Module

Channel Coding

Version 2 ECE IIT, Kharagpur

 Lesson
35

Convolutional Codes

Version 2 ECE IIT, Kharagpur

After reading this lesson, you will learn about
¾ Basic concepts of Convolutional Codes;
¾ State Diagram Representation;
¾ Tree Diagram Representation;
¾ Trellis Diagram Representation;
¾ Catastrophic Convolutional Code;
¾ Hard - Decision Viterbi Algorithm;
¾ Soft - Decision Viterbi Algorithm;

Convolutional codes are commonly described using two parameters: the code rate
and the constraint length. The code rate, k/n, is expressed as a ratio of the number of bits
into the convolutional encoder (k) to the number of channel symbols output by the
convolutional encoder (n) in a given encoder cycle. The constraint length parameter, K,
denotes the "length" of the convolutional encoder, i.e. how many k-bit stages are
available to feed the combinatorial logic that produces the output symbols. Closely
related to K is the parameter m, which indicates how many encoder cycles an input bit is
retained and used for encoding after it first appears at the input to the convolutional
encoder. The m parameter can be thought of as the memory length of the encoder.

Convolutional codes are widely used as channel codes in practical communication

systems for error correction. The encoded bits depend on the current k input bits and a
few past input bits. The main decoding strategy for convolutional codes is based on the
widely used Viterbi algorithm. As a result of the wide acceptance of convolutional codes,
there have been several approaches to modify and extend this basic coding scheme.
Trellis coded modulation (TCM) and turbo codes are two such examples. In TCM,
redundancy is added by combining coding and modulation into a single operation. This is
achieved without any reduction in data rate or expansion in bandwidth as required by
only error correcting coding schemes.

A simple convolutional encoder is shown in Fig. 6.35.1. The information bits are
fed in small groups of k-bits at a time to a shift register. The output encoded bits are
obtained by modulo-2 addition (EXCLUSIVE-OR operation) of the input information
bits and the contents of the shift registers which are a few previous information bits.

Version 2 ECE IIT, Kharagpur

 C1

x(1) D D

+
C2

Fig. 6.35.1 A convolutional encoder with k=1, n=2 and r=1/2

If the encoder generates a group of ‘n’ encoded bits per group of ‘k’ information
bits, the code rate R is commonly defined as R = k/n. In Fig. 6.35.1, k = 1 and n = 2. The
number, K of elements in the shift register which decides for how many codewords one
information bit will affect the encoder output, is known as the constraint length of the
code. For the present example, K = 3.

The shift register of the encoder is initialized to all-zero-state before encoding

operation starts. It is easy to verify that encoded sequence is 00 11 10 00 01 ….for an
input message sequence of 01011….

The operation of a convolutional encoder can be explained in several but

equivalent ways such as, by a) state diagram representation, b) tree diagram
representation and c) trellis diagram representation.

a) State Diagram Representation

A convolutional encoder may be defined as a finite state machine. Contents of the
rightmost (K-1) shift register stages define the states of the encoder. So, the encoder in
Fig. 6.35.1 has four states. The transition of an encoder from one state to another, as
caused by input bits, is depicted in the state diagram. Fig. 6.35.2 shows the state diagram
of the encoder in Fig. 6.35.1. A new input bit causes a transition from one state to
another. The path information between the states, denoted as b/c1c2, represents input
information bit ‘b’ and the corresponding output bits (c1c2). Again, it is not difficult to
verify from the state diagram that an input information sequence b = (1011) generates an
encoded sequence c = (11, 10, 00, 01).

Version 2 ECE IIT, Kharagpur

 10
1/11 1/01
0/10

00 11
1/10
0/00 1/00
0/11
0/01

01

Fig.6.35.2 State diagram representation for the encoder in Fig. 6.35.1

b) Tree Diagram Representation

The tree diagram representation shows all possible information and encoded
sequences for the convolutional encoder. Fig. 6.35.3 shows the tree diagram for the
encoder in Fig. 6.35.1. The encoded bits are labeled on the branches of the tree. Given an
input sequence, the encoded sequence can be directly read from the tree. As an example,
an input sequence (1011) results in the encoded sequence (11, 10, 00, 01).

Version 2 ECE IIT, Kharagpur

 0 0
0 0
1 1
0 0
1 0
1 1
0 1
0 0
0 1 1
1 0
0 0
1 1
0 1
0 1
1 0
0 0
1 1
1 1
1 0
1 0
0 0
0 1
1 1 1 1
1
0 1
0 0
0 1
0 1
1 0
1 0
t=0 t=1 t=2 t=3 t=4

Fig. 6.35.3 A tree diagram for the encoder in Fig. 6.35.1

c) Trellis Diagram Representation

The trellis diagram of a convolutional code is obtained from its state diagram. All
state transitions at each time step are explicitly shown in the diagram to retain the time
dimension, as is present in the corresponding tree diagram. Usually, supporting
descriptions on state transitions, corresponding input and output bits etc. are labeled in
the trellis diagram. It is interesting to note that the trellis diagram, which describes the
operation of the encoder, is very convenient for describing the behavior of the

Version 2 ECE IIT, Kharagpur

 corresponding decoder, especially when the famous ‘Viterbi Algorithm (VA)’ is
followed. Figure 6.35.4 shows the trellis diagram for the encoder in Figure 6.35.1.

t1 00 t2 t3 t4 t5 t6
00 00 00 00
State a = 00
11 11 11 11 11 Codeword
branch
11 11 11
b = 00 00 00 00

10 10 10 10
c = 00
01 01 01 01 01 01 01

d = 00
Legend 10 10 10
Input bit 0
Input bit 1

Fig.6.35.4(a) Trellis diagram for the encoder in Fig. 6.35.1

Input data sequence m 1 1 0 1 1 …

Transmitted codeword U: 11 01 01 00 01 …

Received sequence Z: 11 01 01 10 01 ….
t1 2 t2 1 t3 1 t4 1 t5 1 t
State a = 00
0 1 1 1 1 1 1 1
Branch
metric
b = 00 1 1
1
2 2 0 2

c = 00 0 …
0 2 0
0 2
0
d = 00 2 0 2

Fig.6.35.4(b) Trellis diagram, used in the decoder corresponding to the encoder in Fig.
6.35.1

Version 2 ECE IIT, Kharagpur

Catastrophic Convolutional Code
The taps of shift registers used for a convolutional encoder are to be chosen
carefully so that the code can effectively correct errors in received data stream. One
measure of error correction capability of a convolutional code is its ‘minimum free
distance, dfree’, which indicates the minimum weight (counted in number of ‘1’-s) of a
path that branches out from the all-zero path of the code trellis and again merges with the
all-zero path. For example, the code in Fig.6.35.1 has dfree= 5. Most of the convolutional
codes of present interest are good for correcting random errors rather than correcting
error bursts. If we assume that sufficient number of bits in a received bit sequence are
error free and then a few bits are erroneous randomly, the decoder is likely to correct
these errors. It is expected that (following a hard decision decoding approach, explained
later) the decoder will correct up to (dfree-1)/2 errors in case of such events. So, the taps of
a convolutional code should be chosen to maximize dfree. There is no unique method of
finding such convolutional codes of arbitrary rate and constraint length that ensures
maximum dfree. However, comprehensive description of taps for ‘good ‘ convolutional
codes of practical interest have been prepared through extensive computer search
techniques and otherwise.

While choosing a convolutional code, one should also avoid ‘catastrophic

convolutional code’. Such codes can be identified by the state diagram. The state diagram
of a ‘catastrophic convolutional code’ includes at least one loop in which a nonzero
information sequence corresponds to an all-zero output sequence. Tree diagram of a
‘catastrophic convolutional code’ is shown in Fig.6.35.5.

Sb
1/00

0/00
Sa

1/00 Sc

Fig.6.35.5 Example of a catastrophic code

Hard-Decision and Soft-Decision Decoding

Hard-decision and soft-decision decoding are based on the type of quantization
used on the received bits. Hard-decision decoding uses 1-bit quantization on the received
samples. Soft-decision decoding uses multi-bit quantization (e.g. 3 bits/sample) on the
received sample values.

Version 2 ECE IIT, Kharagpur

Hard-Decision Viterbi Algorithm
The Viterbi Algorithm (VA) finds a maximum likelihood (ML) estimate of a
transmitted code sequence c from the corresponding received sequence r by maximizing
the probability p(r|c) that sequence r is received conditioned on the estimated code
sequence c. Sequence c must be a valid coded sequence.

The Viterbi algorithm utilizes the trellis diagram to compute the path metrics. The
channel is assumed to be memory less, i.e. the noise sample affecting a received bit is
independent from the noise sample affecting the other bits. The decoding operation starts
from state ‘00’, i.e. with the assumption that the initial state of the encoder is ‘00’. With
receipt of one noisy codeword, the decoding operation progresses by one step deeper into
the trellis diagram. The branches, associated with a state of the trellis tell us about the
corresponding codewords that the encoder may generate starting from this state. Hence,
upon receipt of a codeword, it is possible to note the ‘branch metric’ of each branch by
determining the Hamming distance of the received codeword from the valid codeword
associated with that branch. Path metric of all branches, associated with all the states are
calculated similarly.

Now, at each depth of the trellis, each state also carries some ‘accumulated path
metric’, which is the addition of metrics of all branches that construct the ‘most likely
path’ to that state. As an example, the trellis diagram of the code shown in Fig. 6.35.1,
has four states and each state has two incoming and two outgoing branches. At any depth
of the trellis, each state can be reached through two paths from the previous stage and as
per the VA, the path with lower accumulated path metric is chosen. In the process, the
‘accumulated path metric’ is updated by adding the metric of the incoming branch with
the ‘accumulated path metric’ of the state from where the branch originated. No decision
about a received codeword is taken from such operations and the decoding decision is
deliberately delayed to reduce the possibility of erroneous decision.

The basic operations which are carried out as per the hard-decision Viterbi Algorithm
after receiving one codeword are summarized below:
a) All the branch metrics of all the states are determined;

b) Accumulated metrics of all the paths (two in our example code) leading to a state
are calculated taking into consideration the ‘accumulated path metrics’ of the
states from where the most recent branches emerged;

c) Only one of the paths, entering into a state, which has minimum ‘accumulated
path metric’ is chosen as the ‘survivor path’ for the state (or, equivalently ‘node’);

d) So, at the end of this process, each state has one ‘survivor path’. The ‘history’ of a
survivor path is also maintained by the node appropriately ( e.g. by storing the
codewords or the information bits which are associated with the branches making
the path);

Version 2 ECE IIT, Kharagpur

 e) Steps a) to d) are repeated and decoding decision is delayed till sufficient number
of codewords has been received. Typically, the delay in decision making = Lx k
codewords where L is an integer, e.g. 5 or 6. For the code in Fig. 6.35.1, the
decision delay of 5x3 = 15 codewords may be sufficient for most occasions. This
means, we decide about the first received codeword after receiving the 16th
codeword. The decision strategy is simple. Upon receiving the 16th codeword and
carrying out steps a) to d), we compare the ‘accumulated path metrics’ of all the
states ( four in our example) and chose the state with minimum overall
‘accumulated path metric’ as the ‘winning node’ for the first codeword. Then we
trace back the history of the path associated with this winning node to identify the
codeword tagged to the first branch of the path and declare this codeword as the
most likely transmitted first codeword.

The above procedure is repeated for each received codeword hereafter. Thus, the
decision for a codeword is delayed but once the decision process starts, we decide once
for every received codeword. For most practical applications, including delay-sensitive
digital speech coding and transmission, a decision delay of Lx k codewords is acceptable.

Soft-Decision Viterbi Algorithm

In soft-decision decoding, the demodulator does not assign a ‘0’ or a ‘1’ to
each received bit but uses multi-bit quantized values. The soft-decision Viterbi algorithm
is very similar to its hard-decision algorithm except that squared Euclidean distance is
used in the branch metrics instead of simpler Hamming distance. However, the
performance of a soft-decision VA is much more impressive compared to its HDD (Hard
Decision Decoding) counterpart [Fig. 6.35.6 (a) and (b)]. The computational requirement
of a Viterbi decoder grows exponentially as a function of the constraint length and hence
it is usually limited in practice to constraint lengths of K = 9.

Version 2 ECE IIT, Kharagpur

Fig. 6.35.6 (a) Decoded BER vs input BER for the rate – half convolutional codes with
Viterbi Algorithm ; 1) k = 3 (HDD),2) k = 5 (HDD),3) k = 3 (SDD), and 4) k= 5 (SDD).
HDD: Hard Decision Decoding; SDD: Soft Decision Decoding.

Fig. 6.35.5(b) Viterbi decoded BER vs Eb/No for the rate – half convolutional codes; 1) b
= 3 HDD), 2) b = 5 (HDD), 3) b = 3 (SDD), 4) b = 5 (Transfer function bound) and 5)
Uncoded system.

Fig. 6.35.6 (b) Decoded BER vs Eb/No (in dB) for the rate – half convolutional codes with
Viterbi Algorithm ; 1) Uncoded system; 2) with k = 3 (HDD) and 3) k = 3 (SDD). HDD:
Hard Decision Decoding; SDD: Soft Decision Decoding.

Version 2 ECE IIT, Kharagpur