Information Channel 15EC54

MODULE – 3 INFORMATION CHANNELS

STRUCTURE

1. Objectives
2. Introduction
3. Communication channel
4. Channel model and channel capacity
5. Mutual information
6. Review questions.
7. Outcomes.

OBJECTIVES

After completion of this module the student will be able

1. To learn about different Communication channel in communication systems.

2. To find channel capacity of different channels in communication system.
3. To develop channel matrix and to find out mutual information in channel.

Page 1
 Information Channel 15EC54

3.1 COMMUNICATION CHANNELS:

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 impedances. For a passive network the
impedance matrix is always symmetric i.e. ZT = Z, where the superscript indicates
transposition.

Similarly, a communication network may be uniquely described by

specifying the joint probabilities (JPM). Let us consider a simple communication
network comprising of a transmitter (source or input) and a receiver (sink or
output) with the interlinking medium-the channel as shown in Fig 4.1.

Fig 4.1 A Simple Communication System

This simple system may be uniquely characterized by the „ Joint probability

matrix‟ ( JPM),
P(X, Y) of the probabilities existent between the input and output ports.

p( x1 , y1 p( x1 , y2 P( x1 , y3 ) .....
) )
P( x2 , y3 ) .....
P( x2 , y2
P( x2 , y1 P( x3 , y3 ) .....
)
P( X ,Y ) 

) M M .......... (4.1)
P( x3 , y1 P( x3 , y2 M M
)
) P( xm , y3 ) .....
M
M
M
M
P( xm , y2
P( xm , y1 )
)

For jointly continuous random variables, the joint density function satisfies the following:


.2)
∫ ∫ f ( x, y)dxdy  1


 
∫ f ( x, y)dy  f X (
…………. (4.3)
x)

 ……… ………….. (4.4)

∫ f ( x, y)dx  fY ( y)

We shall make use of their discrete counterpart as below:

\ ∑∑ p( xk , y j )  1 ,Sum of all entries of JPM ......... (4.5)

k j

∑ p( xk , y j )  p( xk ) , Sum of all entries of JPM in the kth r …………..(4.6)

j

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

column ............ (4.7)
k
And also

Page 2
Information Channel 15EC54

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.

3.2 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) p(x1 , y1 ) + p(x1, y2) log p(x1 , y2 ) +…+ p(x1, yn) log p(x1 , yn )
log
1 1 1
+ p (x2, y1) log + p(x2,y2) log +…+ p(x2,yn) log

p(x2 , y1 ) p(x2 , y2 ) p(x1 , y1 )

Page 3
Information Channel 15EC54

1 1 1
+… p (xm, y1) log + p(xm,y2) log +… p(xm,yn) log or
p(xm , y1 ) p(xm , y2 ) p(xm , yn )
m n
1
H(X, Y) = ∑∑p(x k , y j )
log ………………(4.9)
k 1 j1 p(x k , y j )

m 1
H(X) = ∑ p(xk ) log
k 1 p( xk )
Using Eq (4.6) only for the multiplication term, this equation can be re-written as:
m n 1
H(X) = ∑ ∑ p( xk , y j ) log ………………… (4.10)
k 1 j1 p( xk )
n m 1
Similarly, H(Y) = ∑ ∑ p( xk , y j ) log ………………. (4.11)
j1k 1 p( y j )

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

P{X  xk ,Y  y j }
P{X = xk | Y = yj} =
P{Y  y j }
i.e., p(xk | y j ) = p(xk , y j ) ) / p (yj)

m 1 m 1
Then ∑ p(xk | y j )= ∑ p(xk ,yj )= . p(y j ) =1 ………. (4.12)
k 1 p( y j ) k 1 p( y j )

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:
m 1
H(X | yj) = ∑ p(xk | y j )log .
k 1 p(xk | y j )
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)
j1
n m 1
= ∑ p( y j ) ∑ p( xk | y j ) log
j1 k 1 p( xk | y j )
nm
1
Or H(X | Y) = ∑∑ p(xk , y j ) log ……………… (4.13)
j 1 k 1 p(xk | y j )

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:

Page 4
Information Channel 15EC54

m n 1
H(Y | X) = ∑ ∑p( xk , y j ) log ……………… (4.14)
k 1 j1 p( y j | xk )
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 ) 1

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

1 p(xk )
Since, p( y | x ) = p(xk , y j )
j k
We can straight away write:
H (Y|X) = ∑∑p(x k , y  ∑∑p(x
1 1
j
)log k , y j )log
k p(y j | x k )
j k j p(x k )
Or H(Y | X) = H(X, Y) – H(X)

That is: H(X, Y) = H(X) + H(Y | X) ……………………….. (4.15)

Similarly, you can show: H(X, Y) = H(Y) + H(X | Y) ………………. (4.16)
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 )
k j k k j

= ∑∑ p(xk , y j p(xk , y j )
)log (4.17) ………

k j p(xk ) . p(y j )
Using the logarithm inequality derived earlier, you can write the above equation as:
H(X) - H(X |Y) =log e ∑∑ p(xk , y p(xk , y j )
j )ln
k j p(xk ) . p(y j )
p(xk ) . p(y j
≥ log e ∑∑ p(xk , y j ) 1 -
kj p(x , y )
k j

≥log e ∑ ∑ p(xk , y j ) - ∑∑ p(xk).p(yj)

k j k j

≥log e ∑ ∑ p(xk , y j ) - ∑ p(xk).∑ p(yj) ≥ 0

k j k j

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

k j k j

H(X) ≥ H (X|Y) ………. (4.18

Page 5
Information Channel 15EC54

Similarly, H(Y)  H (Y|X) ………….. (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
0.1 0.01 0.01 0.01

P(X, Y) = 0 0.02 0.02 0

0.04 0.04 0.01 0.06

0 0.06 0.02 0.2

n
Using p (xk) = ∑ p(xk , y j) , we have, by adding entries of P(X, Y) row-wise we get:
j1
P(X) = [0.4, 0.1, 0.04, 0. 15, 0.28]
Similarly adding the entries column-wise we get:

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

Hence we have:
H(X,Y )  3  0.2 log 1  0.1 log 1  4  0.01 log 1 
0.2 0.1 0.01
3  0.02 log 1  2  0.04 log 1  2  0.06 log 1
0.02 0.04 0.06
= 3.188311023 bits /sym

1 1 1 1 1
H(X) = 0.4 log  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
H(Y) = 0.34 log  0.13 log 0.26 log  0.27 log
0.34 0.13 0.26 0.27
= 1.927127708 bits / sym

P(x k , y j )
Since p (xk | yj) = we have:
P( y j )

Page 6
Information Channel 15EC54

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

0.2 0.2
0 0
0.26
0.1 0.01 0.01
0.26 0.27
0.02
P (X| Y) = 0
0.26
0
0.06
0.26 0.27
0.06 0.02 0.20
0
0.13 0.26 0.27

 H(X | Y)  0.2log 0.34  0.2log 0.26  0.1log 0.34

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

=1.261183315 bits / symbol

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

0.2 0.2
0 0
0.4 0.4
0.1 0.01 0.01 0.01
0.13 0.13 0.13 0.13
0.02 0.02
P(Y | X)  0 0
0.04 0.04
0.04 0.04 0.01 0.06
0.15 0.15 0.15 0.15
0.06 0.02 0.20
0
0.28 0.28 0.28

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)

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
0.2 0.1 0.01 0.02

 2  0.04 log 0.05  0.01 log 0.15  0.06 log 0.15  0.06 log 0.28

Page 7
Information Channel 15EC54
0.04 0.01 0.06 0.06

 2  0.02 log 0.28  1.166376202 bits / sym .
0.02

3.3 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 1
I(x k , y j )  log p(x )  log p(x | y )
k k j

p(xk | y j )
 log ……………………(4.21 a)
p(xk )
p(x k .y j )
OrI (xk, yj)  log ……………… (4.21 b)
p(x k ).p(y j )
Notice from Eq. (4.21a) that
I (xk) = I (xk, xk) = log p(x k | x k )  log 1
p(x k ) p(xk )
This is the definition with which we started our discussion on information theory! Accordingly I (xk)
is also referred to as „Self Information

Page 8
Information Channel 15EC54

p(xk , y j )
It is clear from Eq (3.21b) that, as  p(y j | xk ) ,
p(xk )

p(y j | x k ) 1 1
I (x k , y j )  log = log  log
p(y j ) p(y j ) p(y j | x k )
Or I (xk, yj) = I (yj) – I (y j |xk) …………… (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
∑∑ k

j ). p(xk , y j )
k j
= p(x , y p(xk , y j )
∑∑ k

j ).log …………. (4.24) From Eq

k j p(xk )p(y j )
(4.24) we have:
1 1
1) I(X, Y) = ∑∑p(xk , y ) log .  log = H(X) – H(X | Y)
j
p(x ) p(x | y )
k j k k j
…… (4.25)
I(X, Y) = ∑∑ p(xk , y 1 1
2) j ) [log .  log
k j p(y j ) p(y j | xk )
= H(Y) – H(Y | X) ……………………. (4. 26)
I(X,Y)  ∑∑ p(xk , 1  .∑∑ p(xk , y j 1
3)y j )log ) log 
k j p(xk ) k j p(y j )
∑∑ p(xk , 1
yj ) log
KJ p(xk y j )
Or I(X, Y) = H(X) + H(Y) – H(X, Y) ………………….. (4.27 )

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.

Page 9
Information Channel 15EC54

The entropy of X is represented by the circle on the left and that of Y by the circle on the right. The
overlap between the two circles (dark gray) is the mutual information so that the remaining (light
gray) portions of H(X) and H(Y) represent respective equivocations. Thus we have

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

The joint entropy H(X,Y) is the sum of H(X) and H(Y) except for the fact that the overlap is added
twice so that
H(X, Y) = H(X) + H(Y) - I(X, Y)

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

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

For the JPM given in Example 4.1, I(X, Y) = 0.760751505 bits / sym

Shannon Theorem: Channel Capacity:

Clearly, the mutual information I (X, Y) depends on the source probabilities apart from the
channel probabilities. For a general information channel we can always make I(X, Y) = 0 by choosing
any one of the input symbols with a probability one or by choosing a channel with independent input
and output. Since I(X, Y) is always nonnegative, we thus know the minimum value of the
Transinformation. However, the question of max I(X, Y) for a general channel is not easily answered.

Our intention is to introduce a suitable measure for the efficiency of the channel by making a
comparison between the actual rate and the upper bound on the rate of transmission of information.
Shannon‟s contribution in this respect is most significant. Without botheration about the proof, let us
see what this contribution is.

Shannon’s theorem: on channel capacity(“coding Theo rem”)

It is possible, in principle, to device a means where by a communication system will transmit

information with an arbitrary small probability of error, provided that the information rate R(=r×I
(X,Y),where r is the symbol rate) is less than or equal to a rate „ C‟ called “channel capacity”.

The technique used to achieve this objective is called coding. To put the matter more
formally, the theorem is split into two parts and we have the following statements.

Page 10
Information Channel 15EC54

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):

C = ∆ Max I(X, Y) = Max [H(X) – H (Y|X)] …………. (4.28)

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

Page 11
Information Channel 15EC54

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.

H ( S ) I ( X ,Y )Max
Conversely, if  , it is not possible to transmit information over the T s Tc
channel and reconstruct it with an arbitrarily small probability of error

A communication channel, is more frequently, described by specifying the source

probabilities P(X) & the conditional probabilities P (Y|X) rather than specifying the JPM. The CPM, P
(Y|X), is usually refereed to as the „ noise characteristic‟ of the channel. Therefore unless otherwise
specified, we shall understand that the description of the channel, by a matrix or by a „Channel diagram‟
always refers to CPM, P (Y|X). Thus, in a discrete communication channel with pre-specified noise
characteristics (i.e. with a given transition probability matrix, P (Y|X)) the rate of information
transmission depends on the source that drives the channel. Then, the maximum rate corresponds to a
proper matching of the source and the channel. This ideal characterization of the source depends in turn
on the transition probability characteristics of the given channel.

Redundancy and Efficiency:

A redundant source is one that produces „dependent‟ symbols. (Example: The Markov
source). Such a source generates symbols that are not absolutely essential to convey information. As
an illustration, let us consider the English language. It is really unnecessary to write “U” following
the letter “Q”. The redundancy in English text is e stimated to be 50%(refer J Das etal, Sham
Shanmugam, Reza, Abramson, Hancock for detailed discussion.) This implies that, in the long run,
half the symbols are unnecessary! For example, consider the following sentence.

“ Y.u sh..ld b. abl. t. re.d t.is ev.n tho… sev.r.l l.t..rs .r. m.s..ng ”

However, we want redundancy. Without this redundancy abbreviations would be impossible

and any two dimensional array of letters would form a crossword puzzle! We want redundancy even
in communications to facilitate error detection and error correction. Then how to measure
redundancy? Recall that for a Markov source, H(S) < H(S), where S is an ad- joint, zero memory
source. That is, when dependence creeps in, the entropy of the source will be reduced and this can be
used as a measure indeed!
“ The redundancy of a sequence of symbols is measured by noting the amount by which the entropy
has been reduced”.

When there is no inter symbol influence the entropy at the receiver would be H(X) for any
given set of messages {X} and that when inter symbol influence occurs the entropy would be H (Y|X).
The difference [H(X) –H (Y|X) ] is the net reduction in entropy and is called “ Absolute Redundancy”.
Generally it is measured relative to the maximum entropy and thus we have for the “ Relative
Redundancy” or simply, „ redundancy‟ , E

E = (Absolute Redundancy) ÷ H(X)

Page 12
Information Channel 15EC54

H(Y | X)
Or E  1  ………………………. ( 4.29)
H(X)

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

R
E  1  ………………………… (4.30)
C

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, η for the channel immediately
follows:
η  Actual rate of mutual information
maximum possible rate

R
That is. η  ……… ……………………. (4.31)
C

and η  1 E ……………………………… (4.32)

3.4 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 1
P(Y | X ) 2 2
1 3
4 4

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

Clearly, H(X) = - p log p - (1 - p) log (1 - p)

We shall first find JPM and proceed as below:

p p
p(x1 ). p(y1 | x1 ) p(x1 ). p(y2 | x1 ) 2 2
P(X,Y)  =
p(x2). p(y1 | x 2
)
) p(x2). p(y2 | x2 ) 1 p 3(1  p)
4 4

Adding column-wise, we get:

Page 13
Information Theory and Coding 15EC54

p (y1) = p  1 - p  1  p and p (y2) = p 3(1 - p)  3  p

2 4 4 2 4 4
1p 4 3 -p 4
Hence H(Y) = log  log

4 1p 4 3-p
And H (Y|X) = p log 2  p log 2  1  p log 4  3(1  p) log 4
2 2 4 4 3

3 log 3 3log3 1 p 3(1 - p)

I(X, Y) = H(Y) – H (Y|X) = 1 p  log(1  p)  log (3  p )
4 4 4 4
dI
Writing log x = loge× ln x and setting = 0 yields straight away:
dp
p  3a  1  0.488372093 , Where a =2(4-3log3) = 0.592592593
1a
With this value of p, we find I(X, Y) Max = 0.048821 bits /sym

For other values of p it is seen that I(X, Y) is less than I(X, Y) max

Although, we have solved the problem in a straight forward way, it will not be the case

p . 0.2 . 0.4 .0.5 .0.6 .0.8

I(X,Y) .0.32268399 .0.04730118 .0.04879494 .0.046439344 .0.030518829
Bits / sym

When the dimension 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.

Muroga‟s Theorem :

The channel capacity of a channel whose noise characteristic, P (Y|X), is square and non-
singular, the channel capacity is given by the equation:
i n
C  log ∑2 Qi …………………. (4 .33)
i 1

Where Qi are the solutions of the matrix equation P (Y|X).Q = [h], where h= [h1, h2, h3, h4… h n] t
are the row entropies of P (Y|X).

p11 ..... p Q1 h1
1n
22 ..... p2n

M
pnn Q
n1 n2 n3 ..... n n

Page 14
 Information Theory and Coding 15EC54

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

[p1, p2, p3 … p n] = [p1, p2, p3… p n ] P-1 [Y|X], provided the inverse exists.

However, although the method provides us with the correct answer for Channel capacity, this
value of C may not necessarily lead to physically realizable values of probabilities and if P-1 [Y|X]
does not exist ,we will not have a solution for Qi`s as well. One reason is that we are not able to
incorporate the inequality constraints 0≤ pi ≤ 1 .Still, within certain limits; the method is indeed very
useful.
Example 3.2: Consider a Binary channel specified by the following noise characteristic (channel matrix):

1 1
P(Y | X )  2 2
1 3
4 4

The row entropies are:

h  1 log 2  1 log 2  1 bit / symbol .
1
2 2
h  log 4  log 4  0.8112781 bits / symbol .
1 3
2
4 4 3
3 1
P Y | X 
1

1 2
Q h 1.3774438
Q2
1 P
1
 Y | X  . 1  h 2 0.6225562

 
C  log 2Q1  2 Q 2  0.048821 bits / symbol , as before.

Further , p  2Q1 C  0.372093 and p   2Q2 C  0.627907.
1 2
p p p p .P 1 Y|X 0.488372 0.511628
1 2   1 2     
Giving us p = 0.488372

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54

Example 3.3:

Consider a 33 channel matrix as below:

0.4 0.6 0
PY | X  0.5 0 0.5

0 0.6 0.4
The row entropies are:

h1 = h3 = 0.4 log (1/0.4) + 0.6 log (1/0.6) = 0.9709505 bits / symbol.

h2 = 2  0.5 log (1/0.5) = 1 bit / symbol.

1.25 1  1.25
1
P Y | X  56 2 3 5 6

 1.25 1 1.25
Q1 1
2
Q  1.0193633

Q
3 1
-1 -1.0193633
C = log {2 + 2 + 2-1} = 0.5785369 bits / symbol.

p1 =2-Q1 –C =0.3348213 = p3, p2 = 2-Q2 –C =0.3303574.

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

Suppose we change the channel matrix to:

0.8 0.2 0 0.625 1  0.625

PY | X  0.5 0 0.5 P Y | X 
1
2.5 4 2.5

0 0.2 0.8  0.625 1 0.625

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

This results in:

Q1 = Q3 = 1; Q2 =  0.39036.

C = log {2  21 + 2+0.39036} = 1.2083427 bits / symbol.

p 1 =2-Q1 –C =0.2163827 = p3, p2 = 2-Q2 –C =0.5672345

Giving: p1 = p3 =1.4180863 and p2 = Negative!

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54

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 Muroga‟s 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:

p p

……………
p p p
(4.34)

This channel is said to be Symmetric or Uniform if the second and subsequent rows of the
channel matrix are certain permutations of thenn 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:

p1 p2 p3 ... pn
p
n 1
p2 pn ... p4
P [Y | X ]  p3 p2 p1 ... p5 …………… (4.35)

M M M M M
P p
pn n 1 n 2 ... p1
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 ……………… (4.36)
j 1

The conditional entropy H (Y|X) for this channel can be computed from:
m n 1
H ( Y | X )  ∑ ∑ p( xk , y j )log
k 1 j 1 p( xk , y j )
m n 1
 ∑ p( xk ).∑ p( y j | xk )log
k 1 p( y j | xk )
j 1

However, for the channel under consideration observe that:

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54

m n n
∑p( xk ).∑ p( y j | xk )log 1  ∑ p j log
1  h……… (4.37)
k 1 j 1 p( y j | xk ) j 1 pj
is a constant, as the entropy function is symmetric with respect to its arguments and depends only on
the probabilities but not on their relative locations. Accordingly, the entropy becomes:
m

H (Y | X )  ∑ p( xk ).h  h ……………..(4.38)
k 1
as the source probabilities all add up to unity.

Thus the conditional entropy for such type of channels can be computed from the elements of any
row of the channel matrix. Accordingly, we have for the mutual information:

I(X, Y) = H(Y) – H (Y|X)

= H(Y) – h
Hence, C = Max I(X, Y) =Max
{H(Y) – h} =
Max H(Y) – h

Since, H(Y) will be maximum if and only if all the received symbols are equally probable and as
there are n – symbols at the output, we have:
H(Y) Max = log n

Thus we have for the symmetric channel:

C = log n – h …………… (4.39)

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

We define the channel to be symmetric if the CPM, P (X|Y) has the form:

p1 pm p2 .. pm
p . p
m 1 m 1
p2 p6 ..
.
P( X | Y )  p3 p4 pm ..
p
m 2 ……………….(4.40)
.
M M M M M

p
pm p1 m 3 .. p1
.
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:
nm n m
1 1
H ( X | Y )  ∑∑ p( xk , y j  ∑ p( y j )∑ p( xk | y j )log
)log p( xk | y j ) j 1 k 1 p( xk | y j )
j 1 k 1
n m m
Since ∑ p( y j )  1 and p( x | y )log 1  p log
1  h is a constant, because
k

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:

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54

m
 p log 1
k

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 3.4:

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 )  3366

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):
1 1
H (Y|X) = h =2× log 3 + 2× 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, ATMECE

 Information Theory and Coding 15EC54

0 1
02 1
3 3
1 2
1
3 3
(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 1
P( X | Y )  3
2
3 3
H ( Y | X )  h  log  log 3  log 3   0.918295833 bits / symbol .
3 2 3 3
3 4 1 3 3
H ( X )  log  log4  log4  log 3  2  log 3  0.811278125 bits / symbol . 4
3444

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

2
3 1 3 1 1
 
P( X ,Y ) 3 4 3 4  2 4
1 1 2 3 1 1
3 4 3 4 12 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
7 5
P( X | Y ) 
1
2

7 5
From these values we have:
7 12 5 12
H(Y) log  log  0.979868756 bits / symbol .
12 7 12 5
H ( X ,Y )  log 2  log4  1 log12  log6  1.729573958 bits / symbol .

H ( X | Y )  log log  log7  log

1  0.74970520 bits / symbol
6 4 12 2

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54


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

2 2 4 12 6 2 3
I ( X ,Y )  H ( X )  H ( X | Y )  0.061572924 bits / symbol .
 H ( Y )  h  0.061572923 bits / symbol .
C  log n  h  log 2  h  1  h  0.081704167 bits / symbol .
Efficiency ,  I ( X ,Y )  0.753608123 or 75.3608123%
C
Re dundancy , E  1   0.246391876 or 24.6391876%

To find the source probabilities, let p(x1) = p and p(x2) = q = 1 – p .Then the JPM becomes:

2 p, 1p
P( X ,Y )  3 3
1 ( 1  p ), 2
3 3
Adding columnwise we get: p( y1 )  (1  p) and p( y2)  1 (2  p)
1
3 3
For H(Y) = H(Y) max, we want p (y1) = p (y2) and hence 1+p = 2-p or p 1
2
Therefore the source probabilities corresponding to the channel capacity are: p(x1) =1/2 = p(x2).

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

The channel considered in Example 3.6 is called a „Binary Symmetric Channel‟ or ( BSC). It
is one of the most common and widely used channels. The channel diagram of a BSC is shown in Fig
3.4. Here „ p‟ is called the error probability.

For this channel we have:

H (Y | X )  p log 1  q log 1  H ( (4.43)
p)
p q
1 1
H (Y )  [ p  ( p   [q  ( p  q)]log …(4.44)
q)]log
[ p  ( p  q)] [q  ( p  q)]
I(X, Y) = H(Y) – H (Y|X) and the channel capacity is:

C=1 + p log p +q log q …………(4.45)

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

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

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54

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

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

The observer is a noiseless channel that compares the transmitted and the received symbols.
Whenever there is an error a „ 1‟ is sent to the receiver as a correction signal and appropriate
correction is effected. When there is no error the observer transmits a „ 0‟ indicating no change. Thus
the observer supplies additional information to the receiver, thus compensating for the noise in the
channel. Let us compute this additional information .With P (X=0) = P (X=1) = 0.5, we have:

Probability of sending a „1‟ = Probability of error in the channel .

Probability of error = P (Y=1|X=0).P(X=0) + P (Y=0|X=1).P(X=1)

= p × 0.5 + p × 0.5 = p
Probability of no error = 1 – p = q

Thus we have P (Z = 1) = p and P (Z = 0) =q

Accordingly, additional amount of information supplied is:

1 1
 p log  q log  H ( X | Y )  H ( Y | X ) …….. (4.46)
p q

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, ATMECE

 Information Theory and Coding 15EC54

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.

Binary Erasure Channels (BEC):

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

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

1 1
H (Y | X )   qlog ……………… (4.47)
p q
plog
1 1
H ( X )  log  ( 1   )log ……………. (4.48)
 (1)
The JPM is obtained by multiplying first row of P (Y|X) by  and second row by (1– ).
We get:
P( X ,Y )  q p 0
…………….. (4.49)
0
p( 1   ) q( 1   )
Adding column wise we get: P (Y) = [q, p, q (1– )] ……………. (4.50)
From which the CPM P (X|Y) is computed as:

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54

1  0
P( X | Y ) 
……………… (4.51)
0
(1) 1
 H ( X | Y )  q log1  p log  ( 1   ) p log
1 1  ( 1   )q log1
 (1)
 pH ( X )

I(X, Y) =H(X)– H (X|Y) = (1 – p) H(X) = q ………… (4.52)

H(X)
C = Max I (X, Y) = q bits / symbol. …………… (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) =α × p (xk),

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54

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

 p (yj) = ∑ p( xr , y j ) = p (xk, yj) = α p (xk)

r 1

p( xk , y j )
 p( xk | y j )   1, and p( xr | y j )  0 , r  k ,r  1,2 ,3,...m .
p( y j )
It then follows that H (X|Y) = 0 and I (X, Y) = ……... (4.55)
H(X)

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 3.7:
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)=0kj

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, ATMECE

 Information Theory and Coding 15EC54

For such a channel it follows:

H (X|Y) = H (Y|X) = 0 and I(X, Y) = H(X) = H(Y) = H(X, Y) (4.56)

....…

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
p2
P( X ,Y )  ... p2
M
p
m pm ... m

∑ 1
m
with pk  1
i .e . p( yj )  ,  j  1,2 ,3 ,...n.
k 1 n

This means that there is no correlation between xk and yj and an input xk may be received as any
one of the yj‟s with equal probability. In other words, the input-output statistics are independent!!

This can be verified, as we have p (xk, yj) = pk

m

=npk. ∑ pk = p (xk).p (yj)

k 1

 p(xk|yj) = npk and p(yj|xk) = 1/n

Thus we have:
m 1 m 1 m 1 1
H ( X ,Y )  n. ∑ pk log, H ( X )  ∑ npk log n ∑ pk log  log

k 1 p kk 1 npk k 1 pk n
n 1
H ( Y )  ∑ p( y j )log  logn,
j 1p( y j )
H ( X | Y )  H ( X ), H ( Y | X )  H ( Y ) and I ( X ,Y ) 

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54

Such a channel conveys no information whatsoever. Thus a channel with independent input-
output 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 input-
output 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 man‟s 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.

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54
Additional Illustrations:

Example 3.9

Consider two identical BSC„s cascaded as shown in Fig 4.10. Tracing along the transitions indicated
we find:

p (z1|x1) = p2 + q2 = (p + q) 2 – 2pq =(1 – 2pq) = p(z 2|x2) and p(z1|x2) = 2pq = p(z2|x1)

Labeling ˆp  1  2 pq , qˆ  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:
I(X, U) = 1 – H (3pq 2 + p3)

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
2 2 0
1 1 2
P( Y | X )  6 6 3 P( Z | Y )  1 2 0 , with p(x1) = p(x2) =0.5
1 1 1 3 3
2 4 4 1 2
0
3 3

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

DEPT. OF ECE, ATMECE

 Information Theory and Coding 15EC54

5 5 4
P( Z | X )  36 12 9
1 1 1
3 2 6

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

I(X, Z) = 0.079744508 bits / symbol.

Clearly I(X, Y) > I(X, Z).

Example 3.10: Let us consider yet another cascade of noisy channels described by:

2
 P( Z | Y )  0

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:

1 1 1
P( Z | X )  3 3 3  P(Y | X ) ;
0 1 1
2 2

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

DEPT. OF ECE, ATMECE

Information Theory and Coding 15EC54

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, ¼, ½. 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 BCDEFGH
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 ¾ ¼ 0
y ¼ ½ ¼
z 0 ¼ ¾
Determine the rate at which information is being received.

12. Determine the rate of transmission l(x,y) through a channel whose noise characteristics is
shown in fig. P(A1)=0.6, P(A2)=0.3, P(A3)=0.1
A1 0.5 B1

T
R
0.5
A2 0.5 B2

0.5

DEPT. ECE ATMECE

 Information Theory and Coding 15EC54

OUTCOME:

1. Able to understand different Communication channel in communication systems.

2. Capable of finding channel capacity of different channels in communication system.
3. Able to develop channel matrix and mutual information in channel.
Error Control Coding And Binary Cyclic Code 15EC54

MODULE 4
CHAPTER 1: ERROR CONTROL CODING

STRUCTURE
 Rationale for Coding
 Discrete memory less channel
 Shannon‟s theorem on channel capacity Revisited
 Types of errors
 Types of codes
 Example of Error Control Coding
 Block codes
 Minimum Distance Considerations
 Standard Array and Syndrome Decoding
 Hamming Codes

OBJECTIVE
 To analyze different types of erros
 To develop procedures for designing efficient coding schemes for controlling various types of
errors in digital communication system.
 To Study various methods of detecting and/or correcting error and compare different coding
schemes.

1
Error Control Coding And Binary Cyclic Code 15EC54

INTRODUCTION

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.

4.1 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 2 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 Huffman‟s procedure is:

s1……… 0 s3…… 1 1 0
s2……… 10 s4…… 1 1 1

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 “s1s2s4s1”, which is altogether different than the transmitted sequence. Thus
although the coding provides 100% efficiency in the light of Shannon‟s theorem, it suffers a major
disadvantage. Another disadvantage of a „variable length‟ code lies in the fact that output data rates

2
Error Control Coding And Binary Cyclic Code 15EC54

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.
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. 4.1

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.

3
Error Control Coding And Binary Cyclic Code 15EC54

Fig 4.1: Block diagram of a typical data transmission

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

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

4
Error Control Coding And Binary Cyclic Code 15EC54

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.

4.2 Discrete memory less channel:

Referring to the block diagram in Fig. 4.2 the channel is said to be memory less if the de-
modulator (Detector) output in a given interval depends only on the signal transmitted in the
interval, and not on any previous transmission. Under this condition, we may model (describe) the
combination of the modulator – channel – and the demodulator as a “Discrete memory less
channel”. Such a channel is completely described by the set of transition probabilities p (yj | xk)
where xk is the modulator input symbol.

The simplest channel results from the use of binary symbols (both as input and output). When
binary coding us used the modulator has only „0‟`s and „1‟`s as inputs. Similarly, the inputs to the
demodulator also consists of „0‟`s and „1‟`s provided binary quantization is used. If so we say a
„Hard decision‟ is made on the demodulator output so as to identify which symbol was actually
transmitted. In this case we have a „Binary symmetric channel‟ (BSC). The BSC when derived
from an additive white Gaussian noise (AWGN) channel is completely described by the transition
probability „p‟. The majority of coded digital communication systems employ binary coding with
hard-decision decoding due to simplicity of implementation offered by such an approach.

The use of hard-decisions prior to decoding causes an irreversible loss of information in the
receiver. To overcome this problem “soft-decision” coding is used. This can be done by including a
multilevel quantizer at the demodulator output as shown in Fig. 4.2(a) for the case of binary PSK
signals. The input-output characteristics and the channel transitions are shown in Fig. 4.2(b) and
Fig. 4.2(c) respectively. Here the input to the demodulator has only two symbols „0‟`s and „1‟`s.
However, the demodulator output has „Q‟ symbols. Such a channel is called a “Binary input-Q-ary
output DMC”. The form of channel transitions and hence the performance of the demodulator,
depends on the location of representation levels of the quantizer, which inturn depends on the signal
level and variance of noise. Therefore, the demodulator must incorporate automatic gain control, if an
effective multilevel quantizer is to be realized. Further the soft-decision decoding offers significant
improvement in performance over hard-decision decoding.

5
Error Control Coding And Binary Cyclic Code 15EC54

Fig. 4.2 (a) - Reciever

Fig 4.2 (b) Transfer characteristics (c) Channel Diagram

4.3 Shannon‟s theorem on channel capacity Revisited:

The “Shannon‟s theorem on channel capacity” is re-stated here and call it the “Coding
Theorem”.

“It is possible in principle, to devise a means where by a communication system will transmit
information with an arbitrarily small probability of error, provided the information rate R (=r I(X,Y)
where r-is the symbol rate) is less than or equal to a rate „C‟ called the „channel capacity”. The
technique used to achieve this goal is called “Coding”. For the special case of a BSC, the theorem
tells us that if the code rate, Rc (defined later) is less than the channel capacity, then it is possible to
find a code that achieves error free transmission over the channel. Conversely, it is not possible to
find such a code if the code rate Rc is greater than C.
The channel coding theorem thus specifies the channel capacity as a “Fundamental limit” on
the rate at which reliable transmission (error-free transmission) can take place over a DMC. Clearly,
the issue that matters is not the signal to noise ratio (SNR), so long as it is large enough, but how the
input is encoded.

The most un-satisfactory feature of Shannon‟s theorem is that it stresses only about the
“existence of good codes”. But it does not tell us how to find them. So, we are still faced with the

6
Error Control Coding And Binary Cyclic Code 15EC54

task of finding a good code that ensures error-free transmission. The error-control coding techniques
presented in this and subsequent chapters provide different methods of achieving this important
system requirement.

4.4 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 by 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

4.5 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
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 τb and τc are the bit durations in the encoded and coded
words, i.e. the input and output code words, then it is necessary that
n.τc = k.τb

We define the “rate of the code” by (also called rate efficiency)

7
Error Control Coding And Binary Cyclic Code 15EC54

k
Rc 
n
fb c k
Accordingly, with f b  1 and f c  1 , we have    Rc
b c fc b n

4.6 Example of Error Control Coding:

Better way to understand the important aspects of error control coding is by way of an
example. Suppose that we wish transmit data over a telephone link that has a useable bandwidth of 4
KHZ and a maximum SNR at the out put of 12 dB, at a rate of 1200 bits/sec with a probability of
error less than 10-3. Further, we have DPSK modem that can operate at speeds of 1200, 1400 and
3600 bits/sec with error probabilities 2(10-3), 4(10-3) and 8(10-3) respectively. We are asked to
design an error control coding scheme that would yield an overall probability of error < 10-3. We
have:

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

S S
[C=Blog2 (1+ ).  12dB or 15.85 , B=4KHZ], p = 2(10-3), 4(10-3) and 8(10-3) respectively.
N N
Since Rc < C, according to Shannon‟s theorem, we should be able to transmit data with arbitrarily
small probability of error. We shall consider two coding schemes for this problem.

(i) Error detection: Single parity check-coding. Consider the (4, 3) even parity check code.

Message 000 001 010 011 100 101 110 111

Parity 0 1 1 0 1 0 0 1

Codeword 0000 0011 0101 0110 1001 1010 1100 1111

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

This code is capable of „detecting‟ all single and triple error patterns. Data comes out of the channel
encoder at a rate of 3600 bits/sec and at this rate the modem has an error probability of 8(10-3). 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.

Using binomial probability law, we have with p = 8(10-3):

8
Error Control Coding And Binary Cyclic Code 15EC54

 n
P(X = k) =   p k ( 1  p )n  k
k

4 4 4 4

pd    p( 1  p )3    p 3 ( 1  p ),    4C1  4 ,    4C 3  4
 1  3  1  3

Expanding we get pd  4 p  12 p 2  16 p 3  8 p4

Substituting the value of p we get:

pd = 32 (10-3) - 768 (10-6) +8192 (10-9) – 32768 (10-12) = 0.031240326 > > (10-3)

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 4
pnd  P ( X  2 )  P ( X  4 )    p 2 ( 1  p )2    p 4 ( 1  p )0  6 p 2  12 p 3  7 p 4
 2 4

Substituting the value of p we get pnd=0.410-3  10-3

Thus probability of error is less than 10-3 as required.

(ii) 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 000 001 010 100 011 101 110 111

Triplet

Output 0 0 0 0 1 1 1 1
message

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

 3  3 3
=   p2 (1-p) +   p (1-p) 0 =3p2-2p3
 2  3

9
Error Control Coding And Binary Cyclic Code 15EC54

=190.464 x 10-6=0.19x 10-3 < p= 10-3

Probability of no detection, pnd =P (All 3 bits in error) = p3 =512 x 10-9 < < 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
both examples with out error control coding the probability of error =8(10-3) 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 bonafied 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.

4.7 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 to 2k information blocks possible, we would then have 2k code words of length n > k.
This set of 2k code words is called a “Block code”. For a block code to be useful these 2k 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 vector‟ and „code vector‟ respectively. Notice that encoding equipment
must be capable of storing the 2k 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‟.

4.7.1 Linear Block Codes:

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

Fields with 2m symbols are called „Galois Fields‟ (pronounced as Galva fields), GF (2m).Their
arithmetic involves binary additions and subtractions. For two valued variables, (0, 1).The modulo –
2 addition and multiplication is defined in Fig 4.3.

10
Error Control Coding And Binary Cyclic Code 15EC54

Fig 4.3

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. The arithmetic of GF(2m) 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 (i.e. it does not have a factor of degree  m and >0, for example X3 + X2 + 1, X3 + X + 1,
X4 +X3 +1, X5 +X2 +1 etc. are irreducible polynomials, whereas f(X)=X4+X3+X2+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 4.1: 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.

11
Error Control Coding And Binary Cyclic Code 15EC54

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.

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 g1u2 g2 … uk gk ……………… (4.1)

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) …………………… (4.2)

Thus we have v=uG …………………… (4.3)

Where: u = (u1, u2…uk) …………………… (4.4)

represents the data vector and

 g11 g12  g1n 

 g1   
   g 21 g 22  g 2 n 
G   g2  …………………… (4.5)
  
 g 3   
 gk 1 gk 2  gkn 
is called the “generator matrix”.

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

Example 4.2: The (6, 3) linear code of Example 6.1 has the following generator matrix:

 g1  1 0 0 0 1 1
 
G   g2   0 1 0 1 0 1
 g3  0 0 1 1 1 0 

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

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

= (0,0,0,0,0,0) + (0,1,0,1,0,1) + (0,0,1,1,1,0) = (0, 1, 1, 0, 1, 1)

Thus v = (0 1 1 0 1 1)

12
Error Control Coding And Binary Cyclic Code 15EC54

“v can be computed simply by adding those rows of G which correspond to the locations of
1`s of u.”

4.7.2 Systematic Block Codes (Group Property):

A desirable property of linear block codes is the “Systematic Structure”. Here a code word is
divided into two parts –Message part and the redundant part. If either the first k digits or the last k
digits of the code word correspond to the message part then we say that the code is a “Systematic
Block Code”. We shall consider systematic codes as depicted in Fig.4.5.

Fig 4.5 Systematic format of code word

In the format of Fig.4.5 notice that:

v1 = u1, v2 = u2, v3 = u3 … vk = uk …………… (4.6 a)

vk  1  u1 p11  u2 p21  u3 p31    uk pk1 

vk  2  u1 p12  u2 p22  u3 p32    uk pk2 
 ……………… (4.6 b)
  
vn  u1 p1,n - k  u2 p2,n - k  u3 p3,n - k   ..  uk pk, n - k 


Or in matrix from we have

v1 v2 ... v k vk  1 vk  2 ... v n  

 1 0 0 ... 0 p11 p12 ... p1 ,n  k 
0 1 0 ... p2 ,n  k  …….
0 p21 p22 ... (4.7)
u1 u2 ... uk  
          
 
0 0 0 ... 1 pk 1 pk 2 ... pk ,n  k 

i.e., v = u.G

Where G = [Ik, P] ………………. (4.8)

 p11 p12  p1 ,n  k 
 
p21 p22  p2 ,n  k
Where P =   ………………. (4.9)
    
 
 pk , 1 pk ,2  pk ,n  k 

13
Error Control Coding And Binary Cyclic Code 15EC54

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 (4.6b) are
referred to as parity check equations. Observe that the G matrix of Example 4.2 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:

a.b = (a1, a2…an) (b1, b2 …bn) T = 0.

where, „T‟ represents transposition. Accordingly for any kn matrix, G, with k linearly independent
rows there exists a (n-k)  n matrix H with (n-k) linearly independent rows such that any vector in
the row space of G is orthogonal to the rows of H and that any vector that is orthogonal to the rows
of 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.HT = O”. ……… (4.9a)
(O represents an all zero row vector.)

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.
 p11 p21 ... pk 1 1 0 0 ... 0 
 p p22 ... pk 2 0 1 0 ... 0 
T
H = [P .In-k] =  12
……… (4.10)
          
 
 p1 ,n  k p2 ,n  k ... pk ,n  k 0 0 0 ... 1

The ith row of G is:

gi = (0 0 …1 …0…0 pi,1 pi,2…pi,j…pi, n-k)

 
th
i element (k + j) th element

The jth row of H is:

i th element (k + j) th element
 
hj = ( p1,j p2,j …pi,j ...pk, j 0 0 … 0 1 0 …0)

Accordingly the inner product of the above n – vectors is:

gihj = (0 0 …1 …0…0 pi,1 pi,2…pi,j…pi, n-k) ( p1,j p2,j …pi,j ...pk, j 0 0 … 0 1 0 …0)T
   

14
Error Control Coding And Binary Cyclic Code 15EC54

ith element (k + j) th element ith element (k + j) th element

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

This implies simply that:

G. HT = Ok (n – k) …………………………. (4.11)

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

Further, since the (n – k) rows of the matrix H are linearly independent, the H matrix of
Eq. (4.10) is a parity check matrix of the (n, k) linear systematic code generated by G. Notice that the
parity check equations of Eq. (4.6b) can also be obtained from the parity check matrix using the fact
v.HT = O.

Alternative Method of proving v.HT = O.:

We have v = u.G = u. [Ik: P]= [u1, u2… uk, p1, p2 …. Pn-k]

Where pi =( u1 p1,i + u2 p2,i + u3 p3,i …+ uk pk, i) are the parity bits found from Eq (4.6b).

 P 
Now H T   
 I n k 
 v.HT = [u1 p11 + u2 p21 +…. + …. + uk pk1 + p1, u1 p12 + u2 p22 + ….. + uk pk2 + p2, …
u1 p1, n-k + u2 p2, n-k + …. + uk pk, n-k + pn-k]

= [p1 + p1, p2 + p2… pn-k + pn-k]

15
Error Control Coding And Binary Cyclic Code 15EC54

= [0, 0… 0]

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

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

If this is to be true for any arbitrary message vector v then this implies: G HT = Ok (n – k)

Example 4.3:

Consider the generator matrix of Example 4.2, the corresponding parity check matrix is

0 1 1 1 0 0 
H = 1 0 1 0 1 0 
1 1 0 0 0 1 

4.7.3 Circuit implementation of Block codes:

The implementation of Block codes is very simple. We need only combinational logic circuits.
Implementation of Eq (4.6) is shown in the encoding circuit of Fig.4.6. Notice that pij is either a „0‟
or a „1‟ and accordingly  pij  indicates a connection if pij = 1 only (otherwise no connection).
The encoding operation is very simple. The message u = (u1, u2 … uk) to be encoded is shifted into
the message register and simultaneously into the channel via the commutator. As soon as the entire
message has entered the message register, the parity check digits are formed using modulo -2 adders,
which may be serialized using, another shift register – the parity register, and shifted into the channel.
Notice that the complexity of the encoding circuit is directly proportional to the block length of the
code. The encoding circuit for the (6, 3) block code of Example 2 is shown in Fig 4.7

16
Error Control Coding And Binary Cyclic Code 15EC54

Fig 4.6 Encoding circuit for systematic block code

Fig 4.7 Encoder for the (6,3) block code of example 4.2

4.7.4 Syndrome and Error Detection:

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

e = r – v = (e1, e2… en) …………………… (4.12)

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 1‟s in e are the transmission errors caused by the channel noise. Hence from
Eq (4.12) it follows:

r=ve ………………………………. (4.12a)

Observer 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

17
Error Control Coding And Binary Cyclic Code 15EC54

errors and correct them (FEC – Forward Error Correction) or make a request for re–transmission
(ARQ). When r is received, the decoder computes the following (n-k) tuple:

s = r. HT …………………….. (4.13)
= (s1, s2… sn-k)

It then follows from Eq (4.9a), 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
pattern”. Since there are 2k -1 nonzero code words, it follows that there are 2k -1 error patterns as
well. Hence when an undetectable error pattern occurs the decoder makes a “decoding error”.
Eq. (4.13) can be expanded as below:

s1  r1 p11  r2 p21  ....  rk pk 1  r k  1 


s2  r1 p12  r2 p22  ....  rk pk 2  rk  2 
From which we have  ………… (4.14)
     
sn  k  r1 p1, n  k  r2 p2, n  k  ....  rk pk , n  k  rn 

A careful examination of Eq. (4.14) reveals the following point. The syndrome is simply the vector
sum of the received parity digits (rk+1, rk+2 ...rn) and the parity check digits recomputed from the
received information digits (r1, r2 … rn). Thus, we can form the syndrome by a circuit exactly similar
to that of Fig.6.6 and a general syndrome circuit is as shown in Fig. 4.8.

Example 4.4:
We shall compute the syndrome for the (6, 3) systematic code of Example 4.2. We have

0 1 1 
1 0 1 
 
s = (s1, s2, s3) = (r1, r2, r3, r4, r5, r6) 1 1 0 
 
1 0 0 
0 1 0 
 
0
 0 1 


18
Error Control Coding And Binary Cyclic Code 15EC54

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

The syndrome circuit for this code is given in Fig.4.9.

Fig 4.8 Syndrome circuit for the (n,k) Linear systematic block code

Fig 4.8 Syndrome circuit for the (6,3) systematic block code

In view of Eq. (4.12a), and Eq. (4.9a) we have

s = r.HT = (v e) HT

= v .HT  e.HT

19
Error Control Coding And Binary Cyclic Code 15EC54

or s = e.HT …………… (4.15)

as v.HT= O. Eq. (4.15) 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 = e1 p11 + e2 p21 + …. + ek pk, 1 + ek + 1 

s2 = e1 p12 + e2 p22 + … + ek pk, 2 + ek + 2 
 …………… (4.16)
     
sn - k = e1 p1, n - k + e2 p2, n - k + ….. + ek pk, n - k + en 

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. (4.16) 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
H matrix is k, it follows that there are 2k non-trivial solutions. In other words there exist 2k error
patterns that result in the same syndrome. Therefore to determine the true error pattern is not any easy
task

Example 4.5:

For the (6, 3) code considered in Example 4.2, 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)

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

Then it follows that:

e2 + e3 +e4 = 1
e1 + e3 +e5 =1
e1 + e2 +e6 = 0

There are 23 = 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 (4.16) 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 4.5, notice that the error vector (0 0 1 0 0 0) has

20
Error Control Coding And Binary Cyclic Code 15EC54

the smallest number of nonzero components and hence can be regarded as the most probable error
vector. Then using Eq. (4.12) we have

v̂ = r  e

= (0 1 1 1 0 1) + (0 0 1 0 0 0) = (0 1 0 1 0 1)

Notice now that v̂ indeed is the actual transmitted code word.

4.8 Minimum Distance Considerations:

The concept of distance between code words and single error correcting codes was first
developed by R .W. Hamming. Let the n-tuples,

 = (1, 2 … n),  = (1, 2 … n)

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

n
d( , )   ( j   j ) ……………………. (4.17)
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 4.6: 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 (4.17) we find

d (,) = (0  1) + (1  0) + (1  1) + (1  0) + (0  1) + (1  1)

= 1 + 1 + 0 + 1 + 1 + 0

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

Further, () = 4 and () = 4.

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

21
Error Control Coding And Binary Cyclic Code 15EC54

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(,) ……………. (4.18)

From the discussions made above, we may write

d (,) =  (  ) …………………. (4.19)

Example 4.7: For the vectors  and  of Example 4.6, we have:

   = (01), (10), (11) (10), (01) (11)= (11 0 1 1 0)

 ( ) = 4 = d (,)

If  = (1 0 1 01 0), we have d (,) = 4; d (,) = 1; d (,) = 5

Notice that the above three distances satisfy the triangle inequality:

d (,) + d (,) = 5 = d (,)

d (,) + d (,) = 6 > d (,)

d (,) + d (,) = 9 > d (,)

Similarly, the minimum distance of a linear block code, „C‟ may be mathematically
represented as below:

dmin =Min {d (,):,  C,   } …………. (4.20)

=Min {( ):,  C,   }

=Min {(v), v  C, v  0} ………………. (4.21)

That is d min  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. (4.9a) we have:

22
Error Control Coding And Binary Cyclic Code 15EC54

0 = v.HT

= v1h1  v2h2  ….  vnhn

Here h1, h2 … hn represent the columns of the H matrix. Let vj1, vj2 …vjl be the „l‟ nonzero
components of v i.e. vj1 = vj2 = …. vjl = 1. Then it follows that:

hj1 hj2 …  hjl = OT …………………. (4.22)

That is “if v is a code vector of Hamming weight „l‟, then there 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 … xn) whose nonzero components are xj1, xj2 … xjl. Consider the
product:

x.HT = x1h1  x2h2 …. xnhn = xj1hj1  xj2hj2  ….  xjlhjl = hj1  hj2  …  hjl

If Eq. (4.22) holds, it follows x.HT = 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:

i) If no (d-1) or fewer columns of H add to OT, 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.

0 1 1 1 0 0 
For the H matrix of Example 6.3, i.e. H = 1 0 1 0 1 0  , notice that all columns of H are non
1 1 0 0 0 1
zero and distinct. Hence no two or fewer columns sum to zero vector. Hence the minimum weight of
the code is at least 3.Further notice that the 1st, 2nd and 3rd columns sum to OT. 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 4.1.

4.8.1 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.4.10.

23
Error Control Coding And Binary Cyclic Code 15EC54

Fig 4.10 The distance concept

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.

Further, assume that 2k 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) …………………. (4.23)

dmin is either odd or even. Let „t‟ be a positive integer such that

2t + 1  dmin  2t + 2 ………………… (4.24)

Suppose  be any other code word of the code. Then, the Hamming distances among
, and  satisfy the triangular inequality:

d(,) + d(, )  d(,) ………………… (4.25)

24
Error Control Coding And Binary Cyclic Code 15EC54

Suppose an error pattern of „t ‟ errors occurs during transmission of . Then the received vector 
differs from  in „t ‟ places and hence d(,) = t. Since  and  are code vectors, it follows from
Eq. (6.24).

d(,)  dmin  2t + 1 ………………. (4.26)

Combining Eq. (4.25) and (4.26) and with the fact that d(,) = t, it follows that:

d (,  )  2t + 1- t ……………… (4.27)

Hence if t t, then: d (, ) > t ……………… (4.28)

Eq 4.28 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 d (,) = dmin, and let e1 and e2 be two error patterns such that:

i) e1  e2 =   

ii) e1 and e2 do not have nonzero components in common places. Clearly,

(e1) + (e2) = ( ) = d( ,) = dmin ………………… (4.29)

Suppose,  is the transmitted code vector and is corrupted by the error pattern e1. Then the received
vector is:

 =   e1 ……………………….. (4.30)

and d (,) = (   ) = (e1) ……………………… (4.31)

d (,) = ()

= (    e1) = (e2) ………………………. (4.32)

25
Error Control Coding And Binary Cyclic Code 15EC54

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 ………………………… (4.33)

 d (,)  d (,) …………………………. (4.34)

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.4.11 (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.

Fig. 4.11(a)

On the other hand, if d (,)  2t, the two spheres around  and  intersect and if „‟ is located as in
Fig. 4.11(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

1 
t   ( d min  1 ) …………………………. (4.35)
2 

26
Error Control Coding And Binary Cyclic Code 15EC54

1  1 
where  ( d min  1 ) denotes the largest integer no greater than the number  ( d min  1 ) . The
2  2 
1 
parameter„t‟ =  ( d min  1 ) is called the “random-error-correcting capability” of the code and
2 
the code is referred to as a “t-error correcting code”. The (6, 3) code of Example 4.1 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.

For an (n, k) linear code, observe that, there are 2n-k syndromes including the all zero
syndrome. Each syndrome corresponds to a specific error pattern. If „j‟ is the number of error
 n
locations in the n-dimensional error pattern e, we find in general, there are    nC j multiple error
 j
t
 n
patterns. It then follows that the total number of all possible error patterns =    , where„t‟ is the
j  0 j 

maximum number of error locations in e. Thus we arrive at an important conclusion. “If an (n, k)
linear block code is to be capable of correcting up to„t‟ errors, the total number of syndromes
shall not be less than the total number of all possible error patterns”, i.e.

t
 n
2 n-k
   j  ………………………. (4.36)
j  0 

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”.

.9 Standard Array and Syndrome Decoding:

The decoding strategy we are going to discuss is based on an important property of the
syndrome.

Suppose vj , j = 1, 2… 2k, be the 2k distinct code vectors of an (n, k) linear block code.
Correspondingly let, for any error pattern e, the 2k distinct error vectors, ej, be defined by

ej = e  vj , j = 1, 2… 2k ………………………. (4.37)

The set of vectors {ej, j = 1, 2 … 2k} so defined is called the “co- set” of the code. That is, a
„co-set‟ contains exactly 2k elements that differ at most by a code vector. It then fallows that there are
2n-k co- sets for an (n, k) linear block code. Post multiplying Eq (4.37) by HT, we find

ej HT = eHT  vj HT

= e HT ………………………………………………. (4.38)

27
Error Control Coding And Binary Cyclic Code 15EC54

Notice that the RHS of Eq (4.38) is independent of the index j, as for any code word the term
vj H = 0. From Eq (4.38) it is clear that “all error patterns that differ at most by a code word
T

have the same syndrome”. That is, each co-set is characterized by a unique syndrome.

Since the received vector r may be any of the 2n n-tuples, no matter what the transmitted code
word was, observe that we can use Eq (4.38) to partition the received code words into 2k 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:

Step1: Place the 2k code vectors of the code in a row, with the all zero vector
v1 = (0, 0, 0… 0) = O as the first (left most) element.

Step 2: From among the remaining (2n – 2k) - n – tuples, e2 is chosen and placed below the
all-zero vector, v1. The second row can now be formed by placing (e2  vj),
j = 2, 3… 2k under vj

Step 3: Now take an un-used n-tuple e3 and complete the 3rd row as in step 2.

Step 4: continue the process until all the n-tuples are used.

The resultant array is shown in Fig. 4.12.

Fig 4.12: Standard Array for an (n,k) linear block code

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 slandered
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 appearance in one and only one row”.

28
Error Control Coding And Binary Cyclic Code 15EC54

From the above discussions it is clear that there are 2n-k disjoint rows or co-sets in the standard
array and each row or co-set consists of 2k distinct entries. The first n-tuple of each co-set, (i.e., the
entry in the first column) is called the “Co-set leader”. Notice that any element of the co-set can be
used as a co-set leader and this does not change the element of the co-set - it results simply in a
permutation.

Suppose DjT is the jth column of the standard array. Then it follows

Dj = {vj, e2  vj, e3  vj… e2n-k  vj} ………………….. (4.39)

where vj is a code vector and e2, e3, … e2n-k are the co-set leaders.

The 2k disjoints columns D1T, D2T… D2 k can now be used for decoding of the code. If vj is
T

the transmitted code word over a noisy channel, it follows from Eq (5.39) that the received vector r is
in DjT 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 ê which is not a co-set
leader must be in some co-set and under some nonzero code vector is, say, in the i-th co-set and under
v  0. Then it follows

ê = ei  vl , and the received vector is r = vj  ê = vj  (ei  vl ) = ei  vm

Thus the received vector is in DmT and it will be decoded as vm and a decoding error has been
committed. Hence it is explicitly clear that “Correct decoding is possible if and only if the error
pattern caused by the channel is a co-set leader”. Accordingly, the 2n-k co-set leaders (including
the all zero vector) are called the “Correctable error patterns”, and it follows “Every (n, k) linear
block code is capable of correcting 2n-k error patterns”.

So, from the above discussion, it follows that in order to minimize the probability of a
decoding error, “The most likely to occur” error patterns should be chosen as co-set leaders. For a
BSC an error pattern of smallest weight is more probable than that of a larger weight. Accordingly,
when forming a standard array, error patterns of smallest weight should be chosen as co-set leaders.
Then the decoding based on the standard array would be the „minimum distance decoding‟ (the
maximum likelihood decoding). This can be demonstrated as below.

Suppose a received vector r is found in the jth column and lth row of the array. Then r will be
decoded as vj. We have

d(r, vj) = (r  vj ) = (el  vj  vj ) = (el )

where we have assumed vj indeed is the transmitted code word. Let vs be any other code word, other
than vj. Then

d(r, vs ) = (r vs ) = (el  vj  vs ) (el ) = (el  vi )

29
Error Control Coding And Binary Cyclic Code 15EC54

as vj and vs are code words, vi = vj  vs is also a code word of the code. Since el and (el  vi ) are in
the same co set and, that el has been chosen as the co-set leader and has the smallest weight it follows
(el )  (el  vi ) and hence d(r, vj )  d(r, vs ). Thus the received vector is decoded into a closet
code vector. Hence, if each co-set leader is chosen to have minimum weight in its co-set, the standard
array decoding results in the minimum distance decoding or maximum likely hood decoding.

Suppose “a0, a1, a2 …, an” denote the number of co-set leaders with weights 0, 1, 2… n. This
set of numbers is called the “Weight distribution” of the co-set leaders. Since a decoding error will
occur if and only if the error pattern is not a co-set leader, the probability of a decoding error for a
BSC with error probability (transition probability) p is given by
n
P ( E )  1   a j p j ( 1  p )n  j …………………… (4.40)
j 0

Example 4.8:

For the (6, 3) linear block code of Example 4.1 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 (5.40) it then follows:

P (E) = 1- [(1-p) 6 +6p (1-p) 5 + p2 (1-p) 4]

With p = 10-2, we have P (E) = 1.3643879  10-3

A received vector (010 001) will be decoded as (010101) and a received vector (100 110) will be
decoded as (110 110).

Notice that an (n, k) linear code is capable of detecting (2n -2k) error patterns while it is
capable of correcting only 2n-k error patterns. Further, as n becomes large 2n-k/ (2n-2k) becomes

30
Error Control Coding And Binary Cyclic Code 15EC54

smaller and hence the probability of a decoding error will be much higher than the probability of an
undetected error.

Let us turn our attention to Eq (5.35) and arrive at an interpretation. Let x1and x2 be two n-
tuples of weights„t‟ or less. Then it follows

 (x1  x2)  (x1) + (x2)  2t  dm-n

Suppose x1 and x2 are in the same co-set then it follows that (x1  x2) must be a nonzero code
vector of the code. This is impossible because the weight of (x1  x2) is less than the minimum
weight of the code. Therefore, “No two n-tuples, whose weights are less than or equal to„t‟, can
be in the same co-set of the code and all such n-tuples can be used as co-set leaders”.

Further, if v is a minimum weight code vector, i.e. (v) = dmin and if the n-tuples, x1 and x2
satisfy the following two conditions:

i) x1  x2 = v

ii) x1 and x2 do not have nonzero components in common places

It follows from the definition, x1 and x2 must be in the same co-set and

 (x1) + (x2) = (v) = dmin

Suppose we choose x2 such that (x2) = t + 1. Since 2t+1 dmin  2t+2, we have  (x1) = t or (t+1).
If x1 is used as a co-set leader then x2 cannot be a co-set leader.

The above discussions may be summarized by saying “For an (n , k) linear block code with
1 
minimum distance dmin, all n-tuples of weight t   ( d min  1 ) can be used as co-set leaders of
2 
a standard array. Further, if all n-tuples of weight  t are used as co-set leaders, there is at least
one n-tuple of weight (t + 1) that cannot be used as a co-set leader”.

These discussions once again re-confirm the fact that an (n, k) linear code is capable of
1 
correcting error patterns of  2 ( d min  1 ) or fewer errors but is incapable of correcting all the
 
error patterns of weight (t + 1).

We have seen in Eq. (4.38) 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.
These relationships, then, can be used in preparing a decoding table that is made up of 2n-k 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.HT

31
Error Control Coding And Binary Cyclic Code 15EC54

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. However, as (n-k) becomes large the implementation becomes difficult and
impractical as either a large storage or a complicated logic circuitry will be required.

For implementation of the decoding scheme, one may regard the decoding table as the truth
table of n-switching functions:

e1 = f1 (s1, s2... sn-k); e2 = f2 (s1, s2... sn-k); … en = fn (s1, s2... sn-k)

where s1, s2... sn-k are the syndrome digits and are regarded as the switching variables and e1, e2 … en
are the estimated error digits. The stages can be released by using suitable combinatorial logic
circuits as indicated in Fig 4.13.

Fig. 4.13 General Decoding scheme for an (n,k) linear block code

Example 4.9:

From the standard array for the (6, 3) linear block code of Example 4.8, the following truth table can
be constructed.

32
Error Control Coding And Binary Cyclic Code 15EC54

The two shaded portions of the truth table are to be observed carefully. The top shaded
one corresponds to the all-zero error pattern and the bottom one corresponds to a double error patter
which cannot be corrected by this code. From the table we can now write expressions for the
correctable single error patterns as below:

e1  s1 .s2 s3 e2  s1 s2 s3 e3  s1 s2 s3
e4  s1 s2 s3 e5  s1 s2 s3 e6  s1 s2 s3

The implementation of the decoder is shown in Fig.4.14.

Fig 4.14: Decoding circuit for (6,3) code

33
Error Control Coding And Binary Cyclic Code 15EC54

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
withthe4th column of the H- matrix and hence the 4th – position of the received vector is in error.
Hence the corrected vector is 010101. Similarly, for a received vector (100110), the syndrome is 101
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 r6.).

Notice that for the systematic encoding considered by us (r1 r2 r3) corresponds to the data digits and
(r4 r5 r6) are the parity digits.

Accordingly the correction for the data digits would be

v̂1 = r1 + (s2. s3), v̂ 2 = r2 + (s1. s3), v̂ 3 = r3 + (s1. s2)

Hence the circuit of Fig 6.14 can be modified to have data recovery by removing only the
connections of the outputs vˆ4 , vˆ5 and vˆ6 .

4.10 Hamming Codes:

Hamming code is the first class of linear block codes devised for error correction. The single error
correcting (SEC) Hamming codes are characterized by the following parameters.

Code length: n = (2m-1)

Number of Information symbols: k = (2m – m – 1)

Number of parity check symbols :( n – k) = m

34
Error Control Coding And Binary Cyclic Code 15EC54

Error correcting capability: t = 1, (dmin= 3)

The parity check matrix H of this code consists of all the non-zero m-tuples as its columns. In
systematic form, the columns of H are arranged as follows

H = [Q  Im]

Where Im is an identity (unit) matrix of order m  m and Q matrix consists of

(2m-m-1) columns which are the m-tuples of weight 2 or more. As an illustration for k=4 we have
from k = 2m – m – 1.

m=1 k=0, m=2 k=1, m=3 k=4

Thus we require 3 parity check symbols and the length of the code 23 – 1 = 7. This results in the
(7, 4) Hamming code.

The parity check matrix for the (7, 4) linear systematic Hamming code is then

The generator matrix of the code can be written in the form


G  I 2 m  m  1  QT 
And for the (7, 4) systematic code it follows:

A non systematic Hamming code can be constructed by placing the parity check bits at 2l, l=0, 1,
2…locations. It was the conventional method of construction in switching and computer applications
(Refer, for example „Switching circuits and applications -Marcus).One simple procedure for
construction of such code is as follows:

Step 1: Write the BCD of length (n – k) for decimals from 1 to n.

Step 2: Arrange the sequences in the reverse order in a matrix form.

Step 3: Transpose of the matrix obtained in step 2 gives the parity check matrix H for the code.

35
Error Control Coding And Binary Cyclic Code 15EC54

The code words are in the form

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

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.

Step1:

1 0 0 
0 1 0 
 
1 1 0 
 
Step2: HT =  0 0 1 
1 0 1 
 
0 1 1 
1 1 1 
 

 10 10 10 1 
Step3: H=  0 1 1 0 0 1 1 
 0 0 0 1 1 1 1 

Notice that the parity check bits, from he above H matrix apply to positions.

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…

p4 = 8, 9, 10, 11, 12, 13, 14, 15 and so on

Accordingly, the check bits can be represented as linear combinations of the message bits.
For the (7, 4) code under consideration we have

36
Error Control Coding And Binary Cyclic Code 15EC54

p1 = m1 + m2 +m4

p2 = m1 + m3 +m4

p3 = m2 + m3 + m4

Accordingly, the generator matrix can written as

Notice that the message bits are located at the positions other than 2l, l = 0, 1, 2, 3….
locations. i.e., they are located in the positions of 3, 5, 7, 9, 11, 13, 15, 17, 18….. The k- columns of
the identity matrix Ik are distributed successively to these locations. The Q sub-matrix in the H
matrix can be identified to contain those columns which have weights more than one. The transpose
of this matrix then gives the columns to be filled, in succession, in the G- matrix. For the Example of
the (7, 4) linear code considered, the Q- sub-matrix is:

1 1 0
1 1 0 1 1
1 0 1 1  0 1
Q=   , and hence QT =
0 1 1
0 1 1 1  
1 1 1

The first two columns of this matrix then are the first two columns of the G: matrix and the third
column is the Forth column of the G matrix. Table below gives the codes generated by this method.

37
Error Control Coding And Binary Cyclic Code 15EC54

Observe that the procedure outlined for the code construction starts from selecting the H
matrix which is unique and hence the codes are also unique. We shall consider the correctable error
patterns and the corresponding syndromes listed in the table below.

Table: Error patterns and syndromes for the (7, 4) linear non-systematic code

Error Pattern Syndrome

e1 e2 e3 e4 e5 e6 s1 s2 s3
e
1 0 0 0 0 0 1 0 0
0
0 1 0 0 0 0 0 1 0
0
0 0 1 0 0 0 1 1 0
0
0 0 0 1 0 0 0 0 1
0
0 0 0 0 1 0 1 0 1
0
0 0 0 0 0 1 0 1 1
0
0 0 0 0 0 0 1 1 1
1

38
Error Control Coding And Binary Cyclic Code 15EC54

If the syndrome is read from right to left i.e. if the sequence is arranged as „s3 s2 s1‟ it is interesting to
observe that the decimal equivalent of this binary sequence corresponds to the error location. Thus if
the code vector 1 0 1 1 0 1 0 is received as 1 0 1 0 0 1 0, the corresponding syndrome is „0 0 1‟,
th
which is exactly the same as the 4 column of the H-matrix and also the sequence 100 corresponds to
decimal 4.

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. (4.36)
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. (4.35)

We may delete any „l‟columns from the parity check matrix H of the Hamming code resulting
in the reduction of the dimension of H matrix to m  (2m-l-1).Using this new matrix as the parity
check matrix we obtain a “shortened” Hamming code with the following parameters.

Code length: n = 2m-l-1

Number of Information symbols: k=2m-m-l-1

Number of parity check symbols: n–k=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
obtain an m  2m-1 matrix


H  Q : Im 
Where Q contains (2m-1 -m) columns of odd weight. Clearly no three columns add to zero as all
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 one‟s 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.

39
Error Control Coding And Binary Cyclic Code 15EC54

(3) If s contains even number of one‟s an uncorrectable error pattern has been detected.

Alternatively the SEC Hamming codes may be made to detect double errors by adding an extra
parity check in its (n+1) Th position. Thus (8, 4), (6, 11) etc. codes have dmin = 4 and correct single
errors with detection of double errors.

RECOMMENDATION QUESTIONS
1. Consider a (7, 4) linear code whose generator matrix is
1 0 0 0 1 0 1
0 1 0 0 1 1 1 
G=  
0 0 1 0 1 1 0 
 
0 0 0 1 011 
a) Find all code vectors of this code b) Find the parity check matrix for this code
c) Find the minimum weight of this code

2. For linear (n, k) block code, C, prove that CHT = O, where H is the parity check matrix.

3. The parity check bits of a ( 8,4) block code are generated by

c5 = d1 + d2 + d4
c6 = d1 + d2 + d3
c7 = d1 + d3 + d4
c8 = d2 + d3 + d4

Where d1, d2, d3, and d4 are the message bits

a) Find the generator matrix and parity check matrix for this code.
b) Find the minimum weight of this code
c) Find the error detecting and error correcting capabilities of this code
d) Show through an example that this code can detect 3 errors/code word.

4. Construct an encoder for the code given in problem 3.

5. Construct a syndrome circuit for the code given in problem 3.

6. Let H be the parity check matrix of an (n, k) linear code C that has both odd and even weight
code vectors. Construct new linear codes C1 and C2 with the following parity check matrices
respectively.

(Note that the last row of H1 (H2) consists of all 1‟s)

a) Show that C1 and C2, called extensions of C, are (n+1, k) linear block codes.

40
Error Control Coding And Binary Cyclic Code 15EC54

b) Show that every code vector of C1 (C2) has an even weight.

c) Show that the minimum distance of C1 (C2) is (d+1),), where d is the minimum distance of C.
d) Show that C1 can be obtained from C by adding an extra parity check digit denoted by vo to
the left of each code vector v as follows (1) If v has odd weight then vo = 1 and (2) If v has even
weight vo = 0. The parity check digit vo is called an “overall parity check digit”.

7. Let C be a linear code with both even and odd-weight code words. Show that the number of
even weight code vectors is equal to the number of odd weight code vectors.

8. Since the (8, 4) linear code of problem 3 has a minimum distance 4, it is capable of correcting all
the single error patterns and simultaneously detecting any combination of double errors.
Construct a decoder for this code. The decoder must be capable of correcting any single error
and detecting any double error.

9. The (8, 4) linear code of problem 3 is capable of correcting 16 error patterns (The co-set leaders
of a standard array). Suppose that this code is used for a BSC. Device a decoder for this code
based on the table-look up decoding scheme. The decoder is designed to correct the 16 most
probable error patterns.

10. Verify whether the dual code of the (8, 4) linear code of problem 3 is identical to the code itself.
Is the code self dual?

11. Form a parity check matrix for a (15, 11) systematic Hamming code. Device a decoder for this
code.

12. Let C1 be an (n1, k) linear systematic code with dmin = d1 and generator matrix G1 = [P1, Ik]. Let
C2 be an (n2, k) linear systematic code with dmin = d2 and Generator matrix G2 = [P2, Ik].
Consider an (n1 + n2, k) linear code with the following parity check matrix.

Show that this code has minimum distance at least (d1 +d2).

13. The “design distance” of an (n, k) linear block code is defined as being equal to (n - k +1).
Show that the minimum distance of the code can never exceed its design distance.

14. “Repetition codes” represent the simplest type of linear block codes. The generator matrix
of a (5, 1) repetition code is given as G = [1 1 1 1| 1]
a) Write its parity check matrix.
b) Evaluate the syndrome for:
i) All five possible single error patterns. ii) All 10 possible double error patterns.

15. Show that the decoder for a Hamming code fails if there are two or more than two transmission
errors in the received sequence.

41
Error Control Coding And Binary Cyclic Code 15EC54

OUTCOMES
• Detection of the errors.
• Correction of the errors using different techniques.

RESOURCES
• https://en.wikipedia.org/wiki/Block_code
• www.inference.phy.cam.ac.uk/mackay/itprnn/1997/l1/node7.html
• web.ntpu.edu.tw/~yshan/intro_lin_code.pdf
• users.ece.cmu.edu/~koopman/des_s99/coding/
• elearning.vtu.ac.in/P4/EC63/S11.pdf

42
Error Control Coding And Binary Cyclic Code 15EC54

CHAPTER 2 - BINARY CYCLIC CODES

STRUCTURE
 Generator Polynomial for Cyclic Codes
 Multiplication Circuits 
 Dividing Circuits  
 Systematic Cyclic Codes 
 Generator Matrix for Cyclic Codes
 Syndrome Calculation - Error Detection and Error
Correction

OBJECTIVE
 Discuss about cyclic codes and also study about implementation methods using feedback shift
registers.
 Study about syndrome calculator for cyclic codes.

43
Error Control Coding And Binary Cyclic Code 15EC54

INTRODUCTION

"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
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, 4.1. 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.

Fig 4.1: Illustrating the cyclic property

If the n - tuple, read from „A‟ in the CW direction in Fig 4.1,

v = (vo, v1, v2, v3, vn-2, vn-1) …………………… (4.1)

is a code vector, then the code vector, read from B, in the CW direction, obtained by a one bit cyclic
right shift:

v(1) = (vn-1 , vo, v1, v2, … vn-3,vn-2,) …………………… (4.2)

44
Error Control Coding And Binary Cyclic Code 15EC54

is also a code vector. In this way, the n - tuples obtained by successive cyclic right shifts:

v(2) = (vn-2, vn-1, vn, v0, v1…vn-3) ………………… (4.3a)

v(3) = (vn-3 ,vn-2, vn-1, vn, ... vo, v1, vn-4) ………………… (4.3b)

v(i) = (vn-i, vn-i+1,…vn-1, vo, v1,…. vn-i-1) …………… (4.3c)

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 (6.1) as a vector polynomial as:

V(X) = vo + v1 X + v2 X2 + v3 X3 +…+ vi-1 Xi-1 +... + vn-3 Xn-3 + vn-2 Xn-2 + vn-1 Xn-1 ….. (4.4)

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 5.7.1.

. 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
to the condition Xn = 1. This condition (i) restores XV(X) to the degree (n-1) (ii) Implies that right
most bit is fed-back at the left.

. This special form of multiplication is called "Multiplication modulo “Xn + 1”

. Thus for a single shift, we have

XV(X) = voX + v1 X2 + v2 X3 + ...... + vn-2 Xn-l + vn-l Xn

(+ vn-1 + vn-1) … (Manipulate A + A =0 Binary Arithmetic)

= vn-1 + v0 X + v1 X2 + + vn-2 Xn-1 + vn-1(Xn + 1)

=V (1) (X) = Remainder obtained by dividing XV(X) by Xn + 1

(Remember: X mod Y means remainder obtained after dividing X by Y)

Thus it turns out that

V (1) (X) = vn-1 + vo X + v1 X2 + ... + vn-2 Xn-1 ……………… (4.5)

(1)
I is the code polynomial for v . We can continue in this way to arrive at a general format:

45
Error Control Coding And Binary Cyclic Code 15EC54

X i V(X) = V (i) (X) + q (X) (Xn + 1) …………………… (4.6)

 
Remainder Quotient
Where

V (i) (X) = vn-i + vn-i+1X + vn-i+2X2 + …vn-1X i+ …v0Xi +v1Xi+1+…vn-i-2Xn-2 +vn-i-1Xn- …… (4.7)

4.1 GENERATOR POLYNOMIAL FOR CYCLIC CODES:

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

Theorem 4.1 "If g(X) is a polynomial of degree (n-k) and is a factor of (Xn +1) then g(X) generates
an (n, k) cyclic code in which the code polynomial V(X) for a data vector u = (u0, u1… uk-1) is
generated by

V(X) = U(X) g(X) ………………… (4.8)

Where U(X) = u0 + u1 X + u2 X2 + ... + uk-1 Xk-I ……………….. (4.9)

is the data polynomial of degree (k-1).

The theorem can be justified by Contradiction: - If there is another polynomial of same degree, then
add the two polynomials to get a polynomial of degree < (n, k) (use linearity property and binary
arithmetic). Not possible because minimum degree is (n-k). Hence g(X) is unique

Clearly, there are 2k code polynomials corresponding to 2k data vectors. The code vectors
corresponding to these code polynomials form a linear (n, k) code. We have then, from the theorem

n  k 1
g( X )  1   gi X i  X n  k …………………… (4.10)
i 1

As g(X) = go + g1 X + g2 X2 +…….+ gn-k-1 Xn-k-1 + gn-k Xn-k ………… (4.11)

is a polynomial of minimum degree, it follows that g0 = gn-k = 1 always and the remaining co-
efficients may be either' 0' of '1'. Performing the multiplication said in Eq (4.8) we have:

U (X) g(X) = uo g(X) + u1 X g(X) +…+uk-1Xk-1g(X) …………. (4.12)

46
Error Control Coding And Binary Cyclic Code 15EC54

Suppose u0=1 and u1=u2= …=uk-1=0. Then from Eq (4.8) 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).
Therefore from cyclic property Xi 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 (4.12) 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 (4.8) 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) = uoXk-1 + u1Xk-2 +…… + uk-2X + uk-1 ……………….. (4.13a)

V(X) = v0 Xn-1 + v1Xn-2 +…. + vn-2X + vn-1 ……………… (4.13b)

g(X) = g0Xn-k + g1Xn-k-1 +…..+gn-k-1 X + gn-k …………… (4.13c)

n k
= X n  k   gi X n  k  i  gn  k ………………… (4.13d)
i 1

Other manipulation and implementation procedures remain unaltered.

4.2 MULTIPLICATION CIRCUITS

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.

For the polynomial: A(X) = a0 + a1 X + a2 X2 +...+ an-1Xn-1 ………… (4.14)

where ai‟s 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)

47
Error Control Coding And Binary Cyclic Code 15EC54

= (a0 + a1 X + a2 X2 +… ..................+ an-1Xn-1) (b0 + b1 X + b2X2 +…+ bm-1 Xm-1)

= a0b0+ (a1b0+a0b1) X + (a0b2 + b0a2+a1b1) X2 +…. + (an-2bm-1+ an-1bm-2) Xn+m -3 +an-1bm-1Xn+m -2

This product may be realized with the circuits of Fig 4.2 (a) or (b), where A(X) is the input and the
co-efficient 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 6.2(a). Then the co-efficient of Xn+m-3 is obtained as the sum of {an-2bm-1 + an-1 bm-2}, the first
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, a0, 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 a0b0. The product is now complete and the
contents of the SR's become (0, 0, 0 …0, 0). Fig 4.2(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 4.2(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.

Fig 4.2: Multiplication circuits

Example 4.1: Consider that a polynomial A(X) is to be multiplied by

B(X) = 1 + X + X3 + X4 + X6

The circuits of Fig 4.3 (a) and (b) give the product C(X) = A(X). B(X)

48
Error Control Coding And Binary Cyclic Code 15EC54

Fig 4.3: Circuit to perform C(X)*(1+X2+X3+X4+X6)

Example 4.2: Consider the generation of a (7, 4) cyclic code. Here (n- k) = (7-4) = 3 and we have to
find a generator polynomial of degree 3 which is a factor of Xn + 1 = X7 + 1.

To find the factors of‟ degree 3, divide X7+1 by X3+aX2+bX+1, where 'a' and 'b' are binary
numbers, to get the remainder as abX2+ (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

gl (X) = X3+ X2+ 1 and g2 (X) = X3+X+1

In fact, X7 + 1 can be factored as:

(X7+1) = (X+1) (X3+X2+1) (X3+X+1)

Thus selection of a 'good' generator polynomial seems to be a major problem in the design of cyclic
codes. No clear-cut procedures are available. Usually computer search procedures are followed.

Let us choose g (x) = X3+ X + 1 as the generator polynomial. The encoding circuits are shown
in Fig 4.4(a) and (b).

Fig 4.4 Generation of Non-systematic cyclic codes

49
Error Control Coding And Binary Cyclic Code 15EC54

To understand the operation, Let us consider u = (10 1 1) i.e.

U (X) = 1 +X2+X3.

We have V (X) = (1 +X2+X3) (1 +X+X3).

= 1 +X2+X3+X+X3+X4+X3+X5+X6

= 1 + X + X2+ X3+ X4+ X5+ X6 because (X3+ X3=0)

=> v = (1 1 1 1 1 1 1)

The multiplication operation, performed by the circuit of Fig 6.4(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:
V (X) = 1 +X+X2+X3+X4+X5+X6,
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 4.4(b) in the same manner. Thus the multiplication
circuits of Fig 6.4 can be used for generation of non-systematic cyclic codes.

Table showing sequence of computation

Shift Input Bit Contents of shift Ou Remarks

Number Queue shifted registers. t
IN SRI SR2 SR3 put
0 0001011 - 0 0 0 - Circuit In reset
mode
1 000101 1 1 0 0 1 Co-efficient of X6
2 00010 1 1 1 0 1 Co-efficient of X5
3 0001 0 0 1 1 1 X4 co-efficient
*4 000 1 1 0 1 1 X3 co-efficient
5 00 0 0 1 0 1 X2 co-efficient
6 0 0 0 0 1 1 X1 co-efficient
7 - 0 0 0 0 1 X0co-efficient

4.3 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 4.5. In a division circuit, the first co-efficient of the
quotient is (an-l (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.

50
Error Control Coding And Binary Cyclic Code 15EC54

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.

Fig 4.5: Dividing circuit

It is possible to combine a divider circuit with a multiplier circuit to build a “composite
multiplier-divider circuit” which is useful in various encoding circuits. An arrangement to accomplish
this is shown in Fig 4.6(a) and an illustration is shown in Fig 4.6(b).

We shall understand the operation of one divider circuit through an example. Operation of
other circuits can be understood in a similar manner.

Example 4.3:

Let A(X) = X3+X5+X6, A= (0001011), B(X) = 1 +X+X3. 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 4.7,
drawn using the format of Fig 4.5(a). The operation of the divider circuit is listed in the table:

Fig 4.6 Circuits for Simultaneous Multiplication and division

51
Error Control Coding And Binary Cyclic Code 15EC54

Fig 4.7 Circuits for dividing A(x) by (1 + X + X3)

Table Showing the Sequence of Operations of the Dividing circuit

Shift Input Bit Contents of shift Ou Remarks

Numbe Queue shifted Registers. t
r IN SRI SR SR put
2 3
0 0001011 - 0 0 0 - Circuit in reset
mode
1 000101 1 1 0 0 0 Co-efficient of X6
2 00010 1 1 1 0 0 Co-efficient of X5
3 0001 0 0 1 1 0 X4 co-efficient
4 *000 1 0 1 1 1 X3 co-efficient
5 00 0 1 1 1 1 X2 co-efficient
6 0 0 1 0 1 1 X1 co-efficient
7 - 0 1 0 0 1 Xo co-efficient

The quotient co-efficients will be available only after the fourth shift as the first three shifts
result in entering the first 3-bits to the shift registers and in each shift out put of the last register, SR3,
is zero.

The quotient co-efficient serially presented at the out put are seen to be (1111) and hence the
quotient polynomial is Q(X) =1 + X + X2 + X3. The remainder co-efficients are (1 0 0) and the
remainder polynomial is R(X) = 1.

4.4 SYSTEMATIC CYCLIC CODES:

Let us assume a systematic format for the cyclic code as below:

v = (p0, p1, p2 … pn-k-1, u0, u1, u2… uk-1) …………… (4.15)

The code polynomial in the assumed systematic format becomes:

V(X) = p0 + p1X + p2X2 + … +pn-k-1Xn-k-1 +u0Xn-k + u1Xn-k+1 +… +uk-1Xn-1 ………... (4.16)

= P(X) + Xn-kU(X) …………………… (4.17)

Since the code polynomial is a multiple of the generator polynomial we can write:

52
Error Control Coding And Binary Cyclic Code 15EC54

V (X) = P (X) +Xn-k U (X) = Q (X) g (X) ....................... (4.18)

X n  kU ( X ) P( X )
  Q( X )  ………………. (4.19)
g( X ) g( X )

Thus division of Xn-k 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 remainder polynomial P (X) after dividing Xn-k U (X) by g(X). This division process
can be easily achieved by noting that "multiplication by Xn-k amounts to shifting the sequence by
(n-k) bits". Specifically in the circuit of Fig 4.5(a), if the input A(X) is applied to the Mod-2 adder
after the (n-k) th shift register the result is the division of Xn-k A (X) by B (X).

Accordingly, we have the following scheme to generate systematic cyclic codes. The
generator polynomial is written as:

g (X) = 1 +glX+g2X2+g3X3+…+gn-k-1 Xn-k-1 +Xn-k ………… (4.20)

The circuit of Fig 4.8 does the job of dividing Xn-kU (X) by g(X). The following steps describe the
encoding operation.

Fig 4.8 Syndrome encoding of cyclic codes using (n-k) shift register stages

1. The switch S is in position 1 to allow transmission of the message bits directly to an

out put shift register during the first k-shifts.
2. At the same time the 'GATE' is 'ON' to allow transmission of the message bits into the
(n-k) stage encoding shift register
3. After transmission of the kth message bit the GATE is turned OFF and the switch S is
moved to position 2.
4. (n-k) zeroes introduced at "A" after step 3, clear the encoding register by moving the
parity bits to the output register
5. The total number of shifts is equal to n and the contents of the output register is the
code word polynomial V (X) =P (X) + Xn-k U (X).
6. After step-4, the encoder is ready to take up encoding of the next message input

Clearly, the encoder is very much simpler than the encoder of an (n, k) linear block code and the
memory requirements are reduced. The following example illustrates the procedure.

53
Error Control Coding And Binary Cyclic Code 15EC54

Example 4.4:

Let u = (1 0 1 1) and we want a (7, 4) cyclic code in the systematic form. The generator polynomial
chosen is g (X) = 1 + X + X3

For the given message, U (X) = 1 + X2+X3

Xn-k U (X) = X3U (X) = X3+ X5+ X6

We perform direct division Xn-kU (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)

Fig 4.9 Encoder for the (7,4) cyclic code

The encoder circuit for the problem on hand is shown in Fig 4.9. The operational steps are as follows:

Shift Number Input Bit shifted IN Register Output

Queue contents
0 1011 - 000 -
1 101 1 110 1
2 10 1 101 1
3 1 0 100 0
4 - 1 100 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

54
Error Control Coding And Binary Cyclic Code 15EC54

agrees with the direct hand calculation.

4.5 GENERATOR MATRIX FOR CYCLIC CODES:

The generator polynomial g(X) and the parity check polynomial h(X) uniquely specify the
generator matrix G and the parity check matrix H respectively. We shall consider the construction of
a generator matrix for a (7, 4) code generated by the polynomial g(X) = 1 +X+X3.

We start with the generator polynomial and its three cyclic shifted versions as below:
g(X) = 1 + X + X3
X g(X) = X + X2 + X4
X2g(X) = X2 + X3 + X5
X3g(X) = X3 + X4 + X6
The co-efficients of these polynomials are used as the elements of the rows of a (47) matrix to get
the following generator matrix:

1 1 0 1 0 0 0
0 1 1 0 1 0 0 
G
0 0 1 1 0 1 0
 
0 0 0 1 1 0 1

Clearly, the generator matrix so constructed is not in a systematic format. We can transform this into
a systematic format using Row manipulations. The manipulations are:

First row = First row; Second row = Second row; Third row = First row + Third row; and Fourth row
= First row + second row + Fourth row.

These operations give the following result:

1 1 0 1 0 0 0
0 1 1 0 1 0 0 
G  [ P  I4 ]
1 1 1 0 0 1 0
 
1 0 1 0 0 0 1

Using this generator matrix, which is in systematic form the code word for u = (1 0 1 1) is
v = (1 0 0 1 0 1 1) (obtained as sum of 1st row + Third row + Fourth row of the G-matrix). The result
agrees with direct hand calculation.
To construct H-matrix directly, we start with the reciprocal of the parity check
polynomial defined by Xkh(X-1). Observe that the polynomial Xkh(X-1) is also a factor of the
polynomial Xn+ 1. For the polynomial (X7+1) we have three primitive factors namely, (X + 1),
(X3+X+1) and (X3+X2+1). Since we have chosen (X3+X+1) as the generator polynomial the
other two factors should give us the parity check polynomial.
h(X) = (X +1) (X3+X2+1) = X4+X2+X+1

There fore with h(X) = 1 +X+X2+X4, we have

55
Error Control Coding And Binary Cyclic Code 15EC54

h(X-1) = 1 +X-1+X-2+X-4, and

Xk h(X-1) = X4h(X-1) = X4+X3+X2+1

The two cyclic shifted versions are:

X5h(X-1) = X5 + X4 +X3 + X

X6P(X-1) = X6 + X5 + X4 + X2

Or X4h(X-1) = X4+X3+X2+1

X5h(X-1) = X5 + X4 +X3 + X

X6h(X-1) = X6 + X5 + X4 + X2

Using the co-efficients of these polynomials, we have:

1 0 1 1 1 0 0 
H  0 1 0 1 1 1 0 
0 0 1 0 1 1 1

Clearly, this matrix is in non systematic form. It is interesting to check that for the non-
systematic matrixes obtained GHT = O. We can obtain the H matrix in the systematic format
H = [ I 3  P T ] , by using Row manipulations. The manipulation in this case is simply.
'First row = First row + Third row'. The result is

 1 0 0 1 0 1 1
H  0 1 0 1 1 1 0 
0 0 1 0 1 1 1

Observe the systematic format adopted: G  [ P  I k ] and H  [ I n k  P T ]

4.6 SYNDROME CALCULATION - ERROR DETECTION AND ERROR

CORRECTION :
Suppose the code vector v= (v0, v1, v2 …vn-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
vector be r = (r0, r1, r2…rn-1). The received vector may not be anyone of the 2k 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

56
Error Control Coding And Binary Cyclic Code 15EC54

error correction. Let the received code vector be represented by the polynomial

R(X) = r0+r1X+r2X2+…+rn-1Xn-l

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 )  ……………….. (4.21)
g( X ) g( X )

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:
R(X) =V(X) + E(X) ……………….. (4.22)

And it follows as V(X) = U(X) g(X), that:

E(X) = [A(X) + U(X)] g(X) +S(X) ………………. (4.23)

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. Fig 4.5. A “Syndrome calculator” is shown in Fig 4.10.

Fig 4.10 Syndrome calculator using (n-k) shift register

The syndrome calculations are carried out as below:

1 The register is first initialized. With GATE 2 -ON and GATE1- OFF, the received vector is
entered into the register

2 After the entire received vector is shifted into the register, the contents of the register will be
the syndrome, which can be shifted out of the register by turning GATE-1 ON and GATE-2
OFF. The circuit is ready for processing next received vector.

Cyclic codes are extremely well suited for 'error detection' .They can be designed to
detect many combinations of likely errors and implementation of error-detecting and error correcting
circuits is practical and simple. Error detection can be achieved by employing (or adding) an
additional R-S flip-flop to the syndrome calculator. If the syndrome is nonzero, the flip-flop sets and
provides an indication of error. Because of the ease of implementation, virtually all error detecting
codes are invariably 'cyclic codes'. If we are interested in error correction, then the decoder must be
capable of determining the error pattern E(X) from the syndrome S(X) and add it to R(X) to

57
Error Control Coding And Binary Cyclic Code 15EC54

determine the transmitted V(X). The following scheme shown in Fig 6.11 may be employed for the
purpose. The error correction procedure consists of the following steps:

Step1. Received data is shifted into the buffer register and syndrome registers with switches
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 pattern with an error at the highest order position Xn-l. 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.

Fig 4.11 General decoder for cyclic code

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

58
Error Control Coding And Binary Cyclic Code 15EC54

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
6.10 and Fig6.11 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
such simplicity is in the reduction of code efficiency for a given block size.

A decoder of the form described above operates on the received data bit by bit; and each bit is
tested in turn for error and is corrected whenever an error is located. Such a decoder is called
a“Meggitt decoder”.

For illustration let us consider a decoder for a (7, 4) cyclic code generated by

g(X) = 1 + X + X 3

The circuit implementation of the Meggitt decoder is shown on Fig 6.12. The entire received
vector R(X) is entered in to the SR‟s bit by bit and at the same time it is stored in the buffer memory.
The division process will start after the third shift and after the seventh shift the syndrome will be
stored in the SR‟s. If S(X) = (000) then E(X) = 0 and R(X) is read out of the buffer. Since S(X) can
be found from E(X) with nonzero coefficients, suppose E(X) = (000 0001). Then the SR contents are
given as: (001, 110, 011, 111, 101) showing that S(X) = (101) after the seventh shift. At the eighth
shift, the SR content is (100) and this may be used through a coincidence circuit to correct the error
bit coming out of the buffer at the eighth shift. On the other hand if the error polynomial were
E(X) = (000 1000) then the SR content will be (100) at he eleventh shift and the error will be
corrected when the buffer delivers the error bit at the eleventh shift. The SR contents for different
shifts, for two other error patterns are as shown in the table below:
SR contents for the error patterns (1001010) and (1001111)
Shift Input SR-content for Input SR- content for
Number (1001010) (1001111)
1 0 000 1 100
2 1 100 1 110
3 0 010 1 111
4 1 101 1 001
5 0 100 0 110
6 0 010 0 011
7 1 101 1 011 *Indicates an error
8 0 100 0 111
9 - - 0 101
10 - - 0 100

59
Error Control Coding And Binary Cyclic Code 15EC54

Fig 4.12 Meggitt decoder for (7,4) cyclic code

For R(X) = (1001010), the SR content is (100) at the 8-th shift and the bit in X6 position of
R(X) is corrected giving correct V(X) = (1001011). On the other hand , if R(X) = (1001111), then it
is seen from the table that at the 10-th shift the syndrome content will detect the error and correct the
X4 bit of R(X) giving V(X) = (1001011).

The decoder for the (15, 11) cyclic code, using g(X) = 1 + X + X 4, is shown in Fig 6.13. It is
easy to check that the SR content at the 16-th shift is (1000) for E(X) =X 14. Hence a coincidence
circuit gives the correction signal to the buffer out put as explained earlier.
Although the Meggitt decoders are intended for Single error correcting cyclic codes, they may
be generalized for multiple error correcting codes as well, for example (15, 7) BCH code.
An error trapping decoder is a modification of a Meggitt decoder that is used for certain cyclic
codes.

Fig 4.13 Meggitt decoder for (15,11) cyclic code

The syndrome polynomial is computed as: S(X) = Remainder of [E(X) / g(X)]. If the error
E(X) is confined to the (n-k) parity check positions (1, X, X2… Xn-k-1) of R(X), then E(X) = S(X),
since the degree of E(X) is less than that of g(X). Thus error correction can be carried out by simply
adding S(X) to R(X). Even if E(X) is not confined to the (n-k) parity check positions of R(X) but has
nonzero values clustered together such that the length of the nonzero values is less than the syndrome
length, then also the syndrome will exhibit an exact replica of the error pattern after some cyclic
shifts of E(X). For each error pattern, the syndrome content S(X) (after the required shifts) is
subtracted from the appropriately shifted R(X), and the corrected V(X) recovered.

“If the syndrome of R(X) is taken to be the remainder after dividing X n-k R(X) by g(X),
and all errors lie in the highest-order (n-k) symbols of R(X), then the nonzero portion of the
error pattern appears in the corresponding positions of the syndrome”. Fig 4.14 shows an error
trapping decoder for a (15, 7) BCH code based on the principles described above. A total of 45 shifts

60
Error Control Coding And Binary Cyclic Code 15EC54

are required to correct the double error, 15 shifts to generate the syndrome, 15 shifts to correct the
first error and 15 shifts to correct the second error.

Fig 4.14 Error trapping decoder for (15,7) BCH code

Illustration:

U(X) = X6 + 1; g(X) = X8 + X7 + X6 + X4 + 1

V(X) = X14 + X13 + X12 + X10 + X8 + X7 + X4 + 1

E(X) = X11 + X

R(X) = X14 + X13 + X12 + X11 + X10 + X8 + X7 + X4 + X + 1

r = (110010011011111)
Shift Syndrome Shift Middle Shift Bottom
Number Generator Number Register Number Register
Register
1 10001011 16 01100011 31 00000100
2 01000101 17 10111010 32 00000010
3 00100010 18 01011101 33 00000001
4 10011010 19 10100101 34 00000000
5 11000110 20 11011001 35 ;
6 01100011 21 11100111 36 :
7 00110001 22 11111000 37 :
8 00011000 23 01111100 38 All zeros
9 00001100 24 00111110 39 :
10 00000110 25 00011111 40 :
11 10001000 26 10000100 41 :
12 01000100 27 01000010 42 :

61
Error Control Coding And Binary Cyclic Code 15EC54

13 00100010 28 00100001 43 :
14 10011010 29 00010000 44 :
15 11000110 30 00001000 45 :
Errors trapped at shift numbers 28 and 33.

Some times when error trapping cannot be used for a given code, the test patterns can be
modified to include the few troublesome error patterns along with the general test. Such a modified
error trapping decoder is possible for the (23, 12) Golay code in which the error pattern E(X) will be
of length 23 and weight of 3 or less (t  3). The length of the syndrome register is 11 and if E(X) has
a length greater than 11the error pattern is not trapped by cyclically shifting S(X). In this case, it is
shown that one of the three error bits must have at least five zeros on one side of it and at least six
zeros the other side. Hence all error patterns can be cyclically shifted into one of the following three
configurations (numbering the bit positions, e0, e1, e2 … e22):

(i) All errors (t  3 ) occur in the 11 high-order bits

(ii) One error occurs in position e5 and the other two errors occur in the 11
high-order bits.

(iii) One error occurs in position e6 and the remaining two errors occur in the
11 high-order bits.

In the decoder shown in Fig 4.15, the received code vector R(X) is fed at the rightmost stage
of the syndrome generator (as was done in Fig 6.14), equivalent to multiplying R(X) by X11. Then the
syndrome corresponding to e5 and e6 are obtained (using g1(X) as the generator polynomial) as:

S (e5) = Remainder of [X 16 /g1(X)] = X + X2 + X5 + X6 + X8 + X9 and

S (e6) = Remainder of [X 17 /g1(X)] = X2 + X3 + X6 + X7 + X9 + X10

Fig 4.15 Error trapping decoder for (23,12) Golay code

The syndrome vectors for the errors e5 and e6 will be (01100110110) or (00110011011)
respectively. Two more errors occurring in the 11 high-order bit positions will cause two 1‟s in the

62
Error Control Coding And Binary Cyclic Code 15EC54

appropriate positions of the syndrome vectors, thereby complementing the vector for e5 or e6. Based
on the above relations, the decoder operates as follows:

(i).The entire received vector is shifted into the syndrome generator (with switch G1 closed)
and the syndrome S(X) corresponding to X 11 R(X) is formed.

(ii).If all the three or less errors are confined to X 12, X 13 … X22 of R(X), then the syndrome
matches the errors in these positions. The weight of the syndrome is now 3 or less. This is
checked by a threshold gate and the gate output T0 switches G2 ON and G1 OFF. R(X) is now
received from the buffer and corrected by the syndrome bits (as they are clocked bit by bit)
through the modulo-2 adder circuit.

(iii).If the test in (ii) fails then it is assumed that one error is either at e5 or at e6, and the other
two errors are in the 11 high-order bits of R(X). Then if the weight of S(X) is more than 3 (in
test (ii)), then the weights of [S(X) + S (e5)] and [S(X) + S (e6)] are tested. The decisions
are:
1. If weight of [S(X) + S (e5)]  2 then the decision (T1 = 1) is that one error
is at position e5 and two errors are at positions where [S(X) + S (e5)] are nonzero.

2. If weight of [S(X) + S (e6)]  2 then the decision (T2 = 1) is that one error
is at position e6 and two errors are at positions where [S(X) + S (e6)] are nonzero.
The above tests are arranged through combinatorial switching circuits and the
appropriate corrections in R(X) are made as R(X) is read from the buffer.
(iv). If the above tests fail then with G1 and G3 ON and G2 OFF, the syndrome and buffer
contents are shifted by one bit. Tests (ii) and (iii) are now repeated. Bit by bit shifting of S(X)
and R(X) is continued till the errors are located, and then corrected. A maximum of 23 shifts
will be required to complete the process. After correction of R(X), the corrected V(X) is
further processed through a divider circuit to obtain the message U(X) = V(X) / g(X).

Assuming that upon shifting the block of 23 bits with t  3 cyclically, „at most one error will
lie outside the 11 high-order bits of R(X)‟ at some shift, an alternative decoding procedure can be
devised for a Golay coder – The systematic search decoder. Here the test (ii) is first carried out. If the
test fails, then first bit of R(X) is inverted and a check is made to find if the weight of S(X)  2. If
this test is successful, then the nonzero positions of S(X) give the two error locations (similar to test
(iii) above) and the other error is at first position. If this test fails, then the syndrome content is
cyclically shifted, each time testing for weight of S(X)  3; and if not, invert 2nd, 3rd ……and 12th bit
of R(X) successively and test for weight of S(X)  2. Since all errors are not in the parity check
section, an error must be detected in one of the shifts. Once one error is located and corrected, the
other two errors are easily located and corrected by test (ii). Some times the systematic search
decoder is simpler in hardware than the error trapping decoder, but the latter is faster in operation.
The systematic search decoder can be generalized for decoding other multiple-error-correcting cyclic
codes. It is to be observed that the Golay (23, 12) code cannot be decoded by majority-logic
decoders.

63
Error Control Coding And Binary Cyclic Code 15EC54

RECOMMENDATION QUESTIONS
1. Sketch the shift register circuit for multiplying A(X) by B(X) = 1 + X3 + X4 + X5 + X6. Compute
the circuit output for A(X) = 1 + X + X4 by examining the shift register contents at each shift.

2. Draw a shift register circuit for simultaneous multiplication by B(X) and division by D(X), where
B(X) = 1 + X2 + X5 + X6 and D(X) = 1 + X3 + X4 + X5 + X8
If the input to the circuit is A(X) = 1 + X7, calculate the quotient and the remainder.

3. Determine which, if any, of the following polynomials can generate a cyclic code with code word
length n  7. Find the (n, k) values of any such codes that can be generated.

(a) 1 + X3 + X4 (d) 1 + X + X2 + X4
(b) 1 + X2 + X4 (e) 1 + X3 + X5
(c) 1 + X + X3 +X4

4. A (15, 5) linear cyclic code has a generator polynomial: g(X) = 1+X+X2+X4+X5+X8+X10

a) Draw block diagrams of an encoder and syndrome calculator for this code.
b) Find the code polynomial for the message polynomial U(X) =1+X2+X4 in systematic form.
c) Is V(X) =1+X4+X6+X8+X14 a code polynomial? If not, find the syndrome of v (X).
d) Hard-decision detection gives the received code word as:

R(X) = 1 + X4 + X7 + X8 + X9 + X10 +X14. Locate the errors.

5. Write the H Matrix for the (15, 11) cyclic code using g(X) = 1 + X + X2 + X3 + X4 determine the
code polynomial for U(X) = 1 + X3 + X7 + X10. Construct the decoder for the code.

6. Write the generator polynomial for (31, 26) SEC cyclic code. Write the G and H matrices for the
code. Find the code polynomial for the message:

U(X) =1 + X + X4 + X20. Draw the Meggitt decoder circuit for the code.

7. Write the G matrix for the (15, 11) SEC code. Show that by successively removing some of the
rows of G, one obtains (14, 10), (13, 9), (12, 8), (11, 7), (10, 6) etc. SEC codes. Write H matrix
for the (10, 6) code. Construct the coder and decoder for the code.

8. Construct a Meggitt decoder for the (15, 7) BCH code. What are the error patterns which will
form the test syndromes for the decoder? Is it possible to reduce the test syndromes to only 8
error patterns as below:

0 0 0 0 0 0 0 1, 0 0 0 0 0 0 1 1
0 0 0 0 0 1 0 1, 0 0 0 0 1 0 0 1
0 0 0 1 0 0 0 1, 0 0 1 0 0 0 0 1
0 1 0 0 0 0 0 1, 1 0 0 0 0 0 0 1

What is the modification required in the new decoder? What is the total number of shifts required
to complete the decoding of a block?

64
Error Control Coding And Binary Cyclic Code 15EC54

9. The generator polynomial for a cyclic code is g(X) =1+X4+X6+X7+X8

a) Show that its length is 15.
b) Find the generator matrix and parity check matrix in systematic form.
c) Devise two shift register encoder circuits using k = 7 stages and (n – k) = 8 stages.
d) Find the code vector (In systematic form) for the message polynomial
U(X) =1+X2+X3+X4
e) Assume that the first and last bits of the code vector V(X) for U(X) given in (a)
Suffer transmission errors. Find the syndrome of V(X).

10. The decoder for a class of single error correcting cyclic codes (Hamming codes) is shown in Fig
P7.1 Show by way of an example, that single error in a (15, 11) Hamming code generated by
g(X) = 1+X+X4 can be decoded using the decoder shown in the figure.

11. A t-error correcting code is said to be a “perfect code” if it is possible to form a standard array,
with all patterns of„t‟ or fewer errors and no others as co-set leaders. Show that a (7, 4) linear
block code generated by g(X) =1+X+X3 is a perfect code.

12. The “Expurgated (n, k-1) Hamming code” is obtained from the original (n, k) Hamming code
by discarding some of the code words. Let g(X) Denote the generator polynomial of the original
code. The most common expurgated code is the one generated by g1(X) = (1+X) g(X), where
(1+X) is a factor of (1+Xn). Consider the (7, 4) Hamming code generated by g(X) =1+X2+X3.

a) Construct the eight code words in the expurgated (7, 3) Hamming code, assuming a
systematic format. Hence show that the minimum distance of the code is 4.
b) Determine the generator matrix G and the parity check matrix H of the expurgated
Hamming code.
c) Device the encoder and syndrome calculator for the expurgated Hamming code.
Hence determine the syndrome for the received sequence 0111110.

13. A systematic (7, 4) cyclic code is generated by g(X) = 1 + X2 + X3. The message is
U(X) =1+X+X3, and after detection the effective error polynomial is E(X) = X4. Find the first
syndrome word generated by a Meggitt decoder for decoding the first received symbol.

65
Error Control Coding And Binary Cyclic Code 15EC54

14. For the (7, 4) cyclic code generated by g(X) of problem 9, write the generator matrix and the
parity check matrices. Delete the last three columns in the H matrix in order to generate a (4, 1)
shortened cyclic code. If the code word v = (1101) is received as r = (1111), show that the
decoder corrects the error on the fifth clock pulse.

15. Show that a binary cyclic code of length n generated by g(X) has minimum weight of at least 3 if
n is the smallest integer for which g(X) divides (Xn + 1).

16. Let G = [I: P] be the generator matrix of a cyclic code. If h(X)=1+h1X+h2X2+…+hk –1Xk–1+Xk
is the parity check polynomial of the code, show that the last column of the matrix P is:
(hk – 1, hk – 2 …..h2, h1, 1)

17. A (31, 21) binary double error correcting code has the generator polynomial:

g(X) = 1 + X3 + X5 + X6 + X8 + X9 + X10

i. Show that an error-trapping decoder cannot decode this code to the designed distance.
ii. Show a simple modification of the error-trapping decoder that will decode to the
designed distance.

18. The triple-error-correcting Golay (23, 12) code may be constructed with the generator
polynomial: g(X) = 1 + X + X5 + X6 + X7 + X9 + X11. Find the parity check polynomial h(X).
Construct a decoder for the code. What are the test syndromes to be checked in the decoder?

OUTCOMES

 How the cyclic codes can be implemented using feedback shift registers.
 How the errors can be detected and corrected.

REFERENCE

 www.mcs.csueastbay.edu/~malek/Class/Cyclic.pdf
 www.fi.muni.cz/usr/.../CHAPTER%2003%20-%20Cyclic%20codes.ppt
 elearning.vtu.ac.in/P4/EC63/S11.pdf
 nptel.ac.in/courses/IIT...Of.../Lecture40-41_ErrorControlCoding.pdf

66
Cyclic codes and Convolution codes 15EC54

MODULE 5
CHAPTER 1: SOME IMPORTANT CYCLIC CODES

STRUCTURE

 Golay codes
 BCH codes

OBJECTIVE

 Discuss about BCH codes and Golay codes.

Dept. of ECE, ATMECE Page 1

Cyclic codes and Convolution codes 15EC54

5.1 GOLAY CODES

Golay code is a (23, 12) perfect binary code that is capable of correcting any combination of
three or fewer random errors in a block of 23 bits. It is a perfect code because it satisfies the
Hamming bound with the equality sign for t = 3 as:

 23   23   23   23  12
           2  2
23

 0   1   2   3 

The code has been used in many practical systems. The generator polynomial for the code is
obtained from the relation (X23+1) = (X+ 1) g1(X) g2(X), where:

g1(X) = 1 + X2 + X4 + X5 + X6 + X10 + X11 and g2 (X) = 1 + X + X5 + X6 + X7 + X9 + X11

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.

5.2 BCH CODES

BCH codes can also be used for burst error correction. If correction of single bursts is the only
requirement, implementation would be much simpler. The additional error correcting ability can be
used to error detection (any error pattern that the code would correct, if fully utilized, it could
certainly still detect) or it can be used to correct bursts beyond its guaranteed burst correcting
capability.

5.2.1 Decoding:

Burst error correcting cyclic codes can be decoded with an extremely simple version of the
error trapping decoder. Fig 5.1 illustrates a scheme for decoding an (n, k) cyclic code with burst error
correcting ability „b‟ (The following description assumes a Binary code. Defining suitable q-ary logic
gates it can be easily generalized for q – ary codes). To understand the operation of the circuit we
proceed as below.

 Assume that a burst of length „b‟ or less has occurred.

 The circuit computes a shifted version of the syndrome of the error pattern which is contained
in the received word.
 If the burst is confined to the „b‟ high-order positions of the received word, then the „b‟ right
most stages of the syndrome generator contain the burst and the (n-k-b) left stages will
contain zeros.

Dept. of ECE, ATMECE Page 2

Cyclic codes and Convolution codes 15EC54

 The logical OR gate detects this situation. Since its output is now 0, the feedback path in the
syndrome generator is opened and the path to the binary adder is closed.
 Upon shifting the syndrome generator and buffer register, the syndrome, which is the error
burst, is added to the received word, thereby correcting it.
 If the burst has occurred elsewhere in the word, it would be corrected in exactly the same
manner. For, after shifting, the burst will occupy the „b‟ high-order positions of the buffer. At
this time the „b‟ high-order bits of the syndrome will be identical to the burst as before.
 This decoding procedure does not correct “bursts” that occupy only the „i‟ high-order
positions of the word and (b-i) low-order positions. Such bursts are correctable with a burst-b-
correcting cyclic code, however. It is possible, even simple; to do this if a syndrome generator
which does not pre multiply the received word by Xn-k is used.

Fig 5.1 A decoder for burst error correcting cyclic codes

Example 5.1: Interleaver for a BCH code.

Consider a (15, 7) BCH code generated by g(X) = 1+X+X2+X4+X8. For this code dmin=5,
d 1
t = mim  2 .With λ =5, we can construct a (75, 35) interleaved code with a burst error correcting
2
capability of b= λt=10. 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 515
matrix. The columns of the matrix are transmitted in the sequence shown as a 75-bit long code vector.

 Each rows is 15  bit code word 

1 6 11 …. 31 (36) ….. (66) 71
2 7 12 …. (32) (37) ….. 67 72
3 8 13 …. (33) (38) ….. 68 73
4 (9) 14 …. (34) 39 ….. 69 74
5 10 15 …. (35) 40 …… (70) 75
Fig 5.2 Block Interleaver for a (15, 7) BCH code.

Dept. of ECE, ATMECE Page 3

Cyclic codes and Convolution codes 15EC54

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 de-
interleaver now feeds the rows of Fig 5.1 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.

Example 5.2: The polynomial: G(X) = (1+X+ X4)(1+X7) = 1+X+X4+X7+X8+X11 , generates a binary
fire code of length n=7(24-1)=105, which corrects any single burst of length 4 or less. It has 11 check
symbols and 105-11=94 information symbols. The encoder for the code is shown in Fig 5.3(a) and
the decoder in Fig 5.3(b).

Fig 5.3 Encoder/Decoder for a (105,94)

 The decoder has a local encoder.
 The entire received vector is read into the buffer and simultaneously in to the syndrome
generator with G1 closed and G2 open.
 After the syndrome bits have been formed, the received bits are read out from the buffer
bit by bit and the syndrome bits are also shifted with flushing zero input.
 When allzero bits appear in the first 7 syndrome positions, the test circuit indicates the
allzero pattern, at the same time the error pattern must be in the last 5 positions.
 The erroneous symbols are now ready to leave the buffer and they are corrected by the last
4 syndrome bits by addition while they are shifted bit by bit with gates G2 closed and G1
open.

It may be desirable to shorten a burst error correcting code either because (i) bursts occur too
frequently or (ii) the total length or number of information bits is constrained by other system
requirements. If a code of suitable natural length is not available or it cannot be found, an available

Dept. of ECE, ATMECE Page 4

Cyclic codes and Convolution codes 15EC54

code can be shortened simply by making some high-order information bits completely „0‟ and
omitting them. By doing this, the encoding and parity check calculations are not affected, since the
leading 0’s do not affect them. However, such shortening of codes do affect the error correction
procedure. Suppose we want to form a shortened cyclic code of length (n-s) from an existing cyclic
code of natural length n. This requires omitting „s‟ information symbols from the original n. Then, to
have an unaltered correction procedure, we must first shift the „s‟ omitted information bits which are
assumed to be 0’s before reading the actual received code word out of the buffer. One possible
simplification could be achieved by an automated pre-multiplication by „Xs mod g(X)‟.

OUTCOMES

 To know how the decoding can be done.

 How the interleaving of coded data can be obtained for burst error correction.

REFERENCE
 nptel.ac.in/Clarify_doubts.php?subjectId=117106031&lectureId=16
 elearning.vtu.ac.in/EC63.html
 www.inference.phy.cam.ac.uk/itprnn/book.pdf

Dept. of ECE, ATMECE Page 5

Cyclic codes and Convolution codes 15EC54

CHAPTER 2 CONVOLUTIONAL CODES

STRUCTURE

 Connection Pictorial Representation

 Convolutional Encoding – Time domain approach
 Encoding of Convolutional Codes; Transform Domain Approach
 Systematic Convolutional Codes
 Structural Properties of Convolutional Codes
 Maximum Likely-hood decoding of Convolutional Codes
 Sequential Decoding

OBJECTIVE

 Discuss about Convolution codes which is used to generate encoded sequences for input
sequence on bit by bit basis.
 Discuss about the decoding of convolutional codes with the help of Viterbi algorithm.

Dept. of ECE, ATMECE Page 6

Cyclic codes and Convolution codes 15EC54

INTRODUCTION

In block codes, a block of n-digits generated by the encoder depends only on the block of k-
data 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.

5.1 CONNECTION PICTORIAL REPRESENTATION

The encoder for a (rate 1/2, K = 3) or (2, 1, 2) convolutional code is shown in Fig.5.1. Both
sketches shown are one and the same. While in Fig.5.1 (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.

Fig 5.1 A (2,1,2) Encoder (a) Representation using 3-bit shift register (b) Equivalent representation
requires only two shift register stages

Dept. of ECE, ATMECE Page 7

Cyclic codes and Convolution codes 15EC54

At each input bit time one bit is shifted into the left most stage and the bits that were present
in the 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.5.1 for an input sequence u = (101)
are illustrated diagrammatically in Fig 5.2.

Fig 5.2
From Fig 5.2, 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

Dept. of ECE, ATMECE Page 8

Cyclic codes and Convolution codes 15EC54

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
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 5.2 is
rather tedious. We can approach the encoder in terms of “Impulse response” or “generator sequence”
which merely represents the response of the encoder to a single „1‟ bit that moves through it.

5.2 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 module-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
information sequence. Specifically, the encoder output sequences, in this case v(1) = {v1(1), v2(1), v3(1)
… }and v(2) = {v1(2),v2(2),v3(2) … } are obtained by the discrete convolution of the information
sequence with the encoder "impulse responses'. The impulse responses are obtained by determining
the output sequences of the encoder produced by the input sequence u = (1, 0, 0, 0…).The impulse
responses so defined are called 'generator sequences' of the code. Since the encoder has a m-time
unit memory the impulse responses can last at most (m+ 1) time units (That is a total of (m+ 1) shifts
are necessary for a message bit to enter the shift register and finally come out) and are written as:

g (i) = {g1(i), g2(i),g3(i) …gm+l(i)}.

Fig 5.3 (2,1,3) binary encoder

For the encoder of Fig.5.3, we require the two impulse responses,

g (1) = {g1(1), g2(1), g3 (1), g4 (1)} and

g (2) = {g1(2), g2(2), g3 (2), g4 (2)}

By inspection, these can be written as: g (1) = {1, 0, 1, 1} and g (2) = {1, 1, 1, 1}

Observe that the generator sequences represented here is simply the 'connection vectors' of the

Dept. of ECE, ATMECE Page 9

Cyclic codes and Convolution codes 15EC54

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,
we get the overall impulse response of the encoder, Thus for the encoder of Fig 5.3, 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) ………………. (5.1 a)

v (2) = u * g (2) ……………. (5.1 b)

Where * denotes the „discrete convolution‟, which implies:

m
  ul  i . g i  1
( j) ( j)
vl
i 0
= ul g1
(j)
+ ul – 1 g2 (j) + ul – 2 g3 (j) + ….. +ul – m gm+1(j) ……………. (5.2)

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 5.3, we have:

vl(1) = ul + ul – 2 + ul - 3

vl(2) = 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 5.1:

Suppose the information sequence be u = (10111). Then the output sequences are:

v (1) = (1 0 1 1 1) * (10 1 1)
= (1 0 0 0 0 0 0 1),

v (2) = (1 0 1 1 1) * (1 1 1)
= (1 1 0 1 1 1 0 1),

and the code word is

v = (11, 01, 00, 01, 01, 01, 00, 11)

The discrete convolution operation described in Eq (8.2) is merely the addition of shifted

Dept. of ECE, ATMECE Page 10

Cyclic codes and Convolution codes 15EC54

impulses. Thus to obtain the encoder output we need only to shift the overall impulse response by
'one branch word', multiply by the corresponding input sequence and then add them. This is
illustrated in the table below:

INPUT OUT PUT

1 11 01 11 11
0 0 0 0 0 0 0 0 0 -----one branch word shifted
sequence
1 11 01 11 11 ---Two branch word
shifted
1 11 01 11 11
1 11 01 11 11
Modulo -2 sum 1 1 0 1 0 0 0 1 0 1 0 1 0 0 1 1

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)
g2
(2)
g3
(1)
g3
(2)
g4
(1)
g4
(2)
  
 
G ….. (5.3)
(1) (2) (1) (2) (1) (2) (1) (2)
g1 g1 g2 g2 g3 g3 g4 g4  
 (1) (2)
g1 g1 g2
(1)
g2
(2)
g3
(1)
g3
(2)
g4
(1)
g4
(2)

 
(Blank places are zeros.)

The encoding equations in Matrix form is:

v = u .G …………………. (5.4)

Example 5.2:

For the information sequence of Example 5.1, the G matrix has 5 rows and 2(3 +5) =16 columns and
we have

Dept. of ECE, ATMECE Page 11

Cyclic codes and Convolution codes 15EC54

1 1 0 1 1 1 1 1 0 0 0 0 0 0 0 0
0 0 1 1 0 1 1 1 1 1 0 0 0 0 0 0 

G  0 0 0 0 1 1 0 1 1 1 1 1 0 0 0 0
 
0 0 0 0 0 0 1 1 0 1 1 1 1 1 0 0
0 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1

Performing multiplication, v = u G as per Eq (5.4), 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

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
time and can be written as u = {u1 u1 , u2 u2 , u3 u3 …} or as two separate input
(1) (2) (1) (2) (1) (2)

sequences:
u (1) = {u1 (1), u2 (1), u3 (1) …} and u (2) = {u1 (2), u2 (2), u3 (2) …}.

Fig 5.4 (3,2,1) Convolution encoder

There are three generator sequences corresponding to each input sequence. Letting
gi ( j) = {gi,1 ( j), gi,2 ( j), gi,3 ( j) … gi,m+1 ( j)} represent the generator sequence corresponding to
input i and output j. The generator sequences for the encoder are:

g1 (1) = (1, 1), g1 (2) = (1, 0), g1 (3) = (1, 0)

g2 (1) = (0, 1), g2 (2) = (1, 1), g2 (3) = (0, 0)

The encoding equations can be written as:

v (1) = u (l) * g1 (1) + u (2)* g2 (1) …………………. (5.5 a)

v (2) = u (1) * g1 (2) + u (2) * g2 (2) …………………. (5.5 b)
v (3) = u (l) * g1 (3) + u (2) * g2 (3) ………………… (5.5 c)

The convolution operation implies that:

v l (1) = u l (1) + u l-1 (1) + u l-1 (2)

Dept. of ECE, ATMECE Page 12

Cyclic codes and Convolution codes 15EC54

v l (2) = u l (1) + u l (2) + u l-1 (2)

v l (3) = u l (1)
as can be seen from the encoding circuit.

After multiplexing, the code word is given by:

v = { v 1 ( 1) v 1 ( 2) v 1 ( 3) , v 2 ( 1) v 2 ( 2) v 2 ( 3) , v 3 ( 1) v 3 ( 2) v 3 ( 3) … }

Example 5.3:

Suppose u = (1 1 0 1 1 0). Hence u (1) = (1 0 1) and u (2) = (1 1 0). Then

v (1) = (1 0 1) * (1,1) + (1 1 0) *(0,1) = (1 0 0 1)

v (2) = (1 0 1) * (1,0) + (1 1 0) *(1,1) = (0 0 0 0)

v (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:

 g11( 1 ) g11( 2 ) g11( 3 ) (1)

g12 g12
(2)
g13
(3)

(1)
g1 ,m  1 g1 ,m  1
(2)
g1 ,m  1
(3) 
 (1) (2) (3) (1) (2) (3) (1) (2) (3) 
 g 21 g 21 g 21 g 22 g 22 g 22  g 2 ,m  1 g 2 ,m  1 g 2 ,m  1 
G 
(1)
g11 g11
(2)
g11
(3) (1)
g12 g12
(2)
g12
(3)
  …… (5.6)
 (1) (2) (3) (1) (2) (3)

  g 21 g 21 g 21 g 22 g 22 g 22  
 
     

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 5.4:

For the Example 5.3, we have

u = {u1 (1) u1 (2), u2 (1) u2 (2), u3 (1) u3 (2)} = (1 1, 0 1, 1 0)

The generator matrix is:

Dept. of ECE, ATMECE Page 13

Cyclic codes and Convolution codes 15EC54

1 1 1, 1 0 0 
0 1 0 , 1 1 0 
 
 1 1 1, 1 0 0 
G 
 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

m  Max ki ………………. (5.7)

1 i  k

(i.e. the maximum length of all k-shift registers)

An example of a (4, 3, 2) convolutional encoder in which the shift register lengths are 0, 1 and 2 is
shown in Fig 5.5.

Fig 5.5 (4,3,2) Binary convolution code encoder

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

n A  n(m  1) …………………… (5.8)

„nA‟ is called the 'constraint length" of the code. For example, the constraint lengths of the encoders
of Figures 5.3, 5.4 and 5.5 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

Dept. of ECE, ATMECE Page 14

Cyclic codes and Convolution codes 15EC54

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 (5.8).

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 5.3, 5.4 and 5.5 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 encoder‟s 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  Gm Gm  1 
 G1 G2  Gm  1 Gm Gm  1 
G   …………… (5.9)
 G1  Gm  2 Gm  1 Gm Gm  1 
 
      

Where each Gi is a (k  n) sub matrix with entries as below:

 g1 ,i ( 1 ) g1 ,i
(2)
 g 1 ,i
(n)
 (1) (2) (n )
Gi   2 ,i 
g g 2 ,i  g 2 ,i
………………… (5.10)
    
 (1) (2) (n)

 gk ,i g k ,i  gk ,i 

Notice that each set of k-rows of G are identical to the previous set of rows but shifted n-places to
the right. For an information sequence u = (u1, u2…) where ui = {ui (1), ui (2)…ui (k)}, the code word is
v = (v1, v2…) where vj = (vj (1), vj (2) ….vj (n)) 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 nm 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 clear the encoder memory. 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:

Dept. of ECE, ATMECE Page 15

Cyclic codes and Convolution codes 15EC54

k / n  kL / n( L  m ) m
 …………………….. (5.11)
k/n L m
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%.

5.3 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, ni) code, the encoding equations become:

v (1)(X) = u(X) g(l)(X) …………….. (5.12a)

v(2) (X) = u(X) g(2)(X) ……………..... (5.12b)

Where u(X) = u1 + u2X + u3X2 + … is the information polynomial,

v(1)(X) = v1(1) + v2(1)X + v3(1) X2 +....., and

v(2)(X) = v1(2) + v2(2)X + v3(2) X2 +.....

are the encoded polynomials.

g(1)(X) = g1(1) + g2(1) X + g3(1) X2 + ....., and

g(2)(X) = g1(2) + g2(2) X + g3(2) X2 + .....

are the “generator polynomials” of' the code; and all operations are modulo-2. After multiplexing, the
code word becomes:

v(X) = v(1)(X2) + X v(2)(X2) ………………… (5.13)

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.

Dept. of ECE, ATMECE Page 16

Cyclic codes and Convolution codes 15EC54

Example 5.5:

For the (2, 1, 3) encoder of Fig 8.3, the impulse responses were: g(1)= (1,0, 1, 1), and g(2) = (1,1, 1, 1)

The generator polynomials are: g(l)(X) = 1 + X2 + X3, and g(2)(X) = 1 + X + X2 + X3

For the information sequence u = (1, 0, 1, 1, 1); the information polynomial is: u(X) = 1+X2+X3+X4

The two code polynomials are then:

v(1)(X) = u(X) g(l)(X) = (1 + X2 + X3 + X4) (1 + X2 + X3) = 1 + X7

v(2)(X) = u(X) g(2)(X) = (1 + X2 + X3 + X4) (1 + X + X2 + X3) = 1 + X + X3 + X4 + X5 + X7

From the polynomials so obtained we can immediately write:

v(1) = ( 1 0 0 0 0 0 0 1), and v(2) = (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 (5.13) and write:

v (l) (X2) = 1 + X14 and v (2) (X2) = 1+X2+X6+X8+X10+X14; Xv (2) (X2) = X + X3 + X7 + X9 + X11 +
X15;

and the code polynomial is: v(X) = v (1) (X2) + X v (2) (X2) = 1 + X + X3 + X7 + X9 + X11 + X14 + X15

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.
Specifically, the co-efficient of Xl 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
(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 )  ……………… (5.14)

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:

Kl 
Max
1 j  n
deg gl
( j)

(X ) , 1 l  k ………… (5.15)

Where gl (j) (X) is the generator polynomial relating the l-th input to the j-th output and the encoder
memory order m is:

Dept. of ECE, ATMECE Page 17

Cyclic codes and Convolution codes 15EC54

Max
m
Max
1 l  k

K l  1  j  deg g l
( j)

(X) ………… (5.16)
1 l  k
Since the encoder is a linear system and u (l) (X) represents the l-th input sequence and v (j)
(X) represents the j-th output sequence the generator polynomial gl (j) (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 kn transfer functions which can be represented as a (k  n) "transfer
function matrix".
 g1( 1 ) ( X ), g1( 2 ) ( X ),  g1( n ) ( X )
 (1) (2) (n) 
 g 2 ( X ), g 2 ( X ),  g 2 ( X )
G( X )  …………… (5.17)
     
 (1) (2) (n)

 g k ( X ), g k ( X ),  g k ( 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) ………… (5.18)

U(X) = [u (1) (X), u (2) (X)...u (k) (X)] is the k-vector, representing the information polynomials, and.
V(X) = [v (1) (X), v (2) (X) … v (n) (X)] is the n-vector representing the encoded sequences. After
multiplexing, the code word becomes:

v(X) = v(1)(Xn) + X v(2)(Xn) +X2 v(3)(Xn)+…+ Xn-l v(n)(Xn) ………… (5.19)

Example 5.6:

For the encoder of Fig 5.4, we have:

g 1(1) (X) = 1 + X, g 2(1) (X) = X

g 1(2) (X) = 1, g 2(2) (X) =1+ X

g 1(3) (X) = 1 , g 2(3) (X) = 0

1  X 1 1
 G( X )  
 X 1  X 0
For the information sequence u (1) = (1 0 1), u (2) = (1 1 0), the information polynomials are:
u (1) (X) = 1 + X2, u(2)(X) = 1 + X

Then V(X) = [v (1) (X), v (2) (X), v (3) (X)]

1  X 1 1
= [1 + X2, 1 + X]   = [1 +X3, 0, 1+X2]
 X 1 X 0
Hence the code word is:

Dept. of ECE, ATMECE Page 18

Cyclic codes and Convolution codes 15EC54

v(X) = v(1)(X3) + Xv(2)(X3) + X2v(3)(X3)

= (1 + X9) + X (0) + X2(1 + X6)

= 1 + X2 + X8 + X9

 v = (1 0 1, 0 0 0, 0 0 1, 1 0 0).

This is exactly the same as that obtained in Example 5.3.From Eq (5.17) and (5.18) it follows that:

k
v ( j ) ( X )   u( i ) ( X ) gi ( X )
( j)

i 1
And using Eq (8.19) we have:
n
v( X )   X j  1v ( j ) ( x n )
j 1
n k
X j 1
 u( i ) ( X n ) gi
( j)
(X n)
j 1 i 1
k
 v(X)   u( i ) ( X n ) gi ( X ) …………………. (5.20)
i 1
n
Where gi ( X )   X j  1 gi ( j ) ( X n )
j 1

 gi ( X n )  Xgi ( X n )  X 2 gi ( X n )    X n 1 gi ( X n ) ………….. (5.21)

(1) (2) (3) (n)

is called the "composite generator polynomial" relating the i-th input sequence to v(X).

Example 5.7:

From Example 5.6, we have:

g1(X) = g1 (l) (X3) + Xg1 (2) (X3) + X2gl (3) (X3) = 1 + X + X2 + X3

g2(X) = g2 (1) (X3) + Xg2 (2) (X3) + X2g2(3) (X3)= X + X3 + X4

For the input sequence u (1) (X) = 1 + X2, u (2) (X) = 1 + X, we have

v (X) = u (1) (X3) g1(X) + u (2) (X3) g 2 (X) = 1+ X2 + X8 + X9. This is exactly the same as obtained
before.

5.4 Systematic Convolutional Codes:

In a systematic code, the first k-output sequences are exact replica of the k-input sequences
i.e.
v(i) = u (i) , i = 1,2, ...... k ………………………… (5.22)
and the generating sequences satisfy:
 i, j i  1,2 ,3.... k ………………….
( j)
gi (5.23)

Dept. of ECE, ATMECE Page 19

Cyclic codes and Convolution codes 15EC54

where δi, j is the Kronecker delta, having values: δi, j = 1 … if j = i

= 0 … if j ≠ i
The generator matrix for such codes is given by

I P1 O P2 O P3  O Pm  1 
  I P1 O P2  O Pm O Pm  1 
G  ……………. (5.24)
    I P1  O Pm 1 O Pm O
 
           

Where I is a k  k identity (unit) matrix, O is the k  k all zero matrix and Pi is a k  (n - k) matrix
given by:
 g 1 , i ( k  1 ) g1 , i ( k  2 )  g 1 , i ( n ) 
 ( k 1 ) ( k2 ) ( n )

Pi  
g 2 ,i g 2 ,i  g 2 ,i 
………………… (5.25)
    
 ( k 1 ) ( k2 ) (n)

 gk ,i g k ,i  gk ,i 
Further, the transfer function matrix of the code is given by:

1 0 0  0 g1
( k 1 )
( X )  g1( X )
(n)
 ( k 1 ) (n) 
0 1 0  0 g2 ( X )  g2 ( X )
0 0 1  0 g3
( k 1 )
( X )  g3 ( X )
(n)
G( X )    …………….. (5.26)
        
        
 ( k 1 )

0 0 0  1 gk ( X )  gk ( X )
(n)

The first k-output sequences = Input sequences  Information sequences

Last (n-k) sequences  parity sequences.

Number of sequences required to specify a general (n, k, m) code = kn.

For a systematic code we require only k  (n-k) sequences. Thus systematic codes form a sub class of
the set of all possible convolutional codes. Any code not satisfying Eq (5.22) to Eq (5.26) is said to
be "non systematic".

Example 5.8:

Consider a (2, 1, 3) systematic code whose encoder is shown in Fig 5.6.

Dept. of ECE, ATMECE Page 20

Cyclic codes and Convolution codes 15EC54

Fig 8.6 (2,1,3) systematic encoder

The generator sequences are: g (1) = (1 0 0 0) and g (2) = (1 1 0 1); and the generator matrix is:

1 1 0 1 0 0 0 1 
  1 1 0 1 0 0 0 1 
G 
    1 1 0 1 0 0 0 1
 
           

The transfer function matrix is:

G(X) = [1, 1 + X + X3]

For an input sequence u(X) = 1+X+X3, the information sequence is:

v(1)(X) = u(X) g(1)(X) =( 1 + X + X3)

and the parity sequence is:

v(2)(X) = u(X) g(2)(X) = (1+ X + X3) (1 + X + X3) = (1 + X2 + X6)

One advantage of systematic codes is that encoding is much simpler than for the non
systematic codes - because less hardware is required. For example, the (2, 1, 3) encoder of Fig 5.6
needs only one modulo-2 adder while that of Fig 5.5 requires three such adders. Notice also the total
number of inputs to the adders required. Further, for systematic (n, k, m) codes with K > n – k,
encoding schemes that normally require fewer than K-shift registers exist as illustrated in the
following simple example.

Example 5.9:

Consider a systematic (3, 2, 2) code with the transfer function matrix

1 0 1  X  X 2 
G( X )   
0 1 1 X2 

The straight forward realization requires a total of K=K1+K2=2+2=4 shift registers and is
shown in Fig 5.7(a). However, since the parity sequences are generated by:
v(3)(X) = u(1)(X). g1(3)(X) + u(2)(X) g2(3)(X), an alternative realization as shown in Fig 5.7(b) can be
obtained.

Dept. of ECE, ATMECE Page 21

Cyclic codes and Convolution codes 15EC54

Fig 5.7 Realization of encoder for example 5.9

In majority of situations, the straight forward realization is the most efficient. However, in the
case of systematic codes simpler realizations usually exist as shown in Example 5.9.

Another advantage of' systematic codes is that no inverting circuit is needed for recovering
information from the code word. For information recovery from a non systematic code, inversion is
required in the form of an (n  k) matrix G -1(X) such that
G(X).G -1(X) = Ik X l ……………… (5.27)

for some l  0 and Ik is the (k  k) unit matrix. Then it follows that:

V(X).G -1(X) = U(X) G(X) G -1(X) = U(X).X l …………… (5.28)

and the information sequence can be recovered with an l-time unit delay from the code word by
letting V(X) to be the input to the n-input, k-output linear sequential circuit whose transfer function
matrix is G – 1(X).

For an (n, l, m) code the transfer function matrix G(X) will have a "feed forward" inverse
G – 1 (X) of delay l units if and only if :

G.C.D [g (1) (X), g (2) (X) …g (n) (X)] = X l ……………. (5.29)

for some l  0; where G.C.D denotes the 'greatest common divisor'. For an (n, k, m) code with k >
 n  n
1, let i (X), i=1, 2 …   be the determinants of the   distinct k  k sub-matrices of the transfer
k k
function matrix G(X).Then a feed forward inverse of delay l-units exists if an only if:
 n
GGD [i (X), i=1, 2 …   ] = X l ……………… (5.3
k

Dept. of ECE, ATMECE Page 22

Cyclic codes and Convolution codes 15EC54

Example 5.10:

For the (2, 1, 3) encoder of Fig 5.3, we have, from Example 8.5, the generator matrix as

G(X) = [1 + X2 + X3, 1 +X + X2 + X3)

Its inverse can be computed as:

1  X  X 2 
G 1 ( X )   2 
 XX 

and the implementation of the inverse is shown in Fig 5.8.

Example 5.11:

For the (3, 2, 1) encoder of Fig 5.4, the generator matrix as found in Example 5.6 is:

1  X 1 1
G( X )  
 X 1 X 0

The determinants of the (2  2) sub matrices are 1+ X + X2, X and 1 + X. Their GCD is 1.
A feed forward inverse with no delay exists and can be computed as:

 1 X X 
1 
G (X )   X 1  X  I
 X  X 2 1  X 2 

Implementation of this inverse is shown in Fig 5.9.

Fig 5.8 Feed forward encoder of (2,1,3) code Fig 5.9 Feed forward encoder of (3,2,1) code

To understand what happens when a feed forward inverse does not exist consider an example
of a (2, 1, 2) encoder with generator matrix

G(X) = [1+X, 1 + X2]

Since the GCD of g (1) (X) and g (2) (X) is (1+ X) (not of the form X l), a feed forward inverse does
not exist. Suppose the input sequence is:

Dept. of ECE, ATMECE Page 23

Cyclic codes and Convolution codes 15EC54

1
u( X )   1  X 2  X 3  ... Then the output sequences are: v(1)(X) = 1, v(2)(X) = 1 + X.
1 X
That is, the code word contains only three nonzero bits even though the information sequence has
infinite weight. If this code word is transmitted over a BSC and the three nonzero bits are changed to
zeros by the channel noise, the received sequence will be all zeros. A maximum likely hood decoder
(MLD) will then produce the all-zero code word as its estimate, since this a valid code word and it
agrees exactly with the received sequence. Thus, the estimated information sequence will be
uˆ ( X )  0 , implying an infinite number of decoding errors caused by a finite number (only three in
this case) of channel errors. Clearly this is a very undesirable situation and the code is said to be
subject to "Catastrophic error propagation" and the code is called a "catastrophic code".
Equations (5.29) and (5.30) can be shown to be necessary and sufficient conditions for a code
to be 'non-catastrophic'. Hence any code for which a feed forward inverse exists is non-catastrophic.
Another advantage of systematic codes is that they are always non-catastrophic.

8.5 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   K i as the total encoder -
i 1
memory (m - represents the memory order which we have defined as the maximum length of any
shift register), the encoder state at time unit T', when the encoder inputs are, {u l (1), u l (2)…u l (k)}, 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), u l-2(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 2k different possible states. For a (n, 1, m) code, K = K1 = m and the encoder
state at time unit l is simply {ul-1, ul-2 … ul-m}.
Each new block of k-inputs causes a transition to a new state. Hence there are 2k 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 5.3

State S0 S1 S2 S3 S4 S5 S6 S7
Binary
000 100 010 110 001 101 011 111
Description

Dept. of ECE, ATMECE Page 24

Cyclic codes and Convolution codes 15EC54

Fig 5.10 State diagram of encoder of Fig 5.3

Recall (or observe from Fig 8.3) that the two out sequences are:

v (1) = ul + ul – 2 + ul – 3 and
v (2) = ul + ul – 1 + ul – 2 + ul – 3

Till the reader, gains some experience, it is advisable to first prepare a transition table using the
output equations and then translate the data on to the state diagram. Such a table is as shown below:

State transition table for the encoder of Fig 5.3

Previou Binary Input Next Binary ul ul – 1 ul – 2 ul - 3 Output
s Descriptio State Descriptio
State n n
S0 0 0 0 0 S0 0 0 0 0 0 0 0 0 0
1 S1 1 0 0 1 0 0 0 1 1
S1 1 0 0 0 S2 0 1 0 0 1 0 0 0 1
1 S3 1 1 0 1 1 0 0 1 0
S2 0 1 0 0 S4 0 0 1 0 0 1 0 1 1
1 S5 1 0 1 1 0 1 0 0 0
S3 1 1 0 0 S6 0 1 1 0 1 1 0 1 0
1 S7 1 1 1 1 1 1 0 0 1
S4 0 0 1 0 S0 0 0 0 0 0 0 1 1 1
1 S1 1 0 0 1 0 0 1 0 0
S5 1 0 1 0 S2 0 1 0 0 1 0 1 1 0
1 S3 1 1 0 1 1 0 1 0 1
S6 0 1 1 0 S4 0 0 1 0 0 1 1 0 0
1 S5 1 0 1 1 0 1 1 1 1
S7 1 1 1 0 S6 0 1 1 0 1 1 1 0 1
1 S7 1 1 1 1 1 1 1 1 0

Dept. of ECE, ATMECE Page 25

Cyclic codes and Convolution codes 15EC54

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 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-all-
zero block appended to the information sequence. For example, in Fig 5.10, 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.

Apart from obtaining the output sequence for a given input sequence, the state diagram can be
modified to provide a complete description of the Hamming weights of all nonzero code words. (That
is, the state diagram is useful in determining a weight- distribution for the code).

This is achieved as follows: The state S0 is split into an initial and a final state. The self loop
around S0 is discarded. Each branch is labeled with a 'branch gain', „X i‟, where 'i' is the weight
(number of ones) of the n-encoded bits on that branch. Each path that connects the initial state to the
final state which diverges from and remerges with state S0, exactly once, represents a nonzero code
word.

Those code words that diverge from and remerge with S0 more than once can be regarded as a
sequence of shorter code words. The "path gain" is the product of the branch gains along a path and
the weight of the associated code word is the power of X in the path gain. As an lustration, let us
consider the modified state diagram for the (2, 1, 3) code of Fig 5.3 as shown in Fig 5.11 and another
version of the same as shown in Fig 5.12.

Fig 5.11 Modified state diagram for the (2,1,3) code

Example 5.12: A (2, 1, 2) Convolutional Encoder:

Consider the encoder shown in Fig 5.15. We shall use this example for discussing further
graphical representations viz. Trees, and Trellis.

Dept. of ECE, ATMECE Page 26

Cyclic codes and Convolution codes 15EC54

Fig 5.15 (2,1,2) convolution encoder

For this encoder we have: v l (1) = u l + u l – 1 + u l – 2 and v l (2) = u l + u l – 2

The state transition table is as follows.

State transition table for the (2, 1, 2) convolutional encoder of Example 5.12
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 1 0 1 0 0 1 1
S1 1 0 0 S2 0 1 0 1 0 1 0
1 S3 1 1 1 1 0 0 1
S2 0 1 0 S0 0 0 0 0 1 1 1
1 S1 1 0 1 0 1 0 0
S3 1 1 0 S2 0 1 0 1 1 0 1
1 S3 1 1 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 5.16.

Fig 5.16 State diagram for the (2,1,2) encoder

Dept. of ECE, ATMECE Page 27

Cyclic codes and Convolution codes 15EC54

There are three loops in the augmented state diagram:

S3 → S3: l1 = DLI, S1→ S2 → S1: l2 = DL2I, S1 → S3 → S2 → S1:l3 = D2L3I2

The loops l1 and l2 are non-touching and their gain product is: l1l2 = D2L3I2

  = 1 – (l1 + l2 + l3) – l1 l2

= 1 – DLI (1 + L)

There are two forward paths: F1  S0 → S1 → S2→ S0. Path gain = D5L3I

F2  S0 → S1 → S3→ S2 → S0, Path gain = D6L4I2

The loop l1 does not touch the forward path F1.  1 = 1 – l1 = 1 – DLI.

All the three loops touch the forward path F2.  2 = 1

Now use Mason‟s gain formula to get:

D 5 L3 I ( 1  DLI )  D6 L4 I 2 D 5 L3 I
T ( D , L, I )  
1  DLI ( 1  L ) 1  DLI ( 1  L )
 D 5 L3 I  D6 L4 I 2 ( 1  L )  D7 L5 I 3 ( 1  L )2  ...

Thus there is one code word of weight 5 that has length 3 branches and an information
sequence of weight 1, Two code words of weight 6, of which one has length 4 branches and an
information sequence of weight 2, and the other has length 5 branches and an information sequence
of weight 2 and so on.

8.6 TREE AND TRELLIS DIAGRAMS:

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 5.15 is shown in Fig 5.18. The state transition
table can be conveniently used in constructing the tree graph.

Dept. of ECE, ATMECE Page 28

 Fig 5.18: The tree graph for the (2, 1, 2) encoder of Fig 5.15

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
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

29
"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
5.19 and the complete encoded sequence is (11, 10, 11, 11, 01, 01, 11).

Fig 5.19 Illustration of the “Tail of the tree”

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
number of right most nodes increase as 2L. Hence for a specified sequence length, L, compute 2L.
Mark 2L equally spaced points at the rightmost portion of your page, leaving space to complete the m
tail branches. Join two points at a time to obtain 2L-l 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 5.20 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.

Fig 5.20 Procedure for drawing neat tree diagram

From Fig 5.18, observe that the tree becomes "repetitive' after the first three branches.

30
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 (100u3u4---) 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 5.18 into a new form of Fig 5.21 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).

Fig 5.21 Trellis diagram for encoder of fig 5.15

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 5.15 as shown in Fig 5.21.
The following observations can be made from the Trellis diagram

1. There are no fundamental paths at distance 1, 2 or 3 from the all zero path.

2. There is a single fundamental path at distance 5 from the all zero path. It diverges from the all-zero
path three branches back and it differs from the all-zero path in the single input bit.

3. There are two fundamental paths at a distance 6 from the all zero path. One path diverges from the
all zero path four branches back and the other five branches back. Both paths differ from the all zero
path in two input bits. The above observations are depicted in Fig 8.24(a).

4. There are four fundamental paths at a distance 7 from the all-zero path. One path diverges from the
all zero path five branches back, two other paths six branches back and the fourth path diverges seven
branches back as shown in Fig 8.24(b). They all differ from the all zero path in three input bits. This
information can be compared with those obtained from the complete path enumerator function found
earlier.

31
8.7 THE VITERBI ALGORITHM:

The Viterbi algorithm, when applied to the received sequence r from a DMC finds the path
through the trellis with the largest metric. At each step, it compares the metrics of all paths entering
each state and stores the path with the largest metric called the "survivor" together with its metric.

The Algorithm:

Step: 1. Starting at level (i.e. time unit) j = m, compute the partial metric for the single path
entering each node (state). Store the path (the survivor) and its metric for each state.
Step: 2. Increment the level j by 1. Compute the partial metric for all the paths entering a
state by adding the branch metric entering that state to the metric of the connecting
survivor at the preceding time unit. For each state, store the path with the largest
metric (the survivor), together with its metric and eliminate all other paths.
Step: 3. If j < (L + m), repeat Step 2. Otherwise stop.

Notice that although we can use the Tree graph for the above decoding, the number of
nodes at any level of the Trellis does not continue to grow as the number of incoming message bits
increases, instead it remains a constant at 2m.

There are 2k survivors from time unit „m‟ up to time unit L, one for each of the 2kstates. After
L time units there are fewer survivors, since there are fewer states while the encoder is returning to
the all-zero state. Finally, at time unit (L + m) there is only one state, the all-zero state and hence
only one survivor and the algorithm terminates.

Fig 5.22 Survivor after time unit „j‟

Suppose that the maximum likely hood path is eliminated by the algorithm at time unit j as
shown in Fig 8.22. This implies that the partial path metric of the survivor exceeds that of the
maximum likely hood path at this point. Now, if the remaining portion of the maximum likely hood
path is appended onto the survivor at time unit j, then the total metric of this path will exceed the total
metric of the maximum likely hood path. But this contradicts the definition of the 'maximum likely
hood path' as the 'path with largest metric'. Hence the maximum likely hood path cannot be
eliminated by the algorithm and it must be the final survivor and it follows

32
 M (r | vˆ )  M (r | v ),  v  vˆ .Thus it is clear that the Viterbi algorithm is optimum in the sense that it
always finds the maximum likely hood path through the Trellis. From an implementation point of
view, however, it would be very inconvenient to deal with fractional numbers. Accordingly, the bit
metric M (ri|vi) = ln P (ri|vi) can be replaced by “C2 [ln P (ri|vi) + C1]”, C1 is any real number and C2
is any positive real number so that the metric can be expressed as an integer. Notice that a path v
which maximizes M (r | v )   M ( ri | vi )   ln P ( ri | vi ) also maximizes  C2 ln P ( ri | vi )  C1  .
N N N

i 1 i 1 i 1
Therefore, it is clear that the modified metrics can be used without affecting the performance of the
Viterbi algorithm. Observe that we can always choose C1 to make the smallest metric as zero and
then C2 can be chosen so that all other metrics can be approximated by nearest integers. Accordingly,
there can be many sets of integer metrics possible for a given DMC depending on the choice of C2.
The performance of the Viterbi algorithm now becomes slightly sub-optimal due to the use of
modified metrics, approximated by nearest integers. However the degradation in performance is
typically very low.

Example 5.13:

As an illustration let us consider a binary input-quaternary output DMC shown in Fig 5.23(a).
The bit metrics ln P (ri| vi) are shown in Fig 5.23(b). Choosing C1= − 2.3 and C2 = 7.195 yields the
"integer metric table" shown in Fig 5.23(c).

Fig 5.23 Diagran for example 5.13

Now suppose that a code word from the (2,1,2) encoder of Fig 5.15, whose Trellis diagram is
shown in Fig 5.21, is transmitted over the DMC of Fig 8.26 and the quaternary received sequence is:

r = {y3 y4, y3 y1, y3 y2, y3 y4, y3 y4, y2 y4, y1 y3}

Let us apply Viterbi algorithm to determine the transmitted sequence.

In the first time unit (j = 1) there are two branches originating from the state S0 with output
vectors (00) terminating at S0 and (11) terminating at S1. The received sequence in this time unit is
(y3 y4) and using the integer metric table of Fig 8.23(c) we have:

33
 M [r1|v1 (1)] = M (y3 y4|00) = M (y3|0) + M (Y4|0) = 5 + 0 = 5, and

M [r1|v1 (2)] = M (y3 y4|11) = M (y3|1) + M (Y4|1) = 8 + 10 = 18

These computations are indicated in Fig 8.24(a). The path discarded is shown by a cross. Note that
the branch metrics are also indicated along the branches with in brackets and the state metrics are
indicated at the nodes.

For j = 2 there are single branches entering each state and the received sequence in this time
unit is (y3 y1). The four branch metrics are computed as below.

M1 = M (y3 y1|00) = M (y3|0) + M (y1|0) =5 + 10 =15

M2 = M (y3 y1|11) = M (y3|1) + M (y1|1) =8 + 0 =8

M3= M (y3 y1|10) = M (y3|1) + M (y1|0) =8 + 10 =18

M4 = M (y3 y1|01) = M (y3|0) + M (y1|1) =5 + 0 =5

The metrics at the four states are obtained by adding the branch metrics to the metrics of the previous
states (survivors) and are shown in Fig 8.24(b).

Fig 5.24 Computation for time units j=1, j=2 and j=3

Next for j = 3, notice that there are two branches entering each state as shown in Fig 8.24(c).
The received sequence in this time unit is (y3, y2) and the branch metrics are computed as below:

M1 = M (y3 y2|00) = M (y3|0) + M (y2|0) =5 + 8 =13

M2 = M (y3 y2|11) = M (y3|1) + M (y2|1) =8 + 5 =13

M3 = M (y3 y2|10) = M (y3|1) + M (y2|0) =8 + 8 =16

M4 = M (y3 y2|01) = M (y3|0) + M (y2|1) =5 + 5 =10

Following the above steps, we arrive at the following diagram.

34
 Fig 5.25 Application of Viterbi algorithm
Notice that, in the last step we have ignored the highest metric computed! Indeed, if the
sequence had continued we should take this into account. However, in the last m-time units
remember that the path must remerge with S0.

From the path that has survived, we observe that the transmitted sequence is:

v̂ = (11, 10, 11, 11, 01, 01, 11)

and the information sequence at the encoder input is: û = (1 0 0 1 1)

Notice that “the final m-branches in any trellis path always corresponds to „0‟ inputs and
hence not considered part of the information sequence”.

As already mentioned, the MLD reduces to a 'minimum distance decoder' for a BSC (see
Eq 8.40). Hence the distances can be reckoned as metrics and the algorithm must now find the path
through the trellis with the smallest metric (i.e. the path closest to r in Hamming distance). The
details of the algorithm are exactly the same, except that the Hamming distance replaces the log
likely hood function as the metric and the survivor at each state is the path with the smallest metric.
The following example illustrates the concept.

Example 5.14:

Suppose the rode word r = (01, 10, 10, 11, 01, 01, 11), from the encoder of Fig 5.15 is received
through a BSC. The path traced is shown in Fig 5.25 as dark lines.

35
 Fig 5.26 Viterbi algorithm for a BSC

The estimate of the transmitted code word is

v̂ = (11, 10, 11, 11, 01, 01, 11)

and the corresponding information sequence is: û = (1 0 0 1 1)

Notice that the distances of the code words of each branch with respect to the corresponding
received words are indicated in brackets. Note also that at some states neither path is crossed out
indicating a tie in the metric values of the two paths entering that state. If the final survivor goes
through any of these states there is more than one maximum likely hood path (i.e. there may be more
than one path whose distance from r is a minimum). From an implementation point of view whenever
a tie in metric values occur, one path is arbitrarily selected as survivor, because of the non-
practicability of storing a variable number of paths. However, this arbitrary resolution of ties has no
effect on the decoding error probability. Finally, the Viterbi algorithm cannot give fruitful results
when more errors in the transmitted code word than permissible by the dfree of the code occur. For the
example illustrated, the reader can verify that the algorithm fails if there are three errors. Discussion
and details about the performance bounds, convolutional code construction, implementation of the
Viterbi algorithm etc are beyond the scope of this book.

RECOMMENDATION QUESTIONS
1. Fig P 8.1 shows a convolutional encoder. i) What is the constraint length and rate efficiency?
ii) Find the encoder output produced by the sequence 101101…….. iii) Is the code systematic?

36
2. Consider the convolutional encoder shown in Fig P 8.2. The message bits are shifted into the
encoder two bits at a time. Repeat the questions asked in problem P 8.1.

3. A convolution encoder has a single shift register with two stages (i.e. m=2), three Modulo-2
adders and Output multiplexer. The generator sequences of the encoder are as follows:
g(1)=(1,0,1); g(2)=(1,1,0) ; g(3)=(1,1,1)

Draw the block diagram of the encoder.

4. Fig P8.4 shows the encoder of a rate 1/2, constraint length =4 convolutional encoder.
Determine the encoder out-put produced by the information sequence (10111…) using the following
two approaches: i) Time domain approach, based on convolution. ii) Transform domain approach.

5. For the encoder of problem 3,

a) Find the generator matrix G
b) Find the code word corresponding to the information sequence (11101…)

6. For the encoder of problem 3,

a) Find the transfer function matrix G(X)
b) Find the set of output sequences V(X) and the code word v(X) corresponding to the
information sequence u(X)=1=X2+X3+X4.

7. Determine which of the following rate 1/2 convolutional codes are „catastrophic‟:
(a) g (1) (X) = X2, g (2) (X) = 1 + X + X3
(b) g (1) (X) =1 + X2 + X4, g (2) (X) = 1 + X + X3+X4
(c) g (1) (X) =1 + X + X2 + X4, g (2) (X) = 1 + X3+X4
(d) g (1) (X) =1 + X4 + X5 + X6, g (2) (X) = 1 + X + X3+X5

37
8. Consider the encoder of problem 3,
a) Draw the state diagram of the encoder
b) Draw the modified state diagram
c) Find the generating function T(X)
d) Draw the augmented state diagram
e) Find the Complete Path Enumerator function T (D, L, I)

8. Consider the (3, 1, 5) systematic code with g (2) = (101101), g (3) = (110011)
a) Find the generator matrix
b) Find the parity sequences corresponding to the information sequence
u= (1101…)

9. Consider the (3, 2, 3) systematic code with

g1 (3) (X) = 1+X2+X3 and g2 (3) (X) = 1+X+X3
a) Draw the straight forward realization of the encoder
b) Draw a simple encoder realization which requires only three shift register stages.

10. Consider the (2, 1, 2) code with G(X) = [1+X2, 1+X+X2]

a) Find the GCD of its generator polynomials
b) Find the transfer function matrix G-1(X) of its minimum delay feed forward
inverse.

11. Consider a (2, 1, 3) code with G(X) = [1+X2, 1+X+X2 +X3]

a) Find the GCD of its generator polynomials
b) Draw the encoder state diagram.
c) Find the infinite-weight information sequence that generates a code word of finite
weight
d) Is this code catastrophic?

12. Construct the code tree for the convolutional encoder of Fig P 8.1. Trace the path through the tree
that corresponds to the information sequence (101101…) and compare the output with that
determined in problem 21.

2. The code tree for the encoder of Fig 8.15, assuming that the incoming message sequence has L =
2 is shown in Fig P 8.13.Validate this tree.

38
14. Construct the code tree for the encoder of Fig P 8.4. Trace the path through the tree that
corresponds to the information sequence (10111…). Compare the resulting encoder output with
that found in problem 4.

15. Draw the state diagram and augmented state diagram for the encoder of Fig P.8.4.
i) Show that generating function is
X6  X7  X8
T(X) =
1  2X  X 8
ii) What is the free distance of this code? How many errors it can correct?
iii) Find the path enumerator function T (D, L, I)

16. Construct the Trellis diagram for the encoder of Fig P.8.4, assuming a message sequence of
length 5. Trace the path through the Trellis diagram corresponding to the message sequence
(10111…). Compare the resulting encoder output with that found in problem 4.

17. Consider the encoder of Fig P 8.17.

i) What is its rate efficiency and constraint length?

ii) Using the transform domain approach, determine the code sequence produced by the
information sequence (10111…).
iii) Construct the state diagram.
iv) Starting from the all zero state, trace the path that corresponds to the information sequence
(10111…). Compare your answer with that in part (ii).
18. For a (3, 1, 2) code the transfer function matrix is: G(X) = [1+X, 1+X2, 1+X+X2]

a) Draw the Trellis diagram for an information sequence of length L = 5.

b) Trace the path and write the code word corresponding to information sequence
(11101).

19. Find the integer metric table for the DMC of Fig 8.26(a) with C1 = -2.3 and C2 = 3.0.
Decode the receiver sequence for the encoder of problem 18:
r = (y3y4y1, y4y4y2, y4y4y1, y4y4y4, y1y3y1, y3y2y4, y3y1y4) using the Viterbi algorithm
Also decode the same sequence using the integer metric table of Fig 8.26(c). Compare the results.

20. Consider a binary input 8-ary output DMC with transition probabilities P (ri|vi) given by the
following table.

39
 Find the metric table and integer table for this channel.

21. Consider the (2, 1, 3) code with G(X) = [1+X2 +X3, 1+X+X2 +X3]
a) Draw the trellis diagram for an information sequence of length L = 4.
b) Assume code vector is transmitted over the DMC of problem 20.
Decode the received sequence: r = [y7y8, y7y1, y3y1, y1y6, y7y2, y3y8, y3y2]

22. The DMC of problem 20 is converted to BSC by combining the soft decision outputs
y1, y2, y3 and y4 into a single hard decision output „0‟ and combining the soft decision outputs
y5, y6, y7 and y8 into a single hard decision output „1‟. A code word from code of problem 21 is
transmitted over this channel. The hard decision version of the received sequence is:

r = [11, 10, 00, 01, 10, 01, 00]

Decode this sequence and compare with the result of problem 21.

23. For a rate 1/3 systematic code: g(1,1) = (1,0,1) and g(1,2) = (1,1,1)
a) Draw the tree graph, Trellis and state diagram.
b) Find dfree and„t‟ for the code.
c) For the information sequence u = (00110100….) find v.
d) If the received vector is
r = {001, 110, 110, 010, 100, 001, 011, 000}
Find the transmitted u using Viterbi algorithm.

24. Repeat the problem 23. For a systematic code if g (1, 1) = (1, 0, 1, 1) and g (1, 2) = (1, 1, 0, 1)

25. For a non systematic rate 1/2 code given by: g(1,1) = (1, 1,1) and g(1,2) = (1, 0,1,) Repeat parts
(a) and (b) of problem 23 for this code. Show, by an example, that the code corrects 2 errors in six
channel bits.

26. A rate 1/3 non systematic code is given by the sub generators
g(1,1) = (1,1,0 ,1), g(1,2) = (1,0 ,0,1) and g(3,1) = ( 1,1,1,0)
a) Construct the coder
b) Draw the tree graph, Trellis and state diagrams
c) Find dfree and t.
d) Tabulate the survivor paths and their Hamming distances for a given error vector e = (001,
010, 101, 000)

27. Consider the (2, 1, 3) code of problem 21.

a) Draw the code tree for an information sequence of length L= 4.
b) Find the code word corresponding to an information sequence u = (1001).

40
28. Consider the (2, 1, 3) code of problem 21.
a) For a BSC with p =0.045, find an integer metric table for the Fano metric.
b) Decode the received sequence: r = (11, 00, 11, 00, 01, 10, 11) using the stack algorithm.
Compare the number of decoding steps with the number required by the Viterbi algorithm.
c) Repeat part (b) for the received sequence r = (11, 10, 00,01,10,01,00)

29. Consider again the (2, 1, 3) code of problem 21.

a) For the binary input-8-ary-output DMC of problem 20, find an integer metric table for the
Fano metric.(Hint: Use appropriate scale factor for each metric and round to the nearest
integer).
b) Decode the received sequence: r = [y7y8, y7y1, y3y1, y1y6, y6y2, y3y8, y3y2] using the Stack
algorithm. Compare the final decoded path with the result of problem 21(b) where the same
received sequence is decoded using Viterbi algorithm.

30. Repeat problem 29 using Fano algorithm with threshold increments of  = 5 and
 = 9.Compare the final decoded path and the number of computations with the
Stack algorithm.

31. Refer to the code tree of Fig P 8.13. The branch metrics are specified as below.
 1 if r j  v j and ri  vi


M ( r j ri | v j vi )   4 if r j  v j and ri  vi or r j  v j and ri  vi

 9 if r j  v j and ri  vi

Calculate the path metrics at the nodes of the tree for an all zero transmitted sequence, and the
received sequence (10, 00, 01, 01) with three transmissions errors. Then apply Fano algorithm to
decode the received sequence under the assumptions:
a) The decoder chooses the lower branch in the case of a tie in path metrics.
b) The decoder chooses the upper branch in the case of a tie in path metrics.

OUTCOMES

 To know the encoding of convolutional codes.

 How to decode the convolutional codes using algorithm.

REFERENCE

 www.youtube.com/watch?v=AnyVu5eDhAQ
 nptel.ac.in/courses/117106031/
 elearning.vtu.ac.in/P4/EC63/S11.pdf

41