Stochastic Processes
Stochastic Processes
Stochastic Processes
i
9781107035904ar 2014/1/6 20:35 page 147 #161
9
Poisson processes and waiting times
i
i
i
9781107035904ar 2014/1/6 20:35 page 148 #162
148
as the basis for a widespread stochastic integration technique, called Markov chain Monte
Carlo.
For the sake of clarity we consider only a single random variable X that might, however,
depend on an L-component vector of independent random variables. For fixed , X ()
is an ordinary random variable, for which the CDF and PDF can be defined as usual:
F (x) := P (X () x),
d
F (x).
p (x) :=
dx
(9.1a)
(9.1b)
The really new aspects of stochastic processes concern the dependence on the independent
variable , which, as said before, typically represents time, space or energy.
Stochastic processes are referred to as continuous (discrete) stochastic processes if they
are based on continuous (discrete) input variables . That doesnot, however, imply that
X () as a function of is necessarily continuous (discrete). As a counterexample consider
a (discrete) Bernoulli experiment, whose outcome o {o1 , o2 } defines the global behaviour
cos(t)
for o = o1
Xt (o) :=
exp(t) for o = o2 .
This is a case of a discrete stochastic process which is continuous in t. The opposite situation is given by the following process:
Xt ( ) := 1(t ),
with being a (continuous) exponential random variable. Apparently, Xt is discontinuous
if considered as a function of t, but the process is referred to as continuous.
Interesting aspects of stochastic processes are the -dependence of individual realizations (samples)
X := X ( )
for a fixed , or the -dependence of the mean, i.e.
E( ) = X := dx x p (x)
L
dx x p (x|)
= d p()
(xX ())
=
d L X () p().
In the first line we have introduced the independent random variable via the marginalization rule and in the second we have utilized the fact that x is uniquely determined if and
are specified. Another interesting aspect of stochastic processes is the autocorrelation
in , i.e.
i
i
i
9781107035904ar 2014/1/6 20:35 page 149 #163
A(, ) := X X =
149
d L X () X () p() X X .
The autocorrelation plays a crucial role in MCMC, as outlined in Section 30.3 [p. 544]. As a
simple introductory example we consider a one-dimensional random walk on a lattice, with
sites x Z. In addition, we assume a parabolic potential, in order to avoid an increasing
variance, which we found for the unrestricted random walk in Section 4.2 [p. 57]. Let Xt be
the position of the walker at discrete time t and in each time step the walker has only three
choices: stay or move one step forward or backward. That is, Xt+1 = Xt + X. A possible
choice for the probability P (X|Xt , I) that the walker will be at site Xt+1 in the next time
step can be defined via the MetropolisHastings algorithm (see Section 30.3 [p. 544]),
resulting in
'
(
1
exp 2 Xt , 1 ,
P (X = 1|Xt , I) = min
2
1
10
0.8
A(t)
5
Xt
0
0.4
0.2
5
0
(a)
0.6
0.2
0.4
0.6
t /T
0.8
1
(b)
20
40
60
80
100
Figure 9.1 (a) Three individual random walks, all starting at X0 = 0, are depicted as jagged lines in
different shades of grey. The time is expressed in units of the total time T = 1000. In addition, the
sample mean (thick solid line) and variance (dashed line) of N = 1000 independent random walks
are displayed. (b) The corresponding scaled autocorrelation function A(t) versus lag size t.
i
i
i
9781107035904ar 2014/1/6 20:35 page 150 #164
150
A(t)
:=
t1
1
Xt+t Xt ,
t1 t0 + 1 t=t
0
Xt := Xt Xt .
The autocorrelation is determined in the time window ranging from t0 = 100 to t1 = 900.
The first 100 steps have been discarded, in order to enter the region where Xt+t Xt is
independent of t, then the process is said to be homogeneous. There are a couple more
subtleties to be accounted for in computing autocorrelations numerically, which are, however, immaterial for the basic understanding of this topic. In Figure 9.1 the scaled autocor
N
.
L
(9.2)
x
.
L
i
i
i
9781107035904ar 2014/1/6 20:35 page 151 #165
151
The probability that n PPs land in Ix is therefore binomial. The mean number of points
generated on Ix is
= p N = x .
(9.3)
Next we erase these PPs and start afresh, but this time we double both the length L and the
number of points N . We retain the interval Ix . As the point density is the same, the mean
number () of points in Ix is also the same. Upon repeating the doubling process we fulfil
the requirements of the Poisson theorem ( = pN = fixed and N ). Hence, in the
limit of N ,1 we obtain a Poisson distribution
P (n|x, , I) := ex
(x)n
.
n!
(9.4)
(9.5)
i
i
i
9781107035904ar 2014/1/6 20:35 page 152 #166
152
( dn )n1 e dn .
(n)
(9.6)
n
i .
i=1
The PDF implies that the probability for the next PP being at a distance is given by
p(d |, I) d = (1 d ) d.
This expression can be interpreted as the probability for no event (1 dx) in d times
the probability for one event d in the following infinitesimal interval d . In other words,
the events in infinitesimal intervals are uncorrelated and the probability Pd for an event in
d is analytic in d :
Pd = d.
As we will see in the next section, these features entail the Poisson distribution.
i
i