Block turbo schemes using TCH codes

Nuno Souto1 , João Carlos Silva2, Francisco Cercas3

1
IST/IT, Torre Norte 11-11, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal, [email protected]
2
IST/IT, Torre Norte 11-11, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal, [email protected]
3
ISCTE/IT, Torre Norte 11-08, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal, [email protected]

Abstract deviation σ. The elementary decoder calculates the soft-

output [Y’(m)] and delivers at its output
The application of TCH(n,k) (Tomlinson, Cercas, Hughes)
codes [ 1] in several domains, namely in power efficient [W (m + 1)] = [Y '(m)] − [Y (m)] (0.3)
radio systems
[ 2 ] with simplified receivers [ 4 ] has been searched for.
In particular, it was verified that TCH codes of small length
can be advantageous in some mobile communications
environments. On the other hand, Turbo coding schemes
involving block codes have been studied, significantly
augmenting its overall performance. In this paper we
investigate the application of these turbo schemes with TCH
codes, mainly due to its very efficient decoding algorithm,
and we present graphs showing the obtained performance
through simulation in a AWGN channel.
Figure 1 : Product code structure of BTCs
Keywords – TCH codes, block turbo codes

I. INTRODUCTION
Since block-turbo coding emerged with good results using
BCH codes, an attempt to extrapolate and do something
similar with TCH was done, especially because of the TCH’s
computationally efficient decoding method. Usually block
turbo schemes are employed with systematic codes, a Figure 2 : Decoding structure for BTC
property that TCH codes don’t possess. In [ 3 ], a quasi- Each stage of row-wise or column-wise decoding is called a
systematic bit-mapping was studied for TCH codes (with half-iteration, and each stage of row-wise and column-wise
emphasis on codes of n=16), and those results will be used decoding is called a full-iteration.
for constructing a block turbo scheme using TCH(16,5) and In the classic case, a systematic linear code such as the
TCH(256,9) codes. The best way to implement the code will BCH is used, and thus it is possible to produce a product
be explored. code structure such as the one schematised in Figure 1, where
BTCs (Block Turbo Codes) involve serial concatenation of the information symbols are coded column and row-wise.
two or more block coders separated by simple row/ column However, the TCH code is neither linear nor systematic,
interleaver. This gives them a product code structure as though the “quasi-systematic” property of the TCH, and other
shown in Figure 1. Thus, the decoding procedure for BTCs is schemes involving TCH codes explained in detail later on,
similar to the decoding of product codes. Decoding can be can be exploited in this case.
performed either row-wise first and then column-wise or At each iteration, the closest word to the received symbols
vice.versa. For iterative decoding, the decoding process at is found. There are several methods for doing this, the main
each step should be SISO (Soft Input, Soft Output). A algorithms being Chase [ 5 ] and Kaneko [ 6 ]. In the case of
schematic of the elementary row or column decoder at any TCH’s, these algorithms don’t need to be run, for the result
stage of iteration is shown in Figure 2. of the FFT at the TCH decoder provides automatically the
The input data of the mth elementary decoder is given by correlations between the received word and all other words in
[Y (m)] = [Y ] + β (m) [W (m)] (0.1)
the code set, from which the Euclidean distance is easily
extracted via the expression
where [Y] is the received data corresponding to the Eucl (cw, rw) =
transmitted codeword [X], (0.4)
autocorr (cw) + autocorr (rw) − 2 × crosscorr (cw, rw)
Y = X + G (0.2)
where cw is a codeword out of the codeword set; rw is the
[W(m)] is the extrinsic information computed by the
previous decoder, β(m) is the damping or scaling factor [ 5],
received codeword with noise, Eucl is the Euclidean
distance; autocorr is the autocorrelation and crosscorr is the
and G = (g1,g2,…,gn) are AWGN samples of standard
cross-correlation. Thus, since the FFT result returns all the
  | Y − C c |2 − | Y − D |2 
r 'j =  d j
cross-correlations between the received word and the words
 
in the codeword set, and since the autocorrelation of each 4
codeword is equal to its length, (0.4) can be summarized as (0.9)
Eucl(cw, rw) = where Cc is the concurrent codeword defined as the
codeword closest to the received vector given that dj≠cj and
length(rw) + autocorr(rw) − 2× crosscorr(cw, rw) Cc and D are expressed in antipodal form (-1 and +1).
Although the reliability calculations are approximate, the
(0.5)
results portrayed in [ 5 ] and [ 8] show that this procedure
Note that, for finding the closest word, the Euclidean
leads to a near optimal performance for the decoding
distance needs not to be computed; simply pick the codeword
algorithm (this is true if check on checks is used, otherwise it
whose cross-correlation with the received word is higher.
is not so).
The Euclidean distance was mentioned, for it will be useful
For the TCH specific case, the scheme that was simulated
later on.
took advantage of the quasi-systematic feature of the TCH
Once the optimal decision D for the received vector Y is
polynomials. Note that, for a (16,5) code, only 4 bits are used
obtained, it is necessary to compute the reliability of each bit
in the top left corner. The non-systematic bit is implicitly
that composes codeword D, so that a soft codeword is
present, being defined by the initial bit mapping. In order to
obtained. The soft-output is required so that the extrinsic
increase the decoder’s performance, the non-systematic bits
information generated at any stage of the iteration can be
corresponding to a horizontal codeword should be the same
calculated and fed back to the decoder for the next stage of
as the non-systematic bits that correspond to a vertical
iteration (this feedback must not be positive, in order not to
codeword. The code rate is reduced to R=5/(16+12).
bias the results; all information that is fed back to the decoder
needs to go to a different decoding step than the one that
originated it). The reliability that the decision dj (jth bit of D)
is defined as the LLR (Log Likelihood Ratio) of the
transmitted symbol xj (jth bit of X) which is given by
 pr ( x j = +1) | Y
{ } 
Λ ( d ) = ln 
 pr ( x j = −1) | Y
{ } 
(0.6)
 Figure 3 : Block turbo code for “quasi-systematic” TCH code.
By taking into account that D is one of the codewords of C,
applying Bayes’ rule, and assuming that the different
codewords are uniformly distributed, the LLR can be II. SIMULATION RESULTS

 ∑ p {Y | ( X = C i )} 
expressed as The Monte Carlo method was used for all simulations. The

 C i∈S +j 1 
simulations were run until a sufficient number of errors

Λ ( d ) = ln   (0.7)
occurred (at least 20). All the random generators were

 ∑ p {Y | ( X = C )} 
initialized at the beginning of each value of Eb/N0, and serial

i

 C ∈S j 
decoding was employed.
i −1
Looking at the results for turbo TCH based on TCH(16,5)
where Sj+1 is the set of codewords {Ci} such that cij=1, Sj- and TCH(256,9) (Figure 4 and Figure 5 respectively), the
1 is the set of codewords {Ci} such that cij=-1, and p{Y|X= performance gain was above 0,6dB for a BER of 1e-5, in the
Ci } is the probability density function of Y conditioned on X standard STCH (Systematic Turbo TCH) scheme. A 3D
and is defined by the particular channel in consideration. For scheme [11] using 3 parity sequences (in which the third
an AWGN channel, p{Y|X= Ci } is given by parity sequence is done in such a way that only one
 1   | Y − Ci |2 
n systematic bit is common to every other codeword from the
p{Y | ( X = C )} =   − 
 2πσ   2σ 2 
i other two parity sequences) was also simulated, but results
exp
for this 3D scheme are worse than for a regular (2D) STCH
(0.8)
scheme for low values of Eb/N0. Results for the STCH
scheme based on TCH(256,9) present worse performance
The LLR yields the soft-output for every decision and can
gains, due to using only 6 bits out of 9 [3]. Good results have
be used for calculating the extrinsic information and hence
been obtained with similar schemes, using Hadamard
the soft input for the next stage of iteration. The calculation
biorthogonal codes [10] [11]; but these codes have the
of the LLR using (0.6) can be tedious and often impractical
advantage of being systematic, and thus use the maximum
(though in TCH codes, it is easier to implement). Another
possible number of systematic bits.
method for calculating the soft decision is to use the
concurrent codeword for reliability calculations as used in the
soft-output Viterbi algorithm [ 7 ], and used in [ 5 ] and [ 8 ].
For every bit dj in D, the soft-decision can be obtained as
 Figure 4 : Results for block turbo code based on TCH(16,5) Figure 6 : Effect of different number of iterations

Figure 5 : Results for block turbo code based on TCH(256,9) Figure 7 : Effect of optimizing extra bit
The effect of the number of iterations is also significative.
Several simulations were run for different number of
iterations (Figure 6), and it was seen that 3 iterations are
reasonable. The number of iterations is much lower than
those mentioned in [10] due to the fact that the decoding is
MLE. Note also that, as in [11], the iterations needn’t be
weighed.
The effect of optimizing the first bit’s mapping was also
analysed. Figure 7 quantifies the performance increase
obtained from doing so (0,3dB).
Looking at the error distribution per column, we can see
that the difference between both situations in solely in the
errors attached to the first bit (Figure 8), whereas with Figure 8: STCH(16+12,5) code without mapping optimization
optimization, all bits have the same error rate (Figure 9). on first bit
STCH schemes with different polynomials were also Note that all the presented results are worse than similar
studied; Figure 10 shows that a STCH using as a basis results for Hadamard-Biorthogonal codes [11]. However, the
polynomial BCD0 has slightly worse results to the one using TCH codes’ decoding ease and MLE performance [1] reduce
polynomial DA1C, as predicted in [3]. The main difference the number of decoding iterations and speed up the decoding
between them is that the latter boasts optimum bit-mapping process, boasting good results in channels where multipath is
for quasi-systemacy, and requires an interleaver in order to present, due to their autocorrelation being very small (the
position the bits, whereas codes based on polynomial BCD0 autocorrelation for a TCH code is the same as the cross-
don’t need an interleaver (they use the first 4 bits as correlation , if the code has a single polynomial).
systematic [3]), but don’t have an optimum bit-mapping.
 January 1993, Hilton Palacio del Rio Hotel, San Antonio,
Texas, USA, p.198.
[ 2 ] P.J.A. Sebastião, F.A.B. Cercas e A.V.T. Cartaxo,
”Performance of TCH codes in a land mobile satellite
channel”, PIMRC 2002 – The 13th IEEE Internatiional
Symposium on Personal, Indoor and Mobile Radio
Communications, Lisboa-Portugal, September 15-18 de
2002.
[ 3 ] N.Souto, J.C.Silva, F. Cercas, “Special bit-mapping of
TCH codes for application in turbo schemes”, submitted to
this conference.
[ 4 ] F. A. B. Cercas, “A new family of codes for simple
receiver implementation”, Ph.D. Thesis, Technical University
Figure 9: STCH(16+12,5) code with mapping optimization on of Lisbon, Instituto Superior Técnico, Lisbon, March 1996.
first bit [ 5 ]R. M. Pyndiah, “Near-optimum decoding of product
codes: Block turbo codes”, IEEE Trans. Commun., vol 46,
pp.1003-1010, Aug. 1998
[ 6 ]Sameep Dave, Junghwan Kim and Subhash C. Kwatra,
“An Efficient Decoding Algorithm for Block Turbo Codes”,
IEEE Trans. Commun., vol 49, pp.41-146, Jan. 2001
[ 7 ]J. Hagenauer and P. Hoeher, “A Viterbi Algorithm with
soft-decision outputs and its applications”, in Proc. IEEE
GLOBECOM’89, Dallas, TX, Nov. 1989, pp. 1680-1686
[ 8 ]– “ Iterative decoding of product codes: Block turbo
codes”, in Proc. Int. Symp. Turbo Codes, Brest, France, Sept.
1997, pp. 71-79
[ 9]– Bernard Sklar, “ A Primer on Turbo Code Concepts”,
in IEEE Communication Magazine, pp94-102, December
1997
[ 10 ] – Li Ping and S. Chan, “Iterative decoding of
concatenated Hadamard codes”, Proc. IEEE ICC-98, vol.1,
pp136-140, 1998
Figure 10 : Differences in polynomials [ 11 ]– Li Ping, S. Chan and K.L. Yeung, “Iterative
decoding of multidimensional concatenated single parity
check codes”, Proc. IEEE ICC-98, vol.1, pp131-135, 1998
III. CONCLUSION
A Turbo-scheme taking advantage of TCH codes quasi-
systematic properties was proposed. Simulations were ran for
the turbo TCH based on TCH(16,5) and TCH(256,9) codes.
The best coding gains were obtained for the TCH(16,5) based
codes, due to having 4 out of 5 bits (80%) of information-
passing bits, whereas the TCH(256,9) based codes only have
6 out of 9 bits (66,67%) for information-passing. The effect
of optimizing the extra bits for performance was measured,
and various simulations showed that using a maximum of 3
iterations produced good results. Although the obtained
results are slightly worse than similar ones for Hadamard-
Biorthogonal codes [11], TCH codes’ decoding ease and
MLE performance [1] reduce the number of decoding
iterations and speed up the decoding process, so they can be
a valid option specially for channels where multipath is
present.
IV. REFERENCES
[ 1 ] F. Cercas, M. Tomlinson, A. Albuquerque, “TCH: A
New Family of Cyclic Codes Length 2m”, 1993 IEEE
International Symposium on Information Theory, 17-22