
T/2
[y(t ) y]e
i2f t
dt

2
. (27)
where y is the time average over T . Why do we use this expression? Lets suppose that the mean
has been subtracted off of y, so y = 0. The quantity inside the [ [ is the Fourier transform of the
4
process y
T
(t ) . How does this grow with T ? We can estimate this by supposing the interval T to
be formed of N subintervals of length . The Fourier transform y
T
is then the sum of N transforms
y

of processes y

(t )dened over the interval . For a random process we would expect each of
the y

to be of similar magnitude, but with arbitrary phase, since the latter depends sensitively on
the phasing of the e
i2f t
with respect to the time start of the signal. Adding N complex numbers
with random phase gives a number of magnitude

N and random phase. Thus the transform
of y(t ) y grows as

T and the phase varies over all values as T changes. The spectral density
G
y
(f ) is constructed to be independent of T and to contain all the useful information. Parsevals
theorem now gives us
_

0
G
y
(f )df = lim
T
1
T
_
T/2
T/2
[y(t ) y]
2
dt =
2
y
, (28)
so that the frequency integral of the spectral density is the variance of the signal.
The spectral density is directly related to the Fourier transform of the correlation function C
y
().
Lets set the mean y to zero for simplicity. Then, using assumption of ergodicity to replace the
ensemble average by a time average, the correlation function is
C
y
() = lim
T
1
T
_
T/2
T/2
dt y(t )y(t ) (29)
= lim
T
1
T
_

dt y
T
(t )y
T
(t ) (30)
where the small error in replacing y(t ) by y
T
(t ) is unimportant in the limit. Now inserting
the Fourier transforms and using y

(f ) = y(f )
C
y
() = lim
T
1
T
_

dt
_

df
_

df
/
y
T
(f ) y
T
(f
/
)e
i2f
/

e
i2(f f
/
)t
. (31)
The t integrations is (f f
/
), and using y

(f ) = y(f ) gives
C
y
() = lim
T
1
T
_

df [ y
T
(f )[
2
e
i2f
(32)
= lim
T
2
T
_

0
df [ y
T
(f )[
2
cos 2f (33)
=
_

0
G
y
(f ) cos(2f )df. (34)
Thus we have the inverse pair
C
y
() =
_

0
G
y
(f ) cos(2f )df (35a)
G
y
(f ) = 4
_

0
C
y
() cos(2f )d (35b)
(since C
y
and G
f
are both even functions, we have written the results as cosine transforms only
involving the positive domain). These equations are known as the Wiener-Khintchine theorem.
5
A particularly simple spectral density is a at one, independent of frequency. We describe such
a random process as being white. The corresponding correlation function is a delta function, i.e.
no correlations except for time differences tending to zerp. One strength parameter g is needed to
specify the force
G
F
(f ) = g, (36a)
C
F
() =
g
2
(). (36b)
The Einstein Relation
Einstein showed how to relate the diffusion constant, describing the random uctuations of the
Brownian particle, to its mobility , the systematic response to an externally applied force.
Under an applied force dV/dx the drift velocity of the particle is (the denition of the mobility)
u
d
=
dV
dx
. (37)
For a sphere of radius a in a liquid the viscosity the mobility is given by the Stokes expression
= (6au
d
)
1
, and so is related to the dissipation in the uid.
Consider now the thermodynamic equilibrium of a density n(x) of independent Brownian particles
in the potential V(x). We can dynamically understand the equilibrium in terms of the cancelling
of the particle currents due to diffusion and mobility
D
dn
dx
n
_

dV
dx
_
= 0. (38)
Equilibrium thermodynamics on the other hand tells us n(x) exp[V(x)/kT ]. Substituting into
Eq. (38) gives the Einstein identity
D = kT . (39)
Note the use of equilibrium constraints to relate uctuation quantities (the diffusion constant which
gives us
_
x
2
(t )
_
) and dissipation coefcients ( or ). This is an example of a general approach
known as uctuation dissipation theory, that we will take up again later. The fact that the uctuations
and dissipation of a Brownian particle are related should not be unexpected: both are a reection
of the molecular buffeting, the dissipation given by the net force due to the systematic component
of the collisions coming from the drift of the particle relative to the equilibrium molecular velocity
distribution, and the uctuations coming from the random component.
Fluctuation-Dissipation Theory
The relationship between the dissipation coefcient and the uctuations is made more explicit by
directly evaluating D in terms of the uctuations producing the random walk
D = lim
t
1
2t
_
[x(t ) x(0)]
2
_
. (40)
6
Expressing the displacement as the integral of the stochastic velocity u
x(t ) x(0) =
_
t
0
u(t
1
)dt
1
(41)
leads to
D = lim
t
1
2t
_
t
0
dt
1
_
t
0
dt
2
u(t
1
)u(t
2
)) , (42)
which depends on the velocity correlation function. The integrand is symmetric in t
1
, t
2
and we can
replace the integral over the square by twice the integral over the triangle 0 < t
1
< t, t
1
< t
2
< t ,
and then introducing the time difference = t
2
t
1
D = lim
t
1
t
_
t
0
dt
1
_
t t
1
0
d u(t
1
)u(t
1
)) . (43)
Since the correlation function u(t
1
)u(t
1
)) decays to zero in some nite relaxation time
r
, as
t the limit of the second integral can be replaced by innity for almost all values of t
1
in
the rst integration. Further, u(t
1
)u(t
1
)) = C
u
() is independent of t
1
(u(t ) is a stationary
random process if external conditions are xed). Hence
D =
_

0
d u(0)u()) (44)
and
=
1
kT
_

0
d u(0)u()) (45)
directly relating a dissipation coefcient to a correlation function.
7