Random processes

1 Discrete random processes

Definition 1.1. Let Ω be a sample space. A discrete random process/sequence is a function X : Z ×
(Ω, F, P ) → R. For a fixed n ∈ Z, X(n, .) is a random variable and for ω ∈ Ω, X(n, ω) ∈ R. We can
also define a complex valued random process as X : Z × (Ω, F, P ) → C. Some times we use the notation
Xn to denote a discrete random process.
A separable random process can be factorized as X(n, ω) = X(ω)f (n) where X(ω) is a random
variable and f (n) is a deterministic sequence.
Example 1.2 (Example 8.1.2, [2]). Let X(n, ω) = A(ω) sin(πn/10 + Θ(ω)) where A and Θ are ran-
dom variables. In this case, for each n, we have a function of two random variables given by X(n) =
A sin(πn/10 + Θ) (A, Θ being random variables and X(n) is a function of them).
Example 1.3 (Example 8.1.3, [2]). Let
(
Xn (ω), 1 ≤ n ≤ N,
X(n, ω) = (1)
0, otherwise.

Since X(n, .) = 0 except for n ∈ [1, N ], we say that X(n, .) has finite support. In specific, random vectors
are random processes with finite support.
Definition 1.4. An independent random process/sequence is one whose random variables at any time
n1 , n2 , . . . , nN are jointly independent for all positive integers N ∈ N.
Example 1.5 (Example 8.1.5, [2]). [4] The Bernoulli process is an example of a discrete random process.
Let {W (n)} be a Bernoulli process. Define
n
X
X(n) = αn−i W (i), n ≥ 1. (2)
i=1

Suppose |α| < 1. Let W (n) = 1 with probability p and W (n) = 0 with probability q = 1 − p. Then,
n n
X
n−i
X 1 − αn
E[X(n)] = α E[W (n)] = αn−i p = p .
1−α
i=1 i=1

Note further that

E[X(2)X(1)] = E[(αW (1) + W (2))W (1)] = αE[W (1)2 ] + E[W (1)W (2)] = αp + p2 .

Whereas,
E[X(2)]E[X(1)] = p2 (1 + α) 6= E[X(2)X(1)].
Therefore, X(2) and X(1) are not independent.
n
X n
X
n−i
var(X(n)) = var(α W (i)) = α2(n−i) var(W (i))
i=1 i=1
1 − α2n
= pq.
1 − α2
Observe that X(n) satisfies difference equation X(n) = αX(n − 1) + W (n). For large n, X(n) is the sum
of a large number of independent random variables. Therefore, by the CLT, it tends to Gaussian as n → ∞
n 1−α2n
with mean p 1−α
1−α and variance 1−α2 pq.

1
 2 ), then X(n) =
Pn n−i W (i) is zero
If W (1), . . . , W (n), . . . is zero mean iid Gaussian N (0, σW i=1 α
2n
mean Gaussian with variance 1−α 1−α2 W
σ 2 . The discrete time process X(t) is called Gaussian-Markov pro-
cesses1 .

Statistical specification: A random sequence {Xn } (or {X(n)}) is statistically specified by its N th
order CDF for all N ≥ 1 and for all times n, n + 1, . . . , n + N − 1, i.e., if we know

FX (xn , xn+1 , . . . , xn+N −1 ; n, n+1, . . . , n+N −1) = P (Xn ≤ xn , Xn+1 ≤ xn+1 , . . . , Xn+N −1 ≤ xn+N −1 ),
(3)
where the variables after the semicolon n, n + 1, . . . , n + N − 1, indicate the location of the N random
variables in the joint CDF. We denote the general N th order CDFs as

FX (xn1 , . . . , xnN ; n1 , . . . , nN ) or FX (xn1 , . . . , xnN ). (4)

N th order PDFs are given by

∂ N FX (xn , . . . , xn+N −1 ; n, . . . , n + N − 1)
fX (xn , . . . , xn+N −1 ; n, . . . , n + N − 1) = . (5)
∂xn ∂xn+1 . . . ∂xn+N −1

The mean of a random process is

ˆ ∞
µX (n) := E[Xn ] = xn fX (xn )dxn (6)
−∞

for a sequence of continuous random variables Xn and for a sequence of discrete random variables Xn , it is
∞
X
µX (n) := E[Xn ] = xn PX (Xn = xn ). (7)
xn =−∞

The autocorrelation function of a random sequence is defined as

ˆ ∞ˆ ∞
RXX (k, l) := E[Xk Xl∗ ] = xk x∗l fX (xk , xl ; k, l)dxk dxl . (8)
−∞ −∞

The autocovariance is defined as

KXX (k, l) := E[(Xk − µX (k))(Xl − µX (l))∗ ]. (9)

It follows from definitions that

∗
RXX (k, l) = RXX (l, k) (10)
∗
KXX (k, l) = KXX (l, k). (11)

Furthermore,
KXX (k, l) = RXX (k, l) − µX (k)µ∗X (l). (12)
2 (n) = K
The variance σX XX (n, n).
Note that RXX (k, k) = E[|X(k)|2 ] ≥ 0.

Example 1.6. Consider a Bernoulli process {X(n)}. Then, µX (n) = E[X(n)] = p, RXX (k, l) =
E[X(k)X ∗ (l)] = E[X(k)]E[X(l)] = p2 and KXX (k, l) = 0.
1
Markovianness comes from the dependence of X(n) only on its past sample X(n − 1), Gaussian because X(n) is Gaussian
when W (n) is Gaussian

2
Definition 1.7. A random process {X(n)} is called Gaussian random process if its N th order CDFs (PDFs)
are jointly Gaussian2 for N ≥ 1.

Definition 1.8. A random sequence is said to have independent increments if for all integers n1 < n2 <
. . . < nN , the increments X(n1 ), X(n2 ) − X(n1 ), . . . , X(nN ) − X(nN −1 ) are jointly independent for all
integers N > 1.

Example 1.9 (Random walk). Let {W (n)} be a Bernoulli process with p beingPthe probability of heads. Let
W (k) = s for outcome ω = H and W (k) = −s for ω = T . Define X(n) = nk=1 W (k) with W (0) = 0.
At each time step, a step of size s is taken to the right or to the left. After n time steps, we will be at a
position rs for some integer r. If there are k successes and n − k failures, then
n+r
rs = ks − (n − k)s = (2k − n)s ⇒ k =
2
for those values of r which make k an integer. Let P (H) = P (T ) = 12 . Then,

n+r
P (X(n) = rs) = P ( successes)
( 2
n

1
(n+r)/2 2n , (n + r)/2 an integer r ≤ n
= (13)
0, otherwise.

Since X(n) = W (1) + . . . + W (n) and W (i)s are jointly independent,

n
X
E[X(n)] = E[W (i)] = 0,
i=1

and
n n
X X 1
E[X 2 (n)] = E[W 2 (i)] = (s2 + (−s)2 ) = ns2 . (14)
2
i=1 i=1

Let X̃(n) = √1 X(n),

then by the CLT, X̃(n) converges to Gaussian distribution N (0, s2 ).
n
n

k n−k n

n+r n−r
Let P (H) = p and P (T ) = q = 1−p. Therefore, P (X(n) = rs) = (n+r)/2 p q = (n+r)/2 p 2 q 2 .
The mean is
Xn
E[X(n)] = E[W (i)] = ns(p − q), (15)
i=1

and the variance is

n
X X
2 2
var(X(n)) = E[X (n)] − E[X(n)] = E[W 2 (i)] + 2 E[W (i)W (j)] − n2 s2 (p − q)2
i=1 1≤i<j≤n
2 2 2 2 2 2
= ns (p + q) + n(n − 1)s (p − q) − n s (p − q)
= ns2 − ns2 (p − q)2
= ns2 (1 − (p − q)2 ) = ns2 (1 − p + q)(1 + p − q) = 4ns2 pq. (16)

Notice that we can also use the property of variance of the sum of iid random variables to find

var(X(n)) = nvar(W (i)) = n(s2 − s2 (p − q)2 ).

2
two random variables X, Y are jointly Gaussian if their joint PDF is
x−µ (X−µX )(Y −µY ) y−µ
1√ − 1 {( σ X )2 −2ρ( )+( σ Y )2 }
fXY (x, y) = e 2(1−ρ2 ) X σX σY Y where ρ is the correlation coefficient. For n
2πσX σY 1−ρ 2
1 T
K −1 (x−µ)
Gaussian random variables with random vector X covariance matrix K, it is fX (x) = n
1
√ e− 2 (x−µ) .
(2π) 2 det(K)

3
For a biased random walk,
k
X X
2
E[X (k)] = E[W 2 (i)] + 2 E[W (i)]E[W (j)]
i=1 1≤i<j≤k

= ks2 (p + q) + k(k − 1)s (p − q)2

2

= ks2 (1 + (k − 1)(p − q)2 ). (17)

Furthermore, for k ≤ l,

RXX (k, l) = E[X(k)X(l)] = E[X(k)(X(l) − X(k) + X(k))]

= E[X 2 (k)] + E[X(k)]E[X(l) − X(k)]
= ks2 (1 + (k − 1)(p − q)2 ) + ks(p − q)(l − k)s(p − q)
= ks2 (1 + (k − 1)(p − q)2 ) + k(l − k)s2 (p − q)2 . (18)

For p = q = 21 , RXX (k, l) = ks2 =min(k, l)s2 .

We can write down the following difference equation for the probability of being at position js at time n

P (X(n) = js) = pP (X(n − 1) = (j − 1)s) + qP (X(n − 1) = (j + 1)s), P (X(0) = js) = δj,0 . (19)

Example 1.10. Consider an iid discrete Gaussian process {W (n)} with zero mean and variance σ 2 . Then,

 1 ´ ∞ x e− 2σk2 dx ´ ∞ x e− 2σl2 dx = 0, If k 6= l,

 x2 x2

2 k k −∞ l l
RXX (k, l) = E[X(k)X(l)] = 2πσ −∞ (20)
 1
 ´ ∞ 2
x2
− 2k
2
2πσ 2 −∞ k
x e 2σ dxk = σ , If k = l.

Thus,
RXX (k, l) = σ 2 δ(k − l). (21)
Example 1.11 (Example 8.1.12, [2]). Let W (n) be real valued Gaussian iid with mean µW (n) = 0 and
RW W (k, l) = σ 2 δ(k − l), σ > 0. This forms an independent Gaussian random sequence. Let X(n) =
W (n) + W (n − 1). Then, X(n) is also Gaussian (linear transformation of Gaussian random variables is
Gaussian). Note that µX (n) = 0 and

RXX (k, l) = E[(W (k) + W (k − 1))(W (l) + W (l − 1))]

= RW W (k, l) + RW W (k, l − 1) + RW W (k − 1, l) + RW W (k − 1, l − 1)
= σ 2 (δ(k − l) + δ(k − l + 1) + δ(k − l − 1) + δ(k − l)).

Note that X(n) is not an independent random sequence but has independent increments.
Cross correlation of two random sequences X(n), Y (n) is defind as

RXY (k, l) = E[X(k)Y ∗ (l)] (22)

and cross covariance as

KXY (k, l) = E[(X(k) − µX (k))(Y (l) − µY (l))∗ ]. (23)

Note that
RXY (k, l) = RY∗ X (l, k), KXY (k, l) = RXY (k, l) − µX (k)µ∗Y (l). (24)
Definition 1.12. Two processes {X(k)}, {Y (l)} are orthogonal if

RXY (k, l) = 0, ∀k, l (25)

and are uncorrelated if

CXY (k, l) = 0, ∀k, l. (26)

4
Definition 1.13 (Strict Sense Stationary sequence). If for all orders l and for all shift parameters k, the joint
CDFs of (Xn1 , Xn2 , . . . , Xnl ) and (Xn1 +k , Xn2 +k , . . . , Xnl +k ) are the same, then the random sequence is
said to be stationary or strict sense stationary (SSS) i.e., for all l ≥ 1,

FX (xn1 , . . . , xnl ) = FX (xn1 +k , . . . , xnl +k ) (27)

for all −∞ < k < ∞.

Example 1.14. Consider the random sequence consisting of iid random variables τ (n) for n ≥ 1 each with
an exponential PDF. This is a strict sense stationary process. These are the interarrival times of a Poisson
process.
Consider the arrival time T (n) = ni=1 τ (i). Then, T (1) = τ (1) with distribution fT (t; 1) = fτ (t) =
P
λe−λt u(t) and fT (t; 2) = fT (t)∗fT (t) = λ2 te−λt u(t) 6= fT (t; 1). Thus, T (n) is not strict sense stationary.

Definition 1.15. A stochastic process {Xn } is called wide sense stationary (WSS) if

1. The mean E[Xn ] is constant for all integers n, −∞ < n < ∞, µX (n) = µX (0).

2. For all k, l, −∞ < k, l < ∞, RXX (k, l) = RXX (k − l).

Theorem 1.16. SSS ⇒ WSS.

The random walk is not WSS hence not SSS.

Point processes: A point process is a set of random points ti on the time axis. To every point process, we
can associate a stochastic process x(t) equal to the number of points ti in the interval (0, t). An example is
the Poisson process. To every point process ti , we can associate a sequence of random variables Zn such
that Z1 = t1 , Z2 = t2 − t1 , . . . , Zn = tn − tn−1 . This sequence is called a renewal process.

2 Continuous random processes

Definition 2.1. Let Ω be a sample space. A continuous stochastic process is a function X : R × Ω → R.
For a fixed t ∈ R, X(t, .) is a random variable and for ω ∈ Ω, X(t, ω) ∈ R. We can also define a complex
valued random process as X : R × Ω → C. Some times we use the notation Xt to denote the continuous
time stochastic process where for a fixed t, Xt is a random variable.

Example 2.2 (Example 9.1.2, [2]). {X(t)} where X(t, ω) = A(ω) sin(ω0 t + Θ(ω)) is a stochastic process
where A(ω) and Θ(ω) are random variables and X(t) is a function of these random variables for all t.

Statistical specification: The definitions of autocorrelation, autocovariance, cross-correlation, cross-

covariance, SSS, WSS extend naturally.

µX (t) = E[X(t)], RXX (t1 , t2 ) = E[X(t1 )X ∗ (t2 )], KXX (t1 , t2 ) = E[(X(t1 ) − µX (t1 ))(X(t2 ) − µX (t2 ))∗ ]. (28)

We will also use the notations below

µX (t) = E[Xt ], RXX (t1 , t2 ) = E[Xt1 Xt∗2 ], KXX (t1 , t2 ) = E[(Xt1 − µX (t1 ))(Xt2 − µX (t2 ))∗ ] (29)

where both (28) and (29) mean the same thing.

Definition 2.3. [2] A function of two variables g(t, s) is positive semidefinite if for all N > 0 and all
t1 < · · · < tN , and all complex constants a1 , . . . , aN ,
N X
X N
ai a∗j g(ti , tj ) ≥ 0.
i=1 j=1

5
 Using the identity X X
0 ≤ E[| ai X(ti )|2 ] = ai a∗j E[X(ti )X(tj )],
i i,j

it follows that X
ai a∗j RXX (ti , tj ) ≥ 0. (30)
i,j

It turns out that the converse is also true.

The continuous time autocorrelation function satisfies similar properties to those satisfied by their dis-
crete counterparts.
e−λt (λt)k
Example 2.4. Consider a Poisson arrival process {X(t)} ∼ P (λt) where P (X(t) = k) = k! ,
k = 0, 1, . . .. Then, µX (t) = E[X(t)] = λt, var(X(t)) = λt, E[X(t)2 ] = λt + (λt)2 .

RXX (t1 , t2 ) = E[X(t1 )X ∗ (t2 )] = E[X(t1 )X(t2 )] = E[X(t1 )(X(t2 ) − X(t1 ) + X(t1 ))]
= E[X(t1 )(X(t2 ) − X(t1 ))] + E[X 2 (t1 )]
= E[X(t1 )]E[(X(t2 ) − X(t1 ))] + E[X 2 (t1 )]
= (λt1 )λ(t2 − t1 ) + λt1 + λ2 t21
= λ2 t1 t2 + λt1 , t2 ≥ t1 (31)
2
= λ t1 t2 + λt2 , t2 < t1 (32)
2
= λ t1 t2 + λ min(t1 , t2 ). (33)

and KXX (t1 , t2 ) = λ2 t1 t2 + λ min(t1 , t2 ) − λ2 t1 t2 = λ min(t1 , t2 ).

The Poisson counting process is also denoted by {N (t)} where N (t) denotes the total number of arrival
counts up to time t. It can be expressed as
∞
X
N (t) := u(t − Y (n)), (34)
n=1

whee Y (n) is the time of nth arrival and u(t) is the unit step signal. Notice that

P (N (t) = n) = P (Y (n) ≤ t, Y (n + 1) > t). (35)

For a fixed t, N (t) is a Poisson random variable with parameter λt. Notice that N (t) is continuous time
random process but at time t, the random variable N (t) is a discrete Poisson random variable with parameter
λt. From the independent increments property of the Poisson process, over a given interval (ta , tb ]

(λ(tb − ta ))n
P (N (tb ) − N (ta ) = n) = e−λ(tb −ta ) , n ≥ 0. (36)
n!
Suppose we want to find the probability that P (N (t1 ) = n1 , N (t2 ) = n2 ), t2 > t1 . Then, from the
independent increment property,

P (N (t1 ) = n1 , N (t2 ) = n2 ) = P (N (t1 ) = n1 )P (N (t2 ) − N (t1 ) = n2 − n1 )

(λt1 )n1 −λt1 (λ(t2 − t1 ))n1 −n1 −λ(t2 −t1 )

= e e .
n1 ! (n2 − n1 )!
Remark 2.5. Notice that the Poisson process {N (t)} is continuous time but at each time t, the random
variable N (t) is a discrete valued Poisson random variable. The interarrival times {T (n)} is a discrete
process where for a fixed n, the random variable T (n) takes continuous values i.e., T (n) is an exponential
random variable. Furthermore, the arrival times {Y (n)} for the nth arrival is also a discrete process taking
continuous values for a fixed n. For a fixed n, Y (n) has an Erlang distribution of order n.

6
 Note that arrival sequence Y (n) can also be interpreted as the waiting time for the nth arrival (for
example, a person in a queue at nth position). We are interested in P (Y (n) ≤ t). This is done in the notes
on Bernoulli-Poisson process. It is clear that P (Y (n) ≤ t) = P (N (t) ≥ n) and using Poisson random
variables, we can find P (N (t) ≥ n).

Example 2.6. [1] Consider a Poisson process {N (t)} with parameter λ and t1 < t2 . We want to find
P (N (t1 ) = k|N (t2 ) = n).

P (N (t1 ) = k, N (t2 ) = n) P (N (t1 ) = k, N (t2 − t1 ) = n − k)

P (N (t1 ) = k|N (t2 ) = n) = =
P (N (t2 ) = n) P (N (t2 ) = n)
e−λt1 (λt1 )k e−λ(t2 −t1 ) (λ(t2 − t1 ))n−k n!
=
k! (n − k)! e−λt2 (λt2 )n
 
n t1 k t1
= ( ) (1 − )n−k , k = 0, 1, . . . , n (37)
k t2 t2

which is the binomial random variable with parameters (n, tt12 )

Example 2.7 ([1]). Let {X1 }, {X2 } be two independent Poisson processes with parameters λ1 t and λ2 t
respectively. Let {Y (t)} = {X1 (t)} − {X2 (t)}. Then,
∞
X
P (Y (t) = n) = P (X1 (t) = n + k)P (X2 (t) = k)
k=0
∞
X (λ1 t)n+k −λ2 t (λ2 t)k
= e−λ1 t e
(n + k)! k!
k=0
∞ √
λ1 X (t λ1 λ2 )n+2k
= e−(λ1 +λ2 )t ( )n/2
λ2 k!(n + k)!
k=0
λ1 p
= e−(λ1 +λ2 )t ( )n/2 I|n| (2 λ1 λ2 t), n = 0, ±1, ±2, . . . (38)
λ2
where
∞
X (x/2)n+2k
I|n| (x) := (39)
k!(n + k)!
k=0

represents the modified Bessel function of order n. It turns out that E[Y (t)] = (λ1 − λ2 )t and var(Y (t)) =
(λ1 + λ2 )t. Note that Y (t) is not Poisson.

Definition 2.8 (Strict Sense Stationary process). A stochastic process {Xt } is strict sense stationary (SSS)
(or just stationary) if it has the same nth order CDF as Xt+T i.e.,

FX (xt1 , . . . , xtn ) = FX (xt1 +T , . . . , xtn +T ) (40)

for all T and for all t1 , . . . , tn .

This mean that the set of random variables Xt1 , . . . , Xtn and the set of random variables Xt1 +T , . . . , Xtn +T
have the same joint CDF for all n ≥ 1 and all T ≥ 0. If the CDF is differentiable, then

fX (xt1 , . . . , xtn ) = fX (xt1 +T , . . . , xtn +T ). (41)

This implies that fX (x; t) = fX (x, 0), therefore, uX (t) = uX (0) is a constant. Furthermore,

fX (xt1 , xt2 ; t1 , t2 ) = fX (xt1 −t2 , x0 ; t1 − t2 ; 0),

which implies that RXX (t1 , t2 ) = E[X(t1 )X ∗ (t2 )] = RXX (t1 − t2 , 0) = RXX (t1 − t2 ).

7
Definition 2.9. A stochastic process {Xt } is WSS if µX (t) is a constant and RXX (t1 , t2 ) = RXX (t1 − t2 ).
Is the Poisson process SSS? Is it WSS?
Example 2.10 (Example 9.1.5, [2]). Consider {Xt } where Xt = A sin(ω0 t + Θ) where A, Θ are indepen-
dent and Θ is uniformly distributed over [−π, π]. Then,
ˆ π
1
µX (t) = E[A]E[sin(ω0 t + Θ)] = µA . sin(ω0 t + θ)dθ = 0
2π −π
and

RXX (t1 , t2 ) = E[A2 sin(ω0 t1 + Θ) sin(ω0 t2 + Θ)]

= E[A2 ]E[sin(ω0 t1 + Θ) sin(ω0 t2 + Θ)]
1
= E[A2 ] cos ω0 (t1 − t2 ).
2
Similarly,

RXX (t1 − t2 , 0) = E[A2 sin(ω0 (t1 − t2 ) + Θ) sin(Θ)]

1
= E[A2 ]E[cos(ω0 (t1 − t2 )) + cos(ω0 (t1 − t2 ) + 2Θ)]
2
1
= E[A2 ] cos ω0 (t1 − t2 ).
2
Since RXX (t1 , t2 ) = RXX (t1 − t2 , 0) = RXX (t1 − t2 ), X(t) is WSS.
A Gaussian process {Xt } is one where the samples Xt1 , . . . , Xtn are jointly Gaussian for all n ∈ N.
Consider a zero mean iid Gaussian processes {Xt } with variance σ 2 . Then, for t1 6= t2 ,

RXX (t1 , t2 ) = E[Xt1 Xt2 ] = E[Xt1 ]E[Xt2 ] = 0,

and for t1 = t2 ,
RXX (t1 , t1 ) = E[Xt1 Xt1 ] = E[Xt21 ] = σ 2 .
Therefore, RXX (t1 , t2 ) = σ 2 δ(t1 −t2 ). This is called the white Gaussian process (or white Gaussian noise).
Theorem 2.11. [1] For Gaussian processes, WSS ⇒ SSS.
Definition 2.12. A continuous time Markov process {X} has a conditional PDF that satisfies

fX (xn |xn−1 , . . . , x1 ; tn , . . . , t1 ) = fX (xn |xn−1 ; tn , tn−1 ), ∀x1 , . . . , xn , ∀t1 < · · · < tn , ∀n > 0.

and joint PDF given by

fX (xn , . . . , x1 ; tn , . . . , t1 ) = fX (xn |xn−1 ; tn , tn−1 ) · · · fX (x2 |x1 ; t2 , t1 )fX (x1 ; t1 ). (42)

Examples of Markov processes are random walks, Poisson processes, Wiener process. A random process
which is both Gaussian and Markov is called Gauss-Markov process.
Example 2.13 (Random telegraph/switching signal, [2]). Let {X(t)} be a stochastic process such that
X(0) = ±1 with equal probability and t ≥ 0. Consider a Poisson arrival process {N (t)} independent
of X(0) of rate λ. At each arrival, we switch the sign of X(t). For example, at the first arrival Y (1), we
change the sign for the first time, then again at Y (2) and so on i.e., X(t) = sign(X(0))(−1)N (t) where
N (t) takes values 0, 1, 2, . . .. This model can be used for modeling a random switching phenomena.
From symmetry of the signal X(t), it follows that µX (t) = 0

E[X(t)] = E[X(0)]E[(−1)N (t) ] = 0.

8
Notice that
∞ k ∞
X
k −λt (λt) −λt
X (λt)k
E[(−1) N (t)
]= (−1) e =e (−1)k = e−2λt . (43)
k! k!
k=0 k=0

Note that P (X(t) = 1) = P (X(t) = −1) = 12 . Therefore,

RXX (t1 , t2 ) = E[X(t1 )X(t2 )] = 1.1P (X(t1 ) = 1, X(t2 ) = 1) + 1.(−1)P (X(t1 ) = 1, X(t2 ) = −1)
+(−1).1P (X(t1 ) = −1, X(t2 ) = 1) + (−1).(−1)P (X(t1 ) = −1, X(t2 ) = −1)

1
= P (X(t2 = 1)|P (X(t1 ) = 1)) + P (X(t2 = −1)|P (X(t1 ) = 1)) +
2

P (X(t2 = 1)|P (X(t1 ) = −1)) + P (X(t2 = −1)|P (X(t1 ) = −1)) (44)

where we have used the conditional probabilities and the fact that P (X(t1 ) = 1) = P (X(t1 ) = −1) = 21
Now P (X(t2 = 1)|P (X(t1 ) = 1)) = P (X(t2 = −1)|P (X(t1 ) = −1)) is the probability that there are
even number of zero crossings in the interval (t1 , t2 ] whereas P (X(t2 = −1)|P (X(t1 ) = 1)) = P (X(t2 =
1)|P (X(t1 ) = −1)) is the probability that there are odd number of zero crossings in (t1 , t2 ] (t2 > t1 ). Let
τ := t2 − t1 . Then,
X (λτ )k X (λτ )k
RXX (t1 , t2 ) = e−λτ − e−λτ
k! k!
even
k≥0,k k≥0,k odd
X (λτ )k
= e−λτ (−1)k = e−2λτ , τ > 0. (45)
k!
k≥0

By symmetry, it follows that for τ < 0, RXX (t1 , t2 ) = e−2λτ . Therefore, in general, for a random switching
signal,
RXX (t1 , t2 ) = e−2λ|τ | . (46)
Is the random switching process WSS?
Theorem 2.14. [1] Given an arbitrary function µ(t) and a positive definite function K(t1 , t2 ), there exists
a Gaussian process with mean µ(t) and autocovariance K(t1 , t2 ).
Remark 2.15. One can have a zero mean Gaussian process with RXX (t1 , t2 ) = e−λ|t1 −t2 | . Thus, the
second order statistics are the same for this process as well as the random switching signal. However, the
realizations of the two processes are quite different. Thus, the second order statistics do not capture all
properties of stochastic processes.
Wiener process: Recall the random walk process. We consider the continuous analogue of random
walk. Let
∞
X
XT (t) := W (k)u(t − kT ) (47)
k=1
where W (k) is s with probability 0.5 and −s with probability 0.5 (i.e., Bernoulli random variable) and u(t)
is the unit step signal. Then, XT (nT ) = X(n), the random walk sequence. The Wiener process is given by
limT →0 XT (t). Recall that E[XT (nT )] = 0 and E[XT2 (nT )] = ns2 . Let t = nT , therefore,
s2
var(XT (t)) = E[XT2 (nT )] = t . (48)
T
Set s2 = αT where α > 0. As T → 0, var(XT (t)) = αt and µX (t) = 0. The Wiener process inherits
the independent increment property of random walks. By the CLT, we get a limiting Gaussian distribution.
Therefore,
1 x2
fX (x; t) = √ e− 2αt , t > 0 (49)
2παt

9
The Wiener process is an example of a Gaussian random process (where nth order PDFs are jointly Gaussian
for all n). Suppose t2 ≥ t1 . The increment X(t2 ) − X(t1 ) is independent of X(t1 ) and both have zero
mean. Therefore,

RXX (t1 , t2 ) = E[X(t1 )X(t2 )] = E[X(t1 )(X(t2 ) − X(t1 ) + X(t1 ))]

= E[X(t1 )(X(t2 ) − X(t1 ))] + E[X 2 (t1 )]
= E[X 2 (t1 )] = αt1 . (50)

Therefore,
RXX (t1 , t2 ) = αmin(t1 , t2 ). (51)
Thus, the Wiener process is not stationary or WSS.
d
The derivative of the Wiener process gives white Gaussian process. Let W (t) = dt X(t). Then,
d
µW (t) = dt µX (t) = 0. The covariance KW W is given by
   
∂ ∂ ∂ ∂
KW W (t1 , t2 ) = KXX (t1 , t2 ) = α min(t1 , t2 )
∂t1 ∂t2 ∂t1 ∂t2
 (
∂ ∂ αt2 , if t2 ≤ t1
=
∂t1 ∂t2 αt1 , if t2 ≥ t1
(
∂ α, if t2 ≤ t1
=
∂t1 0, if t2 > t1
∂
= αu(t1 − t2 )
∂t1
= αδ(t1 − t2 ). (52)

Note that RW W (t1 , t2 ) = σ 2 δ(t1 − t2 ) = KW W (t1 , t2 ) for white noise.

Definition 2.16. Let {Xt }, {Yt } be two random processes. They are

• uncorrelated if RXY (t1 , t2 ) = µX (t1 )µY (t2 ) for all t1 , t2 .

• orthogonal if RXY (t1 , t2 ) = 0 for all t1 , t2 .

• independent if for all positive integers n,

FXY (x1 , y1 , . . . , xn , yn ; t1 , . . . , tn ) = FX (x1 , . . . , xn ; t1 , . . . , tn )FY (y1 , . . . , yn ; t1 , . . . , tn ).

• jointly WSS if both {Xt } and {Yt } are WSS and RXY (t1 , t2 ) = RXY (t1 − t2 ).

For jointly Gaussian processes, uncorrelated and independence are equivalent.

A vector stochastic process consists of a vector of n stochastic processes or a vector of n random
variables at each time instance t. The mean for a vector stochastic process {Xt } is given by µXt = E[Xt ] ∈
Rn . The autocorrelation and autocovariance functions for a vector stochastic process Xt are given by

RXX (t1 , t1 ) = E[Xt1 XT T

t2 ], KXX (t1 , t1 ) = E[(Xt1 − µXt1 )(Xt2 − µXt2 ) ]. (53)

Theorem 2.17. [4] Let {φn (t)}, n ∈ Z be a sequence of orthonormal real

P valued functions. Let {Xn } be a
sequence of independent random variables, Xn ∼ N (0, σn2 ). Suppose ln=−l σn2 φ2n (t) < ∞ as l → ∞ for
all t. Then,
Xl
Xt = lim Xn φn (t), t ∈ R (54)
l→∞
n=−l

is a Gaussian process.

10
 One can in general choose φn to be the Fourier basis functions such as complex exponentials or Dirac-
delta functions or integer translates of sinc function. Notice that the Gaussian process in (54) is a seperable
process with Xn being stochastic and φn being deterministic.

Definition 2.18. [4] A circularly symmetric Gaussian process {X} is a complex stochastic process such
that for all positive integers k, and t1 , . . . , tk ∈ R,the complex n random variables X(t1 ), . . . , X(tk ) are
circularly symmetric Gaussian random variables.

To generate a circularly symmetric Gaussian process, one can start with a sequence of independent
circularly symmetric Gaussian random variables {Xn ∼ CN (0, σn2 )}, n ∈ Z and a sequence of orthonormal
functions φn and use Theorem 2.17

Definition 2.19 (Ergodicity). A WSS random process is ergodic in the mean if the time average of {X}
converges to the ensemble average E[X(t)] = µX i.e.,
ˆ T
1
M̂ (T ) := X(t)dt → µX (m.s.) as T → ∞. (55)
2T −T

Refer Theorem 10.4-1 [2] for conditions to check ergodicity using KXX .

Definition 2.20. A stochastic process X(t) is wide sense periodic if there exists T > 0 such that µX (t) =
µX (t + T ) for all t and KXX (t1 , t2 ) = KXX (t1 + T, t2 ) = KXX (t1 , t2 + T ) for all t1 , t2 . The smallest
such T is called period.
A stochastic process X(t) is wide sense cyclostationary if there exists T > 0 such that µX (t) = µX (t +
T ) for all t and KXX (t1 , t2 ) = KXX (t1 +T, t2 +T ) for all t1 , t2 . It is said to be strict sense cyclostationary
if

FX (x(t1 ), . . . , x(tn ); t1 , . . . , tn ) = FX (x(t1 + mT ), . . . , x(tn + mT ); t1 + mT, . . . , tn + mT ) (56)

for all m ∈ Z.

References
[1] A. Papoulis, U. Pillai, Probability, Random variables and Stochastic processes, 4th edition, 2002.

[2] H. Stark, J. Woods, Probability and Random processes with applications to Signal processing, 4th
edition, 2012.

[3] B. Hajek, Random processes for engineers, 2015.

[4] R. Gallager, Stochastic processes: Theory for applications, 2013.

11