Chapter 8

RANDOM SIGNALS

Signals can be divided into two main categories - deterministic and random. The
term random signal is used primarily to denote signals, which have a random in its
nature source. As an example we can mention the thermal noise, which is created
by the random movement of electrons in an electric conductor. Apart from this,
the term random signal is used also for signals falling into other categories, such
as periodic signals, which have one or several parameters that have appropriate
random behavior. An example is a periodic sinusoidal signal with a random phase
or amplitude. Signals can be treated either as deterministic or random, depending
on the application. Speech, for example, can be considered as a deterministic signal,
if one specific speech waveform is considered. It can also be viewed as a random
process if one considers the ensemble of all possible speech waveforms in order to
design a system that will optimally process speech signals, in general.
The behavior of stochastic signals can be described only in the mean. The
description of such signals is as a rule based on terms and concepts borrowed from
probability theory. Signals are, however, a function of time and such description
becomes quickly difficult to manage and impractical. Only a fraction of the signals,
known as ergodic, can be handled in a relatively simple way. Among those signals
that are excluded are the class of the non-stationary signals, which otherwise play
an essential part in practice.
Working in frequency domain is a powerful technique in signal processing. While
the spectrum is directly related to the deterministic signals, the spectrum of a ran-
dom signal is defined through its correlation function. This is a “natural” conse-
quence of the “uncertainty”, which is characteristic to random signals.
Random signals can be both analog and digital. The values of digital signals are
represented with a finite number of digits. This implies that the stochastic terms
used are different for the two signal categories. In the following, however, we will
assume that the values of digital signals are represented with infinite precision, and
that we can describe the two types of signals in similar way. Often such signals are
called ”discrete time signals” rather than digital signals to emphasize the fact that
the signal values are represented with infinite precision.
The main part of the signals that will be processed will be real. This does not
mean that stochastic signals cannot be complex. Complex random signals can be
analyzed the same way as real random signals with very few changes.

8.1 Random variables

In this section we set the framework for the description of the random processes
and the subsequent signal processing. Regarding further details and proofs, the
inquisitive reader is encouraged to read some of the standard texts on probability
and statistics. We assume that the fundamental concepts of the probability theory,
such as “experiment”, “event”, “outcome”, etc. are familiar to the reader.

8-1
8-2 Random signals Chapter 8

8.1.1 Definitions
In this section we will briefly review some of the key concepts and terms necessary
to build the theory around stochastic signal processing.
In the study of probability, any process of observation is referred to as an ex-
periment. The results of an observation are called outcomes of the experiment.
If the outcomes of an experiment cannot be predicted, then it is called random
experiment.
The set of possible outcomes of a random experiment is called the sample space.
An element in the sample space is called a sample point. Each outcome of a random
experiment corresponds to a sample point.
Subsets of the sample space are called events, and events consisting of a single
element (sample point) are called elementary events.
Probability is a number, associated with events according to some appropriate
probability law. The probability assigned to the event A from the sample space S
A ∈ S is denoted as P (A) and has a value between 0 and 1:

P (A), 0 6 P (A) 6 1

In order to be a valid probability assignment, the following three axioms

must be satisfied:
1. P (A) > 0 for every event A ∈ S.
2. P (S) = 1 for the certain event S.
3. For any two mutually exclusive events A1 and A2 ,

P (A1 ∪ A2 ) = P (A1 ) + P (A2 ).

By associating a number to the possible outcomes of an experiment, we define

a random variable. A random variable can also be considered as a mathematical
function that maps the outcomes of random experiment to numbers. For example,
a random variable can be used to describe the process of rolling a fair die and the
possible outcomes { 1, 2, 3, 4, 5, 6 }. Another random variable might describe the
possible outcomes of picking a random person and measuring his or her height. Such
random variables will be denoted with capital letters such as X, Y and Z. The set
of values that are encountered in signal processing is either the set of real numbers
R (analog signals), or the finite subset of rational numbers (digital signals).
In signal processing applications, it is the probabilistic description of the random
variable, rather than the statistical characterization of events in the sample space,
that is generally of interest. It is therefore more convenient to have a probability
law assigned to the random variable itself. For a real-valued random variable X,
one such statistical characterization is the probability distribution function, given
by:
WX (ξ1 ) = P {X 6 ξ1 } (8.1.1)

It is a function of the parameter ξ, and ξ1 is a fixed value of the parameter. It

follows directly from the definition that

0 6 WX (ξ) 6 1, (8.1.2)

and that WX (ξ) never decreases. We have also that:

P {ξ1 < X 6 ξ2 } = WX (ξ2 ) − WX (ξ1 ), (8.1.3)

as shown in the example in Fig. 8.1.

Section 8.1. Random variables 8-3

Probability distribution function (analog) Probability distribution function (discrete)

1 1

0.8 0.8
Wx(ξ)

Wx(ξ)
0.6 0.6

0.4 0.4

0.2 0.2

0 0
−1 0 1 −1 0 1
ξ ξ

Figure 8.1. Examples of two different probability distribution functions.

The multi-dimensional probability distribution function of several random vari-

ables is defined in a similar way. For two random variables X and Y , the two-
dimensional distribution is defined as:

WXY (ξ1 , η1 ) = P {X 6 ξ1 , Y 6 η1 }. (8.1.4)

Another useful statistical characterization of a random variable is the probability

density function, wX (ξ), which is the derivative of WX (ξ):

Z ξ
d
wX (ξ) = WX (ξ), WX (ξ) = wX (θ)dθ (8.1.5)
dξ −∞

For the case, in which ξ is defined within the subset of the rational numbers,
wX (ξ) consists exclusively of δ-functions, because WX (ξ) has a stair-case shape.
If ξ is defined over the set of rational numbers, wX (ξ) is continuous, except for
those points, where WX (ξ) has a discontinuity point. Figure 8.2 shows two exam-
ples of probability density functions that correspond to the probability distribution
functions shown in Fig. 8.1. For the case, in which ξ ∈ R we have the following
relations:
Z ∞
P {ξ1 < X 6 ξ1 + dξ1 } = wX (ξ1 )dξ1 , and wX (ξ)dξ = 1. (8.1.6)
−∞

Multidimensional probability density functions are defined in a similar manner.

Let’s consider the two-dimensional case with two random variables X and Y . We
have:
∂ 2 WXY (ξ, η)
wXY (ξ; η) = . (8.1.7)
∂ξ∂η
In many cases, the time enters as a parameter in the probability distribution-
and density functions. If this is important in a given context, then this dependence
will be explicitly written, e.g.:
wX (ξ, t).
The expected value -the average- of a random variable is given by:
Z ∞
E{X} = ξwX (ξ)dξ. (8.1.8)
−∞
8-4 Random signals Chapter 8

Prbability density function (analog) Probability density function (diskret)

0.015 0.2

0.15
0.01
wx(ξ)

wx(ξ)
0.1
0.005
0.05

0 0
−2 −1 0 1 2 −2 −1 0 1 2
ξ ξ

Figure 8.2. Examples of two different probability density functions.

E{X} is also referred as the first order moment of X. The nth moment of X is
defined as: Z ∞
E{X n } = ξ n wX (ξ)dξ. (8.1.9)
−∞
The n’th central moment is given by:
Z ∞
E{(X − E{X})n } = (ξ − E{X})n wX (ξ)dξ. (8.1.10)
−∞

The second central moment, called also variance is:

Z ∞
2
σ {X} = (ξ − E{X})2 wX (ξ)dξ. (8.1.11)
−∞

If X and Y are two random variables, then:

Z ∞Z ∞
E{XY } = ξηwXY (ξ, η)dξdη (8.1.12)
−∞ −∞

The central moment

E{(X − E{X})(Y − E{Y })} (8.1.13)
is called also covariance of X and Y .
If k, q, and a are constants, it follows that:
E{kX} = kE{X} and E{kX + qY } = kE{X} + qE{Y }
E{X + a} = E{X} + a (8.1.14)
2 2 2
σ {kX + a} = k σ {X}.
Furthermore, for the variance we have:
σ 2 {X} = E{(X − E{X})2 } = E{X 2 } − E 2 {X}. (8.1.15)
The covariance is given by:
E{(X − E{X})(Y − E{Y })} = E{X}E{Y } − E{X}E{Y }. (8.1.16)
If V is a complex random variable, i.e. V = X + jY , then E{V } = E{X} +
jE{Y }. The variance of V is:
σ 2 {V } = E{|V − E{V }|2 }. (8.1.17)
Similarly, the covariance of two complex random variables V and U is:
E{(V − E{V })(U ∗ − E{U ∗ })}. (8.1.18)
Section 8.1. Random variables 8-5

Distribution ξ-region wx (ξ) E{X} σ 2 {X}

1 d+c (d−c)2
Rectangular c<ξ<d 2 12
d−c
(ξ − m)2
 
1
Normal (Gauss) −∞ < ξ < +∞ √ exp − m s2
2πs2  2s2
2

ξ ξ
s π2 (2 − π2 )s2
p
Rayleigh 0 < ξ < +∞ exp − 2
s2 2s
 
1 1 ξ
χ2 (N = 1) 0 < ξ < +∞ √ √ exp − 2 s2 2s4
2πs2 ξ 2s
 
1 N −2 ξ
χ2 0 < ξ < +∞ N ξ 2 exp − N s2 2N s4
2 2 sN Γ( N ) 2s2
2
1√
2D −∞ < ξ, η < +∞ 2πs1 s2 1−r 2
×
h  i
1 (ξ−m1 )2 (η−m2 )2 2r(ξ−m1 )(η−m2 )
Gauss exp − 2(1−r 2) s2
+ s22
− s1 s2
1

Parametre m1 , m2 , s1 , s2 , r (|r| < 1)

Table 8.1. A table of probability distribution functions.

8.1.2 Relations between random variables

Two random variables X and Y are said to be independent if
WXY (ξ, η) = WX (ξ)WY (η), or wXY (ξ, η) = wX (ξ)wY (η) (8.1.19)
This definition can be extended to more random variables.
Two random variables X and Y are said to be uncorrelated if the mean of their
product is equal to the product of their means:
E{XY } = E{X}E{Y }. (8.1.20)
Two variables are orthogonal if the mean of their product is 0:
E{XY } = 0. (8.1.21)
It can be seen that independent random variables are also uncorrelated. The
opposite is not always true. Let’s consider two variables X and Y , which both have
normal distribution (see Table 8.1). If these two variables are not correlated, i.e.
r = 0, then they are also independent.
If a random variable U is the sum of two random variables X and Y , i.e.:
U =X +Y (8.1.22)
we can find the probability density function of U from the two-dimensional proba-
bility density function of X and Y by:
Z ∞
wU (ξ) = wXY (ξ − η, η)dη. (8.1.23)
−∞

If X and Y are independent, then this expression is reduced to

Z ∞
wU (ξ) = wX (η)wY (ξ − η)dη (8.1.24)
−∞
or
wU (ξ) = wX (ξ) ∗ wY (ξ). (8.1.25)
These equations can be extended to an arbitrary number of random variables.
8-6 Random signals Chapter 8

Gauss, m=0, s=1 Rayleigh, s=1

0.4 0.6

0.5
0.3
0.4
w(ξ)

w(ξ)
0.2 0.3

0.2
0.1
0.1

0 0
−4 −2 0 2 4 0 1 2 3 4
ξ ξ
2 2
χ , s=1, N=1 χ , s=1, N=2 og N=6

0.5
0.6

0.5 0.4

0.4 N=2
0.3
w(ξ)

w(ξ)
0.3
0.2
0.2 N=6
0.1
0.1

0 0
0 1 2 3 4 0 2 4 6 8
ξ ξ

Figure 8.3. Examples of different probability density functions.

8.1.3 Function of a random variable

Often, we would like to work not with the original random variable, but with another
random variable that has been derived from the original (the result of processing
a random signal). The conversion can be made either through a look-up table, or
through one or more mathematical manipulations.
In such cases, we want to be able to compute the probability density function
of the new variables, as well as their expected values and other moments.
Let’s consider the random variable X with a probability density function wX (ξ),
and the variable Y which is a function of X:

Y = β(X), (8.1.26)

where η = β(ξ) is the desired conversion of variable. The probability density func-
tion wY (η) can be calculated through an appropriate summation of probabilities.
If wX (ξ) consists exclusively of δ-functions, then placement and strength of these
δ-functions can be found easily through the mapping η = β(ξ).
If wX (ξ) does not contain δ-functions, and the function η = β(ξ) is differentiable,
then wY (η) can be found from the expression:
X
P {Y ∈ [η, η + dη]} = P {X ∈ Iq }, (8.1.27)
q

where Iq denotes those intervals, where X can be found, given that Y is in the
desired interval of values (in the case of surjections there may be several values of
ξ which map to a given η). The magnitude of the probability P {X ∈ Iq } for one of
Section 8.1. Random variables 8-7

Figure 8.4. Mapping of the probability density function.

the intervals within the sum is (see Fig. 8.4):

1
wX (ξq ) , (8.1.28)
|β 0 (ξ q )|

where ξq are the roots of the equation β(ξ) − η = 0, and

dβ(ξ)
β 0 (ξ) = . (8.1.29)
dξ
If β 0 (ξ) is a zero only for a finite number of values of ξ, then the inverse function
ξ = γ(η) of η = β(ξ) exists in these intervals, and we get:
X wX (ξ) X wX (γ(η))
wY (η) = = . (8.1.30)
q
|β 0 (ξ)| q
|β 0 (γ(η))|

In some applications the function η = β(ξ) can be a constant in certain intervals

along the ξ-axis. In such cases wY (η) will contain δ-functions, whose strength can
be found by integrating wX (ξ) over the respective intervals.
When wX (ξ) is of complex character, then it can be treated by combining the
above mentioned two methods.
It can be shown that if Y = β(X), then:
Z ∞
E{Y q } = E{β q (X)} β q (ξ)wX (ξ)dξ (8.1.31)
−∞

Example 8.1 (Amplifier with a non-linear transfer characteristics)

A power amplifier of audio signals has a transfer characteristics given by:

k(g(t) + α) for g(t) < −α

y(t) = k(g(t) − α) for g(t) > α (8.1.32)

0 for |g(t)| 6 α


Here, g(t) is the input signal to the amplifier, y(t) is the corresponding output
signal, and α is a positive real constant. The transfer characteristic is shown below:
8-8 Random signals Chapter 8

y(t)

g(t)
−α α

A noise source x(t) is connected to the input of the amplifier. The probability
density function of x(t) is defined as:
(
β for |ξ| 6 b
wx (ξ) = (8.1.33)
0 otherwise
Find β and the probability density function of the output signal y(t) expressed via
b, α, and k, when b 6 α and b > alpha. Calculate the power of the output signal
for both cases.
The probability density functions:
Rb 1
Since −b wx (ξ)dξ = 1, we get β = 2b .
For the output signal when b > α we get:
η


η


α/b
k(b-α)

wy(η) ξ

-k(b-α)

1
2bk

1
2b

−α α ξ

-b b

wy (η) consists of a delta function and rectangular density. The amplitude of the
delta function is:
1 α
2α =
2b b
The height of rectangular density function is found as:
β(ξ) = k(ξ − α)

β 0 (ξ) = k

X wx (ξ) 1
2b 1
wy (η) = 0 (ξ)|
= =
q
|β |k| 2bk
When b < α the output is always 0, and the probability density function is just
a delta function with amplitude of 1. The power of the noise is 0. When b > α the
power is:
Z +k(b−α)  
2 1 α
Py = η + δ(η) dη
−k(b−α) 2bk b
1 1
= ((k(b − α))3 + (k(b − α))3 )
2bk 3
k 3 (b − a)3 k 2 (b − a)3
= =
3bk 3b
Section 8.2. Random Process 8-9

..................................................................................

8.2 Random Process

In this section we develop the theory behind the characterization of random pro-
cesses. After the terms are defined we look into the class of stationary and ergodic
random processes. The section finishes with the treatment of Gaussian random
processes, known also as normal random process.

8.2.1 Definitions
The word process (lat. movement) denotes a sequence of changes of the properties
of a system or object. These may appear deterministic or stochastic in nature to
the observer. Just as a random variable may be thought of as mapping from the
sample space of a given experiment into the set of real or complex numbers, a
random process represents mapping of the sample space into a set of signals. Thus
a random processes is a collection or ensemble of signals. In the following we will
use the capital letters to denote random processes:

X(t), Y (t) or X(n), Y (n) (8.2.1)

for analog and digital signals, respectively. Another term for a random process is
stochastic process and in the rest of the chapter these will be used interchangeably.
The process must be considered at a fixed time instance, to apply the mathe-
matical apparatus used to treat random variables. In this way we have the values of
the signals for a given time instance, which random values are tied to the outcomes
of the experiment. In other words, the values of the signal at the time instance
t = t0 represent stochastic variables, which can be characterized by the appropriate
probability density functions. The simplest of these are one-dimensional of the type:

wX (ξ; t0 ) or wX (ξ; n0 ). (8.2.2)

The expected value or the mean of a random process is also a function of time
and can be obtained by:
Z ∞
E{X(t)} = ξwX (ξ; t)dξ = µ1 (t) (8.2.3)
−∞

Similarly, the qth moment, calculated as

Z ∞
E{X q (n)} = ξ q wX (ξ; n)dξ = µq (n) (8.2.4)
−∞

will also be time dependent.

A process is linked to signals, not to numbers that indicate the outcome of a
given experiment. Therefore a random process is not described in its entireness by
giving wx (ξ; t) . A complete description would usually require the knowledge of the
properties of the process at different times t1 , t2 , t3 , . . . . Such description is often
impossible to handle in practice.
If a process is considered at two time instances, t1 and t2 , we get a pair of
numbers, whose elements each in its own right is a random variable. The two-
dimensional probability density function wX (ξ1 , ξ2 ; t1 , t2 ) can be used to describe
the behavior of this pair of numbers. It is used in the calculations of the different
moments (expected values), which could be of interest. These expected values will
in general are a function of both parameters t1 and t2 .
8-10 Random signals Chapter 8

Hvis støj (digitalt signal)

4
2
0
−2
−4
0 5 10 15 20 25 30 35 40 45 50
4
2
0
−2
−4
0 5 10 15 20 25 30 35 40 45 50
4
2
0
−2
−4
0 5 10 15 20 25 30 35 40 45 50
4
2
0
−2
−4
0 5 10 15 20 25 30 35 40 45 50
n
Smabåndstøj (analogt signal)

20
0
−20

0 20 40 60 80 100 120 140 160 180 200

20
0
−20

0 20 40 60 80 100 120 140 160 180 200

20
0
−20

0 20 40 60 80 100 120 140 160 180 200

20
0
−20

0 20 40 60 80 100 120 140 160 180 200

t

Figure 8.5. Examples of different realizations of analog and digital random signals.

The expected value E{X(t1 )X(t2 )} plays an essential part in the characteriza-
tion of the process. It is known as the autocorrelation of the signal and is calculated
as:
Z ∞Z ∞
RX (t1 , t2 ) = E{X(t1 )X(t2 )} = ξ1 ξ2 wX (ξ1 , ξ2 ; t1 , t2 )dξ1 dξ2 (8.2.5)
−∞ −∞
Section 8.2. Random Process 8-11

In the following, the autocorrelation will be denoted as RX (t1 , t2 ) or just R(t1 , t2 ).

The corresponding expression for digital signals is:
Z ∞ Z ∞
RX (n1 , n2 ) = ξ1 ξ2 wX (ξ1 , ξ2 ; n1 , n2 )dξ1 dξ2 (8.2.6)
−∞ −∞

Notice that the autocorrelation function of a discrete-time signal is also discrete in

time.
A further increase in the number of time instances used in the description of
signals is straightforward. The practical problems accompanying such approach,
however, do not justify its use. This is why, people usually stop at the autocorrela-
tion and accept the hereby following incompleteness of the signal description.

8.2.2 Stationary random process

If the probability density functions (of arbitrary high order) of a stochastic process
are independent of time, then the process is said to be stationary in the strict sense.
Mathematically, this condition is expressed as:

wX (ξ; t) = wX (ξ). (8.2.7)

As said, a stationary process in the strict sense will have qth order density function
independent of time, and the qth-order moment will also be independent of time:
Z ∞
q
E{X (t)} = ξ q wX (ξ)dξ, q integer. (8.2.8)
−∞

Furthermore, the joint probability density function of a stationary process

wX (ξ1 , ξ2 ; t1 , t2 ) = wX (ξ1 , ξ2 ; |t1 − t2 |) (8.2.9)

will depend only on the difference between t1 and t2 . This difference is often denoted
by τ , and the autocorrelation of the stationary process becomes:
Z ∞ Z ∞
RX (τ ) = ξ1 ξ2 wX (ξ1 , ξ2 ; |τ |)dξ1 dξ2
−∞ −∞
Z ∞ Z ∞ (8.2.10)
RX (k) = ξ1 ξ2 wX (ξ1 , ξ2 ; |k|)dξ1 dξ2 ,
−∞ −∞

for analog and digital signals, respectively.

Continuing to a higher-order joint density functions, a process is said to be
stationary of order p if pth-order probability density functions are independent of
time.
As said, we will be interested only in the mean and the autocorrelation of a pro-
cess. If these are time-independent, we get a wide-sense stationary (WSS) process,
which is defined as follows.
Wide Sense Stationarity: A random process x(n) is said to be wide-sense
stationary if:
1. The mean of the process is constant E{X(t)} = const.
2. The auto-correlation function depends only on the difference in time
RX (t1 , t2 ) = RX (|t1 − t2 |).
2
3. The variance of the process is finite σX < ∞.
8-12 Random signals Chapter 8

8.2.3 Ergodic random process

In many applications it is necessary to find the mean and the autocorrelation of a
random process. As it was shown in the previous section, these measures require
that we have an ensemble of realizations of the random process, which is often
impossible. A certain class of processes, known as ergodic random processes occupy
a special place among random processes. An ergodic process is a process generating
uniform signals, that is signals which share the same statistical properties. In these
signals it is possible to substitute the ensemle averages with averages in time. So
having any of these signals is sufficient to characterize the process completely.
The expected values of the ergodic signals (which require knowledge of the statis-
tical properties of the whole ensemble of signals) are identical to the time averages of
the signal (which require knowledge of only a single of the signals of the ensemble).
We will use the notation h·i to denote a time-average of the signal(s), which
are inside the angle brackets. The time averages of an analog x(t) and digital x(n)
signals are given by:
Z 12 T
1
hx(t)i = lim x(t)dt
T →∞ T − 1 T
2
(8.2.11)
1 X
hx(n)i = lim x(n)
N →∞ N
N
P
where N means that the summation spans over N signal samples, chosen in such
a way, that through the limit transition one encompasses the whole signal.

The fact that a stationary random process process is ergodic means that
the mean values in time of any of the signals generated by the process are
equal to the mean values calculated over the ensemble of realizations. For
example:

E{X(t)} = hx(t)i
(8.2.13)
E{X(n)X(n + k)} = hx(n)x(n + k)i

Example 8.2 (Example of a stochastic process)

Let’s consider the following experiment. We have a box of resistors of the same type
(e.g. 150 Ω ± 10%). Typically these resistor will have different actual resistance,
since there is a variation in the production process. Let’s pick randomly a number
of resistors, and apply voltage on them. These resistors will create electric noise
signal due to the thermal movement of electrons. The measured noise signals is that
signal, which is assigned to the individual outcomes of the experiment. Let’s further
assume that we like fresh air, and we conduct the experiment outdoors, where the
ambient temperature varies.
Since we have assumed that the resistors are of the “same type”, we have gen-
erated a stochastic process X(t).
The stochastic process is not stationary, because the probability density function
wX (ξ; t) (8.2.14)
of the process will have parameters which vary with the ambient temperature, in
other words - the time.
If we subject the resistors to completely identical conditions (temperature, mois-
ture, etc), the process will become stationary. The process is, however, not ergodic.
Let’s assume for the sake of simplicity that:
E{X(t)} = 0, (8.2.15)
Section 8.2. Random Process 8-13

which means that there is no DC component of the current flowing through the
resistors. If we calculate the power of the resulting signal we will find out that:
hx2p (t)i =
6 hx2q (t)i, (8.2.16)
where xp (t) and xq (t) are two randomly chosen signals from the constructed process.
They are different, because the actual resistance of the resistors varies from resistor
to resistor.
In other words, it is not possible to characterize the process based on the study
of only one of the signals assigned to the outcomes of the experiment.
Now, if the process is modified such as the resistors in the experiment are fully
identical, and are kept under the same conditions, then the process becomes ergodic.
Under these conditions we will have:
hx2p (t)i = hx2q (t)i, for all p and q. (8.2.17)
For the sake of clarity we must emphasize that this is not the same as
xp (t) ≡ xq (t) for p 6= q, (8.2.18)
because the signals emanate from two different resistors.
..................................................................................

8.2.4 Stationary, normal random process

The so-called normal distribution plays an important role in nature. It is even part
of the plot of the novel Jurassic Park by Michael Crichton. The people who work at
the park are thrilled because the histogram shows a nice normal distribution, which
they say is what you’d expect from height data from a typical, healthy, biological
population. Many other processes can be modeled well using a normal distribution,
and this is the reason why it is treated in this section.
The one-dimensional probability density function of a normal random process is
given by1

(ξ − m)2
 
1
wX (ξ) = √ · exp − , −∞ < ξ < ∞ (8.2.19)
2π · s2 2s2

(see also table 8.1). The constants m and s are given by:
m = E{X(t)}
(8.2.20)
s2 = E{(X(t) − E{X(t)})2 } = E{X 2 (t)} − E 2 {X(t)}.
If the mean value of the process is 0, E{X(t)} = 0, then
s2 = E{X 2 (t)} = R(0), (8.2.21)
where R(τ ) is the autocorrelation function of the process.
It can be shown that the higher-order even moments of the process are given by:
(2q)! 2q
E{x2q (t)} = s , (8.2.22)
2q q!
while all moments of odd order are equal to zero.
The two-dimensional probability density function of a stationary normal process
with mean value of 0 is:
R(0)ξ12 + R(0)ξ22 − 2R(τ )ξ1 ξ2
 
1
wX (ξ1 , ξ2 ; τ ) = · exp − .
2(R2 (0) − R2 (τ ))
p
2π R2 (0) − R2 (τ )
(8.2.23)
1 Notice that strictly speaking, a digital signal cannot be normally distributed.
8-14 Random signals Chapter 8

From this equation it can be seen that if R(τ ) = 0 either for discrete values of τ or
in intervals of values, then we have:

wX (ξ1 , ξ2 ; τ ) = wX (ξ1 )wX (ξ2 ) . (8.2.24)

In other words, the values of the random signal which are τ seconds apart are
independent under these conditions.
It is possible to show that all multidimensional probability density functions are
functions of R(τ ) (and m, if m 6= 0).

8.3 Correlation functions and power density spectrum

As mentioned in Section 8.2, the first and the second order moments are important
means to characterize random processes in practice. The mean of a process is an
indication of whether there is a DC (direct current) component of the signal, and
the autocorrelation function, which is a function of time, is used as a basis for the
spectral description of random processes.

8.3.1 Autocorrelation functions

The auto correlation functions of the stationary signals X(t) and X(n) are defined
as:

RX (τ ) = E{X(t)X(t + τ )} and RX (k) = E{X(n)X(n + k)} (8.3.1)

They express the average dependence (relation) between the values of the signal
that are τ seconds or k samples apart.
Closely related to the autocorrelation functions is the autocovariance:

CX (τ ) = E{[X(t) − E{X(t)}][X(t + τ ) − E{X(t}]} . (8.3.2)

The auto-covariance becomes identical to the autocorrelation, when there is no DC

component present in the signal, i.e. E{X(t)} = 0. For discrete-time random
signals, the auto-covariance is defined in a similar way.
The autocorrelation functions RX (τ ) and RX (k) are even functions because
the two-dimensional probability density functions wX (ξ1 , ξ2 ; τ ) and wX (ξ1 , ξ2 ; k)
depend only on the value of the time parameter. In other words:

RX (τ ) = RX (−τ ) and RX (k) = RX (−k). (8.3.3)

Analogous to the relations that exist for random variables, the variance of a
discrete-time signal x(n) is defined as:

σ 2 {X(n)} = E{[X(n) − E{X(n)}]2 }. (8.3.4)

The variance of an analog signal is defined in a similar way. Using the fact that
Z ∞
(ξ − E{X(n)})2 wX (ξ)dξ = E{X 2 (n)} − E 2 {X(n)} (8.3.5)
−∞

it can be shown that

RX (0) = E{X 2 (n)} = σ 2 {X(n)} + E 2 {X(n)}. (8.3.6)

Hence the autocorrelation function at lag 0 is always non-negative:

RX (0) > 0 (8.3.7)

Section 8.3. Correlation functions and power density spectrum 8-15

Consider the random signal Y (t) constructed from an arbitrary stationary ran-
dom signal X(t) as:
Y (t) = X(t) ± X(t + τ ) (8.3.8)
The autocorrelation function of Y (t) at lag zero is

RY (0) = E{(X(t) ± X(t + τ ))2 } = 2RX (0) ± 2RX (τ ). (8.3.9)

Since RY (0) > 0 it follows that:

RX (0) > |RX (τ )| for τ 6= 0 (8.3.10)

The autocorrelation functions of most random signals approaches a positive constant

at infinite time lags:

lim R(τ ) = R∞ and lim R(k) = R∞ , (8.3.11)

|τ |→∞ |k|→∞

where R∞ is a positive constant. It can be shown that

R∞ = E 2 {X(t)} and R∞ = E 2 {X(n)} (8.3.12)

This limit value is in other words the DC-power of the signal2 . Hence the AC-power
of the random process must be given by:

σ 2 {X(t)} and σ 2 {X(n)}, (8.3.13)

since the total power P is given by the value of the autocorrelation function at time
lag 0.
As mentioned previously, the time-average values can be used instead of the
expected values. In other words, the power of a signal is

P = hx2q (t)i and P = hx2q (n)i. (8.3.14)

8.3.2 Spectrum of a random signal

The spectrum of a random signal is defined as the Fourier transform of the auto-
correlation function of the signal. For analog signals this means that:
Z ∞
RX (τ ) ↔ SX (f ) = RX (τ )e−j2πf τ dτ. (8.3.15)
−∞

RX (τ ) is an even function, hence the SX (f ) is real and even function of the frequency
f . Using the fact that the power P is P = RX (0) and that e−jπf 0 ≡ 1, we have
Z ∞
P = RX (0) = SX (f )df (8.3.16)
−∞

In other words, it can be seen that SX (f ) is a power density spectrum. It tells us in

which frequencies the random signal has spectral components. It can also be shown
that:
SX (f ) > 0. (8.3.17)
Similarly the spectrum of a random discrete-time signal is given as:
∞
∆T
X
RX (k) ↔ SX (f ) = RX (k)e−j2πf k∆T . (8.3.18)
k=−∞

2 You can imagine that the signal is either voltage applied on or current flowing through a 1Ω

resistor. The power is that one dissipated by the resistor.

8-16 Random signals Chapter 8

SX (f ) is an even function of f in this case too. It is also periodic with a a period

1
of ∆T . For discrete-time signals, SX (f ) is a power spectrum. Often the terms
power-density spectrum and power spectrum are used indiscriminately.
The total effect of a discrete-time random signal is found as:
fs
Z2
1
P = RX (0) = SX (f )df, (8.3.19)
fs
− f2s

1
where fs = ∆T is the sampling frequency.
As it can be seen, the spectral characteristics of random signals are characterized
by using a deterministic signal with finite energy - RX (τ ) or RX (k).

8.3.3 Cross correlation functions

In many practical applications we have random signals which are a combination
of random signals with known properties. To be able to characterize such sig-
nals through their autocorrelation function, we need to introduce the terms cross-
correlation and cross spectrum.
The cross correlation of two stationary random analog signals X(t) and Y (t) is
defined as:
Z ∞Z ∞
RXY (τ ) = E{X(t), Y (t + τ )} = ξηwXY (ξ, η; τ )dξdη. (8.3.20)
−∞ −∞

The cross correlation function expresses the level of connection (coherence) between
two signals X(t) and Y (t).
The corresponding expression for discrete-time signals is:

RXY (k) = E{X(n)Y (n + k)} (8.3.21)

The probability density functions

wXY (ξ, η; τ ) and wXY (ξ, η; k)

are usually not even functions of τ or k. Hence, the cross correlation functions are
not even too:

RXY (τ ) 6= RXY (−τ ) (τ 6= 0) RXY (k) 6= RXY (−k) (k 6= 0). (8.3.22)

On the other hand it can be shown through a simple substitution of variables

that:
RXY (τ ) = RY X (−τ ) and RXY (k) = RY X (−k). (8.3.23)
Consider the equation:

E{(X(t) + a(Y (t + τ ))2 )} = RX (0) + a2 RY (0) + 2aRXY (τ ) > 0 (8.3.24)

where a is a real constant and X(t) and Y (t) are analog random signals. This
equation is a square equation of a. It is a convex function of a. In order to be valid
for all values of a, it must not have any solutions.Therefore, the following condition
must be fulfilled:
(2RXY (τ ))2 − 4RY (0)RX (0) < 0, (8.3.25)
hence
2
RXY < RX (0)RY (0) (8.3.26)
In the special case when a = 1 or a = −1 we get

2|RXY (τ )| < RX (0)RY (0) (8.3.27)

Section 8.4. Linear and non-linear combination of random signals 8-17

The corresponding expressions for discrete-time signals are:

2
RXY (k) < RX (0) + RY (0) (8.3.28)

and
2|RXY (k)| < RX (0) + RY (0) (8.3.29)
Two random signals X(n) and Y (n) are said to be uncorrelated for a given time
lag k = k1 if

RXY (k1 ) = E{X(n)Y (n − k1 )} = E{X(n)}E{Y (n)} (8.3.30)

If the last expression is valid for any value of k, i.e.

RXY (k) = E{X(n)}E{Y (n)}, (8.3.31)

then the two signals are uncorrelated.

X(n) and Y (n) are said to be orthogonal signals if

RXY (k1 ) = 0 and RXY (k) = 0 (for all k) (8.3.32)

Notice that if only one of the signals has a zero DC component, then the two terms
uncorrelated and orthogonal become identical.
While the above conditions impose some restrictions on the expected values of
the signals, two signals are said to be statistically independent if their multidimen-
sional distributions (and pdf) can be written as a product of the distributions (and
pdf):
wXY (ξ, η; k) = wX (ξ) · wY (η). (8.3.33)
Consequently, statistically independent signals are also uncorrelated. The converse
is valid only for signals, whose two-dimensional probability density functions have
normal distribution.
The above is valid for analog signals too. If the signals are ergodic, then the
expected values can be substituted by time averages.

8.3.4 Cross spectra of two random signals

The cross-spectrum of random signals is defined as the Fourier transform of the
cross-correlation function of the two signals:
∆T
RXY (τ ) ↔ SXY (f ) and RXY (k) ↔ SXY (f ). (8.3.34)

It must be noted that SXY (f ) usually is complex and that

∗
SXY (f ) = SXY (−f ) (8.3.35)

and
∗
SY X (f ) = SXY (f ). (8.3.36)
The cross-spectrum carries information about within which parts of the spectrum
the two signals has some resemblance. In other words, the values of the cross-
spectrum in that frequency range must be 6= 0.

8.4 Linear and non-linear combination of random signals

8.4.1 Multiplication of a random signal by a constant
The multiplication of a random signal by a constant is expressed as:

Y (t) = aX(t) (8.4.1)

8-18 Random signals Chapter 8

where a is a real constant which is 6= 0, and E{X(t)} = a

The one-dimensional probability density function of the signal Y (t) can be found
as described in section 8.1.3:
1
wY (η) = wX (η/a) (8.4.2)
|a|

The power PY of the signal Y (t) is easily found as

PY = a2 PX , (8.4.3)

where PX is the power of X(t).

From the expression

RY (τ ) = E{aX(t)aX(t + τ )} = a2 RX (τ ) (8.4.4)

it follows that
SY (f ) = a2 SX (f ). (8.4.5)
It must be noted that if X(t) is stationary then Y (t) is stationary too.
Some applications use a slightly different form of signal combination:

U (t) = aX(t) + b (8.4.6)

where a and b are real constants and a 6= 0.

The above expression means that the signal aX(t) is modified by a constant
(DC) signal b. Repeating the considerations from above we get:
1
wU (η) = wx ((η − b)/a). (8.4.7)
|a|

The power of U (t) is (given that E{X(t)} = 0)

PU = a2 PX + b2 . (8.4.8)

The autocorrelation function of U (t) is

RU (τ ) = a2 RX (τ ) + b2 . (8.4.9)

The power density spectrum is

SU (f ) = a2 SX (f ) + b2 δ(f ) (8.4.10)

All of these expressions are valid for digital signals too.

8.4.2 Linear combination of two random signals

The linear combination of two random signals X(t) and Y (t) creates a new random
signal U (t) which is
U (t) = aX(t) + bY (t), (8.4.11)
where a and b are real constants. The one-dimensional probability density function
of U (t) can be found using the results from section 8.1.3.
The autocorrelation of U (t) is

RU (τ ) = E{(aX(t) + bY (t))(aX(t + τ ) + bY (t + τ ))}

(8.4.12)
= a2 RX (τ ) + b2 RY (τ ) + abRXY (τ ) + abRY X (τ )

which implies that we must know the autocorrelation functions of X(t) and Y (t),
and their cross-correlation to find the autocorrelation of U (t).
Section 8.5. Correlation function for deterministic signals with finite energy 8-19

The total power of U (t) is

PU = a2 RX (0) + b2 RY (0) + 2abRXY (0) (8.4.13)
If RXY (0) is negative and has a certain value, this means that PU could be less
than both PX and PY . The power spectrum of U (t) is
SU (f ) = a2 SX (f ) + b2 SY (f ) + 2abRe[SXY (f )] (8.4.14)
If X(t) and Y (t) are uncorrelated (or statistically independent) and the mean
value of one of the signals is 0 (E{X(t)}=0 or E{Y(t)}=0), then the above expres-
sions are simplified as
RU (τ ) = a2 RX (τ ) + b2 RY (τ ), PU = a2 PX + b2 PY (8.4.15)
and
SU (f ) = a2 SX (f ) + b2 SY (f ) (8.4.16)
There are similar expressions for discrete-time signals.

8.4.3 Product of two random signals

Some non-stationary signals encountered in practical applications, can be modeled
as the product of two random signals X(t) and Y (t), i.e.
V (t) = X(t)Y (t) (8.4.17)
The result is relatively simple only if the two signals are statistically independent.
Its autocorrelation function can be shown to be
RV (τ ) = RX (τ )RY (τ ). (8.4.18)
The same result is valid for digital signals.

8.5 Correlation function for deterministic signals with finite en-

ergy
To understand better the filtration of random signals, it is necessary to introduce
also the autocorrelation function for signals with finite energy. The autocorrelation
Rh (τ ) for the energy signal h(t) is given by:
Z ∞
Rh (τ ) = h(t)h(t + τ )dt (8.5.1)
−∞

Notice that the integral is not normalized as it in the case of ergodic random signals.
An understanding of how the autocorrelation function is created can be obtained
from the following equation:
Z ∞
Rh (τ ) = h(−τ ) ∗ h(τ ) = h(θ)h(θ + τ )dθ (8.5.2)
−∞

It can be seen that the autocorrelation can be found by convolving the signal with
a time-reversed copy of itself. This is illustrated in Fig. 8.6
From the autocorrelation function one can find the energy density spectrum:
Rh (τ ) = h(−τ ) ∗ h(τ )↔H ∗ (f )H(f ) = |H(f )|2 (8.5.3)
The corresponding expressions for a finite-energy digital signal are:
∞
X
Rh (k) = h(n)h(n + k) (8.5.4)
n=−∞

and
∆T
Rh (k) = h(−k) ∗ h(k) = ↔ H ∗ (f )H(f ) = |H(f )|2 . (8.5.5)
8-20 Random signals Chapter 8

1
h(t) 0.5

−0.5

−1
−0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 0.02 0.03 0.04 0.05

0.5
h(−t)

−0.5

−1
−0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 0.02 0.03 0.04 0.05
t

10
R (τ)

0
h

−10

−0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 0.02 0.03 0.04 0.05
τ

Figure 8.6. The autocorrelation function of a finite energy analog signal.

8.6 Correlation function and power spectrum of periodic signals

Sometimes deterministic periodic signals appear together with random signals.
Since the characterization of random signals is based on their autocorrelation func-
tions or their power density spectra, it is useful to define the correlation functions
for periodic signals.

8.6.1 Autocorrelation function of periodic signals

The autocorrelation function of an analog periodic signal is defined as (see Fig. 8.7):
Z T
1 2
Rg (τ ) = g(t)g(t + τ )dt (8.6.1)
T − T2

Rg (τ ) is again a periodic function with the same period T as the signal g(t). This
can easily be verified is τ is replaced by τ + T in (8.6.1).
The autocorrelation function is also an even function of time τ , i.e. Rg (τ ) =
Rg (−τ ). This property can also be verified by substitution of variables. In this case
one uses the fact that the integration must not necessarily start at − T2 , but must
only include an integer number of periods of g(t)3
From the expression
Z T2
1
Rg (0) = g 2 (t)dt, (8.6.2)
T − T2

it can be seen that Rg (0) = P is the power of the signals.

3 Strictly 1
R
speaking Rg (τ ) = lim T1 g(t)g(t + τ )dt
T1 →∞ T1
Section 8.6. Correlation function and power spectrum of periodic signals 8-21

g(t) 0.5

−0.5

−0.025 −0.02 −0.015 −0.01 −0.005 0 0.005 0.01 0.015 0.02 0.025

0.5
g(t+τ1)

−0.5

−0.025 −0.02 −0.015 −0.01 −0.005 0 0.005 0.01 0.015 0.02 0.025
t

20
Rg(τ)

−20
−0.025 −0.02 −0.015 −0.01 −0.005 0 0.005 0.01 0.015 0.02 0.025
τ

Figure 8.7. Autocorrelation function of a periodic signal.

The autocorrelation of a periodic digital signal g(n) is given by

N −1
1 X
Rg (k) = g(n)g(n + N ), (8.6.3)
N n=0

where k is an integer number. Also in this case we have:

Rg (k) = Rg (k + N ), Rg (k) = Rg (−k), and P = Rg (0) (8.6.4)

8.6.2 Power spectrum of periodic signals

The spectrum Sp (f ) of the periodic signals Rg (τ ) is called the power spectrum of
the signal g(t):
T
Rg (τ )↔Sg (m) (8.6.5)
The reason to use the term power spectrum can easily be demonstrated by rewriting
the expression for Rg (τ ) as:
Z T
1 2
Rg (τ ) = g(−t)g(τ − t)dt = g(−τ ) ∗ g(τ ). (8.6.6)
T − T2

Hence,
Sg (m) = G∗ (m)G(m) = |G(m)|2 , (8.6.7)
T
where g(m)↔G(m). The total power can be found by
∞
X
P = Sg (m) (8.6.8)
m=−∞
8-22 Random signals Chapter 8

Since Rg (τ ) is an even function of τ , then the power spectrum Sg (m) will be a real
and even function of m. Notice that Sg (m) is independent of the phase-spectrum
of g(t), which implies that different signals can have the same power spectrum and
hereby the same autocorrelation function. It can also be seen that Sg (m) > 0. The
power spectrum of a periodic digital signal is given as
N
Rg (k)↔Sg (m), (8.6.9)

and as in the case of periodic analog signals the following relations will be valid:
N
X −1
Sg (m) = |G(m)|2 , P = Sg (m), and Sg (m) > 0. (8.6.10)
m=0

8.6.3 Cross-correlation function of two periodic signals

The cross-correlation function Rgh (τ ) of the periodic analog signals g(t) and h(t)
with the same period T is defined by
Z T
1 2
Rgh (τ ) = g(t)h(t + τ )dt (8.6.11)
T − T2

Rgh (τ ) is also periodic in τ with the period T , so that Rgh (τ ) = Rgh (τ + T ).

The cross-correlation function is in general not an even function, except if h(t) is
proportional to g(t). The following relation is, though, valid

Rgh (τ ) = Rhg (−τ ) (8.6.12)

The latter can easily be proven by substituting θ with θ = t + τ in the definition of

the cross correlation. To show it one must also use the fact the at the integration
is carried out over a full period independent of its location in time.
The cross-correlation function for two periodic digital signals g(n) and h(n) is
also periodic with the period N as the two signals g(n) and h(n), and is found from:
N −1
1 X
Rgh (k) = g(n)h(n + k). (8.6.13)
N n=0

It can easily be seen that Rgh (k) = Rhg (−k). Notice that although the given
expression contain integration over a single period, the following relations are also
valid: Z
1
lim g(t)h(t + τ )dt = Rgh (τ ) (8.6.14)
T1 →∞ T1 T
1

and
1 X
lim g(n)h(n + k) = Rgh (k) (8.6.15)
N0 →∞ N0
N0

8.6.4 Cross power spectrum of two periodic signals

The Fourier transform Sgh (f ) of Rgh (τ ) is called cross-power spectrum of the two
signals g(t) and h(t). This spectrum is a normal complex function of the complex
frequency parameter m. From the expression:
Z T
1 2
Rgh (τ ) = g(−t)h(τ − t) = h(τ ) ∗ g(−τ ) (8.6.16)
T − T2

it follows that:
Sgh (m) = H(m)G∗ (m) = H(m)G(−m), (8.6.17)
Section 8.6. Correlation function and power spectrum of periodic signals 8-23

T N
where g(t)↔G(m) and h(t)↔H(f ).
The same relations can be proven for digital signals too:
N
Rgh (k)↔Sgh (m) = H(m)G∗ (m), (8.6.18)
N N
where g(n)↔G(m) and h(n)↔H(m). The cross power spectrum gives information
about those frequencies where both signals have spectral components. It can be seen
that there exist many signal pairs which have the same cross-correlation function,
hence cross power spectrum.

8.6.5 Correlation functions and power spectra of pure tones

A pure-tone signal, analog or digital is given by

g(t) = a1 cos(2πf0 t + θ1 ) (8.6.19)

and
n
+ θ1 ),
g(n) = a1 cos(2π (8.6.20)
N
respectively. In the expressions above a1 and θ1 are constants, and the periods are
T = 1/f0 and N . These signals have spectral components only at m = ±1 (only
first harmonics). In both cases

a21
Sg (±1) = |G(±1)|2 = (8.6.21)
4
It follows that the autocorrelation functions for the two signals are:
1 2 1 2 k
Rg (τ ) = a cos 2πf0 τ and Rg (k) = a1 cos 2π (8.6.22)
2 1 2 N
Notice that the phase angle of the signals θ1 is missing in the expressions of the
autocorrelation functions. This means that all pure tones with the same frequency
and strength have the same autocorrelation function.
The cross-correlation function of two tones with the same frequency can be
found again using simple consideration in frequency domain. Let g(t) and h(n) be
defined by the expressions above and let h(t) and h(n) be h(t) = a2 cos(2πf0 + θ2 )
and h(n) = a2 cos(2πn/N + θ2 ). In other words, the differences lie in the different
amplitude and phase. The cross-power spectrum is:
a1 a2 ±j(θ2 −θ1 )
Sgh (±1) = e , (8.6.23)
4
which implies that:
a1 a2
Rgh (τ ) = cos(2πf0 t + (θ2 − θ1 )) (8.6.24)
2
and
a1 a2 n
Rgh (k) = cos(2π + (θ2 − θ1 )). (8.6.25)
2 N
If the two pure tones do not have the same period, then Rgh (τ ) ≡ 0 and Rgh (k) ≡
0.

8.6.6 Correlation functions of periodic signal with equally large

spectral components
If a periodic band-limited signal has spectral components which are real4 and equally
large, then the maximum of this signal is centered at t = 0 or n = 0, and no other
4 Remember that real signals with even symmetry hare real spectra.
8-24 Random signals Chapter 8

periodic signal consisting of the same spectral components and having the same
power will have values that are larger.
Such a signal (g(t) or g(n)) is further characterized by the fact that the au-
tocorrelation function is proportional to the signal itself, i.e. it has the same
T
shape as the signal. This is a direct sequence of the fact that Rg (τ )↔|G(m)|2
N
and Rg (m)↔|G(m)|2 .
If the analog signal h(t) has spectral components at the same frequencies as
g(t), it follows that the cross-correlation function Rgh (τ ) will have the same shape
as h(t). It can be shown directly from the relation:

Sgh (m) = H(m)G∗ (m) = kx H(m), (8.6.26)

where kx is the constant and real value of the spectral components of g(t).

8.6.7 Autocorrelation function of a periodic rectangular signal

The autocorrelation function of the periodic analog signal
(
a | t |< 21 θ
g(t) = (8.6.27)
0 12 θ <| t |< 12 T

where θ < 12 T , is given for positive values of τ by

1
2 θ−τ
Z
1
a2 dt = a2 θ(1 − τ /θ)/T (8.6.28)
T − 12 θ

as shown in Fig. 8.8. For any value of τ , the autocorrelation function is

 
2 θ |τ |
Rg (τ ) = a 1− , |τ | < θ. (8.6.29)
T θ

The same result can be derived for the digital signal

(
a 0 6 n 6 N0 − 1
g(n) = (8.6.30)
0 N0 6 n 6 N − 1

where the constant N0 is N0 < 21 N . For positive values of k one gets the sum
−1−k
N0X
1 N0 − k 2
a2 = a (8.6.31)
N n=0
N

and hereby  
2 N0 |k|
Rg (k) = a 1− , |k| < N0 (8.6.32)
N N0
These results can be also obtained by first finding the power spectrum of g(t) or
g(n) and then using the inverse Fourier transform to find the correlation function.
Notice that the above expressions are for signals with a duty cycle below 50 %
( Tθ < 12 and NN0 < 21 ). Signals whose duty cycle exceeds 50 % can be treated as the
sum of a constant DC signal and a rectangular signal with a duty cycle < 50%, and
negative amplitude a.

8.6.8 Finding the spectrum of a signal via cross-correlation

The cross-correlation of a periodic signal g(t) and a pure tone makes it possible to
find the spectral component of g(t) at the frequency of the pure tone.
Section 8.6. Correlation function and power spectrum of periodic signals 8-25

1
g(t)

0.5

0
−0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2

1
g(t+τ1)

0.5

0
−0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2
t

0.2
Rg(τ)

0.1

0
−0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2
τ

Figure 8.8. Autocorrelation function of a periodic rectangular signal. The figure

gives the placement of the signals along the time-axis needed to find Rg (τ1 ). The
position of τ1 is indicated by the dashed line.

T
Let the spectrum of g(t) be g(t)↔G(m) = GR (m) + jGI (m), and let the pure
tone be
m0
cos 2π t
T
The cross correlation is
Z T
1 2 m0
Rgc (τ ) = g(t) cos(2π (t + τ ))dt (8.6.33)
T − T2 T

We can manipulate this expression a bit:

1
2T
Z
1
Rgc (τ ) = cos 2πm0 τ /T · g(t) cos 2πm0 t/T dt
T − 12 T
1
2T (8.6.34)
Z
1
− sin 2πm0 τ /T · g(t) sin 2πm0 t/T dt
T − 21 T

=GR (m0 ) cos 2πm0 τ /T + GI (m0 ) sin 2πm0 τ /T

Hence,
 
T
Rgc (0) = GR (m0 ) and Rgc = GI (m0 ) (8.6.35)
4m0

Similar considerations can be carried out for digital signals.

8-26 Random signals Chapter 8

8.7 Linear and non-linear processing of random signals

The most common form of linear processing of random signals in practice, is the
filtration with time-invariant filters. The spectral characteristics of random signals
are defined with the help of their autocorrelation functions and consequently the
description of how filters modify the spectral characteristics of such signals is a bit
different than for deterministic signals.
As far as non-linear signal processing is concerned, there does not exist a general
approach to determining all the characteristics of the result. Non-linear systems
with “memory” (i.e. their present non-linear characteristic is dependent on the
past of the system and its inputs) pose the largest complications. Such systems will
not be treated in this section. Only a couple of very often used in practice systems
without memory will be presented.

8.7.1 Filtration of random signals

Many applications involve the filtration of random signals with filters which have
an impulse response h(t) or h(n) and a transfer function H(f ), where
∆T
h(t) ←→ H(f ) and h(n) ←→ H(f ) . (8.7.1)

From a stochastic point of view, this is equivalent with, that for any outcome of the
given experiment xq (t) or xq (n), a new signal signal yq (t) or yq (n) is created by the
convolution:
yq (t) = h(t) ∗ xq (t) or yq (n) = h(n) ∗ xq (n) . (8.7.2)
In this case xq (t) or xq (n) are input signals to the filter. The output of the filter
is again a random process Y (t) or Y (n), and it is the properties of this random
process, which we will try to find using the properties of the random process X(t)
and X(n) and the characteristics of the filter.
It can be shown that if the input signals are strictly stationary, then the output
of the filter is stationary too. Conversely, if the input is stationary of kth order,
then the output is not necessarily stationary with the same order.
There exists no common calculation procedure to find the probability density
function of the output of the filter. It is possible to show, though5 , that if the
“duration” of the impulse response of a filter is long compared to “duration” of the
autocorrelation function of the input signal, then the probability density function
of the resulting signal will approach a normal (Gaussian) distribution. Here the
loose term “duration” can be the RMS duration for example.
The mean of Y (t) can be directly derived as:
Z ∞ Z ∞
E{Y (t)} = h(Θ)E{X(t−Θ)}dΘ = E{X(t)}· h(Θ)dΘ = E{X(t)}·H(0) ,
−∞ −∞
(8.7.3)
In other words, the mean of Y (t) is equal to the mean of X(t) multiplied by the DC
amplification of the filter. The corresponding expression for the digital signals is:
∞
X
E{Y (n)} = h(q)E{X(n − q)} = E{X(n)}H(0) . (8.7.4)
q=−∞

The autocorrelation function for the output of the filter for the analog case can
be found in the following manner.
From
Y (t) = h(t) ∗ X(t) (8.7.5)
5 See e.g. R.A. Altes: Wide Band Systems and Gaussianity, IEEE Transactions on Information

Theory, IT-21, November 1975, pages 679-82.

Section 8.7. Linear and non-linear processing of random signals 8-27

and
Y (t + τ ) = h(t) ∗ X(t + τ ) , (8.7.6)
one finds that
Z ∞ Z ∞
Y (t)Y (t + τ ) = h(Θ)X(t − Θ) · h(σ) · X(t + τ − σ)dΘdσ , (8.7.7)
−∞ −∞

and hereby
Z ∞ Z ∞
RY (τ ) = E{Y (t)Y (t + τ )} = h(Θ)h(σ)RX (τ − σ + Θ)dΘdσ . (8.7.8)
−∞ −∞

By using variable substitution γ = σ − Θ we get

Z ∞ Z ∞
RY (τ ) = RX (τ − γ) h(Θ)h(Θ + γ)dΘdγ . (8.7.9)
−∞ −∞

As mentioned in section 8.5, the inner of the two integrals is the autocorrelation
function Rh (γ) of the given filter. This means that

RY (τ ) = Rh (τ ) ∗ RX (τ ) . (8.7.10)

Since
Rh (τ ) = h(−τ ) ∗ h(τ ) ↔ H ∗ (f )H(f ) , (8.7.11)
then in frequency domain we get

SY (f ) =| H(f ) |2 SX (f ) . (8.7.12)

These considerations can be carried out also for digital signals and similar results
will be obtained.
Cross-correlation function RXY (τ ) between the input and the output signals of
the filter can be calculated as
Z ∞
RXY (τ ) = E{X(t)Y (t + τ )} = E{X(t) h(Θ)X(t + τ − Θ)dΘ}
−∞
Z ∞
= h(Θ)RX (τ − Θ)dΘ = h(τ ) ∗ RX (τ ) . (8.7.13)
−∞

In practice, this relation between RX (τ ), RXY (τ ) and h(τ ) is used to find h(τ ) and
hereby H(f ) for filters.
In frequency domain, this result is equivalent to

SXY (f ) = H(f )SX (f ) . (8.7.14)

The same expressions are valid for digital signals too.

The power PY of the output signal can be found in many ways. For analog
signals we have:
Z ∞ Z ∞
PY = RX (τ )Rh (τ )dτ = | H(f ) |2 SX (f )df , (8.7.15)
−∞ −∞

whereas for digital signals we have

∞ Z fs /2
X 1
PY = RX (q)Rh (q) = | H(f ) |2 SX (f )df . (8.7.16)
q=−∞
fs −fs /2
8-28 Random signals Chapter 8

2
Spektrum Ækvivalent støjbåndbredde for |H(f)|
6 f b
k 30 e

5
25
4
20

|H(f)|2
|H(f)|

3
15
2 10

1 5

0 0
0 100 200 300 400 500 0 100 200 300 400 500
Frekvens [Hz] Frekvens [Hz]

Figure 8.9. Definition of the equivalent noise bandwidth for a filter. Here fk is
chosen as the mean frequency in the spectrum.

Since PY > 0 for any filter, this implies that for any random signal the power
spectrum can never get negative values, i.e.
S(f ) > 0 . (8.7.17)
In connection with filters, an often used measure is the equivalent noise bandwidth
Be , which is defined as
Z ∞ Z fs /2
| H(f ) |2 | H(f ) |2
Be = 2
df and B e = df (8.7.18)
0 | H(fk ) | 0 | H(fk ) |2
for analog and digital signals, respectively, where fk is one or another characteristic
frequency of the filter, e.g. f = 0, the center frequency or the frequency, where
| H(f ) | has a maximum. Although it is not directly evident from the expression,
Be is a function of both | H(fk ) | and fk .
The equivalent noise bandwidth for a filter is based on the definition of the
equivalent noise bandwidth for a signal x(t). Be is the width of a fictitious rectan-
gular filter (in frequency domain) filtering noise with uniform power density equal
to Sx (fk ) W. Be is chosen such as the power at the output of the rectangular filter
is equal to the power of the signal x(t). The power of x(t) is
Z ∞
Px = Sx (f )df (8.7.19)
0

The equivalent power is:

Pe = Be Sx (fk ) (8.7.20)
where fk is the desired frequency. The equivalent bandwidth is easily derived to be:
Z ∞
1
Be = Sx (f )df (8.7.21)
Sx (fk ) 0
By replacing Sh (f ) = |H(f )|2 one gets the expression given in (8.7.18). An example
of Be is shown in Fig. 8.9.

8.7.2 White and colored noise

The term white noise refers to noise, which power density spectrum is constant for
all frequencies. In other words:
SY (f ) = A , (8.7.22)
Section 8.7. Linear and non-linear processing of random signals 8-29

where A is constant.
The autocorrelation of this noise for analog signals is

RY (τ ) = A · δ(τ ) . (8.7.23)

The power of such a signal is infinite, implying that such a signal cannot be created
physically. This signal has, however, a number of theoretical uses.
The autocorrelation function of digital white noise is

RY (k) = Aδ(k) , (8.7.24)

and the power of this signal is PY = A. Such a signal can be realized and is often
used in digital signal processing.
Band limited white noise is produced by filtering white noise with a filter with
an appropriate transfer function H(f ). The power density spectrum for this signal
is
SX (f ) =| H(f ) |2 A . (8.7.25)
The autocorrelation function for this is signal is then

RX (τ ) = ARh (τ ) . (8.7.26)

From the expression for power of analog signals

Z ∞ Z ∞
PX = SX (f )df = A | H(f ) |2 d(f ) (8.7.27)
−∞ −∞

we get that
PX = A | H(fk ) |2 Be = 2Be SX (fk ) , (8.7.28)
where Be is the equivalent noise bandwidth found for the characteristic frequency
fk (see sec. 8.7.1).
From this it appears that Be can be considered as the bandwidth - for positive
frequencies - of the bandlimited white noise signal with the power density SX (fk ),
which has the same power as the original signal.
The derivation for digital signals is similar and the result is
Z fg
1 Be
PX = SX (f )df = SX (fk ) , (8.7.29)
2fg −fg fg

where fg is the folding or Nyquist frequency, fg = f2s = 2∆T

1
.
The term pink noise is used for noise, whose spectrum has the shape
A
SX (f ) = , (8.7.30)
|f |

where A is a constant. It obvious that the power PX in this case is also infinitely
large. That’s why the pink noise signals used in practice are bandlimited both at
low and at high frequencies. The spectrum of band-limited pink noise is given by

 A
for fn <| f |< fø
SX (f ) = |f | (8.7.31)
 0 ellers

and the power of this signal is

f
PX = 2A ln ø . (8.7.32)
fn
8-30 Random signals Chapter 8

8.7.3 Narrow band noise

The term narrow band noise is applied to a signal V (t), whose power density spec-
trum is not 0 within a narrow frequency band B centered at the frequency f0 .
Usually the bandwidth will be much smaller than the center frequency, B  f0 .
One practical example of narrow band noise is encountered when one tries to
determine the spectrum of a random signal with the help of bandpass filters.
As in the case of deterministic narrow-band signals, it is convenient to consider
the signal V (t) as a sum of two random signals X(t) and Y (t)

V (t) = X(t) cos 2πf0 t + Y (t) sin 2πf0 t . (8.7.33)

Since t is in the argument of the trigonometric factors, it is not obvious that V (t)
is stationary, even in the case that X(t) and Y (t) are.
The condition for V (t) to be stationary in wide sense,is that the expected value
of V (t) and its autocorrelation function are independent of the zero-point of the
time (see sec. 8.2.2). Since

E{V (t)} = E{X(t)} cos 2πf0 t + E{Y (t)} sin 2πf0 t , (8.7.34)

this implies that

E{X(t)} = 0 and E{Y (t)} = 0 . (8.7.35)
From the expression

RV (τ ) = E{(X(t) cos 2πf0 t + Y (t) sin 2πf0 t)(X(t + τ ) cos 2πf0 (t + τ )

+Y (t + τ ) sin 2πf0 (t + τ ))}
1 1
= (RX (τ ) + RY (τ )) cos 2πf0 τ − (RY X (τ ) − RY X (τ )) sin 2πf0 τ (8.7.36)
2 2
1 1
+ (RX (τ ) − RY (τ )) cos 2πf0 (2 + τ ) + (RY X (τ ) + RXY (τ )) sin 2πf0 (2t + τ ) ,
2 2
it follows that if

RX (τ ) = RY (τ ) and RXY (τ ) = −RY X (τ ) , (8.7.37)

then V (t) will be wide sense stationary signal and that the following equation is
fulfilled:
RV (τ ) = RX (τ ) cos 2πf0 τ − RY X (τ ) sin 2πf0 τ . (8.7.38)
From here we can deduce that the spectra of X(t) and Y (t) will be concentrated in
a region around f = 0 and bandwidth of size on the order of 21 B.
If we now create two new random signals
 
p
2 2
Y (t)
A(t) = X (t) + Y (t) ψ(t) = arctan − (8.7.39)
X(t)

from the signals X(t) and Y (t), it can be seen that V (t) can be written as

V (t) = A(t) cos(2πf0 t + ψ(t)) . (8.7.40)

Also the signals A(t) and ψ(t) will roughly be signals with spectra concentrated
around f = 0.
If the signals X(t) and Y (t) are mutually independent and Gaussian, the one-
dimensional probability density functions of both signals A(t) and ψ(t) will be
( α α2
e− 2σ2 α > 0
wA (α) = σ 2 (8.7.41)
0 α60
Section 8.7. Linear and non-linear processing of random signals 8-31

(Rayleigh distribution, see also Table 8.1).

σ 2 = RV (0) = PV , (8.7.42)

and6
1
(
0 6 β 6 2π
wψ (β) = 2π . (8.7.43)
0 otherwise
1 1
E{A(t)} = ( πσ 2 ) 2 (8.7.44)
2
and
1
σ 2 {A(t)} = (2 − π)σ 2 , (8.7.45)
2
and also
RV (τ ) = RX (τ ) cos 2πf0 τ . (8.7.46)

8.7.4 White noise filtered by a low-pass RC-filter

Let X(t) be analog white noise signal with power density spectrum SX (f ) = A,
where A is a constant. Let X(t) be filtered by an RC-filter with a transfer function:
1
H1 (f ) = (8.7.47)
1 + j2πf RC
The power density spectrum of the output signal Y (t) will be
A
SY (f ) = A | H1 (f ) |2 = h i2 , (8.7.48)
f
1+ f1

where f1 = (2πRC)−1 .
The autocorrelation function of SY (f ) is found by applying the inverse Fourier
transform on SY (f ). It is
A − |τ |
RY (τ ) = e RC (8.7.49)
2RC
(see Fig. 8.10). Then, the total power of the signal is
A
PY = . (8.7.50)
2RC

8.7.5 Integrated random signal

In certain applications (e.g. the estimation of the power of a signal) one needs to
create the signal
1 t
Z
Y (t) = X(θ)dθ , (8.7.51)
T t−T
from the analog signal X(t).
Since integration over the interval T is equivalent to low-pass filtering, Y (t) can
be considered the output of this filter when X(t) is the input signal to the filter.
From this immediately follows that

E{Y (t)} = E{X(t)} . (8.7.52)

From the shape of the impulse response h(t) of the integrator, shown in Fig. 8.11,
it follows that 
(1− | τ | /T )/T | τ |< T
Rh (τ ) = (8.7.53)
0 | τ |> T
8-32 Random signals Chapter 8

Effekttæthedsspektret Autokorrelationsfunktionen
0 6
f1 RC
5
−10

RY(τ)/(A/(2 RC))
4
SY(f)

−20 3

2
−30
1

−40 0
−1 0 1 2 −0.2 −0.1 0 0.1 0.2
10 10 10 10
Frekvens [Hz] τ

Figure 8.10. The power density spectrum and the autocorrelation function of a
RC-filtered white noise.

Impulssvar for h(t) |H(f)|2

1
1
0.8
h(t) ⋅ T

0.6

0.4 0.8

0.2

0
0 0.5 1 1.5 2 0.6
2

t/T
|H(f)|

Autokorrelationsfunktionen

1
0.4
0.8
Rh(τ) ⋅ T

0.6
0.2
0.4

0.2

0 0
−2 −1 0 1 2 −3 −2 −1 0 1 2 3
τ/T f⋅T

Figure 8.11. Impulse response, autocorrelation function and spectrum of an inte-

grator.

according to which
RY (τ ) = Rh (τ ) ∗ RX (τ ) . (8.7.54)
If X(t) is white noise signal with a constant power density A, then RY (τ ) and
SY (f ) become
RY (τ ) = ARh (τ ) (8.7.55)
6 It can be shown, that A(t) and ψ(t) are under these conditions statistically independent
Section 8.7. Linear and non-linear processing of random signals 8-33

Figure 8.12. A schematic of a full-wave rectifier using a center-tapped transformer.

and
sin2 πf T
SY (f ) = A , (8.7.56)
(πf T )2
respectively. The power of the signal Y (t) is

PY = A/T . (8.7.57)

If we consider a digital signal X(n), which is filtered with a digital integrator, then
the output signal Y (n) is given by
n
1 X
Y (n) = X(q) . (8.7.58)
N
q=n−(N −1)

It follows directly that

k


1− N /N | k |< N
Rh (k) = . (8.7.59)
0 ellers

If RX (k) = Aδ(k), then SX (f ) = A, and with that

 2
1 sin πf N ∆T
SY (f ) = A · (8.7.60)
N sin πf ∆T
The total power of the output signal is

PY = A/N . (8.7.61)

8.7.6 The square-low rectifier

Rectifiers are often used in practice, for example in DC power supplies. These
circuits convert a signal which has both negative and positive values to a signal that
has only positive values. An example of a full-wave rectifier is shown in Fig. 8.12.
This section treats a square-low rectifier and the next section a linear one.
The square-low rectifier is a non-linear circuit and finds applications in, for
example, measuring the power of a random signal. The relation between the input
value x and the output value y of the systems is

y = ax2 , (8.7.62)

where a is a constant.
Let X(t) be the input signal to the non-linear system, and Y (t) its corresponding
output. The mean value of Y (t) is

E{Y (t)} = aE{X 2 (t)} = aPX . (8.7.63)

In other words, Y (t) has a DC component, whose magnitude gives the total power
of X(t).
8-34 Random signals Chapter 8

If we use the results from section 8.1.3 we find that since

p
x = ± y/a , (8.7.64)

then ( √ √
wX (− η/a)+wX ( η/a)
wY (η) =
√
2 aη η>0 (8.7.65)
0 n<0

We will set a = 1 in the following for the sake of simplicity.

Given that X(t) is a Gaussian noise signal with E{X(t)} = 0, then
2
1 − 2Rξ
wX (ξ) = p ·e X (0) , (8.7.66)
2π · RX (0)

hence
− 2R η (0)
(
√ 1 √1
· η ·e X η>0
wY (η) = 2πRX (0) . (8.7.67)
0 n<0

This probability density function is illustrated in Fig. ??.

The autocorrelation function RY (τ ) of the output signal Y (t) can be found from

RY (τ ) = E{Y (t)Y (t + τ )}
Z ∞ Z ∞
= E{X 2 (t)X 2 (t + τ )} = ξ12 ξ22 wX (ξ1 , ξ2 ; τ )dξ1 dξ2 (, 8.7.68)
−∞ −∞

where wX (ξ1 , ξ2 ; τ ) is as derived in section 8.2.4. If we insert in this expression a

new variable:
RX (τ )
η = ξ2 − ξ1
RX (0)
then wX can be expressed as the product of two probability density functions
of ξ1 and η, respectively. This, combined with the fact that E{X(t)} = 0 and
E{X 4 (t)} = 3RX
2
(0), and some manipulations in time leads to7
2 2
RY (τ ) = RX (0) + 2RX (τ ) . (8.7.69)

From here it can be seen, that Y (t) has the aforementioned DC-component and the
spectrum of Y (t) is
2
SY (f ) = RX (0)δ(f ) + 2SX (f ) ∗ SX (f ) . (8.7.70)

Figure 8.13 shows an example of the spectrum of Y (t), when X(t) is a bandlimited
white noise (ideal filter).
It must be noted that the continuous portion of SY (f ) will always be 6= 0 in the
vicinity of f = 0, independent of the shape of SX (f ).
When the total power of a signal is measured (or its RMS value), the AC-power
of the signal Y (t) must also be known. This AC power is
2
PAC = 2RX (0) , (8.7.71)

which can be found either from wY (η), from R(τ ) or from SY (f ).

7 This expression appears also from the identity E{X1 (t)X2 (t)X3 (t)X4 (t)} =
E{X1 (t)X2 (t)}E{X3 (t)X4 (t)} + E{X1 (t)X3 (t)}E{X2 (t)X4 (t)} + E{X1 (t)X4 (t)}E{X2 (t)X3 (t)},
which is valid for 4 signals with joint gaussian density function (4-dimensional).
Section 8.7. Linear and non-linear processing of random signals 8-35

Effektæthedsspektrum for båndbegrænset hvid støj

0.8
Sx(f)/k

0.6

0.4

0.2

0
−3 −2 −1 0 1 2 3
f/f0

Effektæthedsspektrum for kvadreret hvid støj

R2X(0) δ(f)
1.2
R2X(0)/f0
1

0.8
Sy(f)

0.6

0.4

0.2

0
−3 −2 −1 0 1 2 3
f/f0

Figure 8.13. Power density spectrum of band-limited white noise, and of squared
band-limited white noise.

8.7.7 The linear full-wave (double) rectifier

The linear full-wave rectifier, shown in Fig. 8.12 has a non-linear characteristic given
by:
y=a|x| , (8.7.72)
where the symbols have the same meaning as in section 8.7.6.
If we also let X(t) be the input signal to and Y (t) the output signal from the
rectifier, then the probability density function becomes
 w (−η/a)+w (η/a)
X X
η>0
wY (η) = a . (8.7.73)
0 η<0

(see section 8.1.3).

Using a = 1 and if X(t) is gaussian, then the probability density is
 2
 √ 2 − η
· e 2RX (0) η>0
wY (η) = 2πRX (0) (8.7.74)
 0 η<0 .

Hence,
∞
Z r
2
E{Y (t)} = η · wY (η)dη = · RX (0) . (8.7.75)
0 π
√
This mean value (DC-value) is proportional to PX .
It can be shown that under these assumptions, the autocorrelation function of
Y (t) is
2 4
2 1 RX (τ ) 1 RX (τ )
RY (τ ) = RX (0) +
π
+
π RX (0) 12π RX 3 (0) + · · · . (8.7.76)

The power density spectrum of Y (t) is found by applying the Fourier transform on
the autocorrelation function RY (τ ). This spectrum too has a DC-component and
an AC-component in the vicinity of the the DC-component (near f = 0).
8-36 Random signals Chapter 8

The AC-power of Y (t) is

2
PAC = (1 − )RX (0) . (8.7.77)
π