Contents

4 The Fokker-Planck and Master Equations 1

4.1 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
4.2 Fokker-Planck Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
4.2.1 Forward and backward time equations . . . . . . . . . . . . . . . . . . . . . . . . . 2
4.2.2 Surfaces and boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
4.2.3 One-dimensional Fokker-Planck equation . . . . . . . . . . . . . . . . . . . . . . . 3
4.2.4 Eigenfunction expansions for Fokker-Planck . . . . . . . . . . . . . . . . . . . . . . 5
4.2.5 First passage problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.2.6 Escape from a metastable potential minimum . . . . . . . . . . . . . . . . . . . . . 14
4.2.7 Detailed balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2.8 Multicomponent Ornstein-Uhlenbeck process . . . . . . . . . . . . . . . . . . . . . 18
4.2.9 Nyquist’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 Master Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3.1 Birth-death processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3.2 Examples: reaction kinetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3.3 Forward and reverse equations and boundary conditions . . . . . . . . . . . . . . 26
4.3.4 First passage times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3.5 From Master equation to Fokker-Planck . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3.6 Extinction times in birth-death processes . . . . . . . . . . . . . . . . . . . . . . . . 34

i
ii CONTENTS
Chapter 4

The Fokker-Planck and Master Equations

4.1 References

– C. Gardiner, Stochastic Methods (4th edition, Springer-Verlag, 2010)

Very clear and complete text on stochastic methods, with many applications.

– N. G. Van Kampen Stochastic Processes in Physics and Chemistry (3rd edition, North-Holland,
2007)
Another standard text. Very readable, but less comprehensive than Gardiner.

– Z. Schuss, Theory and Applications of Stochastic Processes (Springer-Verlag, 2010)

In-depth discussion of continuous path stochastic processes and connections to partial differential
equations.

– R. Mahnke, J. Kaupužs, and I. Lubashevsky, Physics of Stochastic Processes (Wiley, 2009)

Introductory sections are sometimes overly formal, but a good selection of topics.

1
2 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

4.2 Fokker-Planck Equation

Here we mainly follow the discussion in chapter 5 of Gardiner, and chapter 4 of Mahnke et al.

4.2.1 Forward and backward time equations

We have already met the Fokker-Planck equation,

∂P (x, t | x0 , t0 ) ∂ h i 1 ∂2 h i
=− Ai (x, t) P (x, t | x0 , t0 ) + Bij (x, t) P (x, t | x0 , t0 ) . (4.1)
∂t ∂xi 2 ∂xi ∂xj

Defining the probability flux,

1 ∂ h i
Ji (x, t | x0 , t0 ) = Ai (x, t) P (x, t | x0 , t0 ) − Bij (x, t) P (x, t | x0 , t0 ) , (4.2)
2 ∂xj

the Fokker-Planck equation takes the form of the continuity equation,

∂P (x, t | x0 , t0 )
+ ∇ · J (x, t | x0 , t0 ) = 0 . (4.3)
∂t

The corresponding backward Fokker-Planck equation is given by

∂P (x, t | x0 , t0 ) 0 0
0 0 ∂P (x, t | x , t ) 1
2 0 0
0 0 ∂ P (x, t | x , t )
− = +A i ( x , t ) + B
2 ij (x , t ) . (4.4)
∂t0 ∂x0i ∂x0i ∂x0j

The initial conditions in both cases may be taken to be

P (x, t | x0 , t) = δ(x − x0 ) . (4.5)

4.2.2 Surfaces and boundary conditions

Forward equation

Integrating Eqn. 4.3 over some region Ω, we have

Z Z
d
dx P (x, t | x , t ) = − dΣ n̂ · J (x, t | x0 , t0 ) ,
0 0
(4.6)
dt
Ω ∂Ω

where n̂ is locally normal to the surface ∂Ω. At surfaces we need to specify boundary conditions. Gen-
erally these fall into one of three types:

(i) Reflecting surfaces satisfy n̂ · J (x, t | x0 , t0 ) Σ

= 0 at the surface Σ.
(ii) Absorbing surfaces satisfy P (x, t | x0 , t0 ) Σ
= 0.
4.2. FOKKER-PLANCK EQUATION 3

(iii) Continuity at a surface entails

P (x, t | x0 , t0 ) Σ+
= P (x, t | x0 , t0 ) Σ−
, n̂ · J (x, t | x0 , t0 ) Σ = n̂ · J (x, t | x0 , t0 ) Σ .
+ −
(4.7)

These conditions on x may be enforced even if the functions Ai (x, t) and Bij (x, t) may be discontinuous
across Σ.

Backward equation

For the backward FPE, we have the following1 :

(i) Reflecting surfaces satisfy ni (x0 ) Bij (x0 ) ∂x∂ 0 P (x, t | x0 , t0 ) Σ

= 0 for x0 ∈ Σ.
j

(ii) Absorbing surfaces satisfy P (x, t | x0 , t0 ) Σ

= 0.

4.2.3 One-dimensional Fokker-Planck equation

Consider the Fokker-Planck equation in d = 1. On an infinite interval x ∈ (−∞, +∞), normalization

requires P (±∞, t) = 0, which generally2 implies ∂x P (±∞, t) = 0. On a finite interval x ∈ [a, b], we may
impose periodic boundary conditions P (a) = P (b) and J(a) = J(b).
Recall that the Fokker-Planck equation follows from the stochastic differential equation

dx = f (x, t) dt + g(x, t) dW (t) , (4.8)

p
with f (x, t) = A(x, t) and g(x, t) = B(x, t) , and where W (t) is a Wiener process. In general3 , a
solution to the above Itô SDE exists and is unique provided the quantities f and g satisfy a Lipschitz
condition, which says that there exists a K > 0 such that f (x, t) − f (y, t) + g(x, t) − g(y, t) < K|x − y|
for all x, y ∈ [a, b]4 . Coupled with this is a growth condition which says that there exists an L > 0 such
that f 2 (x, t) + g 2 (x, t) < L(1 + x2 ) for all x ∈ [a, b]. If these two conditions are satisfied for all t ∈ [0, T ],
then there is a unique solution on this time interval.
Now suppose B(a, t) = 0, so there is no diffusion
√ at the left endpoint. The left boundary is then said to
be prescribed. From the Lipschitz condition on B, this says that B(x, t) vanishes no slower than (x−a)2 ,
which says that ∂x B(a, t) = 0. Consider the above SDE with the condition B(a, t) = 0. We see that

(i) If A(a, t) > 0, a particle at a will enter the region [a, b] with probability one. This is
called an entrance boundary.
(ii) If A(a, t) < 0, a particle at a will exit the region [a, b] with probability one. This is called
an exit boundary.
1
See Gardiner, §5.1.2.
2
I.e. for well-behaved functions which you would take home to meet your mother.
3
See L. Arnold, Stochastic Differential Equations (Dover, 2012).
4
One can choose convenient dimensionless units for all quantities.
4 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

(iii) If A(a, t) = 0, a particle at a remain fixed with probability one. This is called a natural
boundary.

Mutatis mutandis, similar considerations hold at x = b, where A(b, t) > 0 for an exit and A(b, t) < 0 for
an entrance.

Stationary solutions

We now look for stationary solutions P (x, t) = Peq (x). We assume A(x, t) = A(x) and B(x, t) = B(x).
Then
1 d 
J = A(x) Peq (x) − B(x) Peq (x) = constant . (4.9)
2 dx
Define the function
( Zx )
A(x 0)
ψ(x) = exp 2 dx0 , (4.10)
B(x0 )
a

so ψ 0 (x) = 2ψ(x)A(x)/B(x). Then

d B(x) Peq (x)

 
2J
=− , (4.11)
dx ψ(x) ψ(x)

with solution
Zx
B(a) ψ(x) 2J ψ(x) dx0
Peq (x) = · · Peq (a) − . (4.12)
B(x) ψ(a) B(x) ψ(x0 )
a

Note ψ(a) = 1. We now consider two different boundary conditions.

• Zero current : In this case J = 0 and we have

B(a) ψ(x)
Peq (x) = · · P (a) . (4.13)
B(x) ψ(a) eq

Rb
The unknown quantity P (a) is then determined by normalization: dx Peq (x) = 1.
a

• Periodic boundary conditions : Here we invoke P (a) = P (b), which requires a specific value for J,

, Z b
Peq (a) B(a) B(b) dx0

J= − . (4.14)
2 ψ(a) ψ(b) ψ(x0 )
a

This leaves one remaining unknown, Peq (a), which again is determined by normalization.
4.2. FOKKER-PLANCK EQUATION 5

Examples

We conclude this section with two examples. The first is diffusion in a gravitational field, for which the
Langevin equation takes the form
√
dx = −vD dt + 2D dW (t) , (4.15)
where the drift velocity is vD = g/γ, with γ the frictional damping constant (Ffr = −γM ẋ) and g the
acceleration due to gravity. Thus, the Fokker-Planck equation is ∂t P = vD ∂x P + D ∂x2 P , whence the
solution with a reflecting (J = 0) condition at x = 0 is
vD

Peq (x) = exp −vD x/D , (4.16)
D
where we have normalized P (x) on the interval x ∈ [0, +∞). This steady state distribution reflects the
fact that particles tend to fall to the bottom. If we apply instead periodic boundary conditions at x = 0
and x = L, the solution is a constant P (x) = P (0) = P (L). In this case the particles fall through the
bottom x = 0 only to return at the top x = L and keep falling, like in the game Portal 5 .
Our second example is that of the Ornstein-Uhlenbeck process, described by ∂t P = ∂x (βxP ) + D ∂x2 P .
The steady state solution is
Peq (x) = Peq (0) exp −βx2 /2D .


(4.17)
p
This is normalizable over the real line x ∈ (−∞, ∞), in which case Peq (0) = β/2πD . On a finite
interval x ∈ [−a, a] with periodic boundary conditions, we have
2 −x2 )/2D
Peq (x) = Peq (a) eβ(a . (4.18)

4.2.4 Eigenfunction expansions for Fokker-Planck

We saw in §4.2.1 how the (forward) Fokker-Planck equation could be written as

∂P (x, t) ∂ 1 ∂2
= L P (x, t) , L=− A(x) + B(x) , (4.19)
∂t ∂x 2 ∂x2
and how the stationary state solution Peq (x) satisfies J = A Peq − 21 ∂x (B Peq ). Consider the operator

∂ 1 ∂2
Le = +A(x) + B(x) 2 , (4.20)
∂x 2 ∂x
where, relative to L, the sign of the leading term is reversed. It is straightforward to show that, for any
functions f and g,
h ib
f Le g − g L f = g J − f Kg , f (4.21)
a
where
Zb
g L f = dx g(x) L f (x) , (4.22)
a
5
The cake is a lie.
6 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

and Jf = Af − 12 (Bf )0 and Kg = − 21 Bg 0 . Thus we conclude that Le = L† , the adjoint of L, if either (i)
Jf and Kg vanish at the boundaries x = a and x = b (reflecting conditions), or (ii) the functions f and g
vanish at the boundaries (absorbing conditions).
We can use the zero current steady state distribution Peq (x) , for which J = APeq − 12 ∂x (BPeq ) = 0 , to
convert between solutions of the forward and backward time Fokker-Planck equations. Suppose P (x, t)
satisfies ∂t P = LP . Then define Q(x, t) ≡ P (x, t)/Peq (x), in which case

∂t P = Peq ∂t Q = −∂x (APeq Q) + 21 ∂x2 (BPeq Q)

n o n o
= − ∂x (APeq ) + 12 ∂x2 (BPeq ) Q + − A ∂x Q + 21 B ∂x2 Q Peq + ∂x (BPeq ) ∂x Q (4.23)
n o
= A ∂x Q + 12 B ∂x2 Q Peq ,

where we have used ∂x (BPeq ) = 2APeq . Thus, we have that Q(x, t) satisfies ∂t Q = LQ.
e We saw in §4.2.1
how the (forward) Fokker-Planck equation could be written as

∂Q(x, t) ∂ 1 ∂2
= L† Q(x, t) , L† = A(x) + B(x) 2 , (4.24)
∂t ∂x 2 ∂x
which is the backward Fokker-Planck equation when written in terms of the time variable s = −t.
Now let us seek eigenfunctions Pn (x) and Qn (x) which satisfy6

LPn (x) = −λn Pn (x) , L† Qn (x) = −λn Qn (x) . (4.25)

where now A(x, t) = A(x) and B(x, t) = B(x) are assumed to be time-independent. If the functions
Pn (x) and Qn (x) form complete sets, then a solution to the Fokker-Planck equations for P (x, t) and
Q(x, t) is of the form7
X X
P (x, t) = Cn Pn (x) e−λn t , Q(x, t) = Cn Qn (x) e−λn t . (4.26)
n n

To elicit the linear algebraic structure here, we invoke Eqn. 4.25 and write

(λm − λn ) Qm (x) Pn (x) = Qm (x) L Pn (x) − Pn (x) L† Qm (x) . (4.27)

Next we integrate over the interval [a, b], which gives

Zb h ib
(λm − λn ) dx Qm (x) Pn (x) = Qm (x) Jn (x) − Km (x) Pn (x) = 0 , (4.28)
a
a

where Jn (x) = A(x) Pn (x) − 21 ∂x B(x) Pn (x) and Km (x) = − 21 B(x) ∂x Qm (x). For absorbing boundary
 

conditions, the functions Pn (x) and Qn (x) vanish at x = a and x = b, so the RHS above vanishes. For
6 ∂
In the eigensystem, the partial differential operators ∂x in L and L† may be regarded as ordinary differential operators d
dx
.
7
Since Pn (x) = Peq (x) Qn (x), the same expansion coefficients {Cn } appear in both sums.
4.2. FOKKER-PLANCK EQUATION 7

reflecting boundaries, it is the currents Jn and Km (x) which vanish at the boundaries. Thus we have
(λm − λn ) Qm Pn = 0, where the inner product is

Zb
Q P ≡ dx Q(x) P (x) . (4.29)
a

We obtain the familiar result from Sturm-Liouville theory that when the eigenvalues differ, the cor-
responding eigenfunctions are orthogonal. In the case of eigenvalue degeneracy, we can invoke the
Gram-Schmidt procedure, in which case we may adopt the general normalization

Zb Zb Zb
P (x) Pn (x)
Qm Pn = dx Qm (x) Pn (x) = dx Peq (x) Qm (x) Qn (x) = dx m = δmn . (4.30)
Peq (x)
a a a

A general solution to the Fokker-Planck equation with reflecting boundaries may now be written as
X
P (x, t) = Cn Pn (x) e−λn t , (4.31)
n

where the expansion coefficients {Cn } are given by

Zb
Cn = dx Qn (x) P (x, 0) = Qn P (0) . (4.32)
a

Suppose our initial condition is P (x, 0 | x0 , 0) = δ(x − x0 ). Then Cn = Qn (x0 ) , and

X
P (x, t | x0 , 0) = Qn (x0 ) Pn (x) e−λn t . (4.33)
n

We may now take averages, such as

Zb X
Qn (x0 ) Pn (x) e−λn t .


F x(t) = dx F (x) (4.34)
a n

Furthermore, if we also average over x0 = x(0), assuming is is distributed according to Peq (x0 ), we have
the correlator
Zb Zb
x(t) x(0) = dx0 dx x x0 P (x, t | x0 , 0) Peq (x0 )
a a
" Zb #2 (4.35)
2 −λn t
X X
= dx x Pn (x) e−λn t = x Pn e .
n a n
8 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

Absorbing boundaries

At an absorbing boundary x = a , one has P (a) = Q(a) = 0. We may still use the function Peq (x) obtained
from the J = 0 reflecting boundary conditions to convert between forward and backward Fokker-Planck
equation solutions.
Next we consider some simple examples of the eigenfunction formalism.

Heat equation

We consider the simplest possible Fokker-Planck equation,

∂P ∂ 2P
=D 2 , (4.36)
∂t ∂x

which is of course the one-dimensional diffusion equation. We choose our interval to be x ∈ [0, L].

• Reflecting boundaries : The normalized steady state solution is simply Peq (x) = 1/L. The eigen-
functions are P0 (x) = Peq (x) and
√
√
   
2 nπx nπx
Pn (x) = cos , Qn (x) = 2 cos (4.37)
L L L

for n > 0. The eigenvalues are λn = D (nπ/L)2 . We then have

∞    
1 2X nπx0 nπx −λn t
P (x, t | x0 , 0) = + cos cos e . (4.38)
L L L L
n=1

Note that as t → ∞ one has P (x, ∞ | x0 , 0) = 1/L , which says that P (x, t) relaxes to Peq (x). Both
boundaries are natural boundaries, which prevent probability flux from entering or leaking out of
the region [0, L].

• Absorbing boundaries : Now we have

√
√
   
2 nπx nπx
Pn (x) = sin , Qn (x) = 2 sin (4.39)
L L L

and
∞    
2X nπx0 nπx −λn t
P (x, t | x0 , 0) = sin sin e , (4.40)
L L L
n=1

again with λn = D (nπ/L)2 . Since λn > 0 for all allowed n, we have P (x, ∞ | x0 , 0) = 0, and all the
probability leaks out by diffusion. The current is J(x) = −DP 0 (x), which does not vanish at the
boundaries.
4.2. FOKKER-PLANCK EQUATION 9

• Mixed boundaries : Now suppose x = 0 is an absorbing boundary and x = L a reflecting bound-

ary. Then
√
√
   
2 (2n + 1)πx (2n + 1)πx
Pn (x) = sin , Qn (x) = 2 sin (4.41)
L 2L 2L

2
with n ≥ 0. The eigenvalues are λn = D (n + 12 ) π/L .

√
We can write the eigenfunctions in all three cases in the form Pn (x) = L2 sin(kn x+δ), where kn = nπx/L
or (n + 21 )πx/L and δ = 0 or δ = 21 π, with λn = D kn2 . One then has

1

2L reflecting, n = 0
− √8/Lk 2
δ



√ n n,odd reflecting, n > 0
x Pn = n+1 2/k
(4.42)
 (−1) n absorbing, n > 0
(−1)n+1 √2/Lk 2



n half reflecting, half absorbing, n > 0 .

Note that when a zero mode λmin = 0 is part of the spectrum, one has P0 (x) = Peq (x), to which P (x, t)
relaxes in the t → ∞ limit. When one or both of the boundaries is absorbing, the lowest eigenvalue
λmin > 0 is finite, hence P (x, t → ∞) → 0, i.e. all the probability eventually leaks out of the interval.

Ornstein-Uhlenbeck process

The Fokker-Planck equation for the OU process is ∂t P = ∂x (βxP ) + D∂x2 P . Over the real line x ∈ R, the
normalized steady state distribution is Peq (x) = (β/2πD)1/2 exp(−βx2 /2D). The eigenvalue equation
for Qn (x) is
d2 Qn dQn
D − βx = −λn Qn (x) . (4.43)
dx2 dx
Changing variables to ξ = x/`, where ` = (2D/β)1/2 , we obtain Q00n − 2ξQ0n + (2λn /β) Qn = 0, which is
Hermite’s equation. The eigenvalues are λn = nβ, and the normalized eigenfunctions are then
1

Qn (x) = √ Hn x/`
2n n!
(4.44)
1 2 2
Hn x/` e−x /` ,


Pn (x) = √
n
2 n! π`2
√

which satisfy the orthonormality relation hQm |Pn i = δmn . Since H1 (ξ) = 2ξ , one has hx|Pn i = `/ 2 δn,1 ,
hence the correlator is given by x(t) x(0) = 12 `2 e−βt .

4.2.5 First passage problems

Suppose we have a particle on an interval x ∈ [a, b] with absorbing boundary conditions, which means
that particles are removed as soon as they get to x = a or x = b and not replaced. Following Gardiner8 ,
8
See Gardiner §5.5.
10 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

define the quantity

Zb
G(x, t) = dx0 P (x0 , t | x, 0) . (4.45)
a

Thus, G(x, t) is the probability that x(t) ∈ [a, b] given that x(0) = x. Since the boundary conditions
are absorbing, there is no reentrance into the region, which means that G(x, t) is strictly decreasing as a
function of time, and that
∂G(x, t)
− dt = probability, starting from x at t = 0, to exit [a, b] during time interval [t, t + dt] . (4.46)
∂t
If we assume the process is autonomous, then

Zb
G(x, t) = dx0 P (x0 , 0 | x, −t) , (4.47)
a

which satisfies the backward Fokker-Planck equation,

∂G ∂G 1 ∂ 2G
=A + 2B = L† G . (4.48)
∂t ∂x ∂x2

We may average functions of the exit time t according to

Z∞  
∂G(x, t)
f (t) x
= dt f (t) − . (4.49)
∂t
0

In particular, the mean exit time T (x) is given by

Z∞   Z∞
∂G(x, t)
T (x) = htix = dt t − = dt G(x, t) . (4.50)
∂t
0 0

From the Fokker-Planck equation for G(x, t), the mean exit time T (x) satisfies the ODE

1 d2 T dT
B(x) 2 + A(x) = −1 . (4.51)
2 dx dx
∂2
This is derived by applying the operator L† = 1 ∂
2 B(x) ∂x2
+ A(x) ∂x to the above expression for T (x).
R∞
Acting on the integrand G(x, t), this produces ∂G
∂t , according to Eq. 4.48, hence dt∂t G(x, t) = G(x, ∞)−
0
G(x, 0) = −1.
To solve Eqn. 4.51, we once again invoke the services of the function
 x 
Z 0
0 2A(x )

ψ1 (x) = exp dx , (4.52)
 B(x0 ) 
a
4.2. FOKKER-PLANCK EQUATION 11

which satisfies ψ10 (x)/ψ1 (x) = 2A(x)/B(x). Thus, we may reexpress eqn. 4.51 as
ψ10 0 2
0 2 ψ1
T 00 + T =− ⇒ ψ1 T 0 =− . (4.53)
ψ1 B B
We may integrate this to obtain
T 0 (a) ψ (x)
T 0 (x) = − 2 , (4.54)
ψ1 (x) ψ1 (x)
where we have defined
Zx
ψ (x0 )
ψ2 (x) = 2 dx0 1 0 . (4.55)
B(x )
a
Note that ψ1 (a) = 1 and ψ2 (a) = 0. We now integrate one last time to obtain
T (x) = T (a) + T 0 (a) ψ3 (x) − ψ4 (x) , (4.56)
where
Zx Zx
dx0 ψ (x0 )
ψ3 (x) = , ψ4 (x) = dx0 2 0 . (4.57)
ψ1 (x0 ) ψ1 (x )
a a
Note that ψ3 (a) = ψ4 (a) = 0
Eqn. 4.56 involves two constants of integration, T (a) and T 0 (a), which are to be determined by imposing
two boundary conditions. If the boundaries are absorbing, we have T (a) = T (b) = 0, which entail
T 0 (a) = ψ4 (b)/ψ3 (b). The final result for the mean exit time is then
ψ3 (x) ψ4 (b) − ψ3 (b) ψ4 (x)
T (x) = . (4.58)
ψ3 (b)

As an example, consider the case of pure diffusion: A(x) = 0 and B(x) = 2D. Then
ψ1 (x) = 1 , ψ2 (x) = (x − a)/D , ψ3 (x) = (x − a) , ψ4 (x) = (x − a)2 /2D , (4.59)
whence
(x − a)(b − x)
T (x) = . (4.60)
2D
A particle starting in the middle x = 12 (a + b) at time t = 0 will then exit the region in an average time
(b − a)2 /8D.

One absorbing, one reflecting boundary

Suppose the boundary at a is now reflecting, while that at b remains absorbing. We then have the
boundary conditions ∂x G(a, t) = 0 and G(b, t) = 0, which entail T 0 (a) = 0 and T (b) = 0. Then the
general result of Eqn. 4.56 then gives T (x) = T (a) − ψ4 (x). Requiring T (b) = 0 then yields the result
Zb Zy
dy ψ (z)
T (x) = T (b) − ψ4 (x) = 2 dz 1 (x = a reflecting , x = b absorbing) . (4.61)
ψ1 (y) B(z)
x a
12 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

Under the opposite condition, where the boundary at a is absorbing while that at b is reflecting, we have
T (a) = 0 and T 0 (b) = 0. Eqn. 4.56 then gives T (x) = T 0 (a) ψ3 (x) − ψ4 (x) , and imposing T 0 (b) = 0 entails
T 0 (a) = ψ2 (b), hence

Zx Zb
dy ψ (z)
T (x) = ψ2 (b) ψ3 (x) − ψ4 (x) = 2 dz 1 (x = a absorbing , x = b reflecting) . (4.62)
ψ1 (y) B(z)
a y

Escape through either boundary

Define the quantities

Z∞ Z∞  h i
0 0 0 0 1 0
Ga (x, t) = − dt J(a, t | x, 0) = dt − A(a) P (a, t | x, 0) + 2 ∂a B(a) P (a, t | x, 0)
t t
(4.63)
Z∞ Z∞  h i
0 0 0 0 1 0
Gb (x, t) = + dt J(b, t | x, 0) = dt + A(b) P (b, t | x, 0) − 2 ∂b B(b) P (b, t | x, 0) .
t t

Since −J(a, t | x, 0) is the left-moving probability flux at x = a , Ga (x, t) represents the probability that a
particle starting at x ∈ [a, b] exits at a sometime after a time t. The second expression for Gb (x, t) yields
the probability that a particle starting at x exits at b sometime after t. Their sum is

Z∞ Zb  h i
0 0 0 0 0 1 0 0
Ga (x, t) + Gb (x, t) = dt dx ∂x0 A(x ) P (x , t | x, 0) − 2 ∂x0 B(x ) P (x , t | x, 0)
t a
(4.64)
Z∞ Zb h i Zb
= dt dx − ∂t0 P (x , t | x, 0) = dx0 P (x0 , t | x, 0) = G(x, t) ,
0 0 0 0

t a a

which is the total probability starting from x to exit the region after t.
Since P (a, t0 | x, 0) satisfies the backward Fokker-Planck equation, i.e. L† P (a, t0 | x, 0) = ∂t0 P (a, t0 | x, 0),
we have

L† Ga (x, t) = J(a, t | x, 0) = +∂t Ga (x, t)

L† Gb (x, t) = J(b, t | x, 0) = −∂t Gb (x, t) . (4.65)

Now let us evaluate the above equations in the limit t → 0. Since P (x0 , 0 | x, 0) = δ(x − x0 ), there can
only be an infinitesimal particle current at any finite distance from the initial point x at an infinitesimal
value of the elapsed time t. Therefore we have
( )
† ∂ 1 ∂2
L Gc (x, 0) = A(x) + B(x) 2 Gc (x, 0) = 0 . (4.66)
∂x 2 ∂x
4.2. FOKKER-PLANCK EQUATION 13

Thus, Gc (x, 0) is the total probability for exit via c ∈ {a, b} over all time, conditioned at starting at x at
time 0. The boundary conditions here are

Ga (a, 0) = 1 , Ga (b, 0) = 0 ; Gb (b, 0) = 1 , Gb (a, 0) = 0 , (4.67)

which says that a particle starting at a is immediately removed with probability unity and therefore can
Rx
never exit through b, and vice versa. Solving using the function ψ1 (x) = exp dx 2A(x0 )/B(x0 ) , we have
a

Zb , Zb
Ga (x, 0) = dy ψ1 (y) dz ψ1 (x)
x a
(4.68)
Zx , Zb
Gb (x, 0) = dy ψ1 (y) dz ψ1 (x) .
a a

Note Ga (x, 0) + Gb (x, 0) = 1, which says that eventually the particle exits via either a or b. We next
define
Z∞
G (x, t)
Tc (x) = dt c , (4.69)
Gc (x, 0)
0

which is the mean exit time through c, given that the particle did exit through that boundary. This then
satisfies
h i
L† Gc (x, 0) Tc (x) = −Gc (x, 0) . (4.70)

For pure diffusion, A(x) = 0 and B(x) = 2D, and we found ψ1 (x) = 1. Therefore

b−x x−a
Ga (x, 0) = , Gb (x, 0) = . (4.71)
b−a b−a

We may then solve the equations

d2 h 
D G c (x, 0) Tc (x) = −Gc (x, 0) (4.72)
dx2

to obtain
(x − a)(2b − x − a) (b − x)(b + x − 2a)
Ta (x) = , Tb (x) = . (4.73)
6D 6D
Note that
(x − a)(b − x)
Ga (x, 0) Ta (x) + Gb (x, 0) Tb (x) = = T (x) , (4.74)
2D
which we found previously in Eqn. 4.60.
14 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

4.2.6 Escape from a metastable potential minimum

In the presence of a local potential U (x), the local drift velocity is −U 0 (x)/γm, where m is the particle’s
mass and γ its frictional damping (Ffr = −γmẋ). An example potential U (x) is depicted in Fig. 4.1.
Gardiner in §5.5.3 begins with the equation

∂ U 0 (x) ∂ 2P
 
∂P
= P +D 2 , (4.75)
∂t ∂x γm ∂x

which resembles a Fokker-Planck equation for P (x, t) with drift vD (x) = −U 0 (x)/γm. However, Eqn.
4.75 is not a Fokker-Planck equation but rather something called the Smoluchowski equation. Recall that
the position x(t) of a Brownian particle does not execute a Markov process. So where does Eqn. 4.75 come
from, and under what conditions is it valid?
It is the two-component phase space vector ϕ = (x, v) which executes a Markov process, and for whose
conditional probability density we can derive a Fokker-Planck equation, and not the position x alone.
The Brownian motion problem may be written as two coupled first order differential equations,

dx = v dt
" #
1 0 √ (4.76)
dv = − U (x) + γv dt + Γ dW (t) ,
m

where Γ = 2γkB T /m = 2γ 2 D, and where W (t) is a Wiener process. The first of these is an ODE and the
second an SDE. Viewed as a multicomponent SDE, the Fokker-Planck equation for P (x, v, t) is
"  #
∂P ∂ ∂ U 0 (x) γk T ∂ 2P
=− vP ) + + γv P + B . (4.77)
∂t ∂x ∂v m m ∂v 2

Suppose though that the damping γ is large. Then we can approximate the second equation in 4.76 by
assuming v rapidly relaxes, which is to say dv ≈ 0. Then we have
1 √
v dt ≈ − U 0 (x) dt + 2D dW (t) (4.78)
γm
and replacing v dt in the first equation with this expression we obtain the SDE
√
dx = vD (x) dt + 2D dW (t) , (4.79)

which immediately yields the Smoluchowski equation 4.75. This procedure is tantamount to an adia-
batic elimination of the fast variable. It is valid only in the limit of large damping γ = 6πηa/m , which is
to say large fluid viscosity η.
Taking the Smoluchowski equation as our point of departure, the steady state distribution is then found
to be
Peq (x) = C e−U (x)/kB T , (4.80)
where we invoke the result D = kB T /γm from §2.2.2. We now consider the first passage time T (x | x0 ) for
a particle starting at x = x0 escaping to a point x ≈ x∗ in the vicinity of the local potential maximum. We
4.2. FOKKER-PLANCK EQUATION 15

Figure 4.1: Escape from a metastable potential minimum.

apply the result of our previous analysis, with (a, b, x) in Eqn. 4.61 replaced by (−∞, x, x0 ), respectively,
and x > ∗ 0
∼ x . Note that A(x) = −U (x)/γm, and B(x) = 2D, hence
Zx
2A(x0 ) U (a) − U (x)
ln ψ1 (x) = dx0 0
= . (4.81)
B(x ) kB T
a

Formally we may have U (a) = ∞, but it drops out of the expression for the mean exit time,
Zx Zy Zx Zy
1 dy 1
T (x | x0 ) = dz ψ1 (z) = dy e U (y)/kB T
dz e−U (z)/kB T . (4.82)
D ψ1 (y) D
x0 −∞ x0 −∞

The above integrals can be approximated as follows. Expand U (x) about the local extrema at x0 and x∗
as
U (x0 + δx) = U (x0 ) + 21 K0 (δx)2 + . . .
(4.83)
U (x∗ + δx) = U (x∗ ) − 12 K ∗ (δx)2 + . . . ,
where K0 = U 00 (x0 ) and K ∗ = −U 00 (x∗ ). At low temperatures, integrand e−U (z)/kB T is dominated by the
region z ≈ x0 , hence
Zy
2πkB T 1/2 −U (x0 )/k T
 
−U (z)/kB T
dz e ≈ e B . (4.84)
K0
−∞

Similarly, the integrand eU (y)/kB T is dominated by the region y ≈ x∗ , so for x somewhere between x∗
and x1 , we may write9
Zx
2πkB T 1/2 U (x∗ )/k T
 
U (y)/kB T
dy e ≈ e B . (4.85)
K∗
x0

9
We take x > x∗ to lie somewhere on the downslope of the potential curve, on the other side of the barrier from the
metastable minimum.
16 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

We then have
U (x∗ ) − U (x0 )
 
2πkB T
T (x1 | x0 ) ≈ p exp . (4.86)
D K0 K ∗ kB T
Known as the Arrhenius law, this is one of the most ubiquitous results in nonequilibrium statistical
physics, with abundant consequences for chemistry, biology, and many other fields of science. With
∆E = U (x∗ ) − U (x0 ), the energy necessary to surmount the barrier, the escape rate is seen to be propor-
tional to exp(−∆E/kB T ).

4.2.7 Detailed balance

Let ϕ denote a coordinate vector in phase space. In classical mechanics, ϕ = (q, p) consists of all the
generalized coordinates and generalized momenta. The condition of detailed balance says that each in-
dividual transition balances precisely with its time reverse, resulting in no net probability currents in
equilibrium. Note that this is a much stronger condition than conservation of probability.
In terms of joint probability densities, detailed balance may be stated as follows:
T T
P (ϕ, t ; ϕ0 , t0 ) = P (ϕ0 , −t0 ; ϕT , −t) = P (ϕ0 , t ; ϕT , t0 ) , (4.87)

where we have assumed time translation invariance. Here, ϕT is the time reverse of ϕ. This is accom-
plished by multiplying each component ϕi by a quantity εi = ±1. For positions ε = +1, while for
momenta ε = −1. If we define the diagonal matrix εij = εi δij (no sum on i), then ϕTi = εij ϕj (implied
sum on j). Thus we may rewrite the above equation as

P (ϕ, t ; ϕ0 , t0 ) = P (ε ϕ0 , t ; ε ϕ, t0 ) . (4.88)

In terms of the conditional probability distributions, we have

P (ϕ, t | ϕ0 , 0) Peq (ϕ0 ) = P (εϕ0 , t | εϕ, 0) Peq (εϕ) , (4.89)

where Peq (ϕ) is the equilibrium distribution, which we assume holds at time t0 = 0. Now in the limit
t → 0 we have P (ϕ, t → 0 | ϕ0 , 0) = δ(ϕ − ϕ0 ), and we therefore conclude

Peq (εϕ) = Peq (ϕ) . (4.90)

The equilibrium distribution Peq (ϕ) is time-reversal invariant. Thus, detailed balance entails

P (ϕ, t | ϕ0 , 0) Peq (ϕ0 ) = P (εϕ0 , t | εϕ, 0) Peq (ϕ) . (4.91)

One then has

Z
ϕi = dϕ Peq (ϕ) ϕi = εi ϕi
Z Z (4.92)
Gij (t) ≡ ϕi (t) ϕj (0) = dϕ dϕ0 ϕi ϕ0j P (ϕ, t | ϕ0 , 0) Peq (ϕ0 ) = εi εi Gji (t) .

Thus, as a matrix, G(t) = ε Gt (t) ε.

4.2. FOKKER-PLANCK EQUATION 17

The conditions under which detailed balance holds are10

W (ϕ | ϕ0 ) Peq (ϕ0 ) = W (εϕ0 | εϕ) Peq (ϕ)
h i ∂ h i
Ai (ϕ) + εi Ai (εϕ) Peq (ϕ) = Bij (ϕ) Peq (ϕ) (4.93)
∂ϕj
εi εj Bij (εϕ) = Bij (ϕ) (no sum on i and j) .

Detailed balance for the Fokker-Planck equation

It is useful to define the reversible and irreversible drift as

1h i
Ri (ϕ) ≡ Ai (ϕ) + εi Ai (εϕ)
2 (4.94)
1h i
Ii (ϕ) ≡ Ai (ϕ) − εi Ai (εϕ) .
2
Then we may subtract
∂i εi Ai (εϕ) Peq (ϕ) − 12 ∂i ∂j εi εj Bij (εϕ) Peq (ϕ)
   

from
∂i Ai (ϕ) Peq (ϕ) − 12 ∂i ∂j Bij (ϕ) Peq (ϕ)
   

to obtain ( )
X ∂ h i X ∂Ii (ϕ) ∂ ln Peq (ϕ)
Ii (ϕ) Peq (ϕ) = 0 ⇒ + Ii (ϕ) =0. (4.95)
∂ϕi ∂ϕi ∂ϕi
i i

We may now write the second of Eqn. 4.93 as

Ri (ϕ) = 12 ∂j Bij (ϕ) + 21 Bij (ϕ) ∂j ln Peq (ϕ) , (4.96)

or, assuming the matrix B is invertible,

−1
Ri − 12 ∂j Bij ≡ Zk (ϕ) .


∂k ln Peq (ϕ) = 2 Bki (4.97)

Since the LHS above is a gradient, the condition that Peq (ϕ) exists is tantamount to

∂Zi ∂Zj
= (4.98)
∂ϕj ∂ϕi

for all i and j. If this is the case, then we have

Zϕ
Peq (ϕ) = exp dϕ0 · Z (ϕ0 ) . (4.99)

Because of the condition 4.98, the integral on the RHS may be taken along any path. The constant
associated with the undetermined lower limit of integration is set by overall normalization.
10
See Gardiner, §6.3.5.
18 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

Brownian motion in a local potential

Recall that the Brownian motion problem may be written as two coupled first order differential equa-
tions,

dx = v dt
" #
1 0 √ (4.100)
dv = − U (x) + γv dt + Γ dW (t) ,
m

where Γ = 2γkB T /m = 2γ 2 D, and where W (t) is a Wiener process. The first of these is an ODE and the
second an SDE. Viewed as a multicomponent SDE with
  ! !
x v 0 0
ϕ= , Ai (ϕ) = 0 , Bij (ϕ) = 2γkB T . (4.101)
v − U m(x) − γv 0 m

We have already derived in Eqn. 4.77 the associated Fokker-Planck equation for P (x, v, t).
The time reversal eigenvalues are ε1 = +1 for x and ε2 = −1 for v. We then have
  !
0 v
R(ϕ) = , I (ϕ) = 0 . (4.102)
−γv − U m(x)

As the B matrix is not invertible, we appeal to Eqn. 4.96. The upper component vanishes, and the lower
component yields
γk T ∂ ln Peq
−γv = B , (4.103)
m ∂v
which says Peq (x, v) = F (x) exp(−mv 2 /2kB T ). To find F (x), we use Eqn. 4.95, which says

0 0
z}|{ z}|{
∂I ∂I ∂ ln Peq ∂ ln Peq
0 = 1 + 2 +I1 + I2
∂x ∂v ∂x ∂v
(4.104)
U 0 (x)
 
∂ ln F mv
=v − − ⇒ F (x) = C e−U (x)/kB T .
∂x m kB T

Thus,
2 /2k T
Peq (x, v) = C e−mv B e−U (x)/kB T . (4.105)

4.2.8 Multicomponent Ornstein-Uhlenbeck process

In §3.4.3 we considered the case of coupled SDEs,

dϕi = Ai (ϕ) dt + βij (ϕ) dWj (t) , (4.106)

4.2. FOKKER-PLANCK EQUATION 19

where Wi (t) Wj (t0 ) = δij min(t, t0 ). We showed in §3.4.3 that such a multicomponent SDE leads to the
Fokker-Planck equation
∂P ∂
1 ∂2

=− Ai P + Bij P , (4.107)
∂t ∂ϕi 2 ∂ϕi ∂ϕj
where B = ββ t , i.e. Bij = k βik βjk .
P

Now consider such a process with

Ai (ϕ) = Aij ϕj , Bij (ϕ) = Bij , (4.108)

where Aij and Bij are independent of ϕ . The detailed balance conditions are written as εBε = B, and


A + εA ε ϕ = B ∇ ln Peq (ϕ) . (4.109)

This equation says that Peq (ϕ) must be a Gaussian, which we write as
h i
Peq (ϕ) = Peq (0) exp − 21 ϕi Mij−1 ϕj , (4.110)

Obviously we can take M −1 to be symmetric, since any antisymmetric part of M −1 is projected out
in the expression ϕi Mij−1 ϕj . Thus M is also symmetric. Substituting this solution into the stationary
Fokker-Planck equation ∂i Aij ϕj Peq = 12 ∂i ∂j Bij Peq yields
 

h i
1 −1 −1 1 −1 −1


Tr A + 2 Tr BM = ϕi M A + 2 M BM ϕj = 0 . (4.111)
ij

This must be satisfied for all ϕ, hence both the LHS and RHS of this equation must vanish separately.
This entails
A + M At M −1 + BM −1 = 0 . (4.112)

We now invoke the detailed balance condition of Eqn. 4.109, which says

A + εA ε + BM −1 = 0 . (4.113)

Combining this with our previous result, we conclude

εAM = (AM )t ε , (4.114)

which are known as the Onsager conditions. If we define the phenomenological force

F = ∇ ln Peq = −M −1 ϕ , (4.115)

then we have
dhϕi
= A hϕi = −AM F , (4.116)
dt
and defining L = −AM which relates the fluxes J = hϕ̇i to the forces F , viz. Ji = Lik Fk , we have
the celebrated Onsager relations, εLε = Lt . A more general formulation, allowing for the presence of a
magnetic field, is
Lik (B ) = εi εk Lki (−B ) . (4.117)
We shall meet up with the Onsager relations again when we study the Boltzmann equation.
20 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

Figure 4.2: Electrical circuit containing a fluctuating voltage source Vs (t) and a fluctuating current source
Is (t).

4.2.9 Nyquist’s theorem

Consider the electrical circuit in Fig. 4.2. Kirchoff’s laws say that the current flowing through the resistor
r is IS − IB , and that
Q dI
(IS − IB ) r = = VS − L A − RIA (4.118)
C dt
and
dQ
= IA + IB . (4.119)
dt
Thus, we have the coupled ODEs for Q and IA ,

dQ Q
= IA − + IS (t)
dt rC
(4.120)
dIA RI Q V (t)
=− A − + S .
dt L LC L

If we assume VS (t) and IS (t) are fluctuating sources each described by a Wiener process, we may write
p p
VS (t) dt = ΓV dWV (t) , IS (t) dt = ΓI dWI (t) . (4.121)

Then
 
Q p
dQ = −
+ IA dt + ΓI dWI (t)
rC
  (4.122)
Q R IA 1p
dIA = − + dt + ΓV dWV (t) .
LC L L
4.3. MASTER EQUATION 21

We now see that Eqn. 4.122 describes a two component Ornstein-Uhlenbeck process, with ϕt = (Q, IA ),
and    
1/rC −1 ΓI 0
Aij = − , Bij = . (4.123)
1/LC R/L 0 ΓV /L2
 
1 0
The ε matrix for this problem is ε = since charge is even and current odd under time reversal.
0 −1
Thus,  
2/rC 0
A + εAε = − = −BM −1 , (4.124)
0 2R/L
from which we may obtain M −1 and then
 
ΓI rC/2 0
M= . (4.125)
0 ΓV /2LR

The equilibrium distribution is then

( )
2
Q2 RL IA
Peq (Q, IA ) = N exp − − . (4.126)
rCΓI ΓV

We now demand that equipartition hold, i.e.

 2  2 
Q LIA
= = 12 kB T , (4.127)
2C 2
which fixes
ΓV = 2RkB T , ΓI = 2kB T /r . (4.128)

Therefore, the current and voltage fluctuations are given by

2kB T
VS (0) VS (t) = 2kB T R δ(t) , IS (0) IS (t) = δ(t) , VS (0) IS (t) = 0 . (4.129)
r

4.3 Master Equation

In §2.6.3 we showed that the differential Chapman-Kolmogorov equation with only jump processes
yielded the Master equation,
∂P (x, t | x0 , t0 )
Z h i
0 0 0 0
= dy W (x | y , t) P (y , t | x , t ) − W (y | x, t) P (x, t | x , t ) . (4.130)
∂t
Here W (x | y , t) is the rate density of transitions from y to x at time t, and has dimensions T −1 L−d . On
a discrete state space, we have
∂P (n, t | n0 , t0 ) X h i
= W (n | m, t) P (m, t | n0 , t0 ) − W (m | n, t) P (n, t | n0 , t0 ) , (4.131)
∂t m

where W (n | m, t) is the rate of transitions from m to n at time t, with dimensions T −1 .

22 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

4.3.1 Birth-death processes

The simplest case is that of one variable n, which represents the number of individuals in a population.
Thus n ≥ 0 and P (n, t | n0 , t0 ) = 0 if n < 0 or n0 < 0. If we assume that births and deaths happen
individually and at with a time-independent rate, then we may write

W (n | m, t) = t+ (m) δn,m+1 + t− (m) δn,m−1 . (4.132)

Here t+ (m) is the rate for m → m + 1, and t− (m) is the rate for m → m − 1. We require t− (0) = 0, since
the dying rate for an entirely dead population must be zero11 . We then have the Master equation

∂P (n, t | n0 , t0 ) h i
= t+ (n−1) P (n−1, t | n0 , t0 )+t− (n+1) P (n+1, t | n0 , t0 )− t+ (n)+t− (n) P (n, t | n0 , t0 ) .
∂t
(4.133)
This may be written in the form

∂P (n, t | n0 , t0 )
+ ∆J(n, t | n0 , t0 ) = 0 , (4.134)
∂t
where the lattice current operator on the link (n, n + 1) is

J(n, t | n0 , t0 ) = t+ (n) P (n, t | n0 , t0 ) − t− (n + 1) P (n + 1, t | n0 , t0 ) . (4.135)

The lattice derivative ∆ is defined by

∆f (n) = f (n) − f (n − 1) , (4.136)

for any lattice function f (n). One then has

∞
dhnit X  +
t (n) − t− (n) P (n, t | n0 , t0 ) = t+ (n) − t− (n) .

= t
(4.137)
dt
n=0

Steady state solution

We now seek a steady state solution Peq (n), as we did in the case of the Fokker-Planck equation. This
entails ∆n J(n) = 0, where we suppress the initial conditions (n0 , t0 ). Now J(−1) = 0 because t− (0) = 0
and P (−1) = 0, hence 0 = J(0) − J(−1) entails J(0) = 0, and since 0 = ∆n J(n) we have J(n) = 0 for all
n ≥ 0. Therefore
t+ (j)
Peq (j + 1) = − P (j) , (4.138)
t (j + 1) eq
which means
n +
Y t (j − 1)
Peq (n) = Peq (0) . (4.139)
t− (j)
j=1
11
We neglect here the important possibility of zombies.
4.3. MASTER EQUATION 23

4.3.2 Examples: reaction kinetics

First example

Consider the example in Gardiner §11.1.2, which is the reaction

k2
X A. (4.140)
k1

We assume the concentration [A] = a is fixed, and denote the number of X reactants to be n. The rates
are t− (n) = k2 n and t+ (n) = k1 a, hence we have the Master equation


∂t P (n, t) = k2 (n + 1) P (n + 1, t) + k1 a P (n − 1, t) − k2 n + k1 a P (n, t) , (4.141)

with P (−1, t) ≡ 0. We solve this using the generating function formalism, defining
∞
X
P̃ (z, t) = z n P (n, t) . (4.142)
n=0

Note that P̃ (1, t) = ∞ n

P
n=0 P (n, t) = 1 by normalization. Multiplying both sides of Eqn. 4.141 by z and
then summing from n = 0 to n = ∞, we obtain

zP (z,t) P (z,t) ∂z P (z,t) z∂z P (z,t)

z }| { z }| { z }| { z }| {
X∞ X∞ X∞ X∞
∂t P̃ (z, t) = k1 a P (n − 1, t) z n − k1 a P (n, t) z n + k2 (n + 1) P (n + 1, t) z n − k2 n P (n, t) z n
n=0 n=0 n=0 n=0
n o
= (z − 1) k1 a P̃ (z, t) − k2 ∂z P̃ (z, t) .
(4.143)

We now define the function Q̃(z, t) via

P̃ (z, t) = ek1 az/k2 Q̃(z, t) , (4.144)

so that
∂t Q̃ + k2 (z − 1) ∂z Q̃ = 0 , (4.145)
and defining w = − ln(1 − z), this is recast as ∂t Q̃ − k2 ∂w Q̃ = 0, whose solution is

Q̃(z, t) = F (w + k2 t) , (4.146)

where F is an arbitrary function of its argument. To determine the function F (w), we invoke our initial
conditions,
Q̃(z, 0) = e−k1 az/k2 P̃ (z, 0) = F (w) . (4.147)
We then have  
k1 a
(1 − e ) P̃ 1 − e−w , 0 ,
−w


F (w) = exp − (4.148)
k2
24 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

and hence  
k a
P̃ (z, t) = exp − 1 (1 − z)(1 − e−k2 t ) P̃ 1 − (1 − z) e−k2 t , 0 .


(4.149)
k2
We may then obtain P (n, t) via contour integration, i.e. by extracting the coefficient of z n in the above
expression: I
1 dz
P (n, t) = P̃ (z, t) . (4.150)
2πi z n+1
|z|=1

Note that setting t = 0 in Eqn. 4.149 yields the identity P̃ (z, 0) = P̃ (z, 0). As t → ∞, we have the steady
state result
λn −λ
P̃ (z, ∞) = ek1 a(z−1)/k2 ⇒ P (n, ∞) = e , (4.151)
n!
where λ = k1 a/k2 , which is a Poisson distribution. Indeed, suppose we start at t = 0 with the Poisson
distribution P (n, 0) = e−α0 α0n /n!. Then P̃ (z, 0) = exp α0 (z − 1) , and Eqn. 4.149 gives
 
 
k1 a −k2 t
n o
P̃ (z, t) = exp − (1 − z)(1 − e ) exp − α0 (1 − z) e−k2 t = eα(t) (z−1) , (4.152)
k2
where
k1
α(t) = α0 e−k2 t + a 1 − e−k2 t .


(4.153)
k2
Thus, α(0) = α0 and α(∞) = k1 a/k2 = λ. The distribution is Poisson all along, with a time evolv-
ing Poisson parameter α(t). The situation is somewhat reminiscent of the case of updating conjugate
Bayesian priors, where the prior distribution was matched with the likelihood function so that the up-
dated prior retains the same functional form.
If we start instead with P (n, 0) = δn,n , then we have P̃ (z, 0) = z n0 , and
0
( )
k1 a  n0
P̃ (z, t) = exp − (1 − z)(1 − e−k2 t ) 1 − (1 − z) e−k2 t . (4.154)
k2

We then have
∂ P̃ (z, t) k1 a
1 − e−k2 t + n0 e−k2 t


n(t) = =
∂z z=1 k2
!
2
∂ P̃ (z, t) ∂ P̃ (z, t)
2
n (t) = + = hn(t)i2 + hn(t)i − n0 e−2k2 t (4.155)
∂z 2 ∂z
z=1
 
k1 a
+ n0 e−k2 t 1 − e−k2 t .
 

Var n(t) =
k2

Second example

Gardiner next considers the reactions

k2 k3
X A , B + 2X 3X , (4.156)
k1 k4
4.3. MASTER EQUATION 25

Figure 4.3: Geometric interpretation of the ODE in Eqn. 4.160.

for which we have

t+ (n) = k1 a + k3 b n(n − 1)
(4.157)
t− (n) = k2 n + k4 n(n − 1)(n − 2) .

The reason here is that for the second equation to proceed to the left, we need to select three X molecules
to take part in the reaction, and there are n(n − 1)(n − 2) ordered triples (i, j, k). Now Eqn. 4.137 gives

dhni
= k1 a + k3 n(n − 1) − k2 hni − k4 n(n − 1)(n − 2) . (4.158)
dt

For a Poisson distribution Pn = e−λ λn /n! , it is easy to see that

k
n(n − 1) · · · (n − k + 1) = n (Poisson) . (4.159)

Suppose the distribution P (n, t) is Poissonian for all t. This is not necessarily the case, but we assume it
to be so for the purposes of approximation. Then the above equation closes, and with x = hni, we have

dx
= −k4 x3 + k3 x2 − k2 x + k1 a
dt (4.160)
= −k4 (x − x1 )(x − x2 )(x − x3 ) ,

where x1,2,3 are the three roots of the cubic on the RHS of the top equation. Since the coefficients of this
equation are real numbers, the roots are either real or come in complex conjugate pairs. We know that
the product of the roots is x1 x2 x3 = k1 a/k4 and that the sum is x1 + x2 + x3 = k3 /k4 , both of which are
positive. Clearly when x is real and negative, all terms in the cubic are of the same sign, hence there can
be no real roots with x < 0. We assume three real positive roots with x1 < x2 < x3 .
Further examining Eqn. 4.160, we see that x1 and x3 are stable fixed points and that x2 is an unstable
fixed point of this one-dimensional dynamical system. Thus, there are two possible stable equilibria. If
26 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

x(0) < x2 the flow will be toward x1 , while if x(0) > x2 the flow will be toward x3 . We can integrate
Eqn. 4.160 using the method of partial fractions. First, we write
1 A1 A2 A3
= + + , (4.161)
(x − x1 )(x − x2 )(x − x3 ) x − x1 x − x2 x − x3
with (x − x2 )(x − x3 )A1 + (x − x1 )(x − x3 )A2 + (x − x1 )(x − x2 )A3 = 1. This requires
0 = A1 + A2 + A3
0 = (x2 + x3 )A1 + (x1 + x3 )A2 + (x1 + x2 )A3 (4.162)
1 = x2 x3 A1 + x1 x3 A2 + x1 x2 A3 ,
with solution
1 1 1
A1 = , A2 = − , A3 = . (4.163)
(x2 − x1 )(x3 − x1 ) (x2 − x1 )(x3 − x2 ) (x3 − x1 )(x3 − x2 )
Thus, Eqn. 4.160 may be recast as
(x3 − x2 ) d ln(x − x1 ) − (x3 − x1 ) d ln(x − x2 ) + (x2 − x1 ) d ln(x − x3 ) = −k4 (x2 − x1 )(x3 − x1 )(x3 − x2 ) dt .
(4.164)
The solution is given in terms of t(x):
 
1 x0 − x1
t(x) = ln (4.165)
k4 (x2 − x1 )(x3 − x1 ) x − x1
   
1 x0 − x2 1 x0 − x2
− ln + ln ,
k4 (x2 − x1 )(x3 − x2 ) x − x2 k4 (x3 − x1 )(x3 − x2 ) x − x3
where x0 = x(0).
Going back to Eqn. 4.139, we have that the steady state distribution is
n + n
Y t (j − 1) Y k1 a + k3 b (j − 1) (j − 2)
Peq (n) = Peq (0) −
= Peq (0) . (4.166)
t (j) k2 j + k4 j (j − 1) (j − 2)
j=1 j=1

The product is maximized for when the last term with j = n is unity. If we call this value n∗ , then n∗ is
a root of the equation
k1 a + k3 b (n − 1) (n − 2) = k2 n + k4 n (n − 1) (n − 2) . (4.167)
If n  1 and all the terms are roughly the same size, this equation becomes k1 a + k3 b n2
= k2 n + k4 n3 ,
which is the same as setting the RHS of Eqn. 4.160 to zero in order to find a stationary solution.

4.3.3 Forward and reverse equations and boundary conditions

In §2.6.3 we discussed the forward and backward differential Chapman-Kolmogorov equations, from
which, with Aµ = 0 and Bµν = 0 , we obtain the forward and reverse Master equations,
∂P (n, t | · ) X n o
= W (n | m, t) P (m, t | · ) − W (m | n, t) P (n, t | · )
∂t m
(4.168)
∂P ( · | n, t) X n o
− = W (m | n, t) P ( · | m, t) − P ( · | n, t) ,
∂t m
4.3. MASTER EQUATION 27

where we have suppressed the initial conditions in the forward equation and the final conditions in the
backward equation. Consider the one-dimensional version, and take the transition rates to be

W (j 0 | j, t) = t+ (j) δj 0 ,j+1 + t− (j) δj 0 ,j−1 . (4.169)

We may then write

J(n−1 , t | · )
z }| o{
∂P (n, t | · ) n
+ −
= LP (n, t | · ) = t (n − 1) P (n − 1, t | · ) − t (n) P (n, t | · ) (4.170)
∂t
J(n , t | · )
zn }| o{
− t+ (n) P (n, t | · ) − t− (n + 1) P (n + 1, t | · )
K( · | n+1 , t) K( · | n , t)
z }| o{ zn }| o{
∂P ( · | n, t) +
n
−
− = LP ( · | n, t) = t (n) P ( · | n + 1, t) − P ( · | n, t) − t (n) P ( · | n, t) − P ( · | n − 1, t)
e ,
∂t
(4.171)

where we have defined the quantities J(n, t | · ) and K( · | n, t) . Here (Lf )n = Lnn0 fn0 and (Lf
e ) =
n
Lenn0 fn0 , where L and Le are matrices, viz.

Lnn0 = t+ (n0 ) δn0 ,n−1 + t− (n0 ) δn0 ,n+1 − t+ (n0 ) δn0 ,n − t− (n0 ) δn0 ,n
(4.172)
Lenn0 = t+ (n) δn0 ,n+1 + t− (n) δn0 ,n−1 − t+ (n) δn0 ,n − t− (n) δn0 ,n .

Clearly Lenn0 = Ln0 n , hence Le = Lt , the matrix transpose, if we can neglect boundary terms. For n, n0 ∈
Z , we could specify P (±∞, t | · ) = P ( · | ± ∞, t) = 0 .
Consider now a birth-death process where we focus on a finite interval n ∈ {a, . . . , b}. Define the inner
product
Xb


hg|O|f i = g(n) Of (n) . (4.173)
n=a
One then has
h g | L | f i − h f | Le | g i = t− (b + 1) f (b + 1) g(b) − t+ (b) f (b) g(b + 1)
(4.174)
+ t+ (a − 1) f (a − 1) g(a) − t− (a) f (a) g(a − 1) .

Thus, if f (a−1) = g(a−1) = f (b+1) = g(b+1) = 0, we have Le = Lt = L† , the adjoint. In the suppressed
initial and final conditions, we always assume the particle coordinate n lies within the interval.
We now must specify appropriate boundary conditions on our interval. These conditions depend on
whether we are invoking the forward or backward Master equation:

• Forward equation : For reflecting boundaries, we set t− (a) = 0 and t+ (b) = 0, assuring that a
particle starting from inside the region can never exit. We also specify P (a − 1, t | · ) = 0 and
28 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

P (b + 1, t | · ) = 0 so that no particles can enter from the outside. This is equivalent to specifying
that the boundary currents vanish, i.e. J(a − 1, t | · ) = 0 and J(b, t | · ) = 0, respectively. For
absorbing boundaries, we choose t+ (a − 1) = 0 and t− (b + 1) = 0 , which assures that a particle
which exits the region can never reenter. This is equivalent to demanding P (a − 1, t | · ) = 0 and
P (b + 1, t | · ) = 0, respectively.
• Backward equation : From Eqn. 4.171, it is clear that the reflecting conditions t− (a) = 0 and
t+ (b) = 0 are equivalent to K( · | a, t) = 0 and K( · | b + 1, t) = 0, where these functions. Neither of
the quantities in the absorbing conditions t+ (a − 1) = 0 and t− (b + 1) = 0 enter in the backward
Master equation. The effect of these conditions on the data outside the interval is to preserve
P ( · | a − 1, t) = 0 and P ( · | b + 1, t) = 0, respectively.
The situation is summarized in Tab. 4.3.3 below.

conditions equivalent conditions

equation boundary reflecting absorbing reflecting absorbing

FORWARD left t− (a) = 0 t+ (a − 1) = 0 J(a − 1, t | · ) = 0 P (a − 1, t | · )

right t+ (b) = 0 t− (b + 1) = 0 J(b, t | · ) = 0 P (b + 1, t | · )

BACKWARD left t− (a) = 0 t+ (a − 1) = 0 K( · | a, t) = 0 P ( · | a − 1, t)

right t+ (b) = 0 t− (b + 1) = 0 K( · | b + 1, t) = 0 P ( · | b + 1, t)

Table 4.1: Absorbing and reflecting boundary conditions for the Master equation on the interval
{a, . . . , b}.

4.3.4 First passage times

The treatment of first passage times within the Master equation follows that for the Fokker-Planck equa-
tion in §4.2.5. If our discrete particle starts at n at time t0 = 0, the probability that it lies within the
interval {a, . . . , b} at some later time t is
b
X b
X
0
G(n, t) = P (n , t | n, 0) = P (n0 , 0 | n, −t) , (4.175)
n0 =a n0 =a

and therefore −∂t G(n, t) dt is the probability that the particle exits the interval within the time interval
[t, t + dt]. Therefore the average first passage time out of the interval, starting at n at time t0 = 0, is
Z∞   Z∞
∂G(n, t)
T (n) = dt t − = dt G(n, t) . (4.176)
∂t
0 0
4.3. MASTER EQUATION 29

Applying L,
e we obtain
n o n o
Le T (n) = t+ (n) T (n + 1) − T (n) − t− (n) T (n) − T (n − 1) = −1 . (4.177)

Let a be a reflecting barrier and b be absorbing. Since t− (a) = 0 we are free to set T (a − 1) = T (a). At the
right boundary we have T (b + 1) = 0, because a particle starting at b + 1 is already outside the interval.
Eqn. 4.177 may be written
t+ (n) ∆T (n) − t− (n) ∆T (n − 1) = −1 , (4.178)
with ∆T (n) ≡ T (n + 1) − T (n). Now define the function
n
Y t− (j)
φ(n) = , (4.179)
t+ (j)
j=a+1

with φ(a) ≡ 1. This satisfies φ(n)/φ(n − 1) = t− (n)/t+ (n) , and therefore Eqn. 4.178 may be recast as

∆T (n) ∆T (n − 1) 1
= − + . (4.180)
φ(n) φ(n − 1) t (n) φ(n)

Since ∆T (a) = −1/t+ (a) from Eqn. 4.177, the first term on the RHS above vanishes for n = a. We then
have
n
X 1
∆T (n) = −φ(n) , (4.181)
t+ (j) φ(j)
j=a

and therefore, working backward from T (b + 1) = 0, we have

b k
X X 1
T (n) = φ(k) (a reflecting , b absorbing). (4.182)
t+ (j) φ(j)
k=n j=a

One may also derive

n b
X X 1
T (n) = φ(k) (a absorbing , b reflecting). (4.183)
t+ (j) φ(j)
k=a j=k

Example

Suppose a = 0 is reflecting and b = N − 1 is absorbing, and furthermore suppose that t± (n) = t± are
site-independent. Then φ(n) = r−n , where r ≡ t+ /t− . The mean escape time starting from site n is

N −1 k
1 X −k X j
T (n) = + r r
t
k=n j=0 (4.184)
1 n
−N −n
o
= (N − n)(r − 1) + r − r .
(r − 1)2 t+
30 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

If t+ = t− , so the walk is unbiased, then r = 1. We can then evaluate by taking r = 1 + ε with ε → 0, or,
more easily, by evaluating the sum in the first line when r = 1. The result is

1 n1 1
o
T (n) = N (N − 1) − n(n + 1) + N − n (r = 1) . (4.185)
t+ 2 2

By taking an appropriate limit, we can compare with the Fokker-Planck result of Eqn. 4.61, which for
an interval [a, b] with a = 0 reflecting and b absorbing yields T (x) = (b2 − x2 )/2D. Consider the Master
equation,

∂P (n, t) h i ∂ 2P 1 ∂ 4P
= β P (n + 1, t) + P (n − 1, t) − 2P (n, t) = β + 12 β + ... , (4.186)
∂t ∂n2 ∂n4
where β = t+ = t− . Now define n ≡ N x/b, and rescale both time t ≡ N τ and hopping β ≡ N γ ,
resulting in
∂P ∂ 2P Db2 ∂ 4P
=D 2 + + ... , (4.187)
∂τ ∂x 12N 2 ∂x4
where D = b2 γ is the diffusion constant. In the continuum limit, N → ∞ and we may drop all terms
beyond the first on the RHS, yielding the familiar diffusion equation. Taking this limit, Eqn. 4.185 may
be rewritten as T (x)/N = (N/2t+ b2 )(b2 − x2 ) = (b2 − x2 )/2D , which agrees with the result of Eqn. 4.61.

4.3.5 From Master equation to Fokker-Planck

Let us start with the Master equation,

Z
∂P (x, t) h i
= dx0 W (x | x0 ) P (x0 , t) − W (x0 | x) P (x, t) , (4.188)
∂t

and define W (z | z0 ) ≡ t(z − z0 | z0 ), which rewrites the rate W (z | z0 ) from z0 to z as a function of z0

and the distance z − z0 to z . Then the Master equation may be rewritten as
Z
∂P (x, t) h i
= dy t(y | x − y ) P (x − y , t) − t(y | x) P (x, t) . (4.189)
∂t

Now expand t(y | x − y ) P (x − y ) as a power series in the jump distance y to obtain12

∞
(−1)n ∂n
Z
∂P (x, t) X h i
= dy yα1· · · yαn t(y | x) P (x, t)
∂t n! ∂xα · · · ∂xαn
n=1 1
∞ (4.190)
X (−1)n ∂n h i
= Rα1···αn (x) P (x, t) ,
n! ∂xα · · · ∂xαn
n=1 1

where Z
Rα1···αn (x) = dy yα1· · · yαn t(y | x) . (4.191)

12
We only expand the second argument of t(y | x − y ) in y . We retain the full y -dependence of the first argument.
4.3. MASTER EQUATION 31

For d = 1 dimension, we may write

∞
∂P (x, t) X (−1)n ∂ n h i Z
= R (x) P (x, t) , Rn (x) ≡ dy y n t(y | x) . (4.192)
∂t n! ∂xn n
n=1

This is known as the Kramers-Moyal expansion. If we truncate at order n = 2, we obtain the Fokker-
Planck equation,
∂P (x, t) ∂ h i 1 ∂2 h i
=− R1 (x) P (x, t) + R (x) P (x, t) . (4.193)
∂t ∂x 2 ∂x2 2

The problem is that the FPE here is akin to a Procrustean bed. We have amputated the n > 2 terms from
the expansion without any justification at all, and we have no reason to expect this will end well. A more
systematic approach was devised by N. G. van Kampen, and goes by the name of the size expansion. One
assumes that there is a large quantity lurking about, which we call Ω. Typically this can be the total
system volume, or the total population in the case of an ecological or epidemiological model. One
assumes that t(y | x) obeys a scaling form,
 
z0
t(∆z | z0 ) = Ω τ ∆z . (4.194)
Ω

From the second of Eqn. 4.192, we then have

Z  
n x
Rn (x) = Ω dy y τ y ≡ΩR
e (x/Ω)
n . (4.195)
Ω

We now proceed by defining √

x = Ω φ(t) + Ωξ , (4.196)
where φ(t) is an as-yet undetermined function of time, and ξ is to replace x, so that our independent
variables are now (ξ, t). We therefore have
en φ(t) + Ω −1/2 ξ


Rn (x) = Ω R . (4.197)

Now we are set to derive a systematic

√ expansion in inverse
√ powers of Ω . We define P (x, t) = Π(ξ, t),
and we note that dx = Ω φ̇ dt + Ω dξ, hence dξ x = − Ω φ̇ dt , which means

∂P (x, t) ∂Π(ξ, t) √ ∂Π(ξ, t)

= − Ω φ̇ . (4.198)
∂t ∂t ∂ξ
We therefore have, from Eqn. 4.192,
∞
∂Π(ξ, t) √ ∂Π X (−1)n Ω (2−n)/2 ∂ n h i
en φ(t) + Ω −1/2 ξ Π(ξ, t) .


− Ω φ̇ = R (4.199)
∂t ∂ξ n! ∂ξ n
n=1

e (φ + Ω −1/2 ξ) in powers of Ω −1/2 , we obtain

Further expanding Rn
∞ ∞
∂Π(ξ, t) √ ∂Π X X (−1)n Ω (2−n−k)/2 dkR
en (φ) ∂n h k i
− Ω φ̇ = ξ Π(ξ, t) . (4.200)
∂t ∂ξ n! k! dφk φ(t) ∂ξ
n
k=0 n=1
32 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

Let’s define an index l ≡ n + k, which runs from 1 to ∞. Clearly n = l − k , which for fixed l runs from 1
to l. In this way, we can reorder the terms in the sum, according to
∞ X
X ∞ ∞ X
X l
A(k, n) = A(l − n, n) . (4.201)
k=0 n=1 l=1 n=1

The lowest order term on the RHS of Eqn. 4.200 is the term with n = 1 and k = 0, corresponding to
√

l = n = 1 if we eliminate the k index in favor of l. It is equal to − Ω R
e φ(t) ∂ Π, hence if we demand
1 ξ
that φ(t) satisfy
dφ
=R e1 (φ) , (4.202)
dt
these terms cancel from either side of the equation. We then have
∞ l
∂Π(ξ, t) X (2−l)/2 X (−1)n e(l−n)
∂ n h l−n i
= Ω Rn φ(t) ξ Π(ξ, t) , (4.203)
∂t n! (l − n)! ∂ξ n
l=2 n=1

where Ren(k) (φ) = dkR

e /dφk . We are now in a position to send Ω → ∞ , in which case only the l = 2 term
n
survives, and we are left with
∂Π
2
e0 φ(t) ∂ (ξΠ) + 1 R
e φ(t) ∂ Π


= −R1 2 2 , (4.204)
∂t ∂ξ ∂ξ 2
which is a Fokker-Planck equation.

Birth-death processes

Consider a birth-death process in which the states | n i are labeled by nonnegative integers. Let αn
denote the rate of transitions from | n i → | n + 1 i and let βn denote the rate of transitions from | n i →
| n − 1 i. The Master equation then takes the form13
dPn

= αn−1 Pn−1 + βn+1 Pn+1 − αn + βn Pn , (4.205)
dt
where we abbreviate Pn (t) for P (n, t | n0 , t0 ) and suppress the initial conditions (n0 , t0 ).
Let us assume we can write αn = K ᾱ(n/K) and βn = K β̄(n/K), where K  1. Define x ≡ n/K, so the
Master equation becomes
∂P
= K ᾱ(x − K1 )P (x − K1 ) + K β̄(x + K1 )P (x + K1 ) − K ᾱ(x) + β̄(x) P (x)


∂t
(4.206)
∂ h i 1 ∂2 h i
+ O(K −2 ) .



=− ᾱ(x) − β̄(x) P (x, t) + ᾱ(x) + β̄(x) P (x, t)
∂x 2K ∂x2
If we truncate the expansion after the O(K −1 ) term, we obtain
∂P ∂   1 ∂2 h i
=− f (x) P (x, t) + g(x) P (x, t) , (4.207)
∂t ∂x 2K ∂x2
13
We further demand βn=0 = 0 and P−1 (t) = 0 at all times.
4.3. MASTER EQUATION 33

where we have defined

f (x) ≡ ᾱ(x) − β̄(x) , g(x) ≡ ᾱ(x) + β̄(x) . (4.208)

This FPE has an equilibrium solution

Zx
A −KΦ(x) f (x0 )
Peq (x) = e , Φ(x) = −2 dx0 , (4.209)
g(x) g(x0 )
0

where the constant A is determined by normalization. If K is large, we may expand about the minimum
of Φ(x)

2f (x∗ ) 2f (x∗ ) g 0 (x∗ ) − 2g(x∗ ) f 0 (x∗ )

Φ(x) = Φ(x∗ ) − (x − x ∗
) + (x − x∗ )2 + . . .
g(x∗ ) g 2 (x∗ )
(4.210)
2f 0 (x∗ )
= Φ(x∗ ) − (x − x∗ )2 + . . . ,
g(x∗ )

where f (x∗ ) ≡ 0 determines x∗ . Thus, we obtain a Gaussian distribution

r
K −K(x−x∗ )2 /2σ2 2 g(x∗ )
Peq (x) ' e with σ = − . (4.211)
2πσ 2 2f 0 (x∗ )

In order that the distribution be normalizable, we must have f 0 (x∗ ) < 0.

In §4.3.6, we will see how the Fokker-Planck expansion fails to account for the large O(K) fluctuations
about a metastable equilibrium which lead to rare extinction events in this sort of birth-death process.

van Kampen treatment

We now discuss the same birth-death process using van Kampen’s size expansion. Assume √ the distribu-
tion Pn (t) has a time-dependent maximum at n = √ Kφ(t) and a width proportional to K. We expand
relative to this maximum, writing n ≡ Kφ(t) + K ξ and we define Pn (t) ≡ Π(ξ, t). We now rewrite
the Master equation in eqn. 4.205 in terms of Π(ξ, t). Since n is an independent variable, we set
√ √
dn = K φ̇ dt + K dξ ⇒ dξ n
= − K φ̇ dt . (4.212)

Therefore
dPn √ ∂Π ∂Π
= − K φ̇ + . (4.213)
dt ∂ξ ∂t
We now write

αn−1 Pn−1 = K ᾱ φ + K −1/2 ξ − K −1 Π ξ − K −1/2

βn+1 Pn+1 = K β̄ φ + K −1/2 ξ + K −1 Π ξ + K −1/2

(4.214)
αn + βn Pn = K ᾱ φ + K −1/2 ξ Π(ξ) + K β̄ φ + K −1/2 ξ Π(ξ) ,




34 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

and therefore Eqn. 4.205 becomes

√ ∂Π ∂Π √ ∂Π ∂Π ∂ 2Π
+(β̄ 0 − ᾱ0 ) ξ +(β̄ 0 − ᾱ0 )Π + 21 (ᾱ+ β̄) 2 +O K −1/2


− K φ̇+ = K (β̄ − ᾱ) , (4.215)
∂ξ ∂t ∂ξ ∂ξ ∂ξ
√
where ᾱ = ᾱ(φ) and β̄ = β̄(φ). Equating terms of order K yields the equation

φ̇ = f (φ) ≡ ᾱ(φ) − β̄(φ) , (4.216)

which is a first order ODE for the quantity φ(t). Equating terms of order K 0 yields the Fokker-Planck
equation,
∂Π
∂
∂ 2Π
= −f 0 φ(t) ξΠ + 12 g φ(t)


, (4.217)
∂t ∂ξ ∂ξ 2
where g(φ) ≡ ᾱ(φ) + β̄(φ). If in the limit t → ∞, eqn. 4.216 evolves to a stable fixed point φ∗ , then the
stationary solution of the Fokker-Planck eqn. 4.217, Πeq (ξ) = Π(ξ, t = ∞) must satisfy

∂ ∂ 2Πeq 1 2 /2σ 2
−f 0 (φ∗ ) ξ Πeq + 21 g(φ∗ ) e−ξ


=0 ⇒ Πeq (ξ) = √ , (4.218)
∂ξ ∂ξ 2 2πσ 2
where
g(φ∗ )
σ2 = − . (4.219)
2f 0 (φ∗ )
Now both α and β are rates, hence both are positive and thus g(φ) > 0. We see that the condition σ 2 > 0 ,
which is necessary for a normalizable equilibrium distribution, requires f 0 (φ∗ ) < 0, which is saying that
the fixed point in Eqn. 4.216 is stable.
We thus arrive at the same distribution as in Eqn. 4.211. The virtue of this latter approach is that we
have a better picture of how the distribution evolves toward its equilibrium value. The condition of nor-
malizability f 0 (x∗ ) < 0 is now seen to be connected with the dynamics of location of the instantaneous
maximum of P (x, t), namely x = φ(t). If the dynamics of the FPE in Eqn. 4.217 are fast compared with
those of the simple dynamical system in Eqn. 4.216, we may regard the evolution of φ(t) as adiabatic so
far as Π(ξ, t) is concerned.

4.3.6 Extinction times in birth-death processes

In §4.3.1 we discussed the Master equation for birth-death processes,

dPn
= t+ (n − 1) Pn−1 + t− (n + 1) Pn+1 − t+ (n) + t− (n) Pn
 
. (4.220)
dt
P
At the mean field level, we have for the average population n̄ = n n Pn ,

dn̄
= t+ (n̄) − t− (n̄) . (4.221)
dt
Two models from population biology that merit our attention here:
4.3. MASTER EQUATION 35

• Susceptible-infected-susceptible (SIS) model : Consider a population of fixed total size N , among

which n individuals are infected and the remaining N − n are susceptible. The number of possible
contacts between infected and susceptible individuals is then n(N − n), and if the infection rate
per contact is Λ/N and the recovery rate of infected individuals is set to unity14 , then we have
 
n
+
t (n) = Λn 1 − , t− (n) = n . (4.222)
N

• Verhulst model : Here the birth rate is B and the death rate is unity plus a stabilizing term (B/N ) n
which increases linearly with population size. Thus,

Bn2
t+ (n) = Bn , t− (n) = n + . (4.223)
N

The mean field dynamics of both models is the same, with

dn̄ Λn̄2
= (Λ − 1)n̄ − (4.224)
dt N
for the SIS model; take Λ → B for the Verhulst model. This is known as the logistic equation: n̄˙ =
rn̄(K − n̄), with r = Λ/N the growth rate and K = (Λ − 1)/Λ the equilibrium population. If Λ > 1
then K > 0, in which case the fixed point at n̄ = 0 is unstable and the fixed point at n̄ = K is stable.
The asymptotic state is one of an equilibrium number K of infected individuals. At Λ = 1 there is a
transcritical bifurcation, and for 0 < Λ < 1 we have K < 0, and the unphysical fixed point at n̄ = K is
unstable, while the fixed point at n̄ = 0 is stable. The infection inexorably dies out. So the mean field
dynamics for Λ > 1 are a simple flow to the stable fixed point (SFP) at n̄ = K, and those for Λ < 1 are a
flow to the SFP at n̄ = 0. In both cases, the approach to the SFP takes a logarithmically infinite amount
of time.
Although the mean field solution for Λ > 1 asymptotically approaches an equilibrium number of in-
fected individuals K, the stochasticity in this problem means that there is a finite extinction time for the
infection. The extinction time is the first passage time to the state n = 0. Once the population of infected
individuals goes to zero, there is no way for new infections to spontaneously develop. The mean first
passage time was studied in §4.3.4. We have an absorbing boundary at n = 1 , since t+ (0) = 0, and a
reflecting boundary at n = N , since t+ (N ) = 0 , and Eqn. 4.182 gives the mean first passage time for
absorption as
n N
X X 1
T (n) = φ(k) , (4.225)
t+ (j) φ(j)
k=1 j=k

where15
k −
Y t (l)
φ(k) = . (4.226)
t+ (l)
l=1
14
That is, we measure time in units of the recovery time.
15
In §4.3.4, we defined φ(a) = 1 where a = 1 is the absorbing boundary here, whereas in Eqn. 4.226 we have φ(1) =
t+ (1)/t− (1). Since the mean first passage time T (n) does not change when all φ(n) are multiplied by the same constant, we
are free to define φ(a) any way we please. In this chapter it pleases me to define it as described.
36 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

The detailed analysis of T (n) is rather tedious, and is described in the appendices to C. Doering et al.,
Multiscale Model Simul. 3, 283 (2005). For our purposes, it suffices to consider the behavior of the
function φ(n). Let x ≡ n/N ∈ [0, 1]. Then with y ≡ j/N define

t+ (j)
ρ(y) ≡ = Λ(1 − y) , (4.227)
t− (j)

in which case, using the trapezoidal rule, and setting x ≡ n/N ,

n
X
− ln φ(n) = ln ρ(l/N )
l=1
Zx
(4.228)
≈ − 21 ln ρ(0) − 12 ln ρ(x) + N du ln ρ(u)
0
n o
= N ln Λ −(1 − x) ln Λ − (1 − x) ln(1 − x) − x − ln Λ − 21 ln(1 − x) .

In the N → ∞ limit, the maximum occurs at x∗ = (Λ − 1)/Λ, which for Λ > 1 is the scaled mean field
equilibrium population of infected individuals. For x ≈ x∗ , the mean extinction time for the infection is
therefore
T (x∗ ) ∼ eN Φ(Λ) , Φ(Λ) = ln Λ − 1 + Λ−1 . (4.229)
The full result, from Doering et al., is
r
Λ 2π N (ln Λ−1+Λ−1 )  
T (x∗ ) = e × 1 + O(N −1 ) (4.230)
(Λ − 1)2 N

The extinction time is exponentially large in the population size.

Below threshold, when Λ < 1, Doering et al. find

ln(N x)
T (x) = + O(1) , (4.231)
1−Λ

which is logarithmic in N . From the mean field dynamics n̄˙ = (Λ − 1)n̄ − Λn̄2 , if we are sufficiently
close to the SFP at n̄ = 0 , we can neglect the nonlinear term, in which case the solution becomes n̄(t) =
n̄(0) e(Λ−1)t . If we set n̄(T ) ≡ 1 and n̄(0) = N x , we obtain T (x) = ln(N x)/(1 − Λ) , in agreement with
the above expression.

Fokker-Planck solution

Another approach to this problem is to map the Master equation onto a Fokker-Planck equation, as we
did in §4.3.5. The corresponding FPE is

∂P ∂
1 ∂2

=− fP + 2
gP , (4.232)
∂t ∂x 2N ∂x
4.3. MASTER EQUATION 37

where
f (x) = (Λ − 1) x − Λx2 = Λ x (x∗ − x)
(4.233)
g(x) = (Λ + 1) x − Λx2 = Λ x (x∗ + 2Λ−1 − x) .

The mean extinction time, from Eqn. 4.63, is

Zx Z1
dy ψ(z)
T (x) = 2N dz , (4.234)
ψ(y) g(z)
0 y

where ( Zx )
f (y)
ψ(x) = exp 2N dy ≡ e2N σ(x) (4.235)
g(y)
0
and  ∗
x + 2Λ−1 − x

−1
σ(x) = x + 2Λ ln . (4.236)
x∗ + 2Λ−1
Thus,
Zx Z1
2N e2N σ(z) e−2N σ(y)
T (x) = dy dz . (4.237)
Λ z(x∗ + 2Λ−1 − z)
0 y

The z integral is dominated by z ≈ x∗ ,

and the y integral by y ≈ 0. Computing the derivatives for the
Taylor series,  
∗ Λ−1 2 Λ+1
σ(x ) = − ln , σ 0 (x∗ ) = 0 , σ 00 (x∗ ) = − 21 Λ (4.238)
Λ Λ 2
and also σ(0) = 0 and σ 0 (0) = (Λ − 1)/(Λ + 1). One then finds
r
∗ Λ 2π 2N σ(x∗ )
T (x ) ≈ 2
e . (4.239)
(Λ − 1) NΛ

Comparison of Master and Fokker-Planck equation predictions for extinction times

How does the FPE result compare with the earlier analysis of the extinction time from the Master equa-
tion? If we expand about the threshold value Λ = 1 , writing Λ = 1 + ε , we find

Φ(Λ) = ln Λ − 1 + Λ−1 = 1
2 ε2 − 23 ε3 + 34 ε4 − 45 ε5 + . . .
  (4.240)
∗ 2(Λ − 1) 4 Λ+1 1 2 2 3 35 4 181 5
2σ(x ) = − ln = 2 ε − ε +
3 48 ε − 240 ε + ...
Λ Λ 2
The difference only begins at fourth order in ε viz.
 4
11 ε5 11 ε6

ME ∗ FPE ∗ ε
ln T (x ) − ln T (x ) = N − + + . . . + O(1) , (4.241)
48 240 160
38 CHAPTER 4. THE FOKKER-PLANCK AND MASTER EQUATIONS

where the superscripts indicate Master equation (ME) and Fokker-Planck equation (FPE), respectively.
While the term inside the parentheses impressively small when ε  1, it is nevertheless finite, and,
critically, it is multiplied by N . Thus, the actual mean extinction time, as computed from the original
Master equation, is exponentially larger than the Fokker-Planck result.
What are we to learn from this? The origin of the difference lies in the truncations we had to do in order
to derive the Fokker-Planck equation itself. The FPE fails to accurately capture the statistics of large
deviations from the metastable state. D. Kessler and N. Shnerb, in J. Stat. Phys. 127, 861 (2007), show
that the FPE is only valid for fluctuations about the metastable state whose size is O(N 2/3 ) , whereas to
reach the absorbing state requires a fluctuation of O(N ) . As these authors put it, ”In order to get the correct
statistics for rare and extreme events one should base the estimate on the exact Master equation that describes the
stochastic process. . . ”. They also derive a real space WKB method to extract the correct statistics from
the Master equation. Another WKB-like treatment, and one which utilizes the powerful Doi-Peliti field
theory formalism, is found in the paper by V. Elgart and A. Kamenev, Phys. Rev. E 70, 041106 (2004).