www.rejinpaul.

com

Information Theory and Coding

10EC55

Information Theory and Coding

Subject Code
: 10EC55
No. of Lecture Hrs/Week : 04
Total no. of Lecture Hrs. : 52

IA Marks
Exam Hours
Exam Marks

: 25
: 03
: 100

PART - A
Unit 1:
Information Theory: Introduction, Measure of information, Average information content of
symbols in long independent sequences, Average information content of symbols in long
dependent sequences. Mark-off statistical model for information source, Entropy and
information rate of mark-off source.
6 Hours
Unit 2:
Source Coding: Encoding of the source output, Shannons encoding algorithm.
Communication Channels, Discrete communication channels, Continuous channels.
6 Hours
Unit 3:
Fundamental Limits on Performance: Source coding theorem, Huffman coding, Discrete
memory less Channels, Mutual information, Channel Capacity. 6 Hours
Unit 4:
Channel coding theorem, Differential entropy and mutual information for continuous
ensembles, Channel capacity Theorem. 6 Hours
PART - B
Unit 5:
Introduction to Error Control Coding: Introduction, Types of errors, examples, Types of
codes Linear Block Codes: Matrix description, Error detection and correction, Standard arrays
and table look up for decoding. 7 Hours
Unit 6:
Binary Cycle Codes, Algebraic structures of cyclic codes, Encoding using an (n-k) bit shift
register, Syndrome calculation. BCH codes. 7 Hours
Unit 7:
RS codes, Golay codes, Shortened cyclic codes, Burst error correcting codes. Burst and
Random Error correcting codes. 7 Hours
Unit 8:
Convolution Codes, Time domain approach. Transform domain approach.
7Hours
Text Books:
Digital and analog communication systems, K. Sam Shanmugam, John Wiley,
1996. Digital communication, Simon Haykin, John Wiley, 2003.
Reference Books:
ITC and Cryptography, Ranjan Bose, TMH, II edition, 2007
Digital Communications - Glover and Grant; Pearson Ed. 2nd Ed 2008

Dept. of ECE/SJBIT

Page 1

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

INDEX SHEET
Sl
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

Unit & Topic of Discussion

PAGE NO.

PART - A
UNIT 1: INFORMATION THEORY
Introduction
Measure of information
Average information content of symbols in long independent
Sequences
Average information content of symbols in long dependent
Sequences
Mark-off statistical model for information source,
Entropy and information rate of mark-off source.
Review questions
UNIT 2
SOURCE CODING
Encoding of the source output
Shannons encoding algorithm
Communication Channels
Discrete communication channels
Review questions
UNIT 3
FUNDAMENTAL LIMITS ON PERFORMANCE
Source coding theorem
Huffman coding
Discrete memory less Channels
Mutual information
Channel Capacity
Review questions
UNIT 4
Continuous Channel
Differential entropy and mutual information for continuous
Ensembles
Channel capacity Theorem
Review questions
PART B
UNIT 5
INTRODUCTION TO ERROR CONTROL CODING
Introduction
Types of errors
Types of codes
Linear Block Codes: Matrix description.
Error detection and correction
Standard arrays and table look up for decoding

Dept. of ECE/SJBIT

5
5
8
9
11
19
27
29
30
31
44
45
73
74
75
75
81
88
90
110
111
112
119
121
129
130
131
133
133
136
146
149

Page 2

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54

10EC55

Hamming codes
Review questions

153
155
156
157
167
171
178
181
182

UNIT 6
Binary Cyclic Codes
Algebraic structures of cyclic codes
Encoding using an (n-k) bit shift register,
Syndrome calculation.
BCH codes
Review questions
UNIT 7

183
184
188
189
189
191
196
197
198
200
206
216

Introduction
Golay codes and Shortened cyclic codes
R S codes
Burst error correcting codes
Burst and Random Error correcting codes
Review questions
UNIT 8
Convolution Codes
Time domain approach
Transform domain approach.
Review questions

Dept. of ECE/SJBIT

Page 3

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

PART A
Unit 1:

Information Theory

Syllabus:
Introduction, Measure of information, Average information content of symbols in long
independent sequences, Average information content of symbols in long dependent
sequences. Mark-off statistical model for information source, Entropy and information rate of
mark-off source. 6 Hours

Text Books:

Digital and analog communication systems, K. Sam Shanmugam, John Wiley,

1996.

Reference Books:
Digital Communications - Glover and Grant; Pearson Ed. 2nd Ed 2008

Dept. of ECE/SJBIT

Page 4

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Unit 1: Information Theory

1.1 Introduction:

Communication

Communication involves explicitly the transmission of information from one point to another,
through a succession of processes.
Basic elements to every communication system
o

Transmitter

Channel and

Receiver
Communication System

Source
of
information

Transmitter

Message
signal

CHANNEL

Transmitted
Signal

User
of
information

Receiver

Received
signal

Estimate of
message signal

Information sources are classified as:

INFORMATION
SOURCE
ANALOG

DISCRETE

Source definition
Analog : Emit a continuous amplitude, continuous time electrical wave
from. Discrete : Emit a sequence of letters of symbols.
The output of a discrete information source is a string or sequence of symbols.

1.2 Measure the information:

To measure the information content of a message quantitatively, we are required to arrive at
an intuitive concept of the amount of information.
Consider the following examples:
A trip to Mercara (Coorg) in the winter time during evening hours,
1. It is a cold day
2. It is a cloudy day
3. Possible snow flurries

Dept. of ECE/SJBIT

Page 5

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Amount of information received is obviously different for these messages.

o Message (1) Contains very little information since the weather in coorg is cold for most
part of the time during winter season.
o The forecast of cloudy day contains more informat ion, since it is not an event that
occurs often.
o In contrast, the forecast of snow flurries convey s even more information, since the
occurrence of snow in coorg is a rare event.
On an intuitive basis, then with a knowledge of the occurrence of an event, what can be said
about the amount of information conveyed?
It is related to the probability of occurrence of the event.
What do you conclude from the above example with regard to quantity of information?
Message associated with an event least likely to occur contains most information. The
information content of a message can be expressed quantitatively as follows:

The above concepts can now be formed interns of probabilities as follows:

Say that, an information source emits one of q po ssible messages m1, m2 m q with p1, p2
pq as their probs. of occurrence.
th

Based on the above intusion, the information content of the k message, can be written as
I (mk)

1
pk

Also to satisfy the intuitive concept, of information.

I (mk) must

zero as pk

Therefore,
I (mk) > I (mj);
I (mk)

if pk < pj

O (mj);

if pk

I (mk) O;

when

------ I

O < pk < 1

Another requirement is that when two independent messages are received, the total
information content is
Sum of the information conveyed by each of the messages.
Thus, we have
I (mk & mq)

I (mk & mq) = Imk + Imq ------ I

We can define a measure of information as

1
I (mk ) = log

------ III

p
k

Dept. of ECE/SJBIT

Page 6

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Unit of information measure

Base of the logarithm will determine the unit assigned to the information content.
Natural logarithm base : nat
Base - 10

: Hartley / decit

Base - 2

: bit

Use of binary digit as the unit of information?

(p1 = p2 =
Is based on the fact that if two possible binary digits occur with equal proby
) then the correct identification of the binary digit conveys an amount of information.
I (m1) = I (m2) = log 2 ( ) = 1 bit
One bit is the amount if information that we gain when one of two possible and equally
likely events occurs.

Illustrative Example
A source puts out one of five possible messages during each message interval. The probs. of
these messages are p = 1 ; p = 1 ; p = 1 : p = 1 , p 1
1

16
16
2
4
4
What is the information content of these messages?
1
I (m1) = - log2

2
1

I (m2) = - log2

4
1

I (m3) = - log

8
1

I (m4) = - log2

16
1

I (m5) = - log2

16

= 1 bit

= 2 bits

= 3 bits

= 4 bits

= 4 bits

HW: Calculate I for the above messages in nats and Hartley

Dept. of ECE/SJBIT

Page 7

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Digital Communication System:

Source of
information

Estimate of the
Message signal

Message signal

Source
Encoder
Source
code word
Channel
Encoder
Channel
code word

Transmitter

User of
information

Source
decoder
Estimate
of
source codeword
Channel
decoder
of
Estimate
channel codeword

Modulator

Receiver

Demodulator

Waveform

Channel

Received
signal

Entropy and rate of Information of an Information Source /

Model of a Mark off Source
1.3 Average Information Content of Symbols in Long Independence Sequences
Suppose that a source is emitting one of M possible symbols s0, s1 .. s M in a statically
independent sequence
Let p1, p2, .. p M be the problems of occurrence of the M-symbols resply. suppose further
that during a long period of transmission a sequence of N symbols have been generated.
On an average

s 1 will occur NP1 times

S2 will occur NP2 times

:

si will occur NPi times

1
The information content of the i th symbol is I (si) = log

bits

PiN occurrences of si contributes an information content of

i

ii

p
i

Total information content of the message is = Sum of the contribution due to each of

Dept. of ECE/SJBIT

bitsPN.I(s)=PN.log

Page 8

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

M symbols of the source alphabet

1

bits

i.e., Itotal = NP 1 log

i 1

Averageinforamtion content

H=

total

M
NP1 log

per symbol in given by

This is equation used by Shannon

bits per

p
i

i 1

---- IV

symbol

Average information content per symbol is also called the source entropy.
1.4 The average information associated with an extremely unlikely message, with an extremely
likely message and the dependence of H on the probabilities of messages
consider the situation where you have just two messages of probs. p and (1-p).
Average information per message is H = p log

1
1
(1 p) log
p
1p

At p = O, H = O and at p = 1, H = O again,
The maximum value of H can be easily obtained as,
H = log 2 + log 2 = log 2 = 1
max

Hmax = 1 bit / message

Plot and H can be shown below

H

The above observation can be generalized for a source with an alphabet of M symbols.
Entropy will attain its maximum value, when the symbol probabilities are equal,
i.e., when p1 = p2 = p3 = . = p

M=

1
M

Hmax = log2 M bits / symbol

Hmax = p M log
Hmax = p M log

1
pM
1
1

Dept. of ECE/SJBIT

Page 9

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Hmax = M log 2 M log 2 M

Information rate

If the source is emitting symbols at a fixed rate of r s symbols / sec, the average source
information rate R is defined as
R = rs . H bits / sec

Illustrative Examples

1. Consider a discrete memoryless source with a source alphabet A = { so, s1, s2} with
respective probs. p0 = , p1 = , p2 = . Find the entropy of the source.
Solution: By definition, the entropy of a source is given by
M

H =

pi

log

i 1

1
bits/ symbol
pi

H for this example is

2

H (A) =

pi

log

pi
Substituting the values given, we get
i0

H (A) = po log

1 + P
1 p log 1
1 log
2
Po
p2
p1

= log2 4 + log2
=

3
2

4 + log2 2

1.5 bits

if rs = 1 per sec, then

H (A) = rs H (A) = 1.5 bits/sec
2. An analog signal is band limited to B Hz, sampled at the Nyquist rate, and the samples are
quantized into 4-levels. The quantization levels Q1, Q2, Q3, and Q4 (messages) are assumed
independent and occur with probs.
P1 = P2 = 1 and P2 = P3 = 3 . Find the information rate of the source.
8
8
Solution: By definition, the average information H is given by
H = p log
1

1 + p log 1 + p log 1 + p log 1

4
2
3
p1
p2
p3
p4

Substituting the values given, we get

Dept. of ECE/SJBIT

Page 10

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

H = 1 log 8 + 3 log 8 + 3 log 8 + 1 log 8

8
8
3
8
3 8
= 1.8 bits/ message.
Information rate of the source by definition is
R = rs

R = 2B, (1.8) = (3.6 B) bits/sec

3. Compute the values of H and R, if in the above example, the quantities levels are so chosen
that they are equally likely to occur,
Solution:
Average information per message is
H = 4 ( log 4) = 2 bits/message
2

and R = rs H = 2B (2) = (4B) bits/sec

1.5 Mark off Model for Information Sources
Assumption
A source puts out symbols belonging to a finite alphabet according to certain probabilities
depending on preceding symbols as well as the particular symbol in question.

Define a random process

A statistical model of a system that produces a sequence of symbols stated above is and which

is governed by a set of probs. is known as a random process.

Therefore, we may consider a discrete source as a random process
And the converse is also true.
i.e. A random process that produces a discrete sequence of symbols chosen from a finite set
may be considered as a discrete source.

Discrete stationary Mark off process?

Provides a statistical model for the symbol sequences emitted by a discrete source.

General description of the model can be given as below:

1. At the beginning of each symbol interval, the source will be in the one of n possible states 1, 2,
.. n
Where n is defined as

Dept. of ECE/SJBIT

Page 11

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

n (M)

M = no of symbol / letters in the alphabet of a discrete stationery source,

m = source is emitting a symbol sequence with a residual influence lasting
m symbols.
i.e. m: represents the order of the source.
nd

m = 2 means a 2

order source

m = 1 means a first order source.

The source changes state once during each symbol interval from say i to j. The probabilityy of
this transition is Pij. Pij depends only on the initial state i and the final state j but does not depend on
the states during any of the preceeding symbol intervals.
2. When the source changes state from i to j it emits a symbol.
Symbol emitted depends on the initial state i and the transition ij.
3. Let s1, s2, .. s M be the symbols of the alphabet, and let x1, x2, x3, x k, be a sequence of
th

random variables, where xk represents the k symbol in a sequence emitted by the source.
th

Then, the probability that the k symbol emitted is sq will depend on the previous symbols x 1, x2,
x3, , x k1 emitted by the source.
i.e., P (Xk = sq / x1, x2, , x

k1 )

4. The residual influence of

th

x1, x2, , x k1 on xk is represented by the state of the system at the beginning of the k symbol
interval.
i.e. P (xk = sq / x1, x2, , x
th

k1 ) = P (xk = sq / Sk)

When Sk in a discrete random variable representing the state of the system at the beginning of the

k interval.
Term states is used to remember past history or residual influence in the same context as the use
of state variables in system theory / states in sequential logic circuits.
System Analysis with regard to Markoff sources
Representation of Discrete Stationary Markoff sources:
o Are represented in a graph form with the nodes in the graph to represent states and the
transition between states by a directed line from the initial to the final state.

Dept. of ECE/SJBIT

Page 12

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

o Transition probs. and the symbols emitted corresponding to the transition will be shown
marked along the lines of the graph.
A typical example for such a source is given below.

P1(1) = /3

P2(1) = /3
1

P3(1) = /3
C
A

It is an example of a source emitting one of three symbols A, B, and C

o The probability of occurrence of a symbol depends on the particular symbol in question and
the symbol immediately proceeding it.
o

Residual or past influence lasts only for a duration of one symbol.

Last symbol emitted by this source

o The last symbol emitted by the source can be A or B or C. Hence past history can be
represented by three states- one for each of the three symbols of the alphabet.

Nodes of the source

o Suppose that the system is in state (1) and the last symbol emitted by the source was A.
o The source now emits symbol (A) with probability and returns to state (1).
OR
o The source emits letter (B) with probability andgoes to state (3)
OR
o

The source emits symbol (C) with probability andgoes to state (2).

A 1

To state 2

To state 3

State transition and symbol generation can also be illustrated using a tree diagram.
Tree diagram

Tree diagram is a planar graph where the nodes correspond to states and branches
correspond to transitions. Transitions between states occur once every Ts seconds.
Along the branches of the tree, the transition probabilities and symbols emitted will be
indicated.Tree diagram for the source considered

Dept. of ECE/SJBIT

Page 13

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Symbol
probs.

Symbols
emitted

1 AA

C
B

2 AC

/3

/3

3 AB

1 CA

C
B

2 CC

1 BA

C
B

2 BC

3 CB

3 BB

1 AA

C
B

2 AC

C
B

Symbol
sequence

3 AB
1 CA
2 CC
3 CB

1
A

/3 3

1 BA

C
B

2 BC

1 AA

C
B

2 AC

State at the end of the

first symbol internal

Dept. of ECE/SJBIT

3 AB

1 CA

C
B

2 CC

1 BA

C
B

2 BC

Initial
state

3 BB

3 CB

3 BB

State at the end of the

second symbol internal

Page 14

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Use of the tree diagram

Tree diagram can be used to obtain the probabilities of generating various symbol sequences.
Generation a symbol sequence say AB
This can be generated by any one of the following transitions:
1

OR
2

OR
3

Therefore proby of the source emitting the two s ymbol sequence AB is given by
P(AB) =

P ( S1 = 1, S2 = 1, S3 = 3 )
Or

P ( S1 = 2, S2 = 1, S3 = 3 )
Or

----- (1)

P ( S1 = 3, S2 = 1, S3 = 3 )
Note that the three transition paths are disjoint.
Therefore P (AB) = P ( S1 = 1, S2 = 1, S3 = 3 ) + P ( S1 = 2, S2 = 1, S3 = 3 )
+ P ( S1 = 2, S2 = 1, S3 = 3 )

----- (2)

The first term on the RHS of the equation (2) can be written as
P ( S1 = 2, S2 = 1, S3 = 3 )
= P ( S1 = 1) P (S2 = 1 / S1 = 1) P (S3 = 3 / S1 = 1, S2 = 1)
= P ( S1 = 1) P (S2 = 1 / S1= 1) P (S3 = 3 / S2 = 1)

Dept. of ECE/SJBIT

Page 15

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Recall the Markoff property.

Transition probability to S3 depends on S2 but not on how the system got to S2.
1

Therefore, P (S1 = 1, S2 = 1, S3 = 3 ) = /3 x x
Similarly other terms on the RHS of equation (2) can be evaluated.

1
Therefore P (AB) =

1
/3 x x + /3 x x + /3 x x =

48

1
12

Similarly the probs of occurrence of other symbol sequences can be computed.

Therefore,
In general the probability of the source emitting a particular symbol sequence can be
computed by summing the product of probabilities in the tree diagram along all the paths that yield
the particular sequences of interest.
Illustrative Example:
1. For the information source given draw the tree diagram and find the probs. of messages of lengths
1, 2 and 3.

A 1
/4

C
C

2 B3

/4

p1 =

P2 =

Source given emits one of 3 symbols A, B and C

Tree diagram for the source outputs can be easily drawn as shown.

Dept. of ECE/SJBIT

Page 16

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

A
A

A
C

1 AAA

2 AAC
1 ACC

/4

2 ACB

1 CCA

2 CCC
1 CBC

/4

2 CBB

2 CAC
1 CCC

/4

1 CAA

2 CCB

1 BCA

2 BCC
1 BBC

/4

2 BBB

Messages of length (1) and their probs

A
B

x = /8
3

x = /8
11

Cx+x==88

Message of length (2)

How may such messages are there?
Seven
Which are they?
AA, AC, CB, CC, BB, BC & CA
What are their probabilities?
Message AA : x x

Message AC: x x =

9
32

3
and so on.
32

Tabulate the various probabilities

Dept. of ECE/SJBIT

Page 17

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

Messages of Length (1)

3
A

8
B

3
8
1
4

10EC55

Messages of Length (2)

9
AA

AC

CB

CC

BB

BC

CA

32
3
32
3
32
2
32
9
32
3
32
3
32

Messages of Length (3)

27
AAA

AAC

ACC

ACB

BBB

BBC

BCC

BCA

CCA

CCB

CCC

CBC

CAC

CBB

CAA

Dept. of ECE/SJBIT

128
9
128
3
128
9
128
27
128
9
128
3
128
9
128
3
128
3
128
2
128
3
128
3
128
9
128
9
128

Page 18

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

A second order Markoff source

Model shown is an example of a source where the probability of occurrence of a symbol
depends not only on the particular symbol in question, but also on the two symbols proceeding it.
A
2
P1 (1)
1
1/
/8
18
1
2
A
7
3
(AA)
(AA)
/4
P2 (1)
18
7
7
B 3
/8 B
/4
/8
7
P3 (1)
B
(BB)
18
B
2
1
(AB) 4
3 /8
P4 (1)
1
/4
18
B

No. of states: n (M) ;

2
4M
M=2
m = No. of symbols for which the residual influence lasts
(duration of 2 symbols)
or
M = No. of letters / symbols in the alphabet.
Say the system in the state 3 at the beginning of the symbols emitted by the source were BA.
Similar comment applies for other states.
1.6 Entropy and Information Rate of Markoff Sources

Definition of the entropy of the source

Assume that, the probability of being in state i at he beginning of the first symbol interval is
the same as the probability of being in state i at the beginning of the second symbol interval, and so
on.
The probability of going from state i to j also doesnt depend on time, Entropy of state i is
defined as the average information content of the symbols emitted from the i-th state.
n

Hi pij log2

bits / symbol

------ (1)

ij

j 1

Entropy of the source is defined as the average of the entropy of each state.
n

i.e. H = E(Hi) =

pi

j 1

Hi

------ (2)

Where,
Pi = the proby that the source is in state i'.
Using eqn (1), eqn. (2) becomes,

Dept. of ECE/SJBIT

Page 19

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

H = pi

n
p

10EC55

log 1

ij

bits / symbol

------ (3)

p
i 1

ij

j 1

Average information rate for the source is defined as

R = rs . H bits/sec
Where, r s is the number of state transitions per second or the symbol rate of the source.
The above concepts can be illustrated with an example
Illustrative Example:
1. Consider an information source modeled by a discrete stationary Mark off random process shown
in the figure. Find the source entropy H and the average information content per symbol in messages
containing one, two and three symbols.

C
C

A 1
/4

2 B3

p1 =

/4

P2 =

The source emits one of three symbols A, B and C.

A tree diagram can be drawn as illustrated in the previous session to understand the various
symbol sequences and their probabilities.

A
A

C
B

A
C

Dept. of ECE/SJBIT

C
B

1 AAA

2 AAC
1 ACC

/4

2 ACB

1 CCA

2 CCC
1 CBC

/4

2 CBB
1 CAA

2 CAC
1 CCC

B
A

/4

2 CCB

1 BCA

2 BCC
1 BBC

/4

2 BBB

Page 20

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

As per the outcome of the previous session we have

Messages of Length (1)
3
A

8
B

3
8
1
4

Messages of Length (2)

9
AA

AC

CB

CC

BB

BC

CA

32
3
32
3
32
2
32
9
32
3
32
3
32

Messages of Length (3)

27
AAA

AAC

ACC

ACB

BBB

BBC

BCC

BCA

CCA

CCB

CCC

CBC

CAC

CBB

CAA

Dept. of ECE/SJBIT

128
9
128
3
128
9
128
27
128
9
128
3
128
9
128
3
128
3
128
2
128
3
128
3
128
9
128
9
128

Page 21

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

By definition Hi is given by
n

H p

log

ij

ij

j1

Put i = 1,
n2

Hi

1j

log

1j

j 1

log 1 p

11

log 1

12
11

12

Substituting the values we get,

H1

1
1
log 2
3 / 4 4
1/ 4
4 1

4
3

= log

log2

4
3 4
H1 = 0.8113

log2 4

Similarly H2 = 1 log 4 + 3 log 4 = 0.8113

4
4
3
By definition, the source entropy is given by,
n

i 1

i 1

H pi Hi pi Hi

= 1 (0.8113) + 1 (0.8113)
2
2
= (0.8113) bits / symbol
To calculate the average information content per symbol in messages containing two symbols.

How many messages of length (2) are present? And what is the information content of these
messages?

There are seven such messages and their information content is:
1
= log
(BB
)
I (AA) = I (BB) = log ( AA)
1
i.e., I (AA) = I (BB) =
log

= 1.83 bits

(9 /
32)

Similarly calculate for

other messages and
verify that they are
=

I (BB) = I (AC) =
Dept. of ECE/SJBIT

log

3
.
4
1
5
Page 22

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

b
i
t
s
I (CB) = I (CA) =

Dept. of ECE/SJBIT

(3 /
32)

Page 23

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

1 = 4 bits
P(2 / 32)

I (CC) = log =

10EC55

Computation of the average information content of these messages.

Thus, we have
7

H(two)= Pi

1 bits / sym.
Pi

log
i1

Pi . Ii

i1

Where Ii = the Is calculated above for different messages of length two

Substituting the values we get,

H
(two)

(1.83)

x (3.415)

32
32
32
3 x (3.415) 9 x (1.83)
32
32

(3.415)

2
32

(4)

3
32

x (3.415)

2.56 bits
Computation of the average information content per symbol in messages containing two
symbols using the relation.
(two)

GN = Average information content of the messages of length

Number of symbols in the message

Here, N = 2

GN = Average information content of the messages of length

(2) 2

H
( two )

2.56
2

1.28 bits / symbol

G 2 1.28
Similarly compute other Gs of interest for the problem under discussion viz G 1 &
G3. You get them as
G1 = 1.5612 bits / symbol
And G3 = 1.0970 bits / symbol

from the values of Gs calculated

We note that,

Dept. of ECE/SJBIT

Page 24

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

G1 > G2 > G3 > H

Statement
It can be stated that the average information per symbol in the message reduces as the length of
the message increases.
The generalized form of the above statement
If P (mi) is the probability of a sequence mi of N symbols form the source with the average
information content per symbol in the messages of N symbols defined by
P(mi ) log P(mi )
i

GN =

Where the sum is over all sequences mi containing N symbols, then GN is a monotonic decreasing
function of N and in the limiting case it becomes.
Lim GN = H bits / symbol
N

Recall

H = entropy of the source

The above example illustrates the basic concept that the average information content per symbol
from a source emitting dependent sequence decreases as the message length increases.

It can also be stated as,

Alternatively, it tells us that the average number of bits per symbol needed to represent a message
decreases as the message length increases.
Problems:
Example 1
The state diagram of the stationary Mark off source is shown below
Find (i) the entropy of each state
(ii) The entropy of the source
(iii) G1, G2 and verify that G1 G2 H the entropy of the source.
C

P(state1) = P(state2) =

P(state3) = 1/3

C
A

Dept. of ECE/SJBIT

Page 25

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

For

the

Mark

10EC55

off

source

S
L

L 1

p1 =

shown,

S
2

S
R

calculate

3 R

the

information

rate.

/2

P2 =

P3 =
Solution:

By definition, the average information rate for the source is given by

R = rs . H bits/sec

------ (1)

Where, rs is the symbol rate of the source

And H is the entropy of the source.
To compute H
Calculate the entropy of each state using,

Hi

log 1 bits / sym

iJ

----- (2)

ij

j1

For this example,

3

Hi

log

ij

i 1, 2, 3

------ (3)

ij

j 1

Put i = 1
H
i

log p1 j

1j

j1

= - p11 log p11 p 12 log p12 p 13 log p13

Substituting the values, we get
1
H1 = -

x log

1
2

1
2

log

1
2

-0

= + 1 log (2) + 1 log (2)

2
2
H1 = 1 bit / symbol
Put i = 2, in eqn. (2) we get,
3

H2 = - p 2 j

log p 2 j

j1

i.e., H2 = - p21 log p21 p22 log p22 p23 log p23
Substituting the values given we get,

Dept. of ECE/SJBIT

Page 26

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

H2

=-

=+

log

log 4 +

10EC55

1
2

log 2 +

log

4
2
4
1
1
log 2 + + log 4
22

1
4

log

1
4

log 4

H2 = 1.5 bits/symbol

Similarly calculate H3 and it will be

H3 = 1 bit / symbol
With Hi computed you can now compute H, the source entropy, using.

H=

Pi H i
i1

= p1 H1 + p2 H2 + p3 H3
Substituting the values we get,
H=

x1+

4
2
1 1.5 1
+

1
4

x1

4 2 4
1 1.5 2.5
2

x 1.5 +

= 1.25 bits / symbol

H = 1.25 bits/symbol
Now, using equation (1) we have
Source information rate = R = rs 1.25
Taking r s as one per second we get
R = 1 x 1.25 = 1.25 bits / sec

Dept. of ECE/SJBIT

Page 27

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Review questions:
(1) Explain the terms (i) Self information (ii) Average information (iii) Mutual Information.
(2) Discuss the reason for using logarithmic measure for measuring the amount of information.
(3) Explain the concept of amount of information associated with message. Also explain what
infinite information is and zero information.
(4) A binary source emitting an independent sequence of 0s and 1s with pro babilities p and (1p) respectively. Plot the entropy of the source.
(5) Explain the concept of information, average information, information rate and redundancy as
referred to information transmission.
(6) Let X represents the outcome of a single roll of a fair dice. What is the entropy of X?
(7) A code is composed of dots and dashes. Assume that the dash is 3 times as long as the dot and
has one-third the probability of occurrence. (i) Calculate the information in dot and that in a
dash; (ii) Calculate the average information in dot-dash code; and (iii) Assume that a dot lasts
for 10 ms and this same time interval is allowed between symbols. Calculate the average rate
of information transmission.
(8) What do you understand by the term extension of a discrete memory less source? Show that
the entropy of the nth extension of a DMS is n times the entropy of the original source.
(9) A card is drawn from a deck of playing cards. A) You are informed that the card you draw is
spade. How much information did you receive in bits? B) How much information did you
receive if you are told that the card you drew is an ace? C) How much information did you
receive if you are told that the card you drew is an ace of spades? Is the information content of
the message ace of spades the sum of the information contents of the messages spade and
ace?
(10) A block and white TV picture consists of 525 lines of picture information. Assume that each
consists of 525 picture elements and that each element can have 256 brightness levels.
Pictures are repeated the rate of 30/sec. Calculate the average rate of information conveyed by
a TV set to a viewer.
(11) A zero memory source has a source alphabet S= {S1, S2, S3} with P= {1/2, 1/4, 1/4}. Find
the entropy of the source. Also determine the entropy of its second extension and verify that H
2
(S ) = 2H(S).
(12) Show that the entropy is maximum when source transmits symbols with equal probability.
Plot the entropy of this source versus p (0<p<1).
(13) The output of an information source consists OF 128 symbols, 16 of which occurs with
probability of 1/32 and remaining 112 occur with a probability of 1/224. The source emits
1000 symbols/sec. assuming that the symbols are chosen independently; find the rate of
information of the source.

3 BB
A
C
Dept. of ECE/SJBIT

1 AA
Page 28

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Unit - 2:

SOURCE CODING

Syllabus:
Encoding of the source output, Shannons encoding algorithm. Communication Channels,
Discrete communication channels, Continuous channels.

Text Books:

Digital and analog communication systems, K. Sam Shanmugam, John Wiley,

1996.

Reference Books:
Digital Communications - Glover and Grant; Pearson Ed. 2nd Ed 2008

Dept. of ECE/SJBIT

Page 29

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

Unit - 2:

10EC55

SOURCE CODING

2.1 Encoding of the Source Output:

Need for encoding

N

Suppose that, M messages = 2 , which are equally likely to occur. Then recall that average
information per messages interval in H = N.
Say further that each message is coded into N bits,
Average information carried by an individual bit is =

H
N

1 bit

If the messages are not equally likely, then H will be les s than N and each bit will carry
less than one bit of information.

Is it possible to improve the situation?

Yes, by using a code in which not all messages are encoded into the same number of bits. The
more likely a message is, the fewer the number of bits that should be used in its code word.

Source encoding
Process by which the output of an information source is converted into a binary sequence.

Symbol
sequence
emitted
by
the Input
information source

Where, GN =

Source
Encoder

Output

: a binary sequence

If the encoder operates on blocks of N symbols, t he bit rate of the

encoder is given as

Produces an average bit rate of GN bits / symbol

p(mi ) log p(mi )

Ni

p(mi ) = Probability of sequence mi of N symbols from the source,

Sum is over all sequences m i containing N symbols.
GN in a monotonic decreasing function of N and
Lim

N GN = H bits / symbol
Performance measuring factor for the encoder
Coding efficiency: c
Definition of c =

Source inf ormation rate

Average output bit rate of the encoder

c = H(S)
^
HN

Dept. of ECE/SJBIT

Page 30

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

2.2 Shannons Encoding Algorithm:

Formulation of the design of the source encoder

Can be formulated as follows:

One of q
possible
messages
INPUT

SourceOUTPUT
encoder

A message

A unique binary
code word c i of
length n i bits for
the message m i

N-symbols Replaces the input message

symbols by a sequence of
binary digits
: m 1, m2, ..m i, .., m
Probs. of messages : p1, p2, ....p i, ..., p
q messages

q
q

ni : an integer

The objective of the designer

^

To find n i and c i for i = 1, 2, ...., q such that the average number of bits per symbol H N
used in the coding scheme is as close to GN as possible.
^

Where, H N =

n p

i 1

and GN =

pi

log

i 1
pi
i.e., the objective is to have

HN

GN as closely as possible

The algorithm proposed by Shannon and Fano

Step 1: Messages for a given block size (N) m1, m2, ....... mq are to be arranged in decreasing order of
probability.
Step 2: The number of n i (an integer) assigned to message m i is bounded by
log2 1 n i 1 log 2 1
pi
pi
Step 3: The code word is generated from the binary fraction expansion of Fi defined as

Dept. of ECE/SJBIT

Page 31

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

i1

Fi = p k , with F1 taken to be zero.

k 1

Step 4: Choose n i bits in the expansion of step (3)

Say, i = 2, then if ni as per step (2) is = 3 and
If the Fi as per stop (3) is 0.0011011
Then step (4) says that the code word is: 001 for message (2)
With similar comments for other messages of the source.
The codeword for the message mi is the binary fraction expansion of Fi upto ni bits.
i.e., Ci = (Fi)binary, ni bits

Step 5: Design of the encoder can be completed by repeating the above steps for all the messages of
block length chosen.

Illustrative Example
Design of source encoder for the information source given,

A 1
/4

C
C

2 B3

p1 =

/4

P2 =

Compare the average output bit rate and efficiency of the coder for N = 1, 2 & 3.
Solution:
The value of N is to be specified.
Case I: Say N = 3 Block size
Step 1: Write the tree diagram and get the symbol sequence of length = 3.
Tree diagram for illustrative example (1) of session (3)

Dept. of ECE/SJBIT

Page 32

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

A
A
A

C
2

1
2

2 ACB

1 CCA

C
C

2 CCC
1 CBC

/4
B
A

2 CBB
1 CAA

C
C

2 CAC
1 CCC

/4

2 CCB

1 BCA

C
C

2 BCC
1 BBC

2 AAC
1 ACC

/4

C
C

1 AAA

/4
B
2 BBB
From the previous session we know that the source emits fifteen (15) distinct three symbol messages.
They are listed below:
Messages

AAA AAC ACC ACB

CAA

Probability

27

128128

BBB BBC BCC BCA

CCA CCB

CCC CBC CAC

CBB

27

128

128

128

128

128

128

128

128

128

128

128

128

128

Step 2: Arrange the messages m i in decreasing order of probability.

Messages
mi

AAA BBB CAA CBB

CCC

Probability

27

pi

27 9

128128

128

9
128

BCA BBC AAC ACB CBC CAC

BCC ACC

128

128

128

128

128

128

128

128

128

128

CCB CCA

Step 3: Compute the number of bits to be assigned to a message m i using.

pi

Log2 1 n i 1 log 2 1 ; i = 1, 2, . 15
pi
Say i = 1, then bound on n i is

2
7

Dept. of ECE/SJBIT

log

128

n1 1 log

128
27

Page 33

www.rejinpaul.com

2
128

www.rejinpaul.com

Information Theory and Coding

10EC55

i.e., 2.245 < n1 < 3.245

Recall n i has to be an integer
n1 can be taken

as, n1 = 3
Step 4: Generate the codeword using the binary fraction expansion of Fi defined as
i1

Fi = p k ; with F1 = 0
k 1

Say i = 2, i.e., the second message, then calculate n2 you should get it as 3 bits.
21

Next, calculate F2 =

p k

k 1

k 1

27 . Get the binary fraction expansion of 27 . You

128
128

get it as : 0.0011011
Step 5: Since ni = 3, truncate this exploration to 3 bits.
The codeword is: 001

Step 6: Repeat the above steps and complete the design of the encoder for other messages listed above.
The following table may be constructed

Message
pi
mi
AAA
27
BBB

128
27
128
9

Binary expansion of
Fi
.00000

Code word
ci
000

27/128

.001101

001

Fi

ni

CAA

128
9

54/128

0110110

0110

CBB

128
9

63/128

0111111

0111

BCA

128
9

72/128

.1001100

1001

BBC

128
9

81/128

1010001

1010

90/128

1011010

1011

99/128

1100011

1100

AAC
ACB

128
9
128
3
128

Dept. of ECE/SJBIT

Page 34

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

CBC

3
128
3

10EC55

108/128

110110

110110

CAC

128
3

111/128

1101111

110111

CCB

128
3

114/128

1110010

111001

CCA

128
3

117/128

1110101

111010

120/128

1111000

111100

123/128

1111011

111101

126/128

1111110

111111

128

BCC
ACC

2
128

CCC

the average number of bits per symbol used by the encoder

Average number of bits = n i pi

Substituting the values from the table we
get, Average Number of bits = 3.89
^

Average Number of bits per symbol = H N

nn
i
i

Here N = 3,
^

H3 3.89 = 1.3 bits / symbol

3
State entropy is given by
n

Hi = pij

log

1
p

bits / symbol
ij

j 1

Here number of states the source can be in are two

i.e., n = 2
1

Hi = pij

log

p
ij

j 1

Say i = 1, then entropy of state (1) is

1

Hi = pij
j 1

log

p11 log

p
1j

p12

11

1
log

p
12

Substituting the values known we get,

Dept. of ECE/SJBIT

Page 35

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H1 = 4 x log 3 / 4 4 log 1/ 4
=

x log

4 1
3

log 4

H1 = 0.8113

Similarly we can compute, H2 as

2

H2 = p21 log

p21 p22 log

21

j 1

1
22

Substituting we get,
H2 =

x log

4
=

1
4

x log 4

1/ 4
3
4

log

log

1
3/4

4
3

H2 = 0.8113
Entropy of the source by definition is
n

H = pi Hi ;
j 1

th

Pi = Probability that the source is in the i state.

2

H = pi H i ; = p1H1 + p2H2
i 1

Substituting the values, we get,

H = x 0.8113 + x 0.8113 = 0.8113
H = 0.8113 bits / sym.

Dept. of ECE/SJBIT

Page 36

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

What is the efficiency of the encoder?

By definition we have
c = H x 100 H x 100 0.8113 x 100 62.4%
^
^
1.3
H
H
2

c for N = 3

is, 62.4%

Case II
Say N = 2
The number of messages of length two and their probabilities (obtai ned from the tree diagram)
can be listed as shown in the table.
Given below
N=2
Message pi

ni

ci

AA

9/32

00

BB

9/32

01

AC

3/32

1001

CB

3/32

1010

BC

3/32

1100

CA

3/32

1101

CC

2/32

1111

Calculate H N and verify that it is 1.44 bits / sym.

Encoder efficiency for this case is

H
c = ^ x100 H
N

Substituting the values we get,

c = 56.34%
Case III:

N=1

Proceeding on the same lines you would see that

N=1
Message

pi

ni

ci

3/8

00

3/8

01

1/4

11

Dept. of ECE/SJBIT

Page 37

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H1 = 2 bits / symbol and

c = 40.56%
Conclusion for the above example
We note that the average output bit rate of the encoder H N

decreases as N increases and

hence the efficiency of the encoder increases as N increases.

Operation of the Source Encoder Designed:
I. Consider a symbol string ACBBCAAACBBB at the encoder input. If the encoder uses a
block size of 3, find the output of the encoder.

C
C

A 1

/4

2 B3

p1 =

/4

SOURCE
ENCODER

OUTPUT

P2 =

INFORMN. SOURCE
Recall from the outcome of session (5) that for the source given possible three symbol
sequences and their corresponding code words are given by
Message
mi

ni

Codeword
ci

AAA

000

BBB

001

CAA

0110

CBB

0111

BCA

1001

BBC

1010

ACB

1100

CBC

110110

CAC

110111

CCB

111001

CCA

111010

BCC

111100

ACC

111101

AAC

Dept. of ECE/SJBIT

Determination of the
code words and their
size as illustrated in
the previous session

Page 38

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

CCC

10EC55

111111

Output of the encoder can be obtained by replacing successive groups of three input symbols by
the code words shown in the table. Input symbol string is
ACB BCA AAC BBB
123

1100

123

1001

123

1011

011

Encoded version of the symbol string

II. If the encoder operates on two symbols at a time what is the output of the encoder for the
same symbol string?
Again recall from the previous session that for the source given, different two-symbol sequences
and their encoded bits are given by
N=2
ci

Message

No. of bits

mi

ni

AA

00

BB

01

AC

1001

CB

1010

BC

1100

CA

1101

CC

1111

For this case, the symbol string will be encoded as

AC BB CA AA CB BB
{{{{{{

100101110100 1010 01

Encoded message

DECODING

How is decoding accomplished?

By starting at the left-most bit and making groups of bits with the codewords listed in the

table.
Case I: N = 3
i) Take the first 3 bit group viz 110 why?
ii) Check for a matching word in the table.
iii) If no match is obtained, then try the first 4-bit group 1100 and again check for the matching
word.
iv) On matching decode the group.

Dept. of ECE/SJBIT

Page 39

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

NOTE: For this example, step (ii) is not satisfied and with step (iii) a match is found and the decoding
results in ACB.
Repeat this procedure beginning with the fifth bit to decode the remaining symbol groups.
Symbol string would be ACB BCA AAC BCA

Conclusion from the above example with regard to decoding

The effect of bit errors in transmission

It is clear that the decoding can be done easily by knowing the codeword lengths apriori if no
errors occur in the bit string in the transmission process.
Leads to serious decoding problems.

Example: For the case of N = 3, if the bit string, 1100100110111001 was received at the decoder
input with one bit error as
1101100110111001
What then is the decoded message?
Solution: Received bit string is
110110 0110 11101
Error bit
CBCCAACCB

----- (1)

For the errorless bit string you have already seen that the decoded symbol string is
ACB BCA AAC BCA

----- (2)

(1) and (2) reveal the decoding problem with bit error.
Illustrative examples on source encoding
1. A source emits independent sequences of symbols from a source alphabet containing five
symbols with probabilities 0.4, 0.2, 0.2, 0.1 and 0.1.
i) Compute the entropy of the source
ii) Design a source encoder with a block size of two.
Solution: Source alphabet = (s1, s2, s3, s4, s5)
Probs. of symbols = p1, p2, p3, p4, p5
= 0.4, 0.2, 0.2, 0.1, 0.1
5

(i)

Entropy of the source = H = pi log pi bits / symbol

i1

Substituting we get,
H = - [p1 log p1 + p2 log p2 + p3 log p3 + p4 log p4 + p5 log p5 ]
= - [0.4 log 0.4 + 0.2 log 0.2 + 0.2 log 0.2 + 0.1 log 0.1 + 0.1 log 0.1]

Dept. of ECE/SJBIT

Page 40

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H = 2.12 bits/symbol
(ii)

Some encoder with N = 2

Different two symbol sequences for the source are:

(s1s1)

AA

( ) BB

( ) CC

( ) DD

( ) EE

(s1s2)

AB

( ) BC

( ) CD

( ) DE

( ) ED

(s1s3)

AC

( ) BD

( ) CE

( ) DC

( ) EC A total of 25 messages

(s1s4)

AD

( ) BE

( ) CB

( ) DB

( ) EB

(s1s5)

AE

( ) BA

( ) CA

( ) DA

( ) EA

Arrange the messages in decreasing order of probability and determine the number of bits n i as
explained.
Proby.
No. of bits
Messages
pi
ni
AA

0.16

AB
AC
BC
BA
CA
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...

0.08
0.08
0.08
0.08
0.08
0.04
0.04
0.04
0.04
0.04
0.04
0.04
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.01
0.01
0.01
0.01

Calculate H1 =
^

Substituting, H1 = 2.36 bits/symbol

2. A technique used in constructing a source encoder consists of arranging the messages in
decreasing order of probability and dividing the message into two almost equally probable

Dept. of ECE/SJBIT

Page 41

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

groups. The messages in the first group are given the bit O and the messages in the second
group are given the bit 1. The procedure is now applied again for each group separately, and
continued until no further division is possible. Using this algorithm, find the code words for six
messages occurring with probabilities, 1/24, 1/12, 1/24, 1/6, 1/3, 1/3
Solution: (1) Arrange in decreasing order of probability
m5

1/3

m6

1/3

1 division

m4
m2

1/6
1/12

1
1

0
1

m1

1/24

m3

1/24

st

nd

division
division

rd

th

4 division

Code words are

m1 = 1110
m2 = 110
m3 = 1111
m4 = 10
m5 = 00
m6 = 01

Example (3)
a) For the source shown, design a source encoding scheme using block size of two
symbols and variable length code words
^

b) Calculate H 2 used by the encoder

c) If the source is emitting symbols at a rate of 1000 symbols per second, compute the
output bit rate of the encoder.

L 1

S
2

p1 =

Dept. of ECE/SJBIT

S
L

S
R

3 R

/2

P2 =

P3 =

Page 42

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Solution (a)
1. The tree diagram for the source is

1 LL

(1/16)

2 LS

(1/16)

(1/32)
(1/16)
(1/32)

1 LL
2 LS
1 SL

(1/16)

2 SS
3
2 SR
3 RS

1 SL
2 SS
3 SR

(1/32)
(1/16)

2 RS

(1/16)

3 RR (1/16)

SL
2
SS
3
SR

C
3

(1/16)

LL

(1/16)

LS

(1/8)

SL

(1/16)

SS

(1/8)

SR

Different
Messages
of Length
Two

RS
RR

(1/32)

2. Note, there are seven messages of length (2). They are SS, LL, LS, SL, SR, RS & RR.
3. Compute the message probabilities and arrange in descending order.
4. Compute ni, Fi. Fi (in binary) and ci as explained earlier and tabulate the results, with usual
notations.
Message

pi

ni

Fi

Fi (binary)

ci

SS

1/4

.0000

00

LL

1/8

1/4

.0100

010

LS

1/8

3/8

.0110

011

SL

1/8

4/8

.1000

100

SR

1/8

5/8

.1010

101

RS

1/8

6/8

.1100

110

RR

1/8

7/8

.1110

111

mi

Dept. of ECE/SJBIT

Page 43

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

1 pi log 2
2 i 1

G2 =

pi = 1.375 bits/symbol

7
^

(b) H 2

= 1 pi ni
2 i 1

= 1.375 bits/symbol

1
Recall, H N GN + N ; Here N = 2
^

H 2 G2 +

2
(c) Rate = 1375 bits/sec.
2.3 SOURCE ENCODER DESIGN AND COMMUNICATION CHANNELS

The schematic of a practical communication system is shown.

Data Communication Channel (Discrete)
Coding Channel (Discrete)
Modulation Channel (Analog)

Channel
Encoder

Transmitter

Channel
Encoder

Electrical
Communication
channel
OR
Transmissi
on medium

Physical channel

Noise

g
Demodulator

Channel
Decoder

Receiver

Fig. 1: BINARY COMMN. CHANNEL CHARACTERISATION

Communication Channel

Communication Channel carries different meanings and characterizations depending on its

terminal points and functionality.
(i)

Portion between points c & g:

Referred to as coding channel
Accepts a sequence of symbols at its input and produces a sequence of symbols at its
output.
Completely characterized by a set of transition probabilities pij. These probabilities will
depend on the parameters of (1) The modulator, (2) Transmiss ion media, (3) Noise, and
(4) Demodulator

Dept. of ECE/SJBIT

Page 44

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

A discrete channel
(ii)

Portion between points d and f:

Provides electrical connection between the source and the destination.
The input to and the output of this channel are analog electrical waveforms.
Referred to as continuous or modulation channel or simply analog
channel. Are subject to several varieties of impairments
Due to amplitude and frequency response variations of the channel within the
passband.
Due to variation of channel characteristics with time.
Non-linearities in the channel.
Channel can also corrupt the signal statistically due to various types of additive and
multiplicative noise.

2.4 Mathematical Model for Discrete Communication Channel:

Channel between points c & g of Fig. (1)

The input to the channel?

A symbol belonging to an alphabet of M symbols in the general case is the input to the channel.
he output of the channel
A symbol belonging to the same alphabet of M input symbols is the output of the channel.
Is the output symbol in a symbol interval same as the input symbol during the same symbol
interval?
The discrete channel is completely modeled by a set of probabilities

pit

th

Probability that the input to the channel is the i symbol of the alphabet.
(i = 1, 2, . M)
and

p ij

th

th

Probability that the i symbol is received as the j symbol of the alphabet at the output of
the channel.

Discrete M-ary channel

If a channel is designed to transmit and receive one of M possibl e symbols, it is called a
discrete M-ary channel.
discrete binary channel and the statistical model of a binary channel

Dept. of ECE/SJBIT

Page 45

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Shown in Fig. (2).

O

P00

O
P10

Transmitted

Received

digit X

digit X

P01
p11

pij = p(Y = j / X=i)

pot = p(X = o); p1t P(X = 1)

por = p(Y = o); p1r P(Y = 1)

poo + po1 = 1 ; p11 + p10 = 1

Fig. (2)

Its features
X & Y: random variables binary valued
Input nodes are connected to the output nodes by four paths.
(i) Path on top of graph : Represents an input O appearing correctly
as O as the channel output.
(ii)

Path at bottom of graph

(iii) Diogonal path from 0 to 1 : Represents an input bit O appearing

incorrectly as 1 at the channel output
(due to noise)
(iv) Diagonal path from 1 to 0 :

Similar comments

Errors occur in a random fashion and the occurrence of errors can be statistically modelled
by assigning probabilities to the paths shown in figure (2).

A memory less channel:

If the occurrence of an error during a bit interval does not affect the behaviour of the system
during other bit intervals.
Probability of an error can be evaluated as
p(error) = Pe = P (X Y) = P (X = 0, Y = 1) + P (X = 1, Y = 0)
Pe = P (X = 0) . P (Y = 1 / X = 0) + P (X = 1), P (Y = 0 / X= 1)
Can also be written as,
t

Pe = po p01 + p1 p10

------ (1)

We also have from the model

Dept. of ECE/SJBIT

Page 46

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

por pot p00 p1t . p10 , and

----- (2)

p1r pot p01 p1t p11

Binary symmetric channel (BSC)

If, p00 = p11 = p (say), then the channel is called a BSC.

Parameters needed to characterize a BSC

Model of an M-ary DMC.

p11

p12

pit = p(X = i)
prj = p(Y = j)

p = p(Y = j / X = i)
2

INPUT X

ij

OUTPUT Y
j

pij

piM
M

Fig. (3)
This can be analysed on the same lines presented above for a binary channel.
M

r
pj

pi pij
t

----- (3)

i 1

The p(error) for the M-ary channel

Generalising equation (1) above, we have
P(error) Pe pi

i 1

ij

----- (4)

j 1
ji

In a DMC how many statistical processes are involved and which are they?
Two, (i) Input to the channel and
(ii) Noise

Dept. of ECE/SJBIT

Page 47

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Definition of the different entropies for the DMC.

i)

Entropy of INPUT X:
M

H(X).

t bits / symbol

HX pit log pi

----- (5)

i1

ii)

Entropy of OUTPUT Y: H(Y)

M

r bits / symbol

HY pir log pi
i 1

iii)

Conditional entropy: H(X/Y)

M

HX / Y

P (X i, Y j) log p (X i / Y j) bits/symbol

i1

iv)

- (7)

j1

Joint entropy: H(X,Y)

M

HX, Y

P (X i, Y j) log p (X i, Y j) bits/symbol

i1

v)

----- (6)

- (8)

i1

Conditional entropy: H(Y/X)

M

H (Y/X)

P (X i, Y j) log p (Y j / X i) bits/symbol

i1

- (9)

i1

Representation of the conditional entropy

H(X/Y) represents how uncertain we are of the channel input x, on the average, when we know
the channel output Y.

Similar comments apply to H(Y/X)

vi)

Joint Entropy H(X, Y)

H(X) H(Y/X) H(Y) H(X/Y) - (10)

ENTROPIES PERTAINING TO DMC

To prove the relation for H(X Y)

By definition, we have,
M

H(XY) =
i

Dept. of ECE/SJBIT

p ( i, j) log p ( i, j)
j

Page 48

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

associated with variable X, white j

H(XY) =

with variable Y.

p ( i) p ( j / i) log p ( i) p ( j / i)

ij

p ( i) p ( j / i) log p ( i)

p ( i) p ( j / i) log p ( i)

ij

Say, i is held constant in the first summation of the first t erm on RHS, then we can write
H(XY) as
H(XY) = p ( i)1 log p ( i)

p ( ij)

log p ( j / i)

H(XY) H(X) H(Y / X)

Hence the proof.

1. For the discrete channel model shown, find, the probability of error.
p

Since the channel is symmetric,

p(1, 0) = p(0, 1) = (1 - p)
Y

1
Transmitted
digit

Proby. Of error means, situation

when X Y

Received
digit

P(error) = Pe = P(X Y) = P(X = 0, Y = 1) + P (X = 1, Y = 0)

= P(X = 0) . P(Y = 1 / X = 0) + P(X = 1) . P(Y = 0 / X = 1)
Assuming that 0 & 1 are equally likely to occur
P(error) = 1 x (1 p) + 1 (1 p) = 1 - p + 1 - p
2
2 2
2
2 2
P(error) = (1 p)

2. A binary channel has the following noise characteristics:

Dept. of ECE/SJBIT

Page 49

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

P(Y/X)

10EC55

2/3

1/3

1/3

2/3

If the input symbols are transmitted with probabilities & respectively, find H(X), H(Y),
H(XY), H(Y/X).
Solution:
Given = P(X = 0) = and P(Y = 1)
H(X) = p i log pi 3 log
4
i

4 1 log 4 0.811278 bits / symbol

2
3 4

Compute the probability of the output symbols.

Channel model isx1

y1

x2

y2

p(Y = Y1) = p(X = X1, Y = Y1) + p(X = X2, Y = Y1)

----- (1)

To evaluate this construct the p(XY) matrix using.

2 .3
3

. 3

y1

y2

x1

P(XY) = p(X) . p(Y/X) =

1
3

1
4

2
3

----- (2)
2

12

P(Y = Y1) = 1 1 7 -- Sum of first column of matrix (2)

12 12
2
nd
Similarly P(Y2) = 5
sum of 2 column of P(XY)
12
Construct P(X/Y) matrix using

Dept. of ECE/SJBIT

Page 50

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

P(XY) = p(Y) . p(X/Y) i.e., p(X/Y) =

p(XY)

p(Y)

1
12 6
pX 1Y1 2
and so on

7
14 7
pY1
12

X1
p

Y1

6
p(Y/X) =

=?

3
7

-----(3)

H(Y) = pi log 1 7 log 12 5 log 12 0.979868 bits/sym.

pi 12
7
12
5
i
H(XY) =

p(XY) log

1
p(XY)

= 1 log 2 1 log 4 1 log12 1 log 6

6
2
4
12
H(XY) = 1.729573 bits/sym.

H(X/Y) =

p(XY) log

1
p(X / Y)

= 1 log 7 1 log 5 1 log 7 1 log 5

2
6 4
3 12
6
2
H(X/Y) =

p(XY) log

= 1.562

1
p(Y / X)

= 1 log 3 1 log 3 1 log 3 1 log 3

2
2 4
12
6
2
3. The joint probability matrix for a channel is given below. Compute H(X), H(Y), H(XY),
H(X/Y) & H(Y/X)
0.05
P(XY) =

0.2

0.05

0
0

0.1
0

0.1
0.2

0
0.1

0.05

0.05

0.1

Solution:
Row sum of P(XY) gives the row matrix P(X)
P(X) = [0.3, 0.2, 0.3, 0.2]

Dept. of ECE/SJBIT

Page 51

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Columns sum of P(XY) matrix gives the row matrix P(Y)

P(Y) = [0.1, 0.15, 0.5, 0.25]

Get the conditional probability matrix P(Y/X)

1
6
0

1
2

P(Y/X) =

3
1
2
2
5

6
0
2
5

Get the condition probability matrix P(X/Y)

1
2
P(X/Y) =

2
3

1
3

5
1
5
2
5

0
2
5

Now compute the various entropies required using their defining equations.
(i)

H(X) = pX.log
i

2 0.3 log

pX

0.2 log

0.3

0.2

H (X) = 1.9705 bits / symbol

H(Y) pY.log 1

0.1 log 1 0.15 log 1

pY
0.1
0.15
j
1
1
0.5 log
0.25 log
0.5
0.25

(ii)

H (Y) = 1.74273 bits / symbol

(iii)

H(XY) = p(XY) log

4 0.05 log

Dept. of ECE/SJBIT

1
0.05

1
p(XY)
0.1 log

1
0.1

2 0.2

log

1
0.2

Page 52

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H(XY) = 3.12192

H(X/Y) = p(XY) log

(iv)

p(X / Y)

Substituting the values, we get.

H(X/Y) = 4.95 bits / symbol
1
H(Y/X) = p(XY) log
p(Y / X)

(v)

Substituting the values, we get.

H(Y/X) = 1.4001 bits / symbol

4. Consider the channel represented by the statistical model shown. Write the channel matrix
and compute H(Y/X).
1

/3

X
1

Y1

/6
1/3

INPUT

/6

/6

X2

Y2

/3

OUTPUT
Y3

/6

/3

Y4

For the channel write the conditional probability matrix P(Y/X).

P(Y/X) =

x1

y1
1

y2
1

y3
1

y4
1

x26
nd

st

NOTE: 2 row of P(Y/X) is 1 row written in reverse order. If this is the situation, then channel is
called a symmetric one.
1
First row of P(Y/X) . P(X1) =

1
3

1
2

1
3

1
2

1
6

1
2

1
6

Second row of P(Y/X) . P(X2) = 1 x 1 1 x 1 1 x 1 1 x 1

2 6 2 6 6 2 2 3

Dept. of ECE/SJBIT

Page 53

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Recall
P(XY) = p(X), p(Y/X)
P(X1Y1) = p(X1) . p(Y1X1) = 1 . 1
3

P(X1Y2) = 1 , p(X1, Y3) = 1 = (Y1X4) and so on.

3
6
1
1
1
1
6

12

12

12

12

P(X/Y) =
6

1
p(Y / X)

H(Y/X) = p(XY) log

Substituting for various probabilities we get,

H(Y/X) 1 log 3 1 log 3 1 log 6 1 log 6 1 log 6
12
6
6
12
12
1
1
1
log 6
log 3 log 3

12
6
6
=4x

1
1
log 3 4 x
log 6
6
12

= 2 x 1 log 3 1 log 6 = ?
3
3
5. Given joint proby. matrix for a channel compute the various entropies for the input and
output rvs of the channel.
0.2

0.2

0.01
0.02

0.01
0.02

0.01
0

0.04

0.04

0.01

0.06

0.06

0.02

0.2

0.1
P(X . Y) = 0

Solution:
P(X) = row matrix: Sum of each row of P(XY) matrix.
P(X) = [0.4, 0.13, 0.04, 0.15, 0.28]

P(Y) = column sum = [0.34, 0.13, 0.26, 0.27]

1. H(XY) = p(XY) log

Dept. of ECE/SJBIT

1
p(XY)

3.1883 bits/sym.

Page 54

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

2. H(X) =

p(X) log

1
2.0219 bits/sym.
p(X)

3. H(Y) =

p(Y) log

1
1.9271 bits/sym.
p(Y)

Construct the p(X/Y) matrix using, p(XY) = p(Y) p(X/Y)

0.2
0.34
0.1
0.34
or P(X/Y) = p(XY) =
p(Y)

0
0.04
0.34
0

0.2

0.01
0.13

0.36
0.01
0.26

0.02
0.13
0.04
0.13
0.06

0.02
0.26
0.01
0.26
0.02

0.06
0.27
0.2

0.13

0.26

0.27

0.01
0.27
0

4. H(X/Y) = p(XY) log p(X / Y) 1.26118 bits/sym.

Problem:
Construct p(Y/X) matrix and hence compute H(Y/X).
Rate of Information Transmission over a Discrete Channel :

For an M-ary DMC, which is accepting symbols at the rate of rs symbols per second, the
average amount of information per symbol going into the channel is given by the entropy of
the input random variable X.
M

i.e., H(X) = pit log 2 pit

----- (1)

i1

Assumption is that the symbol in the sequence at the input to the channel occur in a statistically
independent fashion.

Average rate at which information is going into the channel is

Din = H(X), rs bits/sec

----- (2)

Is it possible to reconstruct the input symbol sequence with certainty by operating on the
received sequence?

Dept. of ECE/SJBIT

Page 55

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Given two symbols 0 & 1 that are transmitted at a rate of 1000 symbols or bits per second.
With p 0t 1 & p1t 1
2
2
Din at the i/p to the channel = 1000 bits/sec. Assume that the channel is symmetric with the
probability of errorless transmission p equal to 0.95.

Rate of transmission of information:

Recall H(X/Y) is a measure of how uncertain we are of the input X given output Y.
What do you mean by an ideal errorless channel?
H(X/Y) may be used to represent the amount of information lost in the channel.
Define the average rate of information transmitted over a channel (Dt).
Amount of information
Dt

going into the channel

Amount of

information lost

Symbolically it is,
Dt = H (H) H(X / Y) .rs bits/sec.
When the channel is very noisy so that output is statistically independent of the input, H(X/Y) = H(X)
and hence all the information going into the channel is lost and no information is transmitted over the
channel.
DISCRETE CHANNELS:
1. A binary symmetric channel is shown in figure. Find the rate of information transmission
over this channel when p = 0.9, 0.8 & 0.6. Assume that the symbol (or bit) rate is
1000/second.
p
1p
Input
X

1p

p(X = 0) = p(X = 1) =
Output
Y

1
2

Dept. of ECE/SJBIT

Page 56

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Example of a BSC
Solution:
1
log 2 2 log 2 2 1 bit / sym.
2
2
Din rs H(X) 1000 bit / sec

H(X) =

By definition we have,
Dt = [H(X) H(X/Y)]

Where, H(X/Y) = pXY log pX / Y . rs

i

Where X & Y can take values.

X

H(X/Y) = - P(X = 0, Y = 0) log P (X = 0 / Y = 0)

= - P(X = 0, Y = 1) log P (X = 0 / Y = 1)
= - P(X = 1, Y = 0) log P (X = 1 / Y = 0)
= - P(X = 1, Y = 1) log P (X = 1 / Y = 1)
The conditional probability p(X/Y) is to be calculated for all the possible values that X & Y can
take.
Say X = 0, Y = 0, then

P(X = 0 / Y = 0) = p(Y 0 / X 0) p(X

0) p(Y 0)

Where
Y0
p(Y = 0) = p(Y = 0 / X = 0) . p(X = 0) + p (X = 1) . p

X 1

= p . 1 1 (1 p)
2 2

Dept. of ECE/SJBIT

Page 57

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

p(Y = 0) =

1
2

p(X = 0 /Y = 0) = p

Similarly we can calculate

p(X = 1 / Y = 0) = 1 p
p(X = 1 / Y = 1) = p
p(X = 0 / Y = 1) = 1 p

H (X / Y) = -

1
1
p log 2 p (1 p) log 2 (1 p)
2
2
1
2

p log 2 p

(1 p) log (1 p)

= - p log 2 p (1 p) log 2 (1 p)
Dt rate of inforn. transmission over the channel is = [H(X) H (X/Y)] . r s

with, p = 0.9, Dt = 531 bits/sec.

p = 0.8, Dt = 278 bits/sec.
p = 0.6, Dt = 29 bits/sec.

What does the quantity (1 p) represent?

What do you understand from the above example?

2. A discrete channel has 4 inputs and 4 outputs. The input probabilities are P, Q, Q, and P.
The conditional probabilities between the output and input are.
Y

P(y/x)
X

0
1
2
3

0
1

p
(1p)

(1p)
(p)

Write the channel model.

Solution: The channel model can be deduced as shown below:
Given,

P(X = 0) = P

P(X = 1) = Q
P(X = 2) = Q

Dept. of ECE/SJBIT

Page 58

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

P(X = 3) = P
Off course it is true that: P + Q + Q + P = 1
i.e., 2P + 2Q = 1
Channel model is

1
1p=q

Input
X

(1 p) = q
2

Output
Y

What is H(X) for this?

H(X) = - [2P log P + 2Q log Q]
What is H(X/Y)?
H(X/Y)

= - 2Q [p log p + q log q]

=2Q.
1. A source delivers the binary digits 0 and 1 with equal probability into a noisy channel at a
rate of 1000 digits / second. Owing to noise on the channel the probability of receiving a
transmitted 0 as a 1 is 1/16, while the probabi lity of transmitting a 1 and receiving a 0
is 1/32. Determine the rate at which information is received.
Solution:
Rate of reception of information is given by
1

R = H (X) - H (X/Y) bits / sec

-----(1)

Where, H(X) = p(i) log p(i) bits / sym.

i

H(X/Y) =
i

H(X) =

p(ij) log p(i / j) bits / sym.

-----(2)

1
2

Dept. of ECE/SJBIT

log

1
2

1
2

log

1
2

1 bit / sym.

Page 59

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Channel model or flow graph is

15/16

Index i' refers to the I/P of the

channel and index j referes to
the output (Rx)

0
1/32

Input

Output

1/16
1

31/32

Probability of transmitting a symbol (i) given that a symbol 0 was received was received is denoted
as p(i/j).

i0
What do yo u mean by t he pro bability p

j0

How would you compute p(0/0)

Recall the probability of a joint event AB p(AB)
P(AB) = p(A) p(B/A) = p(B) p(A/B)
i.e., p(ij) = p(i) p(j/i) = p(j) p(i/j)
from which we have,
p(i/j) = p(i) p( j / i)
p( j)

-----(3)

What are the different combinations of i & j in the present case?

Say i = 0 and j = 0, then equation (3) is p(0/0)

p(0) p(0 / 0)
j 0)

p(

What do you mean by p(j = 0)? And how to compute this quantity?
Substituting, find p(0/0)
p(0/0) = p(0 / 0) p(0 / 0)
p(0)

Thus, we have,

1 x 15
= 2 16
31
64

= 30
31

= 0.967

p(0/0) = 0.967

Similarly calculate and check the following.

p

1
0

, p

31

Dept. of ECE/SJBIT

31 0
;p

33 1

22
33

Page 60

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Calculate the entropy H(X/Y)

H(X/Y)

p(00) log p

p(01) log p

0
1

p(10) log p

p(11) log p

Substituting for the various probabilities we get,

H(X/Y)

15
32

log

log p

30
31
31
33

1
32
1
64

log

log

2
33
1

31
64

31

Simplifying you get,

H(X/Y) = 0.27 bit/sym.
[H(X) H(X/Y)] . r s
= (1 0.27) x 1000
R = 730 bits/sec.

2. A transmitter produces three symbols ABC which are related with joint probability shown.
p(i)

9/27

16/27

2/27

p(j/i)

4
5

1
5

1
2

1
2

1
2

2
5

1
10

Calculate H(XY)
Solution:
By definition we have
H(XY) = H(X) + H(Y/X)

-----(1)

Where, H(X) = p(i) log p(i) bits / symbol

-----(2)

and H(Y/X) =
i

Dept. of ECE/SJBIT

p(ij) log p( j / i) bits / symbol

-----(3)

Page 61

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

From equation (2) calculate H(X)

H(X) = 1.257 bits/sym.

To compute H(Y/X), first construct the p(ij) matrix using, p(ij) = p(i), p(j/i)
j

p(i, j)

4
15

1
15

8
27

8
27

1
27

4
135

1
135

From equation (3), calculate H(Y/X) and verify, it is

H(Y/X) = 0.934 bits / sym.

Using equation (1) calculate H(XY)
H(XY) = H(X) + H(Y/X)

= 1.25 + 0.934

2.5 Capacity of a Discrete Memory less Channel (DMC):

Capacity

of

noisy

DMC Is Defined as
The maximum possible rate of information transmission over the channel.
In equation form

C Max Dt

-----(1)

P(x)

i.e., maximized over a set of input probabilities P(x) for the discrete
Definition of Dt?
Dt: Ave. rate of information transmission over the channel defined as
Dt H(x) H(x / y)rs

bits / sec.

-----(2)

Eqn. (1) becomes

C Max H(x) H(x / y)rs

-----(3)

P(x)

Dept. of ECE/SJBIT

Page 62

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

What type of channel is this?

Write the channel matrix
Y

P(Y/X)

10EC55

Do you notice something special in this channel?

What is H(x) for this channel?
Say P(x=0) = P & P(x=1) = Q = (1 P)
H(x) = P log P Q log Q = P log P (1 P) lo g (1 P)
What is H(y/x)?
H(y/x) = [p log p + q logq]

DISCRETE CHANNELS WITH MEMORY:

In such channels occurrence of error during a particular symbol interval does not influence the
occurrence of errors during succeeding symbol intervals
No Inter-symbol Influence
This will not be so in practical channels Errors do not occur as independent events but tend to
occur as bursts. Such channels are said to have Memory.
Examples:
Telephone channels that are affected by switching transients and dropouts
Microwave radio links that are subjected to fading
In these channels, impulse noise occasionally dominates the Gaussian noise and errors occur in
infrequent long bursts.
Because of the complex physical phenomena involved, detailed characterization of channels with
memory is very difficult.
GILBERT model is a model that has been moderately successful in characterizing error bursts in
such channels. Here the channel is modeled as a discrete memoryless BSC, where the probability of
error is a time varying parameter. The changes in probability of error are modeled by a Markoff
process shown in the Fig 1 below.

Dept. of ECE/SJBIT

Page 63

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

The error generating mechanism in the channel occupies one of three states.
Transition from one state to another is modeled by a discrete, stationary Mark off process.
-2

For example, when the channel is in State 2 bit error probability during a bit interval is 10 and the
channel stays in this state during the succeeding bit interval with a probability of 0.998.
However, the channel may go to state 1wherein the bit error probability is 0.5. Since the system
stays in this state with probability of 0.99, errors tend to occur in bursts (or groups).
State 3 represents a low bit error rate, and errors in this state are produced by Gaussian noise. Errors
very rarely occur in bursts while the channel is in this state.
Other details of the model are shown in Fig 1.
The maximum rate at which data can be sent over the channel can be computed for each state of the
channel using the BSC model of the channel corresponding to each of the three states.
Other characteristic parameters of the channel such as the mean time between error bursts and mean
duration of the error bursts can be calculated from the model.
2. LOGARITHMIC INEQUALITIES:

Fig 2 shows the graphs of two functions y1 = x - 1 and y2 = ln x .The first function is
a linear measure and the second function is your logarithmic measure. Observe that the log
function always lies below the
linear function , except at
x = 1. Further the straight line is a
tangent to the log function at x = 1.This is true only for the natural logarithms. For example,
y2
= log2x is equal to y1 = x - 1 at two points. Viz. at x = 1 and at x = 2 .In between these two values
y1
> y2 .You should keep this point in mind when using the inequalities that are obtained. From the
graphs shown, it follows that, y1 y2; equality holds good if and only if x = 1.In other words:
ln x (x-1), equality iffy x = 1

Dept. of ECE/SJBIT

(2.1)
Multiplying equation (2.1) throughout by -1 and

Page 64

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

noting that -ln x = ln (1/x), we obtain another inequality as below.

ln

This

1
x

1 x , equality iffy x = 1

..

(2.2)

property of the logarithmic function will be used

in establishing the extremal property of the Entropy function (i.e. Maxima or minima property). As an
additional property, let {p1, p2, p3,..p n } and {q1, q2, q3,..q n } be any two sets of probabilities
such that pi 0 , qj 0 ; i,j and

qi

.log2

i 1

pi

i 1

q j . Then we have:
j 1

qi

n
log2e . pi

pi

.ln

p
i

i 1

Now using Eq (2.1), it follows that:

n
i

p .log

qi
2

log e .

i 1

i 1
n

qi 1 ,i 1,2 ,...n
pi
n

log2 e .qi pi ,
i 1

qi

n
This, then implies pi .log 2

1
.log2

p
i

i 1

p
i

i 1

That is, pi

i 1

1
pi .log2

i 1

(2.3)

equality iffy pi = qi i=1, 2, 3...n ..........

q
i

This inequality will be used later in arriving at a measure for code efficiency
3. PROPERTIES OF ENTROPY:
We shall now investigate the properties of entropy function
1. The entropy function is continuous in each and every independent variable ,pk in the
interval (0,1)
This property follows since each pk is continuous in the interval (0, 1) and that the
logarithm of a continuous function is continuous by itself.
2. The entropy function is a symmetrical function of its arguments; i.e.
H (pk, 1- pk) = H (1- pk, pk) k =1, 2, 3...q.
That is to say that the value of the function remains same irrespective of the locations
(positions) of the probabilities. In other words, as long as the probabilities of the set are
same, it does not matter in which order they are arranged. Thus the sources S1, S2 and S3
with probabilities:

Dept. of ECE/SJBIT

Page 65

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

P1 = {p1, p2, p3, p4}, P2 = {p3, p2, p4, p1} and P3 = {p4, p1, p3, p2} with pk 1 all have
k 1

same entropy, i.e. H (S1) = H (S2) = H (S3)

3. Extremal property:
Consider a zero memory information source with a q-symbol alphabet
S ={s1 , s2 , s3 ,.. ,s q } with associated probabilities P = {p1 , p2 , p3 ,.. ,p q }.
Then we have for the entropy of the source (as you have studied earlier):
q

H ( S ) p .log 1
k

pk

k 1

Consider log q H(S). We have:

log q H(S) = log q - p

q
k

k 1

p .log

= p .log q
k

k 1

.log 1
pk
k

k 1

= log e p .ln q

k 1
q

pk

k 1

= log e . p ( ln q ln
k

k 1
q

.ln 1
pk

1)
pk

= log e . pk ( ln q pk )
k 1

Invoking the inequality in Eq (2.2), we have:

q
log q H(S) log e . p ( 1 1 ) Equality iffy q p k = 1,
q pk
k 1
q
q 1
log e p

Equality iffy p k = 1/q,

q
k

Since p

k 1

k 1

k =1, 2, 3q.

k =1, 2, 3q.

k 1

1, it follows that log q H(S) 0

k 1 q

Or in other words

H(S) log q

(2.4)

The equality holds good iffy p k = 1/q, k =1, 2, 3q. Thus for a zero memory information source,
with a q-symbol alphabet, the Entropy becomes a maximum if and only if all the source symbols are
equally probable . From Eq (2.4) it follows that:
H(S) max = log q iffy p k = 1/q, k =1, 2, 3 q

..

(2.5)

Particular case- Zero memory binary sources:

Dept. of ECE/SJBIT

Page 66

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

For such a source, the source alphabet is S= {0, 1} with P= {q,

p}. Since p+q=1, we have

lim

H(S) = p log (1/p) + q log (1/q) = -p log p-(1-p log (1-p).Further, as

p 0 log p 0 , we define
0log 0=0 and 1 log 1=0. A sketch showing the variation of H(S) with p is shown in Fig.2.5

If the output of the source is certain (i.e. p=0 or p =1) then the source provides no information. The
maximum entropy provided by the source is log 22 =1 bit/binits, occurs iffy the 0 and the 1 are
equally probable. The outputs of a binary source are Binary digits or Binits. Hence, a sequence of
binits from a zero memory binary source with equi-probable 0s and 1s will provide 1 bit of
information per binit. If the 0s and 1s are not equi-probable, then the amount of information
provided by a given binit will be either less than or greater than 1 bit, depending upon its probability.
However the average amount of information provide d by a binit from such a source will always be
less than or equal to 1 bit per binit.
EXTENSION OF A ZERO MEMORY SOURCE:
The questions of extension of sources arise in coding problems. If multi-alphabet source
outputs are to be encoded in to words of a lesser alphabet, then it is necessary to have an extension of
the later. For example , if we are to code four messages with a binary source S={ 0 , 1 } , it is
necessary to have the binary word { 00 , 01 , 10 , 11 ,}, leading to the second extension of the source.
th
Thus, if a zero memory source S has the alphabets { s 1 , s 2 , , s 3
, , s q }, then its n extension,
n
n
n
called S , is a zero memory source with q symbols{ 1 , 2 , 3 , 4 , . , q } as its
higher order alphabet. The corresponding statistics of the extension are given by the probabilities:
P ( i) = P (s i i1, s i2, s i3 . , s

(Problem 2.1.6 Simon Haykin)

in)

where i = { s i1 , s i2 , s i3 . , s
that is , each i corresponds to some sequence of n
in } ,
- symbols, si,of the source. Since it is a zero memory source all the resultant composite symbols are
independent. There fore:
P ( i) = P(s i i1).P(s i2) .P(s i3)
P (s in). Further:
n
q

The condition that P( i ) 1 is satisfied , sin ce ,

q

i 1

P( i ) P( si 1 ).P( si 2 )..........P( sin )

Sn

i 1
q

i1 1

i2 1

P (si1 ). P (si 2 ) ............

P(sin )
in 1

sin ce each (.) 1

It then follows that:

Dept. of ECE/SJBIT

Page 67

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H (S ) P( i ).log
n

P( i )
P( i ).log

.
P(si1 )P(si 2 ).P(si 3 )..........P(sin )
P( i ).log
1 . P(
1 . .............P(
1
n
.
i ).log
i ).log
S
n
n
P(si1 )
P(s
)
P(s
)
S
S
i2
in
S

Thus the Entropy of the extension, S , of the Zero memory is n times the Entropy of the original
source.
Example:
A zero memory source has a source alphabet S = {s1, s2, s3}
, with P = {1/2, 1/4, 1/4}. Find the
entropy of the source. Find the entropy of the second extension and verify Eq (2.6).
We have H(S) = (1/2) log (2) + 2 (1/4) log 4 = 1.5 bits
/sym. The second extension and the
corresponding probabilities are tabulated as below:
S

= {s1s1 , s1s2 , s1s3 , s2s1 , s2s2 , s2s3 , s3s1 , s3s2 , s3s3 }

P( S ) = { 1/4 , 1/8 , 1/8 , 1/8 , 1/16 , 1/16 , 1/8 , 1/16 , 1/16 }

2

Hence, H(S ) = (1/4) log 4 + 4 (1/8) log 8 + 4 (1/16) log 16 = 3.0 bits / sym.
2
2
{H (S )} {H (S)} = 3 / 1.5 = 2 ; and indeed H ( S )= 2. H (S)
SHANNONS FIRST THEOREM (Noiseless Coding Theorem):

Given a code with an alphabet of r-symbols and a source with an alphabet of q-symbols, the average
length of the code words per source symbol may be made as arbitrarily close to the lower bound
H(S)/log r as desired by encoding extensions of the source rather than encoding each source symbol
individually.
The draw back is the increased coding complexity of the encoding procedure caused by the
n
large number (q ) of source symbol with which we must deal and in the increased time required for
encoding and transmitting the signals. Although the theorem has been proved here for zero memory
sources, it is also valid for sources with memory i.e. for Markov sources.
Construction of some Basic Codes:

Dept. of ECE/SJBIT

Page 68

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

So far we have discussed the properties of the codes, bounds on the word lengths and the
Shannons first fundamental theorem. In this section we present some code generating techniques
Shannons encoding procedure, Shannon Fano codes and Huffmans minimum redundancy codes.

Shannon binary encoding procedure:

We present, first, the Shannons procedure for generating binary codes mainly because of its
historical significance. The procedure is based on Eq (5.32). The technique is easy to use and
generates fairly efficient codes. The procedure is as follows:
1. List the source symbols in the order of decreasing probability of occurrence.
S= {s1, s2 sq};

P= {p1, p2 .pq}: p1 p2

pq

2. Compute the sequence:

0 = 0,
1 = p1,
2 = p2+p1 = p2+1
3=p3+p2+p1 = p3 + 2
.
q-1= pq-1+ pq-2+ +p
1 = pq-1 + q-2.
q=pq +pq-1+ +p 1 = pq + q-1 =1

3. Determine the set of integers, lk, which are the smallest integers solution of the inequalities.
l

lk

2 k pk 1, k=1, 2, 3 q . Or alternatively, find lk such that 2

pk .

4. Expand the decimal numbers k in binary form to lk places. i.e., neglect expansion beyond lk digits
5. Removal of the decimal point would result in the desired code.
Example 6.9:
Consider the following message ensemble
S= {s1, s2, s3, s4}, P= {0.4, 0.3, 0.2, 0.1}
Then following Shannons procedure, we have
1) 0.4 > 0.3 > 0.2 > 0.1
2) 0 = 0,
1 = 0.4
2 = 0.4+0.3=0.7
3= 0.7+0.2=0.9
4= 0.9 + 0.1 = 1.0

Dept. of ECE/SJBIT

Page 69

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

l1

3) 2

l 2

2
2

0.4 l1 2

l3

l4

10EC55

0.3 l2 2
0.2 l3 3

0.1 l4 4

4) 0 = 0 = 0.00 0 1
= 0.4 = 0.01 10

2= 0.7 = 0.101 10
3 = 0.9 = 0.1110 01
5) The codes are
s1 00, s2 01, s3 101, s4 1110
The average length of this code is
L = 2 0.4 + 2 0.3 + 3 0.2 + 3 0.1 = 2.4 Bini ts / message
H(S) = 0.4 log

1
1
1
1
+ 0.3 log
+0.2 log
+0.1 log
= 1.84644 bits / message;
0.4
0.3
0.2
0.1

log 2 = 1.0 and c H ( S ) 1.84644 76.935 % ; Ec = 23.065 %

L Log r
2.4 1
Shannon Fano Binary Encoding Method:
Shannon Fano procedure is the simplest available. Code obtained will be optimum if and only if
pk r

l k

.The procedure is as follows:

1. List the source symbols in the order of decreasing probabilities.

2. Partition this ensemble into almost two equi- probable groups.
Assign a 0 to one group and a 1 to the other group. These form the starting code symbols
of the codes.
3. Repeat steps 2 and 3 on each of the subgroups until the subgroups contain only one source
symbol, to determine the succeeding code symbols of the code words.
4. For convenience, a code tree may be constructed and codes read off directly.
Example
Consider the message ensemble S = {s1, s2, s3, s4, s5, s6, s7, s8} with
1 1 1 1 1 1 1 1
P=

, X { 0 ,1 }

4 8 8 16 16 16 16

Dept. of ECE/SJBIT

Page 70

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

The procedure is clearly indicated in the Box diagram shown in Fig 6.3. The Tree diagram for the
steps followed is also shown in Fig 6.4. The codes obtained are also clearly shown. For this example,
L = 2 1 + 2 1 + 3 1 + 3 1 + 4 1 + 4 1 + 4 1 + 4 1 = 2.75 binits / symbol
8
16
4
4
8
16
16
16
H(S) = 2 1 log 4 + 2 1 log 8 + 4 1 log 16 = 2.75 bits/symbol.
4
8
16
H ( S ) 100% and E = 0%
And as log r = log 2 = 1, we have c
c
L log r

Incidentally, notice from tree diagram that the codes originate from the same source and diverge into
different tree branches and hence it is clear that no complete code can be a prefix of any other code word.
Thus the Shannon- Fano algorithm provides us a means for constructing optimum, instantaneous codes.

Dept. of ECE/SJBIT

Page 71

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

In making the partitions, remember that the symbol with highest probability should be made
to correspond to a code with shortest word length. Consider the binary encoding of the following
message ensemble.
Example 6.11
S = {s1, s2, s3, s4, s5, s6, s7, s8}
P = 0.4 , 0.2 , 0.12 , 0.08 , 0.08 , 0.08 , 0.04
Method - I

Method - II

For the partitions adopted, we find L=2.52 binits / sym. for the Method I
L=2.48 binits/sym for the Method - II
For this example, H(S) =2.420504 bits/sym and
For the first method, c1 96.052%
For the second method,c 2 97.6%
This example clearly illustrates the logical reasoning required while making partitions. The
Shannon Fano algorithm just says that the message ensemble should be partitioned into two almost
equi-probable groups. While making such partitions care should be taken to make sure that the
symbol with highest probability of occurrence will get a code word of minimum possible length. In

Dept. of ECE/SJBIT

Page 72

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

the example illustrated, notice that even though both methods are dividing the message ensemble into
two almost equi-probable groups, the Method II as signs a code word of smallest possible length to
the symbol s1.

Review questions :
1. What do you mean by source encoding? Name the functional requirements to be satisfied in the
development of an efficient source encoder.
2. For a binary communication system, a 0 or 1 is transmitted. Because of noise on the channel, a
0 can be received as 1 and vice-versa. Let m 0 and m1 represent the events of transmitting 0
and 1 respectively. Let r 0 and r0 denote the events of receiving 0 and 1 respect ively. Let
p(m0) = 0.5, p(r1/m0) = p = 0.1, P(r0/m1) = q = 0.2
i.
Find p(r0) and p(r1)
ii.
If a 0 was received what is the probability that 0 was sent
iii.
If a 1 was received what is the probability that 1 was sent.
iv.
Calculate the probability of error.
v.
Calculate the probability that the transmitted symbol is read correctly at the receiver.
3. State Shannon-Hartleys law. Derive an equation showing the efficiency of a system in terms of
the information rate per Unit bandwidth. How is the efficiency of the system related to B/W?
4. For a discrete memory less source of entropy H(S), show that the average code-word length for
any distortion less source encoding scheme is bounded as LH(S).
5. Calculate the capacity of a standard 4KHz telephone channel working in the range of 200 to 3300
KHz with a S/N ratio of 30 dB.
6. What is the meaning of the term communication channel. Briefly explain data communication
channel, coding channel and modulation channel.
7. Obtain the communication capacity of a noiseless channel transmitting n discrete message
system/sec.
8. Explain extremal property and additivity property.
9. Suppose that S1, S2 are two memory sources with probabilities p1,p2,p3,pn for source s1
and q1,q2,.qn for source s2 . Show that the entro py of the source s1
n
H(s1) Pk log (1/qk)
K=1
10. Explain the concept of B/W and S/N trade-off with reference to the communication channel.

Dept. of ECE/SJBIT

Page 73

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

Unit 3 : Fundamental Limits on Performance

Syllabus :
Source coding theorem, Huffman coding, Discrete memory less Channels, Mutual
information, Channel Capacity. 6 Hours

Text Books:
Digital and analog communication systems, K. Sam Shanmugam, John Wiley,
1996. Digital communication, Simon Haykin, John Wiley, 2003.

Reference Books:
ITC and Cryptography, Ranjan Bose, TMH, II edition, 2007

Dept. of ECE/SJBIT

Page 74

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Unit 3 : Fundamental Limits on Performance

Source coding theorem:
Compact code: Huffmans Minimum Redundancy code:
The Huffman code was created by American, D. A. Huffman, in 1952. Huffman`s procedure is
applicable for both Binary and Non- Binary encoding. It is clear that a code with minimum average
length, L, would be more efficient and hence would have minimum redundancy associated with it. A
compact code is one which achieves this objective. Thus for an optimum coding we require:
1) Longer code word should correspond to a message with lowest probability.
r 1

2) lk lk 1 k 1,2 ,......q

(3) lp-r = lq-r-1 = l q-r-2 = ..= lq

(4) The codes must satisfy the prefix property.
Huffman has suggested a simple method that guarantees an optimal code even if Eq. (6.13) is
not satisfied. The procedure consists of step- by- step reduction of the original source followed by a
code construction, starting with the final reduced source and working backwards to the original
source. The procedure requires steps, where
q = r + (r-1)

(6.24)

Notice that is an integer and if Eq.(6.24) is not satisfied one has to add few dummy symbols with
zero probability of occurrence and proceed with the procedure or the first step is performed by setting
r1 = q-(r-1) while the remaining steps involve clubbing of the last r messages of the respective stages.
The procedure is as follows:
1. List the source symbols in the decreasing order of probabilities.
2. Check if q = r + (r-1) is satisfied and find the integer . Otherwise add suitable number of
dummy symbols of zero probability of occurrence to satisfy the equation. This step is not
required if we are to determine binary codes.
3. Club the last r symbols into a single composite symbol whose probability of occurrence is
equal to the sum of the probabilities of occurrence of the last r symbols involved in the step.
4. Repeat steps 1 and 3 respectively on the resulting set of symbols until in the final step exactly
r- symbols are left.
5. Assign codes freely to the last r-composite symbols and work backwards to the original
source to arrive at the optimum code.
6. Alternatively, following the steps carefully a tree diagram can be constructed starting from the
final step and codes read off directly.
7. Discard the codes of the dummy symbols.

Dept. of ECE/SJBIT

Page 75

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Before we present an example, it is in order to discuss the steps involved. In the first step,
after arranging the symbols in the decreasing order of probabilities; we club the last r-symbols into a
composite symbol, say 1 whose probability equals the sum of the last r-probabilities. Now we are
left with q-r+1 symbols .In the second step, we again club the last r-symbols and the second reduced
source will now have (q-r+1)-r+1= q-2r+2 symbols .Continuing in this way we find the k-th reduced
source will have q- kr + k = q k(r - 1) symbols. Accordingly, if -steps are required and the final
reduced source should have exactly r-symbols, then we must have r = q - (r - 1) and Eq (5.38) is
proved. However, notice that if Eq (6.24) is not satisfied, we may just start the first step by taking the
last r1=q-( r 1 ) symbols while the second and subsequent reductions involve last r-symbols only.
However, if the reader has any confusion, he can add the dummy messages as indicated and continue
with the procedure and the final result is no different at all.
Let us understand the meaning of working backwards. Suppose k is the composite symbol
th
th
obtained in the k step by clubbing the last r-Symbols of the (k-1) reduced source. Then whatever
code is assigned to k will form the starting code sequence for the code words of its constituents in the
th
(k-1) reduction.
Example 6.12: (Binary Encoding)
S = {s1, s2, s3, s4, s5, s6}, X = {0, 1};
P=

1 1 1 1
,

3 4

8 8 12

1
12

The reduction diagram and the tree diagram are given in Fig 5.7.Notice that the tree diagram can
be easily constructed from the final step of the source reduction and decomposing the composite
symbols towards the original symbols. Further, observe that the codes are originating from the same
source and diverge out into different tree branches thus ensuring prefix property to the resultant code
.Finally, notice that there is no restriction in the allocation of codes in each step and accordingly, the
order of the assignment can be changed in any or all steps. Thus for the problem illustrated there can
be as many as 2. (2.2+2.2) = 16 possible instantaneous code patterns. For example we can take the
compliments of First column, Second column, or Third column and combinations there off as
illustrated below.

Code
s1 0 0

s2 .. 1 0

s3 . 0 1 0
s4 . 0 1 1
s5 1 1 0
s6 . 1 1 1

Dept. of ECE/SJBIT

I
10
11
00
01
11 0 10 0
111 101
010 000
011 001

II

III
11
01
101
100
001
000

Page 76

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

Code I is obtained by taking complement of the first column of the original code. Code II is obtained by
taking complements of second column of Code I. Code III is obtained by taking complements of third
column of Code II. However, notice that, lk, the word length of the code word for sk is a constant for all
possible codes.
For the binary code generated, we have:
6

L pk lk
k 1

1 2 1 2 1 3 1 3 1 3 1 3 = 29 binits/sym=2.4167 binits/sym
12
3
4
8
8
12
12

H(S) = 1 log 3 1 log 4 2 1 log 8 2 1 log 12

3
4
8
12
1
= ( 6 log 3 19 ) bits/sym = 2.3758 bits/ sym
12
c =

6 log 3 19

98.31% ; Ec = 1.69%

29
Example 6.13 (Trinary Coding)
We shall consider the source of Example 6.12. For Trinary codes r = 3, [X = (0, 1, 2)]
Since q = 6, we have from
q = r + (r-1)
= q r 6 3 3 1.5
r1
2
2
Thus is not an integer and hence we require one dummy message which makes = 2.

Dept. of ECE/SJBIT

Page 77

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

For this code, we have

L 1 1 1 1 2 1 2 1 3 1 3 1 19 Trinits / sym .
12 12
3
4
8
8
12
And 6 log 3 19 94.672% , Ec=5.328%
c
19
Example 6.14:
We conclude this section with an example illustrating Shannons noiseless coding theorem.
Consider a source S = {s1, s, s3} with P = {1/2, 1/3, 1/6}
A compact code for this source is: s1 0, s2 10, s3 11
Hence we have
L 1 2 2 1.5
2 3 6
H (S ) 1 log 2 1 log 3 1 log 6
6
2
3
= 1.459147917 bits/sym
c = 97.28%
2

The second extension of this source will have 3 = 9 symbols and the corresponding probabilities are
computed by multiplying the constituent probabilities as shown below
1

s1 s1

4
1

s1 s2

6
1

s2 s 1 6
1
s2 s2 9

s1 s3

12

s2 s3 18

s3 s1 12
1
s3 s2 18
1
s3 s3 36

These messages are now labeled mk and are arranged in the decreasing order of probability.
M = {m1, m2, m3, m4, m5, m6, m7, m8, m9}

Dept. of ECE/SJBIT

Page 78

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

P=

1
4

1 1 1
,

10EC55

6 6 9 12 12

18 18 36

The Reduction diagram and tree diagram for code construction of the second extended source is
shown in Fig 5.9.

For the codes of second extension, we have the following:

2
H (S ) = 2 H(S)
L = 2 1 + 2 1 + 3 1 + 3 1 + 4 1 + 4 1 + 4 1 + 5 1 + 5 1
36
4
6
6
9
12
12
18
18
= 107 binits/symbol = 2.97222222 binits/sym
36
2
H(S )
2 x 1.459147917
c =

= 98.186 % Ec = 1.814 %
L log 2
2.97222222
An increase in efficiency of 0.909 % (absolute) is achieved.
This problem illustrates how encoding of extensions increase the efficiency of coding in
accordance with Shannons noiseless coding theorem.
One non- uniqueness in Huffman coding arises in making decisions as to where to move a
composite symbol when you come across identical probabilities. In Shannon- Fano binary encoding
you came across a situation where you are required to make a logical reasoning in deciding the
partitioning. To illustrate this point, consider the following example.

Dept. of ECE/SJBIT

Page 79

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Example 6.15:
Consider a zero memory source with
S= {s1, s2, s3, s4, s5}; P= {0.55, 0.15, 0.15, 0.10, 0.05}; X= {0, 1}
Construct two different Huffman binary codes as directed below:
(a) Move the composite symbol as high as possible.
(b) Move the composite symbol as low as possible
(c) In each case compute the variance of the word lengths and comment on the results.
(a)We shall place the composite symbol as high as possible. The source reduction and the
corresponding tree diagram are shown in Fig 6.10

Symbols
Codes
lk

s1
0
1

s2
100
3

s3
101
3

s4
110
3

s5
111
3

We compute the average word length and variance of the word lengths as below:

L=0.55+3(0.15+0.15+0.10+0.05) =1.90 binits/symbol

2
2
2
l = 0.55(1-1.90) + 0.45 (3-19) = 0.99 is the variance of the word length.

(a) We shall move the composite symbol as low as possible. The source reduction and the
corresponding tree diagram are shown in Fig 5.11.We get yet another code, completely
different in structure to the previous one.
Symbols

Dept. of ECE/SJBIT

s1

s2

s3

s4

s5

Page 80

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Codes
lk

0
1

11
2

100 1010 1011

3
4
4

For this case we have: L = 0.55 + 0.30 + 0.45 + 0.20= 1.90 binits/symbol
Notice that the average length of the codes is same.

2= 0.55 (1 -1.9) + 0.15 (2 -1.9) + 0.15(3 1.9) + 0.10(4 -1.9) + 0.05(4 -1.9)
= 1.29 is the variance of the word lengths.

Thus, if the composite symbol is moved as high as possible, the variance of the average code
word length over the ensemble of source symbols would become smaller, which, indeed, is desirable.
Larger variance implies larger buffer requirement for storage purposes. Further, if the variance is
large, there is always a possibility of data overflow and the time required to transmit information
would be larger. We must avoid such a situation. Hence we always look for codes that have minimum
possible variance of the word lengths. Intuitively avoid reducing a reduced symbol in the immediate
next step as far as possible moving the composite symbol as high as possible.
DISCRETE MEMORYLESS CHANNELS:
A multi-port electric network may be uniquely described by the impedance matrices, viz,

Dept. of ECE/SJBIT

Page 81

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Observe that the matrix is necessarily a square matrix. The principal diagonal entries are the self
impedances of the respective ports. The off diagonal entries correspond to the transfer or mutual
T
impedances. For a passive network the impedance matrix is always symmetric i.e. Z = Z, where the
superscript indicates transposition.
Similarly, a communication network may be uniquely described by specifying the joint
probabilities (JPM). Let us consider a simple communication network comprising of a transmitter
(source or input) and a receiver (sink or output) with the interlinking medium-the channel as shown in
Fig 4.1.

Fig 4.1 A Simple Communication System

This simple system may be uniquely characterized by the Joint probability matrix ( JPM),
P(X, Y) of the probabilities existent between the input and output ports.

P( X ,Y )

p( x1 , y1 )

p( x1 , y2 )

P( x1 , y3 ) .....

P( x1 , yn )

P( x2 , y1 )
P( x3 , y1 )

P( x2 , y2 )

P( x2 , y3 ) .....

P( x3 , y2 )

P( x3 , y3 ) .....
M
M

P( x2 , yn )
P( x3 , yn )

M
M

M
M

P( xm , y1 )

P( xm , y2 )

..........

(4.1)

P( xm , y3 ) .....

P( xm , yn )
For jointly continuous random variables, the joint density function satisfies the following:

f ( x, y)dxdy 1

.2)

. (4

f ( x, y)dy f X ( x)

(4.3)

f ( x, y)dx fY ( y)

.. (4.4)

We shall make use of their discrete counterpart as below:

\ p( xk , y j ) 1 ,Sum of all entries of JPM

......... (4.5)

k j

p( xk , y j ) p( xk ) , Sum of all entries of JPM in the kth r

..(4.6)

p( xk , y j ) p( y j ) , Sum of all entries of JPM in the jth column

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

(4.7)

And also

Dept. of ECE/SJBIT

Page 82

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

p( xk ) = p( y j ) =1
k

(4.8)

Thus the joint probabilities, as also the conditional probabilities (as we shall see shortly) form
complete finite schemes. Therefore for this simple communication network there are five probability
schemes of interest viz: P(X), P(Y), P(X, Y), P (X|Y) and P (Y|X). Accordingly there are five entropy
functions that can be described on these probabilities:
H(X): Average information per character or symbol transmitted by the source or the entropy of the
source.
H(Y): Average information received per character at the receiver or the entropy of the receiver.
H(X, Y): Average information per pair of transmitted and received characters or the average
uncertainty of the communication system as a whole.
H (X|Y): A specific character yj being received. This may be the result of the transmission of one of the
xk with a given probability. The average value of the Entropy associated with this scheme when yj covers
all the received symbols i.e., E {H (X|yj)} is the entropy H (X|Y), called the Equivocation, a measure of
information about the source when it is known that Y is received.
H (Y|X) : Similar to H (X|Y), this is a measure of information about the receiver.
The marginal Entropies H(X) and H(Y) give indications of the probabilistic nature of the
transmitter and receiver respectively. H (Y|X) indicates a measure of the noise or error in t he
channel and the equivocation H(X |Y) tells about the ability of recovery or reconstruction of the
transmitted symbols from the observed output symbols.
The above idea can be generalized to an n- port communication system, problem being similar
to the study of random vectors in a product space (n-dimensional random variables Theory). In each
product space there are finite numbers of probability assignments (joint, marginal and conditional) of
different orders, with which we may associate entropies and arrive at suitable physical interpretation.
However, concepts developed for a two-dimensional scheme will be sufficient to understand and
generalize the results for a higher order communication system.
Joint and Conditional Entropies:
In view of Eq (4.2) to Eq (4.5), it is clear that all the probabilities encountered in a two
dimensional communication system could be derived from the JPM. While we can compare the JPM,
therefore, to the impedance or admittance matrices of an n-port electric network in giving a unique
description of the system under consideration, notice that the JPM in general, need not necessarily be a
square matrix and even if it is so, it need not be symmetric.
We define the following entropies, which can be directly computed from the JPM.
1
1
1
H(X, Y) = p(x1, y1) log p(x , y ) + p(x1, y2) log p(x , y ) ++ p(x1, yn) log p(x , y )
1

+ p (x2, y1) log

+ p(x2,y2) log

p(x2 , y1 )

Dept. of ECE/SJBIT

++ p(x2,yn) log

p(x2 , y2 )

p(x1 , y1 )

Page 83

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

p (xm, y1) log

p(xm , y1 )
m

or

p(xm , yn )

p(xm , y2 )
(4.9)

p(x k , y j )

k 1 j1

1
p(xm,yn) log

H(X, Y) = p(x k , y j ) log

H(X) = p(xk

+ p(xm,y2) log

) log

p( xk )
Using Eq (4.6) only for the multiplication term, this equation can be re-written as:
1
m n
H(X) = p( xk , y j
) log
(4.10)

p( xk )
k 1 j1
1
n m
k 1

Similarly, H(Y) = p( xk , y j ) log

. (4.11)

p( y j )

j1k 1

Next, from the definition of the conditional probability we have:

P{X = xk | Y = yj} =

P{X xk ,Y y j }
P{Y y j }

i.e., p(xk | y j ) = p(xk , y j ) ) / p (yj)

m

Then p(xk | y j

)=

p(xk

p( y j )

k 1

,yj

)=

1
. p(y j

) =1

(4.12)

p( y j )

k 1

Thus, the set [X | yj] = {x1 | yj, x2 | yj xm | yj}; P [X | yj] = {p(x1 | yj), p(x2 | yj) p (xm | yj)}, forms a
complete finite scheme and an entropy function may therefore be defined for this scheme as below:
1
m
H(X | yj) = p(xk | y j )log
.
p(xk | y j )
k 1
Taking the average of the above entropy function for all admissible characters received, we have the
average conditional Entropy or Equivocation:
H(X | Y) = E {H(X | yj)} j
n

= p( y j ) H(X | yj)
j1

j1
nm

Or H(X | Y) = p(xk , y j ) log

j 1 k 1

= p( y j ) p( xk

| y j ) log

k 1

1
p(xk | y j )

p( xk | y j )
(4.13)

Eq (4.13) specifies the Equivocation . It specifies the average amount of information n eeded to
specify an input character provided we are allowed to make an observation of the output produced by
that input. Similarly one can define the conditional entropy H(Y | X) by:

Dept. of ECE/SJBIT

Page 84

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

m n

k 1 j1

p( y j | xk )

H(Y | X) = p( xk , y j ) log

(4.14)

Observe that the manipulations, made in deriving Eq 4.10, Eq 4.11, Eq 4.13 and Eq 4.14, are
intentional. The entropy you want is simply the double summation of joint probability multiplied by
logarithm of the reciprocal of the probability of interest . For example, if you want joint entropy, then
the probability of interest will be joint probability. If you want source entropy, probability of interest will
be the source probability. If you want the equivocation or conditional entropy, H (X|Y) then probability of
interest will be the conditional probability p (xK |yj) and so on.
All the five entropies so defined are all inter-related. For example, consider Eq (4.14). We have:

p(x k , y j ) log

H(Y | X) =

1
p(y j | x k )

p(xk )

Since, p( y | x ) = p(x , y )
k
j
j
k
We can straight away write:
1

H (Y|X) = p(x k , y j )log

k

Or

p(y j | x k )

p(x k , y j )log
k

1
p(x k )

H(Y | X) = H(X, Y) H(X)

(4.15)

That is:
H(X, Y) = H(X) + H(Y | X)
Similarly, you can show: H(X, Y) = H(Y) + H(X | Y)
.
Consider H(X) - H(X |Y). We have:
1
1
H(X) - H(X |Y) = p(x k , y j ) log
log
p(x )
p(x | y
..

(4.16)

p(xk , y j )
(4.17)
p(xk ) . p(y j )
k
j
Using the logarithm inequality derived earlier, you can write the above equation as:
p(xk , y j )
H(X) - H(X |Y) =log e p(xk , y j )ln
p(xk ) . p(y j )
k j
= p(xk , y j )log

log e p(xk , y j
k j

- p(xk).p(yj)
k

- p(xk). p(yj) 0

p(x , y )

log e p(xk , y j

Because

p(xk ) . p(y j

)1-

log e p(xk , y j

p(xk , y j ) p(xk ) p(y j ) =1. Thus it follows that:

Dept. of ECE/SJBIT

H(X) H (X|Y)

(4.18)

Page 85

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H(Y) H (Y|X)

Similarly,

..

(4.

19)

Equality in Eq (4.18) &Eq (4.19) holds iffy P (xk , yj) = p(xk) .p(yj); i.e., if and only if input symbols
and output symbols are statistically independent of each other.
NOTE : Whenever you write the conditional probability matrices you should bear in mind the property
described in Eq.(4.12), i.e. For the CPM (conditional probability matrix ) P(X|Y), if you add all the
elements in any column the sum shall be equal to unity. Similarly, if you add all elements along any
row of the CPM, P (Y|X) the sum shall be unity
Example 4.1
Determine different entropies for the JPM given below and verify their relationships.

0.2
0
0.2
0.1 0.01 0.01

0
0.01

0 0.02 0.02
0
0.04 0.04 0.01 0.06

P(X, Y) =

0 0.06 0.02

0.2

Using p (xk) = p(xk , y j) , we have, by adding entries of P(X, Y) row-wise we get:

j1

P(X) = [0.4, 0.1, 0.04, 0. 15, 0.28]

Similarly adding the entries column-wise we get:
P(Y) = [0.34, 0.13, 0.26, 0.27]
Hence we have:

1 0.1 log 1 4 0.01 log 1

0.2
0.1
0.01
3 0.02 log 1
2 0.04 log 1
2 0.06 log 1
0.02
0.04
0.06
= 3.188311023 bits /sym

H(X,Y ) 3 0.2 log

1
1
1
1
1
0.13 log
0.04 log
0.15 log
0.28 log
0.4
0.13
0.04
0.15
0.28
= 2.021934821 bits / sym

H(X) = 0.4 log

1
1
1
1
0.13 log
0.26 log
0.27 log
0.34
0.13
0.26
0.27
= 1.927127708 bits / sym

H(Y) = 0.34 log

Since p (xk | yj) =

Dept. of ECE/SJBIT

P(x k , y j )
P( y j )

we have:

Page 86

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

th

(Divide the entries in the j column of the JPM of p (yj)

0.2
0.34
0.1
0.34
P (X| Y) =

0
0.04
0.34
0

0.01
0.13
0.02
0.13
0.04
0.13
0.06

0.2
0.26
0.01
0.26
0.02
0.26
0.01
0.26
0.02

0.06
0.27
0.20

0.13

0.26

0.27

0
0.01
0.27
0

H(X | Y) 0.2log 0.34 0.2log 0.26

0.1log 0.34
0.2
0.2
0.1
0.13
0.26
+ 0.01log
0.01log
0.01log 0.27
0.01
0.01
0.01
+ 0.02log 0.13 0.02log 0.26 0.04log 0.34
0.02
0.02
0.04
+ 0.04log 0.13 0.01log 0.26 0.06 log 0.27
0.04
0.01
0.06
+ 0.06log 0.13 0.02log 0.26 0.2 log 0.27
0.06
0.02
0.2

=1.261183315 bits / symbol

th

Similarly, dividing the entries in the k row of JPM by p (xk,), we obtain the CPM P (Y|X).Then we
have:
0.2
0.4
0.1
0.13
P(Y | X)

0
0.04
0.15
0

Dept. of ECE/SJBIT

0.01
0.13
0.02
0.04
0.04
0.15
0.06

0.2
0.4
0.01
0.13
0.02
0.04
0.01
0.15
0.02

0.06
0.15
0.20

0.28

0.28

0.28

0
0.01
0.13
0

Page 87

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

And H(Y | X) 2 0.2log 0.4 0.1log 0.13 3 0.01 log 0.13 2 0.02 log 0.04
0.01
0.2
0.1
0.02
0.05
0.15
0.15
2 0.04 log
0.01 log
0.06 log
0.06 log 0.28
0.04
0.01
0.06
0.06
2 0.02 log 0.28 1.166376202 bits / sym .
0.02
Thus by actual computation we have
H(X, Y) = 3.188311023 bits/Sym H(X)= 2.02193482 bit/Sym H(Y)= 1.927127708 bits/Sym
H(X | Y) = 1.261183315 bits/Sym H(Y | X) = 1.166376202 bits/Sym
Clearly, H(X, Y) = H(X) + H(Y | X) = H(Y) + H(X | Y)
H(X) > H(X | Y) and H(Y) > H(Y | X)
Mutual information:
On an average we require H(X) bits of information to specify one input symbol. However, if
we are allowed to observe the output symbol produced by that input, we require, then, only H (X|Y)
bits of information to specify the input symbol. Accordingly, we come to the conclusion, that on an
average, observation of a single output provides with [H(X) H (X|Y)] bits of information. This
difference is called Mutual Information or Transinformation of the channel, denoted by I(X, Y).
Thus:
I(X, Y) H(X) - H (X|Y)

.. (4.20)

Notice that in spite of the variations in the source probabilities, p (xk) (may be due to noise in
the channel), certain probabilistic information regarding the state of the input is available, once the
conditional probability p (xk | yj) is computed at the receiver end. The difference between the initial
uncertainty of the source symbol xk, i.e. log 1/p(xk) and the final uncertainty about the same source
symbol xk, after receiving yj, i.e. log1/p(xk |yj) is the information gained through the channel. This
difference we call as the mutual information between the symbols xk and yj. Thus
1
I(x k , y j )

log p(x )
k
log

OrI (xk, yj)

log

1
log p(x | y )
k
j

p(xk | y j )
p(xk )
p(x k .y j )
p(x k ).p(y j )

a)

(4.21

(4.21

b)

Notice from Eq. (4.21a) that

I (xk) = I (xk, xk) = log p(x k | x k ) log 1
p(x k )
p( x k )
This is the definition with which we started our discussion on information theory! Accordingly I (xk)
is also referred to as Self Information.

Dept. of ECE/SJBIT

Page 88

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

p(xk , y j )
p(y j | xk ) ,
p(xk )

It is clear from Eq (3.21b) that, as

p(y j | x k )

I (x k , y j ) log

= log

p(y j )
I (xk, yj) = I (yj) I (y j |xk)

Or

1
p(y j )

log

p(y j | x k )
(4.22)

Eq (4.22) simply means that the Mutual information is symmetrical with respect to its
arguments.i.e.
I (xk, yj) = I (yj, xk)
(4.23)
Averaging Eq. (4.21b) over all admissible characters xk and yj, we obtain the average information
gain of the receiver:
I(X, Y) = E {I (xk, yj)}

=
=

I(x , y

j ). p(xk , y j )

k
j

p(x , y
k

j ).log

(4.24) we have:
1) I(X, Y) = p(xk , y j
k

2)

) log

p(xk , y j )

1
1
. log
p(x )
p(x | y

I(X, Y) = p(xk , y j ) [log

. log

From Eq

. (4.24)

p(xk )p(y j )

= H(X) H(X | Y)

(4.25)

p(y j )
p(y j | xk )
k j
= H(Y) H(Y | X)
.
1
1
. p(xk , y j ) log
3) I(X,Y) p(xk , y j )log

p(xk ) k j
p(y j )
k j
1
p(xk , y j ) log
p(xk y j )
K J
Or I(X, Y) = H(X) + H(Y) H(X, Y)
(4.27
..

(4.

26)

Further, in view of Eq.(4.18) & Eq.(4.19) we conclude that, even though for a particular received
symbol, yj, H(X) H(X | Yj) may be negative, when all the admissible ou tput symbols are covered, the
average mutual information is always non- negative. That is to say, we cannot loose information on an
average by observing the output of a channel. An easy method, of remembering the various relationships,
is given in Fig 4.2.Althogh the diagram resembles a Venn-diagram, it is not, and the diagram is only a
tool to remember the relationships. That is all. You cannot use this diagram for proving any result.

Dept. of ECE/SJBIT

Page 89

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

The entropy of X is represented by the circle on the left and that of Y by the circle on the right. The
overlap between the two circles (dark gray) is the mutual information so that the remaining (light
gray) portions of H(X) and H(Y) represent respective equivocations. Thus we have
H(X | Y) = H(X) I(X, Y) and H (Y| X) = H(Y) I(X, Y)
The joint entropy H(X,Y) is the sum of H(X) and H(Y) except for the fact that the overlap is added
twice so that
H(X, Y) = H(X) + H(Y) - I(X, Y)
Also observe H(X, Y) = H(X) + H (Y|X)
= H(Y) + H(X |Y)
For the JPM given in Example 4.1,

I(X, Y) = 0.760751505 bits / sym

Shannon Theorem: Channel Capacity:

Clearly, the mutual information I (X, Y) depends on the source probabilities apart from the
channel probabilities. For a general information channel we can always make I(X, Y) = 0 by choosing
any one of the input symbols with a probability one or by choosing a channel with independent input
and output. Since I(X, Y) is always nonnegative, we thus know the minimum value of the
Transinformation. However, the question of max I(X, Y) for a general channel is not easily answered.
Our intention is to introduce a suitable measure for the efficiency of the channel by making a
comparison between the actual rate and the upper bound on the rate of transmission of information.
Shannons contribution in this respect is most significant. Without botheration about the proof, let us
see what this contribution is.
Shannons theorem: on channel capacity(coding Theo rem)
It is possible, in principle, to device a means where by a communication system will transmit
information with an arbitrary small probability of error, provided that the information rate R(=rI
(X,Y),where r is the symbol rate) is less than or equal to a rate C called channel capacity.
The technique used to achieve this objective is called coding. To put the matter more
formally, the theorem is split into two parts and we have the following statements.

Dept. of ECE/SJBIT

Page 90

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

Positive statement:
Given a source of M equally likely messages, with M>>1, which is generating information at a
rate R, and a channel with a capacity C. If R C, then there exists a coding technique such that the
output of the source may be transmitted with a probability of error of receiving the message that can be
made arbitrarily small.
This theorem indicates that for R< C transmission may be accomplished without error even in
the presence of noise. The situation is analogous to an electric circuit that comprises of only pure
capacitors and pure inductors. In such a circuit there is no loss of energy at all as the reactors have the
property of storing energy rather than dissipating.
Negative statement:
Given the source of M equally likely messages with M>>1, which is generating information at a
rate R and a channel with capacity C. Then, if R>C, then the probability of error of receiving the message
is close to unity for every set of M transmitted symbols.
This theorem shows that if the information rate R exceeds a specified value C, the error probability
will increase towards unity as M increases. Also, in general, increase in the complexity of the coding
results in an increase in the probability of error. Notice that the situation is analogous to an electric
network that is made up of pure resistors. In such a circuit, whatever energy is supplied, it will be
dissipated in the form of heat and thus is a lossy network.
You can interpret in this way: Information is poured in to your communication channel. You
should receive this without any loss. Situation is similar to pouring water into a tumbler. Once the
tumbler is full, further pouring results in an over flow. You cannot pour water more than your
tumbler can hold. Over flow is the loss.
Shannon defines C the channel capacity of a communication channel a s the maximum
value of Transinformation, I(X, Y):
. (4.28)
C = Max I(X, Y) = Max [H(X) H (Y|X)]
The maximization in Eq (4.28) is with respect to all possible sets of probabilities that could be
assigned to the input symbols. Recall the maximum power transfer theorem: In any network,
maximum power will be delivered to the load only when the load and the source are properly
matched. The device used for this matching purpose, we shall call a transducer . For example, in a
radio receiver, for optimum response, the impedance of the loud speaker will be matched to the
impedance of the output power amplifier, through an output transformer.

This theorem is also known as The Channel Coding Theorem (Noisy Coding Theorem). It may
be stated in a different form as below:
R C or rs H(S) rc I(X,Y)Max or{ H(S)/Ts} { I(X,Y)Max/Tc}
If a discrete memoryless source with an alphabet S has an entropy H(S) and produces
symbols every T s seconds; and a discrete memoryless channel has a capacity I(X,Y)Max and is
used once every Tc seconds; then if

Dept. of ECE/SJBIT

Page 91

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

There exists a coding scheme for which the source output can be transmitted over the channel and
be reconstructed with an arbitrarily small probability of error. The parameter C/T c is called the
critical rate. When this condition is satisfied with the equality sign, the system is said to be
signaling at the critical rate.
Conversely, if

H ( S ) I ( X ,Y )Max

, it is not possible to transmit information over the T s Tc

channel and reconstruct it with an arbitrarily small probability of error

A communication channel, is more frequently, described by specifying the source
probabilities P(X) & the conditional probabilities P (Y|X) rather than specifying the JPM. The CPM,
P (Y|X), is usually refereed to as the noise characteristic of the channel. Therefore unless otherwise
specified, we shall understand that the description of the channel, by a matrix or by a Channel diagram
always refers to CPM, P (Y|X). Thus, in a discrete communication channel with pre-specified noise
characteristics (i.e. with a given transition probability matrix, P (Y|X)) the rate of information
transmission depends on the source that drives the channel. Then, the maximum rate corresponds to a
proper matching of the source and the channel. This ideal characterization of the source depends in turn
on the transition probability characteristics of the given channel.
Redundancy and Efficiency:
A redundant source is one that produces dependent symbols. (Example: The Markov
source). Such a source generates symbols that are not absolutely essential to convey information. As
an illustration, let us consider the English language. It is really unnecessary to write U following
the letter Q. The redundancy in English text is e stimated to be 50%(refer J Das etal, Sham
Shanmugam, Reza, Abramson, Hancock for detailed discussion.) This implies that, in the long run,
half the symbols are unnecessary! For example, consider the following sentence.
Y.u sh..ld b. abl. t. re.d t.is ev.n tho sev.r.l l.t..rs .r. m.s..ng
However, we want redundancy. Without this redundancy abbreviations would be impossible
and any two dimensional array of letters would form a crossword puzzle! We want redundancy even
in communications to facilitate error detection and error correction. Then how to measure
redundancy? Recall that for a Markov source, H(S) < H(S), where S is an ad- joint, zero memory
source. That is, when dependence creeps in, the entropy of the source will be reduced and this can be
used as a measure indeed!
The redundancy of a sequence of symbols is measured by noting the amount by which the entropy
has been reduced.
When there is no inter symbol influence the entropy at the receiver would be H(X) for any
given set of messages {X} and that when inter symbol influence occurs the entropy would be H (Y|X).
The difference [H(X) H (Y|X) ] is the net reduction in entropy and is called Absolute Redundancy.
Generally it is measured relative to the maximum entropy and thus we have for the Relative
Redundancy or simply, redundancy , E
E = (Absolute Redundancy) H(X)

Dept. of ECE/SJBIT

Page 92

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

E1

Or

10EC55

H(Y | X)
H(X)

. (

4.29)

Careful observation of the statements made above leads to the following alternative definition for
redundancy,
E1

R
C

(4.30)

Where R is the actual rate of Transinformation (mutual information) and C is the channel
capacity. From the above discussions, a definition for the efficiency, for the channel immediately
follows:

Actual rate of mutual information

maximum possible rate

That is.
and

R
C

1 E

. (4.31)

(4.32)

Capacity of Channels:
While commenting on the definition of Channel capacity, Eq. (4.28), we have said that
maximization should be with respect to all possible sets of input symbol probabilities. Accordingly,
to arrive at the maximum value it is necessary to use some Calculus of Variation techniques and the
problem, in general, is quite involved.
Example 3.2: Consider a Binary channel specified by the following noise characteristic (channel matrix):
1
P(Y | X ) 2

1
4

1
2
3
4

The source probabilities are: p(x1) = p, p(x2) = q =1-p

Clearly, H(X) = - p log p - (1 - p) log (1 - p)
We shall first find JPM and proceed as below:

P(X,Y)

p(x1 ). p(y1 | x1 )
p(x2). p(y1 | x

p(x1 ). p(y2 | x1 )

2
)

) p(x2). p(y2 | x2 )

p
2

1 p
4

p
2

3(1 p)
4

Adding column-wise, we get:

Dept. of ECE/SJBIT

Page 93

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

p (y1) = p 1 - p 1 p
2
4
4
1p
4

Hence H(Y) =

log

4
And H (Y|X) = p log 2
2

p (y2) = p 3(1 - p) 3 p
2
4
4
3-p
4

and

log

1p
4
p log 2
2

I(X, Y) = H(Y) H (Y|X) =

3-p
1 p log 4
4

3 log 3

3log3

3(1 p) log 4
4
3
1 p

log(1 p)

3(1 - p)

4
4
dI
Writing log x = loge ln x and setting
= 0 yields straight away:
dp
(4-3log3)

p 3a 1 0.488372093 , Where
1a

a =2

log (3 p )

= 0.592592593

With this value of p, we find I(X, Y) Max = 0.048821 bits /sym

For other values of p it is seen that I(X, Y) is less than I(X, Y) max
Although, we have solved the problem in a straight forward way, it will not be the case
p
I(X,Y)
Bits / sym

whe
n
the
dim
ension of the channel matrix is more than two. We have thus shown that the channel capacity of a
given channel indeed depends on the source probabilities. The computation of the channel capacity
would become simpler for certain class of channels called the symmetric or uniform channels.
. 0.2
.0.32268399

. 0.4
.0.04730118

.0.5
.0.04879494

.0.6
.0.046439344

.0.8
.0.030518829

Murogas Theorem :
The channel capacity of a channel whose noise characteristic, P (Y|X), is square and nonsingular, the channel capacity is given by the equation:
i n

C log 2

.33)

. (4

Qi

i 1

Where Qi are the solutions of the matrix equation P (Y|X).Q = [h], where
are the row entropies of P (Y|X).

p11

p p
12
13
p p

21
M

p
n1

22
M

p
n2

Dept. of ECE/SJBIT

23
M

p
n3

.....
.....
M

p Q1
1n
p

2n

p
.....

nn

Q
n

h= [h1, h2, h3, h4 h n]

h1
h

h
n

Page 94

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

n
C log

i 1

Q C

pi 2
pi

, where pi are obtained from: pi = p1.p1i + p2.p2i + p3.p3i + +p n.pni

p1i
p

p2 ..... pn . 2 i
M
p

ni

i-th column of P [Y|X]

Or

[p1, p2, p3 p n ] = [p1, p2, p3 p n] P [Y|X].

From this we can solve for the source probabilities (i.e. Input symbol probabilities):
-1

[p1, p2, p3 p n] = [p1, p2, p3 p n ] P [Y|X], provided the inverse exists.

However, although the method provides us with the correct answer for Channel capacity, this
-1
value of C may not necessarily lead to physically realizable values of probabilities and if P [Y|X]
does not exist ,we will not have a solution for Qi`s as well. One reason is that we are not able to
incorporate the inequality constraints 0 pi 1 .Still, within certain limits; the method is indeed very
useful.
Example 4.2: Consider a Binary channel specified by the following noise characteristic (channel matrix):
1
P(Y | X )

1
2
1

2
3

The row entropies are:

h 1 log 2 1 log 2 1 bit / symbol .
1
2
2
1
h log 4 3 log 4 0.8112781 bits / symbol .
2
4
4
3
3
1
1
P Y | X
1
2
Q
h
1.3774438
1

Y
|
X

.
1
1
Q
h
2

C log 2

Further ,
p

p
1

0.6225562

Q1

Q 2

2
p
2

0.048821 bits / symbol ,

Q1 C

.P 1

0.372093 and p 2Q2 C 0.627907.

Y|X

as before.

0.488372 0.511628

Giving us p = 0.488372

Dept. of ECE/SJBIT

Page 95

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Example 4.3:
Consider a 33 channel matrix as below:
PY | X

0.4
0.5

0.6
0

0 0.6

0
0.5
0.4

The row entropies are:

h1 = h3 = 0.4 log (1/0.4) + 0.6 log (1/0.6) = 0.9709505 bits / symbol.
h2 = 2 0.5 log (1/0.5) = 1 bit / symbol.
1.25
P

Y | X

56

-1

C = log {2 + 2
-Q1 C

p1 =2

2 3 5 6

1.25
1
1
1.0193633

Q1
2

1.25

-1.0193633

1.25

-1

+ 2 } = 0.5785369 bits / symbol.

=0.3348213 = p3, p2 = 2

-Q2 C

=0.3303574.

Therefore, p1 = p3 =0.2752978 and p2 = 0.4494043.

Suppose we change the channel matrix to:
PY | X

0.8
0.5

0.2
0
0 0.5

0.625
P 1Y | X
2.5

0 0.2 0.8

We have:

0.625

1
4

0.625
2.5

0.625

h1 = h3 =0.721928 bits / symbol, and h2 = 1 bit / symbol.

This results in:
Q1 = Q3 = 1; Q2 = 0.39036.
1

C = log {2 2
p1 =2

-Q1 C

Giving:

+0.39036

+2

} = 1.2083427 bits / symbol.

-Q2 C

=0.2163827 = p3, p2 = 2

=0.5672345

p1 = p3 =1.4180863 and p2 = Negative!

Dept. of ECE/SJBIT

Page 96

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Thus we see that, although we get the answer for C the input symbol probabilities computed are not
physically realizable. However, in the derivation of the equations, as already pointed out, had we
included the conditions on both input and output probabilities we might have got an excellent result!
But such a derivation becomes very formidable as you cannot arrive at a numerical solution! You will
have to resolve your problem by graphical methods only which will also be a tough proposition! The
formula can be used, however, with restrictions on the channel transition probabilities. For example,
in the previous problem, for a physically realizable p1, p11 should be less than or equal to 0.64.
(Problems 4.16 and 4.18 of Sam Shanmugam to be solved using this method)
Symmetric Channels:
The Murogas approach is useful only when the noise characteristic P [X|Y] is a square and
invertible matrix. For channels with m n, we can determine the Channel capacity by simple
inspection when the channel is Symmetric or Uniform.
Consider a channel defined by the noise characteristic:
p11

p
12
p

p
13
p

22

23

21

P[Y | X ] p31

32

33

...

1n

...

2n

(4.34)

...

3n

nn
...
This channel is said to be Symmetric or Uniform if the second and subsequent rows of the
channel matrix are certain permutations of the first row. That is the elements of the second and
subsequent rows are exactly the same as those of the first row except for their locations. This is
illustrated by the following matrix:
n1

n2

n3

p2

p3

...

pn

n 1

p2
p2

pn
p1

...
...

p4
p5

...

p1

p1
P [Y | X ] p3
M

pn

n 1

n 2

(4.35)

Remembering the important property of the conditional probability matrix, P [Y|X], that the sum of
all elements in any row should add to unity; we have:
n

pj 1

j 1

(4.36)

The conditional entropy H (Y|X) for this channel can be computed from:
H(Y|X)

m n

p( xk , y j

k 1 j 1
m

p( xk ). p( y j | xk )log

)log

1
p( xk , y j )
1

p( y j | xk )
However, for the channel under consideration observe that:
k 1

Dept. of ECE/SJBIT

j 1

Page 97

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

p( xk ). p( y j | xk )log

p j log

h (4.37)

p( y j | xk )
pj
k 1
j 1
j 1
is a constant, as the entropy function is symmetric with respect to its arguments and depends only on
the probabilities but not on their relative locations. Accordingly, the entropy becomes:
m

H (Y | X ) p( xk ).h h

..(4.38)

k 1

as the source probabilities all add up to unity.

Thus the conditional entropy for such type of channels can be computed from the elements of any
row of the channel matrix. Accordingly, we have for the mutual information:
I(X, Y) = H(Y) H (Y|X)
= H(Y) h
Hence, C = Max I(X, Y) =Max
{H(Y) h} =
Max H(Y) h
Since, H(Y) will be maximum if and only if all the received symbols are equally probable and as
there are n symbols at the output, we have:
H(Y) Max = log n
Thus we have for the symmetric channel:
C = log n h

(4.39)

The channel matrix of a channel may not have the form described in Eq (3.35) but still it can
be a symmetric channel. This will become clear if you interchange the roles of input and output. That
is, investigate the conditional probability matrix P (X|Y).
We define the channel to be symmetric if the CPM, P (X|Y) has the form:
p1

p2

...

p4

p6
pm

...
...

pm

p1

...

p1

pm

p2
P( X | Y ) p3

m 1

m 3

pm
p

m 1

.(4.40)

m 2

That is, the second and subsequent columns of the CPM are certain permutations of the first column.
In other words entries in the second and subsequent columns are exactly the same as in the first
column but for different locations. In this case we have:
n

H ( X | Y ) p( xk , y j )log

p( y j ) p( xk | y j )log

j 1 k 1

p( xk | y j ) j 1
p( xk | y j )
k 1
m
1
1
h is a constant, because
p log
Since p( y j ) 1 and
p( x k | y j )log
j 1
k 1
p( xk | y j ) k 1
pk
all entries in any column are exactly the same except for their locations, it then follows that:
n

Dept. of ECE/SJBIT

Page 98

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H ( X | Y ) h p log

(4.41)

pk
k 1
*Remember that the sum of all entries in any column of Eq (3.40) should be unity.
As a consequence, for the symmetry described we have:
C = Max [H(X) H ( X|Y)] = Max H(X) - h
Or

C = log m - h

(4.42)

Thus the channel capacity for a symmetric channel may be computed in a very simple and
straightforward manner. Usually the channel will be specified by its noise characteristics and the
source probabilities [i.e. P (Y|X) and P(X)]. Hence it will be a matter of simple inspection to identify
the first form of symmetry described. To identify the second form of symmetry you have to first
compute P (X|Y) tedious!
Example 4.4:
Consider the channel represented by the channel diagram shown in Fig 3.3:
The channel matrix can be read off from the channel diagram as:
1
1 1
1
P( Y | X )

3 3 6 6
1 1 1 1
6
6 3
3

Clearly, the second row is a permutation of the first row (written in the reverse order) and hence the
channel given is a symmetric channel. Accordingly we have, for the noise entropy, h (from either of
the rows):
H (Y|X) = h =2 1 log 3 + 2 1 log 6 = 1.918295834 bits / symbol.
3
6
C = log n h = log 4 h =0.081704166 bits / symbol.
Example 4.5:
A binary channel has the following noise characteristic:

Dept. of ECE/SJBIT

Page 99

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

0
0 2

(a) If the input symbols are transmitted with probabilities 3 / 4 and 1 / 4 respectively, find H(X),
H(Y), H(X, Y), H (Y|X) and I(X, Y).
(b) Find the channel capacity, efficiency and redundancy of the channel.
(c) What are the source probabilities that correspond to the channel capacity?
To avoid confusion, let us identify the input symbols as x1 and x2 and the output symbols by y1 and
y2. Then we have:
P(x1) = 3 /4 and p(x2) = 1 / 4
2

P( X | Y ) 3

3
2

1
3

H ( Y | X ) h 2 log 3 1 log 3 log 3 2 0.918295833 bits / symbol .

3
2 3
3
H(X)

4 1

log
3444

log4 log4

log 3 2

log 3 0.811278125 bits / symbol . 4

Multiplying first row of P (Y|X) by p(x1) and second row by p(x2) we get:
2
P( X ,Y ) 3

4
1 1
3 4

3 4
2 3
3 4

2
1
12

1
4
1
6

Adding the elements of this matrix columnwise, we get: p (y1) = 7/12, p (y2) = 5/12.
Dividing the first column entries of P (X, Y) by p (y1) and those of second column by
p (y2), we get:
6 3
P( X | Y )

7
From these values we have:
H(Y)

12

log

12 5

5
log

12

0.979868756 bits / symbol .

12
5
H ( X ,Y ) 1 log 2 1 log4 1 log12 1 log6 1.729573958 bits / symbol .
12
6
2
4
1
7
1
5
1
H ( X | Y ) log log
log7 1 log 5 0.74970520 bits / symbol
2
6 4
3 12
6
2

Dept. of ECE/SJBIT

Page 100

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H ( Y | X ) 1 log 3 1 log 3 1 log 3 1 log 3 log 3 2 h ( asbefore ).

12
6
2
3
2
2
4
I ( X ,Y ) H ( X ) H ( X | Y ) 0.061572924 bits / symbol .
H ( Y ) h 0.061572923 bits / symbol .
C log n h log 2 h 1 h 0.081704167 bits / symbol .
Efficiency , I ( X ,Y ) 0.753608123 or 75.3608123%
C
Re dundancy , E 1 0.246391876 or 24.6391876%
To find the source probabilities, let p(x1) = p and p(x2) = q = 1 p .Then the JPM becomes:
2 p,
1p
3
3
P( X ,Y )
1 ( 1 p ), 2
3
3
Adding columnwise we get: p( y1 ) 1 (1 p) and
3

p( y2) 1 (2 p)
3

For H(Y) = H(Y) max, we want p (y1) = p (y2) and hence 1+p = 2-p or

p 1
2

Therefore the source probabilities corresponding to the channel capacity are: p(x1) =1/2 = p(x2).

Binary Symmetric Channels (BSC): (Problem 2.6.2 S imon Haykin)

The channel considered in Example 3.6 is called a Binary Symmetric Channel or ( BSC). It
is one of the most common and widely used channels. The channel diagram of a BSC is shown in Fig
3.4. Here p is called the error probability.
For this channel we have:
H (Y | X ) p log 1 q log 1 H ( p)
p
q
1
H (Y ) [ p ( p q)]log

(4.43)
[q ( p q)]log

[ p ( p q)]
I(X, Y) = H(Y) H (Y|X) and the channel capacity is:
C=1 + p log p +q log q

(4.44)

[q ( p q)]
(4.45)

This occurs when = 0.5 i.e. P(X=0) = P(X=1) = 0.5

In this case it is interesting to note that the equivocation, H (X|Y) =H (Y|X).

Dept. of ECE/SJBIT

Page 101

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

An interesting interpretation of the equivocation may be given if consider an idealized

communication system with the above symmetric channel as shown in Fig 4.5.

The observer is a noiseless channel that compares the transmitted and the received symbols.
Whenever there is an error a 1 is sent to the receiver as a correction signal and appropriate
correction is effected. When there is no error the observer transmits a 0 indicating no change. Thus
the observer supplies additional information to the receiver, thus compensating for the noise in the
channel. Let us compute this additional information .With P (X=0) = P (X=1) = 0.5, we have:
Probability of sending a 1 = Probability of error in the channel .
Probability of error = P (Y=1|X=0).P(X=0) + P (Y=0|X=1).P(X=1)
= p 0.5 + p 0.5 = p
Probability of no error = 1 p = q
Thus we have P (Z = 1) = p and P (Z = 0) =q
Accordingly, additional amount of information supplied is:
p log

1
1
q log H ( X | Y ) H ( Y | X )
p
q

.. (4.46)

Thus the additional information supplied by the observer is exactly equal to the equivocation of the
source. Observe that if p and q are interchanged in the channel matrix, the trans -information of
the channel remains unaltered. The variation of the mutual information with the probability of error is

Dept. of ECE/SJBIT

Page 102

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

shown in Fig 3.6(a) for P (X=0) = P (X=1) = 0.5. In Fig 4.6(b) is shown the dependence of the mutual
information on the source probabilities.

4.4.4 Binary Erasure Channels (BEC):(Problem 2.6.4 Simon Haykin)

The channel diagram and the channel matrix of a BEC are shown in Fig 3.7.

BEC is one of the important types of channels used in digital communications. Observe that
whenever an error occurs, the symbol will be received as y and no decision will be made about the
information but an immediate request will be made for retransmission, rejecting what have been received
(ARQ techniques), thus ensuring 100% correct data recovery. Notice that this channel also is a symmetric
channel and we have with P(X = 0) =, P(X = 1) = 1 - .
1
1
qlog
p
q
1
H ( X ) log ( 1 )log

H (Y | X ) plog

(4.47)

(4.48)
.
(1)
The JPM is obtained by multiplying first row of P (Y|X) by and second row by (1 ).
We get:
q
p
0
..
(4.49)
P( X ,Y )

p( 1 ) q( 1 )
Adding column wise we get: P (Y) = [q, p, q (1 )]
From which the CPM P (X|Y) is computed as:

Dept. of ECE/SJBIT

(4.50)

Page 103

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

P( X | Y )

10EC55

1
(1)
H ( X | Y ) q log1 p log 1 ( 1 ) p log

(4.51)

( 1 )q log1

(1)

pH ( X )
I(X, Y) =H(X) H (X|Y) = (1 p) H(X) = q H(X)
C = Max I (X, Y) = q bits / symbol.

(4.52)
(4.53)

In this particular case, use of the equation I(X, Y) = H(Y) H(Y | X) will not be correct, as H(Y)
involves y and the information given by y is rejected at the receiver.
Deterministic and Noiseless Channels: (Additional Information)
Suppose in the channel matrix of Eq (3.34) we make the following modifications.
a) Each row of the channel matrix contains one and only one nonzero entry, which necessarily
should be a 1. That is, the channel matrix is symmetric and has the property, for a given k
and j, P (yj|xk) = 1 and all other entries are 0. Hence given xk, probability of receiving it as yj is
one. For such a channel, clearly
H (Y|X) = 0 and I(X, Y) = H(Y)
. (4.54)
Notice that it is not necessary that H(X) = H(Y) in this case. The channel with such a property will
be called a Deterministic Channel.
Example 4.6:
Consider the channel depicted in Fig 3.8. Observe from the channel diagram shown that the
input symbol xk uniquely specifies the output symbol yj with a probability one. By observing the
output, no decisions can be made regarding the transmitted symbol!!

b) Each column of the channel matrix contains one and only one nonzero entry. In this case,
since each column has only one entry, it immediately follows that the matrix P(X|Y) has also
one and only one non zero entry in each of its columns and this entry, necessarily be a 1
because:
If p (yj|xk) =, p (yj | xr) = 0, r k, r = 1, 2, 3 m.
Then p (xk, yj) = p (xk) p (yj|xk) = p (xk),

Dept. of ECE/SJBIT

Page 104

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

p (xr, yj) = 0, r k, r = 1, 2, 3 m.
m

p (yj) = p( xr , y j ) = p (xk, yj) = p (xk)

r 1

p( xk | y j )

p( xk , y j )

1, and p( xr | y j ) 0 , r k ,r 1,2 ,3,...m .

p( y j )
It then follows that H (X|Y) = 0 and I (X, Y) = H(X)

... (4.55)

Notice again that it is not necessary to have H(Y) = H(X). However in this case, converse of (a)
holds. That is one output symbol uniquely specifies the transmitted symbol, whereas for a given
input symbol we cannot make any decisions about the received symbol. The situation is exactly
the complement or mirror image of (a) and we call this channel also a deterministic channel
(some people call the channel pertaining to case (b) as Noiseless Channel, a classification can be
found in the next paragraph). Notice that for the case (b), the channel is symmetric with respect to
the matrix P (X|Y).
Example 4.7:
Consider the channel diagram, the associated channel matrix, P (Y|X) and the conditional
probability matrix P (X|Y) shown in Fig 3.9. For this channel, let
p (x1)=0.5, p(x2) = p(x3) = 0.25.
Then p (y1) = p (y2) = p(y6) =0.25, p(y3) = p(y4) =0.0625 and p(y5) = 0.125.
It then follows I(X, Y) = H(X) =1.5 bits / symbol,
H(Y) = 2.375 bits / symbol, H (Y|X) = 0.875 bits / symbol and H (X|Y) = 0.

c) Now let us consider a special case: The channel matrix in Eq (3.34) is a square matrix and all
entries except the one on the principal diagonal are zero. That is:
p (yk|xk) = 1 and p(yj|xk)=0kj
Or in general, p (yj|xk) =jk, where jk, is the Kronecker delta, i.e. jk =1 if j = k
=0 if j k.
That is, P (Y|X) is an Identity matrix of order n and that P (X|Y) = P (Y|X) and
p(xk, yj) = p(xk) = p(yj) can be easily verified.

Dept. of ECE/SJBIT

Page 105

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

For such a channel it follows:

(4.56)

H (X|Y) = H (Y|X) = 0 and I(X, Y) = H(X) = H(Y) = H(X, Y) ....

We call such a channel as Noiseless Channel. Notice that for the channel to be noiseless, it
is necessary that there shall be a one-one correspondence between input and output symbols. No
information will be lost in such channels and if all the symbols occur with equal probabilities, it
follows then:
C =I(X, Y) Max=H(X) Max=H(Y) Max=log n bits / symbol.
Thus a noiseless channel is symmetric and deterministic with respect to both descriptions
P (Y|X) and P (X|Y).
Finally, observe the major concept in our classification. In case (a) for a given transmitted
symbol, we can make a unique decision about the received symbol from the source end. In case
(b), for a given received symbol, we can make a decision about the transmitted symbol from the
receiver end. Whereas for case (c), a unique decision can be made with regard to the transmitted
as well as the received symbols from either ends. This uniqueness property is vital in calling the
channel as a Noiseless Channel.
d) To conclude, we shall consider yet another channel described by the following JPM:
p1
p1 p1 ... p1

p2

P( X ,Y )

p
with

pk

k 1

p2

...

pm p m

m
m

p2

1
n

...

i .e . p( yj )

p2
M

1
, j 1,2 ,3 ,...n.
n

This means that there is no correlation between xk and yj and an input xk may be received as any
one of the yjs with equal probability. In other words, the input-output statistics are independent!!
This can be verified, as we have p (xk, yj) = pk

=npk. pk = p (xk).p (yj)

k 1

p(xk|yj) = npk and p(yj|xk) = 1/n

Thus we have:
m

pk log, H ( X ) npk log

H ( X ,Y ) n.

k 1
n

pk
1

1
n

k 1

npk

pk log

k 1

1
log

pk

logn,
p( y j )
H ( X | Y ) H ( X ), H ( Y | X ) H ( Y ) and I ( X ,Y ) 0
H ( Y ) p( y j

)log

j 1

Dept. of ECE/SJBIT

Page 106

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

Such a channel conveys no information whatsoever. Thus a channel with independent inputoutput structure is similar to a network with largest internal loss (purely resistive network), in
contrast to a noiseless channel which resembles a lossless network.

Some observations:
For a deterministic channel the noise characteristics contains only one nonzero entry,
which is a 1, in each row or only one nonzero entry in each of its columns. In either case there
exists a linear dependence of either the rows or the columns. For a noiseless channel the rows as
well as the columns of the noise characteristics are linearly independent and further there is only
one nonzero entry in each row as well as each column, which is a 1 that appears only on the
principal diagonal (or it may be on the skew diagonal). For a channel with independent inputoutput structure, each row and column are made up of all nonzero entries, which are all equal and
equal to 1/n. Consequently both the rows and the columns are always linearly dependent!!
Franklin.M.Ingels makes the following observations:
1) If the channel matrix has only one nonzero entry in each column then the channel is termed
as loss-less channel. True, because in this case H (X|Y) = 0 and I(X, Y) =H(X), i.e. the
mutual information equals the source entropy.
2) If the channel matrix has only one nonzero entry in each row (which necessarily should be a
1 ), then the channel is called deterministic channel. In this case there is no ambiguity
about how the transmitted symbol is going to be received although no decision can be made
from the receiver end. In this case H (Y|X) =0, and I(X, Y) = H(Y).
3) An Ideal channel is one whose channel matrix has only one nonzero element in each row
and each column, i.e. a diagonal matrix. An ideal channel is obviously both loss-less and
deterministic. Lay mans knowledge requires equal number of inputs and outputs-you
cannot transmit 25 symbols and receive either 30 symbols or 20 symbols, there shall be no
difference between the numbers of transmitted and received symbols. In this case
I(X,Y) = H(X) =H(Y); and H(X|Y) =H(Y|X) =0
4) A uniform channel is one whose channel matrix has identical rows ex cept for
permutations OR identical columns except for permutations. If the channel matrix is square,
then every row and every column are simply permutations of the first row.
Observe that it is possible to use the concepts of sufficient reductions and make the
channel described in (1) a deterministic one. For the case (4) observe that the rows and
columns of the matrix (Irreducible) are linearly independent.
Additional Illustrations:
Example 4.8:

Dept. of ECE/SJBIT

Page 107

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Consider two identical BSCs cascaded as shown in Fig 4.10. Tracing along the transitions indicated
we find:
2

p (z1|x1) = p + q = (p + q) 2pq =(1 2pq) = p(z 2|x2) and p(z1|x2) = 2pq = p(z2|x1)
Labeling p 1 2 pq , q 2 pq it then follows that:
I(X, Y) = 1 H (q) =1 + p log p + q log q
I(X, Z) = 1 H (2pq) = 1 + 2pq log 2pq + (1 2pq) log (1 2pq).
If one more identical BSC is cascaded giving the output (u1, u2) we have:
2
3
I(X, U) = 1 H (3pq + p )
The reader can easily verify that I(X, Y) I(X, Z) I(X, U)
Example 4.9:
Let us consider the cascade of two noisy channels with channel matrices:
1
1

1
6

P( Y | X )

1
2

6
1

2
3

1
4

P( Z | Y )

2
1
3
0

1
2
2
3
1
3

0
0

, with p(x1) = p(x2) =0.5

2
3

The above cascade can be seen to be equivalent to a single channel with channel matrix:

Dept. of ECE/SJBIT

Page 108

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

5
P( Z | X )

36 12 9
1
1 1
3
6
2

The reader can verify that: I(X, Y) = 0.139840072 bits / symbol.

I(X, Z) = 0.079744508 bits / symbol.
Clearly I(X, Y) > I(X, Z).
Example 4.10: Let us consider yet another cascade of noisy channels described by:
1

1
3
0

P(Y | X )

3 3
1 1
2
2

1
P( Z | Y ) 0

0
2

0
1

3
1

3
2

The channel diagram for this cascade is shown in Fig 4.12. The reader can easily verify in this case
that the cascade is equivalent to a channel described by:

P( Z | X )

1
2

1
3 P(Y | X ) ;

1
2

Inspite of the fact, that neither channel is noiseless, here we have I(X, Y) = I(X, Z).

Dept. of ECE/SJBIT

Page 109

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Review Questions:
1. What are important properties of the codes?
2. what are the disadvantages of variable length coding?
3. Explain with examples:
4. Uniquely decodable codes, Instantaneous codes
5. Explain the Shannon-Fano coding procedure for the construction of an optimum code
6. Explain clearly the procedure for the construction of compact Huffman code.
7. A discrete source transmits six messages symbols with probabilities of 0.3, 0.2, 0.2, 0.15, 0.1,
0.5. Device suitable Fano and Huffmann codes for the messages and determine the average
length and efficiency of each code.
8. Consider the messages given by the probabilities 1/16, 1/16, 1/8, , . Calculate H. Use the
Shannon-Fano algorithm to develop a efficient code and for that code, calculate the average
number of bits/message compared with H.
9. Consider a source with 8 alphabets and respective probabilities as shown:
A B C DE F GH
0.20 0.18 0.15 0.10 0.08 0.05 0.02 0.01
Construct the binary Huffman code for this. Construct the quaternary Huffman and code
and show that the efficiency of this code is worse than that of binary code
10. Define Noiseless channel and deterministic channel.
11. A source produces symbols X, Y,Z with equal probabilities at a rate of 100/sec. Owing to
noise on the channel, the probabilities of correct reception of the various symbols are as
shown:
P (j/i) X
Y
z
X

Determine the rate at which information is being received.

12. Determine the rate of transmission l(x,y) through a channel whose noise characteristics is
shown in fig. P(A1)=0.6, P(A2)=0.3, P(A3)=0.1
A1

0.5

B1

T
R
A2

0.5

0.5
B2
0.5

Dept. of ECE/SJBIT

Page 110

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Unit 4
Syllabus:
Channel coding theorem, Differential entropy and mutual information for continuous
ensembles, Channel capacity Theorem. 6 Hours

Text Books:
Digital and analog communication systems, K. Sam Shanmugam, John Wiley, 1996.
Digital communication, Simon Haykin, John Wiley, 2003.

Reference Books:
ITC and Cryptography, Ranjan Bose, TMH, II edition, 2007
Digital Communications - Glover and Grant; Pearson Ed. 2nd Ed 2008

Dept. of ECE/SJBIT

Page 111

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Unit 4
CONTINUOUS CHANNELS
Until now we have considered discrete sources and discrete channels that pertain to a digital
communication system (pulse code modulation). Although modern trend is to switch over to digital
communications, analog communications can never become obsolete where amplitude and frequency
modulations are used (Radio, television etc). Here the modulating signal X (t) (which is the set of
messages to be transmitted from an information theoretical view point) is invariably a continuous
speech or picture signal. This message can be treated as equivalent to a continuous sample space
whose sample points form a continuum, in contrast to the discrete case. We shall define a continuous
channel as one whose input is a sample point from a continuous sample space and the output is a
sample point belonging to either the same sample space or to a different sample space. Further we
shall define a zero memory continuous channel as the one in which the channel output statistically
depends on the corresponding channels without memory. In what follows, we briefly discuss the
definition of information and entropy for the continuous source, omitting the time dependence of the
messages for brevity and conclude with a discussion of Shannon-Hartley law.
Entropy of continuous Signals: (Differential entropy):
For the case of discrete messages, we have defined the entropy as
H(S)

p( sk )log

.. (5.1)

p( sk )

k 1

For the case of analog or continuous messages, we may wish to extend Eq. (5.1), considering the
analog data to be made up of infinite number of discrete messages. However, as P(X = x) = 0 for a
CRV, X, then direct extension of Eq. (5.1) may lead to a meaningless definition. We shall proceed as
follows:
Let us consider that the continuous source X(a continuous random variable) as a limiting
form of a discrete random variable that assumes discrete values 0, x, 2x,..,etc. Let k.x =xk,
then clearly xk = x. The random variable X assumes a value in the range (xk, xk+x) with
probability f (xk) xk, (recall that P{x X x+dx} = f(x).dx, the alternative definition of the p.d.f. of
an r.v). In the limit as xk 0, the error in the approximation would tend to become zero.
Accordingly, the entropy of the CRV, X, is given by, using Eq. (5.1),

H(X)

lim

lim

xk 0 k

f ( xk )xk log

xk 0 k

Dept. of ECE/SJBIT

f ( xk )xk

log

1
f ( x k ) x k

1
f ( xk )

f ( xk

)xk

log xk

Page 112

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

As xk 0 it follows: xk x and f (xk) xk f(x) dx. The summation would be replaced by

integration, and thus we have:
H(X)

lim
dx
log x
xk 0
f(x)
k

f ( x ) log

f ( x ) log

f(x)

dx

lim

x 0

f ( x )dx

log x

(5.2)

As xk = x, and the second integral in the RHS is unity (property of any density function). Notice
that in the limit as x 0, log x - . Thus, it appears that the entropy of a continuous random
variable is infinite. Indeed it is so. The amount of uncertainty associated with a continuous message is
infinite as a continuous signal assumes uncountab ly infinite values. Thus it seems there can be no
meaningful definition of entropy for a continuous source (notice that we are using the words source,
message and random variable interchangeably as we are referring to statistical data that is being
generated by the transmitter). However, the anomaly can be solved if we consider the first term in the
RHS of Eq. (5.2) as a relative measure while -log x serves as our reference. Since we will be
dealing, in general, with differences in entropies (for example I(X, Y) = H(X) H (X|Y) ), if we select
the same datum for all entropies concerned, the relative measure would be indeed quite meaningful.
However, caution must be exercised to remember that it is only a relative measure and not an
absolute value. Otherwise, this subtle point would lead to many apparent fallacies as can be observed
by the following example. In order to differentiate from the ordinary absolute entropy we call it
Differential entropy. We then define the entropy. H(X), of a continuous source as

1 dx bits / sample .
f(x)

H ( X ) f ( x ) log

(5.3)

Where f(x) is the probability density function (p.d.f.) of the CRV, X.

Example 5.1: Suppose X is a uniform r.v. over the interval (0, 2). Hence
f(x) = 1
2
=0

0x2
Else where
2

Then using Eq. (5.3), we have H ( X ) log 2 dx 1bit / sample

0 2

Suppose X is the input to a linear amplifier whose gain = 4. Then the output of the amplifier would be
Y = 4X. Then it follows f(y) = 1/8, 0 y 8. Therefore the entropy H(Y), of the amplifier output
8
1
becomes H ( Y ) log 8 dy 3bits / sample

That is the entropy of the output is thrice that of the input! However, since knowledge of X uniquely
determines Y, the average uncertainties of X and Y must be identical. Definitely, amplification of a
signal can neither add nor subtract information. This anomaly came into picture because we did not
bother about the reference level. The reference entropies of X and Y are:

Dept. of ECE/SJBIT

Page 113

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

lim

lim
( log x ) and Ry
( log y )
x 0
y 0
lim
Rx R y x 0 ( log x log y ) log dy log 4 2bits / sample .
dx
y 0
Rx

Clearly, the reference entropy of X, Rx is higher than that of Y, R y. Accordingly, if X and Y have
equal absolute entropies, then their relative entropies must differ by 2 bits. We have:
Absolute entropy of X = Rx + H(X)
Absolute entropy of Y = Ry + H(Y)
Since Rx = Ry + 2, the two absolute entropies are indeed equal. This conclusion is true for any reverse
operation also. However, the relative entropies will be, in general, different.
To illustrate the anomalies that one may encounter if the relative characterization of the entropy
function is ignored we shall consider another example.
Example 5.2:
Suppose X is uniformly distributed over the interval (1/4, 1/4) .Then
f (x) = 2 .
1
4

H(X)2
1
4

-1/4 x 1/4 and 0 else where. It then follows from Eq (5.3):

log 1dx log 2 1bit / sample .
2

which is negative and very much contradictory to our concept of information?

Maximization of entropy:
For the case of discrete messages we have seen that entropy would become a maximum when
all the messages are equally probable. In practical systems, the sources, for example, radio
transmitters, are constrained to either average power or peak power limitations. Our objective, then, is
to maximize the entropy under such restrictions. The general constraints may be listed as below:

1) Normalizing constraint: f ( x )dx 1 (Basic property of any density function)

f ( x )dx 1

2) Peak value limitation:

3)

Average value limitation x . f ( x )dx E{ X } , is a constant.

4)

Average power limitation x . f ( x )dx m2 E{ X } , is a constant.

Dept. of ECE/SJBIT

Page 114

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

5) Average power limitation, with unidirectional distribution (causal systems whose response

does not begin before the input is applied) x . f ( x )dx , a constant.

0

Our interest then is to find the p.d.f., f(x) that maximizes the entropy function H(X) defined in
Eq. (5.3). The maximum entropy method (MEM) we are speaking of then is a problem of constrained
optimization and is a particular case of the so-called isoperimetric problem of the calculus of
variations. We shall adopt, here, the Euler- Lagranges method of undetermined co-efficients to
solve the problem on hand. Suppose, we wish to maximize the integral
b

I ( x . f )dx

(5.4)

(5.5)

Subject to the following integral constraints:

b

1 ( x , f )dx 1
a
b

2 ( x , f )dx 2
a

M
M
b

( x , f )dx r

Where 1, 2 r are pre-assigned constants. Then the form of f(x) that satisfies all the above
constraints and makes I a maximum (or minimum) is computed by solving the equation

1 1 2 2 ... r
f

. (5.6)

The undetermined co-efficients, 1, 2 r are called the Lagrangian multipliers (or simply,
Lagrangians) are determined by substituting the value of f(x) in Eq (5.5) successively. (Interested
reader can refer standard books on calculus of variations for such optimization problems).We shall
consider the cases we have already listed and determine the p.d.f, f(x) of the random signal X for these
cases.
Since differentiation and integration of log f(x) is not possible directly, we convert it into
logarithm to base e and proceed as below:
a

log2f = log e. ln f = ln f = ln f , where a=log e, a constant and eventually we will be finding f that
maximizes Eq (5.7) or equivalently f, if we call H1(x) = - f(x) ln f(x) dx, then H(X) = (log e) H1(X). So
if we maximize H1 (X), we will have achieved our goal.
Case I: Peak Signal Limitation:
Suppose that the signal is peak limited to M (Equivalent to peak power limitation, Sm, as in the case
of AM, FM and Pulse modulation transmitters. For example in FM, we use limiting circuits as our
concern is only on frequency deviation.). Then we have,

Dept. of ECE/SJBIT

Page 115

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H1 ( X ) f ( x ) log f ( x )dx

. (5.7)

And the only constraint on f(x) being the unit area condition (property of a p.d.f)
M

f ( x )dx

Here (x, f) = -f ln f

(5.8)

( 1 ln f )

1(x, f) = f

=1

And from Eq. (5.6), we have

- (1+ln f) +=0 f=e

- (1- )

Substituting Eq (5.10) in Eq (5.8), we get e

- (1- )

(5.9)

5.10)

=1/2M =f(x) .Thus it follows

f (x)= 1/2M , -M x M

. (5.11)

You will immediately identify Eq (5.11) as the uniform density function. Hence the entropy, under
peak signal limitation condition, will become a maximum if and only if the signal has a rectangular or
uniform distribution. The maximum value of entropy can be found by substituting Eq (5.11) in Eq
(5.3) as:
H(X) max=log2M bits/sample.

..

If Sm =M , then 2M= 4 Sm and H(X) max= 1 log4Sm bits/sample.

2

(5.12)

(5.13)

Suppose the signal is band limited to B Hz and is sampled at the Nyquist rate i.e.
r=2B samples/sec., then we have
R(X) max=B log 4Sm bits/sec.

(5.14)

For the uniform distribution over (-M, M), we have

2

Mean = 0, and Variance, = (2M) /12 = M /3

2

Since the mean value is zero, the variance can be treated as the average power, S and since M =
Sm, it follows Sm= 3S and Eq (4.14) becomes:
R(X) max=B log 12S bits/sec

5.15)

Case II Average Signal Limitation:

Dept. of ECE/SJBIT

Page 116

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Suppose X has unidirectional distribution with a specified average value (For example: PAM, PLM
or in AM with average carrier amplitude limitation), the constraint equations are:

f ( x )dx 1

(5.16a)

And xf ( x )dx ( say )

. (5.16b)

The equation to be solved now is

1 2
1

Where 1(x, f) = f and 2(x, f) =x. f

This leads tof e

( 1 x )
1

( 1 )
1

.e

ax

(5.17)

..

Where a = -2>0. This step is needed to make the integral converge as you will see later.
Using Eq (5.17) in Eq (5.16a) gives e

-(1-1)

=a. Substituting in Eq (5.17) back results in

-ax

f(x) = ae

Substituting Eq (5.18) in Eq (4.16b) gives

(5.18)

(5.19)

a=1/

Thus the p.d.f that maximizes H(x) is

1
f ( x ) e

, x>0

Which you will recognize as the exponential density function with parameter 1/.The maximum
value of the entropy can be shown as:
Hmax(X) = loge + log =loge bits/sample.

(5.2 0)

The rate of information transmission over a band width of B, assuming the signal is sampled at the
Nyquist rate (i.e. =2B samples/sec.) is:
Rmax =2Blog e bits/sec.

(5.21)

Case III Average Power Limitation (Symmetrical Distribution):

Examples are Random noise with specified variance, audio frequency telephony and similar
situations.
The constraint equations assuming zero mean value are now:

f ( x )dx 1

5.22a)

Dept. of ECE/SJBIT

Page 117

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

. f ( x )dx

(5.22b)

Hence 1(x, f) = f and 2(x, f) =x . f

Setting

1 2 2 0 , we get as a first solution:

1

( 1 x2 )

fe

( 1 )

.e

ax2
.. (5.23)

Where we have set a= -2>0.Substituting Eq (5.23) in Eq (5.22a), the second solution is

a eax 2

.. (5.24)

Use has been made of the formula e

dx

(From properties of Gamma functions.)Final

solution is obtained by substituting Eq (5.24) in Eq (5.22b) and converting the integral into the form
of a Gamma integral. Thus we have:

. e

ax

dx

2 ax

2-ax2

dx , as the integrand, x e

, is an even function.

Letting y=ax and manipulating, this integral reduces to:

3

2
1 1 2 y
1 2
y e dy
y
a 0
a 0

dy

1 3 1 1 1
2
2
2
2

Thus a = 1/2 and the required density function is

x
1
f ( x ) 2 e

2
2

. (5.25)
2

Eq (5.25) is the Gaussian or Normal density function corresponding to N (0, ) and gives maximum
entropy. We have:

H ( X )max f ( x ) log{

ex

log

2 f ( x )dx

/ 2

log e
2

}dx

x2 f ( x )dx

2
1 log 2 1 log e 1 log 2e bits / sample
2
2
2

In simplifying the integral (not evaluating) use is made of Eq (5.22a) and Eq (5.22b). Thus:

Dept. of ECE/SJBIT

Page 118

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H(X) max = (1/2) log (2e ) bits/sample.

(5.26)

The rate of information transmission over a band width of B assuming the signal is sampled at the
Nyquist rate (i.e. =2Bsamples/sec.) is:
2

R(X) max = B log (2e ) bits/sec.

(5.27)

Since represents the signal power, S, Eq (5.27) can also be expressed as below:
R(X) max = B log (2e S) bits/sec

(5.28)

If X represents white Gaussian noise ( WGN), with an average power =N, Eq (4.28) becomes:
R(X) max = B log (2e N) bits/sec

(5.29)

Case IV: Average Power Limitation (Unidirectional Distribution):

If the signal X has a unidirectional distribution with average power limitation (AM with
average carrier power constraint), the constraint equations become:

f ( x )dx 1 , and

x . f ( x )dx P0
0

Then following the steps, as in other cases, you can show that:
f(x)

2
P

exp

x 2 2P

, 0 x

(5.30)

H ( X )max

R( X )max B log

1
2

log

eP0
2

eP0
2

bits/sample

bits/sec.

..

(5.31)

(5.32)

Compared to Eq (5.28), it is clear that the entropy in Eq (5.32) is smaller by 2 bits/sec.

2

NOTE: In case III we have shown that the entropy of Gaussian signals is log 2e log 2eS ,
where S is the signal power and is maximum among all p.d.f `s of continuous signals with average
power limitation. Often times, for the calculation of the transmission rate and channel capacity, it is
convenient to find equivalent entropy of signals and interference which are neither random nor Gaussian.
We define Entropy Power Se of an ensemble of samples limited to the same band width, B, and period
T as the original ensemble, and having the same entropy H bits/sample. Then
2H

H/sample= log 2eSe .Therefore Se =e /2e. Since H for a random signal is maximum for a given
S when it is Gaussian, then Se for any arbitrary signal is less than or equal to the average power of the
signal.
Mutual Information of a Continuous Noisy Channel:

Dept. of ECE/SJBIT

Page 119

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

While remembering that the entropy definition for continuous signals is a relative one and using
the logarithmic property, it follows that all entropy relations that we have studied for discrete signals
do hold for continuous signals as well. Thus we have
H ( X ,Y )

f ( x , y ) log

H(X|Y)

(5.33)

dxdy

(5.34)

dxdy

f(x,y)

f ( x , y ) log

f(x|y)

The mutual information is I(X, Y) = H(X) H (X|Y)

(5.35)

= H(Y) - H (Y|X)

= H(X) + H(Y) H(X, Y)

The channel capacity is

C = Max I(X, Y)

(5.36)
(5.37)
(5.38)

And so on. You can easily verify the various relations among entropies.
Amount of Mutual Information:
Suppose that the channel noise is additive, and statistically independent of the transmitted
signal. Then, f (y|x) depends on (y-x), and not on x or y. Since Y=X+n, where, n is the channel noise,
and f (y|x) = fY (y|X=x).
It follows that when X has a given value, the distribution of Y is identical to that of n, except for a
translation of X. If fn (.) represents the p.d.f. of noise sample, n, then obviously fY (y|x) = fn(y-x)

1
fY ( y | x )

dy fn ( y x ) log

fY ( y | x ) log

dy

fn ( y x )

Letting z = y x , we have then

H ( Y | x ) fn ( z ) log

dz H ( n )
fn ( z )

Accordingly,

H ( Y | X ) f X ( x )H ( Y | x )dx f X ( x )H ( n )dx H ( n ) f X ( x )dx H ( n )

Or

H (Y|X) =H (n)

There fore, the amount of mutual information

(5.39)

=H(Y) H (Y|X)

That is, I(X, Y) =H(Y) H (n)

The equivocation is H (X|Y) = H(X) I(X, Y) = H(X) H(Y) + H (n)

. (5.40)
(5.41)

Since the Channel capacity is defined as

Dept. of ECE/SJBIT

Page 120

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

C = R max =2B I(X, Y) max bits/sec. =[R(Y) R (Y|X)] max

It follows: C = [R(Y) R (n)] max

.. (5.42)

Capacity of band limited channels with AWGN and Average Power

Limitation of signals: The Shannon Hartl ey law:

Consider the situation illustrated in Fig. 5.1. The received signal will be composed of the
transmitted signal X plus noise n. The joint entropy at the transmitter end, assuming signal and
noise are independent, is:
H(X, n) = H(X) + H (n|X).
= H(X) + H (n)

(5.43)

The joint entropy at the receiver end, however, is

H(X, Y) = H(Y) + H (X|Y)

(5.44)

Since the received signal is Y = X + n, and the joint entropy over the channel is invariant, it follows:
H (X, n) = H (X, Y)

(5.45)

The, from Eq. (5.43) and Eq (5.44), one obtains

I (X, Y) = H (X) H ( X | Y)
= H (Y) - H (n)

(5.46)

Alternatively, we could have directly started with the above relation in view of Eq. (5.40).
Hence, it follows, the channel capacity, in bits / second is:
C = {R(Y) R (n)} max

. (5.47)

If the additive noise is white and Gaussian, and has a power N in a bandwidth of B Hz,
then from Eq. (5.29), we have:
R (n) max = B log 2eN

(5.48)

Further, if the input signal is also limited to an average power S over the same bandwidth, and X and
n are independent then it follows:

Dept. of ECE/SJBIT

Page 121

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Y = (S + N)
We have seen that for a given mean square vale, the entropy will become a maximum if the signal is
Gaussian, and there fore the maximum entropy of the output is:
H (Y) max = (1/2) log 2 (S + N) bits/sample
Or, R (Y) max = B log 2 (S + N) bits/sec

(5.49)

If n is an AWGN, then Y will be Gaussian if and only if, X is also Gaussian. This implies
fX( x )

exp

2
2S
Using Eq. (5.48) and (5.49) in Eq. (5.47), one obtains
C B log 1

bits/sec

Or, C = B log (1 + ) bits/sec

(5.50)

(5.51)

Where, = S/N, is the signal to noise power ratio.

This result (Eq. 5.50) is known as Shannon-Hartley law. The primary significance of the
formula is that it is possible to transmit over a channel of bandwidth B Hz perturbed by AWGN at a
rate of C bits/sec with an arbitrarily small probability of error if the signal is encoded in such a
manner that the samples are all Gaussian signals. This can, however, be achieved by Orthogonal
codes.
Bandwidth SNR Tradeoff:
One important implication of the Shannon - Hartley Law is the exchange of bandwidth with
signal to noise power ratio. Suppose, S/N = 7, and B = 4KHz then from Eq. (4.50), C = 12000 bits/sec.
Let S/N be increased to 15, while the bandwidth is reduced to 3 KHz. We see that the channel capacity
remains the same. The Power Spectral Density, Sn(f), of White Gaussian noise is more or less a constant
over the entire frequency range,(-, ), with a two sided p.s.d of N0/2 as shown in Fig 5.3.

From the figure we find the noise power over (-B, B) as N = N0 / 2 .2B or

Dept. of ECE/SJBIT

Page 122

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

N = N0B

(5.52)

That is, the noise power is directly proportional to the band width B. Thus the noise power will be
reduced by reducing the band width and vice-versa. This indicates, then, an increase in the signal
power and vice-versa. For illustration consider the following:
With

S1

=7 and B1 = 4 KHz, we have from Eq (5.50) C = 12000 bits/sec.

N1

Suppose we decrease the band width by 25% and choose the new values of SNR and band width as

S2

=15 and B2 = 3 KHz, we get the same channel capacity as before.

N2

Then

S1
N1

S2

=7/15

N2

Using Eq (5.52), this means

S1 = 7 . B1 7 . 4 28 0.62222 or S = 45 S = 1.607 S
2
1
1
S2
15 B2 15 3 45
28
Thus a 25 % reduction in bandwidth requires a 60.7 % increase in signal power for
maintaining the same channel capacity. We shall inquire into this concept in a different way. Eq.
(5.50) can be written in the form:
B
C

log 1

5.53)

Table below shows the variation of (B/C) with (S/N)

S/N
A plot

0.5

of (B/C) versus (S/N

) is shown

in Fig

10

5.4. Clearly,

15

the same

20

channel

30

capacity may be obtained

B/C
1.71 1.0
63
0.37
0.289
0.25
0.23
0.20
increasing S/N is poor. Use of larger band width for smaller S/N is generally known as coding
upwards and use of smaller B with larger S/N is called coding downwards. One can consider as
examples of coding upwards, the FM, PM and PCM systems where larger band widths are used with
improvements in S/N ratio. Quantization of the signal samples and then, combining the different sample
values into a single pulse, as in the case of multi-level, discrete PAM can be considered as an example of
coding downwards where the band width reduction depends on the signal power available.

Dept. of ECE/SJBIT

Page 123

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

For wide band systems, where (S/N) 1, use of Eq (5.50) leads to:
C B1 log

S1
1

B2

log

S
1 2
N

B1

or 1

S2

N
2

S2

Or

B2

S1

N2

N1

S2

, when

and

N1

(5.54)

N2

Notice that Eq. (5.54) predicts an exponential improvement in (S/N) ratio with band width for an
ideal system. For the conventional demodulation methods used in FM and PM, however, the ( S/N)
ratio varies as the square of the transmission bandwidth, which, obviously, is inferior to the ideal
performance indicated by Eq. (5.54). However, better performance that approaches the Shannonbound in the limit as bandwidth is made infinite, can be obtained by using optimum demodulators
(Phase-locked loop).

Capacity of a Channel of Infinite Bandwidth:

The Shannon-Hartley formula predicts that a noiseless Gaussian channel with (S/N = ) has
an infinite capacity. However, the channel capacity does not become infinite when the bandwidth is
made infinite, in view of Eq. (5.52). From Eq. (5.52), Eq. (5.50) is modified to read:
C

B log

N B bits / sec .
0

N0 B

log 1

S
N B

N0
0
S
S
S . 1 log1 x .... Where x
N0 B
N0 x
1

Or

C S log1 xx N
0

Dept. of ECE/SJBIT

(5.55)

Page 124

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Accordingly, when B, x 0 and we have

C

max

lim

log

1 x x

(5.56)

(5.57)

x 0

N0

Since lim 1 x x e
x 0

lim

and log e 1.442685 , we have

Cmax 1.442685 S bits/sec

N0

Eq (5.57) places an upper limit on channel capacity with increasing band width.
Bandwidth-Efficiency: Shannon Limit:
In practical channels, the noise power spectral density N0 is generally constant. If Eb is the
transmitted energy per bit, then we may express the average transmitted power as:
S = Eb C

..

(5.58)

Using Eq. (5.52) and (5.58), Eq. (5.50) may now be re-formulated as:

C
log

E C
1 b .
N0

. (5.59)

From which one can show:

Eb

C
2

N0

(5.60)

(C/B) is called the bandwidth efficiency of the syste m. If C/B = 1, then it follows that Eb = N0.This
implies that the signal power equals the noise power. Suppose, B = B0 for which, S = N, then Eq.
(5.59) can be modified as:
C

B0
lim

log 1

B0

(5.61)

B
C

And

B B0

1
B0

C max =B0 log e

lim
log

B0
B

1
B0

B0
log e

log e

(5.62)

That is, the maximum signaling rate for a given S is 1.443 bits/sec/Hz in the bandwidth over which the
signal power can be spread without its falling below the noise level.
It follows from Eq. (5.58):

Dept. of ECE/SJBIT

Page 125

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

S = Eb C =N0B0
And
lim

as
Eb

B
lim B
0

Eb

In other words

or
we
B
0

C
max

Eb/N0 =B0/C
have
B
0

And

as

we

have

ln 2

B0 log e

ln 2 0.69314718

(5.63)

0 min

Expressed in decibels, this means:

Eb

10 log 0.69314718 dB 1.59174539 dB

(5.64)

0 min

Eb

1.592 dB is known as the Shannon` s limit for transmission at capacity

The value

min

Cmax=1.443S/N0 and the communication fails otherwise (i.e.Eb/N0 <0.693.).

We define an ideal system as one that transmits d ata at a bit rate R equal to the channel capacity.
Fig 5.5 shows a plot of the bandwidth efficiency, R / B = C/B = b, as a
function of the energy-per-bit to noise power spec tral density ratio Eb/N0. Such a diagram is called
bandwidth efficiency diagram. Clearly shown on th e diagram are:

The capacity boundary defined by the curve R = C.

Region for which R > C shown as dark gray area for which error-free transmission is not possible.
Region for which R < C (light gray portion) in which combinations of system-parameters have
the potential for supporting error-free transmission.

Example 5.3
A vice-grade channel of a telephone network has a bandwidth of 3.4 kHz.

Dept. of ECE/SJBIT

Page 126

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

a) Calculate the channel capacity of the telephone channel for a signal to noise ratio of 30dB.
b) Calculate the minimum signal to noise ratio required to support information transmission
through the telephone channel at the rate of 4800 bits/sec.
c) Solution:
a) B = 3.4kHz,

S/N = 30dB

[S/N] dB =10 log10[S/N] Abs (Remember S/N is a Power ratio)

{[S/N] dB/10}

S/N =10

=10 =1000

C =B log 2{1+ S/N} =33888.56928 bits/sec

b) C = 4800, B =3.4 kHz and S/N = (2

C/B

-1) = 1.660624116 or 2.2 dB

Example 5.4
A communication system employs a continuous source. The channel noise is white and Gaussian. The
bandwidth of the source output is 10 MHz and signal to noise power ratio at the receiver is 100.
a) Determine the channel capacity
b) If the signal to noise ratio drops to 10, how much bandwidth is needed to achieve the same
channel capacity as in (a).
c) If the bandwidth is decreased to 1MHz, what S/N ratio is required to maintain the same
channel capacity as in (a).
Solution:
7

a) C = 10 log (101) = 6.66 10 bits/sec

b) B = C / log2 (1+ S/N)=C/log211=19.25 MHz
C/B

c) S/N =(2

20

-1)=1.10510

=200.43dB

Example 5.5:
Alphanumeric data are entered into a computer from a remote terminal though a voice grade
telephone channel. The channel has a bandwidth of 3.4 kHz and output signal to noise power ratio of
20 dB. The terminal has a total of 128 symbols which may be assumed to occur with equal
probability and that the successive transmissions are statistically independent.
a) Calculate the channel capacity:
b) Calculate the maximum symbol rate for which error-free transmission over the channel
is possible.

Dept. of ECE/SJBIT

Page 127

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Solution:
a) S/N = 10

20/10

= 10 = 100

C = 3.4 x 10 log 2 (1 + 100) = 22637.92 bits/sec

b) H max = log2 128 = 7 bits per symbol
C = rs max. Hmax
Therefore, rs max = C/Hmax = 3233.99 bits/sec = 3.234 kb/sec
Example 5.6
A black and white television picture may be viewed as consisting of approximately
3
5
x10 elements, each one of which may occupy one of 10 distinct brightness levels with equal
probability. Assume the rate of transmission to be 30 picture frames per second, and the signal to
noise power ratio is 30 dB.
Using the channel capacity theorem, calculate the minimum bandwidth required to support the
transmission of the resultant video signal.
Solution:
(S/N) dB=30 dB or (S/N) Abs=1000
No. of different pictures possible = 10

3 x105
5

Therefore, entropy per picture =3 x 10 log 210 = 9.965784285 x 10 bits

5

Entropy rate = rs. H = 30 x H = 298.97 x 10 bits/sec

C= rs. H=B log2 [1+S/N]

Bmin= rs. HB log2 [1+S/N] =3.0 MHz

Note: As a matter of interest, commercial television transmissions actually employ a bandwidth of 4

MHz.

Dept. of ECE/SJBIT

Page 128

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

Review questions:
1. Show that for a AWGN channel C

10EC55

watts/Hz.
2. Consider an
1012
2

1.448

where /2 = noise power spectral density in

AWGN channel with 4 KHz bandwidth with noise power spectral density
watts/Hz. The signal power required at the receiver is 0.1mW. Calculate the

capacity of the channel.

3. If I(xi, yi) = I(xi)-I(xi/yj). Prove that I(xi,yj)=I(yj)-I(yj/xi)

4. Consider a continuous random variable having a distribution as given below

1

5. FX(X)

0XA

0 OTHERWISE

6. Find the differential entropy H(x)

7. Design a single error correcting code with a message block size of 11 and show that by an
example that it can correct single error.
8. If Ci and Cj an two code vectors in a (n,k) linear block code, show that their sum is also a code
vector.
T

9. Show CH =0 for a linear block code.

10. Prove that the minimum distance of a linear block code is the smallest weight of the non-zero
code vector in the code.
.

Dept. of ECE/SJBIT

Page 129

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

PART B
UNIT 5
CONCEPTS OF ERROR CONTROL CODING -- BLOCK CODES

Syllabus:
Introduction, Types of errors, examples, Types of codes Linear Block Codes: Matrix
description, Error detection and correction, Standard arrays and table look up for decoding.
7 Hours

Text Books:
Digital and analog communication systems, K. Sam Shanmugam, John Wiley,
1996. Digital communication, Simon Hay kin, John Wiley, 2003.
Reference Books:
ITC and Cryptography, Ranjan Bose, TMH, II edition, 2007
Digital Communications - Glover and Grant; Pearson Ed. 2nd Ed 2008

Dept. of ECE/SJBIT

Page 130

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

UNIT 5
CONCEPTS OF ERROR CONTROL CODING -- BLOCK CODES
The earlier chapters have given you enough background of Information theory and Source
encoding. In this chapter you will be introduced to another important signal - processing operation,
namely, Channel Encoding, which is used to provide reliable transmission of information over
the channel. In particular, we present, in this and subsequent chapters, a survey of Error control
coding techniques that rely on the systematic addition of Redundant symbols to the transmitted
information so as to facilitate two basic objectives at the receiver: Error- detection and Error
correction. We begin with some preliminary discussions highlighting the role of error control
coding.
5.1 Rationale for Coding:
The main task required in digital communication is to construct cost effective systems for
transmitting information from a sender (one end of the system) at a rate and a level of reliability that
are acceptable to a user (the other end of the system). The two key parameters available are
transmitted signal power and channel band width. These two parameters along with power spectral
density of noise determine the signal energy per bit to noise power density ratio, Eb/N0 and this ratio,
as seen in chapter 4, uniquely determines the bit error for a particular scheme and we would like to
transmit information at a rate RMax = 1.443 S/N. Practical considerations restrict the limit on Eb/N0
that we can assign. Accordingly, we often arrive at modulation schemes that cannot provide
acceptable data quality (i.e. low enough error performance). For a fixed Eb/N0, the only practical
alternative available for changing data quality from problematic to acceptable is to use coding.
Another practical motivation for the use of coding is to reduce the required Eb/N0 for a fixed
error rate. This reduction, in turn, may be exploited to reduce the required signal power or reduce the
hardware costs (example: by requiring a smaller antenna size).
The coding methods discussed in chapter 5 deals with minimizing the average word length of
the codes with an objective of achieving the lower bound viz. H(S) / log r, accordingly, coding is
termed entropy coding. However, such source codes cannot be adopted for direct transmission over
the channel. We shall consider the coding for a source having four symbols with probabilities p (s1)
=1/2, p (s2) = 1/4, p (s3) = p (s4) =1/8. The resultant binary code using Huffmans procedure is:
s1

s2

0
10

s
s

110
111

Clearly, the code efficiency is 100% and L = 1.75 bints/sym = H(S). The sequence s3s4s1 will
then correspond to 1101110. Suppose a one-bit error occurs so that the received sequence is 0101110.
This will be decoded as s1s 2s4s1, which is altogether different than the transmitt ed sequence.
Thus although the coding provides 100% efficiency in the light of Shannons theorem, it suffers a
major disadvantage. Another disadvantage of a variable length code lies in the fact that output data
rates measured over short time periods will fluctuate widely. To avoid this problem, buffers of large
length will be needed both at the encoder and at the decoder to store the variable rate bit stream if a
fixed output rate is to be maintained.
Some of the above difficulties can be resolved by using codes with fixed length. For
example, if the codes for the example cited are modified as 000, 100, 110, and 111. Observe that even
if there is a one-bit error, it affects only one block and that the output data rate will not fluctuate.

Dept. of ECE/SJBIT

Page 131

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

The encoder/decoder structure using fixed length code words will be very simple compared to the
complexity of those for the variable length codes.
Here after, we shall mean by Block codes, the fixed length codes only. Since as discussed
above, single bit errors lead to single block errors, we can devise means to detect and correct these
errors at the receiver. Notice that the price to be paid for the efficient handling and easy
manipulations of the codes is reduced efficiency and hence increased redundancy.
In general, whatever be the scheme adopted for transmission of digital/analog information, the
probability of error is a function of signal-to-noise power ratio at the input of a receiver and the data
rate. However, the constraints like maximum signal power and bandwidth of the channel (mainly the
Governmental regulations on public channels) etc, make it impossible to arrive at a signaling scheme
which will yield an acceptable probability of error for a given application. The answer to this problem
is then the use of error control coding, also known as channel coding. In brief, error control
coding is the calculated addition of redundancy . The block diagram of a typical data transmission
system is shown in Fig. 6.1
The information source can be either a person or a machine (a digital computer). The source
output, which is to be communicated to the destination, can be either a continuous wave form or a
sequence of discrete symbols. The source encoder transforms the source output into a sequence of
binary digits, the information sequence u. If the source output happens to be continuous, this involves
A-D conversion as well. The source encoder is ideally designed such that (i) the number of bints per unit
time (bit rate, rb) required to represent the source output is minimized (ii) the source output can be
uniquely reconstructed from the information sequence u.

The Channel encoder transforms u to the encoded sequence v, in general, a binary

sequence, although non-binary codes can also be used for some applications. As discrete symbols are
not suited for transmission over a physical channel, the code sequences are transformed to waveforms
of specified durations. These waveforms, as they enter the channel get corrupted by noise. Typical
channels include telephone lines, High frequency radio links, Telemetry links, Microwave links, and
Satellite links and so on. Core and semiconductor memories, Tapes, Drums, disks, optical memory
and so on are typical storage mediums. The switching impulse noise, thermal noise, cross talk and
lightning are some examples of noise disturbance over a physical channel. A surface defect on a
magnetic tape is a source of disturbance. The demodulator processes each received waveform and
produces an output, which may be either continuous or discrete the sequence r. The channel
decoder transforms r into a binary sequence,

Dept. of ECE/SJBIT

which gives the estimate of u, and ideally should be

Page 132

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

the replica of u. The source decoder then transforms u into an estimate of source output and delivers
this to the destination.
Error control for data integrity may be exercised by means of forward error correction
(FEC) where in the decoder performs error correction operation on the received information
according to the schemes devised for the purpose. There is however another major approach known
as Automatic Repeat Request ( ARQ), in which a re-transmission of the ambiguous information is
effected, is also used for solving error control problems. In ARQ, error correction is not done at all.
The redundancy introduced is used only for error detection and upon detection, the receiver
requests a repeat transmission which necessitates the use of a return path (feed back channel).
In summary, channel coding refers to a class of signal transformations designed to improve
performance of communication systems by enabling the transmitted signals to better withstand the
effect of various channel impairments such as noise, fading and jamming. Main objective of error
control coding is to reduce the probability of error or reduce the Eb/N0 at the cost of expending more
bandwidth than would otherwise be necessary. Channel coding is a very popular way of providing
performance improvement. Use of VLSI technology has made it possible to provide as much as 8
dB performance improvement through coding, at much lesser cost than through other methods such as
high power transmitters or larger Antennas.
We will briefly discuss in this chapter the channel encoder and decoder strategies, our major
interest being in the design and implementation of the channel encoder/decoder pair to achieve fast
transmission of information over a noisy channel, reliable communication of information and
reduction of the implementation cost of the equipment.
5.2 Types of errors:
The errors that arise in a communication system can be viewed as independent errors and
burst errors. The first type of error is usually encountered b y the Gaussian noise, which is the
chief concern in the design and evaluation of modulators and demodulators for data transmission. The
possible sources are the thermal noise and shot noise of the transmitting and receiving equipment,
thermal noise in the channel and the radiations picked up by the receiving antenna. Further, in
majority situations, the power spectral density of the Gaussian noise at the receiver input is white.
The transmission errors introduced by this noise are such that the error during a particular signaling
interval does not affect the performance of the system during the subsequent intervals. The discrete
channel, in this case, can be modeled by a Binary symmetric channel. These transmission errors due
to Gaussian noise are referred to as independent errors ( or random errors).
The second type of error is encountered due to the impulse noise, which is characterized by
long quiet intervals followed by high amplitude noise bursts (As in switching and lightning). A noise
burst usually affects more than one symbol and there will be dependence of errors in successive
transmitted symbols. Thus errors occur in bursts
5. 3 Types of codes:
There are mainly two types of error control coding schemes Block codes and convolutional
codes, which can take care of either type of errors mentioned above.
In a block code, the information sequence is divided into message blocks of k bits each,
represented by a binary k-tuple, u = (u1, u2 . uk) and each block is called a message. The symbol u,
here, is used to denote a k bit message rather than the entire information sequence . The encoder

Dept. of ECE/SJBIT

Page 133

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

then transforms u into an n-tuple v = (v1, v2 . vn). Here v represents an encoded block rather than the
entire encoded sequence. The blocks are independent of each other.
The encoder of a convolutional code also accepts k-bit blocks of the information sequence u
and produces an n-symbol block v. Here u and v are used to denote sequences of blocks rather than a
single block. Further each encoded block depends not only on the present k-bit message block but also on
m-pervious blocks. Hence the encoder has a memory of order m. Since the encoder has memory,
implementation requires sequential logic circuits.
If the code word with n-bits is to be transmitted in no more time than is required for the
transmission of the k-information bits and if b and c are the bit durations in the encoded and coded
words, i.e. the input and output code words, then it is necessary that
n.c = k.b
We define the rate of the code by (also called rate efficiency)
Rc

Accordingly, with f
b

and f

, we have fb c k R
c
fc b n
c

5.4 Example of Error Control Coding:

Better way to understand the important aspects of error control coding is by way of an
example. Suppose that we wish transmit data over a telephone link that has a useable bandwidth of 4
KHZ and a maximum SNR at the out put of 12 dB, at a rate of 1200 bits/sec with a probability of error
-3
less than 10 . Further, we have DPSK modem that can operate at speeds of 1200, 1400 and 3600 bits/sec
-3

-3

-3

with error probabilities 2(10 ), 4(10 ) and 8(10 ) respectively. We are asked to design an error
-3

control coding scheme that would yield an overall probability of error < 10 . We have:

C = 16300 bits/sec, Rc = 1200, 2400 or 3600 bits/sec.

[C=Blog2 (1+

S S
).

12dB or 15.85 , B=4KHZ], p = 2(10

-3)

-3

-3

, 4(10 ) and 8(10 ) respectively. Since

N N
Rc < C, according to Shannons theorem, we should be able to transmit data with arbitrarily small
probability of error. We shall consider two coding schemes for this problem.
1. Error detection: Single parity check-coding. Consider the (4, 3) even parity check code.

Message

000

001

010

011

100

101

110

111

Parity

0101

0110

1001

1010

1100

1111

Codeword 0000 0011

Parity bit appears at the right most symbol of the codeword.

Dept. of ECE/SJBIT

Page 134

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

This code is capable of detecting all single and triple error patterns. Data comes out of the channel
-3
encoder at a rate of 3600 bits/sec and at this rate the modem has an error probability of 8(10 ). The
decoder indicates an error only when parity check fails. This happens for single and triple errors only.
pd = Probability of error detection.
= p(X =1) + p(X = 3), where X = Random variable of errors.
-3

Using binomial probability law, we have with p = 8(10 ):

n

P( X = k ) =

k
4

p )n k

p (1

p( 1

p)

( 1 p ),

4C 1

4,

1
3
2
3
4
Expanding we get pd 4 p 12 p 16 p 8 p

4C 3 4

Substituting the value of p we get:

-3

-6

-9

-12

pd = 32 (10 ) - 768 (10 ) +8192 (10 ) 32768 (10

-3

) = 0.031240326 > > (10 )

However, an error results if the decoder does not indicate any error when an error indeed has
occurred. This happens when two or 4 errors occur. Hence probability of a detection error = pnd
(probability of no detection) is given by:
4

p nd

2
P( X 2 ) P( X 4 ) p

(1

p)

2
-3
-3
Substituting the value of p we get pnd=0.410 10

( 1 p ) 6 p 12 p 7 p

-3

Thus probability of error is less than 10 as required.

2. Error Correction: The triplets 000 and 111 are transmitted whenever 0 and 1 are inputted. A
majority logic decoding, as shown below, is employed assuming only single errors.

Received
Triplet

000

001

010

100

011

101

110

111

Output
message

Probability of decoding error, pde= P (two or more bits in error)

Dept. of ECE/SJBIT

Page 135

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

(1-p) +

p (1-p) =3p -2p

3
-6

-3

=190.464 x 10 =0.19x 10 < p= 10

3

-3
-9

Probability of no detection, pnd =P (All 3 bits in error) = p =512 x 10 < < pde!
In general observe that probability of no detection, pnd < < probability of decoding error, pde.
The preceding examples illustrate the following aspects of error control coding. Note that in
-3
both examples with out error control coding the probability of error =8(10 ) of the modem.
1. It is possible to detect and correct errors by adding extra bits-the check bits, to the message
sequence. Because of this, not all sequences will constitute bonafide messages.
2. It is not possible to detect and correct all errors.
3. Addition of check bits reduces the effective data rate through the channel.
4. Since probability of no detection is always very much smaller than the decoding error
probability, it appears that the error detection schemes, which do not reduce the rate efficiency
as the error correcting schemes do, are well suited for our application. Since error detection
schemes always go with ARQ techniques, and when the speed of communication becomes a
major concern, Forward error correction (FEC) using error correction schemes would be
desirable.
5.5 Block codes:
We shall assume that the output of an information source is a sequence of Binary digits. In
Block coding this information sequence is segmented into message blocks of fixed length, say k.
Each message block, denoted by u then consists of k information digits. The encoder transforms these
k-tuples into blocks of code words v, each an n- tuple according to certain rules. Clearly, corresponding
k
k
k
to 2 information blocks possible, we would then have 2 code words of length n > k. This set of 2 code
k
words is called a Block code. For a block code to be useful these 2 code words must be distinct, i.e.
there should be a one-to-one correspondence between u and v. u and v are also referred to as the input
k
vector and code vector respectively. Notice that encoding equipment must be capable of storing the 2
code words of length n > k. Accordingly, the complexity of the equipment would become prohibitory if n
and k become large unless the code words have a special structural property conducive for storage and
mechanization. This structural is the linearity.
5.5.1 Linear Block Codes:
k

A block code is said to be linear (n ,k) code if and only if the 2 code words from a k- dimensional
sub space over a vector space of all n-Tuples over the field GF(2).
m

Fields with 2 symbols are called Galois Fields (pronounced as Galva fields), GF
m
(2 ).Their arithmetic involves binary additions and subtractions. For two valued variables, (0, 1).The
modulo 2 addition and multiplication is defined in Fig 6.3.

Dept. of ECE/SJBIT

Page 136

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

The binary alphabet (0, 1) is called a field of two elements (a binary field and is denoted by
GF (2). (Notice that represents the EX-OR operation and represents the AND operation).Further in
binary arithmetic, X=X and X Y = X Y. similarly for 3-valued variables, modulo 3 arithmetic can
be specified as shown in Fig 6.4. However, for brevity while representing polynomials involving binary
addition we use + instead of and there shall be no confusion about such usage.
Polynomials f(X) with 1 or 0 as the co-efficients can be manipulated using the above relations.
m
The arithmetic of GF(2 ) can be derived using a polynomial of degree m, with binary co-efficients
and using a new variable called the primitive element, such that p() = 0.When p(X) is irreducible
3
2
3
4
3
(i.e. it does not have a factor of degree m and >0, for example X + X + 1, X + X + 1, X +X +1,
5
2
4 3 2
X +X +1 etc. are irreducible polynomials, whereas f(X)=X +X +X +1 is not as f(1) = 0 and hence has a
factor X+1) then p(X) is said to be a primitive polynomial.
If vn represents a vector space of all n-tuples, then a subset S of vn is called a subspace if (i)
the all Zero vector is in S (ii) the sum of any two vectors in S is also a vector in S. To be more
specific, a block code is said to be linear if the following is satisfied. If v1 and v2 are any two code
words of length n of the block code then v1 v2 is also a code word length n of the block code.
Example 6.1: Linear Block code with k= 3, and n = 6

Observe the linearity property: With v3 = (010 101) and v4 = (100 011), v3 v4 = (110 110) = v7.
Remember that n represents the word length of the code words and k represents the number of
information digits and hence the block code is represented as (n, k) block code.

Dept. of ECE/SJBIT

Page 137

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Thus by definition of a linear block code it follows that if g1, g2 gk are the k linearly
independent code words then every code vector, v, of our code is a combination of these code words,
i.e.
v = u1 g1u2 g2 uk gk

(6.1)

Where uj= 0 or 1, 1 j k
Eq (6.1) can be arranged in matrix form by nothing that each gj is an n-tuple, i.e.
gj= (gj1, gj2,. gjn)
Thus we have

v=uG

(6.2)

u = ( u1, u2 uk)

Where:

(6.3)
(6.4)

represents the data vector and

g1
G

g2

g3

g11

21

L g1n

12
22

L g2 n

(6.5)

Lg

g
k1

k2

kn

is called the generator matrix.

Notice that any k linearly independent code words of an (n, k) linear code can be used to form
a Generator matrix for the code. Thus it follows that an (n, k) linear code is completely specified by
the k-rows of the generator matrix. Hence the encoder need only to store k rows of G and form linear
combination of these rows based on the input message u.
Example 6.2: The (6, 3) linear code of Example 6.1 has the following generator matrix:
g1
G g

1 0 0 0 1
2

0 1 0 1 0

g3
0 0 1
If u = m5 (say) is the message to be coded, i.e. u = (011)

1
1

1 1 0

We have v = u .G = 0.g1 + 1.g2 +1.g3

= (0,0,0,0,0,0) + (0,1,0,1,0,1) + (0,0,1,1,1,0) = (0, 1, 1, 0, 1, 1)
Thus v = (0 1 1 0 1 1)
v can be computed simply by adding those rows of G which correspond to the locations of
1`s of u.

Dept. of ECE/SJBIT

Page 138

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

5.5.2 Systematic Block Codes (Group Property):

A desirable property of linear block codes is the Systematic Structure. Here a code word is
divided into two parts Message part and the redund ant part. If either the first k digits or the last k
digits of the code word correspond to the message part then we say that the code is a Systematic
Block Code. We shall consider systematic codes as depicted i n Fig.6.5.

In the format of Fig.6.5 notice that:

v1 = u1, v2 = u2, v3 = u3 vk = uk
vk 1 u1 p11 u2 p21 u3 p31 uk pk1

vk 2 u1 p12 u2 p22 u3 p32 uk pk2

M
M

v u p
n

u p

1,n - k

2,n - k

u p
3

3,n - k

(5.6

(5.6

a)

b)

.. u p
k

k, n - k

Or in matrix from we have

v1

v2

v v

... vk

... vn
0 0 ... 0
1 0 ... 0

k 1 k 2

1
0

u1

... uk M

u2

M MMM M

... 1

i.e., v = u. G
Where G = [Ik, P]
p
p11
12

Where

P=

21

22

M M
p
p
k,1

k ,2

11

21

12

22

M
p

k1

k2

...
...

MMM

...

Lp ,

p1 ,n k
2 ,nk

(5.7)

M
p
k ,nk

. (5.8)

Lp ,

1 nk

nk

M
L pk ,

. (5.9)

nk

Ik is the k k identity matrix (unit matrix), P is the k (n k) parity generator matrix, in which pi, j
are either 0 or 1 and G is a k n matrix. The (n k) equations given in Eq (6.6b) are referred to as parity
check equations. Observe that the G matrix of Example 6.2 is in the systematic format. The n-vectors a =
(a1, a2 an) and b = (b1, b2 bn) are said to be orthogonal if their inner product defined by:
T

a.b = (a1, a2a n) (b1, b2 b n) = 0.

where, T represents transposition. Accordingly for any kn matrix, G, with k linearly independent
rows there exists a (n-k) n matrix H with (n-k) linearly independent rows such that any vector in the
row space of G is orthogonal to the rows of H and that any vector that is orthogonal to the rows of

Dept. of ECE/SJBIT

Page 139

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

H is in the row space of G. Therefore, we can describe an (n, k) linear code generated by G alternatively
as follows:
An n tuple, v is a code word generated by G, if and only if v.H
a) (O represents an all zero row vector.)

= O. ..(5.9

This matrix H is called a parity check matrix of the code. Its dimension is (n k) n.
If the generator matrix has a systematic format, the parity check matrix takes the following form.
p

11

H = [P .In-k] =

21

12

22

M
p

1 ,nk

...
...
M

...
2 ,nk

k1

1 0 0

k2
0 1 0
MMM M M M
p
0 0 0

... 0
... 0 (5.10)
MMM

...

k ,nk

th

The i row of G is:

gi = (0 0 1 00

th
i element

pi,1 pi,2 pi,j pi, n-k)

th
(k + j) element

th

The j row of H is:

i

th

th

element
(k + j) element

hj = ( p1,j p2,j p i,j ...pk, j 0 0 0 1 0 0 )

Accordingly the inner product of the above n vectors is:
T
T
gihj = (0 0 1 00 pi,1 pi,2 pi,j pi, n-k) ( p1,j p2,j p i,j ...pk, j 0 0 0 1 0 0 )

th
th
th
th
i element
(k + j) element
i element
(k + j) element

= pij + pij. = 0 (as the pij are either 0 or 1 and in modulo 2 arithmetic

X + X = 0)

This implies simply that:

T

G. H = Ok (n k)

Dept. of ECE/SJBIT

(6.11)

Page 140

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Where Ok (n k) is an all zero matrix of dimension k (n k) .

Further, since the (n k) rows of the matrix H are linearly independent, the H matrix of Eq.
(6.10) is a parity check matrix of the (n, k) linear systematic code generated by G. Notice that the
parity check equations of Eq. (6.6b) can also be obtained from the parity check matrix using the fact
T

v.H = O.
T

Alternative Method of proving v.H = O.:

We have v = u.G = u. [Ik: P]= [u1, u2 u k, p1, p2 . P n-k]
Where pi =( u1 p1,i + u2 p2,i + u3 p3,i + u k pk, i) are the parity bits found from Eq (6.6b).
Now H

P
I

nk

v.H = [u1 p11 + u2 p21 +. + . + uk pk1 + p1, u1 p12 + u2 p22 + .. + uk pk2 +

p2, u1 p1, n-k + u2 p2, n-k + . + uk pk, n-k + pn-k]
= [p1 + p1, p2 + p2 pn-k + pn-k]
= [0, 0 0 ]
T

Thus v. H = O. This statement implies that an n- Tuple v is a code word generated by G if and only
if
T
vH =O
T

Since v = u G, This means that: u G H = O

T

If this is to be true for any arbitrary message vector v then this implies: G H = Ok (n k)
Example 5.3:
Consider the generator matrix of Example 6.2, the corresponding parity check matrix is
0
H=

1 1 1 0 0
0 1 010
1 0 00 1

5.5.3 Syndrome and Error Detection:

Suppose v = (v1, v2 vn) be a code word transmitted over a noisy channel and let:
r = (r1, r2 . rn) be the received vector. Clearly, r may be different from v owing to the channel noise.
The vector sum
e = r v = (e1, e2 en)

Dept. of ECE/SJBIT

(5.12)

Page 141

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

is an n-tuple, where ej = 1 if rj vj and ej = 0 if rj = vj. This n tuple is called the error vector or
error pattern. The 1s in e are the transmission errors caused by the channel noise. Hence from Eq
(6.12) it follows:
(5.12a)
r=ve
Observe that the receiver noise does not know either v or e. Accordingly, on reception of r the
decoder must first identify if there are any transmission errors and, then take action to locate these
errors and correct them (FEC Forward Error Correction) or make a request for retransmission
(ARQ). When r is received, the decoder computes the following (n-k) tuple:
T

s = r. H
= (s1, s2 sn-k)

.. (5.13)

It then follows from Eq (6.9a), that s = 0 if and only if r is a code word and s 0 iffy r is not a
code word. This vector s is called The Syndrome (a term used in medical science referring to
collection of all symptoms characterizing a disease). Thus if s = 0, the receiver accepts r as a valid
code word. Notice that there are possibilities of errors undetected, which happens when e is identical
to a nonzero code word. In this case r is the sum of two code words which according to our linearity
property is again a code word. This type of error pattern is referred to an undetectable error
k
k
pattern. Since there are 2 -1 nonzero code words, it follows that there are 2 -1 error patterns as well.
Hence when an undetectable error pattern occurs the decoder makes a decoding error.
Eq. (6.13) can be expanded as below:

s1 r1 p11 r2 p21 .... rk pk 1 r k 1

From which we have

s2 r1 p12 r2 p22 .... rk pk 2 rk 2

M

r p

nk

M
1,nk

r p
2

2,nk

(5.14)

.... r p
k

k ,nk

r
n

A careful examination of Eq. (6.14) reveals the following point. The syndrome is simply the vector
sum of the received parity digits (rk+1, rk+2 ...r n) and the parity check digits recomputed from the
received information digits (r1, r2 r n).
Example 6.4:
We shall compute the syndrome for the (6, 3) systematic code of Example 5.2. We have

Dept. of ECE/SJBIT

Page 142

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

s = (s1, s2, s3) = (r1, r2, r3, r4, r5, r6)

1
1

0
1

1
0

1
0

0
1

0
0

or s1 = r2 +r3 + r4
s2 = r1 +r3 + r5
s3 = r1 +r2 + r6

In view of Eq. (6.12a), and Eq. (6.9a) we have

T

s = r.H = (v e) H
T

= v .H e.H
or s = e.H

(5.15)

as v.H = O. Eq. (6.15) indicates that the syndrome depends only on the error pattern and not on the
transmitted code word v. For a linear systematic code, then, we have the following relationship
between the syndrome digits and the error digits.
s1 = e1p11 + e2 p 21+ . + ek p k,1 + ek +1
s 2 = e1p12 + e2 p 22 + + ek p k, 2
M

s =ep
n-k

1 1, n-k

+e p
2

M
2, n-k

+e

k +2

+ .. + e p
k

..(5.16)

+e
k, n -k

Thus, the syndrome digits are linear combinations of error digits. Therefore they must provide
us information about the error digits and help us in error correction.
Notice that Eq. (6.16) represents (n-k) linear equations for n error digits an under-determined
set of equations. Accordingly it is not possible to have a unique solution for the set. As the rank of the
k
k
H matrix is k, it follows that there are 2 non-trivial solutions. In other words there exist 2 error patterns
that result in the same syndrome. Therefore to determine the true error pattern is not any easy task
Example 5.5:
For the (6, 3) code considered in Example 6 2, the error patterns satisfy the following equations:
s1 = e2 +e3 +e4 , s2 = e1 +e3 +e5 , s3 = e1 +e2 +e6
Suppose, the transmitted and received code words are v = (0 1 0 1 0 1), r = (0 1 1 1 0 1)
T

Then s = r.H = (1, 1, 0)

Then it follows that:
e2 + e3 +e4 = 1
e1 + e3 +e5 =1

Dept. of ECE/SJBIT

Page 143

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

e1 + e2 +e6 = 0
3

There are 2 = 8 error patterns that satisfy the above equations. They are:
{0 0 1 0 0 0, 1 0 0 0 0, 0 0 0 1 1 0, 0 1 0 0 1 1, 1 0 0 1 0 1, 0 1 1 1 0 1, 1 0 1 0 1 1, 1 1 1 1 1 0}
To minimize the decoding error, the Most probable error pattern that satisfies Eq (6.16) is
chosen as the true error vector. For a BSC, the most probable error pattern is the one that has the
smallest number of nonzero digits. For the Example 6.5, notice that the error vector (0 0 1 0 0 0) has
the smallest number of nonzero components and hence can be regarded as the most probable error
vector. Then using Eq. (6.12) we have
v = r e
= (0 1 1 1 0 1) + (0 0 1 0 0 0) = (0 1 0 1 0 1)
Notice now that v indeed is the actual transmitted code word.
5.6 Minimum Distance Considerations:
The concept of distance between code words and single error correcting codes was first
developed by R .W. Hamming. Let the n-tuples,

= (1, 2 n), = (1, 2 n)

be two code words. The Hamming distance d (,) between such pair of code vectors is defined as
the number of positions in which they differ. Alternatively, using Modulo-2 arithmetic, we have
n

d(,) (j j)

(5.17)

j 1

(Notice that represents the usual decimal summation and is the modulo-2 sum, the EX-OR
function).
The Hamming Weight () of a code vector is defined as the number of nonzero
elements in the code vector. Equivalently, the Hamming weight of a code vector is the distance
between the code vector and the all zero code vector.
Example 6.6: Let = (0 1 1 1 0 1), = (1 0 1 0 1 1)
Notice that the two vectors differ in 4 positions and hence d (, ) = 4. Using Eq (5.17) we find
d (, ) = (0 1) + (1 0) + (1 1) + (1 0) + (0 1) + (1 1)
=

= 4 .. (Here + is the algebraic plus not modulo 2 sum)

Further, () = 4 and () = 4.

Dept. of ECE/SJBIT

Page 144

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

The Minimum distance of a linear block code is defined as the smallest Hamming distance
between any pair of code words in the code or the minimum distance is the same as the smallest
Hamming weight of the difference between any pair of code words. Since in linear block codes, the
sum or difference of two code vectors is also a code vector, it follows then that the minimum
distance of a linear block code is the smallest Hamming weight of the nonzero code vectors in the
code.
The Hamming distance is a metric function that satisfies the triangle inequality. Let, and
be three code vectors of a linear block code. Then
d (, ) + d (, ) d(,)
. (5.18)
From the discussions made above, we may write
d (, ) = ( )
. (5.19)
Example 6.7: For the vectors and of Example 6.6, we have:

= (01), (10), (11) (10), (01) (11)= (11 0 1 1 0)

( ) = 4 = d (,)
If = (1 0 1 01 0), we have d (,) = 4; d (,) = 1; d (,) = 5
Notice that the above three distances satisfy the triangle inequality:
d (, ) + d (,) = 5 = d (,)
d (,) + d (,) = 6 > d (,)
d (,) + d (, ) = 9 > d (,)
Similarly, the minimum distance of a linear block code, C may be mathematically
represented as below:
dmin =Min {d (,):, C, }

.(5.20)

=Min {( ):, C, }
=Min {(v), v C, v 0}

(5 .21)

That is dmin min . The parameter min is called the minimum weight of the linear
code C.The minimum distance of a code, dmin, is related to the parity check matrix, H, of the code in
a fundamental way. Suppose v is a code word. Then from Eq. (6.9a) we have:
0 = v.H

= v1h1 v2h2 . vnhn

Here h1, h2 h n represent the columns of the H matrix. Let vj1, vj2 v jl be the l nonzero
components of v i.e. vj1 = vj2 = . v jl = 1. Then it follows that:

Dept. of ECE/SJBIT

Page 145

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

hj1 hj2 hjl = O

(5.22)

That is if v is a code vector of Hamming weight l, then th ere exist l columns of H such that
the vector sum of these columns is equal to the zero vector. Suppose we form a binary n-tuple of weight
l, viz. x = (x1, x2 x n) whose nonzero components are xj1, xj2 x jl. Consider the product:
T

x.H = x1h1 x2h2 . xnhn = xj1hj1 xj2hj2 . xjlhjl = hj1 hj2 hjl
T

If Eq. (6.22) holds, it follows x.H = O and hence x is a code vector. Therefore, we conclude
that if there are l columns of H matrix whose vector sum is the zero vector then there exists a code
vector of Hamming weight l .
From the above discussions, it follows that:
T

i)

If no (d-1) or fewer columns of H add to O , the all zero column vector, the code has
a minimum weight of at least d.

ii)

The minimum weight (or the minimum distance) of a linear block code C, is the smallest
number of columns of H that sum to the all zero column vector.
011100
For the H matrix of Example 6.3, i.e. H =

101010

, notice that all columns of H are non

110001
zero and distinct. Hence no two or fewer columns sum to zero vector. Hence the minimum weight of
st nd
rd
T
the code is at least 3.Further notice that the 1 , 2 and 3 columns sum to O . Thus the minimum
weight of the code is 3. We see that the minimum weight of the code is indeed 3 from the table of
Example 6.1.
5.6.1 Error Detecting and Error Correcting Capabilities:
The minimum distance, dmin, of a linear block code is an important parameter of the code. To
be more specific, it is the one that determines the error correcting capability of the code. To
understand this we shall consider a simple example. Suppose we consider 3-bit code words plotted at
the vertices of the cube as shown in Fig.6.10.

Dept. of ECE/SJBIT

Page 146

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Clearly, if the code words used are {000, 101, 110, 011}, the Hamming distance between the words
is 2. Notice that any error in the received words locates them on the vertices of the cube which are not
code words and may be recognized as single errors. The code word pairs with Hamming distance = 3
are: (000, 111), (100, 011), (101, 010) and (001, 110). If a code word (000) is received as (100, 010,
001), observe that these are nearer to (000) than to (111). Hence the decision is made that the
transmitted word is (000).
Suppose an (n, k) linear block code is required to detect and correct all error patterns (over a
BSC), whose Hamming weight, t. That is, if we transmit a code vector and the received vector is
= e, we want the decoder out put to be = subject to the condition (e) t.
k

Further, assume that 2 code vectors are transmitted with equal probability. The best decision
for the decoder then is to pick the code vector nearest to the received vector for which the
Hamming distance is the smallest. i.e., d (,) is minimum. With such a strategy the decoder will be
able to detect and correct all error patterns of Hamming weight (e) t provided that the minimum
distance of the code is such that:
dmin (2t + 1)

(5.23)

dmin is either odd or even. Let t be a positive integer such that

(5.24)
2t + 1 dmin 2t + 2
Suppose be any other code word of the code. Then, the Hamming distances
among , and satisfy the triangular inequality:

d(, ) + d(, ) d(,)

(5.25)
Suppose an error pattern of t errors occurs during transmission of . Then the received vector
differs from in t places and hence d(,) = t. Since and are code vectors, it follows from
Eq. (6.24).
d(,) dmin 2t + 1

(5.26)

Combining Eq. (6.25) and (6.26) and with the fact that d(,) = t, it follows that:
d (, ) 2t + 1- t
Hence if t t, then: d (, ) > t

(5.27)

(528)

Eq 6.28 says that if an error pattern of t or fewer errors occurs, the received vector is
closer (in Hamming distance) to the transmitted code vector than to any other code vector of the
code. For a BSC, this means P ( |) > P ( |) for . Thus based on the maximum likelihood
decoding scheme, is decoded as , which indeed is the actual transmitted code word and this
results in the correct decoding and thus the errors are corrected.
On the contrary, the code is not capable of correcting error patterns of weight l>t. To show
this we proceed as below:
Suppose
i)

d (,) = dmin, and let e1 and e2 be two error patterns such that:
e1 e2 =

Dept. of ECE/SJBIT

Page 147

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

ii)

10EC55

e1 and e2 do not have nonzero components in common places. Clearly,

(e1) + (e2) = ( ) = d( ,) = dmin

(5.29)

Suppose, is the transmitted code vector and is corrupted by the error pattern e1. Then the received
vector is:

= e1
and

..

d (, ) = ( ) = (e1)

(5.30)

(5.31)

d (, ) = ()
= ( e1) = (e2)

.(5.32)

If the error pattern e1 contains more than t errors, i.e. (e1) > t, and since 2t + 1 dmin 2t + 2, it
follows

(e2) t- 1

(5.33)

d (, ) d (,)

. (5.34)

This inequality says that there exists an error pattern of l > t errors which results in a received
vector closer to an incorrect code vector i.e. based on the maximum likelihood decoding scheme
decoding error will be committed.
To make the point clear, we shall give yet another illustration. The code vectors and the
received vectors may be represented as points in an n- dimensional space. Suppose we construct two
spheres, each of equal radii, t around the points that represent the code vectors and . Further let
these two spheres be mutually exclusive or disjoint as shown in Fig.6.11 (a).
For this condition to be satisfied, we then require d (,) 2t + 1.In such a case if d (,) t, it
is clear that the decoder will pick as the transmitted vector.

On the other hand, if d (,) 2t, the two spheres around and intersect and if is located as in
Fig. 6.11(b), and is the transmitted code vector it follows that even if d (,) t, yet is as close to
as it is to. The decoder can now pick as the transmitted vector which is wrong. Thus it is
imminent that an (n, k) linear block code has the power to correct all error patterns of weight t or
less if and only if d (,) 2t + 1 for all and. However, since the smallest distance between any
pair of code words is the minimum distance of the code, dmin , guarantees correcting all the error
patterns of

Dept. of ECE/SJBIT

Page 148

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

where

(d

parameter t =

(d

(d

1)

min

1)

min

10EC55

.. (5.35)

denotes the largest integer no greater than the number

min

1
2

( dmin 1 ) . The

1) is called the random-error-correcting capability of the code and the

2
code is referred to as a t-error correcting code. The ( 6, 3) code of Example 6.1 has a minimum
distance of 3 and from Eq. (6.35) it follows t = 1, which means it is a Single Error Correcting
(SEC) code. It is capable of correcting any error pattern of single errors over a block of six digits.
n-k

For an (n, k) linear code, observe that, there are 2

syndromes including the all zero syndrome.
Each syndrome corresponds to a specific error pattern. If j is the number of error locations in the nn
dimensional error pattern e, we find in general, there are
nC j multiple error patterns. It then
j
n
t
follows that the total number of all possible error patterns =
, where t is the maximum number
j
j0
of error locations in e. Thus we arrive at an important conclusion. If an (n, k) linear block code is to
be capable of correcting up tot errors, the total number of syndromes shall not be less than the total
number of all possible error patterns, i.e.
t

2n-k

(5.36)

j
j0
Eq (6.36) is usually referred to as the Hamming bound. A binary code for which the Hamming
Bound turns out to be equality is called a Perfect code.
6.7 Standard Array and Syndrome Decoding:
The decoding strategy we are going to discuss is based on an important property of the
syndrome.
k

Suppose vj , j = 1, 2 2 , be the 2 distinct code vectors of an (n, k) linear block code.

k
Correspondingly let, for any error pattern e, the 2 distinct error vectors, ej, be defined by
ej = e vj , j = 1, 2 2

k
. (5.37)

The set of vectors {ej, j = 1, 2 2 } so defined is called the co- set of the code. That is, a
k
co-set contains exactly 2 elements that differ at most by a code vector. It then fallows that there
n-k
T
are 2 co- sets for an (n, k) linear block code. Post multiplying Eq (6.37) by H , we find
T

ej H = eH vj H
=eH

T
(5.38)

Notice that the RHS of Eq (6.38) is independent of the index j, as for any code word the term
vj H = 0. From Eq (6.38) it is clear that all error patterns that differ at most by a code word have the
same syndrome. That is, each co-set is characterized by a uniqu e syndrome.
T

Dept. of ECE/SJBIT

Page 149

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Since the received vector r may be any of the 2 n-tuples, no matter what the transmitted code
k
word was, observe that we can use Eq (6.38) to partition the received code words into 2 disjoint sets
and try to identify the received vector. This will be done by preparing what is called the standard
array. The steps involved are as below:
k

Step1: Place the 2 code vectors of the

code in a row, with the all
v1 = (0, 0, 0 0 ) = O as the first (left most) element.
n

zero vector

Step 2: From among the remaining (2 2 ) - n tuples, e2 is chosen and placed below the allzero vector, v1. The second row
can now be formed by placing (e2 vj),
k
j = 2, 3 2 under vj
rd

Step 3: Now take an un-used n-tuple e3 and complete the 3

row as in step 2.

Step 4: continue the process until all the n-tuples are used.
The resultant array is shown in Fig. 5.12.

Since all the code vectors, vj, are all distinct, the vectors in any row of the array are also
distinct. For, if two n-tuples in the l-th row are identical, say el vj = el vm, j m; we should have
vj = vm which is impossible. Thus it follows that no two n-tuples in the same row of a standard array
are identical.
Next, let us consider that an n-tuple appears in both l-th row and the m-th row. Then for some
j1 and j2 this implies el vj1 = em vj2, which then implies el = em (vj2 vj1); (remember that X
X = 0 in modulo-2 arithmetic) or el = em vj3 for some j3. Since by property of linear block codes vj3 is
also a code word, this implies, by the construction rules given, that el must appear in the m-th row, which
is a contradiction of our steps, as the first element of the m-th row is em and is an unused vector in the
previous rows. This clearly demonstrates another important property of the array: Every n-tuple appears
in one and only one row.
n-k

From the above discussions it is clear that there are 2

disjoint rows or co-sets in the
k
standard array and each row or co-set consists of 2 distinct entries. The first n-tuple of each co-set,
(i.e., the entry in the first column) is called the Co-set leader. Notice that any element of the co-set
can be used as a co-set leader and this does not change the element of the co-set - it results simply in
a permutation.
T

th

Suppose Dj is the j column of the standard array. Then it follows

Dj = {vj, e2 vj, e3 vj e 2

Dept. of ECE/SJBIT

n-k

vj}

..

6.39)

Page 150

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

n-k

where vj is a code vector and e2, e3, e 2

k

are the co-set leaders.

T

The 2 disjoints columns D1 , D2 D2 k can now be used for decoding of the code. If vj is
the transmitted code word over a noisy channel, it follows from Eq (6.39) that the received vector r is
T
in Dj if the error pattern caused by the channel is a co-set leader. If this is the case r will be decoded
correctly as vj. If not an erroneous decoding will result for, any error pattern e which is not a co-set
leader must be in some co-set and under some nonzero code vector, say, in the i-th co-set and under v
0. Then it follows

= ei vl , and the received vector is r

= vj e = vj (ei vl ) = ei vm
e

received vector is in Dm and it will

Thus
the
be decoded as vm and a decoding error has been committed. Hence it is explicitly clear that Correct
decoding is possible if and only if the error pattern caused by the channel is a co-set leader .
n-k
Accordingly, the 2
co-set leaders (including the all zero vector) are called the Correctable error
n-k
patterns, and it follows Every (n, k) linear block code is capable of correcting 2
error patterns.
So, from the above discussion, it follows that in order to minimize the probability of a
decoding error, The most likely to occur error patterns should be chosen as co-set leaders . For a
BSC an error pattern of smallest weight is more probable than that of a larger weight. Accordingly, when
forming a standard array, error patterns of smallest weight should be chosen as co-set leaders. Then the
decoding based on the standard array would be the minimum distance decoding (the maximum
likelihood decoding). This can be demonstrated as below.
Suppose a received vector r is found in the j
decoded as vj. We have

th

column and l

th

row of the array. Then r will be

d(r, vj) = (r vj ) = (el vj vj ) = (el )

where we have assumed vj indeed is the transmitted code word. Let vs be any other code word, other
than vj. Then
d(r, vs ) = (r vs ) = (el vj vs ) = (el vi )
as vj and vs are code words, vi = vj vs is also a code word of the code. Since el and (el vi ) are in
the same co set and, that el has been chosen as the co-set leader and has the smallest weight it follows
(el ) (el vi ) and hence d(r, vj ) d(r, vs ). Thus the received vector is decoded into a closet code
vector. Hence, if each co-set leader is chosen to have minimum weight in its co-set, the standard array
decoding results in the minimum distance decoding or maximum likely hood decoding.
Suppose a 0, a1, a2 , an denote the number of co-set leaders with weights 0, 1, 2 n. This
set of numbers is called the Weight distribution of the co-set leaders. Since a decoding error wil l
occur if and only if the error pattern is not a co-set leader, the probability of a decoding error for a
BSC with error probability (transition probability) p is given by
n

P(E) 1 a j p (1 p)
j

n j

(5.40)

j 0

Example 6.8:

Dept. of ECE/SJBIT

Page 151

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

For the (6, 3) linear block code of Example 6.1 the standard array, along with the syndrome
table, is as below:

The weight distribution of the co-set leaders in the array shown are a0 = 1, a1 = 6, a2 = 1, a3 = a4 = a5
= a6 = 0.From Eq (6.40) it then follows:
6

P (E) = 1- [(1-p) +6p (1-p) + p (1-p) ]

-2

-3

With p = 10 , we have P (E) = 1.3643879 10

A received vector (010 001) will be decoded as (010101) and a received vector (100 110) will be
decoded as (110 110).
We have seen in Eq. (6.38) that each co-set is characterized by a unique syndrome or there is
a one- one correspondence between a co- set leader (a correctable error pattern) and a syndrome.
n-k
These relationships, then, can be used in preparing a decoding table that is made up of 2
co-set
leaders and their corresponding syndromes. This table is either stored or wired in the receiver. The
following are the steps in decoding:
Step 1: Compute the syndrome s = r. H

Step 2: Locate the co-set leader ej whose syndrome is s. Then ej is assumed to be the error pattern caused
by the channel.
Step 3: Decode the received vector r into the code vector v = r ej
This decoding scheme is called the Syndrome decoding or the Table look up decoding.
Observe that this decoding scheme is applicable to any linear (n, k) code, i.e., it need not necessarily
be a systematic code.
Comments:
1) Notice that for all correctable single error patterns the syndrome will be identical to a
column of the H matrix and indicates that the received vector is in error corresponding to
that column position.
For Example, if the received vector is (010001), then the syndrome is (100). This is identical
th
th
withthe4 column of the H- matrix and hence the 4 position of the received vector is in error.
Hence the corrected vector is 010101. Similarly, for a received vector (100110), the syndrome is 101

Dept. of ECE/SJBIT

Page 152

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

and this is identical with the second column of the H-matrix. Thus the second position of the received
vector is in error and the corrected vector is (110110).
2) A table can be prepared relating the error locations and the syndrome. By suitable combinatorial
circuits data recovery can be achieved. For the (6, 3) systematic linear code we have the following
table for r = (r1 r2 r3 r4 r5 r 6.).

5.8 Hamming Codes:

Hamming code is the first class of linear block codes devised for error correction. The single error
correcting (SEC) Hamming codes are characterized by the following parameters.
m

Code length: n = (2 -1)

m

Number of Information symbols: k = (2 m 1)

Number of parity check symbols :( n k) = m
Error correcting capability: t = 1, (dmin= 3)
The parity check matrix H of this code consists of all the non-zero m-tuples as its columns. In
systematic form, the columns of H are arranged as follows
H = [Q M Im]
Where Im is an identity (unit) matrix of order m m and Q matrix consists of
m

(2 -m-1) columns which are the m-tuples of weight 2 or more. As an illustration for k=4 we have from k
m
= 2 m 1.
m=1

k=0, m=2 k=1, m=3 k=4

3

Thus we require 3 parity check symbols and the length of the code 2 1 = 7 . This results in the (7,
4) Hamming code.
The parity check matrix for the (7, 4) linear systematic Hamming code is then

Dept. of ECE/SJBIT

Page 153

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

The generator matrix of the code can be written in the form

G I2 m m 1 M Q

And for the (7, 4) systematic code it follows:

p1 p2 m1 p3 m2 m3 m4 p4 m5 m6

m7 m8

m9 m10 m11 p5

m12

Where p1, p2, p3 are the parity digits and m1, m2, m3 are the message digits. For example, let us
consider the non systematic (7, 4) Hamming code.
p1 = 1, 3, 5, 7, 9, 11, 13, 15
p2 = 2, 3, 6, 7, 10, 11, 14, 15
p3 = 4, 5, 6, 7, 12, 13, 14, 15
It can be verified that (7, 4), (15, 11), (31, 26), (63, 57) are all single error correcting Hamming
codes and are regarded quite useful.
An important property of the Hamming codes is that they satisfy the condition of Eq. (6.36)
with equality sign, assuming that t=1.This means that Hamming codes are single error correcting
binary perfect codes. This can also be verified from Eq. (6.35)
We may delete any l columns from the parity check matrix H of the Hamming code resulting
m
in the reduction of the dimension of H matrix to m (2 -l-1).Using this new matrix as the parity
check matrix we obtain a shortened Hamming code with the following parameters.
m

Code length: n = 2 -l-1

m

Number of Information symbols:

k=2 -m-l-1

Number of parity check symbols:

nk=m

Minimum distance:

dmin 3

Notice that if the deletion of the columns of the H matrix is proper, we may obtain a Hamming code
with dmin = 4.For example if we delete from the sub-matrix Q all the columns of even weight, we
m-1
obtain an m 2
matrix

H Q :Im

Dept. of ECE/SJBIT

Page 154

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

m-1

-m) columns of odd weight. Clearly no three columns add to zero as all
Where Q contains (2
columns have odd weight .However, for a column in Q , there exist three columns in Im such that four

columns add to zero .Thus the shortened Hamming codes with H as the parity check matrix has
minimum distance exactly 4.
The distance 4 shortened Hamming codes can be used for correcting all single error patterns while
simultaneously detecting all double error patterns. Notice that when single errors occur the
syndromes contain odd number of ones and for double errors it contains even number of ones.
Accordingly the decoding can be accomplished in the following manner.
(1) If s = 0, no error occurred.
(2) If s contains odd number of ones, single error has occurred .The single error pattern pertaining
to this syndrome is added to the received code vector for error correction.
(3) If s contains even number of ones an uncorrectable error pattern has been detected.
Alternatively the SEC Hamming codes may be made to detect double errors by adding an extra
Th
parity check in its (n+1)
position. Thus (8, 4), (6, 11) etc. codes have dmin = 4 and correct
single errors with detection of double errors.
Review Questions:
1. Design a single error correcting code with a message block size of 11 and show that by an
example that it can correct single error.
2. If Ci and Cj an two code vectors in a (n,k) linear block code, show that their sum is also a code
vector.
T

3. Show CH =0 for a linear block code.

4. Prove that the minimum distance of a linear block code is the smallest weight of the non-zero
code vector in the code.
5. What is error control coding? Which are the functional blocks of a communication system that
accomplish this? Indicate the function of each block. What is the error detection and
correction on the performance of communication system?
6. Explain briefly the following:
a. Golay code
b. BCH Code
7. Explain the methods of controlling errors
8. List out the properties of linear codes.
9. Explain the importance of hamming codes & how these can be used for error detection and
correction.
10. Write a standard array for (7.4) code

Dept. of ECE/SJBIT

Page 155

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

UNIT 6

Syllabus:
Binary Cycle Codes, Algebraic structures of cyclic codes, Encoding using an (n- k) bit shift
register, Syndrome calculation. BCH codes. 7 Hours

Text Books:
Digital and analog communication systems, K. Sam Shanmugam, John Wiley,
1996. Digital communication, Simon Haykin, John Wiley, 2003.

Reference Books:
ITC and Cryptography, Ranjan Bose, TMH, II edition, 2007
Digital Communications - Glover and Grant; Pearson Ed. 2nd Ed 2008

Dept. of ECE/SJBIT

Page 156

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

UNIT 6
We are, in general, not very much concerned in our every daily life with accurate transmission
of information. This is because of the redundancy associated with our language-in conversations,
lectures, and radio or telephone communications. Many words or even sentences may be missed still
not distorting the meaning of the message.
However, when we are to transmit intelligence-more information in a shorter time, we wish to
eliminate unnecessary redundancy. Our language becomes less redundant and errors in transmission
become more serious. Notice that while we are talking about numerical data, misreading of even a
single digit could have a marked effect on the intent of the message. Thus the primary objective of
coding for transmission of intelligence would be two fold increase the efficiency and reduce the
transmission errors. Added to this we would like our technique to ensure security and reliability. In
this chapter we present some techniques for source encoding and connection between coding and
information theory in the light of Shannons investigation. The problem of channel encoding- coding
for error detection and correction will be taken up in the next chapter.
6.1 Definition of Codes:
Encoding or Enciphering is a procedure for asso ciating words constructed from a finite alphabet
of a language with given words of another language in a one-to- one manner.
Let the source be characterized by the set of symbols
S= {s1, s2... sq}

(6.1)

We shall call S as the Source alphabet. Consider another set, X, comprising of r symbols.
X={x1, x2x r}

(6.2)

We shall call X as the code alphabet. We define coding as the

mapping of all possible
sequences of symbols of S into sequences of symbol of X. In other words coding means
representing each and every symbol of S by a sequence of symbols of X such that there shall be a oneto-one relationship Any finite sequence of symbols from an alphabet will be called a Word. Thus any
sequence from the alphabet X forms a code word. The total number of symbols contained in the
word will be called word length. For example the sequences { x1 ; x1x3x4 ; x3x5x7x9 ; x1x1x2x2x2}
form code words. Their word lengths are respectively1; 3; 4; and 5.The sequences
{100001001100011000} and {1100111100001111000111000} are binary code words with word lengths
18 and 25 respectively.
6.2

Basic properties of codes:

The definition of codes given above is very broad and includes many undesirable properties.
In order that the definition is useful in code synthesis, we require the codes to satisfy certain
properties. We shall intentionally take trivial examples in order to get a better understanding of the
desired properties.

Dept. of ECE/SJBIT

Page 157

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

1. Block codes:
A block code is one in which a particular message of the source is always encoded into the
same fixed sequence of the code symbol. Although, in general, block m eans a group having
identical property we shall use the word here to mean a fixed sequence only. Accordingly, the code can
be a fixed length code or a variable length code and we shall be concentrating on the latter type in
this chapter. To be more specific as to what we mean by a block code, consider a communication system
with one transmitter and one receiver. Information is transmitted using certain set of code words. If the
transmitter wants to change the code set, first thing to be done is to inform the receiver. Other wise the
receiver will never be able to understand what is being transmitted. Thus, until and unless the receiver is
informed about the changes made you are not permitted to change the code set. In this sense the code
words we are seeking shall be always finite sequences of the code alphabet-they are fixed sequence
codes.
Example 6.1: Source alphabet is S = {s1, s2, s3, s4}, Code alphabet is X = {0, 1} and The Code words
are: C = {0, 11, 10, 11}
2. Non singular codes:
A block code is said to be non singular if all the words of the code set X1, are distinct. The
codes given in Example 6.1 do not satisfy this property as the codes for s2 and s4 are not different.
We can not distinguish the code words. If the codes are not distinguishable on a simple inspection we
say the code set is singular in the small. We modify the code as below.
Example 6.2: S = {s1, s2, s3, s4}, X = {0, 1}; Codes, C = {0, 11, 10, 01}
However, the codes given in Example 6.2 although appear to be non-singular, upon transmission
would pose problems in decoding. For, if the transmitted sequence is 0011, it might be interpreted as
s1 s1 s4 or s2 s4. Thus there is an ambiguity about the code. No doubt, the code is non-singular in the
small, but becomes Singular in the large.
3. Uniquely decodable codes:
A non-singular code is uniquely decipherable, if every word immersed in a sequence of
th
words can be uniquely identified. The n extension of a code, that maps each message into the code
words C, is defined as a code which maps the sequence of messages into a sequence of code words.
This is also a block code, as illustrated in the following example.
Example 6.3: Second extension of the code set given in Example 6.2.
2

S ={s1s1,s1s2,s1s3,s1s4; s2s1,s2s2,s2s3,s2s4,s3s1,s3s2,s3s3,s3s4,s4s1,s4s2,s4s3,s4s4}
Source
Codes
Source
Codes
Source
Codes
Source
Codes
Symbols
Symbols
Symbols
Symbols
s1s1
s2s1
s3s1
s4s1
00
110
100
010
s1s2
011
s2s2
1111
s3s2
1011
s4s2
0111
s1s3
010
s2s3
1110
s3s3
1010
s4s3
0110
s 1s 4
001
s2s4
1101
s 3s 4
1001
s4s4
0101
Notice that, in the above example, the codes for the source sequences, s1s3 and s4s1 are not
distinct and hence the code is Singular in the Large. Since such singularity properties introduce

Dept. of ECE/SJBIT

Page 158

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

ambiguity in the decoding stage, we therefore require, in general, for unique decidability of our codes
th
that The n extension of the code be non-singular for every finite n.
4. Instantaneous Codes:
A uniquely decodable code is said to be instantaneous if the end of any code word is
recognizable with out the need of inspection of succeeding code symbols. That is there is no time lag
in the process of decoding. To understand the concept, consider the following codes:
Example 6.4:
Source symbols
s1
s2
s3
s4

Code A
00
01
10
11

Code B
0
10
110
1110

Code C
0
01
011
0111

Code A undoubtedly is the simplest possible uniquely decipherable code. It is non- singular and all the
code words have same length. The decoding can be done as soon as we receive two code symbols without
any need to receive succeeding code symbols.
Code B is also uniquely decodable with a special feature that the 0`s indicate the termination of a code
word. It is called the comma code. When scanning a sequence of code symbols, we may use the
comma to determine the end of a code word and the beginning of the other. Accordingly, notice that the
codes can be decoded as and when they are received and there is, once again, no time lag in the decoding
process.
Where as, although Code C is a non- singular and uniquely decodable code it cannot be decoded word
by word as it is received. For example, if we receive 01, we cannot decode it as s2 until we
receive the next code symbol. If the next code symbol is 0,indeed the previous word corresponds to
s2, while if it is a 1 it may be the symbol s3; which can be concluded so if only if we receive a 0in the
fourth place. Thus, there is a definite time lag before a word can be decoded. Such a time waste is not
there if we use either Code A or Code B. Further, what we are envisaging is the property by which a
sequence of code words is uniquely and instantaneously decodable even if there is no spacing between
successive words. The common English words do not posses this property. For example the words
FOUND, AT and ION when transmitted without spacing yield, at the receiver, an altogether new
word FOUNDATION! A sufficient condition for such property is that
No encoded word can be obtained from each other by the addition of more letters . This property
is called prefix property.
Let Xk = xk1xk2.x km, be a code word of some source symbol sk. Then the sequences of
code symbols, (xk1xk2.x k j), j m, are called prefixes of the code word. Notice tha t a code
word of length m will have m prefixes. For example, the code word 0111 has four prefixes, viz;
0, 01, 011 and 0111.The complete code word is also regarded as a prefix.
Prefix property: A necessary and sufficient condition for a code t o be instantaneous is
that no complete code word be a prefix of some other code word.

Dept. of ECE/SJBIT

Page 159

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

The sufficiency condition follows immediately from the definition of the word Instantaneous. If
no word is a prefix of some other word, we can decode any received sequence of code symbols
comprising of code words in a direct manner. We scan the received sequence until we come to
subsequence which corresponds to a complete code word. Since by assumption it is not a prefix of
any other code word, the decoding is unique and there will be no time wasted in the process of
decoding. The necessary condition can be verified by assuming the contrar y and deriving its
contradiction. That is, assume that there exists some word of our code, say xi, which is a prefix of
some other code word xj. If we scan a received sequence and arrive at a subsequence that corresponds
to xi, this subsequences may be a complete code word or it may just be the first part of code word xj.
We cannot possibly tell which of these alternatives is true until we examine some more code symbols
of the sequence. Accordingly, there is definite time wasted before a decision can be made and hence
the code is not instantaneous.
5. Optimal codes:
An instantaneous code is said to be optimal if it has minimum average word length, for a source
with a given probability assignment for the source symbols. In such codes, source symbols with
higher probabilities of occurrence are made to correspond to shorter code words. Suppose that a
source symbol si has a probability of occurrence Pi and has a code word of length li assigned to it,
while a source symbol sj with probability Pj has a code word of length lj. If Pi >Pj then let li<l j. For
the two code words considered, it then follows, that the average length L1 is given by
L1 = Pili + Pjlj

..

(6.3)

Now, suppose we interchange the code words so that the code word of length lj corresponds to si and
that of length li corresponds to sj. Then, the average length becomes
L2 = Pilj + Pjli

(6 .4)

It then follows, L2 L 1 = Pi (lj l i) + Pj (li l j)

= (Pi P j) (lj li)

(6 .5)

Since by assumption Pi>Pj and li<lj, it is clear that (L2 L 1) is positive. That is assignment of source
symbols and code word length corresponding to the average length L1 is shorter, which is the
requirement for optimal codes.
A code that satisfies all the five properties is called an irreducible code.
All the above properties can be arranged as shown in Fig 5.1 which serves as a quick reference of the
basic requirements of a code. Fig 5.2 gives the requirements in the form of a Tree diagram. Notice
that both sketches illustrate one and the same concept.

Dept. of ECE/SJBIT

Page 160

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

6.3 Construction of Instantaneous Codes:

Consider encoding of a 5 symbol source into Binary instantaneous codes i.e.
S = {s1, s2, s3, s4, s5}; X = {0, 1}
We may start by assigning 0 to s1
i.e. s1 0
If this is the case, to have prefix property, all other source symbols must correspond to code words
beginning with 1. If we let s2 correspond to 1, we would be left with no code symbol for encoding
the remaining three source symbols. We might have
s2 10
This in turn would require the remaining code words to start with 11. If
s3 110;
then the only 3 bit prefix unused is 111 and we might set
s4 1110
s5 1111

Dept. of ECE/SJBIT

Page 161

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

In the above code, notice that the starting of the code by letting s1 correspond 0 has cut down the
number of possible code words. Once we have taken this step, we are restricted to code words
starting with 1. Hence, we might expect to have more freedom if we select a 2-binit code word for
s1. We now have four prefixes possible 00, 01, 10 and 11; the first three can be directly assigned to s1, s2
and s3. With the last one we construct code words of length 3. Thus the possible instantaneous code is
s1 00
s2 01
s3 10
s4 110
s5 111

Thus, observe that shorter we make the first few code words, the longer we will have to make the
later code words.
One may wish to construct an instantaneous code by pre-specifying the word lengths. The
necessary and sufficient conditions for the existence of such a code are provided by the Kraft
Inequality .
6.3.1 Kraft Inequality:
Given a source S = {s1, s2s q}.Let the word lengths of the codes corresponding to these symbols be
l1, l2 .l q and let the code alphabet be X = {x1, x2 x
r}. Then, an instantaneous code for the
source exists iffy
q

lk

..

k 1

(6.6)

Eq (6.6) is called Kraft Inequality (Kraft 1949).

Example 6.5:
A six symbol source is encoded into Binary codes shown below. Which of these codes are
instantaneous?
Source
symbol
s1
s2
s3
s4
s5
s6
6
2

l k

k 1

Code A
00
01
10
110
1110
1111
1

Code B

Code C

Code D

Code E

0
1000
1100
1110
1101
1111
13
<1
16

0
10
110
1110
11110
11111

0
1000
1110
111
1011
1100
7 <1

0
10
110
1110
11110
1111
1 1 > 1

32

As a first test we apply the Kraft Inequality and the result is accordingly tabulated. Code E does not
satisfy Kraft Inequality and it is not an instantaneous code.

Dept. of ECE/SJBIT

Page 162

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Next we test the prefix property. For Code D, notice that the complete code word for the symbol s4 is
a prefix of the code word for the symbol s3. Hence it is not an instantaneous code. However, Code A,
Code B and Code C satisfy the prefix property and are therefore they are instantaneous codes.
Example 6.6:
Given S={s1,s2,s3,s4,s5,s6,s7,s8,s9} and X={0,1}. Further if l1=l2=2 and
l3=l4=l5 =l6=l7=l8=l9=k. Then from Kraft inequality, we have
-2

-2

-k

2 +2 +72
-k

1 7 2 0.5 or

0.5

-k

kmin = 4
k log214=3.807
Clearly, if k <4, it is not possible to construct an instantaneous binary code. Thus if
k 4, Kraft inequality tells us that an instantaneous code does exist but does not tell us how to
construct such a code. The codes for the symbols when k=4 are shown below:
s1 0 0
s4 1 0 0 1
s2 0 1
s5 1 0 1 0
s3 1 0 0 0 s6 1 0 1 1

s7 1 1 0 0
s8 1 1 1 0
s9 1 1 1 1

6.3.2 McMillans Inequality:

Since Instantaneous codes form a sub-class of uniquely decodable codes, we can construct a uniquely
decodable code with word lengths l1, l2. lq satisfying the Kraft inequality. McMillan, (1956) first
proved the necessary part of the inequality and t he inequality is also called by his name.
Both inequalities are one and the same. Only difference is in the approach to derive the inequality.
Since the ideas were totally different and derived independently, the inequality is famous, now-adays, as Kraft-McMillan inequality or the K-M inequality. The two implications of the K-M
inequality are:
(i) Given a set of word lengths, is it possible to construct an instantaneous code?
Yes-if and only if the word lengths satisfy the K-M inequality.
(ii) Whether an already existing code is uniquely decodable?
Yes if and only if it satisfies the K-M inequality.
Observe the importance of the second implication- For a code to be uniquely decodable its
th
n extension should be nonsingular. It is not possible to construct all extensions to test this property. Just
apply the K-M inequality!
6.3.3 Code Efficiency and Redundancy:
Consider a zero memory source, S with q-symbols {s1, s2 sq} and symbol probabilities {p1, p2
pq} respectively. Let us encode these symbols into r- ary codes (Using a code alphabet of

Dept. of ECE/SJBIT

Page 163

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

r- symbols) with word lengths l1, l2l q .We shall find a lower bound for the average length of the code
words and hence define efficiency and redundancy of the code.
q

Let Q1, Q2 Qq be any set of numbers such that Qk 0 and Qk 1 . Consider the quantity
k 1

H(S)p

log 1
Qk

k 1
q
pk log

pk

k 1
q

log

pk

k 1

Qk
p

loge pk ln

k 1

1 p
k
Qk k 1
q
Qk

loge

p
k

Qk
P

pk

k 1
q

log

loge Qk

k 1

pk

k 1

.........( as Pk 1 and Qk 1,by assumption )

k 1

k 1

q
H ( S ) P log 1
(6.7)
k
Qk
k 1
Equality holds iffy Qk = pk. Eq. (5.21) is valid for any set of numbers Qk that are non negative and
sum to unity. We may then choose:

Thus it follows that

l k

qr
r

(6.8)

l k

k 1

and obtain
H(s)

p log r

k 1

lk

l k

log r log r

k 1

q
lk

k 1
lk

q
k 1

i .e H ( S ) log r pk lk log r
k 1

l k

.. (6.9)

k 1

Defining
q

L p k lk

(6.10)

k 1

Which gives the average length of the code words, and observing that the second term in Eq.(5.23) is
q
r lk

1 and logarithm of a number

either negative or at most zero (as from Kraft Inequality
k 1

which is less than unity is always negative) we conclude:

(6.11)
H(S) L log r
Or L H ( S )
.
(6.12)
log r
Notice that achievement of the lower bound is possible for the particular case when:
q

i) r

lk

and (ii)

k 1

Dept. of ECE/SJBIT

pk r

lk

Page 164

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

As, if this is the case Qk

that:
1
lk
p = pk, hence for the equality to hold it follows that we must choose lk such
logr

(6.13)

and

lk must be an integer for all k = 1, 2 q.

Eq. (6.11) can be re-written as:

H ( S ) log r
L

(6.14)

The LHS of Eq. (6.14) is simply (The entropy of the source in bits per source symbol) (no. of code
symbols per source symbol) or bits per code symbol; which is nothing but the actual entropy of the
code symbols. RHS is the maximum value of this entropy when the code symbols are all equiprobable. Thus we can define the code efficiency based on Eq. (6.14) as below:
Code efficiency is the ratio of the average information per symbol of the encoded language to the
maximum possible information per code symbol. Mathematically, we write
Code efficiency, H ( S ) : log r
c

Or

L
H(S)

(6.15)

L log r
Accordingly, Redundancy of the code, Ec=1-c

(6.16).

Example 6.7:
1
Let the source have four messages S= {s1, s2, s3, s4} with P=

1 1
,

4 8

1
8

H(S) = 1 log 2 log 4 2 x 1 log 8 =1.75 bits/sym.

2
8
If S itself is assumed to be the alphabet then we have
L=1, r=4 and 1.75 0.875 , i.e. 87.5% , and Ec=12.5%
c
c
log 4
Suppose the messages are encoded into a binary alphabet, X= {0, 1} as
pk
s1
s2
s3
s4

Code
1/2
1/4
1/8
1/8
4

We have L= l

k 1

lk
0
10
110
111

l1=1
l2=2
l3=3
l4=3

p =1. 1 2. 1 3. 1 3. 1 =1.75binits/symbol
2
4
8
8
k

Dept. of ECE/SJBIT

Page 165

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

1.75
Since r=2, c H ( S )
=1
L log r 1.75 log 2 2
i.e.c 100% , and hence Ec=1- c 0%
Thus by proper encoding, the efficiency can be increased.
Example 6.8:
1 1 1 1

Suppose the probabilities of the messages in Example 6.7 are changed to P=

Then, H(S) =2

1
3

log 3 2

1
6

log 6

=1.918
bits/sym 3
For the codes listed in Example 6.7, we have:
L= 1. 1 2. 1 3. 1 3. 1 = 2binits / symbol, and
log 3+

, , ,

3 3 6 6

c H ( S )
L log r

1.918 0.959 or 95.9%

2 log 2 2

Ec =1-c 0.041 or 4.1%

Notice that in Example 6.7, the equality L= H(S) /log r is strict since lk = log1/pk and in Example 6.8,
the inequality L> H(S)/log r is strict as lk log1/pk.

6.3.4 Shannons First Theorem (Noiseless Coding Theorem):

Eq. (6.12) suggests the lower bound on L, the average word length of the code, expressed as a
fraction of code symbols per source symbol. However, we know that each individual code word will
have an integer number of code symbols. Thus, we are faced with the problem of what to choose for
the value of lk, the number of code symbols in the code word corresponding to a source symbol sk,
1
when the quantity in Eq. (6.13)

viz. lk = logr

is not an integer. Su ppose we choose lk to be

the next integer value greater than log r

log r

1
pk

, i.e.

pk

lk logr

1
1
pk

(6.17)

Eq. (6.17) also satisfies the Kraft inequality, since the left inequality gives:
q
q
1 r lk , or p r lk . from which
p 1 r lk
pk

Further , since logr

k 1

log

k 1

1
2

pk

, Eq (6.17) can be re-written as:

pk

log2 r
log( 1 / pk ) lk log( 1 / pk ) 1
. (6.18)
logr
logr
Multiplying Eq. (6.18) throughout by pk and summing for all values of k, we have

Dept. of ECE/SJBIT

Page 166

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

pk log 1
pk

pk lk

logr

k 1

10EC55

k 1

pk log 1
pk

k 1

logr

pk , or
k 1

H(S) L H(S) 1
logr
logr

(6.19)

..

th

To obtain better efficiency, one will use the n extension of S, giving Ln as the new average word
n
length. Since Eq. (5.33) is valid for any zero- memory source, it is also valid for S , and hence, we
have
n

H ( S ) L H ( S ) 1
n
logr
logr

(6.20)

(6.21)

Since H (S ) = n H(S), Eq. (6.20 reduces to

L
H(S)
H(S) 1
n

logr
n
logr
n
It follows from Eq. (6.21) with n

lim Ln H ( S )
logr
n n

..

(6.22)

Here Ln/n is the average number of code alphabet symbols used per single symbol of S, when the
n
Ln L, where L is the
input to the encoder is n-symbol message for the extended source S . But
n
Ln L. The code capacity is now
average word length for the source S and in general
n
Ln
log r C bits/message of the channel and for successful transmission of messages through

n
the channel we have:
H(S)

L
n

logr C bits/message

.. (6.23)

Eq. (6.21) is usually called the Noiseless coding Theorem and is the essence of Shannons First
not considered any effects
Fundamental Theorem. Notice that in the above discussions we have
of noise on the codes. The emphasis is only on how to most efficiently
encode our source. The
theorem may be stated as belo
CYCLIC CODES
"Binary cyclic codes form a sub class of linear block codes. Majority of important linear
block codes that are known to-date are either cyclic codes or closely related to cyclic codes. Cyclic
codes are attractive for two reasons: First, encoding and syndrome calculations can be easily

Dept. of ECE/SJBIT

Page 167

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

implemented using simple shift registers with feed back connections. Second, they posses well
defined mathematical structure that permits the design of higher-order error correcting codes.
A binary code is said to be "cyclic" if it satisfies:
1. Linearity property sum of two code words is also a code word.
2. Cyclic property Any lateral shift of a code word is also a code word.
The second property can be easily understood from Fig, 7.1. Instead of writing the code as a
row vector, we have represented it along a circle. The direction of traverse may be either clockwise or
counter clockwise (right shift or left shift).
For example, if we move in a counter clockwise direction then starting at A the code word
is 110001100 while if we start at B it would be 011001100. Clearly, the two code words are related in that
one is obtained from the other by a cyclic shift.

If the n - tuple, read from A in the CW direction in Fig 7.1,

v = (vo, v1, v2, v3, vn-2, vn-1)
is a code vector, then the code vector, read from B, in the CW direction, obtained by a one bit cyclic
right shift:
(7.1)

(1)

= (vn-1 , vo, v1, v2, v n-3,vn-2,)

(7.2)

is also a code vector. In this way, the n - tuples obtained by successive cyclic right shifts:

v(2) = (vn-2, vn-1, vn, v0, v1

vn-3)

v(3) = (vn-3 ,vn-2, vn-1, vn, .... vo, v1, vn-4)

M
(i)
v = (vn-i, vn-i+1,v n-1, vo, v1,. v n-i-1)

a)

(7.3

(7.3
(7.3

b)
c)

are all code vectors. This property of cyclic codes enables us to treat the elements of each code vector
as the co-efficients of a polynomial of degree (n-1).
This is the property that is extremely useful in the analysis and implementation of these
codes. Thus we write the "code polynomial' V(X) for the code in Eq (7.1) as a vector polynomial as:

Dept. of ECE/SJBIT

Page 168

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

V(X) = vo + v1 X + v2 X + v3 X ++ v i-1 X

i-1

+... + vn-3 X

n-3

+ vn-2 X

n-2

+ vn-1 X

n-1

.. (7.4)

Notice that the co-efficients of the polynomial are either '0' or '1' (binary codes), i.e. they belong to
GF (2) as discussed in sec 6.7.1.
. Each power of X in V(X) represents a one bit cyclic shift in time.
. Therefore multiplication of V(X) by X maybe viewed as a cyclic shift or rotation to the right subject
n
to the condition X = 1. This condition (i) restores XV(X) to the degree (n-1) (ii) Implies that right
most bit is fed-back at the left.
n

. This special form of multiplication is called "Multiplication modulo X + 1

. Thus for a single shift, we have
2

XV(X) = voX + v1 X + v2 X + ........ + vn-2 X

n-1

+ vn-1 X

(+ vn-1 + vn-1) (Manipulate A + A =0 Binary Arithmetic)

2

= vn-1 + v0 X + v1 X + + vn-2 X

n-1

+ vn-1(X + 1)

(1)

=V
(X) = Remainder obtained by dividing XV(X) by X + 1
(Remember: X mod Y means remainder obtained after dividing X by Y)
Thus it turns out that
V

(1)

(X) = vn-1 + vo X + v1 X +..... + vn-2 X

n-1

I
(1)
is the code polynomial for v . We can continue in this way to arrive at a general format:
i

X V(X) = V

(i)

(X) + q (X) (X + 1)

Remainder Quotient

(7.5)

(7.6)

Where
(i)

V (X) = vn-i + vn-i+1X + vn-i+2X + v n-1X + v 0X

(7.7)
7.1 Generator Polynomial for Cyclic Codes:

i-1

+v1X

i+1

+v n-i-2X

n-2

+vn-i-1X

n-

An (n, k) cyclic code is specified by the complete set of code polynomials of degree (n-1)
and contains a polynomial g(X), of degree (n-k) as a factor, called the "generator polynomial" of the
code. This polynomial is equivalent to the generator matrix G, of block codes. Further, it is the only
polynomial of minimum degree and is unique. Thus we have an important theorem
n

Theorem 7.1 "If g(X) is a polynomial of degree (n-k) and is a factor of (X +1) then g(X) generates an
(n, k) cyclic code in which the code polynomial V(X) for a data vector u = (u0, u1 u k -1) is generated
by
V(X) = U(X) g(X)

Dept. of ECE/SJBIT

..

(7.8)

Page 169

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

Where

10EC55

U(X) = u0 + u1 X + u2 X + ... + uk-1 X

k-I

..

(7.9)

is the data polynomial of degree (k-1).

The theorem can be justified by Contradiction: - If there is another polynomial of same degree, then
add the two polynomials to get a polynomial of degree < (n, k) (use linearity property and binary
arithmetic). Not possible because minimum degree is (n-k). Hence g(X) is unique
k

Clearly, there are 2 code polynomials corresponding to 2 data vectors. The code vectors
corresponding to these code polynomials form a linear (n, k) code. We have then, from the theorem
nk 1

g( X ) 1 gi X X
i

nk

(7.10)

i 1

Asg(X) = go + g1 X + g2 X +.

+ gn-k-1 X

n-k-1

+ gn-k X

n-k

(7.11)

is a polynomial of minimum degree, it follows that g0

= gn-k = 1 always and the remaining coefficients may be either' 0' of '1'. Performing the multiplication said in Eq (7.8) we have:
k-1

U (X) g(X) = uo g(X) + u1 X g(X) ++u k-1X

g(X)

(7.12)

Suppose u0=1 and u1=u2= =u k-1=0. Then from Eq (7.8) it follows g(X) is a code word polynomial
of degree (n-k). This is treated as a basis code polynomial (All rows of the G matrix of a block code,
being linearly independent, are also valid code vectors and form Basis vectors of the code).
i
Therefore from cyclic property X g(X) is also a code polynomial. Moreover, from the linearity
property - a linear combination of code polynomials is also a code polynomial. It follows therefore
that any multiple of g(X) as shown in Eq (7.12) is a code polynomial. Conversely, any binary
polynomial of degree (n-1) is a code polynomial if and only if it is a multiple of g(X). The code
words generated using Eq (7.8) are in non-systematic form. Non systematic cyclic codes can be
generated by simple binary multiplication circuits using shift registers. .
In this book we have described cyclic codes with right shift operation. Left shift version can
be obtained by simply re-writing the polynomials. Thus, for left shift operations, the various
polynomials take the following form
U(X) = uoX

k-1

+ u 1X

n-1

V(X) = v0 X
g(X) = g0X

n-k

k-2

+ + u

n-2

+ v 1X

+ g1X
nk

n-k-1

= X nk gi X
i 1

+. + v

k-2X
n-2X

+ uk-1

+ vn-1

+..+g n-k-1 X + gn-k

nk i

gnk

..

(7.13
(7.13

a)
b)

(7.13

c)
(7.13d)

Other manipulation and implementation procedures remain unaltered.

7.2 Multiplication Circuits:

Dept. of ECE/SJBIT

Page 170

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Construction of encoders and decoders for linear block codes are usually constructed with
combinational logic circuits with mod-2 adders. Multiplication of two polynomials A(X) and B(X)
and the division of one by the other are realized by using sequential logic circuits, mod-2 adders and
shift registers. In this section we shall consider multiplication circuits.
As a convention, the higher-order co-efficients of a polynomial are transmitted first. This is
the reason for the format of polynomials used in this book.
2

n-1

For the polynomial: A(X) = a0 + a1 X + a2 X +...+ an-1X

(7.14)

where ais are either a ' 0' or a '1', the right most bit in the sequence (a0, a1, a2 ... an-1) is transmitted
first in any operation. The product of the two polynomials A(X) and B(X) yield:
C(X) = A(X) B(X)
2

= (a0 + a1 X + a2 X + .................. + a n-1X

2

n-1

m-1

) (b0 + b1 X + b2X ++ b m-1 X

= a0b0+ (a1b0+a0b1) X + (a0b2 + b0a2+a1b1) X +. + (a n-2bm-1+ an-1bm-2) X

n+m -3

+an-1bm-1X

n+m -2

This product may be realized with the circuits of Fig 7.2 (a) or (b), where A(X) is the input and the coefficient of B(X) are given as weighting factor connections to the mod - 2 .adders. A '0' indicates no
connection while a '1' indicates a connection. Since higher order co-efficients are first sent, the
highest order co-efficient an-1 bm-1 of the product polynomial is obtained first at the output of Fig
n+m-3
is obtained as the sum of {an-2bm-1 + an-1 bm-2}, the first
7.2(a). Then the co-efficient of X
term directly and the second term through the shift register SR1. Lower order co-efficients are then
generated through the successive SR's and mod-2 adders. After (n + m - 2) shifts, the SR's contain {0,
0 0, a 0, a1} and the output is (a0 b1 + a1 b0) which is the co-efficient of X. After (n + m-1) shifts,
the SR's contain (0, 0, 0,0, a0) and the out put is a0b0. The product is now complete and the contents
of the SR's become (0, 0, 0 0, 0 ). Fig 7.2(b) performs the multiplication in a similar way but the
arrangement of the SR's and ordering of the co-efficients are different (reverse order!). This
modification helps to combine two multiplication operations into one as shown in Fig 7.2(c).
From the above description, it is clear that a non-systematic cyclic code may be generated using (n-k)
shift registers. Following examples illustrate the concepts described so far.

Dept. of ECE/SJBIT

Page 171

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Example 7.1: Consider that a polynomial A(X) is to be multiplied by

3

B(X) = 1 + X + X + X + X

The circuits of Fig 7.3 (a) and (b) give the product C(X) = A(X). B(X)

Example 7.2: Consider the generation of a (7, 4) cyclic code. Here (n- k) = (7-4) = 3 and we have to
n
7
find a generator polynomial of degree 3 which is a factor of X + 1 = X + 1.
7

To find the factors of degree 3, divide X +1 by X +aX +bX+1, where 'a' and 'b' are binary
2
numbers, to get the remainder as abX + (1 +a +b) X+ (a+b+ab+1). Only condition for the remainder to
be zero is a +b=1 which means either a = 1, b = 0 or a = 0, b = 1. Thus we have two possible
polynomials of degree 3, namely
3

g1 (X) = X + X + 1 and g2 (X) = X +X+1

Dept. of ECE/SJBIT

Page 172

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

In fact, X + 1 can be factored as:

7

(X +1) = (X+1) (X +X +1) (X +X+1)

Thus selection of a 'good' generator polynomial seems to be a major problem in the design of cyclic
codes. No clear-cut procedures are available. Usually computer search procedures are followed.
3

Let us choose g (X) = X + X + 1 as the generator polynomial. The encoding circuits are shown in
Fig 7.4(a) and (b).

To understand the operation, Let us consider u = (10 1 1) i.e.

2

U (X) = 1 +X +X .
We have

V (X) = (1 +X +X ) (1 +X+X ).
3

= 1 +X +X +X+X +X +X +X +X
2

=1+X+X +X +X +X +X

because (X + X =0)

=> v = (1 1 1 1 1 1 1)
The multiplication operation, performed by the circuit of Fig 7.4(a), is listed in the Table below step
by step. In shift number 4, 000 is introduced to flush the registers. As seen from the tabulation the
product polynomial is:
2

V (X) = 1 +X+X +X +X +X +X ,
and hence out put code vector is v = (1 1 1 1 1 1 1), as obtained by direct multiplication. The reader
can verify the operation of the circuit in Fig 7.4(b) in the same manner. Thus the multiplication
circuits of Fig 7.4 can be used for generation of non-systematic cyclic codes.

Table showing sequence of computation

Dept. of ECE/SJBIT

Page 173

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

Shift
Number
0
1
2
3
*4
5
6
7
7.3

10EC55

Input
Queue
0001011
000101
00010
0001
000
00
0
-

Bit
shifted
IN
1
1
0
1
0
0
0

Contents of shift
registers.
SRI
SR2
SR3
0
0
0
1
0
0
1
1
0
0
1
1
1
0
1
0
1
0
0
0
1
0
0
0

Out
put
1
1
1
1
1
1
1

Remarks

Circuit In reset mode

6
Co-efficient of X
5
Co-efficient of X
4
X co-efficient
3
X co-efficient
2
X co-efficient
1
X co-efficient
0
X co-efficient

Dividing Circuits:

As in the case of multipliers, the division of A (X) by B (X) can be accomplished by using shift
registers and Mod-2 adders, as shown in Fig 7.5. In a division circuit, the first co-efficient of the
quotient is (an-1 (bm -1) = q1, and q1.B(X) is subtracted from A (X). This subtraction is carried out by
the feed back connections shown. This process will continue for the second and subsequent terms.
However, remember that these coefficients are binary coefficients. After (n-1) shifts, the entire
quotient will appear at the output and the remainder is stored in the shift registers.

It is possible to combine a divider circuit with a multiplier circuit to build a composite

multiplier-divider circuit which is useful in vari ous encoding circuits. An arrangement to
accomplish this is shown in Fig 7.6(a) and an illustration is shown in Fig 7.6(b).
We shall understand the operation of one divider circuit through an example. Operation of
other circuits can be understood in a similar manner.
Example7.3:
3

Let A(X) = X +X +X , A= (0001011), B(X) = 1 +X+X . We want to find the quotient and
remainder after dividing A(X) by B(X). The circuit to perform this division is shown in Fig 7.7, drawn
using the format of Fig 7.5(a). The operation of the divider circuit is listed in the table:

Dept. of ECE/SJBIT

Page 174

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Table Showing the Sequence of Operations of the Dividing circuit

Shift
Number
0
1
2
3
4
5
6
7

Input
Queue
0001011
000101
00010
0001
*000
00
0
-

Bit
shifted
IN
1
1
0
1
0
0
0

Contents of shift
Registers.
SRI
SR2 SR3
0
0
0
1
0
0
1
1
0
0
1
1
0
1
1
1
1
1
1
0
1
1
0
0

Out
put
0
0
0
1
1
1
1

Remarks

Circuit in reset mode

6
Co-efficient of X
5
Co-efficient of X
4
X co-efficient
X3 co-efficient
2
X co-efficient
X1 co-efficient
o
X co-efficient

The quotient co-efficients will be available only after the fourth shift as the first three shifts
result in entering the first 3-bits to the shift registers and in each shift out put of the last register, SR3,
is zero.
The quotient co-efficient serially presented at the out put are seen to be (1111) and hence the
2
3
quotient polynomial is Q(X) =1 + X + X + X . The remainder co-efficients are (1 0 0) and the
remainder polynomial is R(X) = 1. The polynomial division steps are listed in the next page.
Division Table for Example 7.3:

Dept. of ECE/SJBIT

Page 175

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

7.4

Systematic Cyclic Codes:

Let us assume a systematic format for the cyclic code as below:

v = (p0, p1, p2 p n-k-1, u0, u1, u2 u k-1)

(7.15)

The code polynomial in the assumed systematic format becomes:

2

V(X) = p0 + p1X + p2X + +p n-k-1X

= P(X) + X

n-k

n-k-1

U(X)

+u0X

n-k

+ u1 X

n-k+1

+ +u k-1X

n-1

...

(7.16)
(7.17)

Since the code polynomial is a multiple of the generator polynomial we can write:

V (X) = P (X) +X

nk

n-k

U(X)

U (X) = Q (X) g (X)

Q( X )

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

P( X )

(7.19)

g( X )

g( X )

(7.18)

n-k

Thus division of X U (X) by g (X) gives us the quotient polynomial Q (X) and the remainder
polynomial P (X). Therefore to obtain the cyclic codes in the systematic form, we determine the
n-k
remainder polynomial P (X) after dividing X
U (X) by g(X). This division process can be easily
n-k
achieved by noting that "multiplication by X
amounts to shifting the sequence by (n-k) bits".
Specifically in the circuit of Fig 7.5(a), if the input A(X) is applied to the Mod-2 adder

Dept. of ECE/SJBIT

Page 176

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

after the (n-k)

th

10EC55

n-k

shift register the result is the division of X

A (X) by B (X).

Accordingly, we have the following scheme to generate systematic cyclic codes. The generator
polynomial is written as:
2

n-k-1

g (X) = 1 +glX+g2X +g3X ++g n-k-1 X

+X

n-k

7.20)

n-k

The circuit of Fig 7.8 does the job of dividing X

encoding operation.

U (X) by g(X). The following steps describe the

1. The switch S is in position 1 to allow transmission of the message bits directly to an

out put shift register during the first k-shifts.
2. At the same time the 'GATE' is 'ON' to allow transmission of the message bits into the
(n-k) stage encoding shift register
th
3. After transmission of the k message bit the GATE is turned OFF and the switch S is
moved to position 2.
4. (n-k) zeroes introduced at "A" after step 3, clear the encoding register by moving the
parity bits to the output register
5. The total number of shifts is equal to n and the contents of the output register is the
n-k
code word polynomial V (X) =P (X) + X U (X).
6. After step-4, the encoder is ready to take up encoding of the next message input
Clearly, the encoder is very much simpler than the encoder of an (n, k) linear block code and the
memory requirements are reduced. The following example illustrates the procedure.
Example 7.4:
Let u = (1 0 1 1) and we want a (7, 4) cyclic code in the systematic form. The generator polynomial
chosen is g (X) = 1 + X + X3
2

For the given message, U (X) = 1 + X +X

X

n-k

U (X) = X U (X) = X + X + X
n-k

We perform direct division X U (X) by g (X) as shown below. From direct division observe that
p0=1, p1=p2=0. Hence the code word in systematic format is:
v = (p0, p1, p2; u0, u1, u2, u3) = (1, 0, 0, 1, 0, 1, 1)

Dept. of ECE/SJBIT

Page 177

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

The encoder circuit for the problem on hand is shown in Fig 7.9. The operational steps are as follows:
Shift Number
0
1
2
3
4

Input Queue Bit shifted IN

1011
101
1
10
1
1
0
1

Register contents
000
110
101
100
100

Output
1
1
0
1

After the Fourth shift GATE Turned OFF, switch S moved to position 2, and the parity bits
contained in the register are shifted to the output. The out put code vector is v = (100 1011) which
agrees with the direct hand calculation.
7.6

Syndrome Calculation - Error Detection and Error Correction:

Suppose the code vector v= (v0, v1, v2 v n-1) is transmitted over a noisy channel. Hence the
received vector may be a corrupted version of the transmitted code vector. Let the received code
k
vector be r = (r0, r1, r 2r n-1). The received vector may not be anyone of the 2 valid code vectors.
The function of the decoder is to determine the transmitted code vector based on the received vector.
The decoder, as in the case of linear block codes, first computes the syndrome to check whether or
not the received code vector is a valid code vector. In the case of cyclic codes, if the syndrome is
zero, then the received code word polynomial must be divisible by the generator polynomial. If the
syndrome is non-zero, the received word contains transmission errors and needs error correction. Let
the received code vector be represented by the polynomial

Dept. of ECE/SJBIT

Page 178

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

R(X) = r0+r1X+r2X ++r n-1X

n-1

Let A(X) be the quotient and S(X) be the remainder polynomials resulting from the division of
R(X) by g(X) i.e.
R( X )
S( X )
A( X )
g( X )
g( X )

..

(7.21)

The remainder S(X) is a polynomial of degree (n-k-1) or less. It is called the "Syndrome polynomial".
If E(X) is the polynomial representing the error pattern caused by the channel, then we have:
(7.22)
R(X) =V(X) + E(X)
..
And it follows as V(X) = U(X) g(X), that:
E(X) = [A(X) + U(X)] g(X) +S(X)

(7.23)
That is, the syndrome of R(X) is equal to the remainder resulting from dividing the error pattern by
the generator polynomial; and the syndrome contains information about the error pattern, which can
be used for error correction. Hence syndrome calculation can be accomplished using divider circuits
discussed in Sec 7.4, Fig7.5. A Syndrome calculator is shown in Fig 7.10.

The syndrome calculations are carried out as below:

1 The register is first initialized. With GATE 2 -ON and GATE1- OFF, the received vector is
entered into the register
2 After the entire received vector is shifted into the register, the contents of the register will be
the syndrome, which can be shifted out of the register by turning GATE-1 ON and GATE-2
OFF. The circuit is ready for processing next received vector.
Cyclic codes are extremely well suited for 'error detection' .They can be designed to detect
many combinations of likely errors and implementation of error-detecting and error correcting
circuits is practical and simple. Error detection can be achieved by employing (or adding) an
additional R-S flip-flop to the syndrome calculator. If the syndrome is nonzero, the flip-flop sets and
provides an indication of error. Because of the ease of implementation, virtually all error detecting
codes are invariably 'cyclic codes'. If we are interested in error correction, then the decoder must be
capable of determining the error pattern E(X) from the syndrome S(X) and add it to R(X) to determine
the transmitted V(X). The following scheme shown in Fig 7.11 may be employed for the purpose. The
error correction procedure consists of the following steps:
Step1. Received data is shifted into the buffer register and syndrome registers with switches

Dept. of ECE/SJBIT

Page 179

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

SIN closed and SOUT open and error correction is performed with SIN open and SOUT
closed.
Step2. After the syndrome for the received code word is calculated and placed in the syndrome
register, the contents are read into the error detector. The detector is a combinatorial circuit
designed to output a 1 if and only if the syndrome corresponds to a correctable error
n-l
pattern with an error at the highest order position X . That is, if the detector output is a '1'
then the received digit at the right most stage of the buffer register is assumed to be in
error and will be corrected. If the detector output is '0' then the received digit at the right
most stage of the buffer is assumed to be correct. Thus the detector output is the estimate
error value for the digit coming out of the buffer register.

Step3. In the third step, the first received digit in the syndrome register is shifted right once. If the
first received digit is in error, the detector output will be '1' which is used for error
correction. The output of the detector is also fed to the syndrome register to modify the
syndrome. This results in a new syndrome corresponding to the altered received code
word shifted to the right by one place.
Step4. The new syndrome is now used to check and correct the second received digit, which is
now at the right most position, is an erroneous digit. If so, it is corrected, a new syndrome
is calculated as in step-3 and the procedure is repeated.
Step5. The decoder operates on the received data digit by digit until the entire received code word
is shifted out of the buffer.
At the end of the decoding operation, that is, after the received code word is shifted out of the
buffer, all those errors corresponding to correctable error patterns will have been corrected, and the
syndrome register will contain all zeros. If the syndrome register does not contain all zeros, this
means that an un-correctable error pattern has been detected. The decoding schemes described in Fig
7.10 and Fig7.11 can be used for any cyclic code. However, the practicality depends on the
complexity of the combinational logic circuits of the error detector. In fact, there are special classes
of cyclic codes for which the decoder can be realized by simpler circuits. However, the price paid for

Dept. of ECE/SJBIT

Page 180

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

such simplicity is in the reduction of code efficiency for a given block size.
7.7

Bose- Chaudhury - Hocquenghem (BCH) Codes:

One of the major considerations in the design of optimum codes is to make the block size
n smallest for a given size k of the message block so as to obtain a desirable value of dmin. Or for
given code length n and efficiency k/n, one may wish to design codes with largest dmin. That means
we are on the look out for the codes that have 'best error correcting capabilities". The BCH codes, as
a class, are one of the most important and powerful error-correcting cyclic codes known. The most
common BCH codes are characterized as follows. Specifically, for any positive integer m 3, and t <
m
2 - 1) / 2, there exists a binary BCH code (called 'primitive' BCH code) with the following
parameters:
m

Block length : n = 2 -l
Number of message bits : k n - mt
Minimum distance : dmin 2t + 1
Clearly, BCH codes are "t - error correcting codes". They can detect and correct up to t
random errors per code word. The Hamming SEC codes can also be described as BCH codes. The
BCH codes are best known codes among those which have block lengths of a few hundred or less. The
major advantage of these codes lies in the flexibility in the choice of code parameters viz: block length
and code rate. The parameters of some useful BCH codes are given below. Also indicated in the table are
the generator polynomials for block lengths up to 31.
NOTE: Higher order co-efficients of the generator polynomial are at the left. For example, if we are
interested in constructing a (15, 7) BCH code from the table we have (111 010 001) for the co-efficients
of the generator polynomial. Hence
4

g(X) = 1 + X + X + X + X
n
7
15
15
15
31
31
31
31
31

k
4
11
7
5
26
21
16
11
6

t
1
1
2
3
1
2
3
5
7

Generator Polynomial
1 011
10 011
111 010 001
10 100 110 111
100 101
11 101 101 001
1 000 111 110 101 111
101 100 010 011 011 010 101
11 001 011 011 110 101 000 100 111

For further higher order codes, the reader can refer to Shu Lin and Costello Jr. The alphabet of
m
a BCH code for n = (2 -1) may be represented as the set of elements of an appropriate Galois field,
m
GF(2 ) whose primitive element is .The generator polynomial of the t-error correcting BCH code is the
least common multiple (LCM) of Ml(X), M2(X), M2t(X), where Mi(X) is the minimum polynomial of
i
, i = 1, 22t . For further details of the procedure and discussions the reader can refer to J.Das etal.
There are several iterative procedures available for decoding of BCH codes. Majority of them
can be programmed on a general purpose digital computer, which in many practical applications form

Dept. of ECE/SJBIT

Page 181

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

an integral part of data communication networks. Clearly, in such systems software implementation
of the algorithms has several advantages over hardware implementation
Review questions:
1. Write a standard array for Systematic Cyclic Codes code
2. Explain the properties of binary cyclic codes.
3. With a neat diagrams explain the binary cyclic encoding and decoding
4. Explain how meggit decoder can be used for decoding the cyclic codes.
5. Write short notes on the following
(iii)
BCH codes
6. Draw the general block diagram of encoding circuit using (n-k) bit shift register and explain
its operation.
7. Draw the general block diagram of syndrome calculation circuit for cyclic codes and explain
its operation.

Dept. of ECE/SJBIT

Page 182

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

UNIT 7
Syllabus:
RS codes, Golay codes, Shortened cyclic codes, Burst error correcting codes. Burst and
Random Error correcting codes. 7 Hours

Text Books:
Digital and analog communication systems, K. Sam Shanmugam, John Wiley,
1996. Digital communication, Simon Haykin, John Wiley, 2003.

Reference Books:
ITC and Cryptography, Ranjan Bose, TMH, II edition, 2007
Digital Communications - Glover and Grant; Pearson Ed. 2nd Ed 2008

Dept. of ECE/SJBIT

Page 183

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

UNIT 7
Cyclic Redundancy Check (CRC) codes:
Cyclic redundancy .check codes are extremely well suited for "error detection". The two important
reasons for this statement are, (1) they can be designed to detect many combinations of likely errors.
(2) The implementation of both encoding and error detecting circuits is practical. Accordingly, all
error detecting codes used in practice, virtually, are of the CRC -type. In an n-bit received word if a
contiguous sequence of b-bits in which the first and the last bits and any number of intermediate
bits are received in error, then we say a CRC "error burst' of length b has occurred. Such an error
burst may also include an end-shifted version of the contiguous sequence.
In any event, Binary (n, k) CRC codes are capable of detecting the following error patterns:
1. All CRC error bursts of length (n-k) or less.
2. A fraction of (1 - 2
3. A fraction (1-2

(n k - 1)

(n k)

) of CRC error bursts of length (n k + 1 ).

) of CRC error bursts of length greater than (n k + 1 ).

4. All combinations of (dmin 1 ) or fewer errors.

5. All error patterns with an odd number of errors if the generator polynomial
g (X) has an even number of non zero coefficients.
Generator polynomials of three CRC codes, internationally accepted as standards are listed below. All
three contain (1 +X) as a prime factor. The CRC-12 code is used when the character lengths is 6-bits.
The others are used for 8-bit characters.
2

* CRC-12 code: g (X) = 1 + X + X +X + X

2

15

*CRC-16 code: g (X) = 1 +X + X

+X

*CRC-CCITT code: g (X) = 1 + X + x

11

+X

12

16

12

+X

16

(Expansion of CCITT: "Commit Consultaitif International Tlphonique etTlgraphique " a

Geneva-based organization made up of telephone companies from all over the world)
Maximum Length codes:
For any integer m 3, the maximum length codes exist with parameters:
m

Block length : n = 2 - 1
Message bits : k = m
m-1
Minimum distance : dmin = 2

Dept. of ECE/SJBIT

Page 184

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Maximum length codes are generated by polynomials of the form g( X )

1Xn
p( X )

Maximum length codes are generated by polynomials of degree 'm'. Notice that any cyclic code
generated by a primitive polynomial is a Hamming code of dmin = 3. It follows then that the
maximum length codes are the 'duals' of Hamming codes. These codes are also referred to as 'pseudo
Noise (PN) codes' or "simplex codes".
Majority Logic Decodable Codes:
These codes form a smaller sub-class of cyclic codes than do the BCH codes. Their error
correcting capabilities, for most interesting values code length and efficiency, are much inferior to
BCH codes. The main advantage is that the decoding can be performed using simple circuits. The
concepts are illustrated here with two examples.
Consider a (7, 3) simplex code, which is dual to the (7, 4) Hamming code. Here dmin=4 and t = 1.
This code is generated by G and corresponding parity check matrix H given below:

1 0 1 1 1 0 0
1 1 1 0 0 1 0
0 1 1 1 0 0

1
H

0 0 0
1 0 0

1 1 0
0 1 1

0 1 1 0 1

1 1

The error vector e= (e0, e1, e2, e3, e4, e5, e6) is checked by forming the syndromes:
s0 = e0 + e4 + e5;

s2 = e2 + e4 + e5 + e6;

s1 = e1 + e5 + e6;

s3 = e3 + e4 + e6

Forming the parity check sum as:

A1 = s1 = e1 + e5 + e6
A2 = s3 = e3 + e4 + e6
A3 = s0 + s2 = e0 + e2 + e6
It is observed that all the check sums check for the error bit e6 and no other bit is checked by
more than one check sum. Then a majority decision can be taken that e6 = 1 if two or more Ai's are
non-zero. If e6 = 0 and any other bit is in error then only one of the Ai's will be non-zero. It is said that
the check sums Ai's are orthogonal on the error bit e6. A circulating SR memory circuit along with a
few logic circuits shown in Fig 7.16 forms the hardware of the decoder.
Initially, the received code vector R(X) is loaded into the SR's and check sums A1, A2 and A3
are formed in the circuit. If e6 is in error then the majority logic output is '1 'and is corrected as it is
shifted out of the buffer. If e6 is correct, then e5 is checked after one shift of the SR content.

Dept. of ECE/SJBIT

Page 185

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Thus all the bits are checked by successive shifts and the corrected V(X) is reloaded in the buffer. It is
possible to correct single errors only by using two of the check sums. However, by using three check
sums, the decoder also corrects some double error patterns. The decoder will correct all single errors
and detect all double error patterns if the decision is made on the basis of
(i). A1 = 1, A2 = 1, A3 = 1 for single errors (ii).
One or more checks fail for double errors.

We have devised the majority logic decoder assuming it is a Block code. However we should
not forget that it is also a cyclic code with a generator polynomial
2

g(X) = 1 + X + X + X .
Then one could generate the syndromes at the decoder by using a divider circuit as already
discussed. An alternative format for the decoder is shown in Fig 7.17. Successive bits are checked for
single error in the block. The feed back shown is optional - The feed back will be needed if it is
desired to correct some double error patterns.

Dept. of ECE/SJBIT

Page 186

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Let us consider another example This time the ( 7, 4) Hamming code generated by the
3
polynomial g(X) =1 + X + X .
1
0 0 1 0 1 1
Its parity check matrix is: H 0
1 0 1 1 1 0
0

0 1

0 1 1 1

The Syndromes are seen to be

s0 = e0 + e3 + (e5 + e6)
s1 = e1 + e3 + e4 + e5
s2 = e2 + e4 + (e5 + e6) = e2 + e5 + (e4 + e6)

Check sum A1 = (s0 + s1) = e0 + e1 + (e4 + e6)

It is seen that s0 and s2 are orthogonal on B1 = (e5 + e6), as both of them provide check for this
sum. Similarly, A1 and s2 are orthogonal on B2 = (e4 + e6). Further B1 and B2 are orthogonal on
e6. Therefore it is clear that a two-step majority vote will locate the error on e6.The corresponding
decoder is shown in Fig 7.18, where the second level majority logic circuit gives the correction signal
and the stored R(X) is corrected as the bits are read out from the buffer. Correct decoding is achieved
if t < d / 2 = 1 error (d = no. of steps of majority vote). The circuit provides majority vote '1' when
the syndrome state is {1 0 1}. The basic principles of both types of decoders, however, are the same.
Detailed discussions on the general principles of Majority logic decoding may be found in Shu-Lin
and Costello Jr., J.Das etal and other standard books on error control coding. The idea of this section
was "only to introduce the reader to the concept of majority logic decoding.
m

The Hamming codes (2 -1, 2 -m-1), m any integer, are majority logic decodable, (l5, 7)
BCH code with t 2 is 1-step majority logic decodable. Reed-Muller codes, maximum length
(simplex) codes Difference set codes and a sub-class of convolutional codes are examples majority
logic decodable codes.

Dept. of ECE/SJBIT

Page 187

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Shortened cyclic codes:

The generator polynomials for the cyclic codes, in general, are determined from among the
n
divisors of X + 1. Since for a given n and k, there are relatively few divisors, there are usually very
few cyclic codes of a given length. To overcome this difficulty and to increase the number of pairs ( n,
k) for which useful codes can be constructed, cyclic codes are often used in shortened form. In this
form the last j information digits are always taken to be zeros and these are not transmitted. The
decoder for the original cyclic code can decode the shortened cyclic codes simply by padding the
received (n-j) tuples with 'j zeros. Hence, we can always construct an ( n-j, k-j) shortened cyclic code
starting from a (n, k) cyclic code. Therefore the code thus devised is a sub-set of the cyclic code from
which it was derived - which means its minimum distance and error correction capability is at least as
great as that of the original code. The encoding operation, syndrome calculation and error correction
procedures for shortened codes are identical to those described for cyclic codes. This impliesshortened cyclic codes inherit nearly all of the implementation advantages and much of the
mathematical structure of cyclic codes.
Golay codes:
Golay code is a (23, 12) perfect binary code that is capable of correcting any combination of
three or fewer random errors in a block of 23 bits. It is a perfect code because it satisfies the
Hamming bound with the equality sign for t = 3 as:
23

23

23

23

12

23

0
1
2
The code has been used in many practical systems. The generator polynomial for the code is
23
obtained from the relation (X +1) = (X+ 1) g1(X) g2(X), where:

Dept. of ECE/SJBIT

Page 188

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

g1(X) = 1 + X + X + X + X + X

10

+X

11

11

and g2 (X) = 1 + X + X + X + X + X + X

The encoder can be implemented using shift registers using either g1 (X) or g2 (X) as the divider
polynomial. The code has a minimum distance, dmin =7. The extended Golay code, a (924, 12) code
has dmin =8. Besides the binary Golay code, there is also a perfect ternary (11, 6) Golay code with
dmin = 5.
Reed-Solomon Codes:
The Reed-Solomon (RS) codes are an important sub class of BCH codes, where the symbols
m
m
are from GF (q), q 2 in general, but usually taken as 2 . The encoder for an RS code differs from
a binary encoder in that it operates on multiple bits rather than individual bits. A t-error correcting
RS code has the following parameters.
Block length: n = (q - 1) symbols
Number of parity Check symbols: r = (n - k) = 2t
Minimum distance: dmin = (2t + 1)
The encoder for an RS (n, k) code on m-bit symbols groups the incoming binary data stream into
blocks, each km bits long. Each block is treated as k symbols, with each symbol having m-bits. The
encoding algorithm expands a block of k symbols by adding (n - k) redundant symbols. When m is an
integer power of 2, the m - bit symbols are called 'Bytes'. A popular value of m is 8 and 8-bit RS codes
are extremely powerful. Notice that no (n, k) linear block code can have dmin > (n - k + 1). For the RS
code the block length is one less than the size of a code symbol and minimum distance is one greater
than the number of parity symbols - "The dmin is always equal to the design distance of the code". An
(n, k) linear block code for which dmin = (n-k-l) is called 'Maximum - distance separable' code.
Accordingly, every RS code is maximum - distance separable' code-They make highly efficient use of
redundancy and can be adjusted to accommodate wide range of message sizes. They provide wide range of
code rates (k / n) that can be chosen to optimize performance. Further, efficient decoding techniques are
available for use with RS codes (usually similar to those of BCH codes).
Reed-Muller Codes (RM codes) are a class of binary group codes, which are majority logic
decodable and have a wide range of rates and minimum distances. They are generated from the
Hadamard matrices. (Refer J. Das etal).
CODING FOR BURST ERROR CORRECTION

The coding and decoding schemes discussed so far are designed to combat random or independent
errors. We have assumed, in other words, the channel to be Memory less. However, practical
channels have memory and hence exhibit mutually d ependent signal transmission impairments. In a
fading channel, such impairment is felt, particul arly when the fading varies slowly compared to one
symbol duration. The multi-path impairment involv es signal arrivals at the receiver over two or
more paths of different lengths with the effect that the signals arrive out of phase with each other
and the cumulative received signal are distorted. High-Frequency (HF) and troposphere propagation
in radio channels suffer from such a phenomenon. Further , some channels suffer from switching
noise and other burst noise (Example: Telephone channels or channels disturbed by pulse jamming
impulse noise in the communication channel causes transmission errors to cluster into bursts). All
of these time-correlated impairments results in statistical dependence among successive symbol
transmissions. The disturbances tend to cause errors that occur in bursts rather than isolated events.

Dept. of ECE/SJBIT

Page 189

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

Once the channel is assumed to have memory, the errors that occur can no longer be
characterized as single randomly distributed errors whose occurrence is independent from bit to bit.
Majority of the codes: Block, Cyclic or Convolutional codes are designed to combat such random or
independent errors. They are, in general, not efficient for correcting burst errors. The channel
memory causes degradation in the error performance.
Many coding schemes have been proposed for channels with memory. Greatest problem faced
is the difficulty in obtaining accurate models of the frequently time-varying statistics of such
channels. We shall briefly discuss some of the basic ideas regarding such codes. (A detailed
discussion of burst error correcting codes is beyond the scope of this book). We start with the
definition of burst length, b and requirements on a ( n, k) code to correct error burst. An error
burst of length b is defined as a sequence of error symbols confined to b consecutive bit positions
in which the first and the last bits are non-zero
For example, an error vector (00101011001100) is a burst of b = 10. The error vector
(001000110100) is a burst of b = 8. A code that is capable of correcting all burst errors of length b
or less is called a b-burst-error-correcting code. Or the code is said to have a burst error correcting
capability = b. Usually for proper decoding, the b-symbol bursts are separated by a guard space of g
symbols. Let us confine, for the present, ourselves for the construction of an (n, k) code for a given n and
b with as small a redundancy (n - k) as possible. Then one can make the following observations.
Start with a code vector V with an error burst of length 2b or less. This code vector then may
be expressed as a linear combination (vector sum) of the vectors V1 and V2 of length b or less.
Therefore in the standard array of the code both V1 and V2 must be in the same co-set. Further, if one
of these is assumed to be the co-set leader (i.e. a correctable error pattern), then the other vector
which is in the same co-set turns out to be an un-correctable error pattern. Hence, this code will not
be able to correct all error bursts of length b or less. Thus we have established the following assertion:
Assertion-1: A necessary condition for a (n ,k) linear code to b e able to correct all error bursts of
length b or less is that no error burst of length 2b or less be a code vector.
Next let us investigate the structure of code vectors whose non zero components are confined
b
to the first b bits. There are, clearly, 2 such code vectors. No two such vectors can be in the same
co-set of the standard array; otherwise their sum, which is again a burst of b or less, would be a code
b
b
vector. Therefore these 2 vectors must be in 2 distinct co-sets. For an (n, k) code we know that
(n-k)
there are 2
co-sets. This means that (n-k) must be at least equal to b. Thus we have established
another important assertion.
Assertion-2:The number of parity check bits of an (n, k) linear code that has no bursts of length
b or less as a code vector is at least b, i.e. (n -k) b
Combining the two assertions, now we can conclude that: The number of parity check
bits of a b-burst error correcting code must be at least 2b
i. e. (n- k) 2b

(9.1)

From Eq. (9.1) it follows that the burst-error-correcting capability of an (n, k) code is at most (n-k)//2.
That is, the upper bound on the burst-error-correcting capability of an (n, k) linear code is governed by:

Dept. of ECE/SJBIT

Page 190

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

b (n-k)/2

10EC55

(9.2)

This bound is known by the name Reiger Bound and it is used to define the burst correcting
efficiency, z, of an (n, k) codes as
z = 2b/ (n-k)

..

(9.3)

Whereas most useful random error correcting codes have been devised using analytical techniques,
for the reasons mentioned at the beginning of this section, the best burst- error correcting codes have
to be found through computer aided search procedures. A short list of high rate burst-error- correcting
cyclic codes found by computer search is listed in Table-9.1.
If the code is needed for detecting error bursts of length b, then the number of check bits must
satisfy:
(n - k) b

(9.4)
Some of the famous block/cyclic and convolution codes designed for correcting burst errors are
Burton, Fire, R-S, Berlekemp - Preparata-Massey , Iwadare and Adaptive Gallager codes. Of these
Fire codes have been extensively used in practice. A detailed discussion on these codes is available in
Shu-Lin et-all and J.Das et-all. (Refer Bibliogr aphy)
Burst and Random Error Correcting Codes:
In most practical systems, error occurs neither independently, at random, nor in well-defined
bursts. As a consequence codes designed for random error correction or single-burst-error correction
will become either inefficient or in adequate for tackling a mixture of random and burst errors. For
channels in which both types of error occur, it is better to design codes that can correct both types of
errors. One technique, which only requires knowledge of the duration or span of channel memory, not
its exact statistical characterization, is the use of Time diversity or Interleaving.
Interleaving the code vectors before transmission and de-interleaving after reception causes
the burst errors to be spread out in time and the decoder can handle them as if they were random
errors. Since, in all practical cases, the channel memory decreases with time separation, the idea
behind interleaving is only to separate the code word symbols in time. The interleaving times are
similarly filled by symbols of other code words. Separating the symbols in time effectively transforms
a channel with memory to a memory less channel , and there by enable the random error correcting
codes to be useful in a bursty-noise channel.
The function of an interleaver is to shuffle the code symbol over a span of several block
lengths (for block codes) or several constraint lengths (for convolutional codes). The span needed is
usually determined from the knowledge of the burst length. Further the details of the bitredistribution pattern must be known at the receiver end to facilitate de-interleaving and decoding.
Fig 9.1 illustrates the concept of interleaving.
The un-interleaved code words shown in Fig 9.1(a) are assumed to have a single error correcting
capability with in each six-symbol sequence. If the memory span of the channel is one code word in
duration, a six symbol time noise burst could destroy the information contained in one or two code
words. On the contrary, suppose encoded data were interleaved as shown in Fig 9.1(b), such that each
code symbol of each code word is separated from its pre-interleaved neighbors by a span of six
symbol times. The result of an error burst as marked in Fig.9.1 is to affect one code symbol from each
of the original six code words. Upon reception, the stream is de-interleaved and decoded as

Dept. of ECE/SJBIT

Page 191

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

though a single-random error has occurred in each code word. Clearly the burst noise has no degrading effect on the final sequence.

Block Interleaving:
Given an (n, k) cyclic code, it is possible to construct a (n, k) cyclic interlaced code by
simply arranging code vectors of the original code into rows of a rectangular array and then
transmitting them column by column. The parameter is called Degree of interlacing. By such
an arrangement, a burst of length or less will affect no more than one symbol in each row since
transmission is done on a column by column basis. If the original code (whose code words are the
rows of the (n k) matrix) can correct single errors, then the interlaced code can correct single
bursts of length or less. On the other hand if the original code has an error correcting capability of
t, t >1, then the interlaced code is capable of correcting any combination of t-error bursts of length
or less. The performance of the (n, k) interleaved cyclic code against purely random errors is
identical to that of the original (n, k) cyclic code from which it was generated.
The block interleaver accepts the coded symbols in blocks from the encoder, permutes the
symbols and then feeds the re-arranged symbols to the modulator. The usual permutation of block is
accomplished by filling the rows of a row by n- column array with the encoded sequence.
After the array is completely filled, the symbols are then fed to the modulator one column at a time
and transmitted over the channel. At the receiver the code words are re-assembled in a complimentary
manner. The de-inter leaver accepts the symbols from the de-modulator, de-interleaves them and
feeds them to the decoder - symbols are entered into the de interleaver array by columns and removed
by rows. The most important characteristics of block interleaver may be summarized as follows:
1)

Any burst of length less than contiguous channel symbol errors result in isolated errors at
the de-interleaver output that are separated from each other by at least n-symbols.

2) Any q. burst of errors, where q >1, results in output bursts from the de-interleaver of not more
than q symbol errors. Each output burst is separated fr om the other burst by not less than nq symbols. The notation q means the smallest integer not less than q and q means the largest
integer not greater than q.

Dept. of ECE/SJBIT

Page 192

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

3) A periodic sequence of single errors spaced - symbols apart results in a single burst of errors of
length n at the de-interleaver output.
4) The interleaver/ de-interleaver end to end delay is approximately 2n symbol time units to be
filled at the receiver before decoding begins. Therefore, the minimum end- to- end delay is
(2n-2n+2) symbol time units. This does not include any channel propagation delay.
5) The memory requirement, clearly, is n symbols for each location (interleaver and deinterleaver). However, since the n array needs to be (mostly) filled before it can be read out,
a memory of 2n symbols is generally implemented at each location to allow the emptying of
one n array while the other is being filled , and vice versa.
Finally a note about the simplest possible implementation aspect- If the original code is cyclic
then the interleaved code is also cyclic. If the original code has a generator polynomial g(X), the

interleaved code will have the generator polynomial g (X ). Hence encoding and decoding can be
done using shift registers as was done for cyclic codes. The modification at the decoder for the
interleaved codes is done by replacing each shift register stage of the original decoder by stages
without changing other connections. This modification now allows the decoder to look at successive
rows of the code array on successive decoder cycles. It then follows that if the decoder for the
original cyclic code is simple so will it be for the interleaved code. Interleaving technique is inde ed
an effective tool for deriving long powerful codes from short optimal codes.
Example 9.1:
Let us consider an interleaver with n = 4 and =6. The corresponding (64) array is shown in Fig.
9.2(a). The symbols are numbered indicating the sequence of transmission. In Fig 9.2(b) is shown an
error burst of five-symbol time units-The symbols shown encircled suffer transmission errors. After
de- interleaving at the receiver, the sequence is:

Observe that in the de-interleaved sequence, each code word does not have more than one error. The
smallest separation between symbols in error is n = 4.
Next, with q = 1.5, q= 9. Fig 9.2(c) illustrates an example of 9-symbol error burst. After deinterleaving at the receiver, the sequence is:

Dept. of ECE/SJBIT

Page 193

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

The encircled symbols are in error. It is seen that the bursts consists of no more than 1.5 =2
contiguous symbols per code word and they are separated by at least n - 1.5 = 4 1 = 3 symbols.
Fig.9.2 (d) illustrates a sequence of single errors spaced by = 6 symbols apart. After de-interleaving
at the receiver, the sequence is:

It is seen that de-interleaved sequence has a single error burst of length n = 4 symbols. The minimum
end to end delay due to interleaver and de-interl eaver is (2n 2n+2 ) = 42 symbol time units.
Storage of n = 24 symbols required at each end of the channel. As said earlier, storage for 2n = 48
symbols would generally be implemented.
Example 9.2: Interleaver for a BCH code.
2

Consider a (15, 7) BCH code generated by g(X) = 1+X+X +X +X . For this code dmin=5,

dmim 1

2 .With =5, we can construct a (75, 35) interleaved code with a burst error correcting
2
capability of b= t=10. The arrangement of code words, similar to Example 9.1, is shown in Fig 9.3.
A 35-bit message block is divided into five 7-bit message blocks and five code words of length 15 are
generated using g(X).These code words are arranged as 5-rows of a 515 matrix. The columns of the
matrix are transmitted in the sequence shown as a 75-bit long code vector.

t=

Each rows is 15 bit code word

1
2
3
Dept. of ECE/SJBIT

6
7
8

11 . 31
(36) ..
12 . (32) (37) ..
13 . (33) (38) ..

(66)
71
67
72
68
73
Page 194

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

4 (9) 14 . (34) 39
.. 69
74
5 10
15 . (35) 40

(70) 75
Fig 9.3 Block Interleaver for a (15, 7) BCH code.
To illustrate the burst and random error correcting capabilities of this code, we have put the bit
positions 9, 32 to 38, 66 and 70 in parenthesis, indicating errors occurred in these positions .The deinterleaver now feeds the rows of Fig 9.3 to the decoder. Clearly each row has a maximum of two
errors and since the (15, 7) BCH code, from which the rows were constructed, is capable of correcting
up to two errors per row. Hence the error pattern shown in parenthesis in the Figure can be corrected.
The isolated errors in bit positions 9, 66 and 70 may be thought of as random errors while the cluster
of errors in bit positions 32 to 38 as a burst error.
Convolutional Interleaving:
Convolution interleavers are somewhat simpler and more effective compared to block
interleavers. A (b n) periodic (convolutional) interleaver is shown in Fig 9.4.The code symbols are
shifted sequentially into the bank of n-shift registers. Each successive register introduces a delay b.
i.e., the successive symbols of a codeword are delayed by {0, b, 2b ( n-1) b} symbol units
respectively. Because of this, the symbols of one codeword are placed at distances of b-symbol units
in the channel stream and a burst of length b separated by a guard space of ( n-1)b symbol units
only affect one symbol per codeword. In the receiver, the code words are reassembled through
complementary delay units and decoded to correct single errors so generated. If the burst length l > b
but l 2b, then the (n, k) code should be capable of correcting two errors per code words. To
economize on the number of SRs as shown in Fig 9.4 and clock them at a period of nTo, where To =
symbol duration. This would then ensure the required delays of {b, 2b ( n-1) b} symbol units.

As illustrated in Example 9.2, if one uses a (15, 7) BCH code with t=2, then a burst of length 2b can
be corrected with a guard space of (n-1) b = 14b. This 14 to 2 guard space to burst length ratio is
too large, and hence codes with smaller values of n are preferable. Convolutional codes with
interleaver may also be used. The important advantage of convolutional interleaver over block
interleaver is that, with convolutional interleaving the end-to-end delay is (n-1) b symbol units and
the memory required at both ends of the channel is b (n-1)/2. This means, there is a reduction of one
half in delay and memory over the block interleaving requirements.

Dept. of ECE/SJBIT

Page 195

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

Review Questions:
1. What are RS codes? How are they formed?
2. Write down the parameters of RS codes and explain those parameters with an example.
3. List the applications of RS codes.
4. Explain why golay code is called as perfect code.
5. Explain the concept of shortened cyclic code.
6. What are burst error controlling codes?
7. Explain clearly the interlacing technique with a suitable example.
8. What are Cyclic Redundancy Check (CRC) codes

Dept. of ECE/SJBIT

Page 196

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

UNIT 8: CONVOLUTIONAL CODES

Syllabus:
Convolution Codes, Time domain approach. Transform domain approach.

7 Hours

Text Books:
Digital and analog communication systems, K. Sam Shanmugam, John Wiley,
1996. Digital communication, Simon Haykin, John Wiley, 2003.

Reference Books:
ITC and Cryptography, Ranjan Bose, TMH, II edition, 2007
Digital Communications - Glover and Grant; Pearson Ed. 2nd Ed 2008

Dept. of ECE/SJBIT

Page 197

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

Unit 8
CONVOLUTIONAL CODES
In block codes, a block of n-digits generated by the encoder depends only on the block of kdata digits in a particular time unit. These codes can be generated by combinatorial logic circuits. In a
convolutional code the block of n-digits generated by the encoder in a time unit depends on not only
on the block of k-data digits with in that time unit, but also on the preceding m input blocks. An ( n,
k, m) convolutional code can be implemented with k-input, n-output sequential circuit with input memory
m. Generally, k and n are small integers with k < n but the memory order m must be made large to
achieve low error probabilities. In the important special case when k = 1, the information sequence is not
divided into blocks but can be processed continuously.
Similar to block codes, convolutional codes can be designed to either detect or correct errors.
However, since the data are usually re-transmitted in blocks, block codes are better suited for error
detection and convolutional codes are mainly used for error correction.
Convolutional codes were first introduced by Elias in 1955 as an alternative to block codes.
This was followed later by Wozen Craft, Massey, Fano, Viterbi, Omura and others. A detailed
discussion and survey of the application of convolutional codes to practical communication channels
can be found in Shu-Lin & Costello Jr., J. Das etal and other standard books on error control coding.
To facilitate easy understanding we follow the popular methods of representing convolutional
encoders starting with a connection pictorial - needed for all descriptions followed by connection
vectors.
8.1 Connection Pictorial Representation:
The encoder for a (rate 1/2, K = 3) or (2, 1, 2) convolutional code is shown in Fig.8.1. Both
sketches shown are one and the same. While in Fig.8.1 (a) we have shown a 3-bit register, by noting
that the content of the third stage is simply the output of the second stage, the circuit is modified
using only two shift register stages. This modification, then, clearly tells us that" the memory
requirement m = 2. For every bit inputted the encoder produces two bits at its output. Thus the
encoder is labeled (n, k, m) (2, 1, 2) encoder.

At each input bit time one bit is shifted into the left most stage and the bits that were present in the

Dept. of ECE/SJBIT

Page 198

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

registers shifted to the right by one position. Output switch (commutator /MUX) samples the output
of each X-OR gate and forms the code symbol pairs for the bits introduced. The final code is obtained
after flushing the encoder with "m" zero's where 'm'- is the memory order (In Fig.8.1, m = 2). The
sequence of operations performed by the encoder of Fig.8.1 for an input sequence u = (101) are
illustrated diagrammatically in Fig 8.2.

From Fig 8.2, the encoding procedure can be understood clearly. Initially the registers are in
Re-set mode i.e. (0, 0). At the first time unit the input bit is 1. This bit enters the first register and
pushes out its previous content namely 0 as shown, which will now enter the second register and
pushes out its previous content. All these bits as indicated are passed on to the X-OR gates and the
output pair (1, 1) is obtained. The same steps are repeated until time unit 4, where zeros are
introduced to clear the register contents producing two more output pairs. At time unit 6, if an
additional 0 is introduced the encoder is re-set and the output pair (0, 0) obtained. However, this
step is not absolutely necessary as the next bit, whatever it is, will flush out the content of the second
register. The 0 and the 1 indicated at the output of second register at time unit 5 now vanishes.
Hence after (L+m) = 3 + 2 = 5 time units, the output sequence will read v = (11, 10, 00, 10, 11). (Note:
L = length of the input sequence). This then is the code word produced by the encoder. It is very
important to remember that Left most symbols represent earliest transmission.
As already mentioned the convolutional codes are intended for the purpose of error correction.
However, it suffers from the problem of choosing connections to yield good distance properties. The
selection of connections indeed is very complicated and has not been solved yet. Still, good codes
have been developed by computer search techniques for all constraint lengths less than 20. Another
point to be noted is that the convolutional codes do not have any particular block size. They can be
periodically truncated. Only thing is that they require m-zeros to be appended to the end of the input
sequence for the purpose of clearing or flushing or re-setting of the encoding

Dept. of ECE/SJBIT

Page 199

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

shift registers off the data bits. These added zeros carry no information but have the effect of reducing
the code rate below (k/n). To keep the code rate close to (k/n), the truncation period is generally made
as long as practical.
The encoding procedure as depicted pictorially in Fig 8.2 is rather tedious. We can approach
the encoder in terms of Impulse response or gene rator sequence which merely represents the
response of the encoder to a single 1 bit that moves through it.
8.2 Convolutional Encoding Time domain approach:
The encoder for a (2, 1, 3) code is shown in Fig. 8.3. Here the encoder consists of m=3 stage
shift register, n=2 modulo-2 adders (X-OR gates) and a multiplexer for serializing the encoder
outputs. Notice that module-2 addition is a linear operation and it follows that all convolution
encoders can be implemented using a linear feed forward shift register circuit.
The information sequence u = (u1, u2, u3 .) enters the encoder one bit at a time starting from
u1. As the name implies, a convolutional encoder operates by performing convolutions on the
(1)
(1)
(1)
(1)
information sequence. Specifically, the encoder output sequences, in this case v = {v1 , v2 , v3
}and v

(2)

= {v1

(2)

,v2

(2)

(2)

,v3

} are obtained by the discrete convolution of the information sequence

with the encoder "impulse responses'. The impulse responses are obtained by determining the output
sequences of the encoder produced by the input sequence u = (1, 0, 0, 0) .The impulse responses so
defined are called 'generator sequences' of the code. Since the encoder has a m-time unit memory the
impulse responses can last at most (m+ 1) time units (That is a total of (m+ 1) shifts are necessary for
a message bit to enter the shift register and finally come out) and are written as:

(i)

(i)

(i)

(i)

= {g1 , g2 ,g3

(i)

g m+1 }.

For the encoder of Fig.8.3, we require the two impulse responses,

g
g

(1)
(2)

= {g1
= {g1

(1)
(2)

(1)

(1)

(2)

(2)

, g2 , g3
, g2 , g3

, g4
, g4

(1)
(2)

} and
}

By inspection, these can be written as: g

(1)

= {1, 0, 1, 1} and g

(2)

= {1, 1, 1, 1}

Observe that the generator sequences represented here is simply the 'connection vectors' of the
encoder. In the sequences a '1' indicates a connection and a '0' indicates no connection to the
corresponding X - OR gate. If we group the elements of the generator sequences so found in to pairs,

Dept. of ECE/SJBIT

Page 200

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

we get the overall impulse response of the encoder, Thus for the encoder of Fig 8.3, the over-all
impulse response will be:
v = (11, 01, 11, 11)
The encoder outputs are defined by the convolution sums:

v (1) = u * g (1)

(8.1 a)

v (2) = u * g (2)

(8.1 b)

Where * denotes the discrete convolution, which i mplies:

vl

(j)

ul i .gi 1 ( j )
i 0

= ul g1

(j)

(j)

+ ul 1 g 2

+ ul 2 g 3

(j)

+ .. +u l m gm+1 (j)

(8.2)

for j = 1, 2 and where ul-i = 0 for all l < i and all operations are modulo - 2. Hence for the encoder of
Fig 8.3, we have:

vl
vl

(1)
(2)

= ul + ul 2 + ul - 3
= ul + ul 1 + ul 2 + ul - 3

This can be easily verified by direct inspection of the encoding circuit. After encoding, the
two output sequences are multiplexed into a single sequence, called the "code word" for transmission
over the channel. The code word is given by:

v = {v1

(1)

v1

(2)

, v2

(1)

v2

(2)

, v3

(1)

v3

(2)

Example 8.1:
Suppose the information sequence be u = (10111). Then the output sequences are:
v
v

(1)

(2)

= (1 0 1 1 1) * (10 1 1)
= (1 0 0 0 0 0 0 1),
= (1 0 1 1 1) * (1 1 1)
= (1 1 0 1 1 1 0 1),

and the code word is

v = (11, 01, 00, 01, 01, 01, 00, 11)
The discrete convolution operation described in Eq (8.2) is merely the addition of shifted
impulses. Thus to obtain the encoder output we need only to shift the overall impulse response by
'one branch word', multiply by the corresponding input sequence and then add them. This is illustrated in
the table below:

Dept. of ECE/SJBIT

Page 201

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

INPUT

OUT PUT

1
0
1
1
1

11 01 11 11
0 0 0 0 0 0 0 0 -----one branch word shifted sequence
11 01 1111
---Two branch word shifted
11 01 11 11
11 01 11 11
11 01 00 01 01 01 00 11

Modulo -2 sum

The Modulo-2 sum represents the same sequence as obtained before. There is no confusion at all
with respect to indices and suffices! Very easy approach - super position or linear addition of shifted
impulse response - demonstrates that the convolutional codes are linear codes just as the block codes
and cyclic codes. This approach then permits us to define a 'Generator matrix' for the convolutional
encoder. Remember, that interlacing of the generator sequences gives the overall impulse response
and hence they are used as the rows of the matrix. The number of rows equals the number of
information digits. Therefore the matrix that results would be Semi-Infinite. The second and
subsequent rows of the matrix are merely the shifted versions of the first row -They are each shifted
with respect to each other by "One branch word". If the information sequence u has a finite length,
say L, then G has L rows and n (m +L) columns (or (m +L) branch word columns) and v has a length
of n (m +L) or a length of (m +L) branch words. Each branch word is of length 'n'. Thus the
Generator matrix G, for the encoders of type shown in Fig 8.3 is written as:
g1

(1)

g1

(2)

g2

(1)

g1

(1)

g2

(2)

g1

(2)

g3
g2

(1)
(1)
(1)

g3
g2

(2)
(2)
(2)

g4

(1)

g3

(1)

g4

g3

(1)

g1
g1
g2
(Blank places are zeros.)

(2)
(2)
(2)

g2

L
g 4(1)g
g (1 )g
3

(2)
(2)

L
L
(1)
g
g
4

L
(2)

.. (8.3)

The encoding equations in Matrix form is:

v = u .G

(8.4)

Example 8.2:
For the information sequence of Example 8.1, the G matrix has 5 rows and 2(3 +5) =16 columns and
we have
1
1 0 1 1 1 1 1 0 0 0 0 0 0 0 0
0 1 1 0 1 1 1 1 1 0 0 0 0 0 0
0
G

0 0
0 0

0 0

0 0

1 1 0 1 1 1 1 1 0 0 0 0
0 0 1 1 0 1 1 1 1 1 0 0
0 0 0 0

1 1 0 1 1

1 1 1

Performing multiplication, v = u G as per Eq (8.4), we get: v = (11, 01, 00, 01, 01, 00, 11) same as
before.
As a second example of a convolutional encoder, consider the (3, 2, 1) encoder shown in

Dept. of ECE/SJBIT

Page 202

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Fig.8.4. Here, as k =2, the encoder consists of two m = 1 stage shift registers together with n = 3
modulo -2 adders and two multiplexers. The information sequence enters the encoder k = 2 bits at a
(2)
(2)
time and can be written as u = {u1 (1) u1
, u2 (1) u2
, u3 (1) u3 (2) } or as two separate input
sequences:
(1)

(1)

= {u1

, u2

(1)

, u3 (1) } and u

(2)

= {u1

(2)

, u2

(2)

, u3 (2)

}.

represent the generator sequence corresponding to

There are three generator sequences corresponding to each input sequence. Letting
( j)
( j)
( j)
( j)
( j)
= {gi,1 , gi,2 , gi,3
g i,m+1 }
gi

input i and output j. The generator sequences for the encoder are:
(1)

g1

= (1, 1), g1

(1)

g2

= (0, 1), g2

(2)
(2)

= (1, 0), g1
= (1, 1), g2

(3)
(3)

= (1, 0)
= (0, 0)

The encoding equations can be written as:

(1) * g1 (1) + u
v (1) = u(1)

v (2) = u

* g1

(2) +

(2)

u (2)

* g2 (1)
* g2 (2)

v (3) = u (1) * g1 (3) + u (2) * g2 (3)

(8

.5 a)
(8.5 b)
(8.5 c)

The convolution operation implies that:

vl

(1)

=ul

vl
(2)

(2)

(1)

+ u l-1

=ul

vl

(3)

(1)

=ul

(1)

+ul

+ u l-1

(2)

(2)

+ u l-1

(1)

as can be seen from the encoding circuit.

After multiplexing, the code word is given by:

v ={v1

( 1)

v1

( 2)

v1

( 3)

,v2

( 1)

v2

( 2)

v2

( 3)

,v3

( 1)

v3

( 2)

v3

( 3)

Example 8.3:

Dept. of ECE/SJBIT

Page 203

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Suppose u = (1 1 0 1 1 0). Hence u

(1)

= (1 0 1) and u

(2)

= (1 1 0). Then

(1)

= (1 0 1) * (1,1) + (1 1 0) *(0,1) = (1 0 0 1)

(2)

= (1 0 1) * (1,0) + (1 1 0) *(1,1) = (0 0 0 0)

(3)

= (1 0 1) * (1,0) + (1 1 0) *(0,0) = (1 0 1 0)

v = (1 0 1, 0 0 0, 0 0 1, 1 0 0).
The generator matrix for a (3, 2, m) code can be written as:
(1)
(2)
(3)
(1)
g
g
g
g
11

(1)

21

11

(2)

21

11

(3)

21

12

(1)

22

O
O

(2)
(3)
g
g
12

13

(2)

22

(1)
11

(2)
11

( 1 )g ( 2 )g ( 3 )

21

21

(3)

22

(3)

(1)

11

21

12

( 1 )g

22

L
L
g

(2)

12
( 2 )g

22

(1)
1,m1

(1)

2,m1

(3)

(2)
1,m1

2,m1

(2)

g
g

(3)
1,m1

(3)

2,m1

L
L

12

(3)
22

(8.6)

The encoding equations in matrix form are again given by v = u G. observe that each set of k = 2
rows of G is identical to the preceding set of rows but shifted by n = 3 places or one branch word to
the right.
Example 8.4:
For the Example 8.3, we have

u = {u1 (1) u1

(2)

(2)

, u2 (1) u2 , u3 (1) u3

(2)

} = (1 1, 0 1, 1 0)

The generator matrix is:

1
0

1 1, 1
1 0, 1

0 0
1 0

1 1,

1 0

1 0,

1 1 0
1 1 1, 1 0 0
0 1 0, 1 1 0

*Remember that the blank places in the matrix are all zeros.
Performing the matrix multiplication, v = u G, we get: v = (101,000,001,100), again agreeing with
our previous computation using discrete convolution.
This second example clearly demonstrates the complexities involved, when the number of
input sequences are increased beyond k > 1, in describing the code. In this case, although the encoder
contains k shift registers all of them need not have the same length. If ki is the length of the i-th shift
register, then we define the encoder "memory order, m" by

Dept. of ECE/SJBIT

Page 204

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

m Max ki

(8.7)

1i k

(i.e. the maximum length of all k-shift registers)

An example of a (4, 3, 2) convolutional encoder in which the shift register lengths are 0, 1 and 2 is
shown in Fig 8.5.

Since each information bit remains in the encoder up to (m + 1) time units and during each
time unit it can affect any of the n-encoder outputs (which depends on the shift register connections)
it follows that "the maximum number of encoder outputs that can be affected by a single information
bit" is

nA n(m 1)

8.8)

nA is called the 'constraint length" of the code. For example, the constraint lengths of the encoders
of Figures 8.3, 8.4 and 8.5 are 8, 6 and 12 respectively. Some authors have defined the constraint
length (For example: Simon Haykin) as the number of shifts over which a single message bit can
influence the encoder output. In an encoder with an m-stage shift register, the memory of the
encoder equals m-message bits, and the constraint length nA = (m + 1). However, we shall adopt the
definition given in Eq (8.8).
The number of shifts over which a single message bit can influence the encoder output is
usually denoted as K. For the encoders of Fig 8.3, 8.4 and 8.5 have values of K = 4, 2 and 3
respectively. The encoder in Fig 8.3 will be accordingly labeled as a rate 1/2, K = 4 convolutional
encoder. The term K also signifies the number of branch words in the encoders impulse response.
Turning back, in the general case of an (n, k, m) code, the generator matrix can be put in the
form:
G1 G2 G3
G

G1 G 2
G1

L Gm

LG
LG

m 1

m 2

m 1

Gm

m 1

m 1

Gm

(8.9)

m 1

Where each Gi is a (k n) sub matrix with entries as below:

Dept. of ECE/SJBIT

Page 205

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Notice that each set of k-rows of G are identical to the previous set of rows but shifted n-places to
(1)
(2)
(k)
the right. For an information sequence u = (u1, u2) where ui = {ui
, ui
u i }, the code word is
(1) (2)
(n)
v = (v1, v2) where vj = (vj , vj .v j
) and v = u G. Since the code word is a linear combination
of rows of the G matrix it follows that an (n, k, m) convolutional code is a linear code.
Since the convolutional encoder generates n-encoded bits for each k-message bits, we define R =
k/n as the "code rate". However, an information sequence of finite length L is encoded into a code word
of length n (L +m), where the final nm outputs are generated after the last non zero information block
has entered the encoder. That is, an information sequence is terminated with all zero blocks in order to
(j)
clear the encoder memory. (To appreciate this fact, examine 'the calculations of vl
for the Example 8.l
and 8.3). The terminating sequence of m-zeros is called the "Tail of the message". Viewing the
convolutional-code as a linear block code, with generator matrix G, then the block code rate is given by
kL/n(L +m) - the ratio of the number of message bits to the length of the code word. If L >> m, then, L/
(L +m) 1 and the block code rate of a convolutional code and its rate when viewed as a block code
would appear to be same. Infact, this is the normal mode of operation for convolutional codes and
accordingly we shall not distinguish between the rate of a convolutional code and its rate when viewed as
a block code. On the contrary, if L were small, the effective rate of transmission indeed is kL/n (L + m)
and will be below the block code rate by a fractional amount:
k / n kL / n( L m ) m
(8.11
)
..
Lm
k/n
and is called "fractional rate loss". Therefore, in order to keep the fractional rate loss at a minimum
(near zero), L is always assumed to be much larger than m. For the information 'sequence of
Example 8.1, we have L = 5, m =3 and fractional rate loss = 3/8 = 37.5%. If L is made 1000, the
fractional rate loss is only 3/1003 0.3%.
8.3 Encoding of Convolutional Codes; Transform Domain Approach:
In any linear system, we know that the time domain operation involving the convolution
integral can be replaced by the more convenient transform domain operation, involving polynomial
multiplication. Since a convolutional encoder can be viewed as a 'linear time invariant finite state
machine, we may simplify computation of the adder outputs by applying appropriate transformation.
As is done in cyclic codes, each 'sequence in the encoding equations can' be replaced by a
corresponding polynomial and the convolution operation replaced by polynomial multiplication. For
example, for a (2, 1, m) code, the encoding equations become:
(1)

v
v

(2)

(1)

(X) = u(X) g (X)

(2)

(X) = u(X) g (X)

2

Where u(X) = u1 + u2X + u3X +

(1)

v (X) = v1

(1)

Dept. of ECE/SJBIT

+ v2

(1)

X + v3

(1)

..
.....

(8.12a)
(8.12

is the information polynomial,

2

X +....., and

Page 206

www.rejinpaul.com

b)

www.rejinpaul.com

Information Theory and Coding

(2)

(2)

(X) = v1

10EC55

(2)

+ v2

X + v3

(2)

X +.....

are the encoded polynomials.

(1)

(1)

(2)

(2)

(X) = g1

(X) = g1

(1)

+ g2

(2)

+ g2

(1)

X + g3

(2)

X + g3

X + ....., and
2

X + .....

are the generator polynomials of' the code; and all operations are modulo-2. After multiplexing, the
code word becomes:
(1)

v(X) = v

(2)

(X ) + X v

(X )

(8.13)

The indeterminate 'X' can be regarded as a unit-delay operator, the power of X defining the
number of time units by which the associated bit is delayed with respect to the initial bit in the
sequence.
Example 8.5:
(2)

(1)

= (1,0, 1, 1), and g

For the (2, 1, 3) encoder of Fig 8.3, the impulse responses were: g
(l)

(2)

The generator polynomials are: g (X) = 1 + X + X , and g

(X) = 1 + X + X + X

= (1,1, 1, 1)

3
2

For the information sequence u = (1, 0, 1, 1, 1); the information polynomial is: u(X) = 1+X +X +X
The two code polynomials are then:
(1)

(l)

(2)

(2)

v
v

(X) = u(X) g (X) = (1 + X + X + X ) (1 + X + X ) = 1 + X

(X) = u(X) g

(X) = (1 + X + X + X ) (1 + X + X + X ) = 1 + X + X + X + X + X

From the polynomials so obtained we can immediately write:

(1)

(2)

v = ( 1 0 0 0 0 0 0 1), and v = (1 1 0 1 1 1 0 1)
Pairing the components we then get the code word v = (11, 01, 00, 01, 01, 01, 00, 11).
We may use the multiplexing technique of Eq (8.13) and write:
v

(1)

14

(X ) = 1 + X

and v

(2)

10

14

(X ) = 1+X +X +X +X +X ; Xv

and the code polynomial is: v(X) = v

(1)

(X ) + X v

(2)

(2)

11

11

(X ) = X + X + X + X + X
3

(X ) = 1 + X + X + X + X + X

14

+X

15

+X ;
15

+X

Hence the code word is: v = (1 1, 0 1, 0 0, 0 1, 0 1, 0 1, 0 0, 1 1); this is exactly the same as obtained
earlier.
The generator polynomials of an encoder can be determined directly from its circuit diagram.
l
Specifically, the co-efficient of X is a '1' if there is a "connection" from the l-th shift register stage to
the input of the adder of interest and a '0' otherwise. Since the last stage of the shift register in an

Dept. of ECE/SJBIT

Page 207

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

(n, l) code must be connected to at least one output, it follows that at least one generator polynomial
should have a degree equal to the shift register length 'm', i.e.
m Max
1 j n

deg g( j ) ( X )

(8.14)

In an (n, k) code, where k > 1, there are n-generator polynomials for each of the k-inputs, each
set representing the connections from one of the shift registers to the n-outputs. Hence, the length Kl
of the l-th shift register is given by:
Max
Kl 1jn

deg gl ( j ) ( X ), 1 l k

(8.15)

(j)

Where gl (X) is the generator polynomial relating the l-th input to the j-th output and the encoder
memory order m is:
Max
Max
Kl
(j)
1 j deg gl
(X)

(8.16)
m 1 l k
1lk
(l)
(j)
Since the encoder is a linear system and u (X) represents the l-th input sequence and v (X)
(j)
represents the j-th output sequence the generator polynomial gl (X) can be regarded as the 'encoder
transfer function' relating the input - l to the output j. For the k-input, n- output linear system there
are a total of kn transfer functions which can be represented as a (k n) "transfer function matrix".

(1)

G( X )

(n)

(2)

g1 ( X ), g1
( X ), L g1
(2)
( X ), L g
g ( 1 ) ( X ), g 2
2

2(n)

M
M
M
(2)
( X ), L g
g ( 1 ) ( X ), g
k

(X)
(X)

(n)

(8.17)

(X)

Using the transfer function matrix, the encoding equations for an (n, k, m) code can be expressed as
V(X) = U(X) G(X)
(1)

(2)

(8.18)

(k)

U(X) = [u
(X), u
(X)...u
(X)] is the k-vector, representing the information polynomials, and.
(1)
(2)
(n)
V(X) = [v
(X), v
(X) v
(X)] is the n-vector representing the encoded sequences. After
multiplexing, the code word becomes:
(1)

v(X) = v

(2)

(X ) + X v

2 (3)

(X ) +X v

(X )++ X

n-l (n)

(X )

(8.19)

Example 8.6:
For the encoder of Fig 8.4, we have:
g1

(1)

(X) = 1 + X, g 2

(1)

(X) = X

g 1 (2) (X) = 1,

g 2 (2) (X) =1+ X

g 1 (3) (X) = 1 ,

g 2 (3) (X) = 0

Dept. of ECE/SJBIT

Page 208

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

1X
G( X )
X

1X
0
(1)
(2)
For the information sequence u
= (1 0 1), u
= (1 1 0), the information polynomials
are: u

(1)

(2)

(X) = 1 + X , u

Then V(X) = [v

(1)

(X), v

(2)

(X) = 1 + X

(X), v

= [1 + X , 1 + X]

(3)

(X)]

1X
X

(1)

(2)

(X ) + Xv

1X 0
2 (3)

(X ) + X v

= [1 +X , 0, 1+X ]

Hence the code word is:

v(X) = v

(X )

= (1 + X ) + X (0) + X (1 + X )
2

=1+X +X +X

v = (1 0 1, 0 0 0, 0 0 1, 1 0 0).
This is exactly the same as that obtained in Example 8.3.From Eq (8.17) and (8.18) it follows that:
8.5.1 State Diagrams:
The state of an encoder is defined as its shift register contents. For an (n, k, m) code with k > 1,
k

i-th shift register contains Ki previous information bits. Defining K Ki as the total encoder i 1

memory (m - represents the memory order which we have defined as the maximum length of any
(1)
(2)
(k)
shift register), the encoder state at time unit T', when the encoder inputs are, {u l , u l u l },
are the binary k-tuple of inputs:
{u l-1

(1)

u l-2

(1)

, u l-3

(1)

u l-k

(1)

; u l-1

(2)

(2

, u l-2 , u l-3

(2)

u l-k

(2)

; ; u l-1

(k)

u l-2

(k)

, u l-3

(k)

u l-k

(k)

},

and there are a total of 2 different possible states. For a (n, 1, m) code, K = K1 = m and the encoder
state at time unit l is simply {ul-1, u l -2 u l-m}.
k

Each new block of k-inputs causes a transition to a new state. Hence there are 2 branches
leaving each state, one each corresponding to the input block. For an (n, 1, m) code there are only two
branches leaving each state. On the state diagram, each branch is labeled with the k-inputs causing
the transition and the n-corresponding outputs. The state diagram for the convolutional encoder of Fig
8.3 is shown in Fig 8.10. A state table would be, often, more helpful while drawing the state diagram
and is as shown.
State table for the (2, 1, 3) encoder of Fig 8.3
State
Binary
Description

Dept. of ECE/SJBIT

S0

S1

S2

S3

S4

S5

S6

S7

000

100

010

110

001

101

011

111

Page 209

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

Recall (or observe from Fig 8.3) that the two out sequences are:
(1)

v = ul + ul 2 + ul 3 and
(2)
v = ul + ul 1 + ul 2 + ul 3
Till the reader, gains some experience, it is advisable to first prepare a transition table using the
output equations and then translate the data on to the state diagram. Such a table is as shown below:

State transition table for the encoder of Fig 8.3

ul 1 u l 2 u l - 3

Binary
Description

Input

Next
State

S0

0 0 0

S0

0 0

S1

1 0

0 1

S2
S3

1 1

110

S4

0 0

S5

1 0

101

S6

0 1

S7

1 1

111

S0

0 0

S1

1 0

100

S2

0 1

S3

1 1

110

S4

0 0

S5

1 0

101

S6

0 1

S7

1 1

111

S1
S2
S3
S4
S5
S6
S7

1 0 0
0 1 0
1 1 0
0 0 1
1 0 1
0 1 1
1 1 1

Binary
Description

ul

Previous
State

0
1
0
1
0
1

1
1
0
0
1
1

Output

For example, if the shift registers were in state S5, whose binary description is 101, an input 1
causes this state to change over to the new state S3 whose binary description is 110 while producing

Dept. of ECE/SJBIT

Page 210

www.rejinpaul.com

www.rejinpaul.com

Information Theory and Coding

10EC55

an output (0 1). Observe that the inputs causing the transition are shown first, followed by the
corresponding output sequences shown with in parenthesis.
Assuming that the shift registers are initially in the state S0 (the all zero state) the code word
corresponding to any information sequence can be obtained by following the path through the state
diagram determined by the information sequence and noting the corresponding outputs on the branch
labels. Following the last nonzero block, the encoder is returned to state S0 by a sequence of m-allzero block appended to the information sequence. For example, in Fig 8.10, if u = (11101), the code
word is v = (11, 10, 01, 01, 11, 10, 11, 10) the path followed is shown in thin gray lines with arrows
and the input bit written along in thin gray. The m = 3 zeros appended are indicated in gray which is
much lighter compared to the information bits.
Example 8.12: A (2, 1, 2) Convolutional Encoder:
Consider the encoder shown in Fig 8.15. We shall use this example for discussing further graphical
representations viz. Trees, and Trellis.

(1)

For this encoder we have: v l = u l + u l 1 + u l 2 and v l

The state transition table is as follows.

(2)

=ul+ul2

State transition table for the (2, 1, 2) convolutional encoder of Example 8.12
Previous
Binary
Input Next
Binary
u l u l 1 u l - 2 Output
state
description
State description
S0
0 0
0
S0
0 0
0
0
0
0 0
1
S1
10
1
0
0
1 1
S1
1 0
0
S2
0 1
0
1 0
1 0
1
S3
11
1
1
0
0 1
S2
0 1
0
S0
0 0
0
0
1
1 1
1
S1
10
1
0
1
0 0
S3
1 1
0
S2
0 1
0
1 1
0 1
1
S3
11
1
1
1
1 0
The state diagram and the augmented state diagram for computing the complete path
enumerator function for the encoder are shown in Fig 8.16.

Dept. of ECE/SJBIT

Page 211

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

8.5.3 Catastrophic Code:

We shall re-consider the catastrophic codes considered earlier. The state diagram of a (2, 1, 2)
code is shown in Fig 8.17. Notice that for a catastrophic code "the state diagram consists of a loop of
zero weight other than the self loop around the state So".
It is the characteristic of a catastrophic
8.5.4TreeandTrellisDiagrams:
code that an information sequence of infinite
weight produces a finite weight code word.
In a non-catastrophic code which
contains no zero weight loops other than the
self loop around the all zero state So, all
infinite weight information sequences must
generate infinite weight code words, and
minimum weight code word always has a
finite length.
The best achievable dfree for a convolutional
code with a given rate and constraint length
has not been determined exactly. However,
results are available with respect to the lower
and upper bounds on dfree for the best code,
obtained using random coding approach. It is
observed that more free distance is available
with non-systematic codes of a given rate and
constraint length compared to the systematic
codes and thus has important consequences
when a code with large dfree must be selected
for use with either the Viterbi or Sequential
decoding.

Dept. of ECE/SJBIT

Page 212

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

Let us now consider other graphical means of portraying convolutional codes. The state
diagram can be re-drawn as a 'Tree graph'. The convention followed is: If the input is a '0', then the
upper path is followed and if the input is a '1', then the lower path is followed. A vertical line is called
a 'Node' and a horizontal line is called 'Branch'. The output code words for each input bit are shown
on the branches. The encoder output for any information sequence can be traced through the tree
paths. The tree graph for the (2, 1, 2) encoder of Fig 8.15 is shown in Fig 8.18. The state transition
table can be conveniently used in constructing the tree graph.

Following the procedure just described we find that the encoded sequence for an information
sequence (10011) is (11, 10, 11, 11, 01) which agrees with the first 5 pairs of bits of the actual

Dept. of ECE/SJBIT

Page 213

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

encoded sequence. Since the encoder has a memory = 2 we require two more bits to clear and re-set
the encoder. Hence to obtain the complete code sequence corresponding to an information sequence
of length kL, the tree graph is to extended by n(m-l) time units and this extended part is called the
"Tail of the tree", and the 2kL right most nodes are called the "Terminal nodes" of the tree. Thus the
extended tree diagram for the (2, 1, 2) encoder, for the information sequence (10011) is as in Fig 8.19
and the complete encoded sequence is (11, 10, 11, 11, 01, 01, 11).

At this juncture, a very important clue for the student in drawing tree diagrams neatly and
correctly, without wasting time appears pertinent. As the length of the input sequence L increases the
L
L
number of right most nodes increase as 2 . Hence for a specified sequence length, L, compute 2 .
L
Mark 2 equally spaced points at the rightmost portion of your page, leaving space to complete the m
L-l
tail branches. Join two points at a time to obtain 2 nodes. Repeat the procedure until you get only
one node at the left most portion of your page. The procedure is illustrated diagrammatically in Fig
8.20 for L = 3. Once you get the tree structure, now you can fill in the needed information either
looking back to the state transition table or working out logically.

Dept. of ECE/SJBIT

Page 214

www.rejinpaul.com

Information Theory and Coding

www.rejinpaul.com
10EC55

From Fig 8.18, observe that the tree becomes "repetitive' after the first three branches. Beyond the
third branch, the nodes labeled S0 are identical and so are all the other pairs of nodes that are
identically labeled. Since the encoder has a memory m = 2, it follows that when the third information
bit enters the encoder, the first message bit is shifted out of the register. Consequently, after the third
branch the information sequences (000u3u4---) and (100u3u 4---) generate the same code symbols and
the pair of nodes labeled S0 may be joined together. The same logic holds for the other nodes.
Accordingly, we may collapse the tree graph of Fig 8.18 into a new form of Fig 8.21 called a
"Trellis". It is so called because Trellis is a tree like structure with re-merging branches (You will
have seen the trusses and trellis used in building construction).

The Trellis diagram contain (L + m + 1) time units or levels (or depth) and these are labeled
from 0 to (L + m) (0 to 7 for the case with L = 5 for encoder of Fig 8.15 as shown in Fig8.21. The
convention followed in drawing the Trellis is that "a code branch produced by an input '0' is drawn as
a solid line while that produced by an input '1' is shown by dashed lines". The code words produced by
the transitions are also indicated on the diagram. Each input sequence corresponds to a specific path
through the trellis. The reader can readily verify that the encoder output corresponding to the sequence u =
(10011) is indeed v = (11, 10, 11, 11, 01, 01, 11) the path followed being as shown in Fig 8.22.

Dept. of ECE/SJBIT

Page 215

www.rejinpaul.com

Information Theory and Coding

10EC55

www.rejinpaul.com

A trellis is more instructive

than a tree graph in that the finite
state behaviour of the encoder is
explicitly brought-out. Assuming
that the encoder always starts in state
S0 and returns to S0, the first m time units (levels or depth)
correspond
to
the
encoders
departure from S0 and the last mlevels corresponds to its return to S0.
Clearly, not all the states can be
reached in these two portions of the
Trellis. However, in the centre portion
of the Trellis, all states are possible
and each level (time unit) contains a
replica of the state diagram and is
shown in Fig 8.23. There are two
branches leaving and entering each
state.
In the general case of an (n, k, m) code and an information sequence of length kL,
k
kL
there are 2 branches leaving and entering each state and 2 distinct paths through the
kL
trellis corresponding to the 2 code words.
The following observations can be made from the Trellis diagram
1. There are no fundamental paths at distance 1, 2 or 3 from the all zero path.
2. There is a single fundamental path at distance 5 from the all zero path. It diverges from
the all-zero path three branches back and it differs from the all-zero path in the single
input bit.
3. There are two fundamental paths at a distance 6 from the all zero path. One path
diverges from the all zero path four branches back and the other five branches back. Both
paths differ from the all zero path in two input bits. The above observations are depicted
in Fig 8.24(a).
4. There are four fundamental paths at a distance 7 from the all-zero path. One path
diverges from the all zero path five branches back, two other paths six branches back and
the fourth path diverges seven branches back as shown in Fig 8.24(b). They all differ
from the all zero path in three input bits. This information can be compared with those
obtained from the complete path enumerator function found earlier.

Dept. of ECE/SJBIT

Page 216

www.rejinpaul.com

Information Theory and Coding

10EC55

Dept. of ECE/SJBIT

www.rejinpaul.com

Page 217

www.rejinpaul.com