EEM 306 Introduction to Communications

Lecture 9

Department of Electrical and Electronics Engineering

Anadolu University

May 20, 2014

Lecture 9 1/22
Last Time

I Random Processes
I Analytic Description
I Statistical Description
I Mean and Autocorrelation Function
I Stationary Processes
I Strictly Stationary
I Wide-Sense Stationary (WSS)
I Cyclostationary
I Ergodic Processes
I Power and Energy
I Multiple Random Processes

Lecture 9 2/22
Random Processes and Linear Systems

Let X(t) be a stationary process.

Theorem: If a stationary process X(t) with mean mX and auto-
correlation function RX (τ ) is passed through an LTI system with
impulse response h(t), the input and output processes X(t) and
Y (t) will be jointly stationary with
Z∞
mY = mX h(t)dt
−∞

RXY (τ ) = RX (τ ) ∗ h(−τ )
RY (τ ) = RX (τ ) ∗ h(τ ) ∗ h(−τ )

Lecture 9 3/22
Random Processes in the Frequency Domain

How can frequency-domain analysis techniques be applied to

the case of LTI systems with random inputs?

Look at the input-output power spectra!

Power Spectrum of Stochastic Processes: Let X(t) denote a ran-
dom process and x(t; wi ) denote a sample function of this process.
Truncate the sample function

x(t; wi ), |t| < T /2
xT (t; wi ) =
0, otherwise

The power spectrum:

|XT (f )|2 E[|XT (f )|2 ]

 
SX (f ) = E lim = lim
T →∞ T T →∞ T

Lecture 9 4/22
Example:
Let the random process X(t) be defined by X(t) = X, where
X is a random variable uniformly distributed on [−1, 1]. Find
SX (f ).
Solution: Consider the truncated random signal
 
t
XT (t) = Xrect
T
Taking the FT
XT (f ) = XT sinc(T f )
The power spectrum
E[|XT (f )|2 ]
SX (f ) = lim = lim E[X 2 ]T sinc2 (T f )
T →∞ T T →∞

Note that E[X 2 ] = 1/3 and lim T sinc2 (T f ) = δ(f ). Thus,

T →∞

1
SX (f ) = δ(f )
3
Lecture 9 5/22
Power Spectrum of Stochastic Processes

Theorem: If X(t) is a stationary process, then

SX (f ) = F{RX (τ )}
To find the total power in the process:
Z∞ " #
1 T /2 2
Z
PX = SX (f )df = E lim X (t)dt
T →∞ T −T /2
−∞

Previous Example:
Let the random process X(t) be defined by X(t) = X, where X
is a random variable uniformly distributed on [−1, 1].
1
RX (τ ) = E[X(t + τ )X(t)] = E[X 2 ] =
3
Hence,  
1 1
SX (f ) = F = δ(f )
3 3
Lecture 9 6/22
Transmission over LTI Systems

We have seen that

Z∞
mY = mX h(t)dt
−∞

RXY (τ ) = RX (τ ) ∗ h(−τ )
RY (τ ) = RX (τ ) ∗ h(τ ) ∗ h(−τ )
Translation of these relations into the frequencyR domain is straight-
∞
forward. Noting that F{h(−τ )} = H ∗ (f ) and −∞ h(t)dt = H(0),
we have
mY = mX H(0)
SXY (f ) = SX (f )H ∗ (f )
SY (f ) = SX (f )|H(f )|2
Since RXY (τ ) = RY X (−τ ), we have
∗
SY X (f ) = SXY (f ) = SX (f )H(f )
Lecture 9 7/22
Example:
The random process X(t) is defined by X(t) = A cos(2πf0 t + θ),
where θ is a random variable uniformly distributed on [0, 2π).
X(t) passes through a differentiator, find SY (f ) and SXY (f )
Solution: Recall that H(f ) = j2πf and

A2
RX (τ ) = cos(2πf0 τ )
2
Taking the FT

A2
SX (f ) = [δ(f − f0 ) + δ(f + f0 )]
4
Hence,
SY (f ) = π 2 f 2 A2 [δ(f − f0 ) + δ(f + f0 )]

−jA2 πf
SXY (f ) = (−j2πf )SX (f ) = [δ(f − f0 ) + δ(f + f0 )]
2
Lecture 9 8/22
Power-Spectral Density of a Sum Process

In practice, we often encounter the sum of two random processes.

For example, in the case communication over a channel with
additive noise.
Assume that Z(t) = X(t) + Y (t) where X(t) and Y (t) are jointly
stationary random processes. Z(t) is a stationary process with

RZ (τ ) = RX (τ ) + RY (τ ) + RXY (τ ) + RY X (τ )

Taking the FT of both sides

SZ (f ) = SX (f ) + SY (f ) + SXY (f ) + SY X (f )
| {z }
∗ (f )
SXY
= SX (f ) + SY (f ) + 2Re[SXY (f )]

Lecture 9 9/22
Gaussian Processes

Gaussian processes play an important role in communication sys-

tems.
Gaussian Random Variable: The Gaussian random variable with
mean m and variance σ 2 has the following pdf
1 (x−m)2
fX (x) = √ e− 2σ2
2πσ

Lecture 9 10/22
Gaussian Processes cont’d

Jointly Gaussian Random Variables:

Define the random vector X = (X1 , X2 , ..., Xn ), and the vector
of means m = (m1 , m2 , ..., mn ), and the n × n covariance matrix
C such that
Ci,j = cov(Xi , Xj ),
then the random variables {Xi } are jointly Gaussian if
 
1 1 −1 T
f (x1 , x2 , ..., xn ) = p exp − (x − m)C (x − m)
(2π)n det(C) 2

Gaussian Processes:
A random process X(t) is a Gaussian process if for all n and
all (t1 , t2 , ..., tn ), the random variables {X(ti )}ni=1 have a jointly
Gaussian density function.

Lecture 9 11/22
Gaussian Processes cont’d

I For a Gaussian processes, knowledge of the mean and

autocorrelation; i.e, mX (t) and RX (t1 , t2 ) gives a complete
statistical description of the process.
I If the Gaussian process X(t) is passed through an LTI
system, then the output process Y (t) will also be Gaussian
process.
I For Gaussian processes, WSS and strict stationarity are
equivalent.
I For jointly Gaussian processes, uncorrelatedness and
independence are equivalent.

Lecture 9 12/22
White Processes

All frequency components appear with equal power!

A process X(t) is called a white process if it has a flat spectral

density, i.e., if SX (f ) is constant for all f .
Power spectral density of a white process:

I Thermal noise can be closely modeled as a white process

I Information sources are modeled as the output of LTI
systems driven by a white process
R∞
The power content of white process: PX = −∞ SX (f )df = ∞
Lecture 9 13/22
White Processes cont’d

Thermal Noise

k: Boltzmann’s constant,1.38 1023 Joules/Kelvin

T: temperature in degrees Kelvin

N0
The value kT is usually denoted by N0 , so Sn (f ) = 2
The autocorrelation function for a white process:
 
−1 N0 N0
Rn (τ ) = F = δ(τ )
2 2

I If we sample a white process at two points t1 and t2

(t1 6= t2 ), the resulting random variables will be
uncorrelated.
Lecture 9 14/22
Example:
Let X(t) be a white noise and
r ZT
2
Y (t) = X(t) cos(2πfc t)dt
T
0

Find the power of the modulated process Y (t).

" #
T RT
RY (t1 , t2 ) = E T2
R
X(t1 )X(t2 ) cos(2πfc t1 ) cos(2πfc t2 )dt1 dt2
0 0

2
RT RT
= T E [X(t1 )X(t2 )] cos(2πfc t1 ) cos(2πfc t2 )dt1 dt2
0 0
2
RT RT N0
= T 2 δ(t1 − t2 ) cos(2πfc t1 ) cos(2πfc t2 )dt1 dt2
0 0
2 N0
RT N0
= T 2 cos2 (2πfc t)dt = 2
0

Lecture 9 15/22
Noise-Equivalent Bandwidth

white Gaussian noise ⇒ F ilter ⇒ Gaussian, not white

The filter characteristics determine the spectral properties of the

output process.
N0
SY (f ) = SX (f )|H(f )|2 = |H(f )|2
2
The power content of the output process:
Z ∞
N0 ∞
Z
PY = SY (f )df = |H(f )|2 df
−∞ 2 −∞
Define Bneq , the noise-equivalent bandwidth of a filter with fre-
quency response H(f ) as
R∞
|H(f )|2 df
Bneq = −∞ 2
2Hmax
Hmax : the maximum of |H(f )| in the passband of the filter.
Lecture 9 16/22
Noise-Equivalent Bandwidth cont’d

For a typical filter:

Thus,
2
PY = N0 Bneq Hmax
The noise-equivalent bandwidth of filters and amplifiers is usually
provided by the manufacturer.

Lecture 9 17/22
Example:
Find the noise-equivalent bandwidth of a lowpass RC filter.
1
Recall the frequency response for this filter H(f ) = 1+j2πf RC
Define τ = RC

1
Answer: Bneq = 4RC

Lecture 9 18/22
Bandpass Processes

Bandpass random processes are the equivalents

of bandpass deterministic signals.

X(t) is a bandpass process,

if SX (f ) = 0 for |f − f0 | ≥ W , where W < f0

We have seen earlier that bandpass signals can be expressed in

terms of equivalent lowpass signals
⇒ Generalize those results to the case of random processes.
Lecture 9 19/22
Bandpass Processes cont’d

Let X(t) be a bandpass process. Then RX (τ ) is a deterministic

bandpass signal. If X(t) is passed through the Hilbert filter with
1
impulse response πt , the output process is the HT of the input
process with
Z∞
1
mY = mX dt
πt
−∞

1
RXY (τ ) = RX (τ ) ∗ = −R̂X (τ )
−πt
RY (τ ) = RX (τ ) Exercise!
Parallel to the deterministic case, define

Xc (t) = X(t) cos(2πf0 t) + X̂(t) sin(2πf0 t)

Xs (t) = X̂(t) cos(2πf0 t) − X(t) sin(2πf0 t)

Lecture 9 20/22
Bandpass Processes cont’d

X(t) = Xc (t) cos(2πf0 t) − Xs (t) sin(2πf0 t)

Properties:
1. If X(t) is zero-mean, then Xc (t) and Xs (t) are also
zero-mean.
2. If X(t) is Gaussian, then Xc (t) and Xs (t) are jointly
Gaussian.
3. If X(t) is WSS, then Xc (t) and Xs (t) are jointly WSS with

RXc (τ ) = RXs (τ ) = RX (τ ) cos(2πf0 τ ) + R̂X (τ ) sin(2πf0 τ )
RXc Xs (τ ) = RX (τ ) sin(2πf0 τ ) − R̂X (τ ) cos(2πf0 τ )

Note the cross correlation function is an odd function,

RXc Xs (0) = 0. This means that at any time instant t0 , the
rvs Xc (t0 ) and Xs (t0 ) are uncorrelated.
Lecture 9 21/22
Bandpass Processes cont’d

4. Power spectrum of the in-phase and quadrature

components

SX (f − f0 ) + SX (f + f0 ), |f | < f0
SXc (f ) = SXs (f ) =
0, otherwise

The cross spectral density

SXc Xs (f ) = F[R
 Xc Xs (τ )]
j [SX (f + f0 ) − SX (f − f0 )] , |f | < f0
=
0, otherwise

5. The power contents

PXc = PXs = RXc (τ )|τ =0 = RX (0) = PX

Lecture 9 22/22