Ece-V-Information Theory & Coding (10ec55) - Notes
Ece-V-Information Theory & Coding (10ec55) - Notes
com
10EC55
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
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
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
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
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:
Reference Books:
Digital Communications - Glover and Grant; Pearson Ed. 2nd Ed 2008
Dept. of ECE/SJBIT
Page 4
www.rejinpaul.com
www.rejinpaul.com
10EC55
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
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.
Dept. of ECE/SJBIT
Page 5
www.rejinpaul.com
www.rejinpaul.com
10EC55
Based on the above intusion, the information content of the k message, can be written as
I (mk)
1
pk
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)
1
I (mk ) = log
------ III
p
k
Dept. of ECE/SJBIT
Page 6
www.rejinpaul.com
www.rejinpaul.com
10EC55
: Hartley / decit
Base - 2
: bit
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
Dept. of ECE/SJBIT
Page 7
www.rejinpaul.com
www.rejinpaul.com
10EC55
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
bits
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
10EC55
bits
i 1
Averageinforamtion content
H=
total
M
NP1 log
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
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
1
pM
1
1
Dept. of ECE/SJBIT
Page 9
www.rejinpaul.com
www.rejinpaul.com
10EC55
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 (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
Dept. of ECE/SJBIT
Page 10
www.rejinpaul.com
www.rejinpaul.com
10EC55
Dept. of ECE/SJBIT
Page 11
www.rejinpaul.com
www.rejinpaul.com
10EC55
n (M)
m = 2 means a 2
order source
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 )
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
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
o The probability of occurrence of a symbol depends on the particular symbol in question and
the symbol immediately proceeding it.
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
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
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
Page 14
www.rejinpaul.com
www.rejinpaul.com
10EC55
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
10EC55
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
A 1
/4
C
C
2 B3
/4
p1 =
P2 =
Dept. of ECE/SJBIT
Page 16
www.rejinpaul.com
www.rejinpaul.com
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
x = /8
3
x = /8
11
Cx+x==88
Message AC: x x =
9
32
3
and so on.
32
Dept. of ECE/SJBIT
Page 17
www.rejinpaul.com
www.rejinpaul.com
8
B
3
8
1
4
10EC55
AC
CB
CC
BB
BC
CA
32
3
32
3
32
2
32
9
32
3
32
3
32
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
10EC55
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
H = pi
n
p
10EC55
log 1
ij
bits / symbol
------ (3)
p
i 1
ij
j 1
C
C
A 1
/4
2 B3
p1 =
/4
P2 =
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
10EC55
8
B
3
8
1
4
AC
CB
CC
BB
BC
CA
32
3
32
3
32
2
32
9
32
3
32
3
32
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
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
1
1
log 2
3 / 4 4
1/ 4
4 1
4
3
= log
log2
4
3 4
H1 = 0.8113
log2 4
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)
I (BB) = I (AC) =
Dept. of ECE/SJBIT
log
3
.
4
1
5
Page 22
www.rejinpaul.com
www.rejinpaul.com
10EC55
b
i
t
s
I (CB) = I (CA) =
Dept. of ECE/SJBIT
(3 /
32)
Page 23
www.rejinpaul.com
www.rejinpaul.com
1 = 4 bits
P(2 / 32)
I (CC) = log =
10EC55
H(two)= Pi
1 bits / sym.
Pi
log
i1
Pi . Ii
i1
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)
Here, N = 2
H
( two )
2.56
2
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
Dept. of ECE/SJBIT
Page 24
www.rejinpaul.com
www.rejinpaul.com
10EC55
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
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.
P(state1) = P(state2) =
P(state3) = 1/3
C
A
Dept. of ECE/SJBIT
Page 25
www.rejinpaul.com
www.rejinpaul.com
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:
------ (1)
Hi
iJ
----- (2)
ij
j1
Hi
log
ij
i 1, 2, 3
------ (3)
ij
j 1
Put i = 1
H
i
log p1 j
1j
j1
x log
1
2
1
2
log
1
2
-0
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
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
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 +
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
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
10EC55
Unit - 2:
SOURCE CODING
Syllabus:
Encoding of the source output, Shannons encoding algorithm. Communication Channels,
Discrete communication channels, Continuous channels.
Text Books:
Reference Books:
Digital Communications - Glover and Grant; Pearson Ed. 2nd Ed 2008
Dept. of ECE/SJBIT
Page 29
www.rejinpaul.com
www.rejinpaul.com
Unit - 2:
10EC55
SOURCE CODING
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.
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
Ni
N GN = H bits / symbol
Performance measuring factor for the encoder
Coding efficiency: c
Definition of c =
c = H(S)
^
HN
Dept. of ECE/SJBIT
Page 30
www.rejinpaul.com
www.rejinpaul.com
SourceOUTPUT
encoder
A message
A unique binary
code word c i of
length n i bits for
the message m i
q
q
ni : an integer
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
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
10EC55
i1
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
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
Probability
27
128128
CCA CCB
CBB
27
128
128
128
128
128
128
128
128
128
128
128
128
128
Probability
27
pi
27 9
128128
128
9
128
BCC ACC
128
128
128
128
128
128
128
128
128
128
CCB CCA
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
10EC55
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
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
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
nn
i
i
Here N = 3,
^
Hi = pij
log
1
p
bits / symbol
ij
j 1
Hi = pij
log
p
ij
j 1
Hi = pij
j 1
log
p11 log
p
1j
p12
11
1
log
p
12
Dept. of ECE/SJBIT
Page 35
www.rejinpaul.com
www.rejinpaul.com
10EC55
H1 = 4 x log 3 / 4 4 log 1/ 4
=
x log
4 1
3
log 4
H1 = 0.8113
H2 = p21 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
H = pi H i ; = p1H1 + p2H2
i 1
Dept. of ECE/SJBIT
Page 36
www.rejinpaul.com
www.rejinpaul.com
10EC55
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
H
c = ^ x100 H
N
N=1
pi
ni
ci
3/8
00
3/8
01
1/4
11
Dept. of ECE/SJBIT
Page 37
www.rejinpaul.com
www.rejinpaul.com
10EC55
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
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
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
AC BB CA AA CB BB
{{{{{{
100101110100 1010 01
Encoded message
DECODING
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
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
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)
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
10EC55
H = 2.12 bits/symbol
(ii)
(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 =
^
Dept. of ECE/SJBIT
Page 41
www.rejinpaul.com
www.rejinpaul.com
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
Example (3)
a) For the source shown, design a source encoding scheme using block size of two
symbols and variable length code words
^
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
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
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
Channel
Encoder
Transmitter
Channel
Encoder
Electrical
Communication
channel
OR
Transmissi
on medium
Physical channel
Noise
g
Demodulator
Channel
Decoder
Receiver
Communication Channel
Dept. of ECE/SJBIT
Page 44
www.rejinpaul.com
www.rejinpaul.com
10EC55
A discrete channel
(ii)
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.
Dept. of ECE/SJBIT
Page 45
www.rejinpaul.com
www.rejinpaul.com
P00
O
P10
Transmitted
Received
digit X
digit X
P01
p11
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)
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).
Pe = po p01 + p1 p10
------ (1)
Dept. of ECE/SJBIT
Page 46
www.rejinpaul.com
www.rejinpaul.com
----- (2)
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
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
10EC55
Entropy of INPUT X:
M
H(X).
t bits / symbol
HX pit log pi
----- (5)
i1
ii)
r bits / symbol
HY pir log pi
i 1
iii)
HX / Y
P (X i, Y j) log p (X i / Y j) bits/symbol
i1
iv)
- (7)
j1
HX, Y
P (X i, Y j) log p (X i, Y j) bits/symbol
i1
v)
----- (6)
- (8)
i1
H (Y/X)
P (X i, Y j) log p (Y j / X i) bits/symbol
i1
- (9)
i1
H(X/Y) represents how uncertain we are of the channel input x, on the average, when we know
the channel output 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
10EC55
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)
1. For the discrete channel model shown, find, the probability of error.
p
1
Transmitted
digit
Received
digit
Dept. of ECE/SJBIT
Page 49
www.rejinpaul.com
www.rejinpaul.com
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
y1
x2
y2
----- (1)
. 3
y1
y2
x1
1
4
2
3
----- (2)
2
12
Dept. of ECE/SJBIT
Page 50
www.rejinpaul.com
www.rejinpaul.com
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)
p(XY) log
1
p(XY)
H(X/Y) =
p(XY) log
1
p(X / Y)
p(XY) log
= 1.562
1
p(Y / X)
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
10EC55
1
2
P(Y/X) =
3
1
2
2
5
6
0
2
5
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(Y) pY.log 1
(ii)
(iii)
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
10EC55
H(XY) = 3.12192
(iv)
p(X / Y)
(v)
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
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
Dept. of ECE/SJBIT
Page 53
www.rejinpaul.com
www.rejinpaul.com
10EC55
Recall
P(XY) = p(X), p(Y/X)
P(X1Y1) = p(X1) . p(Y1X1) = 1 . 1
3
12
12
12
12
P(X/Y) =
6
1
p(Y / X)
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]
Dept. of ECE/SJBIT
1
p(XY)
3.1883 bits/sym.
Page 54
www.rejinpaul.com
www.rejinpaul.com
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)
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
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
----- (1)
i1
Assumption is that the symbol in the sequence at the input to the channel occur in a statistically
independent fashion.
----- (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
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.
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
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
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)]
= - 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
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
p(Y = 0) =
1
2
p(X = 0 /Y = 0) = 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
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)
P(X = 0) = P
P(X = 1) = Q
P(X = 2) = Q
Dept. of ECE/SJBIT
Page 58
www.rejinpaul.com
www.rejinpaul.com
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
= - 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
-----(1)
H(X/Y) =
i
H(X) =
-----(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
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
-----(3)
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
1
0
, p
31
Dept. of ECE/SJBIT
31 0
;p
33 1
22
33
Page 60
www.rejinpaul.com
www.rejinpaul.com
10EC55
p(00) log p
p(01) log p
0
1
p(10) log p
p(11) log p
15
32
log
log p
30
31
31
33
1
32
1
64
log
log
2
33
1
31
64
31
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)
-----(2)
and H(Y/X) =
i
Dept. of ECE/SJBIT
-----(3)
Page 61
www.rejinpaul.com
www.rejinpaul.com
10EC55
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
= 1.25 + 0.934
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)
-----(3)
P(x)
Dept. of ECE/SJBIT
Page 62
www.rejinpaul.com
www.rejinpaul.com
P(Y/X)
10EC55
Dept. of ECE/SJBIT
Page 63
www.rejinpaul.com
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
10EC55
This
1
x
1 x , equality iffy x = 1
..
(2.2)
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
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)
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
10EC55
P1 = {p1, p2, p3, p4}, P2 = {p3, p2, p4, p1} and P3 = {p4, p1, p3, p2} with pk 1 all have
k 1
H ( S ) p .log 1
k
pk
k 1
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
Since p
k 1
k 1
k =1, 2, 3q.
k =1, 2, 3q.
k 1
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)
Dept. of ECE/SJBIT
Page 66
www.rejinpaul.com
www.rejinpaul.com
10EC55
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
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
i 1
i 1
q
i1 1
i2 1
P(sin )
in 1
Dept. of ECE/SJBIT
Page 67
www.rejinpaul.com
www.rejinpaul.com
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
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
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.
P= {p1, p2 .pq}: p1 p2
pq
3. Determine the set of integers, lk, which are the smallest integers solution of the inequalities.
l
lk
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
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
l k
, X { 0 ,1 }
4 8 8 16 16 16 16
Dept. of ECE/SJBIT
Page 70
www.rejinpaul.com
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
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
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
www.rejinpaul.com
10EC55
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
10EC55
2) lk lk 1 k 1,2 ,......q
(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
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
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
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
10EC55
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
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.
= 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
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:
(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
10EC55
Codes
lk
0
1
11
2
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
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.
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)
......... (4.5)
k j
..(4.6)
............
(4.7)
And also
Dept. of ECE/SJBIT
Page 82
www.rejinpaul.com
www.rejinpaul.com
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,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
10EC55
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) = 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
. (4.11)
p( y j )
j1k 1
P{X xk ,Y y j }
P{Y y j }
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
= 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
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
Or
p(y j | x k )
p(x k , y j )log
k
1
p(x k )
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
Dept. of ECE/SJBIT
H(X) H (X|Y)
(4.18)
Page 85
www.rejinpaul.com
www.rejinpaul.com
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
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
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
Dept. of ECE/SJBIT
P(x k , y j )
P( y j )
we have:
Page 86
www.rejinpaul.com
www.rejinpaul.com
10EC55
th
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
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
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
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)
Dept. of ECE/SJBIT
Page 88
www.rejinpaul.com
www.rejinpaul.com
10EC55
p(xk , y j )
p(y j | xk ) ,
p(xk )
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
. 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
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,
Dept. of ECE/SJBIT
Page 90
www.rejinpaul.com
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
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
Dept. of ECE/SJBIT
Page 92
www.rejinpaul.com
www.rejinpaul.com
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:
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
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
Dept. of ECE/SJBIT
Page 93
www.rejinpaul.com
www.rejinpaul.com
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
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
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
h1
h
h
n
Page 94
www.rejinpaul.com
www.rejinpaul.com
10EC55
n
C log
i 1
Q C
pi 2
pi
p2 ..... pn . 2 i
M
p
ni
From this we can solve for the source probabilities (i.e. Input symbol probabilities):
-1
1
2
1
2
3
Y
|
X
.
1
1
Q
h
2
C log 2
Further ,
p
p
1
0.6225562
Q1
Q 2
2
p
2
Q1 C
.P 1
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
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
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
=0.3348213 = p3, p2 = 2
-Q2 C
=0.3303574.
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
C = log {2 2
p1 =2
-Q1 C
Giving:
+0.39036
+2
=0.2163827 = p3, p2 = 2
=0.5672345
Dept. of ECE/SJBIT
Page 96
www.rejinpaul.com
www.rejinpaul.com
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
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
(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
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
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
4 1
log
3444
log4 log4
log 3 2
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
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
10EC55
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).
(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)
Dept. of ECE/SJBIT
Page 101
www.rejinpaul.com
www.rejinpaul.com
10EC55
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
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.
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
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
10EC55
p (xr, yj) = 0, r k, r = 1, 2, 3 m.
m
p( xk | y j )
p( xk , y j )
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
10EC55
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
Thus we have:
m
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
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
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
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
10EC55
5
P( Z | X )
36 12 9
1
1 1
3
6
2
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
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
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
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
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
10EC55
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)
0x2
Else where
2
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
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
f ( x )dx 1
3)
4)
Dept. of ECE/SJBIT
Page 114
www.rejinpaul.com
www.rejinpaul.com
10EC55
5) Average power limitation, with unidirectional distribution (causal systems whose response
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)
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
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
- (1- )
- (1- )
(5.9)
5.10)
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.
..
(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)
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)
Dept. of ECE/SJBIT
Page 116
www.rejinpaul.com
www.rejinpaul.com
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)
. (5.16b)
1 2
1
( 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)
-ax
f(x) = ae
(5.18)
(5.19)
a=1/
, 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)
f ( x )dx 1
5.22a)
Dept. of ECE/SJBIT
Page 117
www.rejinpaul.com
www.rejinpaul.com
10EC55
. f ( x )dx
(5.22b)
( 1 x2 )
fe
( 1 )
.e
ax2
.. (5.23)
.. (5.24)
dx
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.
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
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
10EC55
(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
(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)
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)
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
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)
(5.35)
= H(Y) - H (Y|X)
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 )
H ( Y | x ) fn ( z ) log
dz H ( n )
fn ( z )
Accordingly,
Or
H (Y|X) =H (n)
(5.39)
=H(Y) H (Y|X)
. (5.40)
(5.41)
Dept. of ECE/SJBIT
Page 120
www.rejinpaul.com
www.rejinpaul.com
10EC55
.. (5.42)
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)
(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)
(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
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
(5.50)
(5.51)
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
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
N1
Suppose we decrease the band width by 25% and choose the new values of SNR and band width as
S2
N2
Then
S1
N1
S2
=7/15
N2
log 1
5.53)
S/N
A plot
0.5
) is shown
in Fig
10
5.4. Clearly,
15
the same
20
channel
30
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
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).
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
10EC55
max
lim
log
1 x x
(5.56)
(5.57)
x 0
N0
Since lim 1 x x e
x 0
lim
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)
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
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
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
(5.64)
0 min
Eb
The value
min
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:
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
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 =10
=10 =1000
C/B
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
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
10EC55
Solution:
a) S/N = 10
20/10
= 10 = 100
3 x105
5
Dept. of ECE/SJBIT
Page 128
www.rejinpaul.com
www.rejinpaul.com
Review questions:
1. Show that for a AWGN channel C
10EC55
watts/Hz.
2. Consider an
1012
2
1.448
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
5. FX(X)
0XA
0 OTHERWISE
Dept. of ECE/SJBIT
Page 129
www.rejinpaul.com
www.rejinpaul.com
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
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
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.
Dept. of ECE/SJBIT
Page 132
www.rejinpaul.com
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
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
-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:
S S
).
-3)
-3
-3
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
Dept. of ECE/SJBIT
Page 134
www.rejinpaul.com
www.rejinpaul.com
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
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
-6
-9
-12
-3
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
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
Dept. of ECE/SJBIT
Page 135
www.rejinpaul.com
www.rejinpaul.com
10EC55
(1-p) +
3
-6
-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
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
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)
g2
g3
g11
21
L g1n
12
22
L g2 n
(6.5)
Lg
g
k1
k2
kn
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
Dept. of ECE/SJBIT
Page 138
www.rejinpaul.com
www.rejinpaul.com
10EC55
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
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
Dept. of ECE/SJBIT
Page 139
www.rejinpaul.com
www.rejinpaul.com
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
th
i element
th
(k + j) element
th
th
th
element
(k + j) element
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)
G. H = Ok (n k)
Dept. of ECE/SJBIT
(6.11)
Page 140
www.rejinpaul.com
www.rejinpaul.com
10EC55
v.H = O.
T
P
I
nk
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
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
Dept. of ECE/SJBIT
(5.12)
Page 141
www.rejinpaul.com
www.rejinpaul.com
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:
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
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
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
Dept. of ECE/SJBIT
Page 143
www.rejinpaul.com
www.rejinpaul.com
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,
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)
=
Dept. of ECE/SJBIT
Page 144
www.rejinpaul.com
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:
.(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
Dept. of ECE/SJBIT
Page 145
www.rejinpaul.com
www.rejinpaul.com
10EC55
(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
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
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)
(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
ii)
10EC55
(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
where
(d
parameter t =
(d
(d
1)
min
1)
min
10EC55
.. (5.35)
1
2
( dmin 1 ) . 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
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
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
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
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
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
th
Dept. of ECE/SJBIT
n-k
vj}
..
6.39)
Page 150
www.rejinpaul.com
www.rejinpaul.com
10EC55
n-k
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
= vj e = vj (ei vl ) = ei vm
e
th
column and l
th
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
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
-3
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
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.).
(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
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
10EC55
G I2 m m 1 M Q
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
k=2 -m-l-1
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
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
Dept. of ECE/SJBIT
Page 155
www.rejinpaul.com
www.rejinpaul.com
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
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)
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
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
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
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)
(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
www.rejinpaul.com
10EC55
Dept. of ECE/SJBIT
Page 161
www.rejinpaul.com
www.rejinpaul.com
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)
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
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
Dept. of ECE/SJBIT
Page 163
www.rejinpaul.com
www.rejinpaul.com
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
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:
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
(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
10EC55
(6.13)
and
(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
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
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
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
viz. lk = logr
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
k 1
log
k 1
1
2
pk
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
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)
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
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.
(1)
(7.2)
is also a code vector. In this way, the n - tuples obtained by successive cyclic right shifts:
vn-3)
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
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
n-1
+ vn-1 X
= 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)
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)
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
Where
10EC55
k-I
..
(7.9)
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)
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
nk i
gnk
..
(7.13
(7.13
a)
b)
(7.13
c)
(7.13d)
Dept. of ECE/SJBIT
Page 170
www.rejinpaul.com
www.rejinpaul.com
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
(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
n-1
m-1
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
10EC55
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
Dept. of ECE/SJBIT
Page 172
www.rejinpaul.com
www.rejinpaul.com
10EC55
Let us choose g (X) = X + X + 1 as the generator polynomial. The encoding circuits are shown in
Fig 7.4(a) and (b).
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.
Dept. of ECE/SJBIT
Page 173
www.rejinpaul.com
www.rejinpaul.com
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
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.
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
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
0
1
1
1
1
1
1
0
1
1
0
0
Out
put
0
0
0
1
1
1
1
Remarks
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
7.4
(7.15)
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)
.........................
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
th
10EC55
n-k
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
+X
n-k
7.20)
n-k
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
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
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
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
10EC55
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.
Dept. of ECE/SJBIT
Page 179
www.rejinpaul.com
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
10EC55
such simplicity is in the reduction of code efficiency for a given block size.
7.7
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
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
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
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)
15
+X
11
+X
12
16
12
+X
16
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
10EC55
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
Dept. of ECE/SJBIT
Page 185
www.rejinpaul.com
www.rejinpaul.com
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
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 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
10EC55
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
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
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
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
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
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
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=
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
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
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
www.rejinpaul.com
10EC55
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
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
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
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
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 }.
(1)
(2)
= {g1
= {g1
(1)
(2)
(1)
(1)
(2)
(2)
, g2 , g3
, g2 , g3
, g4
, g4
(1)
(2)
} and
}
(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
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)
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),
Dept. of ECE/SJBIT
Page 201
www.rejinpaul.com
www.rejinpaul.com
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)
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
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)
}.
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)
v (2) = u
* g1
(2) +
(2)
u (2)
* g2 (1)
* g2 (2)
(8
.5 a)
(8.5 b)
(8.5 c)
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)
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
10EC55
(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)
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
10EC55
m Max ki
(8.7)
1i k
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
Dept. of ECE/SJBIT
Page 205
www.rejinpaul.com
www.rejinpaul.com
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)
v (X) = v1
(1)
Dept. of ECE/SJBIT
+ v2
(1)
X + v3
(1)
..
.....
(8.12a)
(8.12
X +....., and
Page 206
www.rejinpaul.com
b)
www.rejinpaul.com
(2)
(2)
(X) = v1
10EC55
(2)
+ v2
X + v3
(2)
X +.....
(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)
For the (2, 1, 3) encoder of Fig 8.3, the impulse responses were: g
(l)
(2)
(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 + X ) = 1 + X + X + X + X + X
(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
(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
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 1 (3) (X) = 1 ,
g 2 (3) (X) = 0
Dept. of ECE/SJBIT
Page 208
www.rejinpaul.com
www.rejinpaul.com
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 ]
(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
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:
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
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)
(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
www.rejinpaul.com
10EC55
Dept. of ECE/SJBIT
Page 212
www.rejinpaul.com
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
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
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
www.rejinpaul.com
Dept. of ECE/SJBIT
Page 216
www.rejinpaul.com
Dept. of ECE/SJBIT
www.rejinpaul.com
Page 217
www.rejinpaul.com