EE450 - Wireless Communications: Professor Ha Nguyen

EE450 Wireless Communications
Professor Ha Nguyen
December 2014
EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam
1/291
INTRODUCTION
2/291
Communication Systems (Analog or Digital)
Source
(User)
Message
signal
Transmitted
signal
Transmitter
Received
signal
Channel
Detected
signal
Receiver
Sink
(User)
Distortion
and noise
Source: Originates a message (e.g., human voice, TV picture, email message)

and converts it to an electrical waveform, referred to as a message signal.
Transmitter (Modulator): Modifies the message signal for efficient transmission.
Channel: A physical medium of choice that can convey the electrical signals at
the transmitter output over a distance.
Receiver (Demodulator): Processes the signal received from the channel by
reversing the signal modifications made at the transmitter and removing the
distortion made by the channel.
Sink: Converts the electrical signal at the output of the receiver to its original
form the message.
3/291
Analog, Digital, Continuous-Time, Discrete-Time?

x(t)
x(t)
(a) Analog continuous-time signal
(b) Digital continuous-time signal
Ts
t
0
Ts
(c) Analog discrete-time signal
(d) Digital discrete-time signal
In signal classification, the adjectives analog and digital refer to the

amplitude property of the signals, while continuous-time and discrete-time
refer to the time property.
continuous-time and discrete-time are also commonly used to refer to the
type of signal processing in hardware.
4/291
What is Digital Communication?

The message has to be in a digital format, i.e., it can be represented as a string
of bits 0 and 1. A message can be analog (such as human voice) or digital (such
as a computer-generated text document). An analog message can always be
represented in a digital format through the processes of sampling and
quantization.
BASK signal Digital message
AM signal
Analog message
Digital Modulation: Over a finite duration, the transmitted signal belongs to a

finite set of continuous-time waveforms. Note that each waveform in the set itself
can be an analog continuous-time signal!
1
0
1
0
0.5
1.5
2.5
t
3.5
4.5
0.5
1.5
2.5
t
3.5
4.5
0.5
1.5
2.5
t
3.5
4.5
0.5
1.5
2.5
t
3.5
4.5
1
0
1
0
1
0
1
0
1
0
1
0
5/291
Received BASKTransmitted BASK Received AM Transmitted AM
Why Digital Communications?

1
0
1
0
0.5
1.5
2.5
t
3.5
4.5
0.5
1.5
2.5
t
3.5
4.5
0.5
1.5
2.5
t
3.5
4.5
0.5
1.5
2.5
t
3.5
4.5
2
0
2
0
1
0
1
0
2
0
2
0
Since the transmitted signal in digital communications belong to a finite set of

waveforms The distorted signal can be recovered to its ideal shape, hence removing
all the noise.
6/291
Analog-to-Digital Conversion = Sampling + Quantization
Sampling does not introduce information loss if it satisfies the Nyquist sampling
theorem.
Quantization always introduces information loss, but the loss can be made
arbitrarily small by increasing the number of quantization levels (i.e., using more
bits).
7/291
Types of Modulation and Hardware Realization (Implementation)
Continuous-time implementation is mostly adopted in this course to explain the

theory of digital communications (e.g., why do we use a certain modulation and
demodulation methods?)
Discrete-time implementation is most common in practice, but only briefly
discussed in this course. Discrete-time implementation has to do with how
efficiently one can build modulation/demodulation methods in hardware.
8/291
Block Diagram of a Digital Communication System

Synchronization
Source
(User)
Transmitter
Channel
Receiver
Sink
(User)
(a)
Transmitter
Source
Encoder
Channel
Encoder
Receiver
Demodulator
Modulator
Channel
Decoder
Source
Decoder
(b)
9/291
Digital vs. Analog Communications
Advantages:
Digital signals are much easier to be regenerated.
Digital circuits are less subject to distortion and interference.
Digital circuits are more reliable and can be produced at a lower cost than analog
circuits.
It is more flexible to implement digital hardware than analog hardware.
Digital signals are beneficial from digital signal processing (DSP) techniques.
Disadvantages:
Heavy signal processing.
Synchronization is crucial.
Larger transmission bandwidth.
Non-graceful degradation.
10/291
REVIEW OF PROBABILITY,
RANDOM VARIABLES, RANDOM
PROCESSES
11/291
Random Experiment, Sample Space, Events
Random experiment: its outcome, for some reason, cannot be predicted with
certainty.
Examples: throwing a die, flipping a coin and drawing a card from a deck.
Sample space: the set of all possible outcomes, denoted by . Outcomes are
denoted by s and each lies in , i.e., .
A sample space can be discrete or continuous.
Events are subsets of the sample space for which measures of their occurrences,
called probabilities, can be defined or determined.
12/291
Example of Throwing a Fair Die
Various events can be defined: the outcome is even number of dots, the outcome
is smaller than 4 dots, the outcome is more than 3 dots, etc.
13/291
Three Axioms of Probability
For a discrete sample space , define a probability measure P on as a set function

that assigns nonnegative values to all events, denoted by E, in such that the
following conditions are satisfied
Axiom 1: 0 P (E) 1 for all E (on a % scale probability ranges from 0 to
100%. Despite popular sports lore, it is impossible to give more than 100%).
Axiom 2: P () = 1 (when an experiment is conducted there has to be an
outcome).
1
Axiom
mutually
S3: For
P exclusive events E1 , E2 , E3 ,. . . we have
E
=
P
(E
).
P
i
i=1 i
i=1
The events E1 , E2 , E3 ,. . . are mutually exclusive if Ei Ej = for all i 6= j, where is the null set.
14/291
Important Properties of the Probability Measure
1. P (E c ) = 1 P (E), where E c denotes the complement of E. This property

implies that P (E c ) + P (E) = 1, i.e., something has to happen.
2. P () = 0 (again, something has to happen).
3. P (E1 E2 ) = P (E1 ) + P (E2 ) P (E1 E2 ). Note that if two events E1 and
E2 are mutually exclusive then P (E1 E2 ) = P (E1 ) + P (E2 ), otherwise the
nonzero common probability P (E1 E2 ) needs to be subtracted off.
4. If E1 E2 then P (E1 ) P (E2 ). This says that if event E1 is contained in E2

then occurrence of E1 means E2 has occurred but the converse is not true.
15/291
Conditional Probability
We observe or are told that event E1 has occurred but are actually interested in
event E2 : Knowledge that E1 has occurred generally changes the probability of
E2 occurring.
If it was P (E2 ) before, it now becomes P (E2 |E1 ), the probability of E2
occurring given that event E1 has occurred.
This conditional probability is given by
P (E2 E1 )
,
P (E2 |E1 ) =
P (E1 )
0,
if P (E1 ) 6= 0
otherwise
If P (E2 |E1 ) = P (E2 ), or P (E2 E1 ) = P (E1 )P (E2 ), then E1 and E2 are said
to be statistically independent.
Bayes rule
P (E2 |E1 ) =
P (E1 |E2 )P (E2 )

,
P (E1 )
16/291
Total Probability Theorem
The events {Ei }n

i=1 partition the sample space if:
(i)
n
[
Ei =
(1a)
i=1
(ii)
Ei Ej =
for all 1 i, j n and i 6= j
(1b)
If for an event A we have the conditional probabilities {P (A|Ei )}n

i=1 , P (A) can
be obtained as
n
X
P (A) =
P (Ei )P (A|Ei ).
i=1
Bayes rule:
P (Ei |A) =
P (A|Ei )P (Ei )
P (A|Ei )P (Ei )
= Pn
.
P (A)
j=1 P (A|Ej )P (Ej )
17/291
Examples on Probability Measures
Example 1: In the example of throwing a fair die, based on the three axioms of
probability, it is easy to show that the probability of each face occurring is 1/6.
Example 2: In the example of throwing a fair die, one can define the following
events:
E1 = the outcome is even
E2 = the outcome is smaller than 4
Compute P (E1 ), P (E2 ), P (E2 |E1 ), P (E1 |E2 ). Are the events E1 and E2
statistically independent?
18/291
Random Variables
x(4 )
x(1 )
x(3 )
x(2 )
A random variable is a mapping from the sample space to the set of real
numbers.
We shall denote random variables by boldface, i.e., x, y, etc., while individual or
specific values of the mapping x are denoted by x().
19/291
Random Variable in the Example of Throwing a Fair Die
R
1
There could be many other random variables defined to describe the outcome of this
random experiment!
20/291
Cumulative Distribution Function (cdf)
cdf gives a complete description of the random variable. It is defined as:

Fx (x) = P ( : x() x) = P (x x).
The cdf has the following properties:
1. 0 Fx (x) 1 (this follows from Axiom 1 of the probability measure).
2. Fx (x) is nondecreasing: Fx (x1 ) Fx (x2 ) if x1 x2 (this is because event
x() x1 is contained in event x() x2 ).
3. Fx () = 0 and Fx (+) = 1 (x() is the empty set, hence an impossible
event, while x() is the whole sample space, i.e., a certain event).
4. P (a < x b) = Fx (b) Fx (a).
21/291
Typical Plots of cdf

A random variable can be discrete, continuous or mixed.
Fx ( x )
1.0
(a)
0
Fx ( x )
1.0
(b)
22/291
Fx ( x )
1.0
(c)
23/291
Probability Density Function (pdf)
The pdf is defined as the derivative of the cdf:

fx (x) =
dFx (x)
.
dx
It follows that:
P (x1 x x2 ) = P (x x2 ) P (x x1 )
Z x2
= Fx (x2 ) Fx (x1 ) =
fx (x)dx.
x1
Basic properties of pdf:

1. f
(x) 0.
Rx
fx (x)dx = 1.
2.
R
3. In general, P (x A) = A fx (x)dx.
For discrete random variables, it is more common to define the probability mass
function (pmf): pi = P (x = xi ).
X
Note that, for all i, one has pi 0 and
pi = 1.
i
24/291
Bernoulli Random Variable
fx ( x )
Fx ( x )
1
1 p
(1 p )
( p)
A discrete random variable that takes two values 1 and 0 with probabilities p and
1 p.
Good model for a binary data source whose output is 1 or 0.
Can also be used to model the channel errors.
25/291
Binomial Random Variable

fx ( x )
0.30
0.25
0.20
0.15
0.10
0.05
x
0
A discrete random variable that gives the number of 1s in a sequence of n

independent Bernoulli trials.
n
n
X
n!
n k
=
p (1 p)nk (x k), where
fx (x) =
.
k
k
k!(n
k)!
k=0
26/291
Uniform Random Variable
fx ( x )
Fx ( x )
1
1
ba
x
a
x
a
A continuous random variable that takes values between a and b with equal
probabilities over intervals of equal length.
The phase of a received sinusoidal carrier is usually modeled as a uniform random
variable between 0 and 2. Quantization error is also typically modeled as
uniform.
27/291
Gaussian (or Normal) Random Variable
fx ( x )
Fx ( x )
1
2 2
1
1
2
x
0
A continuous random variable whose pdf is:

1
(x )2
fx (x) =
exp
,
2
2
2
2
and 2 are parameters. Usually denoted as N (, 2 ).
Most important and frequently encountered random variable in communications.
28/291
Outcome
Outcome
Uniform or Gaussian?
0
1
2
0
100
50
Trial number
100
50
Trial number
100
5
Outcome
Outcome
1
2
50
Trial number
5
0
50
Trial number
100
5
0
29/291
Gaussian RVs with Different Average (Mean) Values

0.4
6
4
Outcome
f x(x)
0.3
0.2
2
0
2
0.1
4
0
5
6
0
20
40
60
80
100
80
100
Trial number
0.4
6
4
Outcome
f x(x)
0.3
0.2
2
0
2
0.1
4
0
5
6
0
20
40
60
Trial number
30/291
Gaussian RVs with Different Average (Mean) Squared Values

0.4
6
4
Outcome
f x(x)
0.3
0.2
2
0
2
0.1
4
0
5
6
0
20
40
60
80
100
80
100
Trial number
0.4
6
4
Outcome
f x(x)
0.3
0.2
2
0
2
0.1
4
0
5
6
0
20
40
60
Trial number
31/291
Histogram of Observed Values

300
2
Count
Outcome
0
2
4
0
500
Trial number
0
Bin
150
2
Count
Outcome
100
0
4
1000
0
2
4
0
200
500
Trial number
1000
100
50
0
2
0
Bin
32/291
Q-function

2

1 2
e
2

1
Q(x)
2

Area = Q ( x )
exp
2
2
d.
10
10
Q(x)
10
10
10
10
10
3
x
33/291
Functions of A Random Variable
The function y = g(x) is itself a random variable.

From the definition, the cdf of y can be written as
Fy (y) = P ( : g(x()) y).
Assume that for all y, the equation g(x) = y has a countable number of solutions
and at each solution point, dg(x)/dx exists and is nonzero. Then the pdf of
y = g(x) is:
X
fx (xi )
,

fy (y) =

i dg(x)
dx x=x
i
where {xi } are the solutions of g(x) = y.
A linear function of a Gaussian random variable is itself a Gaussian random

variable.
34/291
Expectation of Random Variables

Statistical averages, or moments, play an important role in the characterization of the
random variable.
The expected value (also called the mean value, first moment) of the random
variable x is defined as
Z
xfx (x)dx,
mx = E{x}
where E denotes the statistical expectation operator.

In general, the nth moment of x is defined as
Z
E{xn }
xn fx (x)dx.
For n = 2,
E{x2 }
is known as the mean-squared value of the random variable.
The nth central moment of the random variable x is:

Z
E{y} = E{(x mx )n } =
(x mx )n fx (x)dx.
2:
When n = 2 the central moment is called the variance, commonly denoted as x
Z
2
(x mx )2 fx (x)dx.
x
= var(x) = E{(x mx )2 } =
35/291
The variance provides a measure of the variables randomness.

The mean and variance of a random variable give a partial description of its pdf.
Relationship between the variance, the first and second moments:
2
x
= E{x2 } [E{x}]2 = E{x2 } m2x .
An electrical engineering interpretation: The AC power equals total power minus

DC power.
The square-root of the variance is known as the standard deviation, and can be
interpreted as the root-mean-squared (RMS) value of the AC component.
36/291
Example
The noise voltage in an electric circuit can be modeled as a Gaussian random variable
with zero mean and variance 2 .
(a) Show that the probability
that the noise voltage exceeds some level A can be

.
expressed as Q A
(b) Given that the noise voltage is negative, express in terms of the Q function the
probability that the noise voltage exceeds A, where A < 0.
(c) Using the MATLAB function Q.m available from the courses website, evaluate
the expressions found in (a) and (b) for 2 = 108 and A = 104 .
37/291
Solution
fx (x) =
1
2
(x)2 (=0)
22
=
1
2
2
x2
2
The Q-function is defined as:

1
Q(t) =
2
2
2
The probability that the noise voltage exceeds some level A is

Z
fx (x)dx
P (x > A) =
A
=
=
Z
2
1
x
e 22 dx
2 A

A
.
Q
x)
(=
)
(d= dx
2
2
38/291
Given the noise voltage is negative, the probability that the noise voltage exceeds
some level A (A < 0) is:
P (x > A|x < 0)
P (A < x < 0)
P (x > A , x < 0)
=
P (x < 0)
P (x < 0)

A
A

Q Q(0)
Q 12
A
=
= 2Q
1.
Q(0)
1/2
f x ( x | x < 0)
2
2 2
Area=P (x > A | x < 0)
fx ( x )
1
2 2
With 2 = 108 andA = 104

,
4
P (x > 104 ) = Q 10 8 = Q(1) = 0.8413 and
10
P (x > 104 |x < 0) = 2Q(1) 1 = 0.6827.

39/291
Example
The Matlab function randn defines a Gaussian random variable of zero mean and unit
variance. Generate a vector of L = 106 samples according to a zero-mean Gaussian
random variable x with variance 2 = 108 . This can be done as follows:
x=sigma*randn(1,L).
(a) Based on the sample vector x, verify the mean and variance of random variable x.
(b) Based on the sample vector x, compute the probability that x > A, where
A = 104 . Compare the result with that found in Question 2-(c).
(c) Using the Matlab command hist, plot the histogram of sample vector x with 100
bins. Next obtain and plot the pdf from the histogram. Then also plot on the
same figure the theoretical Gaussian pdf (note the value of the variance for the
pdf). Do the histogram pdf and theoretical pdf fit well?
40/291
Solution
L=10^6; % Length of the noise vector
sigma=sqrt(10^(-8)); % Standard deviation of the noise
A=-10^(-4);
x=sigma*randn(1,L);
%[2] (a) Verify mean and variance:
mean_x=sum(x)/L; % this is the same as mean(x)
variance_x=sum((x-mean_x).^2)/L; % this is the same as var(x);
% mean_x = -3.7696e-008
% variance_x = 1.0028e-008
%[3] (b) Compute P(x>A)
P=length(find(x>A))/L;
% P = 0.8410
%[5] (c) Histogram and Gaussian pdf fit
N_bins=100;
[y,x_center]=hist(x,100); % This bins the elements of x into N_bins equally spaced containers
% and returns the number of elements in each container in y,
% while the bin centers are stored in x_center.
dx=(max(x)-min(x))/N_bins; % This gives the width of each bin.
hist_pdf= (y/L)/dx; % This approximate the pdf to be a constant over each bin;
pl(1)=plot(x_center,hist_pdf,color,blue,linestyle,--,linewidth,1.0);
hold on;
x0=[-5:0.001:5]*sigma; % this specifies the range of random variable x
true_pdf=1/(sqrt(2*pi)*sigma)*exp(-x0.^2/(2*sigma^2));
pl(2)=plot(x0,true_pdf,color,r,linestyle,-,linewidth,1.0);
xlabel({\itx},FontName,Times New Roman,FontSize,16);
ylabel({\itf}_{\bf x}({\itx}),FontName,Times New Roman,FontSize,16);
legend(pl,pdf from histogram,true pdf,+1);
set(gca,FontSize,16,XGrid,on,YGrid,on,GridLineStyle,:,MinorGridLineStyle,none,FontName,Times New Roman);
41/291
4000
pdf from histogram
true pdf
3500
3000
f (x)
2500
2000
1500
1000
500
0
5
0
x
5
x 10
42/291
Multiple Random Variables
Often encountered when dealing with combined experiments or repeated trials of

a single experiment.
Multiple random variables are basically multidimensional functions defined on a
sample space of a combined experiment.
Examples:
Consider the random experiment of launching a dart on a circular dartboard. The
random variables x and y can be used to map a (physical) point (i ) on the dartboard
to a point in the plane, which is within the unit circle: {x(i ), y(i )}.
Consider choosing a student at random from a population of our university students.
Two random variables can be defined to map from the sample space of university
students to the measurements of height and weight: h(i ) and w(i ).
43/291
Joint cdf and Joint pdf

Let x and y be the two random variables defined on the same sample space .
The joint cumulative distribution function is defined as
Fx,y (x, y) = P (x x, y y).
Similarly, the joint probability density function is:
fx,y (x, y) =
2 Fx,y (x, y)
.
xy
When the joint pdf is integrated over one of the variables, one obtains the pdf of
other variable, called the marginal pdf:
Z
fx,y (x, y)dx = fy (y),
Z
fx,y (x, y)dy = fx (x).
Note that:
Z
fx,y (x, y)dxdy = F (, ) = 1
Fx,y (, ) = Fx,y (, y) = Fx,y (x, ) = 0.
44/291
The conditional pdf of the random variable y, given that the value of the random
variable x is equal to x, is defined as
fx,y (x, y)
, fx (x) 6= 0
.
fy (y|x) =
fx (x)
0,
otherwise
Two random variables x and y are statistically independent if and only if
fy (y|x) = fy (y)
or equivalently
The joint moment is defined as

Z
E{xj yk } =
fx,y (x, y) = fx (x)fy (y).
xj y k fx,y (x, y)dxdy.
The joint central moment is

E{(x mx )j (y my )k } =
(x mx )j (y my )k fx,y (x, y)dxdy
where mx = E{x} and my = E{y}.
45/291
The most important moments are

Z Z
xyfx,y (x, y)dxdy
E{xy}
cov{x, y}
(correlation)
E{(x mx )(y my )}
E{xy} mx my
(covariance).
2 and 2 be the variance of x and y. The covariance normalized w.r.t.

Let x
y
x y is called the correlation coefficient:
cov{x, y}
x,y =
.
x y
x,y indicates the degree of linear dependence between two RVs.

It can be shown that |x,y | 1.
x,y = 1 implies an increasing/decreasing linear relationship.

If x,y = 0, x and y are said to be uncorrelated.
It is easy to verify that if x and y are independent, then x,y = 0: Independence

implies lack of correlation.
However, lack of correlation (no linear relationship) does not in general imply
statistical independence.
46/291
Examples of Uncorrelated and Correlated Random Variables
1
y
1.5
1.5
0.5
0
0
0.5
0.5
1.5
0
0
0.5
1.5
47/291
Other Examples of Uncorrelated and Correlated Random Variables
5
5
3
2
2
1
5
2
0
x
48/291
Interpretation of Correlation Coefficient

(a) x,y 0.71
(b) x,y 0.71
100
y
150
1.5
0.5
0
0
50
0.5
0
0
1.5
(c) x,y 0.97
50
100
150
x
(d) x,y 0.97
1.5
0.5
1
0
0.5
0.5
0
0
0.5
1.5
0.5
1.5
49/291
Examples of Uncorrelated Dependent Random Variables
Example 1: Let x be a discrete random variable that takes on {1, 0, 1} with

probabilities { 14 , 21 , 14 }, respectively. The random variables y = x3 and z = x2
are uncorrelated but dependent.
Example 2: Let x be an uniformly random variable over [1, 1]. Then the
random variables y = x and z = x2 are uncorrelated but dependent.
Example 3: Let x be a Gaussian random variable with zero mean and unit
variance (standard normal distribution). The random variables y = x and z = |x|
are uncorrelated but dependent.
Example 4: Let u and v be two random variables (discrete or continuous) with
the same probability density function. Then x = u v and y = u + v are
uncorrelated dependent random variables.
50/291
Example 1
x {1, 0, 1} with probabilities {1/4, 1/2, 1/4}

y = x3 {1, 0, 1} with probabilities {1/4, 1/2, 1/4}
z = x2 {0, 1} with probabilities {1/2, 1/2}
Since both y and z are related (generated) from x, they are dependent. To see if they
are uncorrelated, we will check whether cov{y, z} = 0.
my = (1) 14 + (0) 12 + (1) 14 = 0; mz = (0) 12 + (1) 12 = 21 .
The joint pmf (similar to pdf) of y and z:
1
2
1
4
1
4
P (y = 1, z = 0) = 0
P (y = 1, z = 1) = P (x = 1) = 1/4
P (y = 0, z = 0) = P (x = 0) = 1/2
P (y = 0, z = 1) = 0
1
0
P (y = 1, z = 0) = 0
P (y = 1, z = 1) = P (x = 1) = 1/4
Therefore, E{yz} = (1)(1) 14 + (0)(0) 12 + (1)(1) 14 = 0

cov{y, z} = E{yz} my mz = 0 (0)1/2 = 0! Thus y and z are uncorrelated.
51/291
Jointly Gaussian Distribution (Bivariate)
fx,y (x, y) =
"
1
exp
q
2x y 1 2x,y
1
2(1 2x,y )
(x mx )2
(y my )2
2x,y (x mx )(y my )
+
2
2
x
x y
y
#
where mx , my , x , y are the means and variances.

x,y is indeed the correlation coefficient.
2 ) and f (y) N (m , 2 ).
Marginal density is Gaussian: fx (x) N (mx , x
y
y
y
When x,y = 0 fx,y (x, y) = fx (x)fy (y) random variables x and y are
statistically independent.
For joint Gaussian random variables, uncorrelatedness means statistically
independent!
Weighted sum of two jointly Gaussian random variables is also Gaussian.
52/291
Joint pdf and Contours for x = y = 1 and x,y = 0
=0
x,y
0.15
2.5
x,y
f (x,y)
0.1
0.05
0
0
1
2
0
2
y
3
2
2.5
0
x
53/291
Joint pdf and Contours for x = y = 1 and x,y = 0.3
=0.30
x,y
a crosssection
0.15
2.5
fx,y(x,y)
0.1
1
0.05
0
0
1
2
0
2
y
3
2
2.5
0
x
54/291
=0.70
x,y
0.2
2.5
2
a crosssection
0.1
0.05
fx,y(x,y)
0.15
0
1
2
0
2
y
3
2
2.5
0
x
55/291
=0.95
x,y
0.5
2.5
0.4
fx,y(x,y)
0.3
1
0.2
y
0.1
0
1
2
0
2
y
3
2
2.5
0
x
56/291
Multivariate Gaussian pdf
Define
x = [x1 , x2 , . . . , xn ], a vector of the means
m = [m1 , m2 , . . . , mn ], and
the n n covariance matrix C with
Ci,j = cov(xi , xj ) = E{(xi mi )(xj mj )}.
The random variables {xi }n
i=1 are jointly Gaussian if:
fx1 ,x2 ,...,xn (x1 , x2 , . . . , xn ) = p
(2)n det(C)

1
x
m)C 1 (
x
m) .
exp (
2
If C is diagonal (i.e., the random variables {xi }n

i=1 are all uncorrelated), the joint
pdf is a product of the marginal pdfs: Uncorrelatedness implies statistical
independent for multiple Gaussian random variables.
57/291
Random Processes
2
Real number
x(t,)
x(t ,)
k
x (t, )
1
x2(t,2)
x (t, )
M
...
.
.
.
.
.
.
.
.
.
.
.
.
Time
A mapping from a sample space to a set of time functions.

Ensemble: The set of possible time functions that one sees.
Denote this set by x(t), where the time functions x1 (t, 1 ), x2 (t, 2 ), x3 (t, 3 ),
. . . are specific members of the ensemble.
58/291
At any time instant, t = tk , we have random variable x(tk ).

At any two time instants, say t1 and t2 , we have two different random variables
x(t1 ) and x(t2 ).
Any relationship between them is described by the joint pdf
fx(t1 ),x(t2 ) (x1 , x2 ; t1 , t2 ).
A complete description of the random process is determined by the joint pdf
fx(t1 ),x(t2 ),...,x(tN ) (x1 , x2 , . . . , xN ; t1 , t2 , . . . , tN ).
The most important joint pdfs are the first-order pdf fx(t) (x; t) and the
second-order pdf fx(t1 )x(t2 ) (x1 , x2 ; t1 , t2 ).
59/291
Examples of Random Processes
(b) Uniform phase
(a) Thermal noise
60/291
(d) Binary random data

+V
V
(c) Rayleigh fading process
Tb
61/291
Classification of Random Processes
Based on whether its statistics change with time: the process is non-stationary or
stationary.
Different levels of stationarity:
Strictly stationary: the joint pdf of any order is independent of a shift in time.
N th-order stationarity: the joint pdf does not depend on the time shift, but depends on
time spacings:
fx(t1 ),x(t2 ),...x(tN ) (x1 , x2 , . . . , xN ; t1 , t2 , . . . , tN ) =
fx(t1 +t),x(t2 +t),...x(tN +t) (x1 , x2 , . . . , xN ; t1 + t, t2 + t, . . . , tN + t).
The first- and second-order stationarity:

fx(t1 ) (x, t1 ) = fx(t1 +t) (x; t1 + t) = fx(t) (x)
fx(t1 ),x(t2 ) (x1 , x2 ; t1 , t2 ) = fx(t1 +t),x(t2 +t) (x1 , x2 ; t1 + t, t2 + t)
= fx(t1 ),x(t2 ) (x1 , x2 ; ),
= t2 t1 .
62/291
Statistical Averages or Joint Moments
Consider N random variables x(t1 ), x(t2 ), . . . x(tN ). The joint moments of these
random variables is
Z
Z
E{xk1 (t1 ), xk2 (t2 ), . . . xkN (tN )} =

x1 =
xN =
xk1 1 xk2 2 xNN fx(t1 ),x(t2 ),...x(tN ) (x1 , x2 , . . . , xN ; t1 , t2 , . . . , tN )

dx1 dx2 . . . dxN ,
for all integers kj 1 and N 1.
Shall only consider the first- and second-order moments, i.e., E{x(t)}, E{x2 (t)}
and E{x(t1 )x(t2 )}. They are the mean value, mean-squared value and
(auto)correlation.
63/291
Mean Value or the First Moment
The mean value of the process at time t is

Z
mx (t) = E{x(t)} =
xfx(t) (x; t)dx.
The average is across the ensemble and if the pdf varies with time then the mean
value is a (deterministic) function of time.
If the process is stationary then the mean is independent of t or a constant:
Z
xfx (x)dx.
mx = E{x(t)} =
64/291
Mean-Squared Value or the Second Moment
This is defined as
MSVx (t) = E{x2 (t)} =
MSVx = E{x2 (t)} =
x2 fx(t) (x; t)dx (non-stationary),
x2 fx (x)dx (stationary).
The second central moment (or the variance) is:

2
x
(t) = E [x(t) mx (t)]2 = MSVx (t) m2x (t) (non-stationary),
2
x

E [x(t) mx ]2 = MSVx m2x (stationary).
65/291
Correlation
The autocorrelation function completely describes the power spectral density of

the random process.
Defined as the correlation between the two random variables x1 = x(t1 ) and
x2 = x(t2 ):
Rx (t1 , t2 ) = E{x(t1 )x(t2 )}
Z
Z
=
x1 x2 fx1 ,x2 (x1 , x2 ; t1 , t2 )dx1 dx2 .
x1 =
x2 =
For a stationary process:

Rx ( ) = E{x(t)x(t + )}
Z
Z
x1 x2 fx1 ,x2 (x1 , x2 ; )dx1 dx2 .
=
x1 =
x2 =
Wide-sense stationarity (WSS) process: E{x(t)} = mx for any t, and

Rx (t1 , t2 ) = Rx ( ) for = t2 t1 .
66/291
Properties of the Autocorrelation Function
1. Rx ( ) = Rx ( ). It is an even function of because the same set of product

values is averaged across the ensemble, regardless of the direction of translation.
2. |Rx ( )| Rx (0). The maximum always occurs at = 0, though there maybe
other values of for which it is as big. Further Rx (0) is the mean-squared value
of the random process.
3. If for some 0 we have Rx (0 ) = Rx (0), then for all integers k,
Rx (k0 ) = Rx (0).
4. If mx 6= 0 then Rx ( ) will have a constant component equal to m2x .
5. Autocorrelation functions cannot have an arbitrary shape. The restriction on the

shape arises from the fact that the Fourier transform of an autocorrelation
function must be greater than or equal to zero, i.e., F {Rx ( )} 0.
67/291
Example of a Random Process

A random process is generated as follows: x(t) = ea|t| , where a is a random variable
with pdf fa (a) = u(a) u(a 1) (1/seconds).
(a) Sketch several members of the ensemble.
(b) For a specific time, t, over what values of amplitude does the random variable
x(t) range?
(c) Find the mean and mean-squared value of x(t). Is the process x(t) wide-sense
stationary (WSS)?
(d) Determine the first-order pdf of x(t).
68/291
(a) Plot of several member functions of x(t) = ea|t| :

a=0.82141
1
0.5
0
2
1.5
0.5
0
0.5
t
a=0.4447
1.5
1
0.5
0
2
1.5
0.5
0
0.5
t
a=0.61543
1.5
1
0.5
0
2
1.5
0.5
0
0.5
t
a=0.79194
1.5
1
0.5
0
2
1.5
0.5
1.5
0
t
0.5
69/291
(b) Since a ranges between 0 and 1, for a specific time, t, the random variable x(t)
ranges from e|t| (when a = 1) to 1 (when a = 0).
(c) x is considered as a function of a with t a parameter. The mean and
mean-squared values of x(t) are:
E{x(t)}
x(t)fa (a)da =
x (t)fa (a)da =
ea|t| da =
E{x (t)}
Z1

1
1 e2|t| .
2|t|
Z1
0
2a|t|

1
ea|t| 1
1 e|t| .
=
|t| 0
|t|
da =

e2a|t| 1
2|t| 0
Since both the mean and mean-squared value are functions of time t, the process
x(t) is not wide-sense stationary.
70/291
(d) For x e|t| or x > 1, we have fx(t) (x; t) = 0.
For e|t| < x 1, the equation x = g(a) = ea|t| has only one solution
1
a = |t|
ln x, e|t| < x 1. Furthermore

dg(a)

= |t|ea|t|
= |t|x.

da

a= 1 ln x
a= 1 ln x
|t|
|t|
Therefore
To conclude:
1
fa (a = |t|
ln x)
1
=
.
fx(t) (x; t) =

|t|x
dg(a)

da

a= 1 ln x
|t|
fx(t) (x; t) =
1
,
|t|x
0,
e|t| < x 1
otherwise
71/291
Power Spectral Density of a Random Process; Effects of an LTI System on

Random Processes
Power Spectral Density (PSD): Determine how the average power of the process
is distributed in frequency.
It can be shown that the power spectral density and the autocorrelation function
are a Fourier transform pair:
Z
Z
Rx ( )ej2f d.
Sx (f )ej2f df Sx (f ) =
Rx ( ) =
=
f =
Input
x(t )
Linear, Time-Invariant
(LTI) System
h(t )
H( f )
Output
y (t )
Rx , y ( )
my , Ry ( )
Sy ( f )
mx , Rx ( )
Sx ( f )
Z
my
E{y(t)} = E
Sy (f )
Ry ( )
|H(f )|2 Sx (f )

h()x(t )d = mx H(0)
h( ) h( ) Rx ( ).
72/291
Thermal Noise in Communication Systems
A natural noise source is thermal noise, whose amplitude statistics are well
modeled to be Gaussian with zero mean.
The autocorrelation and PSD are well modeled as:
Rw ( )
Sw (f )
e| |/t0
t0
2kG
1 + (2f t0 )2
kG
(watts),
(watts/Hz).
where k = 1.38 1023 joule/0 K is Boltzmanns constant, G is conductance of

the resistor (mhos); is temperature in degrees Kelvin; and t0 is the statistical
average of time intervals between collisions of free electrons in the resistor (on
the order of 1012 sec).
73/291
(a) Power Spectral Density, Sw(f)

N /2
S (f) (watts/Hz)
White noise
Thermal noise
0
15
10
0
f (GHz)
10
15
(b) Autocorrelation, R ()
w
N /2()
R () (watts)
0.1
White noise
Thermal noise
0.08 0.06 0.04 0.02
0
0.02
(picosec)
0.04
0.06
0.08
0.1
74/291
The noise PSD is approximately flat over the frequency range of 0 to 10 GHz
let the spectrum be flat from 0 to :
Sw (f ) =
N0
2
(watts/Hz),
where N0 = 4kG is a constant.

Noise that has a uniform spectrum over the entire frequency range is referred to
as white noise
The autocorrelation of white noise is
Rw ( ) =
N0
( )
2
(watts).
Since Rw ( ) = 0 for 6= 0, any two different samples of white noise, no matter

how close in time they are taken, are uncorrelated.
Since the noise samples of white noise are uncorrelated, if the noise is both white
and Gaussian (for example, thermal noise) then the noise samples are also
independent.
75/291
Example
Suppose that a (WSS) white noise process, x(t), of zero-mean and power spectral
density N0 /2 is applied to the input of the filter.
(a) Find and sketch the power spectral density and autocorrelation function of the
random process y(t) at the output of the filter.
(b) What are the mean and variance of the output process y(t)?
L
x(t )
y (t )
76/291
H(f ) =
Sy (f ) =
R
1
=
.
R + j2f L
1 + j2f L/R
N0
N0 R (R/L)| |
1
Ry ( ) =
e
.

2 1 + 2L 2 f 2
4L
R
Ry ( ) (watts)
S y ( f ) (watts/Hz)
N0
2
N0R
4L
f (Hz)
(sec)
77/291
Bandlimiting White Gaussian Noise

The input noise x(t) is modeled as a wide-sense stationary, white, Gaussian random
process with a zero mean and two-sided power spectral density N0 /2.
Low-pass filter
1
x(t )
H( f )
y (t )
(a) Find and sketch the power spectral density of y(t).

(b) Find and sketch the autocorrelation function of y(t).
(c) What are the average DC level and the average power of y(t)?
(d) Suppose that the output noise is sampled every Ts seconds to obtain the noise
samples y(kTs ), k = 0, 1, 2, . . .. Find the smallest values of Ts so that the noise
samples are statistically independent. Explain.
(a) The power spectral density (PSD) of y(t) is
Sy (f )
Sx (f ) |H(f )|2 =
N0
,
2
0,
|f | W
otherwise
78/291
Sx ( f )
Sy ( f )
N0
2
N0
2
f
(b) The autocorrelation can be found as the inverse Fourier transform of the PSD:
Ry ( )
=
=
=
F 1 {Sy (f )} =
N0
4
Sy (f )ej2f df =
ej2f (2 )df
x=2 f
N0
4
W
W
N0 j2f
e
df
2
2W
ejx dx
2W
N0 sin(2W )
N0
2 sin(2W ) =
= N0 W sinc(2W )
4
2
79/291
1
0.8
Ry()/(N0W)
0.6
0.4
0.2
0
0.2
0.4
3
0
W
(c) The DC level of y(t) is

my = E{y(t)} = mx H(0) = 0 1 = 0
The average power of y(t) is
2
y
= E{y2 (t)} = Ry (0) = N0 W
(d) Since the input x(t) is a Gaussian process, the output y(t) is also a Gaussian
process and the samples y(kTs ) are Gaussian random variables. For Gaussian
random variables, statistical independence is equivalent to uncorrelatedness. Thus
one needs to find the smallest value for (or Ts ) so that the autocorrelation is
zero. From the graph of Ry ( ), the answer is:
min = (Ts )min =
1
2W
80/291
Random Sequences
A random sequence can be obtained by sampling a continuous-time random
process. How to characterize a random sequence?
Let x[1], x[2], . . . be a sequence of indexed random variables. Various definitions
of continuous-time random processes can be applied, except that the time
variable t is replaced by the sequence index n.
Mean function: mx [n] = E{x[n]}
2
Variance function: x
[n] = E{(x[n] mx [n])2 }
Correlation function: Rx (n, k) = E{x[n]x[k]}
Covariance function: Cx (n, k) = E{(x[n] mx [n])(x[k] mx [k]}
A Gaussian random sequence is a sequence of random variables, for which any

finite number of members of the sequence are jointly Gaussian.
When the autocorrelation is only a function of only the difference between n and
k, i.e., Rx (n, k) = Rx [n k], or Rx [m] = E{x[n]x[n m]}, then the DTFT of
Rx [m] gives the power spectral density (PSD) of the random sequence x[n]:
Sx (ej
)=
Rx [m]ej m
m=
An important case is the white random sequence, where Rx [k] = 2 [k], and
) = 2 : The sequence is completely uncorrelated and all the average
Sx (ej
power in the sequence is equally shared by all frequencies in the sequence!
81/291
Effects of an LTI System on Random Sequences
h[n ]
H (e j )
x[n ]
y[ n ]
Rx ,y [k ]
my
my , Ry [k ]
mx , Rx [k ]
Sx (e j )
S y (e j )
X

E{y[n]} = E
h[k]x[n k] = mx
h[k] = mx H(ej
)
=0
Sy (ej
)
Ry [k]
k=
) Sx (ej
)
H(ej
h[k] h[k] Rx [k].
82/291
Example: Digital Filtering of a Random Sequence

Rx [k ]
Let x[n] be a white random sequence with
2 [n] = 2 . Suppose x[n] is the
mx [n] = 0 and x
input to a digital filter (LTI system) with impulse
response h[n] = an u[n]. Find and sketch the power
spectral density and autocorrelation function of the
output sequence y[n].
Sx (e j )
Solution: The frequency response of this filter is

j
H(e
)=
n=0
Ry [k ]
n jn
a e
1
.
=
1 aej
) = 2 , it follows that
Since the input PSD is Sx (ej
Sy (ej
)
Ry [k]
|H(ej
)| Sx (ej
) =
IDFT{Sy (ej
)} =
2
1 + a2 2a cos
2
2
1

1 aej
2
a|k|
1 a2
S y (e j )
83/291
Sampling a Continuous-Time Random Process with an ADC
H( f )
t = nTs
w (t )
w (t )
F2s
Fs
2
Sw ( f )
Fs =
N0 2
1
Ts
S w (e j )
(N0 2)Fs
Sw ( f )
w[n ]
N0 2
F2s
Fs
2
power of noise sequence

1
2
(e j ) d =
N0
Fs
2
84/291
Sending a Random Sequence to a DAC
1
0
Ts
2 Ts
h (t)
w (t )
w[n ]
S w (e j )
Fs =
N0 2
power of noise sequence

1
2
(e j ) d =
Ts
1
Ts
Sw ( f )
( N0 2)Fs
N0
2
2 Ts
w (t )
Ts
f
H ( f ) = Ts sinc
Fs
Fs
Sw ( f )
( N0 2)Ts
0
total noise power
( f )d f =
N 0 Fs
2
Fs
H( f ) d f =
2
N0
2
85/291
Relationship Between the PSD of w[n] and the PSD of w(t)

The random process w(t) is constructed from the random sequence {w[n]} as
w(t) =
n=
w[n](t nTs ).
To find the PSD of w(t), truncate the random process to a time interval of T = N Ts
PN
to T = N Ts to obtain wT (t) =
n=N w[n](t nTs ).
Take the Fourier transform of the truncated process:
WT (f ) =
N
X
n=N
w[n]F {(t nTs )} =
N
X
w[n]e
j2f nTs
n=N
Apply the basic definition of PSD:

E |WT (f )|2
Sw (f ) =
lim
T
2T

2

N
X
1
1
j2f nTs

w[n]e
E
=
lim
= T

N (2N + 1)Ts
s

n=N
Rw [k]e
j2kf Ts
k=
where Rw [k]) = E {w[n]w[n k]} is the autocorrelation function of random sequence

{w[n]}. Note also that Rw [k] = Rw [k].
P
jk
Let
= 2f Ts and recognize that Sw (ej
)=
is exactly the PSD
k= Rw [k]e
of random sequence w[n]. Then the PSD of w(t) is

1
j

Sw (f ) =
Sw (e )
Ts
=2f
Ts
86/291
BINARY DATA TRANSMISSION

OVER AN AWGN CHANNEL
87/291
Block Diagram of Binary Communication Systems

m(t )

(t )
m
!"#$%
1 $-**#$
{bk }
)"('*
+,'$-.#** /
b k = 0 s1 (t )
b k = 1 s2 (t )
!&'$$ (
0 ."('*
{ }
b k
+1 #2 /
r( t )
w (t )
Bits in two different time slots are statistically independent.

a priori probabilities: P [bk = 0] = P1 , P [bk = 1] = P2 .
Signals s1 (t) and s2 (t) have a duration of Tb seconds and finite energies:
R
R
E1 = 0Tb s21 (t)dt, E2 = 0Tb s22 (t)dt.
Noise w(t) is stationary Gaussian, zero-mean white noise with two-sided power
spectral density of N0 /2 (watts/Hz):

N0
N0
E{w(t)} = 0, E{w(t)w(t + )} =
( ), w(t) N 0,
.
2
2
88/291
m(t )
345678 94:;<=
(t )
m
I874<EB657B;4<
{bk }
A4:5@?B46
CD6?<EF;BB86G
b k = 0 s1 (t )
b k = 1 s2 (t )
9>?<<8@
H8F4:5@?B46
{ }
b k
CI878;J86G
r( t )
w (t )
Received signal over [(k 1)Tb , kTb ]:

r(t) = si (t (k 1)Tb ) + w(t),
(k 1)Tb t kTb .
Objective is to design a receiver (or demodulator) such that the probability of

making an error is minimized.
Shall reduce the problem from the observation of a time waveform to that of
observing a set of numbers (which are random variables).
89/291
Geometric Representation of Signals s1 (t) and s2 (t) (I)
Wish to represent two arbitrary signals s1 (t) and s2 (t) as linear combinations of two
orthonormal basis functions 1 (t) and 2 (t).
1 (t) and 2 (t) are orthonormal if:
Z
Tb
1 (t)2 (t)dt = 0 (orthogonality),
Tb
21 (t)dt =
Tb
22 (t)dt = 1 (normalized to have unit energy).
The representations are
where
s1 (t)
s11 1 (t) + s12 2 (t),
s2 (t)
s21 1 (t) + s22 2 (t).
sij =
Tb
si (t)j (t)dt,
i, j {1, 2},
90/291
Geometric Representation of Signals s1 (t) and s2 (t) (II)
2 (t )
s12
s1 (t )
s 22
s1 (t) = s11 1 (t) + s12 2 (t),

s2 (t) = s21 1 (t) + s22 2 (t),
Z Tb
sij =
si (t)j (t)dt, i, j {1, 2},
0
s2 (t )
s11
R Tb
0
s 21
1 (t )
si (t)j (t)dt is the projection of si (t) onto j (t).
How to choose orthonormal functions 1 (t) and 2 (t) to represent s1 (t) and
s2 (t) exactly?
91/291
Gram-Schmidt Procedure
s (t)
Let 1 (t) p1 . Note that s11 = E1 and s12 = 0.
E1
s (t)
Project s2 (t) = p2
onto 1 (t) to obtain the correlation coefficient:
E2
=
Tb
3
4
s2 (t)
1
1 (t)dt =
E2
E1 E2
Subtract 1 (t) from s2 (t) to obtain 2 (t) =
Tb
s1 (t)s2 (t)dt.
s2 (t)
E2
1 (t).
Finally, normalize 2 (t) to obtain:
2 (t)
2 (t)
(t)
= p2
2
Tb
1 2
2 (t) dt
0

1
s1 (t)
s2 (t)

.
p
E2
E1
1 2
qR
92/291
Gram-Schmidt Procedure: Summary
2 (t )
s2 (t )
2 (t )
2 (t)
s21
s22
d21
s2 (t )
s 22
E2
1 (t)
d 21
1 (t )
E1
s 21 s1 (t )
1 = cos( ) 1
1 (t )
p
p
1 2
E2 ,
sZ
Tb
s1 (t)
,
E1

s1 (t)
s2 (t)
1

,
p
E2
E1
1 2
Z T
p
b
s2 (t)1 (t)dt = E2 ,
[s2 (t) s1 (t)]2 dt
p
E1 2 E1 E2 + E2 .
93/291
Gram-Schmidt Procedure for M Waveforms {si (t)}M

i=1
s1 (t)
1 (t)
qR
i (t)
qR
i (t)
ij
s2 (t)dt
1
i (t)
2 ,

i (t) dt
i = 2, 3, . . . , N,
i1
si (t) X
ij j (t),
Ei
j=1
Z
si (t)
j (t)dt, j = 1, 2, . . . , i 1.
Ei
If the waveforms {si (t)}M

i=1 form a linearly independent set, then N = M . Otherwise
N < M.
94/291
Example 1
s1 (t )
s 2 (t )
Tb
Tb
0
V
LMN
1 (t )
Tb
s1 (t )
s 2 (t )
Tb
PQR
1 (t )
STU
(a) Signal set. (b) Orthonormal function. (c) Signal space representation.
95/291
Example 2
s1 (t )
s 2 (t )
V
Tb
Tb
V
WXY
2 (t )
1 (t )
1 Tb
1 Tb
Tb
0
Tb
1 Tb
[\]
2 (t )
96/291
Example 3
s1 (t )
s 2 (t )
Tb
Tb
0
V
2 (t, )
=
d
b
b
c
E 3E
,
2
2
=0
Tb
2
s 2 (t )
a
_
_
`
increasing ,
=
E ,0
Tb
4
E
=
s1 (t )
E ,0
1
E
Tb
s2 (t)s1 (t)dt
1
V 2 Tb
2
1
Tb
h
i
V 2 V 2 (Tb )
1 (t )
97/291
Example 4
s1 (t )
s 2 (t )
3V
Tb
2 (t )
3
Tb
Tb
Tb
fgh
1 (t )
Tb 2
Tb
Tb
Tb 2
3
Tb
jkl
98/291
2 (t )
s 2 (t )
3E 2 , E 2
s1 (t )
E ,0
1 (t )
3
2 3
V t V dt =
,
Tb
2
0
0
#
"

s1 (t)
1
s2 (t)
2
3
s
(t)
,
s
(t)
2 (t) =
=
1
2
1
2
E
E
E
(1 43 ) 2
1
3
s21 =
E, s22 =
E.
2
2
=
1
E
d21 =
Tb
Z
s2 (t)s1 (t)dt =
Tb
0
2
E
Tb /2
r
1

2
[s2 (t) s1 (t)]2 dt
=
2 3 E.
99/291
Example 5
2 (t )
= 3 2
=0
s1 (t)
s2 (t)
2
cos(2fc t),
Tb
2
cos(2fc t + ).
Tb
k , k an integer.
where fc = 2T
b
s1 (t )
0
=
= 1
1 (t )
locus of s2 (t ) as
varies from 0 to 2 .
s 2 (t )
= 2
=0
100/291
Representation of Noise with Walsh Functions
x1(t)
5
0
5
5
x2(t)
0
5
2
1(t)
1
2(t)
0
2
0
2
2
3(t)
0
2
2
4(t)
0
2
0
0.1
0.2
0.3
0.4
0.5
t
0.6
0.7
0.8
0.9
Exact representation of noise with 4 Walsh functions is not possible.
101/291
The First 16 Walsh Functions
0.1
0.2
0.3
0.4
0.5
t
0.6
0.7
0.8
0.9
Exact representations might be possible with many more Walsh functions.
102/291
The First 16 Sine and Cosine Functions

Can also use sine and cosine functions (Fourier representation).
1.5
1.5
0
0.25
0.5
t
0.75
103/291
Representation of the Noise

To represent the random noise signal, w(t), in the time interval [(k 1)Tb , kTb ],
need to use a complete orthonormal set of known deterministic functions:
w(t) =
wi i (t),
where
i=1
wi =
Tb
w(t)i (t)dt.
The coefficients wi s are random variables and understanding their statistical

properties is imperative in developing the optimum receiver.
104/291
When w(t) is zero-mean and white, then:

o
nR T
RT
b w(t) (t)dt
1 E{wi } = E
= 0 b E{w(t)}i (t)dt = 0.
i
0
(
o
nR
N0 ,
RT
Tb
2 E{wi wj } = E
dw()i () 0 b d w( )j ( ) =
2
0
0,
{w1 , w2 , . . .} are zero-mean and uncorrelated random variables.
i=j .
i 6= j
If w(t) is not only zero-mean and white, but also Gaussian {w1 , w2 , . . .} are Gaussian
and statistically independent!!!
The above properties do not depend on how the set {i (t), i = 1, 2, . . .} is chosen.
Shall choose as the first two functions the functions 1 (t) and 2 (t) used to represent the
two signals s1 (t) and s2 (t) exactly. The remaining functions, i.e., 3 (t), 4 (t), . . . , are
simply chosen to complete the set.
105/291
Optimum Receiver
Without any loss of generality, concentrate on the first bit interval. The received
signal is
r(t)
=
=
=
=
=
si (t) + w(t), 0 t Tb

s1 (t) + w(t),
if a 0 is transmitted
.
s2 (t) + w(t),
if a 1 is transmitted
[si1 1 (t) + si2 2 (t)]
|
{z
}
si (t)
+ [w1 1 (t) + w2 2 (t) + w3 3 (t) + w4 4 (t) + ]

{z
}
|
w(t)
(si1 + w1 )1 (t) + (si2 + w2 )2 (t) + w3 3 (t) + w4 4 (t) +
r1 1 (t) + r2 2 (t) + r3 3 (t) + r4 4 (t) +
106/291
where rj =
Tb
r(t)j (t)dt, and
r1
si1 + w1
r2
si2 + w2
r3
w3
r4
=
..
.
w4
Note that rj , for j = 3, 4, 5, . . ., does not depend on which signal (s1 (t) or s2 (t))
was transmitted.
The decision can now be based on the observations r1 , r2 , r3 , r4 , . . ..
The criterion is to minimize the bit error probability.
Consider only the first n terms (n can be very very large), ~
r = {r1 , r2 , . . . , rn }
Need to partition the n-dimensional observation space into decision regions.
107/291
Observation space
Decide a "1" was transmitted

m
if r falls in this region.
Decide a "0" was transmitted

m
if r falls in this region.
108/291
P [error] = P [(0 decided and 1 transmitted) or

(1 decided and 0 transmitted)].
=
P [0D , 1T ] + P [1D , 0T ]
=
=
P [0D |1T ]P [1T ] + P [1D |0T ]P [0T ]

Z
Z
f (~
r |0T )d~
r
f (~
r |1T )d~
r + P1
P2
P2
P2
P2 +
f (~
r |1T )d~
r + P1
f (~
r |1T )d~
r+
f (~
r |0T )d~
r
[P1 f (~
r |0T ) P2 f (~
r |1T )]d~
r
[P1 f (~
r |0T ) P2 f (~
r |1T )] d~
r.
The minimum error probability decision rule is

P1 f (~
r |0T ) P2 f (~
r |1T ) 0
P1 f (~
r |0T ) P2 f (~
r |1T ) < 0
decide 0 (0D )
.
decide 1 (1D )
109/291
t = Tb
Pr[error] =
Tb
( ) dt
P2 +
r1 = si1 + w1
1 ( t )
t = Tb
( ) dt
0
t = Tb
r3 = 0 + w 3
AWGN,
PSD
g (r )
Tb
( ) dt
) P2 f ( r |1T )]dr
g (r )
g(r )
>
0
<
1 0 D
g (r ) > 0
3 ( t )
N0
2
r2 = si 2 + w 2
2 ( t )
si (t ) + w (t )
= P1 + [ P2 f ( r |1T ) P2 f ( r | 0T )]dr
Tb
r (t ) =
[P f (r | 0
t = Tb
Tb
( ) dt
0
g (r ) < 0
rn = 0 + w n
n ( t )
g (r ) = 0
2 1D
110/291
Equivalently,
1D
f (~
r |1T )
f (~
r |0T )
The expression
0D
P1 .
P2
(2)
f (~
r |1T )
is called the likelihood ratio.
f (~
r |0T )
The decision rule in (2) was derived without specifying any statistical properties
of the noise process w(t).
Simplified decision rule when the noise w(t) is zero-mean, white and Gaussian:
(r1 s11 )2 + (r2 s12 )2
1D
0D

1
(r1 s21 )2 + (r2 s22 )2 + N0 ln P
P2 .
For the special case of P1 = P2 (signals are equally likely):

(r1 s11 )2 + (r2 s12 )2
1D
0D
(r1 s21 )2 + (r2 s22 )2 .
minimum-distance receiver!
111/291
Minimum-Distance Receiver
1D
(r1 s11 )2 + (r2 s12 )2
0D
(r1 s21 )2 + (r2 s22 )2 .
r2 2 (t )
s 2 (t )
d2
( s21 , s22 )
r (t )
d1
( r1, r2 )
s1 (t )
( s11 , s12 )
1 (t )
0
Choose s2 (t )
Choose s1 ( t )
r1
112/291
Correlation Receiver Implementation

t = Tb
Tb
o
0
()dt
r1
Compute
(r1 si1 )2 + (r2 si 2 )2
r (t ) = si (t ) + w (t )
N 0 ln( Pi )
1 (t )
t = Tb
Tb
n
0
()dt
r2
Decision
for i = 1, 2
and choose
the smallest
2 (t )
t = Tb
Tb
q
0
()dt
r1
Form
the
r (t )
1 (t )
dot
product
t = Tb
Tb
p
0
()dt
r2
2 (t )
| |
rsttuv
N0
E
ln(P1 ) 1
2
2
Decision
wsv
xyz{vuw
r si
N0
E
ln( P2 ) 2
2
2
113/291
Receiver Implementation using Matched Filters
t = Tb
r1
h1 (t ) = 1 (Tb t )
r (t )
}~
t = Tb
h2 (t ) = 2 (Tb t )
Decision
r2
114/291
Example 5.6
s1 (t )
s 2 (t )
1.5
0.5
0
0.5
1 (t )
2 (t )
0.5
0.5
115/291
2 (t )
s 2 (t )
s1 (t )
0.5
0.5
s1 (t)
s2 (t)
0.5
1 (t )
1
2 (t),
2
1 (t) + 2 (t).
1 (t) +
116/291
2 (t ) r2
2 (t ) r2
s 2 (t )
s 2 (t )
1
s1 (t )
s1 (t )
0.5
0.5
Choose s1 (t )
Choose s2 (t )
1
0.5
0.5
Choose s1 (t )
Choose s2 (t )
1 (t )
r1
0.5
0.5
1 (t )
r1
2 (t ) r2
s 2 (t )
1
Choose s2 (t )
s1 (t )
0.5
Choose s1 (t )
0.5
0.5
1 (t )
r1
(a) P1 = P2 = 0.5, (b) P1 = 0.25, P2 = 0.75. (c) P1 = 0.75, P2 = 0.25.
117/291
Example 5.7
s2 (t) = 1 (t) + 2 (t),

s1 (t) = 1 (t) 2 (t).
2 (t )
1 (t )
1
1
0
1
2
118/291
2 (t ) r2
s 2 (t )
Choose s 2 (t )
N 0 P1
ln
4 P2
1 (t )
0
r1
1
Choose s1 (t )
s1 (t )
119/291
t = Tb
Tb
r (t )
()dt
r2
r2 T choose s2 (t )
r2 < T choose s1 (t )
2 (t )
T=
Tb
h2 (t )
r (t )
t = Tb
N 0 P1
ln
4 P2
r2
Tb
r2 T choose s2 (t )
r2 < T choose s1 (t )
T=
N 0 P1
ln
4 P2
120/291
Implementation with One Correlator/Matched Filter

Always possible by a judicious choice of the orthonormal basis.
2 (t )
s2 (t )
2 (t )
1 (t )
s11 = s21
s1 (t )
s22
1 (t )
s12
1 (t)
2 (t)
cos
sin
sin
cos

1 (t)
2 (t)
121/291
2 (t )
s2 (t )
1 (t )
2 (t )
s11 = s21
s1 (t )
s22
1 (t )
s12
f (
s21 + w
1 )f (
s22 + w
2 )f (w
3 ) . . .
f (
r1 , r2 , r3 , . . . , |1T )
=
f (
r1 , r2 , r3 , . . . , |0T )
f (
s11 + w
1 )f (
s12 + w
2 )f (w
3 ) . . .
1D
r2
0D
s22 + s
12 +
2
N0 /2
s22 s12
ln
P1
P2
1D
0D
P1
P2
T.
122/291
t = Tb
Tb
r (t )
()dt
r2
2 ( t )
r2 T
1D
r2 < T
0D
r2 T
1D
r2 < T
0D
P1
P2
Threshold T
t = Tb
r (t )
r2
h (t ) = 2 (Tb t )
Threshold T
2 (t) =
s2 (t) s1 (t)
s22 + s12
, T
+
1
2
(E2 2 E1 E2 + E1 ) 2
N0 /2
s22 s12
ln
123/291
Example 5.8
2 (t )
1 (t )
s2 ( t )
2 (t )
s11 = s21
= /4
E
1 (t)
2 (t)
s1 (t )
1 (t )
1
[1 (t) + 2 (t)],
2
1
[1 (t) + 2 (t)].
2
124/291
t = Tb
Tb
r (t )
r2
()dt
r2 T
1D
r2 < T
0D
2 ( t )
Threshold T
2
Tb
Tb 2
Tb
t = Tb
h(t )
r (t )
2
Tb
r2
r2 T
1D
r2 < T
0D
Tb 2
Threshold T
125/291
Receiver Performance
To detect bk , compare
r2 =
T =
s
12 +
s22
2
N0
2(
s22
s12 )
kTb
r(t)2 (t)dt to the threshold
(k1)T
b
P1
.
ln P
2

f (r2 0T )
f (r2 1T )
s12
choose 0T
P [error]
s22
r2
choose 1T
T
P [(0 transmitted and 1 decided) or (1 transmitted and 0 decided)]
P [(0T , 1D ) or (1T , 0D )].
126/291

f (r2 0T )

f (r2 1T )

!

"!#
!
"!#
$! %$
$! %$
s22
s12

r2
'
choose 0T
& choose 1T
T
P [error]
=
=
P [0T , 1D ] + P [1T , 0D ] = P [1D |0T ]P [0T ] + P [0D |1T ]P [1T ]

Z
Z T
P1
f (
r2 |0T )d
r2 +P2
f (
r2 |1T )d
r2
T
{z
}
|
{z
}
|
Area B
P1 Q
T s12
p
N0 /2
Area A
"
+ P2 1 Q
T s22
p
N0 /2
!#
127/291
Q-function
=;4)
@;7.0
@;7.0
=;//;728 >;?-7540
>;?-7540 12-34-78 A,,-B;4-78
1 2
e
2
9(+)
:+;8<9
()*+,-.-/0
12-345-,x67-*)88
0
1
Q(x)
2
Area = Q ( x )
exp
2
2
d.
10
10
Q(x)
10
10
10
10
10
3
x
128/291
Performance when P1 = P2
P [error] = Q
s22 s12
p
2 N0 /2
=Q
distance between the signals

2 noise RMS value
Probability of error decreases as either the two signals become more dissimilar
(increasing the distances between them) or the noise power becomes less.
To maximize the distance between the two signals one chooses them so that they
are placed 180 from each other s2 (t) = s1 (t), i.e., antipodal signaling.
The error probability does not depend on the signal shapes but only on the
distance between them.
129/291
Relationship Between Q(x) and erfc(x).
The complementary error function is defined as:

Z
2
exp(2 )d
erfc(x) =
x
= 1 erf(x).
erfc-function and the Q-function are related by:

1
x
Q(x) =
erfc
2
2
erfc(x) = 2Q( 2x).

Let Q1 (x) and erfc1 (x) be the inverses of Q(x) and erfc(x), respectively.
Then
Q1 (x) = 2erfc1 (2x).
130/291
Example 5.9
2 (t )
s1 (t ) 0T
1
2
1T s 2 (t )
1 (t )
(a) Determine and sketch the two signals s1 (t) and s2 (t).
131/291
(b) The two signals s1 (t) and s2 (t) are used for the transmission of equally likely bits 0 and 1,
respectively, over an additive white Gaussian noise (AWGN) channel. Clearly draw the
decision boundary and the decision regions of the optimum receiver. Write the expression for
the optimum decision rule.
1 (t) and
2 (t) such that the optimum
(c) Find and sketch the two orthonormal basis functions
receiver can be implemented using only the projection
r2 of the received signal r(t) onto the
2 (t). Draw the block diagram of such a receiver that uses a matched filter.
basis function
(d) Consider now the following argument put forth by your classmate. She reasons that since the
1 (t) is not useful at the receiver in determining which bit
component of the signals along
was transmitted, one should not even transmit this component of the signal. Thus she
modifies the transmitted signal as follows:

(M)
1 (t)
s1 (t) = s1 (t) component of s1 (t) along

(M)
1 (t)
s2 (t) = s2 (t) component of s2 (t) along
(M)
(M)
Clearly identify the locations of s1 (t) and s2 (t) in the signal space diagram. What is the
average energy of this signal set? Compare it to the average energy of the original set.
Comment.
132/291
s 2 (t )
s1 (t )
133/291
2 (t )
0
CDEFGFHI JHKILMNO
0D
1D
2 ( t )
s1M (t )
s1 (t ) 0T
1
2
=
s2M (t )
1T s 2 (t )
1 (t)
2 (t)
cos(/4)
sin(/4)
1 (t )
1
1 (t) = [1 (t) 2 (t)],
2

1 (t)
2 (t)
sin(/4)
cos(/4)
1 (t )
"
1
2
1
2
2
1
#
1 (t)
2 (t)
1
2 (t) = [1 (t) + 2 (t)].
2
134/291
1 (t )
2 (t )
2
1/ 2
1/ 2
h (t ) = 2 (1 t )
t =1
r (t )
1/ 2
r2 0 P 0D
r2 < 0 P 1D
135/291
PSD Derivation of Arbitrary Binary Modulation

Applicable to any binary modulation with arbitrary a priori probabilities, but
restricted to statistically independent bits.
sT (t )
s1 ( t 2Tb )
s1 (t )
2Tb
Tb
Tb
2Tb
3Tb
4Tb
s2 (t 3Tb )
sT (t) =
k=
gk (t),
gk (t) =
s1 (t kTb ),
s2 (t kTb ),
with probability P1
.
with probability P2
The derivation on the next slide shows that:

2

n
n
P S

X
1 1 Tb + P2 S2 Tb
P1 P2
2

f n .
SsT (f ) =
|S1 (f ) S2 (f )| +

Tb
Tb
Tb

n=
136/291
sT (t) = E{sT (t)} + sT (t) E{sT (t)} = v(t) + q(t)

{z
}
| {z } |
DC
AC
v(t) = E{sT (t)} =
k=
Sv (f ) =
n=
|Dn |2
[P1 s1 (t kTb ) + P2 s2 (t kTb )]
n
Tb
, Dn =
1
Tb
P1 S1
n
Tb
+ P2 S2
n
Tb

2

n
n
P S

X
1 1 Tb + P2 S2 Tb

f n .
Sv (f ) =

Tb
Tb

n=
To calculate Sq (f ), apply the basic definition of PSD:

Sq (f ) = lim
P1 P2
E{|GT (f )|2 }
2
= =
|S1 (f ) S2 (f )| .
T
Tb
2

n
P S

n

X
1 1 Tb + P2 S2 Tb
P1 P2
2

f n .
SsT (f ) =
|S1 (f ) S2 (f )| +

Tb
T
T
b
b

n=
For the special, but important case of antipodal signalling, s2 (t) = s1 (t) = p(t), and equally
likely bits, P1 = P2 = 0.5, the PSD of the transmitted signal is solely determined by the
Fourier transform of p(t):
|H(f )|2
|P (f )|2
=
SsT (f ) =
Tb
Tb
137/291
Building A Binary (Antipodal) Comm. System in Labs #4 and #5

low-rate sequence, index k
high-rate sequences, index n
...
...
...
t
s[n ] = i[n ] * h[n ] = a[k ] p(nT - kTsym )

data
bits
S/P
bits-tolevels
mapping
N
N=
Tsym
T
pulse
shaping
filter
DAC
correction
filter
hT [n ] = p(nT )
i[n ] = i (nT ) = a[k ]d (nT - kTsym )

k
DAC
cos( w c n )
NCO
fc
wc = w c Fs
s(t ) = a ( k ) p(t - kTs ) cos(wct )

k
Figure 1: Block diagram of the transmitter.
138/291
s (t )
x[ n d ]
x[ n]
hR [n ] = p( nT )
xc [ n ]
k =n
Tsym
T
cos( c n )
fc
sin( c n )
c d =
x s [n ]
Figure 2: Block diagram of the receiver.
139/291
Baseband Message Signals with Different Pulse Shaping Filters

Information bits or amplitude levels
1
0
1
2
10
12
14
16
18
20
18
20
18
20
18
20
n
Output of the transmit pulse shaping filter Rectangular
2
0
2
10
12
14
16
t/Tb
Output of the transmit pulse shaping filter Halfsine
2
0
2
10
12
14
16
t/Tb
Output of the transmit pulse shaping filter SRRC ( =0.5)
2
0
2
10
12
14
16
t/Tb
140/291
Antipodal Baseband Signalling with Rectangular Pulse Shaping
bk
h(t ) = p(Tb t )
141/291
Antipodal Baseband Signalling with Half-Sine Pulse Shaping
bk
h(t ) = p (Tb t )
142/291
Outputs of the Matched/Mismatched Filters (No-Noise Scenario)

Clean received signals for rect and half-sine (HS) shaping filters
Output of a matched filter: rect/rect matching
Outputs of HS/HS matched filter (red) and HS/rect mismatched filter (pink)
143/291
Introduction to Baseband Data Transmission
Bits are mapped into two voltage levels for direct transmission without any
frequency translation.
Various baseband signaling techniques (line codes) were developed to satisfy
typical criteria:
1
2
3
4
5
Signal interference and noise immunity

Signal spectrum
Signal synchronization capability
Error detection capability
Cost and complexity of transmitter and receiver implementations
Four baseband signaling schemes to be considered: nonreturn-to-zero-level

(NRZ-L), return-to-zero (RZ), bi-phase-level or Manchester, and delay
modulation or Miller.
144/291
Baseband Signaling Schemes
Binary Data
Clock
Tb
V
(a) NRZ Code
Time
(b) NRZ - L
145/291
V
(c) RZ Code
(d) RZ - L
V
(e) Bi - Phase
146/291
(f) Bi - Phase - L
V
(g) Miller Code
V
(h) Miller - L
147/291
Miller Code
Has at least one transition every two bit interval and there is never more than two
transitions every two bit interval.
Bit 1 is encoded by a transition in the middle of the bit interval. Depending on
the previous bit this transition may be either upward or downward.
Bit 0 is encoded by a transition at the beginning of the bit interval if the
previous bit is 0. If the previous bit is 1, then there is no transition.
V
(f) Bi - Phase - L
V
(g) Miller Code
V
(h) Miller - L
148/291
NRZ-L Code
s1 (t )
s 2 (t )
UVWXYU
UYZWU
1 (t )
V
Tb
0
Tb
Tb
\]^
QRS
Choose 0T
b Choose 1T
s 2 (t )
s1 (t )
ENRZ-L
Tb
1 (t )
ENRZ-L
_`a
P [error]NRZ-L = Q
p

2ENRZ-L /N0 .
149/291
RZ-L Code
1 (t )
2 (t )
s 2 (t )
s1 (t )
gkjlhg
ghijg
1 Tb
Tb
Tb
Tb 2 Tb
cde
Tb
Tb 2 Tb
mno
1 Tb
2 (t )
ERZ-L
s 2 (t )
Choose 0T
Choose 1T
s1 (t )
P [error]RZ-L = Q

q
ERZ-L /N0 .
1 (t )
ERZ-L
qrs
150/291
Bi-Phase-Level (Bi-L) Code

1 (t )
s 2 (t )
s1 (t )
x|}zx
xyz{|x
V
Tb
Tb
Tb
0
V
tuv
1 Tb
Choose 0T
Choose 1T
s1 (t )
EBi -L
Tb
s 2 (t )
1 (t )
E Bi -L
P [error]Bi-L = Q
q
2EBi-L /N0
151/291
Miller-Level (M-L)
s1 ( t )
s 2 (t )
Tb
Tb
s3 ( t )
s4 ( t )
Tb
0
Tb
0
1 Tb
1 (t )
Tb
Tb
2 (t )
Tb
Tb
152/291
2 (t )
s 2 (t )
s 3 (t )
s1 (t )
1 (t )
s 4 (t )
153/291
2 (t )
r2
r2
s 2 (t )
0.5E M-L
s 3 (t )
0
450
s1 (t )
1 (t )
EM - L
s4 ( t )
r1
r1
h
p
i2
P [error]M-L = 1 1 Q
EM-L /N0
154/291
Performance Comparison
10
10
ML
RZL
P[error]
10
NRZL and BiL

10
10
10
6
8
E /N (dB)
b
10
12
14
ENRZ-L = ERZ-L = EBi-L = EM-L = V
Tb Eb (joules/bit).
v
u
u 2Eb
,
P [error]NRZ-L = P [error]Bi-L = Q t
N0
u
u
u Eb
uE
, P [error]
t b .
P [error]RZ-L = Q t
M-L 2Q
N0
N0
155/291
Spectrum
SNRZ-L (f )
1
sin2 (f Tb )
=
(1 2P )2 (f ) + 4P (1 P )
.
E
Tb
(f Tb )2
SBi (f )
E

X
n
1
2 2
f
(1 2P )2
Tb
n
Tb
n=
sin4 (f Tb /2)
.
4P (1 P )
(f Tb /2)2
n6=0
SM-L (f )
1
=
(23 2 cos 22 cos 2 12 cos 3
E
2 2 (17 + 8 cos 8)
+5 cos 4 + 12 cos 5 + 2 cos 6 8 cos 7 + 2 cos 8),
where = f Tb and P2 = P1 = 0.5.
156/291
2.5
MillerL
Normalized PSD (S (f)/E)
1.5
NRZL
1
BiL
0.5
0
0
0.5
1
1.5
Normalized frequency (fT )
157/291
Introduction to Basic Digital Passband Modulation
Baseband transmission is conducted at low frequencies.

Passband transmission happens in a frequency band toward the high end of the
spectrum.
Satellite communication is in the 68 GHz band, while mobile phones systems are
in the 800 MHz2.0 GHz band.
Bits are encoded as a variation of the amplitude, phase or frequency, or some
combination of these parameters of a sinusoidal carrier.
The carrier frequency is much higher than the highest frequency of the
modulating signals (or messages).
Shall consider binary amplitude-shift keying (BASK), binary phase-shift keying
(BPSK) and binary frequency-shift keying (BFSK): Error performance, optimum
receivers, spectra.
Extensions to quadrature phase-shift keying (QPSK), offset QPSK (OQPSK) and
minimum shift keying (MSK).
158/291
Examples of Binary Passband Modulated Signals

(a) Binary data
(b) Modulating signal

0
m(t )
t
0
Tb
2Tb
3Tb
4Tb
5Tb
6Tb
7Tb
8Tb
9Tb
(c) BASK signal
V
V
(d) BPSK signal
V
V
(e) BFSK signal
159/291
Binary Amplitude-Shift Keying (BASK)

s1 (t) = 0,
s2 (t) = V cos(2fc t),
0T
, 0 < t Tb , fc = n/Tb
1T
s1 (t )
s 2 (t )
1 (t )
EBASK
t = Tb
Tb
r1 Th 1D
r1
()dt
r (t )
Comparator
r1 < Th 0 D
1 (t )
Th =
(b)
Choose 0T
E BASK
P
N0
ln 1 +
2
2 E BASK P2
Choose 1T
s1 (t )
s 2 (t )
Tb
r1 = r (t )1 (t )dt
0
EBASK 2
P [error]BASK = Q
EBASK
2N0
160/291
PSD of BASK
SBASK (f )

V2
(f fc ) + (f + fc ) +
16

sin2 [Tb (f + fc )]
sin2 [Tb (f fc )]
+
.
2
2
2
2
Tb (f + fc )
Tb (f fc )
SBASK ( f )
V2
16
f
0
fc
1
Tb
fc
fc +
1
Tb
Approximately 95% of the total transmitted power lies in a band of 3/Tb (Hz),
centered at fc .
161/291
Binary Phase-Shift Keying (BPSK)
s1 (t) = V cos(2fc t),

s2 (t) = +V cos(2fc t),
s1 (t )
s 2 (t )
EBPSK
V2
4
1 (t ) =
2
cos(2 f ct )
Tb
EBPSK
P [error]BPSK = Q
SBPSK (f ) =
if 0T
, 0 < t Tb ,
if 1T
2EBPSK
N0

sin2 [(f fc )Tb ]
sin2 [(f + fc )Tb ]
+
.
2 (f fc )2 )Tb
2 (f + fc )2 Tb
Similar to that of BASK, but no impulse functions at fc .
162/291
Binary Frequency-Shift Keying (BFSK)
s (t )
s1 (t) = V cos(2f1 t + 1 ),
s2 (t) = V cos(2f2 t + 2 ),
if 0T
,
if 1T
0 < t Tb .
(i) Minimum frequency separation for coherent orthogonality (1 = 2 ):

[coherent]
(f )min
1
.
2Tb
(ii) Minimum frequency separation for noncoherent orthogonality (1 6= 2 ):

[noncoherent]
(f )min
1
.
Tb
163/291
s1 (t)
1 (t) =
,
EBFSK
s2 (t)
2 (t) =
.
EBFSK
r2 2 (t )
[2 (t ) 1 (t )]
Decision boundary
s 2 (t )
( 0,
EBFSK
when P2 = P1
Choose 0T
Choose 1T
P [error]BFSK = Q
EBFSK , 0
s1 (t )
EBFSK
N0
r1
1 (t )
164/291
PSD of BFSK
SBFSK (f )
V2
sin2 [Tb (f + f2 )]
sin2 [Tb (f f2 )]
(f f2 ) + (f + f2 ) +
+
16
2 Tb (f + f2 )2
2 Tb (f f2 )2
sin2 [Tb (f f1 )]
sin2 [Tb (f + f1 )]
V2
.
+
(f f1 ) + (f + f1 ) +
16
2 Tb (f + f1 )2
2 Tb (f f1 )2
Bandwidth W = ( f 2 f1 ) + 3 / Tb
V2
16
f1
1.5
Tb
f1
V2
16
f1 +
1
Tb
f2
1
Tb
f2
f2 +
1.5
Tb
165/291
Performance Comparison of BASK, BPSK and BFSK

1
10
10
BASK and BFSK

3
P[error]
10
BPSK
4
10
10
10
P [error]BPSK = Q
2
s
2Eb
N0
4
!
6
8
Eb/N0 (dB)
10
, P [error]BASK = P [error]BFSK = Q
12
s
14
Eb
N0
166/291
Quadrature Phase Shift Keying (QPSK)
Basic idea behind QPSK: cos(2fc t) and sin(2fc t) are orthogonal over [0, Tb ] when
fc = k/Tb , k integer Can transmit two different bits over the same frequency band at
the same time.
The symbol signaling rate (i.e., the baud rate) is rs = 1/Ts = 1/(2Tb ) = rb /2
(symbols/sec), i.e., halved.
Bit Pattern
Message
00
01
11
10
m1
m2
m3
m4
Signal Transmitted
s1 (t)
s2 (t)
s3 (t)
s4 (t)
m1 = 00
= V cos(2fc t),
= V sin(2fc t),
= V cos(2fc t),
= V sin(2fc t),
m2 = 01
m3 = 11
0
0
0
0
t
t
t
t
Ts
Ts
Ts
Ts
=
=
=
=
2Tb
2Tb
2Tb
2Tb
m4 = 10
2Tb
4Tb
6Tb
8Tb
167/291
Signal Space Representation of QPSK

Z
Ts
si (t)dt =
s1 (t)
1 (t) =
,
Es
V2
2
Ts = V Tb = Es ,
2
s2 (t)
2 (t) =
.
Es
2 (t )
s 2 (t )
Es
s1 (t )
s3 (t )
0
1 (t )
s4 ( t )
168/291
Optimum Receiver for QPSK

m-dimensional observation space
r = (r1 , r2 , , rm )
1
Choose s1 (t ) or m1
4
P [correct]
P1 f (~
r |s1 (t))d~
r+
P3 f (~
r |s3 (t))d~
r+
P2 f (~
r|s2 (t))d~
r
P4 f (~
r|s4 (t))d~
r.
Choose si (t) if Pi f (~
r |si (t)) > Pj f (~
r|sj (t)), j = 1, 2, 3, 4; j 6= i.
169/291
Simplified Decision Rule and Receiver Implementation
Choose si (t) if
N0 ln P + r s + r s > N0 ln P + r s + r s
1 i1
2 i2
1 j1
2 j2
i
j
2
2
j = 1, 2, 3, 4; j 6= i.
t = Ts = 2Tb
Ts
r1
( ) dt
0
r (t )
1 (t )
t = Ts = 2Tb
Ts
( ) dt
r2
Compute
N
r1 s j1 + r2 s j 2 + 0 ln( Pj )
2
for j = 1, 2,3, 4
and choose
Decision
the largest
2 (t )
170/291
Minimum-Distance Receiver
Choose si (t) if (r1 si1 )2 + (r2 si2 )2 is the smallest

2 (t ), r2
s 2 (t )
/4
s 3 (t )
s1 (t )
1 (t ), r1
s4 ( t )
171/291
Symbol (Message) Error Probability of QPSK

r2
s 2 (t )
r1
r2

/4
s3 ( t )
s1 (t )
r1

/2
Es
/2
s4 ( t )

"
P [error] = P [error|si (t)] = 1 P [correct|si (t)] = 1 1 Q
Es
N0
!#2
172/291
Joint pdf f(
r1 , r2 )
Illustration of Symbol Error Probability Analysis for QPSK

0.15
0.1
0.05
0
4
Observation r2
Observation r1
Observation r2
2
1
0
1
2
2
1
0
1
Observation r1
173/291
Bit Error Probability of QPSK

r2
s 2 (t )
r1
r2
Es
/4
s3 ( t )
s1 (t )
r1
&&%$
#"
!

P [m2 |m1 ]
P [m3 |m1 ]
P [m4 |m1 ]
/2

/2
s4 ( t )

P [bit error]
=
=
v
u
u
u Es
u Es
1 Q t
Q t
N0
N0
v
u
E
u
s
2
,
Q t
N0
v
v
u
u
u Es
u Es
1 Q t
Q t
N0
N0
0.5P [m2 |m1 ] + 0.5P [m4 |m1 ] + 1.0P [m3 |m1 ]

s
!
Es
Q
.
N0
Gray mapping: Nearest neighbors are mapped to the bit pairs that differ in only one bit.
174/291
An Alternative Representation of QPSK
V cos(2f c t )
a I (t )
'()*+,-.+(/(0
a(t )
s (t )
aQ (t )
Bit sequence
a0
a1
a2
V sin(2f ct )
a3
a4
a5
a6
a7
a8
7Tb
8Tb
9Tb
1
a (t )
Tb
2Tb
3Tb 4Tb
5Tb 6Tb
1
a0
a2
a6
a4
a8
1
a I (t )
t
0
Tb
2Tb
3Tb 4Tb
5Tb 6Tb
7Tb
a3
a5
a7
3Tb 4Tb
5Tb
8Tb
9Tb
1
a1
1
aQ (t )
t
0
Tb
2Tb
6Tb
7Tb 8Tb
9Tb
Figure 3: Examples of aI (t) and aQ (t) in QPSK modulation.

175/291
a (t)Vcos(2f t)
I
V
0
V
0
2Tb
4Tb
6Tb
a (t)Vsin(2f t)
Q
8Tb
10Tb
8Tb
10Tb
V
0
V
0
2Tb
4Tb
6Tb
(t)=a (t)Vcos(2f t)+a (t)Vsin(2f t)
QPSK
V
0
V
s(t)
=
=
aI (t)V
cos(2f
t)
0
2Tb
4Tb sin(2fc
6Tb
8Tb
10Tb
c t) + aQ (t)V
t
!!
r
1 aQ (t)
2
2
a (t) + a (t)V cos 2fc t tan
2V cos[2fc t (t)].
=
I
Q
aI (t)
(t) =
/4,
/4,
3/4,
3/4,
if
if
if
if
aI
aI
aI
aI
=
=
=
=
+1, aQ
+1, aQ
1, aQ
1, aQ
=
=
=
=
+1 (bits are 11)

1 (bits are 10)
+1 (bits are 01)
1 (bits are 00)
176/291
Signal Space Representation of QPSK
1 (t) =
2 (t) =
V cos(2fc t)
V 2 Tb
V sin(2fc t)
0 < t < Ts = 2Tb ,
V 2 Tb
2 (t ) =
2
sin(2 fc t )
Ts
aI = 1, aQ = 1 1 (t ) = / 4
aI = 1, aQ = 1 3 (t ) = 3 / 4
V 2Ts
/4
0
aI = 1, aQ = 1 4 (t ) = 3 / 4
1 (t ) =
2
cos(2 fc t )
Ts
aI = 1, aQ = 1 2 (t ) = / 4
177/291
Receiver Implementation of QPSK

Inphase bit stream
Quadrature bit stream

s2(Q ) (t )
s1(Q ) (t )
V 2Ts
s2( I ) (t )
2 (t )
V 2Ts
2
s1( I ) (t )
V 2Ts
1 (t )
V 2Ts
2
t = Ts
Ts
r1
( ) dt
Threshold = 0
a I (t )
r (t )
Multiplexer
1 (t )
t = Ts
r2
Ts
( ) dt
a (t )
aQ (t )
Threshold = 0
2 (t )
s
P [bit error] = Q
V 2 Ts
= = Q
N0
2Eb
N0
178/291
M -ARY COMMUNICATION
SYSTEMS FOR AWGN CHANNELS
179/291
Introduction to M -ary Signaling Techniques
There are benefits to be gained when M -ary (M = 4) signaling methods are used
rather than straightforward binary signaling.
In general, M -ary communication is used when one needs to design a
communication system that is bandwidth efficient.
Unlike QPSK and its variations, the gain in bandwidth is accomplished at the
expense of error performance.
To use M -ary modulation, the bit stream is blocked into groups of bits the
number of bit patterns is M = 2 .
The symbol transmission rate is rs = 1/Ts = 1/(Tb ) = rb / symbols/sec
there is a bandwidth saving of 1/ compared to binary modulation.
Shall consider M -ary ASK, PSK, QAM (quadrature amplitude modulation) and
FSK.
180/291
Optimum Receiver for M -ary Signaling
56789:
mi
;6<7=>?68
@A8>BCDE??:8F
r (t )
si (t )
G:D6<7=>?68
@H:9:EI:8F
m i
w (t )
w(t) is zero-mean white Gaussian noise with power spectral density of

(watts/Hz).
N0
2
Receiver needs to make the decision on the transmitted signal based on the
received signal r(t) = si (t) + w(t).
The determination of the optimum receiver (with minimum error) proceeds in a
manner analogous to that for the binary case.
181/291
Represent M signals by an orthonormal basis set, {n (t)}N

n=1 , N M :
si (t)
sik
si1 1 (t) + si2 2 (t) + + siN N (t),

Z Ts
si (t)k (t)dt.
0
Expand the received signal r(t) into the series

r(t)
si (t) + w(t)
r1 1 (t) + r2 2 (t) + + rN N (t) + rN+1 N+1 (t) +
For k > N , the coefficients rk can be discarded.

Need to partition the N -dimensional space formed by ~r = (r1 , r2 , . . . , rN ) into
M regions so that the message error probability is minimized.
182/291
N dimensiona l observatio n space
r = (r1 , r2 , K , rM )
1
Choose s M (t ) or mM
2
The optimum receiver is also the minimum-distance receiver:
PN
Choose mi if
P
2
sik )2 < N
k=1 (rk sjk ) ;
j = 1, 2, . . . , M ; j 6= i.
k=1 (rk
183/291
M -ary Coherent Amplitude-Shift Keying (M -ASK)
si (t)
2
cos(2fc t), 0 t Ts
Ts
s
2
[(i 1)]1 (t), 1 (t) =
cos(2fc t), 0 t Ts ,
Ts
Vi
i = 1, 2, . . . , M.
s1 (t )
s2 ( t )
s3 (t )
sk (t )
M sM 1 (t )
(k 1)
(M 2)
sM (t )
1 (t )
(M 1)
t = kTs
si (t )
r (t )
kTs
O ( )dt
( k 1) Ts
r1
PQRSTSUV
PQWSRQ
m i
w (t )
1 (t )
N
WGN, strength 0 watts/Hz
2
184/291
Minimum-Distance Decision Rule for M -ASK
sk (t),
Choose
s (t),
1
sM (t),
if
if
if

k 32 < r1 < k 12 , k = 2, 3, . . . , M 1
.
r1 <
2

r1 > M 23
f (r1 sk (t ) )
r1
0
(k 1)
Choose s1 (t )
X Choose sM (t )
Choose sk (t )
185/291
Error Performance of M -ASK

f (r1 sk (t ) )
r1
P [error] =
M
X
P [si (t)]P [error|si (t)].
i=1

p
P [error|si (t)] = 2Q / 2N0 , i = 2, 3, . . . , M 1.

p
P [error|si (t)] = Q / 2N0 , i = 1, M.
P [error] =

2(M 1) p
Q / 2N0 .
M
186/291
Modified M -ASK Constellation

The maximum and average transmitted energies can be reduced, without any sacrifice in error
probability, by changing the signal set to one which includes the negative version of each signal.
s
2
si (t) = (2i 1 M)
cos(2fc t), 0 t Ts , i = 1, 2, . . . , M.
2
Ts
{z
}
|
Vi
\]^
\_^
[
Es
Eb =
PM
i=1
Ei
3
2
1 (t )
1 (t )
M
2 X
(M 2 1)2
.
(2i 1 M)2 =
4M i=1
12
Es
(M 2 1)2
=
=
log2 M
12 log2 M
(12 log2 M)Eb

M2 1
187/291
Probability of Symbol Error for M -ASK
P [error] =
2(M 1)
M
P [bit error] =
v
u
u
Q t
v
u
6Es
2(M 1)
u 6 log2 M Eb
=
.
Q t
2
2
(M 1)N0
M
M 1 N0
P [symbol error] =
2(M 1)
M log2 M
10
10
v
u
u 6 log2 M Eb
(with Gray mapping)
Q t
M 2 1 N0
P[symbol error]
M=16
(W=1/4T )
10
M=8
(W=1/3T )
10
M=4
(W=1/2T )
10
M=2
(W=1/T )
10
10
10
15
20
25
Eb/N0 (dB)
W is obtained by using the W Ts = 1 rule-of-thumb. Here 1/Tb is the bit rate (bits/s).
188/291
Example of 2-ASK (BPSK) and 4-ASK Signals

Baseband information signal
1
0
1
0
Tb
2Tb
3Tb
4Tb
5Tb
6Tb
7Tb
8Tb
9Tb
10Tb
7Tb
8Tb
9Tb
10Tb
BPSK Signalling
2
0
2
0
Tb
2Tb
3Tb
4Tb
5Tb
6Tb
4ASK Signalling
2
0
2
0
2Tb
4Tb
6Tb
8Tb
10Tb
189/291
M -ary Phase-Shift Keying (M -PSK)

(i 1)2
si (t) = V cos 2fc t
,
M
0 t Ts ,
i = 1, 2, . . . , M ; fc = k/Ts , k integer; Es = V 2 Ts /2 joules

(i 1)2
(i 1)2
cos(2fc t) + V sin
sin(2fc t).
si (t) = V cos
M
M
V cos(2fc t)
V sin(2fc t)
, 2 (t) =
.
Es
Es

p
p
(i 1)2
(i 1)2
, si2 = Es sin
.
si1 = Es cos
M
M
The signals lie on a circle of radius Es , and are spaced every 2/M radians around
the circle.
1 (t) =
190/291
Signal Space Plot of 8-PSK

(i 1)2
,
M
0 t Ts ,
2 (t )
s3 (t ) 011
010 s4 (t )
s2 (t ) 001
Es
110 s5 (t )
0
s1 ( t ) 000
1 (t )
s8 (t ) 100
111 s6 (t )
s7 (t ) 101
191/291
Signal Space Plot of General M -PSK

(i 1)2
,
M
0 t Ts ,
2 (t )
s 2 (t )
Es
2 M
0
s1 (t )
2 M
1 (t )
s M (t )
192/291
Optimum Receiver for M -PSK

t = Ts
r1
Ts
a ()dt
Compute
(r1 si1 )2 + (r2 si 2 )2
r (t )
1 (t )
for i = 1, 2,`, M
t = Ts
and choose
r2
Ts
b ()dt
m i
the smallest
2 (t )
r2
P [error]
Region 2
Choose s2 (t )
P [error|s1 (t)]
ZZ
f (r1 , r2 |s1 (t))dr1 dr2 .
r1 ,r2 Region 1
s2 ( t )
Es
M
0
s1 (t )
r1
Region 1
Choose s1 (t )
193/291
Lower Bound of P [error] of M -PSK
r2
Region 2
Choose s2 (t )
r2
s2 ( t )
s2 ( t )
1
Es
Es sin ( M )
s1 (t )
r1
Region 1
Choose s1 (t )
P [error|s1 (t)]
>
P [error|s1 (t)]
>
s1 (t )
Es ,0
r1
P [r1 , r2 fall in 1 |s1 (t)], or

n p
o
Q sin
2Es /N0 .
M
194/291
Upper Bound of P [error] of M -PSK

r2
s2 ( t )
Es sin ( M )
r2
Region 2
Choose s2 (t )
s1 (t )
M
0
Es ,0
Es ,0
r1
s2 ( t )
Es
M
0
r2
s1 (t )
r1
Region 1
Choose s1 (t )
P [error]
<
P [error]
<
r1
s1 (t )
Es sin ( M )
s M (t )
P [r1 , r2 fall in 1 |s1 (t)] + P [r1 , r2 fall in 2 |s1 (t)], or

p

2Q sin
2Es /N0 ,
M
195/291
Symbol Error Probability of M -PSK

1
10
M=32
2
10
M=16
P[symbol error]
10
M=8
Exact
10
M=4
10
M=2
6
10
10
Lower bound
Upper bound
10
15
20
25
Eb/N0 (dB)
With a Gray mapping, the bit q

error probability is approximated
as:
2Eb
sin2 M
.
P [bit error]M -PSK log1 M Q
N
2
196/291
M -ary Quadrature Amplitude Modulation (M -QAM)
M-QAM constellations are two-dimensional and they involve inphase (I) and quadrature (Q)
carriers:
s
2
cos(2fc t), 0 t Ts ,
I (t) =
Ts
s
2
Q (t) =
sin(2fc t), 0 t Ts ,
Ts
The ith transmitted M-QAM signal is:
s
s
2
2
cos(2fc t) + VQ,i
sin(2fc t),
si (t) = VI,i
Ts
Ts
s
p
2
=
Ei
cos(2fc t i )
Ts
0 t Ts
i = 1, 2, . . . , M
VI,i and VQ,i are the information-bearing discrete amplitudes of the two quadrature carriers,
2
2
Ei = VI,i
+ VQ,i
and i = tan1 (VQ,i /VI,i ).
197/291
(1,7)
(4,4)
Rectangle
Triangle
M =8
M =4
(1,3)
Rectangle
198/291
R2
(4,12)
(8,8)
Triangle
R2
R1
(1,5,10)
M = 16
R1
R1
R2
Hexagon
Rectangle
199/291
A Simple Comparison of M -QAM Constellations
With the same minimum distance of all the constellations, a more efficient signal
constellation is the one that has smaller average transmitted energy.
pqrstuvw
hijklmnoi
xyzy{
cdefg
M =8
Es for the rectangular, triangular, (1,7) and (4,4) constellations are found to be
1.502 , 1.1252 , 1.1622 and 1.1832 , respectively.
200/291
Rectangular M -QAM
Q (t )
M = 64
M = 16
M = 32
M =8
I (t )
M =4
The signal components take value from the set of discrete values {(2i 1 M )/2},
.
i = 1, 2, . . . , M
2
201/291
Modulation of Rectangular M -QAM

Each group of = log2 M bits can be divided into I inphase bits and Q quadrature bits,
where I + Q = .
Inphase bits and quadrature bits modulate the inphase and quadrature carriers independently.
Inphase bits
Select VI ,i
Infor. bits
|}~
}}
2
cos(2f c t )
Ts
Inphase ASK
si (t )
Quadrature bits
Select VQ,i
2
sin( 2f c t )
Ts
Quadrature ASK
}
202/291
Demodulation of Rectangular M -QAM

Due to the orthogonality of the inphase and quadrature signals, inphase and quadrature bits can
be independently detected at the receiver.
t = Ts
Inphase
ASK
decision
Ts
( ) dt
0
r (t ) = si (t ) + w (t )
I ( t ) =
2
cos(2 f c t )
Ts
Ts
( ) dt
0
Q (t ) =
2
sin(2 f ct )
Ts
Decision
Multiplexer
t = Ts
Quadrature
ASK
decision
(b) Receiver
The most practical rectangular QAM constellation is one which I = Q = /2, i.e., M is a
perfect square and the rectangle is a square.
203/291
Symbol Error Probability of M -QAM
For square constellations:

2
P [error] = 1 P [correct] = 1 1 PM [error] ,
P
1
[error] = 2 1
M
M
3Es
(M 1)N0
where Es /N0 is the average SNR per symbol.

For general rectangular constellations:
P [error]
"
1 1 2Q
4Q
3Es
(M 1)N0
!
!#2
3Eb
(M 1)N0
where Eb /N0 is the average SNR per bit.
204/291
10
M=256
10
P[symbol error]
M=64
10
M=16
10
10
M=4
5
M=2
10
10
Exact performance
Upper bound
5
10
15
20
25
Eb/N0 (dB)
205/291
Performance Comparison of M -ASK, M -PSK, M -QAM

1
10
P[symbol error]
10
10
4QAM or QPSK
4
10
M=4, 8, 16, 32
5
10
10
MASK
MPSK
MQAM
5
10
E /N (dB)
b
15
20
206/291
M -ary Coherent Frequency-Shift Keying (M -FSK)

si (t) =
0 t Ts
, i = 1, 2, . . . , M,
elsewhere
V cos(2fi t),
0,
where fi are chosen to have orthogonal signals over [0, Ts ].

1
(k i)
, (coherently orthogonal)
2Ts

fi =
(k i) 1 ,
(noncoherently orthogonal)
Ts
, i = 0, 1, 2, . . .
2 (t )
s 2 (t )
Es
0
s3 ( t )
Es
s1 (t )
1 (t )
Es
3 (t )
207/291
Minimum-Distance Receiver of M -FSK
M
X
Choose mi if
M
X
(rk sjk )2
(rk sik )2 <
k=1
k=1
ri > rj ,
j = 1, 2, . . . , M ; j 6= i,
Choose mi if
j = 1, 2, . . . , M ; j 6= i.
t = Ts
( )dt
r1
Ts
r (t )
1 (t ) =
s1 (t )
Es
t = Ts
Ts
()dt
Decision
rM
M ( t ) =
s M (t )
Es
208/291
Symbol Error Probability of M -FSK
P [error] = P [error|s1 (t)] = 1 P [correct|s1 (t)].
P [correct|s1 (t)] = P [(r2 < r1 ) and and (rM < r1 )|s1 (t) sent].
Z
P [(r2 < r1 ) and and (rM < r1 )|{r1 = r1 , s1 (t)}]f (r1 |s1 (t))dr.
r1 =
P [(r2 < r1 ) and and(rM < r1 )|{r1 = r1 , s1 (t)}] =

P [rj < r1 |{r1 = r1 , s1 (t)}] =
P [correct]
"Z
r1
M
Y
P [(rj < r1 )|{r1 = r1 , s1 (t)}].
j=2
r1
1
N0
exp
exp
2
N0
N0
(
)
1
(r1 Es )2
exp
dr1 .
N0
N0
r1 =
2
N0
d.
#M 1
209/291
10
10
M=64
P[symbol error]
10
M=4
M=32
M=2
10
M=16
5
10
M=8
6
10
10
6
8
Eb/N0 (dB)
10
12
14
16
"
#M 1
u
1 Z
1
1 Z y
u 2 log2 M Eb
x2 /2
dy.
P [error] = 1
e
dx
exp y t
2
2
2
N0
210/291
Bit Error Probability of M -FSK
Due to the symmetry of M -FSK constellation, all mappings from sequences of

bits to signal points yield the same bit error probability.
For equally likely signals, all the conditional error events are equiprobable and
occur with probability Pr[symbol error]/(M 1) = Pr[symbol error]/(2 1).

There are
ways in which k bits out of may be in error The average
k
number of bit errors per -bit symbol is
Pr[symbol error]
X
21
=
Pr[symbol error].
k
2 1
2 1
k
k=1
The probability of bit error is simply the above quantity divided by :

Pr[bit error] =
21
Pr[symbol error].
2 1
211/291
Introduction
We have considered that the transmitted signal is only degraded by AWGN and it
might be subjected to filtering.
There are communication channels (e.g., mobile wireless channels) where the
received signal is not subjected to a known transformation or filtering.
Typically the gain and/or phase of a digitally modulated transmitted signal is not
known precisely at the receiver.
It is common to model these parameters as random.
Shall consider channel models where the amplitude and/or phase of the received
signal is random.
212/291
Demodulation with Random Amplitude
r(t) = as(t) + w(t),

where a is a random variable with known pdf fa (a).
r2
r2
r2
0D
a E
2
0D
1D
a E
0 R 0T
0D
1D
E a E
a E
1T
0R
a E
1D
r1
0T
0R
1R
1T
a E
r1 = r2
r1
1R
0T
r1
1R
1T
213/291
Performance of BPSK and BFSK
P[error]
P[error]
E{P[error]}
E{P[error]}
!
2Eb
a
(antipodal),
N0
s
!
Eb
a
(orthogonal).
N0
s
!
2Eb
fa (a)da (antipodal),
N0
s
!
Eb
fa (a)da (orthogonal).
a
N0
a
214/291
Optimum Demodulation of BASK
Optimum receiver is determined by the maximum likelihood ratio:

fr1 (r1 |1T ) 1D
R 1.
fr1 (r1 |0T ) 0D
fr1 (r1 |0T ) is N (0, N0 /2), while
fr1 (r1 |1T )
fr1 (r1 |1T , a = a)fa (a)da
E {fr1 (r1 |1T , a = a)} .
Need to know fa (a) to proceed further.

In general the threshold (and hence the decision regions) is a balance between the
different regions given by the values that a takes on weighted by the probability
that a takes on these values.
215/291
M -ary Demodulation with Random Amplitude
If all the signal points lie at distance of Es from the origin (i.e., equal energy), then the
optimum decision regions are invariant to any scaling by a, provided that a 0.
The matched-filter or correlation receiver structure is still optimum, one does not even need
to know fa (a).
r2
The error performance, however,
depends crucially on a and fr2a (a).
a =1
r1
r1
a = 0.5
216/291
Demodulation with Random Phase

Phase uncertainty can be modeled as a uniform random variable (over [0, 2] or [, ]). It does
not change the energy of the received signal.
Q (t ) =
2
sin(2 f c t )
Tb
locus of s2R (t )
Q ( t )
s2R (t ) 1R
E
s1R (t ) = s1 (t )
locus of s1R (t ), s2R (t )
s2R (t ) 1R
s2 ( t )
0 R 0T
s1 (t )
I (t ) =
1T
s 2 (t )
0T
I (t )
1T
2
cos(2 f ct )
Tb
s1R (t ) 0 R

1,Q (t )
s1R (t ) 0 R
2,Q (t )
s2R (t ) 1R
2,I (t )
E
s1 (t )
0T
s2 (t )
1,I (t )
1T
217/291
Optimum Receiver for Noncoherent BASK
w(t),
if 0T
q 2
.
E T cos(2fc t ) + w(t), if 1T
b

wI ,
if 0T
wQ ,
if 0T
rI =
, rQ =
,
E cos + wI , if 1T
E sin + wQ , if 1T
r(t) =
where wI and wQ are statistically independent N (0, N0 /2).

r2 +r2
1
f (rI , rQ |0T ) = f (rI |0T )f (rQ |0T ) = N
exp IN Q .
0
f (rI , rQ |1T )
E
N0
I0
2
0
rI
1
exp
N0
E cos
2

2
+ rQ E sin
1
d
N0
2
r2 +r2
1 IN Q NE
2 Eq 2
2
0
0 I0
e
e
.
rI + rQ
=
N
N0
!1
q0
E
1D
q
D
2 E
N0 1
2
2 R
R 1
e N0 .
rI2 + rQ
rI2 + rQ
I0
N0
2 E
0
0
D
218/291
Decision Regions of BASK with Random Phase

Q ( t )
locus of s2R (t )
s2R (t ) 1R
1 1D
E
0 0D
0R 0T
1T
rI2 + rQ2 = Th =
I (t )
N 0 1 E N 0
I0 e
2 E
t = kTb
kTb
( ) dt
rI
( k 1)Tb
r (t )
1D
I (t ) =
2
cos(2 f ct )
Tb
>
()
<
0D
t = kTb
kTb
( ) dt
rQ
( k 1)Tb
Q (t ) =
Th =
N 0 1 E N 0
I0 e
2 E
Th
2
sin(2 f c t )
Tb
219/291
Different Implementation of BASK Demodulation
y(t)
=
=
2
cos(2fc (t ))[u(t ) u(t Tb ]d
Tb

q
yQ (t)
2 (t) cos 2f
,
t tan1
yI2 (t) + yQ
c
yI (t)
r()
q
2 (t) is the envelope. At the sampling instant, t = kT , then y (kT ) = r
where yI2 (t) + yQ
I
I
b
b
and yQ (kTb ) = rQ .
r (t )
2
cos(2 f ct ), 0 t Tb y ( t )
h(t ) = Tb
0,
elsewhere
t = kTb
1D
r +r
2
I
2
Q
Th =
>
N0
2 E
I 01 e
<
0D
E
N0
Th
220/291
Error Performance
P [error|0T ]
ZZ
f (rI , rQ |0T )drI drQ =
P [error|1T ]
1
N0
=
where Q(, ) =
xe
ZZ
=0
2
N0
2 +r2
rI
Q
N0
drI drQ ,
dd = eTh /N0 .
=Th
1 P [correct|1T ] = 1
1Q
1
e
N0
2E
,
N0
ZZ
f (rI , rQ |1T )drI drQ
2
Th
N0
x2 +2
2
I0 (x)dx is Marcums Q-function.

s
s
"
!#
2
1
2E
2
1
,
Th
P [error] = eTh /N0 +
1Q
.
2
2
N0
N0
221/291
There is about 0.3qdB penalty in power when using such a simpler suboptimum
Eb
threshold, 2E =
.
2
0
10
Noncoherent BASK
with threshold of 0.5E1/2
1
10
Noncoherent BASK
with the optimum threshold
10
Coherent BASK & BFSK
P[error]
Noncoherent BFSK
3
10
Noncoherent DBPSK
4
10
Coherent BPSK
5
10
10
5 6 7 8 9 10 11 12 13 14 15
SNR per bit (Eb/N0) in dB
222/291
Optimum Receiver for Noncoherent BFSK
r(t) =
E T2 cos(2f1 t ) + w(t),
q 2b
E T cos(2f2 t ) + w(t),
b
r1,Q
0T
E cos + w1,I
E sin + w1,Q
r2,I
w2,I
r2,Q
w2,Q
r1,I
if 0T
if 1T
r1,I
r1,Q
r2,I
r2,Q =
r2 +r2
w
1,I
1T
w1,Q
E cos + w2,I
E sin + w2,Q
1 1,IN 1,Q
0
e
f (r1,I , r1,Q , r2,I , r2,Q |0T ) =
N0
!
2
q
r
+r2
1 2,IN 2,Q
2 E
E
2
2
0
e N0 I0
r1,I
+ r1,Q
,
e
N0
N0
r2
+r2
1 1,IN 1,Q
0
e
N0
!
q
2 E
E
2
2
e N0 I0
.
r2,I
+ r2,Q
N0
f (r1,I , r1,Q , r2,I , r2,Q |1T ) =

1
e
N0
2 +r2
r2,I
2,Q
N0
223/291
Optimum Demodulator for BFSK with Random Phase
I0
I0

q
2 E
2
r2,I
N0
q
2 E
2
r1,I
N0
2
+ r2,Q
2
+ r1,Q
1D
R 1
0D
1D
2
2
r2,I
+ r2,Q
R
0D
2
2
r1,I
+ r1,Q
.
Bandpass filter centered at f 2
2
cos(2 f 2 t ), 0 t Tb

h2 (t ) = Tb

0,
elsewhere

t = kTb
r2,2 I + r2,2 Q
+
r (t )
Bandpass filter centered at f1
2
cos(2 f1t ), 0 t Tb

h1 (t ) = Tb

0,
elsewhere
t = kTb
1D
>
<
0D
r1,2I + r1,2Q
The demodulator finds the envelope at the two frequencies and chooses the larger one at the
sampling instant.
224/291
Error Performance of BFSK with Random Phase

By symmetry P [error] = P [error|0T ] = P

r22,I + r22,Q r21,I + r21,Q .
h
i

P error0T , r21,I + r21,Q = R2 =

P (r2,I , r2,Q ) falls outside the circle of radius R0T
1
=
N0
P [error]
E
Z
=0
=0
r1,I =
2
r1,I
N0
2
r1,Q
=e

2 +r2
r1,I
1,Q
N0
2 +r2
r1,I
1,Q
N0
E cos ,
N0
E sin ,
2
f (r1,I , r1,Q |0T )dr1,I dr1,Q
N0
2
dr1,I
sin 1
1 Eb
1 E 2N
1 E
0
d = e 2N0 = e 2N0 .
e
2
2
2
2
cos
1 E 2N
0
e
2
N0
r1,Q =
=0
r1,I =
dd =
2
R
e N0
r1,Q =
2
N0
=R
dr1,Q f ()d.
225/291
Noncoherent BASK is about 0.3 dB more power efficient than noncoherent BFSK.
0
10
Noncoherent BASK
1
10
Noncoherent BASK
10
P[error]
Noncoherent BFSK
3
10
Noncoherent DBPSK
4
10
Coherent BPSK
5
10
10
5 6 7 8 9 10 11 12 13 14 15
226/291
Differential BPSK
Coherent BPSK is 3 dB better than coherent BASK or BFSK: Is it possible to use BPSK on
a channel with phase uncertainty?
Possible if a phase reference can be established at the receiver that is matched to the
received signal.
If the phase uncertainty changes relatively slowly with time, the received signal in one bit
interval can act as a phase reference for the succeeding bit interval.
r ( t ) = Eb
t = kTb
2
cos(2 f c t ) + w(t )
Tb
kTb
( ) dt
rk

( k 1)Tb
Delay Tb
1
Eb
If rk >0, present bit equals previous bit

If
r
k 0, present bit is complement of previous bit
2
1
cos(2 f c (t Tb ) ) +
w (t Tb )
Tb
Eb
But the above method can lead to error propagation!
227/291
Differential BPSK Modulation and Demodulation
0T : no phase change,
1T : phase change.
The decision rule is:
1D
rk R 0.
0D
which is independent of the previous decision.

Since DBPSK is orthogonal signaling, the error analysis for noncoherent BFSK
therefore applies to DBPSK:
P [error]DBPSK =
1 Eb /N0
e
.
2
The only difference is rather than Eb joules/bit, the energy in DBPSK becomes
2Eb . This is because the received signal over two bit intervals is used to make a
decision.
228/291
DBPSK shows about 1 dB degradation over coherent BPSK.

0
10
Noncoherent BASK
1
10
Noncoherent BASK
10
P[error]
Noncoherent BFSK
3
10
Noncoherent DBPSK
4
10
Coherent BPSK
5
10
10
5 6 7 8 9 10 11 12 13 14 15
229/291
Detection over Fading Channels

Fading channel model arises when there are multiple transmission paths from the transmitter to
the receiver.

!
$% &'!( &&'! !) '!
"!!#!
-.+/01
9741 4:7:/.8
*+, -.+/01 2/310144 5678810
230/291
Rayleigh Fading Channel Model
q
Consider the transmitted signal sT (t) = s(t) cos(2fc t), where s(t) = Eb T2 over
b
the bit interval with bit rate rb fc (lowpass signal).
The received signal is:

X
X
r(t) =
rj (t) =
s(t tj )j cos(2fc (t tj ))
j
s(t)
X
j
j cos (2fc t 2fc tj ) = s(t)
X
j
j cos (2fc t j )
where j represents the attenuation and tj the delay along the jth path, which are random
variables. Also because s(t) is lowpass, we approximate s(t) s(t tj ).
Since tj 1/fc , the random phase j lies in the range [0, 2). Now
X
X
r(t) = s(t)
j cos j cos(2fc t) +
j sin j sin(2fc t) .
j
231/291
Rayleigh Fading Channel Model
nF,I =
P
j cos j
E {nF,I } =
E
X
j
and nF,Q =
P
j sin j
have the following moments:
E{j }E{cos j } = 0, E {nF,Q } =
n
o
o
X n 2o n
2
n2F,I =
E j E cos2 j = F ,
2
j
X
j
E{j }E{sin j } = 0,
n
o
X n 2o n 2 o
2
2
nF,Q =
E j E sin j = F ,
2
j
X
X
k sin k
j cos j
E {nF,I nF,Q } = E
XX
j
E {j k } E {cos j sin k } = 0,
{z
}
|
=0
Since the number of multipaths is large, the central limit theorem says that nF,I , nF,Q are
Gaussian random variables.
232/291
nF,I and nF,Q are statistically independent Gaussian random variables, zero-mean, variance
2
F
/2:
!
!
2
2
fnI ,nQ (nI , nQ ) = fnI (nI )fnQ (nQ ) = N 0, F N 0, F .
2
2
The received signal is therefore:
r(t)
s(t) [nF,I cos(2fc t) + nF,Q sin(2fc t)]
s(t) [ cos(2fc t )] ,

n
q
2
2
where = nF,I + nF,Q , = tan1 nF,Q and
=
F,I
f ()
f ()
1
2
(uniform),
(
)
2
2
exp
u()
2
2
F
F
(Rayleigh).
The term Rayleigh fading comes from the envelope distribution.

The phase of the received signal severely degraded but that the amplitude is affected as well:
The incoming signals add not only constructively but also destructively.
233/291
Noncoherent Demodulation of BFSK in Rayleigh Fading
s(t) =
Eb T2 cos(2f1 t),
q 2b
Eb T cos(2f2 t),
b
r(t)
if 0T
if 1T
Eb T2 cos(2f1 t ) + w(t), if 0T
.
q 2b
Eb T cos(2f2 t ) + w(t), if 1T
b
s
s
2
2
+
cos(2f
t)
sin(2f1 t) +w(t),
E
n
E
n
1
F,I
F,Q
b
b
Tb
Tb
{z
}
{z
}
|
|
1,I (t)
1,Q (t)
s
s
2
2
E n
cos(2f2 t) + Eb nF,Q
sin(2f2 t) +w(t),
b F,I
Tb
Tb
|
|
{z
}
{z
}
2,I (t)
0T ,
1T ,
2,Q (t)
The transmitted signal lies entirely within the signal space spanned by 1,I (t), 1,Q (t), 2,I (t)
and 2,Q (t).
234/291
r1,I
r1,Q
0T
p
Eb nF,I + w1,I
p
Eb nF,Q + w2,Q
1T
r1,I
w1,I
r1,Q
=
=
=
w2,Q
p
Eb nF,I + w2,I
p
Eb nF,Q + w2,Q
r2,I
w2,I
r2,I
r2,Q
w2,Q
r2,Q
w1,I , w1,Q , w2,I , w2,Q are due to thermal noise, are Gaussian, statistically independent,
zero-mean, and variance N0 /2.
2
nF,I and nF,Q , are also Gaussian, statistically independent, zero-mean and variance F
/2.
The sufficient statistics are Gaussian, statistically independent, zero-mean, with a variance of
2
either N0 /2 or Eb F
/2 + N0 /2, depending on whether a 0T or 1T .
Computing the likelihood ratio gives the following decision rule:
1D
2
2
2
2
R r1,I
+ r1,Q
,
r2,I
+ r2,Q
0D
235/291
Equivalently the decision rule can be expressed as:

1D q
q
2 + r2
2 + r2 .
r2,I
r1,I
2,Q R
1,Q
0D
which is identical to that for noncoherent BFSK!

Bandpass filter centered at f 2
J 2
cos(2 f 2 t ), 0 t Tb
K
h2 (t ) = L Tb
K
0,
elsewhere
M
t = kTb
;<=>?@A>
B>C>DC@E
r2,2 I + r2,2 Q
+
r (t )
Bandpass filter centered at f1
F 2
cos(2 f1t ), 0 t Tb
G
h1 (t ) = H Tb
G
0,
elsewhere
I
t = kTb
;<=>?@A>
B>C>DC@E
1D
>
<
0D
r1,2I + r1,2Q
236/291
Error Probability of Noncoherent BFSK

P [error] = P [error|0T ] = P

q
hq
i

r22,I + r22,Q r21,I + r21,Q 0T .
Fix the value of r21,I + r21,Q at a specific value, say R2 and compute
P
ZZ

q
hq
i

r22,I + r22,Q R 0T , r21,I + r21,Q = R =
Z
=0
=R
e N0 dd = e
N0

2 +r2
r1,I
1,Q
N0
1
e
N0
2 +r2
r2,I
2,Q
N0
dr2,I dr2,Q
r2,Q
r2, I
Find volume under

Z
f ( r2, I , r2,Q | 0T ) in region

Z outsize the circle
237/291
Average over all possible values of r1,I , r1,Q :

2

2
r1,I +r1,Q

N
0
E e
0T
r1,I =
1
2
2 + F
Eb
N0
2 +r
r1,I
1,Q
N0
2
r1,Q =
f (r1,I , r1,Q |0T )dr1,I dr1,Q
2
Eb F
can be interpreted as the received energy per bit.
The behavior is P [error]

P [error] eSNR .
1
SNR ,
a much much slower rate of decay as compared to
In the log-log plot of the P [error] versus SNR in dB, the error performance curve appears to
be a straight line of slope 1 in the high SNR region.
238/291
Compared to noncoherent demodulation of BFSK in random phase only, at an error probability of

103 about 19 dB more power is needed for noncoherent demodulation of BFSK in Rayleigh
fading!
0
10
Noncoherent FSK
10
Coherent FSK
Coherent PSK
2
P[error]
10
3 Noncoherent
10
FSK
Noncoherent
DPSK
Random phase
10
10
Random amplitude &2phase

(Rayleigh fading, F=1)
Coherent PSK
Coherent ASK & FSK
10
10
15
20
25
Received SNR per bit (dB)
30
35
239/291
BFSK and BPSK with Coherent Demodulation
If the random phase introduced by fading can be perfectly estimated, then coherent
demodulation can be achieved The situation is the same as detection in random
amplitude.
With a Rayleigh fading channel, is a Rayleigh random variable.
0D
For BFSK, the optimum decision rule is r1 R r2 and

1D
P [error] = E
Eb
N0
!)
u
u 2 Eb
1
u
F N0
= 1 t
.
E
2
2 + 2 b
F N0
1D
For BPSK, the optimum decision rule is r1 R 0 and

0D
u
u 2 Eb
1
u
F N0
.
P [error] = 1 t
E
2
1 + 2 b
F N0
240/291
Coherent BPSK is 3 dB more efficient that coherent BFSK, which in turn is 3 dB more efficient
than the noncoherent BFSK.
0
10
Noncoherent FSK
10
Coherent FSK
Coherent PSK
2
P[error]
10
3 Noncoherent
10
FSK
Noncoherent
DPSK
Random phase
10
10
Random amplitude &2phase

(Rayleigh fading, F=1)
Coherent PSK
Coherent ASK & FSK
10
10
15
20
25
30
35
241/291
Diversity
All communication schemes over a Rayleigh fading channel have the same
1
discouraging performance behavior of P [error] SNR
.
The reason is that it is very probable for the channel to exhibit what is called a
deep fade, i.e, the received signal amplitude becomes very small.
Diversity technique: multiple copies of the same message are transmitted over
independent fading channels in the hope that at least one of them will not
experience a deep fade.
Time diversity: Achieved by transmitting the same message in different time slots.
Frequency diversity: Accomplished by sending the message copies in different frequency
slots.
Antenna diversity: Achieved with the use of antenna arrays
242/291
Optimum Demodulation of Binary FSK with Diversity

Consider N transmissions of BFSK over a fading channel:
q q
cos(2f1 t),
E
q b q Tb
s(t) =
Eb T cos(2f2 t),
b
rj (t)
if 0T
if 1T
q q
E
j cos(2f1 t j ) + w(t), 0T
q b q Tb
Eb T j cos(2f2 t j ) + w(t), 1T
b
s
s
q
q
2
2
cos(2f1 t) + Eb nj,Q
sin(2f1 t) +w(t),
Eb nj,I
Tb
Tb
|
|
{z
}
{z
}
(1)
(1)
(t)
(t)
j,I
j,Q
s
s
q
q
2
2
Eb nj,I
cos(2f2 t) + Eb nj,Q
sin(2f2 t) +w(t),
Tb
Tb
{z
}
{z
}
|
|
(2)
(2)
(t)
(t)
j,I
0T
1T
j,Q
for (j 1)Tb t jTb and j = 1, . . . , N .
243/291
r
r
(1)
2 cos(2f t), (1) (t) =
2 sin(2f t),
(t) =
1
1
j,I
T
j,Q
T
r b
r b
(2)
2 cos(2f t), (2) (t) =
2 sin(2f t),
(t) =
2
2
j,I
j,Q
Tb
Tb
(j 1)Tb t jTb , j = 1, . . . , N.
0T
(1)
=
1,I
(1)
r
=
N,I
(1)
(1)
E n
+w
1,I
b 1,I
(1)
=
1,Q
(1)
r
=
N,Q
.
.
.
(1)
(1)
E n
+w
b N,I
N,I
(2)
(2)
= w
1,I
1,I
.
.
.
(2)
(2)
r
= w
N,I
N,I
(1)
(1)
E n
+w
1,Q
b 1,Q
.
.
.
(1)
(1)
E n
+w
b N,Q
N,Q
(2)
(2)
=w
1,Q
1,Q
.
.
.
(2)
(2)
r
=w
N,Q
N,Q
1T
(1)
(1)
= w
1,I
1,I
.
.
.
(1)
(1)
r
= w
N,I
N,I
r
(2)
(2)
(2)
r
=
+w
E n
1,I
1,I
b 1,I
r
(2)
r
=
N,I
.
.
.
(2)
(2)
E n
+w
b N,I
N,I
(1)
(1)
=w
1,Q
1,Q
.
.
.
(1)
(1)
r
=w
N,Q
N,Q
r
(2)
(2)
(2)
r
=
E n
+w
1,Q
1,Q
b 1,Q
r
(2)
r
=
N,Q
.
.
.
(2)
(2)
E n
+w
b N,Q
N,Q
244/291
Number the sufficient statistics corresponding to f1 from 1 to 2N ; sufficient statistics

associated with f2 from 2N + 1 to 4N . The likelihood ratio test is
f (r1 , . . . , rN ; r2N+1 , . . . , r4N |1T )
=
f (r1 , . . . , rN ; r2N+1 , . . . , r4N |0T )
2N
Q
j=1
2N
Q
j=1
2
2
1
erj /(2w )
2w
r 2 /(2t2 )
1
e j
2t
which can be reduced to
4N
X
4N
Q
j=2N+1
4N
Q
j=2N+1
2
2
1
erj /(2t ) 1
2t
D
2N
1D X
rj2 R
j=2N+1
R 1,
2 ) 0
r 2 /(2w
D
1
e j
2w
rj2 .
0D j=1
245/291
Chi-Square Probability Density Function
Consider y = x21 + x22 + + x2N where the xi s are zero-mean, statistically independent
Gaussian random variables with identical variances, 2 . To find fy (y) determine the characteristic
function y (f ) and then inverse transform it.
)
( N
N

N
j2 P x2k
o
n
Y j2f x2
Y
2
j2f y
k=1
k
e
E ej2f xk .
=
=E
=E e
y (f ) = E e
k=1
k=1
n
R j2f x2 x2 /(22 )
1
ke
k
1
e
dxk =
. Therefore
2
1j42 f
R
j2yf
1
1
and
f
(y)
=
e
df , where y 0. From
y (f ) =
y
(1j42 f )N/2
(1j42 f )N/2
R
2p1 ep
ipx
the identity ( ix) e
dx = N 1
u(p),
where R() > 0 and R() > 0,
2
()
ey/(2 )
y2
the pdf is
u(y),
fy (y) =

N
2 2 N N
2
R x1 t
where (x) = 0 t
e dt = (x 1)! for x integer.
E
ej2f xk
246/291
Error Performance of BFSK with Diversity
Define 1 =
P4N
j=2N+1
r2j and 0 =
P2N
2
j=1 rj .
The decision rule is:
1D
1 R 0 .
0D
P [error] = P [error|0T ] =
f (0 |0T )
Z

f (1 |0T )d1 d0 .
f (1 |0T ) and f (0 |0T ) are chi-square distributions:

2
f (1 |0T ) =
N1 e0 /(2t )
N1
e1 /(2w )
1
u(1 ), f (0 |0T ) = 0 N 2N
u(0 ).
2N (N )
2N w
2 t (N )
It can be shown that

P [error] =
N 2 Nj
X
j=1
2
w

1+
t2
2
w
2Nj
(2N j)
.
(N )(N j + 1)
247/291
2 /N as the averaged SNR per transmission. Recognize that

Define T = Eb F
0
t2
2
w
= 1 + T and (x) = (x 1)! for integer x. Then

P [error] =
N1
X N 1 + k 1 + T k
1
.
N
k
(2 + T ) k=0
2 + T
For large values of SNR, 1 + T 2 + T T and

P [error]
N1
X N 1 + k
1 2N 1
1
.
=
N
N
k
(T ) k=0
(T )N
The error performance now decays inversely with the N th power of the received
SNR.
The exponent N of the SNR is generally referred to as the diversity order of the
modulation scheme.
248/291
Compared to no diversity, there is a significant improvement in performance with diversity.

0
10
10
N=1 (No diversity)

2
P[error]
10
N=6
N=2
10
10
N=8
Noncoherent FSK
in random phase only
10
N=4
10
10
15
20
25
Received SNR per transmission (dB)
30
35
249/291
Optimum Diversity
As the diversity order N increases the error performance improves.

This improvement comes at the expense of a reduced data rate in the case of time diversity.
If the transmitters power or equivalently the energy expended per information bit is
constrained to Eb joules then increasing N does not necessarily lead to a better error
performance.
With increased N we increase the probability of avoiding a deep fade, at the same time the
energy, Eb , of each transmission is reduced. Therefore the SNR of each transmission is

reduced which in turn increases the error probability.
There is an optimum value for the diversity order N at each level of error probability. An
empirical relationship is:
Nopt = Ke10 log10 T .
where K is some constant.
250/291
P [error] versus the averaged received SNR per bit, 10 log10
2
E b F
N0

.
10
10
N=1 (No diversity)

2
10
P[error]
N=2
3
10
N=4
4
10
N=6
Noncoherent FSK
in random phase only
10
N=8
6
10
10
15
20
25
30
35
251/291
Determining The Optimum Diversity Order
10
(b)
(a)
14
SNR per bit ranges from 2 to 16 dB in 1dB steps
12
10
10
Optimum diversity order Nopt
P[error]
10
10
8
6
4
2
10
5
10
15
Diversity order N (number of transmissions per bit)
20
0
2
6
8
10
12
Average received SNR per bit (dB)
14
16
252/291
Central Limit Theorem
Central limit theorem states that under certain general conditions the sum of n
statistically independent continuous random variables has a pdf that approaches a
Gaussian pdf as n increases.
P
Let x = n
independent
random variables
i=1 xi . where the xi s are statistically

2
2
with mean, E{xi } = mi , and variance,
i . Then x is a
Pn E (xi mi ) 2 =
P
2
random variable with mean mx = i=1 mi , variance x = n
i=1 i and a pdf of
fx (x) = fx1 (x) fx2 (x) fxn (x).
By the central limit theorem fx (x) approaches a Gaussian pdf as n increases, i.e.,
fx (x)
1
2x
(xmx )2
2
2x
253/291
Example 1: fxi (xi ) are Zero-Mean Uniform

(a)
0.5
Actual distribution
Gaussian fit
n=2
0.45
0.4
n=4
0.35
fx(x)
0.3
0.25
n=6
0.2
n=10
0.15
0.1
0.05
0
8
0
x
254/291
Example 2: fxi (xi ) are Laplacian

(b)
0.4
Actual distribution
Gaussian fit
0.35
0.3
n=2
f (x)
0.25
n=4
0.2
n=6
0.15
n=10
0.1
0.05
0
5
10
x
15
20
25
255/291
INTRODUCTION TO SPACE-TIME
CODING
256/291
A System Model for Wireless Communications

x[m] = VI [m] + jVQ [m ]
Pulse
shaping
filter
VI [ m]
Information bits
sI ( t )
p(t )
Symbol
mapper
QAM signal
Pulse
shaping
filter
VQ [ m]
sQ ( t )
p(t )
Slowly varying
flat fading
Channel, h(t)
s (t )
cos(2 f ct )
sin(2 f ct )
t = mTs
Matched
filter
Inphase
decision
p (Ts t )
cos(2 f ct )
r (t ) = h (t ) s (t ) + w(t )
y[m] = h[m]x[m] + w[m ]
Detected bits
Mux
t = mTs
Matched
filter
Quad.
decision
p (Ts t )
sin(2 f ct )
257/291
Input/Output Model
y[m] = h[m]x[m] + w[m]

The simplest model for the channel gain h[m] is circularly symmetric complex
2 ).
Gaussian random variable, denoted by CN (0, h
The magnitude |h[m]| is a Rayleigh random variable with density:
f|h| (x) =
2x
exp
2
h
x2
2
h
x0
(3)
The squared magnitude |h[m]|2 is exponentially distributed with density

f|h|2 (x) =
1
exp
2
h
x
2
h
x0
(4)
The Rayleigh fading model applies if there are many small reflectors.
258/291
If the line-of-sight path (or a specular path) is large and has a known magnitude,
and that there are also a large number of independent paths, then h[m] can be
modeled as
r
r
1
2
h[m] =
)
h ej +
CN (0, h
(5)
+1
+1
=energy in the specular path/energy in the scattered paths (called K-factor).
The magnitude of h[m] is said to have a Rician distribution:
"
#
( + 1)x2
2x( + 1)
exp
I0
2
2
h
h
2x
( + 1)
2
h
x0
(6)
259/291
where I0 () is the modified Bessel function of zeroth order, define as:

I0 (z) =
1
2
ez cos() d =
ez cos d
z
e
,
2z
z2
1+ 4 ,
z 1,
z 1.
(7)
Nakagami-m fading is even a more general model, which includes both Rayleigh
and Rician (approximately) fading as special cases.
Modified Bessel Function of 0th Order

80
I (z)
60
40
20
0
0
3
z
260/291
Rician Distributions (h2 = 1)

The case = 0 corresponds to Rayleigh fading, whereas = gives an AWGN
channel: Small implies severe fading and large means relatively mild fading.
2.5
=0
=1
=4
=16
Rician Distribution
1.5
0.5
0
0
0.5
1.5
x
2.5
261/291
Nakagami-m Distributions
"
#
2mm x2m1
mx2
exp 2
,
2m
(m)h
h
Z
tm1 et dt
(m) =
m 0.5
(8)
(9)
= (m 1)! for m integer
(10)
Gamma Function
(m)
150
100
50
0
0
3
m
262/291
Nakagami-m Distributions
The case m = 1 reduces to Rayleigh fading. For m = ( + 1)2 /(2 + 1) gives an
approximately Rician fading. Also m = yields an AWGN channel.
2.5
m=1
m=2
m=4
m=8
Nakagamim Distribution
1.5
0.5
0
0
0.5
1.5
x
2.5
263/291
Detection under AWGN Channel
AWGN channel model:

y[m] = x[m] + w[m]
(11)
With antipodal signalling, x[m] = a and the sufficient statistic is {y[m]}.

The (maximum-likelihood) detection rule is
u = {y[m]}
x[m]=a
x[m]=a
Then
Pe
=
=
=
1
1
P (u > 0|x[m] = a) + P (u < 0|x[m] = +a)
2
2
!

a
Q p
=Q
2SNR where SNR = a2 /N0 = Eb /N0
N0 /2
To get Pe = 103 , SNR = 7 dB.
Q(x) function decays exponentially with x2 .
264/291
Q-function
caZO
fa]TV
fa]TV
caUUa]X^ daeS][ZV
daeS][ZV WXSYZS]^ gRRShaZS]^
1 2
e
2
_NQO
`Qa^b_
NOPQRSTSUV WXSYZ[SR \]SPO^^

x
Area = Q ( x )
1
Q(x)
2
exp
2
2
d.
10
10
Q(x)
10
10
10
10
10
3
x
265/291
Coherent Detection in Rayleigh Flat Fading

y[m] = h[m]x[m] + w[m]
Assume now {h[m]} can be perfectly tracked (e.g., using pilot symbols for
channel estimation).
Complex detection problem can be converted to real detection in AWGN:
y = hx + w r = {(h /|h|)y} = |h|x + z with z N (0, N0 /2)
For a given h, the error probability is:
!

q
a|h|
Q p
2|h|2 SNR
with SNR = a2 /N0
=Q
N0 /2
(12)
(13)
To find the overall error performance, need to average over the random gain h.
For Rayleigh fading, model h as h CN (0, 1) so that the average received SNR
is a2 /N0 , the same as in the case of AWGN channel.
The distribution of |h|2 is exp(x).
266/291
Let v = 2|h|2 SNR, then the distribution of v is:

fv (v) =

1
v
exp
2SNR
2SNR
Use the Craigs formula for the Q-function:
1
Pe|h = Pe|v = Q( v) =
/2

( v)2
exp
d
2
2 sin
267/291
Pe
Q( v)fv (v)dv
=2
1
2

exp

1
v2
v
ddv
exp
2
2 sin 2SNR
2SNR
0
0

Z /2 Z
1
1
1
+
dv d
exp v
2
2SNR
2SNR
2
sin
0
0
1
Z /2
Z /2
1
SNR
sin2
d =
d
1+
2
sin
SNR
+ sin2
0
0
"
#
Z /2
Z
2SNR
SNR
1 /2
1
1
d
d =
1cos(2)
2
0
(2SNR + 1) cos(2)
0
SNR +
2
Z
2SNR
1
d
0 (2SNR + 1) cos
Z
/2
Finally, make use of the following identity:

1
Pe =
2
dx
0 1+a cos x
SNR
1 + SNR
1a2
1
4SNR
we obtain:
(14)
Observation: Error probability decays only inversely with SNR!

268/291
Error Performance Comparison

0
10
pe
10
10
10
15
10
20
BPSK over AWGN

Noncoherent (Orthogonal), fading
Coherent (BPSK), fading
10
10
SNR (dB)
20
30
40
269/291
Diversity Techniques for Fading Channels
Over fading channels, the root cause of poor performance is that there is only
one signal path & there is a significant probability that this path will be in a deep
fade.
Diversity solution: Ensure that the information symbols pass through multiple
signal paths, each of which fades independently .
Typical diversity techniques:
Time Diversity (e.g., coding and interleaving)
Frequency Diversity (e.g., OFDM)
Space Diversity (e.g., MIMO)
Signal Space Diversity
Macro Diversity (e.g, in cellular network)
A wireless system typically uses several types of diversity.

The key issue is how to exploit the diversity in an efficient manner (exploiting all
the available degrees of freedom).
270/291
Time Diversity
Time diversity is achieved by averaging the fading over time.
Typically, the coherence time is of the order of hundreds of symbols. Therefore,
the channel is highly correlated across consecutive symbols.
To ensure the symbols are transmitted through independent fading gains,
interleaving over codewords is required.
hl
L=4
x0
x1
x2
x3
271/291
Time Diversity with Repetition Coding

Consider flat fading with a codeword x = [x1 , . . . , xL ] of L symbols. The
received signal is:
yl = h l x l + w l
(15)
Using interleaving, {xl } are transmitted sufficiently far apart in time and we can
assume hl s are independent.
L is called the number of diversity branches.
The simplest code is repetition code, where xl =x1 y = hx1 + w for all l.
The detection problem becomes the scalar detection:
h
h
h
y = ||h||x1 +
w, where
w CN (0, N0 )
||h||
||h||
||h||
(16)
This is called a matched filter , or a maximal ratio combiner : It weighs the

received signal in each branch in proportion to the signal strength.
p

With BPSK, the conditional error probability is Q
2||h||2 SNR , where
P
2
||h||2 = L
l=1 |hl | .
272/291
||h||2 is Chi-square distributed with 2L degrees of freedom:

1
xL1 exp (x) , x 0
f (x) =
(L 1)!
(17)
22L Distribution
0.5
L=1
L=2
L=3L=4
L=5
0
0
Averaging over ||h||2 , we have:

Z

Q
Pe =
2xSNR f (x)dx
10
1
2
L L1
X
l=0
L1+l
l

1+
2
l
with =
SNR
1 + SNR
273/291
10
L=1
10
pe
L=2
10
L=3
10
L=4
10
15
20
BPSK over AWGN

BPSK over flat Rayleigh fading
10
10
SNR (dB)
20
30
40
274/291
Spatial (or Antenna) Diversity
Time diversity might not be possible when there is a strict delay

constraint and/or the coherence time is large.
Diversity be obtained by placing multiple antennas at the transmitter
and/or the receiver.
Antennas need to be placed sufficiently far apart:
For a mobile: half to one carrier wavelength.
For a base station: several to tens of wavelengths.
275/291
Receive Diversity
y1[m] = h1[m]x[m] + w1[m]

h1[m]
x[m]
y 2 [ m] = h2 [ m]x[ m] + w2 [ m]
h2 [ m]
With L receive antennas:

yl [m] = hl [m]x[m] + wl [m],
y[m] = h[m]x[m] + w[m]
l = 1, . . . , L
This is exactly the same detection problem as in repetition

coding/interleaving: Can achieve diversity order of L with coherent
maximum ratio combiner.
276/291
Transmit Diversity
y[ m] = ( h1[ m] + h2 [ m]) x[ m] + w[ m]
x[m]
x[m]
h1[m]
y[ m] = ( h1* [ m] + h2*[ m]) x[ m] + w[ m]
h2 [ m]
x[ m]
= h1* [ m], h2* [ m]
+ w[ m]
[
]
x
m
h* [ m ]
x[ m ]
= h [ m]x[ m] + w[ m]
*
If only the receiver knows the channel (h1 [m] and h2 [m]), only
diversity order of 1 is achieved!
If the transmitter also knows the channel, transmit beamforming can
be done to achieve diversity order of 2:
x[m] =
h[m]
x[m] y[m] = kh[m]k x[m] + w[m]
kh[m]k
What happens if the transmitter does not know the channel?

277/291
Transmit Diversity with Space-Time Coding
Transmitting the same symbol simultaneously at the antennas does

not work.
Using the antennas one at a time and sending the same symbol over
different antennas (just like repetition coding) Achieve full
diversity order L, but inefficient.
Space-time coding: Achieve full diversity by using all antennas
simultaneously and without repetition Highly efficient.
There are simple but clever solutions!
278/291
Alamouti Scheme
Time slot 1 ( m = 1)
x1[ 2] = u 2*
x1[1] = u1
h1
x2 [1] = u 2
Time slot 2 (m = 2)
x2 [2] = u1*
h2
[y[1] y[2]] =

y[1]
=
y [2]
h1
y[1]
y[ 2]
h2

u1 u2
[h1 h2 ]
+ [w[1] w[2]]
u2
u1

h1
h2
u1
w[1]
+
h2 h1
u2
w [2]
Transmit two complex symbols (u1 and u2 ) over two time slots.
279/291
Alamouti Scheme
Time slot 2 (m = 2)
x1[ 2] = u 2*
x1[1] = u1
h1
x2 [1] = u 2
y[1]
y [2]
x2 [2] = u1*
h2
h1
y[1]

|
h1
h2
y[ 2]
h2

u1
w[1]
h2
+
u2
w [2]
h1
{z
}
H
Projecting on to the orthogonal columns of H yields:

ri = khk ui + wi ,
i = 1, 2
280/291
Alamouti Scheme
Time slot 2 (m = 2)
x1[ 2] = u 2*
x1[1] = u1
h1
x2 [1] = u 2
h1
y[1]
h2
x2 [2] = u1*
y[ 2]
h2

y[1]
= = |h1 |2 + |h2 |2 u1 + w
1
y [2]

y[1]
2
= = |h1 |2 + |h2 |2 u2 + w
r2 = [h2 h1 ]
y [2]
r1 = [h1 h2 ]
281/291
Space-Time Coding for MIMO Channels
x1[1] = u1
x2 [1] = u 2
y1[1]
y2 [1]
Time slot 2 (m = 2)
x1[ 2] = u 2*
x2 [2] = u1*
y1[2]
y2 [ 2]
u
X= 1
u 2
u 2*
u1*
Designing a space-time block code is to design the set of codeword

matrices {X}.
The design goals are to achieve the maximum diversity order L and
coding gain, while allowing a simple receiver implementation.
Many good orthogonal and quasi-orthogonal designs have been
found.
282/291
Design Criteria for Space-Time Coding
Space-time code: A set of L N matrices {Xi }, where L is the #

of transmit antennas, N is the block length.
Input/output channel model: y = h X + w .
Conditional pairwise error probability of confusing XB with XA :
!
kh (XA XB )k
p
Pr (XA XB |h) = Q
2 N0 /2
With SNR = 1/N0 , the pairwise error probability (PEP) is:

!#
"
p
SNR h (XA XB )(XA XB ) h
Pr (XA XB ) = E Q
2
Apply unitary transformation: (XA XB )(XA XB ) = UU ,
where U is unitary and = diag(21 , 22 , . . . , 2L ), with l s are the
singular values of (XA XB ).
283/291
Design Criteria for Space-Time Coding
q
PL 2 2
L
Y
SNR l=1 |h
l | l
1

Pr (XA XB ) = E Q
1 + SNR2l /4
2
l=1
4L
4L
= U h
=
, where h
Q
L
SNRL det[(XA XB )(XA XB ) ]
SNRL l=1 2l
Rank Criterion: If all the 2l are strictly positive, the the maximum
diversity order of L is achieved Should maximize the number of
positive eigenvalues 2l , which is also the rank of the codeword
difference matrix.
Determinant Criterion: The coding gain is determined by the
minimum of the determinant det[(XA XB )(XA XB ) ] over all
codeword pairs Should maximize the minimum determinant.
284/291
Combining Methods for Receive Diversity

Consider a detection problem where with L diversity branches (i.e., with
repetition coding over time or with receive antenna array). This means
that there are L received signals that are related to the same information
symbol x as
yl = hl x + wl , l = 1, 2, . . . L.
(18)
where hl s are independent random variables, wl s are i.i.d. CN (0, N0 ).
Let Es = E{|x|2 }. Then SNR = Es /N0 and the instantaneous SNR of
the lth branch is l = |hl |2 (Es /N0 ) = |hl |2 SNR.
(a) It has been shown that the coherent ML detector is the maximum
ratio combiner (MRC). Show that the instantaneous SNR at the
PL
output of the MRC is given by = l=1 l . How does the above
result depend on the fading statistics of the diversity branches?
(b) Equal gain combiner (EGC) is simpler than MRC since it only
co-phases the signal on each diversity branch and then combines
them with equal weighting. Find the expression for the SNR at the
output of the EGC.
285/291
(c) Now consider selection combiner (SC) which picks the diversity
branch with the highest instantaneous SNR and uses it for detection.
Such a combining scheme requires only one RF chain (as opposed to
L RF chains in both MRC and EGC) in the receiver, hence reducing
the implementation cost. Assume that the fading on all diversity
branches are identically distributed with hl CN (0, 1) (i.e., Rayleigh
fading). Show that the pdf of the instantaneous SNR at the output
of the SC is:
iL1
L h
e SNR
(19)
1 e SNR
f () =
SNR
(d) Although the pdf of the instantaneous SNR needs to known to

compute the error performance of any diversity scheme considered
above, the average SNR is also a very useful and informative
performance index. Assume Rayleigh fading with hl CN (0, 1) on
all branches, obtain the expressions for the average SNR at the
output of the MRC, EGC and SC.
286/291
(e) Finally consider the hybrid combiner (HC) which uses J RF chains
where 1 < J < L and L > 2. The HC chooses J receive signals with
the highest instantaneous SNR and combines them using MRC. It
can be shown that for Rayleigh fading with hl CN (0, 1) on all
branches, the average SNR at the output of the HC is:
"
#
L
X
1
HC = SNR J 1 +
(20)
m
m=J+1
The following figure plots the SNR gains in dB as functions of the

number of diversity branches achieved by the above four diversity
schemes. Here the SNR gain is with respect to the case of single
diversity branch, whose average received SNR is SNR).
287/291
Average SNR Gains of Different Combining Methods

10
9
8
SNR Gain (dB)
7
6
5
4
3
MRC
EGC
SC
HC J=2
HC J=4
2
1
0
0
4
6
L (# of diversity branches)
10
All schemes improve with the number of diversity branches.

HC provides a good compromise between MRC and EGC (look at L = 3 and
J = 2).
SC is the worst in performance but simplest in implementation. Its the other way
around for MRC.
288/291
Laboratory Exercise
This laboratory exercise is concerned with Matlab simulation of a wireless
communication system using QPSK modulation together with the
Alamouti space-time code.
You will be provided a complete Matlab program that simulates QPSK
modulation over an AWGN channel. Please study the Matlab program
carefully and understand all major functionalities (such as setting SNR
values, generating information bits, forming QPSK symbols, simulate
AWGN channel, performing demodulation, counting error bits, etc.).
The specific tasks are as follows:
1. Modify the program to simulate the transmission and detection of
QPSK signals over a Rayleigh fading channel. The only change is to
include a random channel gain h(m) to simulate Rayleigh fading and
how to handle this gain as the receiver. Note that you are supposed
to know h(m) at the receiver.
As a specific requirement for this part, you need to obtain and plot
the BER curves, both by theory and simulation, on the same graph
for SNR = [0 : 5 : 30] dB. If your program is correct, the results
should look the same as the red curve on page 13 of the lecture slides.
289/291
2. Modify the program to simulate the transmission and detection of

QPSK signals with Alamouti space-time code over a Rayleigh fading
channel. It is assumed that there are 2 transmit antennas and 1
receive antenna, i.e., a 2 1 MISO system. The major changes you
need to do in this part are to simulate Alamouti space-time code and
its decoding.
As in Part 1, you need to obtain and plot the BER curves, both by
theory and simulation, on the same graph for SNR = [0 : 5 : 30] dB.
If your program is correct, the results should look the same as the
curve labeled with L = 2 on page 18 of the lecture slides.
290/291
THANK YOU!
291/291

EE450 - Wireless Communications: Professor Ha Nguyen

Caricato da

Informazioni sul documento

Descrizione originale:

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

EE450 - Wireless Communications: Professor Ha Nguyen

Caricato da

Copyright:

Formati disponibili

EE450 Wireless Communications

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Communication Systems (Analog or Digital)

Source: Originates a message (e.g., human voice, TV picture, email message)

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Analog, Digital, Continuous-Time, Discrete-Time?

(a) Analog continuous-time signal

(b) Digital continuous-time signal

(c) Analog discrete-time signal

(d) Digital discrete-time signal

In signal classification, the adjectives analog and digital refer to the

What is Digital Communication?

BASK signal Digital message

Digital Modulation: Over a finite duration, the transmitted signal belongs to a

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Received BASKTransmitted BASK Received AM Transmitted AM

Why Digital Communications?

Since the transmitted signal in digital communications belong to a finite set of

Analog-to-Digital Conversion = Sampling + Quantization

Types of Modulation and Hardware Realization (Implementation)

Continuous-time implementation is mostly adopted in this course to explain the

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Block Diagram of a Digital Communication System

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Digital vs. Analog Communications

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Random Experiment, Sample Space, Events

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Example of Throwing a Fair Die

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Three Axioms of Probability

For a discrete sample space , define a probability measure P on as a set function

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Important Properties of the Probability Measure

1. P (E c ) = 1 P (E), where E c denotes the complement of E. This property

4. If E1 E2 then P (E1 ) P (E2 ). This says that if event E1 is contained in E2

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

P (E1 |E2 )P (E2 )

Total Probability Theorem

The events {Ei }n

for all 1 i, j n and i 6= j

If for an event A we have the conditional probabilities {P (A|Ei )}n

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Examples on Probability Measures

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Random Variable in the Example of Throwing a Fair Die

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Cumulative Distribution Function (cdf)

cdf gives a complete description of the random variable. It is defined as:

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Typical Plots of cdf

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Probability Density Function (pdf)

The pdf is defined as the derivative of the cdf:

Basic properties of pdf:

EE450 Wireless Communications, Duy Tan University, Da Nang, Vietnam

Bernoulli Random Variable

With 2 = 108 andA = 104