M.Phys.

C6 Theory Option
Lecture Notes for 2020-2021
Lectures given by Sid Parameswaran in Michaelmas 2020
Lecture Notes written by F.H.L. Essler with contributions by Sid Parameswaran
The Rudolf Peierls Centre for Theoretical Physics
Oxford University, Oxford OX1 3PU, UK

November 4, 2020

These notes cover Part V of the lectures (Random Systems and Stochastic Processes), which is only part
of M.Phys. C6 Theory Option.
Please report errors and typos to [email protected].
2015
c F.H.L. Essler.

Contents

V Random Systems and Stochastic Processes 3

1 Random Variables 3
1.1 Some Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Discrete-time random walk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 The central limit theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Stochastic Processes: Definitions 6

3 Markov Processes 6
3.1 Examples of Markov Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 Markov Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

4 Brownian Motion 9
4.1 Langevin Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.2 Fokker-Planck Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.3 Diffusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

5 Brownian motion via the Path Integral 14

5.1 Homework Questions 22-26 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

Some general remarks:

These notes aim to be self-contained. Homework questions are marked in red, and are placed at appro-
priate positions in the text, i.e. to work them out you will require only the preceeding material. Passages
marked in blue give details on derivations we don’t have time to go through in the lectures, or present ma-
terial that goes beyond the core of the course. In some cases this material will be very useful for particular
homework problems. All of the material covered in the course can be found in some form or other in a
variety of books. These is no book that covers everything. Some useful references for this section are

1
• Basic ideas of probability, kinetic theory, etc.
M. Kardar, Statistical Physics of Particles, Cambridge.

• Stochastic Processes
N.G. van Kampen, Stochastic Processes in Physics and Chemistry, North-Holland/Elsevier.

2
Part V
Random Systems and Stochastic Processes
In nature there are many phenomena, in which some quantity varies in a random way. An example is
Browninan motion, which refers to the motion of a small particle suspended in a fluid. The motion, observed
under a microscope, looks random. It is hopeless to try to compute the position in detail, but certain average

Figure 1: Cartoon of a path of a particle undergoing Brownian motion.

features obey simple laws. Averaging over a suitable time interval is difficult and one therefore replaces time
averaging of a single irregularly varying function of time by averaging over an ensemble of functions. The
latter must of course be chosen in such a way that the two results agree

time average −→ ensemble average. (1)

1 Random Variables
A random variable is an object X defined by

1. a set {xj } of possible values (either discrete or continuous);

2. a probability distribution PX (xj ) over this set

X
PX (xj ) ≥ 0 , PX (xj ) = 1 (this becomes an integral in the continuous case.). (2)
j

Example: Let X be the number of points obtained by casting a die

1
{xj } = {1, 2, 3, 4, 5, 6} , PX (xj ) = . (3)
6

1.1 Some Definitions

The probability distribution PX (x) can be characterized by the moments of X
Z
n
hX i ≡ dx xn PX (x).
(4)

The average value of X is the first moment hXi, while the variance is σ 2 = hX 2 i − hXi2 .

3
 The Fourier transform of PX (x) is called the characteristic function

∞
(ik)n
Z X
φX (k) = dx PX (x)eikx = hX n i ≡ heikX i.
n!
n=0
(5)

The last equality shows that the characteristic function is the generating function for the moments
dn 
hX n i = (−i)n n  φX (k). (6)
dk k=0
The cumulants of PX (x) are defined as

dn 
Cn (X) = (−i)n


n
ln φX (k) .
dk k=0

(7)

The first few cumulants are

C1 (X) = hXi , C2 (X) = hX 2 i − hXi2 , C3 (X) = hX 3 i − 3hX 2 ihXi + 2hXi3 . (8)

The Gaussian distribution is defined by

1 (x−x0 )2
PX (x) = √ e− 2σ 2 . (9)
2πσ 2
Its characteristic function is
1 2 2
φX (k) = eikx0 − 2 k σ
. (10)
The cumulants are C1 (X) = x0 , C2 (X) = σ 2 , Cn>2 (X) = 0. Hence the Gaussian distribution is completely
determined by its first two cumulants.

1.2 Discrete-time random walk

To see the above definitions in action we consider a tutor on (without loss of generality) his way back to
Summertown after a long night out (needless to say it must be a humanities tutor). He moves along Banbury
road, by making each second a step forward/backward with equal probability. Modelling Banbury road by
a line, his possible positions are all integers −∞ < n < ∞ (assuming for simplicity that Banbury road is
infinitely long, which is probably how it feels to our tutor), and we want to know the probability PN (n) for
her to be at position n after N steps, starting from n = 0.
• Each step is a random variable Xj (j = 1, . . . , N ) taking the values ±1 with probabilities 1/2.

• The position after r steps is Y = N

P
j=1 Xj .

Clearly we have
hY i = 0 , hY 2 i = N hXj2 i = N , (11)
where we have used that the steps are mutually independent. To obtain pN (n) we employ the characteristic
function
φXj (k) = PXj (1)eik1 + PXj (−1)eik(−1) = cos(k). (12)
The characteristic function of the random variable Y is
PN N
Y
N
φY (k) = heikY i = heik j=1 Xj
i= heikXj i = φXj (k)
j=1
N  
1  ik −ik
N 1 X N ik(N −2r)
= e +e = N e . (13)
2N 2 r
r=0

4
On the other hand we have by definition of the characteristic function
X N
X
ikn
φY (k) = PY (n)e = pN (n)eikn . (14)
n n=−N

Equating (13) with (14) gives  

1 N
pN (n) = N N −n , (15)
2 2
where the binomial coefficient is taken to be zero if (N − n)/2 is not an integer between 0 and N .

1.3 The central limit theorem

Let Xj j = 1, . . . , N be independent random variables with identical distributions PX (x) such that hXi and
hX 2 i are finite. and consider their sum
XN
SN = Xj . (16)
j=1
The central limit theorem states that, for large N , the random variable
SN − N hXi
Y ≡ √ (17)
N
is distributed as a Gaussian with zero mean and variance hX 2 i − hXi2 , irrespective of the form of PX (x),
provided that hXi and hX 2 i are finite.
To see that this is the case, let us consider the characteristic function of Y :
√
i √k SN
φY (k) = he N ie−ik N hXi
(18)
Now,
k
N
PN  
i √k SN i √k Xj ik √X N
he N = he N i = he N i = φX √
j=1
, (19)
N
where we have used that the random variables are independent and have the same distribution. Thus, we
have
k
N −ik√N hXi
 
φY (k) = φX √ e (20)
N
The cumulants of Y are
n  √
 
n d  k


Cn (Y ) = (−i) N ln φX √ − ik N hXi . (21)
dk n k=0

N
We see that these cumulants are related to the cumulants Cn (X) of PX (x) by
√
Cn (Y ) = N 1−n/2 Cn (X) − δn,1 N hXi.
(22)
In particular, we see that C1 (Y ) = 0, C2 (Y ) = C2 (X), and all higher moments Cn>2 (Y ) = N 1−n/2 Cn (X)
vanish for N  1. Hence, for very large N , PY (y) tends to a Gaussian distribution with average 0 and
variance C2 (X) = σ 2 , i.e.
 1  
1 2 1 2
PY (y) −→ exp − 2 y .
2πσ 2 2σ
(23)
Note that by rescaling, this can also be phrased in terms of the distribution of the arithmetic mean of the
Xj s (which is a Gaussian with O(1) mean and O(1/N ) variance, rather than the O(N ) mean and a O(1)
variance of the sum), as in previous editions of these notes. However, in my opinion the proof for the sum
— particularly the vanishing of higher cumulants — is cleaner.

5
2 Stochastic Processes: Definitions
A function YX (t) of time t and a random variable X is called a stochastic process (SP). Examples are
the position x(t) or the velocity v(t) of a particle undergoing Brownian motion. A SP is characterized by
probability densities

P1 (y1 , t1 ) = probability that Y (t1 ) = y1 ;

..
.
Pn (y1 , t1 ; . . . ; yn , tn ) = joint probability that Y (t1 ) = y1 , . . . , Y (tn ) = yn . (24)

These are

• normalized Z
dy1 . . . dyn Pn (y1 , t1 ; . . . ; yn , tn ) = 1 ; (25)

• reducible Z
dyn Pn (y1 , t1 ; . . . ; yn , tn ) = Pn−1 (y1 , t1 ; . . . ; yn−1 , tn−1 ). (26)

Another way of characterizing a SP is through correlation functions

Z
hY (t1 )i = dy1 y1 P (y1 , t1 ) ,
Z
hY (t1 )Y (t2 )i = dy1 dy2 y1 y2 P2 (y1 , t1 ; y2 , t2 ) ,
... (27)

A stochastic process is called stationary if for all n and any t

Pn (y1 , t1 ; . . . ; yn , tn ) = Pn (y1 , t1 + τ ; . . . ; yn , tn + τ ). (28)

The conditional probability that Y = y2 at time t2 , given that Y = y1 at time t1 is denoted by

P1|1 (y2 , t2 |y1 , t1 ). (29)

This is normalized Z
dy2 P1|1 (y2 , t2 |y1 , t1 ) = 1 , (30)

and related to unconditional probabilities by

P2 (y1 , t1 ; y2 , t2 ) = P1|1 (y2 , t2 |y1 , t1 ) P1 (y1 , t1 ). (31)

The conditional probability that Y = yn at time tn , given that Y = yj at time tj for j = 1, . . . , n − 1 is

denoted by
P1|n−1 (yn , tn |y1 , t1 ; y2 , t2 ; . . . ; yn−1 , tn−1 ). (32)

3 Markov Processes
Perhaps the most important stochastic processes are so-called Markov processes (MP). Their defining prop-
erty is that t1 < t2 < . . . < tn

P1|n−1 (yn , tn |y1 , t1 ; y2 , t2 ; . . . ; yn−1 , tn−1 ) = P1|1 (yn , tn |yn−1 , tn−1 ).

(33)

6
This means that at time tn−1 one can predict the state of the system at time tn on the basis of present
information, i.e. yn−1 , only! The history of how the system arrived at yn−1 at time tn−1 is irrelevant.
A Markov process is completely determined by the two functions P1 (y1 , t1 ) and P1|1 (y2 , t2 |y1 , t1 ), and
this makes Markov processes tractable. Any P1 (y1 , t1 ) and P1|1 (y2 , t2 |y1 , t1 ) define a MP, provided that they
fulfil the following two consistency conditions

• Chapman-Kolmogorov equation
Z
P1|1 (y3 , t3 |y1 , t1 ) = dy2 P1|1 (y3 , t3 |y2 , t2 )P1|1 (y2 , t2 |y1 , t1 ) , t3 > t2 > t1 .
(34)

• Evolution equation
Z
P1 (y2 , t2 ) = dy1 P1|1 (y2 , t2 |y1 , t1 )P1 (y1 , t1 ) , t2 > t1 .
(35)

A MP is called stationary if P1 (y, t) is time-independent and P1|1 (y2 , t2 |y1 , t1 ) depends only on the time
difference t2 − t1 (and y1,2 ).

3.1 Examples of Markov Processes

• The discrete-time random walk is a MP.

• The Wiener process, defined by

2
1 (y −y )
− 2 1
P1|1 (y2 , t2 |y1 , t1 ) = p e 2(t2 −t1 ) ,
2π(t2 − t1 )
1 − y2
P1 (y, t) = √ e 2t , (36)
2πt
is a non-stationary Markov process. It was originally invented for describing the stochastic behaviour
of the position of a Brownian particle, and is also called a continuous-time random walk.

3.2 Markov Chains

A Markov chain is a MP, in which the random variable only takes a finite number of values and in-
volves discrete time steps. The probability P1 (y, t) can then be represented as a N -component vector,
and P1|1 (y2 , t2 |y1 , t1 ) ≡ T as an N × N matrix. The normalization condition (30) becomes

N
X
Tjk = 1,
j=1
(37)

which is often referred to as probability conservation.

Example: Two state process with Y = 1 or Y = 2, with conditional probabilities

P1|1 (1, t + 1; 1, t) = 1 − q ,
P1|1 (2, t + 1; 1, t) = q ,
P1|1 (1, t + 1; 2, t) = r ,
P1|1 (2, t + 1; 2, t) = 1 − r . (38)

7
Introducing a two-component vector  
P1 (1, t)
p~(t) = , (39)
P1 (2, t)
the evolution equation for the process can be expressed as a vector equation
 
1−q r
p~(t + 1) = p~(t).
q 1−r
| {z }
T
(40)

T is a square matrix with non-negative entries, that is in general not symmetric. The matrix elements Tij
are the rates for transitions from state j to state i. It is useful to rewrite the equation in components
X
pn (t + 1) = Tnj pj (t). (41)
j

Then X X X
pn (t + 1) − pn (t) = Tnj pj (t) − Tjn pn (t) = Tnj pj (t) − Tjn pn (t). (42)
j j j
| {z }
1

Now consider our time interval to be small instead of 1. Then (42) turns into a differential equation, called
a Master equation
dpn (t) X
= Tnj pj (t) − Tjn pn (t).
dt
j
(43)
This has a nice physical interpretation as a “loss/gain” equation for probabilities: the first term on the
right-hand side is the rate of transitions from state j to state n times the probability of j being realized, i.e.
the total gain of probability for state n. The second term on the right-hand side is the rate of transitions
from state n to state j times the probability of n being realized, i.e. the total loss of probability for state n.
Let us now return to the discrete form (40), which can be iterated to give

p~(t + 1) = T t+1 p~(0). (44)

While T is generally not symmetric, it is nevertheless often diagonalizable. Then there exist left and right
eigenvectors such that

T |Rα i = λα |Rα i , hLα |T = λα hLα | , hLα |Rβ i = δα,β , (45)

and T can be represented in the form X

T = λα |Rα ihLα |. (46)
α
Using the orthonormality of left and right eigenvectors we have
X
T t+1 = λt+1
α |Rα ihLα |.
α
(47)

In our example  
1 r/q
λ1 = 1 , hL1 | = (1, 1) , |R1 i = . (48)
1 + r/q 1
 
q 1 −1
λ2 = 1 − q − r , hL2 | = (− , 1) , |R2 i = . (49)
r 1 + r/q 1

8
So for large t we have  
t+1 1 r r
T ≈ |R1 ihL1 | = , (50)
r+q q q
and hence  
1 r r
p~(∞) = p~(0). (51)
r+q q q

4 Brownian Motion
We now want to think of Brownian motion as a Markov process. Let v1 , v2 , . . . be the velocities of the
particle at different time steps. Then vk+1 depends only on vk , but not on v1 , . . . , vk−1 .

4.1 Langevin Equation

One approach to Brownian motion is via a stochastic differential equation, the Langevin equation for the
velocity v(t) (more precisely the velocity in D = 1 or a component of the velocity in D > 1)

dv(t)
= −γv(t) + η(t).
dt (52)
Here the first term on the right hand side is damping term linear in v, while the second term represents the
remaining random force with zero average hη(t)i = 0. This is often referred to as “noise”. For simplicity we
will assume collisions to be instantaneous, so that forces at different times are uncorrelated
hη(t)η(t0 )i = Γδ(t − t0 ). (53)
Given our assumptions about the noise, we can calculate noise-averaged quantities quite easily. We have
 
d  0 γt0  dv 0 0
0
v(t )e = 0
+ γv eγt = η(t0 )eγt , (54)
dt dt
where in the last step we used the Langevin equation (52). Integrating both sides of this equation between
t = 0 and t0 = t, we obtain
Z t
−γt 0
v(t) = v(0)e + dt0 η(t0 )e−γ(t−t ) .
0
(55)
Averaging this over the noise, we find the average velocity
Z t
0
hv(t)i = v(0)e−γt + dt0 hη(t0 )i e−γ(t−t ) = v(0)e−γt .
0 | {z }
=0
(56)
Similarly we have
 Z t  Z t 
2 −γt 0 0 −γ(t−t0 ) −γt 00 00 −γ(t−t00 )
hv (t)i = h v(0)e + dt η(t )e v(0)e + dt η(t )e i
0 0
Z t
0 00
= v 2 (0)e−2γt + e−2γt dt0 dt00 hη(t0 )η(t00 )i eγ(t +t ) (57)
0 | {z }
Γδ(t0 −t00 )

Carrying out the time integrals this becomes

Γ
hv 2 (t)i = v 2 (0)e−2γt + (1 − e−2γt ).
2γ
(58)

9
The displacement of the particle is
Z t Z t Z t0
0 0 v(0) −γt 0 0 00
x(t) − x(0) = dt v(t ) = (1 − e ) + dt dt00 η(t00 )e−γ(t −t ) . (59)
0 γ 0 0

Assuming x(0) = 0, the average position of the particle is

v(0)
hx(t)i = (1 − e−γt ).
γ
(60)

Finally, we want to determine the particle’s mean square deviation

2
h x(t) − hx(t)i i = hx2 (t)i − hx(t)i2 .

(61)

Substituting (59) and using again that x(0) = 0 this becomes

Z t Z t1 Z t Z t0 Z t Z t1 Z t
1 0 0 0
dt1 dt2 e−γ(t1 −t2 ) dt01 dt02 e−γ(t1 −t2 ) hη(t2 )η(t02 )i = dt1 dt2 e−γ(t1 −2t2 ) dt01 e−γt1 . (62)
0 0 0
|0 {z } 0 0
| t2 {z }
0
ΓΘ(t01 −t2 )e−γ(t1 −t2 ) γ −1 (e−γt2 −e−γt )

Carrying out the remaining two integrals we find

2 Γ Γ Γ
h x(t) − hx(t)i i = 2 t − 3 (1 − e−γt ) − 3 (1 − e−γt )2 .

γ γ 2γ
(63)

So at very late times we have

Γ
hx2 (t)i = t + ...
γ2
(64)
√
The displacement grows like t, which is characteristic of diffusion. Finally, we may relate Γ/γ to the
temperature of the fluid by noting that
Γ
hv 2 (t → ∞)i = . (65)
2γ
On the other hand, by equipartition we have
m 2 kB T
hv i ∼ (66)
2 2
Combining these two equations, we arrive at

Γ kB T
= .
2γ m
(67)

4.2 Fokker-Planck Equation

We now want to derive a differential equation for the probability P1 (v, t) of our particle having velocity v
at time t from the Langevin equation
Z t
−γt 0
v(t) = v(0)e + dt0 e−γ(t−t ) η(t0 ). (68)
0

Our starting point is the general evolution equation

Z
P1 (v, t + ∆t) = duP1|1 (v, t + ∆t|u, t) P1 (u, t). (69)

10
It is convenient to consider the integral
Z
Ω = dv [P1 (v, t + ∆t) − P1 (v, t)] h(v), (70)

where h(v) is is test function (infinitely many time differentiable, h(v) and all of its derivatives going to zero
at infinity etc). On the one hand, we have to linear order in ∆t
Z Z
∂P1 (v, t)
dv [P1 (v, t + ∆t) − P1 (v, t)] h(v) = dv ∆t h(v). (71)
∂t

On the other hand, using (69) we have

Z Z
Ω = du dv h(v) P1|1 (v, t + ∆t|u, t) P1 (u, t) − dv h(v) P1 (v, t). (72)

Relabelling
R the integration variable from v to u in the second term, and using that normalization condition
dv P1|1 (v, t + ∆t|u, t) = 1, we obtain
Z Z
Ω = du P1 (u, t) dv P1|1 (v, t + ∆t|u, t) [h(v) − h(u)] . (73)

Expanding h(v) around u in a Taylor series gives

∞
(v − u)n
Z X Z
(n)
Ω= du P1 (u, t) h (u) dv P1|1 (v, t + ∆t|u, t) . (74)
n!
n=1 | {z }
D(n) (u)

Integrating the n’th term in the sum n times by parts, and using the nice properties of the function h(u),
then leads to the following expression
∞ 
∂ n
Z X 
Ω= du h(u) − P1 (u, t)D(n) (u). (75)
∂u
n=1

Using that (71) and (75) have to be equal for any test function h(u), we conclude that

∞ 
∂ n

∂P1 (v, t) X
∆t = − P1 (v, t)D(n) (v).
∂t ∂v
n=1
(76)

This starts looking like our desired differential equation. What remains is to determine the quantities D(n) (v)

(w − v)n zn
Z Z
(n)
D (v) = dw P1|1 (w, t + ∆t|v, t) = dz P1|1 (v + z, t + ∆t|v, t)
n! n!
1
= h[v(t + ∆t) − v(t)]n i. (77)
n!

We see that D(n) (v) are related to the moments of the velocity difference distribution! We can use the
Langevin equation to determine them, and the result is

hv(t + ∆t) − v(t)i = −γv(0)e−γt ∆t ,

2
h v(t + ∆t) − v(t) i = Γ∆t + O (∆t)2 ,



n
h v(t + ∆t) − v(t) i = O (∆t)2 , n ≥ 3.


(78)

11
Substituting these into (76) and then taking the limit ∆t → 0, we arrive at the Fokker-Planck equation

∂ ∂ Γ ∂2
P1 (v, t) = γ vP1 (v, t) + P1 (v, t).
∂t ∂v 2 ∂v 2 (79)

This is a second order linear PDE for P1 (v, t) and can be solved by standard methods. For initial conditions
P1 (v, 0) = δ(v − v0 ) we find

2 !
1 v − v̄(t)
P1 (v, t) = p exp − ,
2πσ 2 (t) 2σ 2 (t)
(80)

where
Γ
σ 2 (t) = (1 − e−2γt ) , v̄(t) = v0 e−γt . (81)
2γ
In the limit t → ∞ this turns into the Maxwell distribution
r r 2
γ − γv2 m − mv
P1 (v, t) = e Γ = e 2kB T . (82)
πΓ 2πkB T

1: Moments of the Velocity Difference Probability Distribution

Let us see how to derive (78), starting from the Langevin equation
Z t
0
v(t) = v(0)e−γt + dt0 e−γ(t−t ) η(t0 ). (83)
0

For a very small time interval ∆t we have

Z t+∆t
−γ(t+∆t) 0
v(t + ∆t) = v(0)e + dt0 e−γ(t+∆t−t ) η(t0 )
0
Z t Z t+∆t
−γt 0 −γ(t−t0 ) 0 0
dt0 e−γ(t−t ) η(t0 ) + O (∆t2 )


= v(0)e (1 − γ∆t) + dt e η(t )(1 − γ∆t) +
0 t
Z t+∆t
0
dt0 e−γ(t−t ) η(t0 ) + O (∆t2 ) .


= (1 − γ∆t) v(t) + (84)
t

Hence ∆v(t) = v(t + ∆t) − v(t) is given by

Z t+∆t
0
dt0 e−γ(t−t ) η(t0 ) + O (∆t2 ) .


∆v = −γv∆t + (85)
t

To derive (78) we rewrite (85) in the form

Z
∆v(t) = v0 (t)∆t + dt0 K(t, t0 ) η(t0 ), (86)

where

v0 (t) = −γv(0)e−γt ,
0 0
K(t, t0 ) = −γΘ(t − t0 )Θ(t0 )e−γ(t−t ) ∆t + Θ(t0 − t)Θ(t + ∆t − t0 )e−γ(t−t ) . (87)

Let us now assume that η(t) is Gaussian distributed. Then the probability distribution for the noise
is the functional
1 2
R
P [η(t)] = e− 2Γ dt η (t) . (88)

12
 Averages are then given by the path integral
Z
1
R 0
η 2 (t0 )
hη(t1 ) . . . η(tn )i = Dη(t) η(t1 ) . . . η(tn ) e− 2Γ dt . (89)

As this is Gaussian, we may use Wick’s theorem to calculate averages

hη(t)i = 0 ,
hη(t)η(t0 )i = Γδ(t − t0 ) ,
hη(t1 )η(t2 )η(t3 )i = hη(t1 )η(t2 )ihη(t3 )i + hη(t3 )η(t1 )ihη(t2 )i + hη(t2 )η(t3 )ihη(t1 )i = 0 ,
hη(t1 )η(t2 )η(t3 )η(t4 )i = hη(t1 )η(t2 )ihη(t3 )η(t4 )i + . . . (90)

The probability distribution for ∆v(t) can be obtained from the generating function
Z
1
R 0 2 0 0 0
Z(λ) = heλ∆v(t) i = eλv0 (t)∆t Dη(t) e− 2Γ dt [η (t )−2λΓK(t,t )η(t )] . (91)

Changing variables to
η̃(t0 ) = η(t0 ) − λΓK(t, t0 ), (92)
the generating function becomes
λ2 Γ
dt0 K 2 (t,t0 )
R
Z(λ) = eλv0 (t)∆t+ 2 . (93)

It is then a straightforward matter to calculate the moments

dZ 
h∆vi = = v0 (t)∆t ,
dλ λ=0

Z t+∆t
d2 Z 
Z
0 −γ(t−t0 )

2 0 2 0 2


h ∆v i = = Γ dt K (t, t ) = Γ dt e + O (∆t)
dλ2 λ=0

t
2


= Γ∆t + O (∆t) ,

n dn Z 
= O (∆t)2 , n ≥ 3.


h ∆v i = n
(94)
dλ λ=0


4.3 Diffusion Equation

Finally, we would like to obtain a differential equation for P1 (x, t). The difficulty is that x(t) is not a Markov
process, because x(t + ∆t) depends on both x(t) and v(t). On can treat the combined evolution of x(t) and
v(t), but we will follow a simpler route. Recalling that
Z t
0
v(t) = v(0)e−γt + dt0 e−γ(t−t ) η(t0 ), (95)
0

we see that v(t) becomes a random variable for t  γ −1 with

Γ −γ|t−t0 |
hv(t)i = 0 , hv(t)v(t0 )i = e . (96)
2γ

Now let us imagine that we observe the Brownian particle only at sufficiently long time intervals t, t0  γ −1 ,
and describe only these coarse grained positions. Then we may replace
Γ −γ|t−t0 | Γ
hv(t)v(t0 )i = e → 2 δ(t − t0 ). (97)
2γ γ

13
 Γ −γ|t−t0 |
This is because 2γ e is substantially different from zero only if |t − t0 | < γ −1 , and
Z ∞
Γ −γ|t−t0 | Γ
dt e = 2. (98)
−∞ 2γ γ

The differential equation for the position

dx(t)
= v(t), (99)
dt
then turns into a special case (of no damping) of the Langevin equation we solved for the velocity. We
therefore can use our previous results to conclude that

∂ Γ ∂2
P1 (x, t) = 2 2 P1 (x, t).
∂t 2γ ∂x
(100)

This is the diffusion equation with diffusion coefficient

Γ
D= . (101)
2γ 2

Its solution for initial conditions P1 (x, 0) = δ(x − x0 ) is

2 !
1 x − x0
P1 (x, t) = p exp − .
2πD|t − t0 | 4D|t − t0 |
(102)

Observe that the diffusion equation is a special case of the Fokker-Planck equation in the absence of a
viscous damping term (the fact that there is a γ in the expression should not worry you – the equation for
x has no damping term, which is the important thing.)

5 Brownian motion via the Path Integral

We now give a brief description of the link between stochastic processes and path integrals, focusing on
Markov processes, and specifically working with the example of Brownian motion studied above.
Consider a Markov process; it is characterized by the two functions P1 (y1 , t1 ) and P1|1 (y2 , t2 |y1 , t1 ),
satisfying the Chapman-Kolmogorov (34) and evolution (103) equations.
For a stochastic process the ‘state’ of a system at time t0 is captured by the probability distribution
function P1 (y0 , t0 ) of the stochastic variable at time t0 and the evolution equation describes how this evolves
to the state at time t: Z
P1 (y, t) = dy0 P1|1 (y, t|y0 , t0 )P1 (y1 , t1 ) , t2 > t1 . (103)

In quantum mechanics, we have an analogous equation :

|Ψ(t)i = U (t; t0 )|Ψ(t0 )i (104)

Z
hx|Ψ(t)i = dx0 hx|U (t; t0 )|x0 ihx0 Ψ(t0 )i (105)

where in the second step, we have chosen to work in the position basis. Evidently, there is a formal resem-
blance between equations (103) and (105): namely, P1|1 (y, t|y0 , t0 ) plays a role analogous to the propagator
U (t; t0 ) in quantum mechanics. Given the state of the system at a given time t0 , it ‘propagates’ this forward
in time to time t > t0 . Of course, (105) describes probability amplitudes rather than probabilities, and

14
therein lies a key difference between quantum and classical physics. However, we can exploit the formal
resemblance between the classical evolution of probabilities in a Markov process and the quantum evolution
of amplitudes under unitary time evolution to recast the former in path integral form. We now carry out
this procedure, specializing to the case of Brownian motion.
Recall that the fundamental idea in the path integral approach to QM was to write the time evolution
operator as a path integral by inserting a complete set of states after each infinitesimal time step. The
analogous idea for the Markov process is to repeatedly use the Chapman-Kolmogorov equation to split up
the time evolution into infinitesimal chunks. Let us split the time interval from t0 to t into N equal time
slices, defining  = t−t N , tn = t0 + n, so that tN = t. Repeatedly applying (34), we have
0

Z
P1|1 (y, t|y0 , t0 ) = dy1 dy2 . . . dyN −1 P1|1 (y, t|yN −1 , tN −1 ) . . . P1|1 (y2 , t2 |y1 , t1 )P1|1 (y1 , t1 |y0 , t0 ) (106)
Z N
Y −1
= dy1 dy2 . . . dyN −1 P1|1 (yn+1 , tn+1 |yn , tn ). (107)
n=0

Just as for the QM case, we have broken up the evolution into infinitesimal evolutions. The reason is that it’s
frequently easier to compute the infinitesimal time evolution. Let’s consider the case of Brownian motion,
where we have from (36) that
−yn )2
1 (y
− 2(tn+1 1 − (y(tn+1 )−y(tn ))2
P1|1 (yn+1 , tn+1 |yn , tn ) = p e n+1 −tn ) =√ e 2 , (108)
2π(tn+1 − tn ) 2π
where in the second step we have used the fact that in our case tn − tn−1 =  by construction, and defined
yn = y(tn ). Substituting this into (107) and rewriting, we have
N −1
1 − (y(tn+1 )−y(tn ))2
Z Y
P1|1 (y, t|y0 , t0 ) = dy1 dy2 . . . dyN −1 √ e 2

n=0
2π
 N Z PN −1 (y(tn+1 )−y(tn ))2
1
= √ dy1 dy2 . . . dyN −1 e− n=0 2 . (109)
2π
(y(tn+1 )−y(tn ))
We now let N → ∞, which defines a trajectory or path y(t). We recognise once again that  '
ẏ(tn ), and hence that
N −1
ẏ(t0 )2
Z t
X (y(tn+1 ) − y(tn ))2
' dt0 . (110)
2 0 2
n=0
Thus we find that we have expressed the propagator for Brownian motion as a path integral,
Z R t 0 ẏ(t0 )2
P1|1 (y, t|y0 , t0 ) = N Dy(t0 )e− 0 dt 2 . (111)

which is the analogue of Eq.(53) of the first part of the notes.

Observe now that if we define an ‘imaginary time’ t̃ = −it, we find that this is identical to the free-particle
path integral (Eq. (64) from the first part) for ~ = m = 1. Note that this transformation also maps between
the transition probability for the Wiener process (36) and the free-particle propagator for t0 = 0 (Eq. (70)
of the first part). Amusingly, we see that this same transformation changes the diffusion equation (100)
to the Schrödinger equation, which, more generally, links the Fokker-Planck and path integral descriptions.
The course has indeed come full circle (pun intended).

5.1 Homework Questions 22-26

Question 22. This question is concerned with the central limit theorem.
(i) Show explicitly that for N  1, pN  1 the binomial distribution
N!
PN (n) = pn q N −n , p+q =1
n!(N − n)!

15
becomes
−(n− < n >)2
 
1
PN (n) = √ exp
2πσ 2 2σ 2
where σ 2 = N pq. Check that the same result follows from the central limit theorem.
(ii) Consider a random walk in one dimension, for which the probability of taking a step of length x → x + dx is
1 γ
f (x)dx = dx.
π x + γ2
2

Find the probability distribution for the total displacement after N steps. Does it satisfy the central limit theorem?
Should it? What are the cumulants of this distribution?

Question 23. Let y = ±1. Show that

1n 0
o 1n 0
o
P1|1 (y, t | y 0 , t0 ) = 1 + e−2γ(t−t ) δy,y0 + 1 − e−2γ(t−t ) δy,−y0 (112)
2 2
obeys the Chapman-Kolmogorov equation.
Show that
1
P1 (y, t) = (δy,1 + δy,−1 ) (113)
2
is a stationary solution. Write P1|1 as a 2 × 2 matrix and formulate the Chapman-Kolmogarov equation as a
property of that matrix.

Question 24. This question is about a continuous random walk, also known as a Wiener process.
Show that for −∞ < y < ∞ and t2 > t1 the Chapman-Kolmogarov equation is satisfied for

(y2 − y1 )2
 
1
P1|1 (y2 , t2 | y1 , t1 ) = p exp − . (114)
2π(t2 − t1 ) 2(t2 − t1 )

Choose P1 (y1 , 0) = δ(y1 ). Show that for t > 0

−y 2
 
1
P1 (y, t) = √ exp . (115)
2πt 2t

Show that P1 (y, t) obeys the diffusion equation

∂P ∂2P
=D 2 (116)
∂t ∂y

for D = 12 . What is the solution for arbitrary D > 0?

Question 25. A particle suspended in a fluid undergoes Brownian motion in one dimension with position
x(t) and velocity v(t). This motion is modelled by the Langevin equation
dv
= −γv + η(t),
dt
where η(t) is a Gaussian random variable characterised completely by the averages hη(t)i = 0 and hη(t1 )η(t2 )i =
Γδ(t1 − t2 ). Discuss the physical origin of each of the terms in the Langevin equation.
What is meant by the term Markov process? Illustrate your answer by discussing which of the following are
Markov processes: (a) v(t) alone; (b) x(t) alone; (c) v(t) and x(t) together.
Show that for t > 0
Z t Z t1
v(0) −γt
x(t) = (1 − e ) + dt1 dt2 e−γ(t1 −t2 ) η(t2 )
γ 0 0

16
is a solution of the Langevin equation with initial condition x(0) = 0. Calculate the average hx(t) v(t)i and
discuss its limiting behaviour at short and long times.

Question 26. The time evolution of a stochastic system is represented by a master equation of the form

dpn (t) X
= Wnm pm (t) .
dt m

Explain briefly the meaning of this equation and discuss the assumptions on which it is based. What general
conditions should the matrix elements Wnm satisfy?
A molecule lies between two atomic-scale contacts and conducts charge between them. A simple model of
this situation is illustrated below. The model has three states: the molecule may be uncharged, or may carry a
single charge at either site A or site B but not both. Charges hop between these sites, and between the sites and
the contacts, at the rates indicated in the figure. (For example, a charge at site A has probability f2 per unit
time of hopping to site B.)
f1 f2 f3

A B
Write down a master equation for this model. For the system in equilibrium, calculate the occupation probabilities
of the three states, and show that the average number of charges flowing through the molecule per unit time is
f1 f2 f3
.
f1 f2 + f1 f3 + f2 f3
Consider the case f1 = f2 = f3 ≡ f . The molecule is uncharged at time t = 0. Show that the probability
p(t) for it to be uncharged at a later time t is
  √ !
1 2 3 3
p(t) = + exp − f t cos ft .
3 3 2 2

17