This action might not be possible to undo. Are you sure you want to continue?
1 Stochastic Processes
Arandom variable is a mapping function which assigns outcomes of a random
experiment to real numbers (see Fig. 1). Occurrence of the outcome follows
certain probability distribution. Therefore, a random variable is completely
characterized by its probability density function (PDF).
Figure 1. Arandom variable is a function mapping outcomes of a random
experiment to real numbers.
Astochastic process { } T t t X e ), ( (or{ } T t X
t
e , ) is a family of random variables
where the index set T may be discrete ( } , 2 , 1 , 0 { = T ) or continuous ( ) , 0 [ · = T ).
The set of possible values which random variables T t t X e ), ( may assume is
called the state space of the process.
Acontinuous time stochastic process { } T t t X e ), ( is said to have independent
increments if for all choices of
n
t t t < < <
1 0
, the n random variables
) ( ) ( , ), ( ) ( ), ( ) (
1 1 2 0 1 ÷
÷ ÷ ÷
n n
t X t X t X t X t X t X
are independent. The process is said to have stationary independent increments if
in addition ) ( ) (
1
s t X s t X
i i
+ ÷ +
÷
has the same distribution as ) ( ) (
1 ÷
÷
i i
t X t X
for all 0 and ,
1
> e
÷
s T t t
i i
.
Arandom variable X can be assigned a number ) (e x based on the outcome e
of a random experiment. Similarly, a random process { } T t t X e ), ( can assume
values { } T t t x e ), , ( e depending on the outcome of a random experiment. Each
possible{ } T t t x e ), , ( e is called a realization of the random process{ } T t t X e ), ( .
Atotality of all realizations is called the ensemble of the random process.
The term “stochastic processes” appears mostly in statistical textbooks; however,
the term “random processes” are frequently used in books of many engineering
applications.
2 Characterizations of a Stochastic Processes
Firstorder densitiesof arandomprocess
Astochastic process is defined to be completelyor totallycharacterizedif the
joint densities for the random variables ) ( ), ( ), (
2 1 n
t X t X t X are known for all
times
n
t t t , , ,
2 1
and all n. In general, a complete characterization is practically
impossible, except in rare cases. As a result, it is desirable to define and work with
various partial characterizations. Depending on the objectives of applications, a
partial characterization often suffices to ensure the desired outputs.
For a specific t, X(t) is a random variable with distribution ] ) ( [ ) , ( x t X p t x F s = .
The function ) , ( t x F is defined as the firstorder distribution of the random
variable X(t). Its derivative with respect to x
x
t x F
t x f
c
c
=
) , (
) , ( (11)
is the firstorder density of X(t). If the firstorder densities defined for all time t,
i.e. f(x,t), are all the same, then f(x,t) does not depend on t and we call the resulting
density the firstorder density of the random process { } ) (t X ; otherwise, we have
a family of firstorder densities.
The firstorder densities (or distributions) are only a partial characterization of the
random process as they do not contain information that specifies the joint densities
of the random variables defined at two or more different times.
Meanandvarianceof arandomprocess
The firstorder density of a random process, f(x,t), gives the probability density of
the random variables X(t) defined for all time t. The mean of a random process,
m
X
(t), is thus a function of time specified by
}
+·
· ÷
= = =
t t t t X
dx t x f x X E t X E t m ) , ( ] [ )] ( [ ) ( (12)
For the case where the mean of X(t) does not depend on t, we have
constant). (a )] ( [ ) (
X X
m t X E t m = = (13)
The variance of a random process, also a function of time, is defined by
{ }
2 2 2 2
)] ( [ ] [ )] ( ) ( [ ) ( t m X E t m t X E t
X t X X
÷ = ÷ = o (14)
Secondorder densitiesof arandomprocess
For any pair of two random variables X(t
1
) and X(t
2
), we define the secondorder
densities of a random process as ) , ; , (
2 1 2 1
t t x x f or ) , (
2 1
x x f .
Nthorder densitiesof arandomprocess
The nth order density functions for { } ) (t X at times
n
t t t , , ,
2 1
are given by
) , , , ; , , , (
2 1 2 1 n n
t t t x x x f or ) , , , (
2 1 n
x x x f .
Autocorrelationandautocovariancefunctionsof randomprocesses
Given two random variables X(t
1
) and X(t
2
), a measure of linear relationship
between them is specified by E[X(t
1
)X(t
2
)]. For a random process, t
1
and t
2
go
through all possible values, and therefore, E[X(t
1
)X(t
2
)] can change and is a
function of t
1
and t
2
. The autocorrelation function of a random process { } ) (t X is
thus defined by
  ) , ( ) ( ) ( ) , (
1 2 2 1 2 1
t t R t X t X E t t R = = (15)
The autocovariance function of a random process { } ) (t X is defined by
( )( )  
) ( ) ( ) , (
) ( ) ( ) ( ) ( ) , (
2 1 2 1
2 2 1 1 2 1
t m t m t t R
t m t X t m t X E t t C
X X
X X
÷ =
÷ ÷ =
(16)
The normalized autocovariance function is defined by
) , ( ) , (
) , (
) , (
2 2 1 1
2 1
2 1
t t C t t C
t t C
t t = µ (17)
Stationarityof randomprocesses
Arandom process { } ) (t X is called strictlystationary(or strictsense stationary,
SSS) if the sets of random variables ) ( ), ( ), (
2 1 n
t X t X t X and
) ( ), ( ), (
2 1
A + A + A +
n
t X t X t X have the same probability density functions for
all t
i
, all n and all A, i.e.,
( ) ( ) A + A + A + =
n n n n
t t t x x x f t t t x x x f , , , ; , , , , , , ; , , ,
2 1 2 1 2 1 2 1
. (18)
Strictsense stationarity seldom holds for random processes, except for some
Gaussian processes. Therefore, weaker forms of stationarity are needed.
Arandom process is called Nth order stationary (or stationary of order N) if
the condition of Eq. (18) holds for all N n s for N a fixed integer.
Amuch weaker form of stationarity, even weaker than stationarity of order two, is
weaksense stationarity (or widesense stationarity, WSS). Arandom process
{ } ) (t X is widesense stationary if
  . all for constant) ( ) ( t m t X E = (19)
( ) ( ) . and all for , ) , (
2 1 1 2 1 2 2 1
t t t t R t t R t t R ÷ = ÷ = (110)
Equalityandcontinuityof randomprocesses
Equality Two random processes { } ) (t X and { } ) (t Y are equal (everywhere) if
their respective samples x(t, e
i
) = y(t, e
i
) for every e
i
.
[Note that “x(t, e
i
) = y(t, e
i
) for every e
i
” is not the same as “x(t, e
i
) =
y(t, e
i
) with probability 1”.]
Mean square equality Two random processes { } ) (t X and { } ) (t Y are equal in
the meansquare sense if
  0 ) ( ) (
2
= ÷ t Y t X E (111)
for every t. Equality in the MS sense leads to the following conclusions: We
denote by A
t
the set of outcomes e
i
such that x(t, e
i
) = y(t, e
i
) for a specific t, and
by A
·
the set of outcomes e
i
such that x(t, e
i
) = y(t, e
i
) for every t. From Eq.
(111) it follows that x(t, e
i
) – y(t, e
i
) = 0 with probability 1, hence P(A
t
) = P(O)
=1 where O represents the sample space. It does not follow, however, that P(A
·
)
=1. In fact, P(A
·
) is the intersection of all sets A
t
as t ranges over the entire axis
and P(A
·
) might even equal 0.
Stochastic continuity
(1) Continuous in probability
Arandom process { } ) (t X is called continuous in probability at t if for any c
> 0,
( ) 0  ) ( ) ( 
0 ÷
÷ > ÷ +
h
t X h t X P c (112)
(2) Continuous in meansquare sense
Arandom process { } ) (t X is called meansquare (MS) continuous at t if
  { } 0 ) ( ) (
0
2
÷
÷ ÷ +
c
c t X t X E . (113)
Arandom process { } ) (t X is MS continuous if its autocorrelation function
is continuous.
Proof
  { } ) , ( ) , ( 2 ) , ( ) ( ) (
2
t t R t t R t t R t X t X E + + ÷ + + = ÷ + c c c c (114)
If ) , (
2 1
t t R is continuous, then the RHS of Eq. (114) approaches zero as
. 0 ÷ c [Note that ) , (
2 1
t t R is used in the above equation to simply the
expression.]
Suppose that Eq. (113) holds for every t in an interval I. It follows that
almost all samples (or realizations) of { } ) (t X will be continuous for a
particular point of I. It does not follow, however, that these samples (or
realizations) of { } ) (t X will be continuous for every point in I.
If { } ) (t X is MS continuous, then its mean is continuous, i.e.,
) ( ) (
0
t m t m
X X
÷
÷ +
c
c (115)
Is it possible to have a realization which is discontinuous at time t whereas
the random process is meansquare continuous?
x(t)
time, t
Proof
  { }   { }
2 2
) ( ) ( ) ( ) ( t X t X E t X t X E ÷ + > ÷ + c c (116)
Therefore,   0 ) ( ) (
0 ÷
÷ ÷ +
c
c t X t X E
[Note:   0 ) ( ) ( ) (
2 2
> ÷ = X E X E X Var ]
(3) Almostsurely continuous
Arandom process { } ) (t X is called almostsurely continuous at t if
1 0  ) ; ( ) ; (  :
lim
0
= 
.

\

= ÷ +
÷
e e e t X h t X P
h
(117)
If { } ) (t X is continuous in probability (meansquare continuous, almostsurely
continuous) at every t, then it is said to be continuous in probability (meansquare
continuous, almostsurely continuous).
Example 1 Suppose that X(t) is a random process with 5 ) ( = t m
X
and
  6 . 0
2 1
2 1
3 25 ) , (
t t
e t t R
÷ ÷
+ = . Determine the mean, the variance and the covariance of
the random variables U=X(6) and V=X(9).
[Solution] 5 ) 6 ( )] 6 ( [ ) ( = = =
X
m X E U E , 5 ) 9 ( )] 9 ( [ ) ( = = =
X
m X E V E
  { } 3 25 28 25 ) 6 , 6 ( ) 6 ( } )] 6 ( {[ ) (
2 2
= ÷ = ÷ = ÷ = R X E X E U Var
  { } 3 25 28 25 ) 9 , 9 ( ) 9 ( } )] 9 ( {[ ) (
2 2
= ÷ = ÷ = ÷ = R X E X E V Var
  496 . 0 3 ) 9 ( )] 6 ( [ ) 9 , 6 ( ) 9 , 6 ( ) , (
8 . 1
= = ÷ = =
÷
e X E X E R C V U Cov
Example 2 The integral of a stochastic process X(t) is a random variable. Let
}
=
b
a
dt t X S ) ( it yields
 
} } }
= =
(
¸
(
¸
= =
b
a
X
b
a
b
a
S
dt t m dt t X E dt t X E S E m ) ( ) ( ) ( ) (
} }
=
b
a
b
a
dt dt t X t X S
2 1 2 1
2
) ( ) (
{ }
} }
} } } }
=
= =
b
a
b
a
b
a
b
a
b
a
b
a
dt dt t t R
dt dt t X t X E dt dt t X t X E S E
2 1 2 1
2 1 2 1 2 1 2 1
2
) , (
)} ( ) ( { ) ( ) ( ) (
StochasticConvergence
Arandom sequence or a discretetime random process is a sequence of random
variables {X
1
(e), X
2
(e), …, X
n
(e),…} = {X
n
(e)}, e e O.
For a specific e, {X
n
(e)} is a sequence of numbers that might or might not
converge. The notion of convergence of a random sequence can be given several
interpretations:
Sure convergence (convergence everywhere)
The sequence of random variables {X
n
(e)} converges surely to the random
variable X(e) if the sequence of functions X
n
(e) converges to X(e) as n ÷ · for
all e e O, i.e.,
X
n
(e) ÷ X(e) as n ÷ · for all e e O. (118)
Asequence of real numbers x
n
converges to the real number x if, given any
c > 0, we can always specify an integer N such that for all values of n
beyond N we can guarantee that x
n
– x < c.
Convergence of a sequence of numbers
Sure convergence requires that the sample sequence corresponding to
every e converges. However, it does not require that all the sample
sequences converge to the same value; that is the sample sequences for
different e and e' can converge to different values.
Almostsure convergence (convergence almost everywhere, convergence with
probability 1)
The sequence of random variables {X
n
(e)} converges almost surely to the random
variable X(e) if the sequence of functions X
n
(e) converges to X(e) as n ÷ · for
all e e O, except possibly on a set of probability zero; i.e.,
1 ) ( ) ( : =
(
¸
(
¸
÷
· ÷
e e e X X P
n
n
. (119)
2c
N
n
x
n
x
Almostsure convergence
Meansquare convergence
The sequence of random variables {X
n
(e)} converges in the mean square sense to
the random variable X(e) if
( )   0 ) ( ) (
2
÷ ÷ e e X X E
n
as · ÷ n (120)
Convergence in probability
The sequence of random variables {X
n
(e)} converges in probability to the random
variable X(e) if, for any c > 0,
  0 ) ( ) ( ÷ > ÷ c e e X X P
n
as · ÷ n (121)
Convergence in probability for the case where the limiting
random variable is a constant x
Convergence in distribution
The sequence of random variables {X
n
(e)} with cumulative distribution functions
{F
n
(x)} converges in distribution to the random variable X(e) with cumulative
distribution functions F(x) if
) ( ) ( x F x F
n
÷ as · ÷ n (122)
for all x at which F(x) is continuous.
2c
N
n
x
n
x
With probability 0
2c
n
x
x
n
n
0
Remarks
(1) The principal difference in the definitions of convergence in probability and
convergence with probability one is that the limit is outside the probability in
the former and inside the probability in the latter.
(2) Convergence with probability one applies to the individual realizations of the
random process. Convergence in probability does not.
(3) The weak law of large numbers is an example of convergence in probability.
(4) The strong law of large numbers is an example of convergence with
probability 1.
(5) The central limit theorem is an example of convergence in distribution.
TheCentral LimitTheorem
Let f(．) be a density with mean μand finite variance σ
2
.
Let
n
X be the sample mean of a random sample of size n from f(．). Then
n
X
Z
n
n
o
µ ÷
=
approaches the standard normal distribution as n approaches infinity. [Note:
It is equivalent to say that
n
X approaches a normal distribution with
expected value µ and variance o
2
/n as n approaches infinity.]
WeakLawof LargeNumbers(WLLN)
Let f(．) be a density with finite mean μ and finite variance. Let n X be
the sample mean of a random sample of size n from f(．), then for any ε> 0,
1 ] [ ÷ < ÷ < ÷ c µ c
n
X P as · ÷ n
StrongLawof LargeNumbers(SLLN)
Let f(．) be a density with finite mean μ and finite variance. Let n X be
the sample mean of a random sample of size n from f(．), then for any ε> 0,
1
lim
=
(
¸
(
¸
=
· ÷
µ
n
n
X P
Sure Convergence A.S. Convergence M.S. Convergence
Convergence in Probability Convergence in Distribution
Example 3 Let e be selected at random from the interval S = [0, 1], where we
assume that the probability that e is in a subinterval of S is equal to the length of
the subinterval. For n = 1, 2, … we define the following five sequences of random
variables:
n U
n
/ ) ( e e = , 
.

\

÷ =
n
V
n
1
1 ) ( e e ,
n
n
e W · = e e) ( ,
te e n Y
n
2 cos ) ( = ,
) 1 (
) (
÷ ÷
=
e
e
n n
n
e Z
Determine the stochastic convergence of these random sequences and identify the
limiting random variable.
[Solution]
0 ) ( ) ( = ÷
· ÷
e e U U
n
n
for every eeS. Therefore, it converges surely to a
constant 0.
e e e = ÷
· ÷
) ( ) ( V V
n
n
for every eeS. Therefore, it converges surely to a random
variable which is uniformly distributed over [0, 1].
( )  
2
1
0
2
2
2
2
3
1
) (
n
d
n n
E V E
n
= =
(
(
¸
(
¸

.

\

= ÷
}
e
e e
e e .
( )   0 ) (
2
· ÷
÷ ÷
n
n
V E e e . Thus, the sequence ) (e
n
V converges in the meansquare
sense.
) (e
n
W converges to 0 for e = 0, but diverges to infinite for all other values of e.
Therefore, it does not converge.
) (e
n
Y converges to 1 for e = 0 and e = 1, but oscillates between –1 and 1 for all
other values of e. Therefore, it does not converge.
· + ÷ = =
· ÷ n
n
n
e Z ) 0 (e
,
0 ) (
· ÷
÷
n
n
Z e
for e > (1/n).
  1 0 = > e P . Thus, ) (e
n
Z converges almost surely to 0.
( )     ( )
2 2
2
2
2
1
0
2 2 ) 1 ( 2 2
1
2
0 ) (
n
n
n n n n
n
e
n
e
d e e e E Z E
÷ ÷ ÷ ÷
÷ = = = ÷
}
e e
e e
.
As n approaches infinity, the rightmost term in the above equation approaches
infinity. Therefore, the sequence ) (e
n
Z does not converge in the mean square
sense even though it converges almost surely.
ErgodicTheorem
Adiscrete time random process {X
n
, n = 0,1, 2, …} is said to satisfy an ergodic
theorem if there exists a random variable X such that in some sense
X n X
n
n
i
i
÷·
÷
=
÷
¿
/
1
0
(123)
The type of convergence determines the type of the ergodic theorem. For example,
if the convergence is in mean square sense, the result is called a mean ergodic
theorem. If the convergence is with probability one, it is called an almost sure
ergodic theorem.
Acontinuous time random process {X(t)} is said to satisfy an ergodic theorem if
there exists a random variable X such that
X dt t X
T
T
T
÷·
÷
}
0
) (
1
(124)
where again the type of convergence determines the type of the ergodic theorem.
Note that we only require the time average to converge, however, it does not need
to converge to some constant, for example the common expectation of the random
process. In fact, ergodic theorem can hold even for nonstationary random
processes where E[X(t)] does depend on time t.
The MeanSquare Ergodic Theorem
Let {X
n
} be a random process with mean function E[X
n
] and covariance
function C
X
(k,j). (The process need not to be even weakly stationary.)
Necessary and sufficient conditions for the existence of a constant m such
that
0
1
2
1
÷
(
(
¸
(
¸

.

\

÷
¿
=
m X
n
E
n
i
i
as · ÷ n (125)
are that
m X E
n
n
i
i
n
=
¿
= · ÷ 1
) (
1
lim
(126)
and
0 ) , (
1
1 1
2 lim
=
¿¿
= = ÷·
n
i
n
k
X
n
k i C
n
(127)
The above theorem shows that one can expect a sample average to
converge to a constant in mean square sense if and only if the average of
the means converges and if the memory dies out asymptotically, that is , if
the covariance decreases as the lag increases in the sense of (127).
MeanErgodic Processes
Arandom process )} ( { t X with constant mean µ is said to be
meanergodic if it satisfies
1 ) (
2
1
=
)
`
¹
¹
´
¦
÷ }
· ÷
÷
µ
T
T
T
dt t x
T
P . (128)
Strong or Individual Ergodic Theorem
Let {X
n
} be a strictly stationary random process with E[X
n
] < ·. Then the
sample mean n X
n
i
i
/
1
¿
=
converges to a limit with probability one.
Let )} ( { t X be a widesense stationary random process with constant
mean µ and covariance function C(t). Then )} ( { t X is meanergidic if and
only if
0
2
1 ) (
2
1
2
2
÷ }
· ÷
÷


.

\

÷
T
T
T
d
T
C
T
t
t
t (129)
or equivalently,
0
2
1 ) (
1
2
0
÷ }
· ÷

.

\

÷
T
T
d
T
C
T
t
t
t (130)
Let )} ( { t X be a widesense stationary random process with constant
mean µ and covariance function C(t) and · <
}
·
· ÷
t t d C  ) (  , then
)} ( { t X is meanergidic.
Let )} ( { t X be a widesense stationary random process with constant
mean µ and covariance function C(t) and · < ) 0 ( C and 0 ) (
÷
· ÷ t
t C , then
)} ( { t X is meanergidic.
3 Examples of Stochastic Processes
iidrandomprocess
Adiscrete time random process {X(t), t = 1, 2, …} is said to be independent and
identically distributed (iid) if any finite number, say k, of random variables X(t
1
),
X(t
2
), …, X(t
k
) are mutually independent and have a common cumulative
distribution function F
X
(·) . The joint cdf for X(t
1
), X(t
2
), …, X(t
k
) is given by
( )
) ( ) ( ) (
, , , ) , , , (
2 1
2 2 1 1 2 1 , , ,
2 1
k X X X
k k k X X X
x F x F x F
x X x X x X P x x x F
k
=
s s s =
(131)
It also yields
) ( ) ( ) ( ) , , , (
2 1 2 1 , , ,
2 1
k X X X k X X X
x p x p x p x x x p
k
= (132)
where p(x) represents the common probability mass function.
Let {X
n
, n = 0, 1, 2, …} be a sequence of independent Bernoulli random variables
with parameter p. It is therefore an iid Bernoulli random process and   p X E
n
=
and   ) 1 ( p p X Var
n
÷ = .
Randomwalkprocess
Let ç
1
, ç
2
, … be integervalued random variables having common probability
mass function f(·) . Let X
0
be an integervalued random variable that is
independent of the ç
i
’
s
and let X
n
be the sum of these random variables, i.e.,
¿
=
+ =
n
i
i n
X X
1
0
ç . (133)
The sequence {X
n
, n > 0} is called a random walk process.
Let t
0
denote the probability mass function of X
0
. The joint probability of X
0
, X
1
,
. X
n
is
( )
)  ( )  ( ) (
) ( ) ( ) (
) ( ) ( ) (
, , ,
) , , , (
1 0 1 0 0
1 0 1 0 0
1 0 1 1 0 0
1 0 1 1 0 0
1 1 0 0
÷
÷
÷
÷
=
÷ ÷ =
÷ = ÷ = = =
÷ = ÷ = = =
= = =
n n
n n
n n n
n n n
n n
x x P x x P x
x x f x x f x
x x P x x P x X P
x x x x x X P
x X x X x X P
t
t
ç ç
ç ç
(134)
Thus,
)  (
)  ( )  ( ) (
)  ( )  ( )  ( ) (
) , , , (
) , , , , (
) , , ,  (
1
1 0 1 0 0
1 1 0 1 0 0
1 1 0 0
1 1 1 1 0 0
1 1 0 0 1 1
n n
n n
n n n n
n n
n n n n
n n n n
x x P
x x P x x P x
x x P x x P x x P x
x X x X x X P
x X x X x X x X P
x X x X x X x X P
+
÷
+ ÷
+ +
+ +
=
·
=
= = =
= = = =
=
= = = =
t
t
(135)
The property
)  ( ) , , ,  (
1 1 1 0 0 1 1 n n n n n n n n
x X x X P x X x X x X x X P = = = = = = =
+ + +
(136)
is known as the Markov property.
Aspecial case of random walk: the Brownian motion
Gaussianprocess
Arandom process {X(t)} is said to be a Gaussian random process if all finite
collections of the random process, X
1
=X(t
1
), X
2
=X(t
2
), …, X
k
=X(t
k
), are jointly
Gaussian random variables for all k, and all choices of t
1
, t
2
, …, t
k
.
Joint pdf of jointly Gaussian random variables X
1
, X
2
, …, X
k
:
( )
(
¸
(
¸
÷ ÷ ÷ =
÷
) (
2
1
exp
) 2 (
1
) , , (
1
1 , ,
1
m X C m X
C
x x f
T
k
k X X
k
t
(137)
where ) , , (
1 k
T
X X X = , )) ( , ), ( (
1 k
T
X E X E m = , and
(
(
(
(
¸
(
¸
=
) , ( ) , ( ) , (
) , ( ) , ( ) , (
) , ( ) , ( ) , (
1 1
2 2 2 1 2
1 2 1 1 1
k k k k
k
k
t t C t t C t t C
t t C t t C t t C
t t C t t C t t C
C
(138)
Timeseries– AR andMArandomprocesses
Awidesense stationary Autoregressive (AR(k)) model
) ( ) ( ) (
1
t i t X a t X
k
i
i
c + ÷ =
¿
=
(139)
where   0 ) ( = t X E ,   ( ) s t R s X t X R ÷ = ) ( ), ( , and ) , 0 ( ~ ) (
2
c
o c N t .
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.