04 e PDF

Random processes - Chapter 4 Random process 1
Random processes
Chapter 4 Random process
4.1 Random process
4.1 Random process

Random process
Random process, stochastic process

The infinite set {Xt, t T } of random variables is called a random process, where
the index set T is an infinite set.
In other words, a random vector with an infinite number of elements is called a

random process.
Discrete time process, continuous time process

A random process is said to be discrete time if the index set is a countably infinite set.
When the index set is an uncountable set, the random process is called a continuous
time random process.
4.1 Random process

Discrete (alphabet) process, continuous (alphabet) process

A random process is called a discrete alphabet, discrete amplitude, or discrete state
process if all finite length random vectors drawn from the random process are discrete
random vectors. A process is called a continuous alphabet, continuous amplitude, or
continuous state process if all finite length random vectors drawn from the random
process are continuous random vectors.
A random process {X()} maps an element of the sample space on a time

function X(, t).
A random process {X()} is the collection of time functions X(t) called sample
functions. This collection is called an ensemble. The value of the time function
X(t) is a random variable at t.
4.1 Random process

A random process and sample functions
4.1 Random process

Since a random process is a collection of random variables with X(t) denoting a

random variable at t, the statistical characteristics of the random process can be
considered vi the cdf and pdf of X(t).
For example, the first-order cdf, first-order pdf, second-order cdf, and nth-order
cdf of the random process {X(t)} are
FX(t)(x) = Pr{X(t) x},
dFX(t)(x)
fX(t)(x) = ,
dx
FX(t1),X(t2)(x1, x2) = Pr{X(t1) x1, X(t2) x2},
and
FX(t1), ,X(tn)(x1, , xn) = Pr{X(t1) x1, , X(tn) xn},
respectively.
4.1 Random process

Mean function
The mean function mX (t) of a random process {X(t)} is defined by
mX (t) = Z
E{X(t)}

= xfX(t)(x)dx.

Autocorrelation function
The autocorrelation function RX (t1, t2) of a random process {X(t)} is defined by
RX (t1, t2) = E{X(t1)X (t2)}.
4.1 Random process

Known signal
An extreme example of a random process is a known or deterministic signal. When
X(t) = s(t) is a known signal, we have
m(t) = E{s(t)}
= s(t),
R(t1, t2) = E{s(t1)s(t2)}
= s(t1)s(t2).
Consider the random process {X(t)} with mean E{X(t)} = 3 and autocorrelation
function R(t1, t2) = 9 + 4 exp(0.2|t1 t2|). If Z = X(5), W = X(8), we can
easily obtain E{Z} = E{X(5)} = 3, E{W } = 3, E{Z 2} = R(5, 5) = 13,
E{W 2} = R(8, 8) = 13, Var{Z} = 13 32 = 4, Var{W } = 13 32 = 4, and
E{ZW } = R(5, 8) = 9 + 4e0.6 11.195. In other words, the random variables
Z and W have the variance 2 = 4 and covariance Cov(5, 8) = 4e0.6 2.195.
4.1 Random process

The autocorrelation function of X(t) = Aejt defined with a random

variable A can be obtained as
RX (t1, t2) = E{Aejt1 Aejt2 }
= ej(t1t2)E{|A|2}.
Autocovariance function
The autocovariance function KX (t1, t2) of a random process {X(t)} is defined by
KX (t1, t2) = E{[X(t1) mX (t1)][X (t2) mX (t2)]}.
In general, the autocovariance and autocorrelation functions are functions of t1

and t2.
The autocovariance function can be expressed in terms of the autocorrelation and
mean functions as
KX (t1, t2) = RX (t1, t2) mX (t1)mX (t2).
4.1 Random process

Uncorrelated random process

A random process {Xt} is said to be uncorrelated if RX (t, s) = E{Xt}E{Xs} or
KX (t, s) = 0 for t 6= s.
If a random process {X(t)} is uncorrelated, the autocorrelation and autocovari-

ance functions are
RX (t, s) = E{XtXs}
(
E{|Xt|2}, t = s,
=
E{Xt}E{Xs}, t =
6 s,
and
(
2
X t
, t = s,
KX (t, s) =
0, t 6= s,
respectively.
4.1 Random process

Correlation coefficient function

The correlation coefficient function X (t1, t2) of a random process {X(t)} is defined
by
KX (t1, t2)
X (t1, t2) = p p
KX (t1, t1) KX (t2, t2)
KX (t1, t2)
= ,
(t1)(t2)
where (ti) is the standard deviation of X(ti).
We can show that

|X (t1, t2)| 1,
X (t, t) = 1.
4.1 Random process

Crosscorrelation function
The crosscorrelation function RXY (t1, t2) of random processes {X(t)} and {Y (t)}
is defined by
RXY (t1, t2) = E{X(t1)Y (t2)}.
The autocorrelation and crosscorrelation functions satisfy

RX (t, t) = E{X(t)X (t)}
2
= X (t) + |E{X(t)}|2 0,

RX (t1, t2) = RX (t2, t1),
RXY (t1, t2) = RY X (t2, t1).
The autocorrelation function RX (t1, t2) is positive semi-definite. In other words,

PP
aiaj R(ti, tj ) 0 for non-negative constants {ak }.
i j
4.1 Random process

Crosscovariance function
The crosscovariance function KXY (t1, t2) of random processes {X(t)} and {Y (t)}
is defined by
KXY (t1, t2) = E{[X(t1) mX (t1)][Y (t2) mY (t2)]}
= RXY (t1, t2) mX (t1)mY (t2).
Two random processes which are uncorrelated

The random processes {X(t)} and {Y (t)} are said to be uncorrelated if RXY (t1, t2) =
E{X(t1)}E{Y (t2)} or KXY (t1, t2) = 0 for all t1 and t2.
Orthogonality
The two random processes {X(t)} and {Y (t)} are said to be orthogonal if
RXY (t1, t2) = 0 for all t1 and t2.
4.1 Random process

Random process
4.2 Properties of random processes
4.2 Properties of random processes

Stationary process and independent process

A random process is said to be stationary if the probabilistic properties

do not change under time shifts.
Stationary process
A random process {X(t)} is stationary, strict-sense stationary (s.s.s.), or strongly-
stationary if the joint cdf of {X(t1), X(t2), , X(tn)} is the same as the joint cdf
of {X(t1 + ), X(t2 + ), , X(tn + )} for all n, , t1, t2, , tn.
4.2 Properties of random processes / 4.2.1 Stationary process and independent process
Wide-sense stationary (w.s.s.) process

A random process {X(t)} is w.s.s., weakly-stationary, or second-order stationary if (1)
the mean function is constant and (2) the autocorrelation function RX (t, s) depends
only on t s but not on t and s individually.
The mean function mX and the autocorrelation function RX of a w.s.s. process

{X(t)} are thus
mX (t) = m
and
RX (t1, t2) = R(t1 t2),
respectively.
In other words, the autocorrelation function RX (t1, t2) of a w.s.s. process {X(t)}
is a function of = t1 t2. For all t and , we have
RX ( ) = E{X(t + )X (t)}.
When = 0, RX (0) = E{|X(t)|2}.
Consider two sequences of uncorrelated random variables A0, A1, , Am and

B0, B1, , Bm having mean zero and variance i2. Assume that the two se-
quences are uncorrelated with each other. Let 0, 1, , m be distinct frequen-
P
m
cies in [0, 2), and let Xn = {Ak cos nk +Bk sin nk } for n = 0, 1, 2, .
k=0
Then we can obtain
m
X
E{XnXn+l } = k2 cos lk ,
k=0
E{Xn} = 0.
Thus, {Xn} is w.s.s..
Properties of the autocorrelation function RX ( ) of a real stationary

process {X(t)}

RX ( ) = RX ( ) : RX ( ) is an even function.
|RX ( )| RX (0) : RX ( ) is maximum at the origin.
If RX ( ) is continuous = 0, then it is also continuous at every value of .
If there is a constant T > 0 such that RX (0) = RX (T ), then RX ( ) is periodic.
Independent random process

A random process is said to be independent if the joint cdf satisfies
n
Y
FXt1 ,Xt2 , ,Xtn (x) = FXti (xi)
i=1
for all n and t1, t2, , tn, x1, x2, , xn .
Independent and identically distributed (i.i.d.) process

A random process is said to be i.i.d. if the joint cdf satisfies
n
Y
FXt1 ,Xt2 , ,Xtn (x) = FX (xi)
i=1
for all n and t1, t2, , tn, x1, x2, , xn.
The i.i.d. process is sometimes called a memoryless process or a white

noise. The i.i.d. process is the simplest process, and yet it is the most
stochastic process in that past outputs do not have any information
on the future.
Bernoulli process
An i.i.d. random process with two possible values is called a Bernoulli process. For
example, consider the random process {Xn} defined by
(
1, if the nth result is head,
Xn =
0, if the nth result is tail,
when we toss a coin infinitely. The random process {Xn} is a discrete-time discrete-
amplitude random process. The success (head) and failure (tail) probabilities are
P {Xn = 1} = p
and
P {Xn = 0} = 1 p,
respectively.
We can easily obtain

mX (n) = E{Xn}
= p
and
KX (n1, n2) = E{Xn1 Xn2 } mX (n1)mX (n2)
(
p(1 p), n1 = n2,
=
0, n1 6= n2.
The mean function mX (n) of a Bernoulli process is not a function of time but a
constant. The autocovariance KX (n1, n2) depends not on n1 and n2 individually, but
only on the difference n1 n2.
Clearly, the Bernoulli process is w.s.s..
Two random processes independent of each other

The random processes {X(t)} and {Y (t)} are said to be independent of each
other if the random vector (Xt1 , Xt2 , , Xtk ) is independent of the random vector
(Ys1 , Ys2 , , Ysl ) for all k, l, and t1, t2, , tk , s1, s2, , sl .
Normal process, Gaussian process

A random process {Xt} is said to be normal if (Xt1 , Xt2 , , Xtk ) is a kdimensional
normal random vector for all k and t1, t2, , tk .
A stationary process is always w.s.s., but the converse is not always true. On the
other hand, a w.s.s. normal process is s.s.s.. This result can be obtained from the
pdf

1 1 T 1
fX (x) = exp (x mX ) K X (x mX )
(2)n/2|K X |1/2 2
of a jointly normal random vector.
Jointly w.s.s. processes

Two random processes are said to be jointly w.s.s. if (1) the mean functions are
constants and and (2) the autocorrelation functions and crosscorrelation function are
all functions only of time differences.
If two random processes {X(t)} and {Y (t)} are jointly w.s.s., then {X(t)} and
{Y (t)} are both w.s.s.. The crosscorrelation function of {X(t)} and {Y (t)} is
XY ( )
RXY (t + , t) = R
= E{X(t + )Y (t)}

= E{Y (t)X (t + )}
Y X ( ).
= R
The crosscorrelation function RXY of two jointly w.s.s. random processes has the
following properties:

1. RY X ( ) = RXY ( ).
p
2. |RXY ( )| RXX (0)RY Y (0) 12 {RXX (0) + RY Y (0)}.
3. RXY ( ) is not always maximum at the origin.
Linear transformation and jointly normal process

Two processes X(t) and Y (t) are w.s.s. if the linear combination Z(t) = aX(t) +
bY (t) is w.s.s. for all a and b. The converse is also true.
Moving average (MA) process

Let a1, a2, , al be a sequence of real numbers and W0, W1, W2, be a sequence
of uncorrelated random variables with mean E{Wn} = m and variance Var{Wn} =
2. Then the following process {Xn} is called a moving average process.
Xn = a1Wn + a2Wn1 + + al Wnl+1
Xl
= aiWni+1.
i=1
The mean and variance of Xn are

E{Xn} = (a1 + a2 + + al )m,
Var{Xn} = (a21 + a22 + + a2l ) 2.
2} = 2 when X
Since E{X n = Wn m, {Xn} is w.s.s. from
n
X l X l
Cov(Xn, Xn+k ) = E Xn m ai Xn+k m ai
i=1 i=1
X
l X
l
= E ni+1
ai X n+kj+1
aj X
i=1 j=1

2
E al alk Xn+kl+1 + al1alk1

= 2
X 2 ,
+ + ak+1a1X k l 1,
n+kl+2 n

0, k l,
(
(al alk + + ak+1a1) 2, k l 1,
=
0, k l.
Autoregressive (AR) process

Let the variance of an uncorrelated zero-mean random sequence Z0, Z1, be
( 2

1 2, n = 0,
Var{Zn} =
2, n 1,
where 2 < 1. Then the random process {Xn} defined by
X0 = Z0, Xn = Xn1 + Zn, n 1.
is called the first order autoregressive process. We can obtain
Xn = (Xn2 + Zn1) + Zn
= 2Xn2 + Zn1 + Zn
...
n
X
= niZi.
i=0
Thus the autocovariance function of {Xn} is

n n+m
!
X X
Cov(Xn, Xn+m) = Cov niZi, n+miZi
i=0 i=0
n
X
= nin+miCov(Zi, Zi)
i=0
n
!
1 X
= 22n+m + 2i
1 2 i=1
2 m

= .
1 2
Now, from the result above and the fact that the mean of {Xn} is E{Xn} = 0, it
follows that {Xn, n 0} is w.s.s..
Square-law detector
Let Y (t) = X 2(t) where {X(t)} is a Gaussian random process with mean 0 and
autocorrelation RX ( ). Then the expectation of Y (t) is E{Y (t)} = E{X 2(t)} =
RX (0). Since X(t+ ) and X(t) are jointly Gaussian with mean 0, the autocorrelation
of Y (t) can be found as
RY (t, t + ) = E{X 2(t)X 2(t + )}
= E{X 2(t + )}E{X 2(t)} + 2E 2{X(t + )X( )}
2 2
= RX (0) + 2RX ( ).
Thus E{Y 2(t)} = RY (0) = 3RX

2
(0) and Y2 = 3RX
2 2
(0) RX 2
(0) = 2RX (0). Clearly,
{Y (t)} is w.s.s..
Limiter
Let {Y (t)} = {g(X(t))} be a random process which is defined by a random process
{X(t)} and a limiter

1, x > 0,
g(x) =
1, x < 0.
Then we can easily obtain P {Y (t) = 1} = P {X(t) > 0} = 1 FX (0) and
P {Y (t) = 1} = P {X(t) < 0} = FX (0). Thus the mean and autocorrelation of
{Y (t)} are
E{Y (t)} = 1 P {Y (t) = 1} + (1) P {Y (t) = 1}
= 1 2FX (0),
RY ( ) = E{Y (t)Y (t + )}
= P {Y (t)Y (t + ) = 1} P {Y (t)Y (t + ) = 1}
= P {X(t)X(t + ) > 0} P {X(t)X(t + ) < 0}.
Now, if {X(t)} is a stationary Gaussian random process with mean 0, X(t + ) and
X(t) are jointly Gaussian with mean 0, variance RX (0), and correlation coefficient
RX ( )/RX (0). Clearly, FX (0) = 1/2.
We have (refer to (3.100), p. 156, Random Processes, Park, Song, Nam, 2004)
n X(t) o
P {X(t)X(t + ) < 0} = P <0
X(t + )
= FZ (0)
1 1 r1
= + tan1
2 1 1 r 2
1 1
= sin1 r
2
1 1 RX ( )
= sin1 ,
2 RX (0)
P {X(t)X(t + ) > 0} = 1 P {X(t)X(t + ) < 0}
1 1 RX ( )
= + sin1 .
2 RX (0)
Thus the autocorrelation of the limiter output is
2 1 RX ( )
RY ( ) = sin ,
RX (0)
from which we have E{Y 2(t)} = RY (0) = 1 and Y2 = 1 {1 2FX (0)}2 = 1.
Power of a random process
Power spectrum
The Fourier transform of the autocorrelation function of a w.s.s. random process is
called the power spectrum, power spectral density, or spectral density. It is usually
assumed that the mean is zero.
When the autocorrelation is RX ( ), the power spectrum is

SX () =
F{RX }
X

RX (k)ejk , discrete time,

= Zk

RX ( )ej d, continuous time.

If the mean is not zero, the power spectral density is defined by the Fourier trans-
form of the autocovariance instead of the autocorrelation.
4.2 Properties of random processes / 4.2.2 Power of random process

White noise, white process

Suppose that a discrete time random process {Xn} is uncorrelated so that RX (k) =
2k . Then we have
X
SX () = 2k ejk
k
2
= .
Such a process is called a white noise or white process. If {Xn} is Gaussian in addition,
it is called a white Gaussian noise.

Telegraph signal
Consider Poisson points with parameter . Let N (t) be the number of points in the
interval (0, t]. As shown in the figure above, consider the continuous-time random
process X(t) = (1)N (t) with X(0) = 1. Here, Wi is the time between adjacent
Poisson points. Assuming > 0, the autocorrelation of X(t) is
RX ( ) = E{X(t + )X(t)}
= 1 P {the number of points in the interval (t, t + ] is even}
+(1) P {the number of points in the interval (t, t + ] is odd}
n ( )2 o n ( )3 o

= e 1+ + e t + +
2! 3!
= e cosh e sinh
= e2 .

We can obtain a similar result when < 0, Combining the two results, we have
RX ( ) = e2| |. Consider a random variable A which is independent of X(t) and
takes on 1 or 1 with equal probability. Let Y (t) = AX(t). We then have
E{Y (t)} = E{A}E{X(t)}
= 0,
E{Y (t1)Y (t2)} = E{A2}E{X(t1)X(t2)}
= E{A2}RX (t1 t2)
= e2|t1t2|
since E{A} = 0, E{A2} = 1. Thus {Y (t)} is w.s.s., and the power spectral density
of {Y (t)} is
SY () = F{e2| |}
4
= 2 .
+ 42

Band limited noise, colored noise

When W > 0, let us consider a random process with the power spectral density
(
1, [W, W ],
SX () =
0, otherwise.
Such a process is called a colored noise. The autocorrelation of a colored noise is

thus
RX ( ) = F 1{SX ()}
sin(W )
= .


Power spectral density is not less than 0. That is,
SX () 0.
Cross power spectral density

The cross power spectral density SXY () of jointly w.s.s. processes {X(t)} and
{Y (t)} is
Z
SXY () = RXY ( )ej d.

Thus SXY () = SY X (), and the inverse Fourier transform of SXY () is

Z
1
RXY ( ) = SXY ()ej d.
2

Time delay process

Consider a w.s.s. process {X(t)} of which the power spectral density is SX ().
Letting Y (t) = X(t d), we have
RY (t, s) = E{Y (t)Y (s)}
= E{X(t d)X (s d)}
= RX (t s).
Thus the process {Y (t)} is w.s.s. and the power spectral density SY () equals to
SX (). In addition, the crosscorrelation and cross power spectral density of {X(t)}
and {Y (t)} are
RXY (t, s) = E{X(t)Y (s)} = E{X(t)X (s d)}
= RX (t s + d) = RX ( + d)
and
Z
SXY () = F{RX ( + d)} = RX ( + d)ej d

Z
= RX (u)ejuejddu = SX ()ejd.

That is, {X(t)} and {Y (t)} are jointly w.s.s..

Random process and linear systems
If the input random process is two-sided and w.s.s., the output of a

linear time invariant (LTI) filter is also w.s.s..
If the input random process is one-sided and w.s.s., however, the output
of an LTI filter is not w.s.s. in general.
4.2 Properties of random processes / 4.2.3 Random process and linear systems
Let h(t) be the impulse response of an LTI system and let

H() = F{h(t)} be the transfer function. Then the crosscorrelation
RXY (t1, t2) of the input random process {X(t)} and output random
process {Y (t)} and autocorrelation RY (t1, t2) of the output are
n Z o

RXY (t1, t2) = E{X(t1)Y (t2)} = E X(t1) X (t2 )h ()d
Z Z
= E{X(t1)X (t2 )}h()d = RX (t1, t2 )h()d
and
RY (t1, t2) = E{Y (t1)Y (t2)}
nZ Z o

= E X(t1 )h()d X (t2 )h ()d
Z Z
= RX (t1 , t2 )h()h()dd
Z
= RXY (t1 , t2)h()d,
respectively.
If the input and output are jointly w.s.s., we can obtain

Z
RXY ( ) = RX (t + )h()d
= RX ( ) h( ),
RY ( ) = RXY ( ) h( ).
since RX (t1, t2 ) = RX (t1 t2 + ) = RX ( + ) and RXY (t1 , t2) =
RXY (t1 t2 ) = RXY ( ).
The cross power spectral density and power spectral density of output are
SXY () = SX ()H (),
SY () = SXY ()H(),
respectively.
We can express the autocorrelation and power spectral density of the output in
terms of those of the input. Specifically, we have
RY ( ) = RX ( ) h( )
and
SY () = SX ()|H()|2,
where h(t) is called the deterministic autocorrelation of h(t) and is defined by
h(t) = F 1(|H()|2)

= h(t)
Z h (t)
= h(t + )h( )d.
Let SY () be the power spectral density of the output process {Yt}. Then we can
obtain
RY ( ) = F 1{SY ()}
= F 1{|H()|2SX ()}.
Coherence function
A measure of the degree to which two w.s.s. processes are related by an LTI trans-
formation is the coherence function ()
defined by
SXY ()
XY () = .
[SX ()SY ()]1/2

Here, |()| = 1 if and only if {X(t)} and {Y (t)} are the linearly related, that is,

Y (t) = X(t) h(t). Note the similarity between the coherence function () and
the correlation coefficient , a measure of the degree to which two random variables
are linearly related. The coherence function exhibits the property
|()|
1
similar to the correlation coefficient between two random variables.
Ergodic theorem*
such that
If there exists a random variable X
n1
1X
Xi X, discrete time random process,
n i=0 n
Z T
1
X(t)dt X, continuous time random process.
T 0 T
{Xn, n I} is said to satisfy ergodic theorem.
When a process satisfies an ergodic theorem, the sample mean con-

verges to something, which may be different from the expectation.
In some cases, a random process with time-varying mean satisfies an
ergodic theorem as shown in the example below.
4.2 Properties of random processes / 4.2.4 Ergodic theorem*

Suppose that nature at the beginning of time randomly selects one of two coins
with equal probability, one having bias p and the other having bias q. After the
coin is selected it is flipped once per second forever. The output random process
is a one-zero sequence depending on the face of a coin. Clearly, the time average
will converge: it will converge to p if the first coin was selected and to q if the
second coin was selected. That is, the time average will converge to a random
variable. In particular, it will not converge to the expected value p/2 + q/2.
If lim E{(Yn Y )2} = 0, Yn, n = 1, 2, is said to converge to Y

n
in mean square, which is denoted as
l.i.m. Yn = Y,
n
where l.i.m. denotes limit in the mean.

Mean ergodic theorem

Let {Xn} be an uncorrelated discrete time random process with finite mean E{Xn} =
2 2
P
n1
m and finite variance Xn = X . Then the sample mean Sn = Xi/n converges
i=0
to the expected value E{Xn} = m in mean square. That is,
n1
1X
l.i.m. Xi = m.
n n
i=0
A sufficient condition for a w.s.s. discrete time random process

{Xn, n I} to satisfy a mean ergodic theorem is KX (0) < and
lim KX (n) = 0.
n

Let {Xn} be a discrete time random process with mean E{Xn} and
autocovariance function KX (i, j). The process need not be stationary
in any sense. A necessary and sufficient condition for
n1
1X
l.i.m. Xi = m
n n
i=0
is
n1
1X
lim E{Xi} = m
n n
i=0
and
n1 n1
1 XX
lim KX (i, k) = 0.
n n2
i=0k=0
That is, if and only if a process is asymptotically uncorrelated and its

sample averages converge, the sample mean converge in mean square.

Mean square ergodic theorem

P
n
Let Xn = Xi/n where {Xn, n 1} is a second order stationary process with
i=1
n m)2} = 0
mean m and autocovariance K(i) = Cov(Xn, Xn+i). Then lim E{(X
n
P
n1
if and only if lim K(i)/n = 0.
n i=0
Let K(i) be the autocovariance of {Xn}, a second order stationary Gaussian

process with mean 0. If
T
1X 2
lim K (i) = 0,
T T
i=1
then
T (i) K(i)|2} = 0
lim E{|K
T
T (i) = PXl Xl+i/T is the sample autocovariance.

T
for i = 1, 2, , where K
l=1

Ergodicity*
Shift operator
An operator T for which T = {xt+1, t I}, where = {xt, t I} is an infinite
sequence in the probability space (RI , B(R)I , P ), is called a shift operator.
Stationary process
A discrete time random process with process distribution P is stationary if
P (T 1F ) = P (F ) for any element F in B(R)I.
4.2 Properties of random processes / 4.2.5 Ergodicity*

Invariant event
An event F is said to be invariant with respect to the shift operator T if and only if
T 1F = F .
Ergodicity, ergodic process

A random process is said to be ergodic if P (F ) = 0 or P (F ) = 1 for any invariant
event F .
Consider a two-sided process with distribution

P ( , x1 = 1, x0 = 0, x1 = 1, x2 = 0, ) = p,
P ( , x1 = 0, x0 = 1, x1 = 0, x2 = 1, ) = 1 p.
Clearly, F = {sequence of alternating 0 and 1} is an invariant event, and has
probability P (F ) = 1. Any other invariant event - for example, the all 1 sequence
- that does not include F has probability 0. Thus the random process is ergodic.
Ergodicity has nothing to do with stationarity or convergence of sample averages.
4.2 Properties of random processes / 4.2.5 Ergodicity*

Random processes
4.3 Process with i.s.i.

Process with i.s.i.
Process with independent increments

A random process {Yt, t I} is said to have independent increments or to be an
independent increment process if for all choices of k = 1, 2, and all choices of
ordered sample times {t0, t1, , tk }, the k increments Yti Yti1 , i = 1, 2, , k
are independent random variables.
Process with stationary increments

When the increments {Yt Ys} are stationary, the random process {Yt} is called a
stationary increment random process.
Process with i.s.i.

A random process is called an independent and stationary increment (i.s.i.) process
if its increments are independent and stationary.

A discrete time random process is an i.s.i. process if and only if it can

be represented as the sum of i.i.d. random variables.
Mean and autocovariance of i.s.i. process

The mean and autocovariance of a discrete time i.s.i. process are
E{Yt} = tE(Y1), t0
and
KY (t, s) = Y21 min(t, s), t, s 0.
Clearly, an i.s.i. process itself is not stationary.

Let the process {Xt, t T } be an i.s.i. process. If

m0 = E{X0}, m1 = E{X1} m0,
02 = E{(X0 m0)2}, 12 = E{(X1 m1)2} 02,
we have
E{Xt} = m0 + m1t,
Var{Xt} = 02 + 12t.

Point process and counting process

A sequence T1 T2 T3 of ordered random variables is called a point
process. For example, the set of times defined by Poisson points is a point process.
A counting process Y (t) can be defined as the number of points in the interval [0, t).
We have, with T0 = 0,
Y (t) = i, Ti t < Ti+1, i = 0, 1, .
.

A counting process
A process constructed by summing the outputs of a Bernoulli process
Let {Xn, n = 1, 2, } be a Bernoulli process with parameter p. Define the

random process {Yn, n = 1, 2, } as
Y0 = 0,
n
Yn = Xi
i=1
= Yn1 + Xn, n = 1, 2, .
Since the random variable Yn represents the number of 1s in {X1, X2, , Xn},
we have
Yn = Yn1 or Yn = Yn1 + 1, n = 2, 3, .
In general, a discrete time process satisfying this relation is called a counting
process since it is nondecreasing, and when it jumps, it is always with an increment
of 1.

Properties of the random process {Yn}
E{Yn} = np, Var{Yn} = np(1 p), KY (k, j) = p(1 p) min(k, j)
Marginal pmf for Yn

pYn (y) = Pr{ there are exactly y ones in X1, X2, , Xn.}

n y
= p (1 p)ny , y = 0, 1, , n.
y
Since the marginal pdf is binomial, the process {Yn} is called a binomial counting
process.
The process {Yn} is not stationary since the marginal pmf depends on the time
index n.

Random walk process
One dimensional random walk, random walk
Consider the modified Bernoulli process for which the event failure is represented
by 1 instead of 0.

+1, for success in the nth trial,
Zn =
1, for failure in the nth trial.
Let p = Pr{Zn = 1}, and consider the sum

n

Wn = Zi
i=1
of the variables Zn. The process {Wn} is referred to as one dimensional random
walk or random walk.

Since
Zi = 2Xi 1,
it follows that the random walk process {Wn} is related to the binomial counting
process {Yn} by
Wn = 2Yn n.
Using the results on the mean and autocorrelation functions of the binomial count-
ing process and the linearity of expectation, we have
mW (n) = (2p 1)n,
KW (n1, n2) = 4p(1 p) min(n1, n2),
2
W n
= 4p(1 p)n.

Random process
4.4 Discrete time process with i.s.i.

Discrete time process with i.s.i.
Discrete time discrete alphabet process {Yn} with i.s.i.

As mentioned before, {Yn} can be dened by the sum of i.i.d. random variables {Xi}.
Let us consider the following conditional pmf.

pYn|Y n1 (yn|y n1) = pYn|Y n1 (yn|yn1, , y1)
= Pr(Yn = yn|Y n1 = y n1),
where Y n = (Yn, Yn1, , Y1) and y n = (yn, yn1, , y1).
The conditioning event {Yi = yi, i = 1, 2, , n 1} above is the same as the

event {X1 = y1, Xi = yi yi1, i = 2, , n 1}. In addition, under the
conditioning event, we have Yn = yn if and only if Xn = yn yn1.

Assuming y0 = 0,
pYn|Y n1 (yn|y n1) = Pr(Yn = yn|X1 = y1, Xi = yi yi1, i = 2, 3, , n 1)
= Pr(Xn = yn yn1|Xi = yi yi1, i = 1, 2, , n 1)
= pXn|X n1 (yn yn1|yn1 yn2, , y2 y1, y1),
where X n1 = (Xn1, Xn2, , X1).
If {Xn} are i.i.d.,
pYn|Y n1 (yn|y n1) = pX (yn yn1)
since Xn is independent of Xk for k < n.
Thus the joint pmf is
pY n (y n) = pYn|Y n1 (yn|y n1) pY n1 (y n1)
...
n n

= pY1 (y1) pYi|Yi1, ,Y1 (yi|yi1, , y1) = pX (yi yi1).
i=2 i=1

Applying the result above to the binomial counting process, we obtain

n

n
pY n (y ) = p(yiyi1)(1 p)1(yiyi1),
i=1
where yi yi1 = 0 or 1 for i = 1, 2, , n and y0 = 0.
Properties of processes with i.s.i.

We can express the conditional pmf of Yn given Yn1 as follows:
pYn|Yn1 (yn|yn1) = Pr(Yn = yn|Yn1 = yn1)
= Pr(Xn = yn yn1|Yn1 = yn1).
The conditioning event {Yn1 = yn1} depends only on Xk for k < n, and Xn is
independent of Xk for k < n. Thus, this conditioning event does not aect Xn.
Consequently,
pYn|Yn1 (yn|yn1) = pX (yn yn1).

Discrete time i.s.i. processes (such as the binomial counting process and discrete
random walk) has the following property:
pYn|Y n1 (yn|y n1) = pYn|Yn1 (yn|yn1),
Pr{Yn = yn|Y n1 = y n1} = Pr{Yn = yn|Yn1 = yn1}.
Roughly speaking, given the most recent past sample (or the current sample), the
other past samples do not aect the probability of what happens next.
A discrete time discrete alphabet random process with this property is called a
Markov process. Thus all i.s.i. processes are Markov processes.

Gamblers ruin problem

A person wants to buy a new car of which the price is N won. The person has k
(0 < k < N ) won, and he intends to earn the dierence from gambling. The game
this person is going to play is that if a toss of a coin results in a head, he will earn 1
won, and if it results in a tail, he will lose 1 won. Let p represent the probability of
heads, and q = 1 p. Assuming the man continues to play the game until he earns
enough money for a new car or lose all the money he has, what is the probability that
the man loses all the money he has?
Let Ak be the event that the man loses all the money when he has started with k
won and B be the event the man wins a game. Then,
P (Ak ) = P (Ak |B)P (B) + P (Ak |B c)P (B c).
Since the game will start again with k + 1 won if a toss of a coin results in a head
and k 1 won if a toss of a coin results in a tail, it is easy to see that
P (Ak |B) = P (Ak+1),
P (Ak |B c) = P (Ak1).

Let pk = P (Ak ), then p0 = 1, pN = 0, and

pk = ppk+1 + qpk1, 1 k N 1.
Assuming pk = k , we get from the equation above
p2 + q = 0,
which gives 1 = 1 and 2 = q/p.
If p = 0.5, pk = A11k + A22k . Thus using the boundary conditions p0 = 1 and

pN = 0, we can nd
(q/p)k (q/p)N
pk = .
1 (q/p)N
If p = 0.5, we have pk = (A1 + A2k)1k since q/p = 1. Thus using the boundary
conditions p0 = 1 and pN = 0, we can nd
k
pk = 1 .
N

Discrete time Wiener process
Discrete time Wiener process

Let {Xn} be an i.i.d. zero-mean Gaussian process with variance 2. The discrete
time Wiener process {Yn} is dened by
Y0 = 0,
n
Yn = Xi
i=1
= Yn1 + Xn, n = 1, 2, .
The discrete time Wiener process is also called the discrete time diusion process
or discrete time Brownian motion.
Since the discrete time Wiener process is formed as sums of an i.i.d. process, it
has i.s.i.. Thus we have E{Yn} = 0 and KY (k, j) = 2 min(k, j).
The Wiener process is a rst-order autoregressive process.
4.4 Discrete time process with i.s.i. / 4.4.3 discrete time Wiener process
The discrete time Wiener process is a Gaussian process with mean func-
tion m(t) = 0 and autocovariance function KX (t, s) = 2 min(t, s).
Since the discrete time Wiener process is an i.s.i. process, we have
fYn|Yn1 (yn|yn1) = fX (yn yn1),
fYn|Y n1 (yn|y n1) = fYn|Yn1 (yn|yn1)
= fX (yn yn1).
As in the discrete alphabet case, a process with this property is called a Markov
process.
Markov process
A discrete time random process {Yn} is called a rst order Markov process if it
satises
Pr{Yn yn|yn1, yn2, } = Pr{Yn yn|yn1}.
for all n, yn, yn1, yn2, .
4.4 Discrete time process with i.s.i. / 4.4.3 discrete time Wiener process
Random processes
4.5 Continuous time i.s.i. process

Continuous time i.s.i. process
Continuous time i.s.i. process

When we deal with a continuous time process with i.s.i. we need to consider more
general collection of sample times than in the case of discrete time process.
In the continuous time case, we assume that we are given the cdf of the increments
as
FYtYs (y) = FY|ts|Y0 (y)
= FY|ts| (y), t > s.

The joint probability functions of a continuous time process

Dene the random variable {Xn} by
Xn = Ytn Ytn1 .
Then {Xn} are independent and
n

Ytn = Xi,
i=1
Pr{Ytn yn|Ytn1 = yn1, Ytn2 = yn2, } = FXn (yn yn1)

= FYtn Ytn1 (yn yn1).
As in the case of discrete time processes, these can be used to nd the joint pmf or
pdf as
n

pYt1 , ,Ytn (y1, , yn) = pYti Yti1 (yi yi1),
i=1
n
fYt1 , ,Ytn (y1, , yn) = fYti Yti1 (yi yi1).
i=1

If a process {Yt} has i.s.i., and the cdf pdf, or pmf for Yt = Yt Y0 is
given, the process can be completely described as shown above.
As in the discrete time case, a continuous time random process {Yt} is

called a Markov process if we have
Pr{Ytn yn|Ytn1 = yn1, Ytn2 = yn2, } = Pr{Ytn yn|Ytn1 = yn1},
fYtn |Ytn1 , ,Yt1 (yn|yn1, , y1) = fYtn |Ytn1 (yn|yn1),
or
pYtn |Ytn1 , ,Yt1 (yn|yn1, , y1) = pYtn |Ytn1 (yn|yn1)
for all n, yn, yn1, , and t1, t2, , tn .
A continuous time i.s.i. process is a Markov process.

Wiener process
Wiener process
A process is called Wiener process if it satises
The initial position is zero. That is, W (0) = 0.
The mean is zero. That is, E{W (t)} = 0, t 0.
The increments of W (t) are independent, stationary, and Gaussian.
Wiener process is a continuous time i.s.i. process.
The increments of Wiener process are Gaussian random variables with

zero mean.
The rst order pdf of Wiener process is

1 x2
fWt (x) = exp 2 .
2t 2 2t
4.5 Continuous time i.s.i. process / 4.5.1 Wiener process

Wiener process, Brownian motion

Wiener process is the limit of the random-walk process.
Properties of Wiener process
The distribution of Xt2 Xt1 , t2 > t1, depends only on t2 t1, not t1 and t2
individually.

Let us show that the Wiener process is Gaussian.

From the denition of the Wiener process, the random variables W (t1), W (t2)
W (t1), W (t3) W (t2), , W (tk ) W (tk1) are independent Gaussian random
variables. Thus the random variables W (t1), W (t2), W (t3), , W (tk ) can be
obtained from the following linear transformation of W (t1) and the increments
W (t1) = W (t1),
W (t2) = W (t1) + {W (t2) W (t1)},
W (t3) = W (t1) + {W (t2) W (t1)} + {W (t3) W (t2)},
...
W (tk ) = W (t1) + {W (t2) W (t1)} +
+{W (tk ) W (tk1)}.
Since W (t1), W (t2), W (t3), , W (tk ) is jointly Gaussian, {W (t)} is Gaussian.

Poisson counting process

Poisson counting process
A continuous time counting process {Nt, t 0} with the following properties is called
the Poisson counting process.
N0 = 0.
The process has i.s.i.. Hence, the increments over non-overlapping time intervals
are independent random variables.
In a very small time interval, the probability of an increment of 1 is proportional to
the length of the interval, and the probability of an increment larger than 1 is 0.
Thus, Pr{Nt+t Nt = 1} = t + o(t), Pr{Nt+t Nt 2} = o(t), and
Pr{Nt+t Nt = 0} = 1 t + o(t), where is a proportionality constant.
4.5 Continuous time i.s.i. process / 4.5.2 Poisson counting process

The Poisson counting process is a continuous time discrete alphabet i.s.i. pro-
cess.
We have obtained the Wiener process as the limit of a discrete time discrete
amplitude random-walk process. Similarly, the Poisson counting process can
be derived as the limit of a binomial counting process using the Poisson ap-
proximation.

The probability mass function pNtN0 (k) = pNt (k) of the increment
Nt N0 between the starting time 0 and t > 0
Let us use the notation
p(k, t) = pNtN0 (k), t > 0.
Using the independence of increments and the third property of the Poisson counting
process, we have
k

p(k, t + t) = Pr{Nt = n}Pr{Nt+t Nt = k n|Nt = n}
n=0
k

= Pr(Nt = n)Pr(Nt+t Nt = k n)
n=0
p(k, t)(1 t) + p(k 1, t)t,

which yields
p(k, t + t) p(k, t)
= p(k 1, t) p(k, t).
t
When t 0, the equation above becomes the following dierential equation

d
p(k, t) + p(k, t) = p(k 1, t), t > 0,
dt
where the initial conditions are

0, k = 0,
p(k, 0) =
1, k=0
since Pr{N0 = 0} = 1. Solving the dierential equation gives
pNt (k) = p(k, t)
(t)k et
= , k = 0, 1, 2, , t 0.
k!
The pmf for k jumps in an arbitrary interval (s, t), t s

((t s))k e(ts)
pNtNs (k) = , k = 0, 1, , t s.
k!

Martingale
Martingale property
An independent increment process {Xt} with zero mean satises
E {X(tn) X(tn1)|X(t1), X(t2), , X(tn1)} = 0
for all t1 < t2 < < tn and integer n 2. This property can be rewritten as
E {X(tn)|X(t1), X(t2), , X(tn1)} = X(tn1),
which is called the martingale property.
Martingale
A process {Xn, n 0} with the following properties is called a martingale process.
E{|Xn|} < .
E{Xn+1|X0, , Xn} = Xn.
4.5 Continuous time i.s.i. process / 4.5.3 Martingale

Random processes
4.6 Compound process*
4.6 Compound process*

Discrete time compound process*
Discrete time compound process

Let {Nk , k = 0, 1, } be a discrete time counting process such as the binomial
counting process, and let {Xk , k = 0, 1, } be an i.i.d. random process. Assume
that the two processes are independent of each other. Dene the random process
{Yk , k = 0, 1, } by
Y0 = 0,
Nk

Yk = Xi, k = 1, 2, ,
i=1
where we assume Yk = 0 if Nk = 0. The process {Yk } is called a discrete time

compound process. The process is also referred to as a doubly stochastic process
because of the two sources {Nk } and {Xk } of randomness.
4.6 Compound process* / 4.6.1 Discrete time compound process*

The expectation of Yk can be obtained using the conditional probability as

Nk
E{Yk } = E Xi (1)
i=1
= E{E{Yk |Nk }}

= pNk (n)E{Yk |Nk = n}
n

= pNk (n)nE{X}
n
= E{Nk }E{X}.
Let X1, X2, be an i.i.d. sequence of random variables and GX be

their common moment generating function. Let the random variable
N be independent of Xi and GN be the moment generatingN function
of N . Then the moment generating function of S = i=1 Xi is
GS (t) = GN (GX (t)).
4.6 Compound process* / 4.6.1 Discrete time compound process*
Continuous time compound process*
When a continuous time counting process {Nt, t 0} and an i.i.d.

process {Xi, i = 1, 2, } are independent of each other, the process
Nt

Yt = Y (t) = Xi
i=1
is called a continuous time compound process. Here, we put Y (t) = 0

when Nt = 0.
We have
E{Yt} = E{Nt}E{X},
MYt (u) = E{MXNt (u)}.
4.6 Compound process* / 4.6.2 Continuous time compound process*

A compound process is continuous and dierentiable except at the

jumps, where a new random variable is added.
If {Nt} is a Poisson counting process, we have
E{Yt} = tE{X},

(t)k et
MYt (u) = MXk (u)
k!
k=0

(tMX (u))k
= et
k!
k=0
= et(1MX (u)).
4.6 Compound process* / 4.6.2 Continuous time compound process*

04 e PDF

Uploaded by

Copyright:

Available Formats

04 e PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

04 e PDF

Uploaded by

Copyright:

Available Formats

Random processes - Chapter 4 Random process 1

Chapter 4 Random process

4.1 Random process

4.1 Random process

Random process, stochastic process

In other words, a random vector with an infinite number of elements is called a

Discrete time process, continuous time process

4.1 Random process

Discrete (alphabet) process, continuous (alphabet) process

A random process {X()} maps an element of the sample space on a time

4.1 Random process

A random process and sample functions

4.1 Random process

Since a random process is a collection of random variables with X(t) denoting a

4.1 Random process

RX (t1, t2) = E{X(t1)X (t2)}.

4.1 Random process

4.1 Random process

The autocorrelation function of X(t) = Aejt defined with a random

In general, the autocovariance and autocorrelation functions are functions of t1

4.1 Random process

Uncorrelated random process

If a random process {X(t)} is uncorrelated, the autocorrelation and autocovari-

4.1 Random process

Correlation coefficient function

We can show that

4.1 Random process

The autocorrelation and crosscorrelation functions satisfy

The autocorrelation function RX (t1, t2) is positive semi-definite. In other words,

4.1 Random process

Two random processes which are uncorrelated

4.1 Random process

Chapter 4 Random process

4.2 Properties of random processes

4.2 Properties of random processes

Stationary process and independent process

A random process is said to be stationary if the probabilistic properties

Wide-sense stationary (w.s.s.) process

The mean function mX and the autocorrelation function RX of a w.s.s. process

Consider two sequences of uncorrelated random variables A0, A1, , Am and

Properties of the autocorrelation function RX ( ) of a real stationary

Independent random process

for all n and t1, t2, , tn, x1, x2, , xn .

Independent and identically distributed (i.i.d.) process

for all n and t1, t2, , tn, x1, x2, , xn.

The i.i.d. process is sometimes called a memoryless process or a white

We can easily obtain

Two random processes independent of each other

Normal process, Gaussian process

Jointly w.s.s. processes

Linear transformation and jointly normal process

Moving average (MA) process

The mean and variance of Xn are

Autoregressive (AR) process

Thus the autocovariance function of {Xn} is

Thus E{Y 2(t)} = RY (0) = 3RX

Power of a random process

When the autocorrelation is RX ( ), the power spectrum is

4.2 Properties of random processes / 4.2.2 Power of random process