Chapter 5.

Performance Analysis of Turbo Codes

The turbo code, as described in Chapters 3 and 4, is a very complex channel
coding scheme. The turbo code encoder is a parallel concatenation of two recursive
systematic convolutional (RSC) codes. The turbo code decoder is an iterative serial
concatenation of two soft output Viterbi algorithm (SOVA) decoders. In addition, the
presence of interleavers in both the encoder and the decoder further complicates this
coding scheme.

The two main methods to evaluate the performance of turbo codes are theoretical
analysis and computer simulation. Theoretical analysis of a turbo code is very difficult
due to the structure of the coding scheme. A few journal papers have analyzed turbo
codes with this theoretical approach; however, their results do not match closely (too
optimistic) to computer simulation results. Also, the theoretical analyses presented in
these papers are not clearly described and thus are difficult to follow. Furthermore, the
theoretical approach often requires tremendous computer run time to find the complete
weight distribution of the turbo code codewords.

In this thesis, due to the difficulties presented in the theoretical analysis, the
performance of turbo codes is evaluated through computer simulation. MATLAB is used
to construct the computer code of a turbo code, and the simulations were carried out SUN
and HP workstations.

5.1 Simulation Setup

The simulation setup is composed of three distinct parts, namely the encoder, the
channel, and the decoder. The simulation of the turbo code encoder is based on its
description in Chapter 3. The simulated turbo code encoder is composed of two identical
RSC component encoders. These two component encoders are separated by a random
interleaver. The random interleaver is a random permutation of bit order in a bit stream.
This random permutation of bit order is stored so that the interleaved bit stream can be
deinterleaved at the decoder.

In the literature, the termination aspects of a turbo code are not very well
described. At the turbo code encoder, the RSC encoders need to be properly terminated
by returning the code memory, of size m, to the all zero state. To perform this task, the
systematic code stream is augmented by the addition of m tail bits. These m tail bits
cannot be easily predetermined and depend on the code memory. The strategy to obtain
these m tail bits for the systematic code stream is discussed in Chapter 3.

The output of the turbo code encoder is described by three streams, one systematic
(uncoded) bit stream and two coded bit streams. The systematic bit stream can only have
Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 68
one set of m tail bits from one of the two recursive encoders. Figure 5.1 and Figure 5.2
show the two possible scenarios for these m tail bits.

Complete RSC 1

Recursive Encoder 1

Interleaver

Recursive Encoder 2

Figure 5.1: Termination of the systematic bit stream is associated with recursive
encoder 1.

Recursive Encoder 1

Interleaver
Complete RSC 2

Recursive Encoder 2

Figure 5.2: Termination of the systematic bit stream is associated with recursive
encoder 2.

In the literature, the termination scheme of Figure 5.1 is commonly used.

However, the literature also suggests acquisition of the decoded bit decisions at the
second component decoder of the turbo code decoder. This is not advisable because the
m tail bits for the systematic bit stream are associated with recursive encoder 1. As a
result, it is possible that the final path chosen in the second component decoder is not the
Maximum Likelihood (ML) path due to the erroneous m tail bits of the systematic bit
stream. Thus, if the encoder structure of Figure 5.1 is used, then the decoded bit
decisions should be acquired from the first component decoder. Also, if the encoder
Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 69
structure of Figure 5.2 is used, then the decoded bit decisions should be acquired from the
second component decoder. In the simulation, the encoder structure of
Figure 5.2 is used.

In its basic form, the turbo code encoder is rate 1/3. However, in many journal
papers, the published computer simulations of turbo codes often use rate 1/2. This is
accomplished by puncturing the coded bit streams of the turbo code. The puncturing
pattern is that for one coded bit stream, the odd bits are punctured out, and for the other
coded bit stream, the even bits are punctured out.

In the simulation, the Gaussian channel model is used because it is a fairly good
model for different transmission mediums. The Gaussian channel (Gaussian noise) is
fairly easy to construct from the basic Gaussian distribution with mean of zero and
standard deviation of one. In order to use this model, the turbo code encoder output bit
streams must be mapped from {0,1} to {-1,+1} domain.

The simulation of the turbo code decoder is based on its description in Chapter 4.
If the encoder structure of Figure 5.2 is used, then the systematic bit stream must be
deinterleaved before passing to the first component decoder of the turbo code decoder.
Also, the initial (first iteration) a priori values for the first component decoder are set to
zero because there are no extrinsic information available from the second component
decoder.

It has been found that the turbo code (SOVA) decoder, implemented as described
in the literature, did not perform up to the published results. The cause of this
performance degradation has been traced to the extrinsic/a priori information that is
passed between the component decoders. After many computer simulation runs, it has
been determined that the extrinsic/a priori information must be limited (reduced) before
passing to the next component decoder. The two main strategies that were devised to
reduce the extrinsic/a priori information are:
1. Introduce a limiter on the extrinsic/a priori information so that the values do not go
over a predetermined bound.
2. Introduce a “numerical factor” to scale down the extrinsic/a priori information.

After extensive trials and tests, it has been determined that the most effective
reduction technique was to introduce a numerical factor Eb/No (signal to noise ratio per
bit) to scale down the extrinsic/a priori information. This factor is idealistically
reasonable in the sense that for low Eb/No, the extrinsic/a priori values are relatively
small so no drastic reduction is required. However, for high Eb/No, the extrinsic/a priori
values are relatively large so greater reduction is required. In the simulation, the scale
down factor of Eb/No is used on the extrinsic/a priori information.

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 70

5.2 Simulation Results
Simulation results for a turbo code are based on bit error rate (BER) performance
over a range of Eb/No. The following table shows the rate 1/2 component RSC codes
used in the simulation results.

Table 5.1: Rate 1/2 RSC Component Codes Used in Simulation Results
Constraint Length (K) Feedforward Feedback
Generator (in Octal) Generator (in Octal)
3 5 7
4 15 17
5 23 35
6 53 75

First, the performance of these rate 1/2 RSC codes is shown for soft decision
Viterbi decoding. Then, the simulation results for two specific cases of a turbo code
using SOVA decoding are compared to the published journal paper results [Jun96]. In
addition, other turbo code simulation results are shown for different rate (R), constraint
lengths (K), and frame sizes (FS) using the rate 1/2 RSC component codes. These results
are first shown in table format and then in graphical format.

Table 5.2: Rate 1/2 RSC Code BER Performance in Soft Decision Viterbi Decoding
Eb/No (dB) Uncoded K=3 K=4 K=5 K=6
1 5.7x10-2 5.3x10-2 4.2x10-2 3.5x10-2 2.9x10-2
2 3.9x10-2 2.0x10-2 1.4x10-2 8.9x10-3 5.0x10-3
3 2.4x10-2 6.1x10-3 2.7x10-3 1.1x10-3 8.0x10-4
4 1.3x10-2 1.0x10-3 4.9x10-4 1.8x10-4 3.9x10-5

Table 5.3: Comparison of Turbo Code BER Performance for 10 Decoding

Iterations (R=1/2, K=3, FS=192)
Eb/No (dB) Uncoded Simulated Published
1 5.7x10-2 4.5x10-2 5.0x10-2
2 3.9x10-2 4.9x10-3 4.3x10-3
3 2.4x10-2 5.9x10-4 1.5x10-4

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 71

 Table 5.4: Comparison of Turbo Code BER Performance for 10 Decoding
Iterations (R=1/2, K=4, FS=192)
Eb/No (dB) Uncoded Simulated Published
1 5.7x10-2 9.0x10-2 7.0x10-2
2 3.9x10-2 6.1x10-3 4.5x10-3
3 2.4x10-2 1.7x10-4 5.4x10-5

Table 5.5: Turbo Code BER Performance for 10 Decoding Iterations (R=1/2, K=3)
Eb/No (dB) Uncoded FS=50 FS=192 FS=1000
1 5.7x10-2 4.9x10-2 4.5x10-2 4.1x10-2
2 3.9x10-2 1.8x10-2 4.9x10-3 3.9x10-4
3 2.4x10-2 3.6x10-3 5.9x10-4 -
4 1.3x10-2 5.9x10-4 6.8x10-5 -

Table 5.6: Turbo Code BER Performance for 10 Decoding Iterations (R=1/3, K=3)
Eb/No (dB) Uncoded FS=50 FS=192 FS=1000
1 5.7x10-2 5.8x10-2 1.7x10-2 1.5x10-3
2 3.9x10-2 1.8x10-2 1.8x10-3 1.3x10-4
3 2.4x10-2 2.3x10-3 1.7x10-4 -
4 1.3x10-2 4.0x10-4 3.5x10-5 -

Table 5.7: Turbo Code BER Performance for 10 Decoding Iterations (R=1/2, K=4)
Eb/No (dB) Uncoded FS=50 FS=192
1 5.7x10-2 7.8x10-2 9.0x10-2
2 3.9x10-2 2.8x10-2 6.1x10-3
3 2.4x10-2 2.9x10-3 1.7x10-4

Table 5.8: Turbo Code BER Performance for 10 Decoding Iterations (FS=50)
Eb/No (dB) Uncoded R=1/2, K=3 R=1/3, K=3 R=1/2, K=4
1 5.7x10-2 4.9x10-2 5.8x10-2 7.8x10-2
2 3.9x10-2 1.8x10-2 1.8x10-2 2.8x10-2
3 2.4x10-2 3.6x10-3 2.3x10-3 2.9x10-3
4 1.3x10-2 5.9x10-4 4.0x10-4 -

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 72

 Table 5.9: Turbo Code BER Performance for 10 Decoding Iterations (FS=192)
Eb/No (dB) Uncoded R=1/2, K=3 R=1/3, K=3 R=1/2, K=4
1 5.7x10-2 4.5x10-2 1.7x10-2 9.0x10-2
2 3.9x10-2 4.9x10-3 1.8x10-3 6.1x10-3
3 2.4x10-2 5.9x10-4 1.7x10-4 1.7x10-4
4 1.3x10-2 6.8x10-5 3.5x10-5 -

Table 5.10: Turbo Code BER Performance for 10 Decoding Iterations (FS=1000)
Eb/No (dB) Uncoded R=1/2, K=3 R=1/3, K=3
1 5.7x10-2 4.1x10-2 1.5x10-3
1.5 4.6x10-2 3.5x10-3 3.8x10-4
2 3.9x10-2 3.9x10-4 1.3x10-4

Table 5.11: Turbo Code BER Performance Gain (R=1/2, K=3)

Eb/No Uncoded FS=50 FS=50 FS=192 FS=192 FS=1000 FS=1000
(dB) Iter=1 Iter=10 Iter=1 Iter=10 Iter=1 Iter=10
1 5.7x10-2 6.7x10-2 4.9x10-2 8.5x10-2 4.5x10-2 9x10-2 4.1x10-3
2 3.9x10-2 4.0x10-2 1.8x10-2 3.5x10-2 4.9x10-3 2.7x10-2 3.9x10-4
3 2.4x10-2 7.9x10-3 3.6x10-3 7.0x10-3 5.9x10-4
4 1.3x10-2 1.8x10-3 5.9x10-4 1.2x10-3 6.8x10-5

Table 5.12: Turbo Code BER Performance Gain (R=1/3, K=3)

Eb/No Uncoded FS=50 FS=50 FS=192 FS=192 FS=1000 FS=1000
(dB) Iter=1 Iter=10 Iter=1 Iter=10 Iter=1 Iter=10
1 5.7x10-2 6.1x10-2 5.8x10-2 6.5x10-2 1.7x10-2 7.5x10-2 1.5x10-3
2 3.9x10-2 3.9x10-2 1.8x10-2 2.7x10-2 1.8x10-3 2.1x10-2 1.3x10-4
3 2.4x10-2 7.3x10-3 2.3x10-3 4.4x10-3 1.7x10-4
4 1.3x10-2 1.0x10-3 4.0x10-4 5.0x10-4 3.5x10-5

Table 5.13: Turbo Code BER Performance Gain (R=1/2, K=4)

Eb/No (dB) Uncoded FS=50 FS=50 FS=192 FS=192
Iter=1 Iter=10 Iter=1 Iter=10
1 5.7x10-2 9.2x10-2 7.8x10-2 1.0x10-1 9.0x10-2
2 3.9x10-2 4.8x10-2 2.8x10-2 3.4x10-2 6.1x10-3
3 2.4x10-2 7.0x10-3 2.9x10-3 5.5x10-3 1.7x10-4

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 73

 Rate 1/2 RSC Code BER Performance in Soft Decision
Viterbi Decoding

1.00E-01

1.00E-02

Uncoded
K=3
BER

1.00E-03 K=4
K=5
K=6

1.00E-04

1.00E-05
1 2 3 4
Eb/No (dB)

Figure 5.3: Rate 1/2 RSC code BER performance in soft-decision Viterbi decoding.

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 74

 Comparison of Turbo Code BER Performance
for 10 Decoding Iterations (R=1/2, K=3, FS=192)

1.00E-01

1.00E-02 Uncoded
Simulated
Published
BER

1.00E-03

1.00E-04
1 2 3
Eb/No (dB)

Figure 5.4: Comparison of turbo code BER performance for 10 decoding

iterations (R=1/2, K=3, FS=192).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 75

 Comparison of Turbo Code BER Performance
for 10 Decoding Iterations (R=1/2, K=4, FS=192)

1.00E-01

1.00E-02

Uncoded
Simulated
BER

1.00E-03 Published

1.00E-04

1.00E-05
1 2 3
Eb/No (dB)

Figure 5.5: Comparison of turbo code BER performance for 10 decoding

iterations (R=1/2, K=4, FS=192).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 76

 Turbo Code BER Performance for 10 Decoding Iterations
(R=1/2, K=3)

1.00E-01

1.00E-02

Uncoded
BER

FS=50
1.00E-03
FS=192
FS=1000

1.00E-04

1.00E-05
1 2 3 4
Eb/No (dB)

Figure 5.6: Turbo code BER performance for 10 decoding iterations (R=1/2, K=3).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 77

 Turbo Code BER Performance for 10 Decoding Iterations
(R=1/3, K=3)

1.00E-01

1.00E-02

Uncoded
BER

FS=50
1.00E-03
FS=192
FS=1000

1.00E-04

1.00E-05
1 2 3 4
Eb/No (dB)

Figure 5.7: Turbo code BER performance for 10 decoding iterations (R=1/3, K=3).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 78

 Turbo Code BER Performance for 10 Decoding Iterations
(R=1/2, K=4)

1.00E-01

1.00E-02

Uncoded
BER

FS=50
FS=192

1.00E-03

1.00E-04
1 2 3
Eb/No (dB)

Figure 5.8: Turbo code BER performance for 10 decoding iterations (R=1/2, K=4).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 79

 Turbo Code BER Performance for 10 Decoding Iterations
(FS=50)

1.00E-01

1.00E-02

Uncoded
BER

R=1/2, K=3
R=1/3, K=3
R=1/2, K=4
1.00E-03

1.00E-04
1 2 3 4
Eb/No (dB)

Figure 5.9: Turbo code BER performance for 10 decoding iterations (FS=50).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 80

 Turbo Code BER Performance for 10 Decoding Iterations
(FS=192)

1.00E-01

1.00E-02

Uncoded
BER

R=1/2, K=3
1.00E-03
R=1/3, K=3
R=1/2, K=4

1.00E-04

1.00E-05
1 2 3 4
Eb/No (dB)

Figure 5.10: Turbo code BER performance for 10 decoding iterations (FS=192).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 81

 Turbo Code BER Performance for 10 Decoding Iterations
(FS=1000)

1.00E-01

1.00E-02

Uncoded
BER

R=1/2, K=3
R=1/3, K=3

1.00E-03

1.00E-04
1 1.5 2
Eb/No (dB)

Figure 5.11: Turbo code BER performance for 10 decoding iterations (FS=1000).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 82

 Turbo Code BER Performance Gain (R=1/2, K=3)

1.00E-01

1.00E-02

Uncoded
FS=50, Iter=1
FS=50, Iter=10
BER

1.00E-03 FS=192, Iter=1

FS=192, Iter=10
FS=1000, Iter=1
FS=1000, Iter=10

1.00E-04

1.00E-05
1 2 3 4
Eb/No (dB)

Figure 5.12: Turbo code BER performance gain (R=1/2, K=3).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 83

 Turbo Code BER Performance Gain (R=1/3, K=3)

1.00E-01

1.00E-02

Uncoded
FS=50, Iter=1
FS=50, Iter=10
BER

1.00E-03 FS=192, Iter=1

FS=192, Iter=10
FS=1000, Iter=1
FS=1000, Iter=10

1.00E-04

1.00E-05
1 2 3 4
Eb/No (dB)

Figure 5.13: Turbo code BER performance gain (R=1/3, K=3).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 84

 Turbo Code BER Performance Gain (R=1/2, K=4)

1.00E-01

1.00E-02

Uncoded
FS=50, Iter=1
BER

FS=50, Iter=10
FS=192, Iter=1
FS=192, Iter=10

1.00E-03

1.00E-04
1 2 3
Eb/No (dB)

Figure 5.14: Turbo code BER performance gain (R=1/2, K=4).

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 85

5.3 Simulation Analysis
Before analyzing the performance of a turbo code, the performance of its rate 1/2
RSC component code is first examined. The performance of the rate 1/2 RSC code in
soft decision Viterbi decoding for different constraint lengths is shown in Figure 5.3. In
this figure, it can be seen that as the constraint length increases, the performance of the
code also increases, resulting in lower BER. This is the typical characteristic of any
convolutional code.

To validate the turbo code simulation, Figure 5.4 and Figure 5.5 show the
simulated and published performance results for two specific cases. In Figure 5.4, the
difference between the simulated and published performance results is negligible for low
Eb/No (<2), but for higher Eb/No (>2), this difference becomes greater. This can also be
seen in Figure 5.5. However, the simulated turbo code performance results shown in this
thesis were independently validated by another researcher, Matthew Valenti, at Virginia
Tech (via personal communications).

The difference between the simulated and published performance results are
unknown. This problem is believed to be caused by two independent factors. First, it has
been suggested that the SOVA decoding algorithm is very sensitive to numerical
inaccuracies. The computer may introduce round off error in many of its calculations
such that the final output result is not very precise compared to the true value. This may
be the case for BER at high Eb/No. As can be seen from Figure 5.4 and 5.5, the relatively
significant difference between the simulated and published performance results is located
at very low BER. At this low BER, a few bit errors caused by computer inaccuracies will
be magnified greatly.

It is speculated that the source of the round off error originates from the iterative
decoding process. As small errors are introduced from one decoding iteration, it will be
accumulated with other errors from previous decoding iterations. This error may grow as
the number of iterations increases. It should be noted that most of the published turbo
code simulation results were carried out on supercomputers.

The second factor that may cause the difference between the simulated and
published performance results comes from the lack of detail in the SOVA decoding
algorithm. Two researchers, J. Hagenauer and P. Jung (listed in the reference), who are
well known for their work on turbo codes, were contacted and questioned via personal
communications about the details of the SOVA decoding algorithm. However, their help
did not resolve the discrepancy between the simulated and published performance results.

The simulated performance results of turbo codes with the same component codes
but different frame sizes are shown in Figure 5.6, Figure 5.7, and Figure 5.8. As shown

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 86

in these figures, the turbo code performance improves as the frame size increases. This
conclusion can be expected because the larger the frame size, the more accurate are the
reliability estimations in the SOVA decoding algorithm.

The simulated performance results of turbo codes with fixed frame sizes but
different component codes and rates are shown in Figure 5.9, Figure 5.10, and Figure
5.11. From these figures, it can be seen that for a fixed constraint length, a decrease in
code rate increases the turbo code performance. It can also be seen that for a fixed code
rate, an increase in constraint length improves the turbo code performance. These
conclusions are similar to that for convolutional codes.

The overall iterative (10 iterations) decoding gain for a turbo code with the same
component codes and rates but different frame sizes are shown in Figure 5.12, Figure
5.13, and Figure 5.14. As shown in these figures, the overall iterative decoding gain
increases as the frame size increases. This and the above mentioned characteristics of the
simulated turbo code performance results are in the same direction with current published
research work.

5.4 Comparison of Turbo Code Decoding and Viterbi Decoding

Turbo code decoding and Viterbi decoding are often compared in terms of bit
error rate (BER) performance and computational complexity. In Figure 5.3, BER
performance of soft-decision Viterbi decoding for rate 1/2 RSC code is shown. This
figure shows that Viterbi decoding is effective for large constraint length codes (e.g. K=5
or K=6). However, this is not the case for turbo code decoding. From Figures 5.6, 5.7,
and 5.8, it can be seen that turbo code decoding is able to provide comparable or better
BER performance for small constraint length codes (e.g. K=3 or K=4) than Viterbi
decoding of large constraint length codes (e.g. K=5 or K=6) under the condition that
relatively large frame sizes (FS ≥ 192 bits) are used.

Fairly high computational complexity is required for turbo code decoding. The
SOVA component decoder is approximately two times as complex as the Viterbi decoder
(Figure 4.11). Turbo code decoding requires two SOVA component decoders; thus, one
iteration of turbo code decoding requires complexity up to four times that of the Viterbi
decoder (Figure 4.12). As one can see, the complexity of the turbo code decoder
dramatically increases as the number of decoding iterations increases. In this chapter,
most of the turbo code simulation results are based on ten decoding iterations, which are
approximately equivalent to complexity up to forty times that of Viterbi decoding.
Typically, turbo code simulations require up to three days for frame size of 192 bits and
over a week for frame size of 1000 bits running MATLAB on the background of SUN
SPARC 10 workstations. Due to the decoding complexity of turbo code, most of the

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 87

BER performance results are for component code with small constraint lengths, high code
rates, and short frame sizes.

Fu-hua Huang Chapter 5. Performance Analysis of Turbo Codes 88