Partition Decoding for Reed-Solomon Codes Based

on Bit Reliability
Ming-Hua Chang*, Ta-Hsiang Hu**, Ing-Jiunn Su***and Zhong-Yu Li**
*Department of Electrical Engineering, Chin-Min Institute of Technology, Tou-Fen, Miao-Li, Taiwan 351, ROC
**Department of Electrical Engineering, Da-Yeh University, Da-Tsuen, Changhua ,Taiwan 515, ROC
***Department of Electrical Engineering, Chung Cheng Institute of Technology, National Defense University, Taoyuan,
Taiwan 335, ROC

Abstract-This study presents a partition decoding encoding process, and (3) to generate a codeword and
algorithm for an (mN, mK) binary image of an (N, K) Reed then calculate its Euclidean distance with respect to a
Solomon code over GF(2m). A permutation is performed on
a binary generator matrix of an RS code and a received received vector.
vector, both of which are then partitioned into two This paper is organized as follows. The binary image
submatrices and two subvectors. The first subvector of structure for an RS code is presented in Section II. Based
length N(m-1) associate with the first submatrix and the on the like of a binary image code structure of an RS
second subvector with the least subvector reliability value code, a proposed partition decoding is presented, and its
relates to the second submatrix. An MLD based on the first
submatrix or a subfield subcode BCH generator matrix is performance analysis is described in Sections III and IV,
employed to decode these two submatrices one after another. respectively. The GMD decoding algorithm is a
A codeword is finally outputted after performing the inverse suboptimum algorithm [1], which asymptotically
permutation on a concatenation of code vectors decoded approaches to the optimum performance. Simulation
from these two decoding. Simulation results show that results of proposed partition decoding, MLD[8], hybrid
under a constraint of decoding complexity, a solution to
decode binary images of long RS codes could be proposed decoding[13] and GMD decoding are discussed in
partition decoding, which provides superior performance to Sections V.
GMD decoding with much lower complexity than the MLD.
II. BINARY IMAGE STRUCTURE OF RS CODES
I. INTRODUCTION
The code structure of a binary image for an (N, K, D)
Channel coding has a problem of decoding complexity RS code over GF(2m ) contains a concatenation of m
in long transmission codes. Many works on soft decision binary (N, k, dmin,bch) BCH codes and sums up some
decoding (SDD) algorithm [1]-[10] have indicated the glue-vector codewords. The structure for a codeword is
difficulty in reducing decoding complexity. The decoding given by [4][10]
complexity of several small and medium size codes, such (1)
c = (cbch (m)
,", cbch ) + c glue − vector
as Goray code and Reed Muller (RM) codes can be   

m ' s BCH codewords
reduced by intensively exploiting the code structure. . (1)
Nonetheless, maximum likelihood decoding (MLD) is Its corresponding binary generator matrix is expressed
impractical in large block codes, because of the time as follows:
limitation and hardware complexity. Since two-stage  Gbch
(1)
0 " 0 
decoding reduces decoding complexity, the closest coset  ( 2)

 0 Gbch " 0 
decoding (CCD) can be regarded as a two-stage decoding
algorithm for RM codes [9]. In the closet coset decoding, G (b ) =  %
 = (G (b) G ( b) " G (b) )
 1 2 m

an MLD algorithm based on RM code trellis is  (m )

Gbch 
implemented in each stage. However, this class of  (1) ( 2) (m ) 

G
 gv G gv " G gv 
two-stage decoding algorithm cannot be adopted for the
(2)
other error-correcting codes, like Reed Solomon codes,
 0 
which do not have as good trellises as RM codes.  
Ordered statistic decoding with order-reprocessing is  # 
For 1 ≤ i ≤ m,  G (i) 
used in the MLD algorithm for these binary
= 
bch
Gi( b )
codes[5]-[8][10]. Ordered statistic decoding requires bit  0 
reliability of a received vector reordering that the  # 
 
generator matrix G of a binary code C is manipulated  G (i ) 
with respect to this new order to decode received vector.  gv 
(1) ( 2) ( m)
There are 3 steps employed in order-reprocessing: (1) to and Ggv = (Ggv , G gv ,", Ggv ) (3)
permute and systematize a generator matrix based on (i )
G bch and G (i )
gv
are the i-th k×N BCH subcode
received bit reliability, (2) to examine an error pattern in
generator matrix and the i-th m( K − k ) × N
glue-subvector generator submatrix in a glue-vector III. PARTITION DEAODING
generator matrix G gv . An example of a binary generator
For BPSK
transmission, the codeword
is given as follows: c = (c1 , c 2 ," , c mN ) ∈ (mN, mK, dmin)RS , is mapped
(b)
According to generator matrix structure in (2), a binary
information vector u can be divided as follows: into the bipolar sequence x = ( x1 , x 2 , " , x mN ) , with
u = (u (1 )
," , u (m )
,u (g)
), (4) xi = −2ci + 1 ∈ {±1} . A received vector at the output
of the sampler in the demodulator is
where, for 1 ≤ i ≤ m, a subvector u (i ) associates with the mN
r = (r1 , r2 , ", rmN ) ∈ R with ri = xi + wi , and
(i )
i-th BCH subcode generator matrix Gbch and subvector for 1 ≤ i ≤ mN , wi ’s are independent Gaussian random
(g)
u relates to a glue-vector generator matrix G gv . Let variables. The bit reliability is proportional to the value of
(mN , mK , d min )C (b ) be a binary image for an (N, K, D)
| ri | . According to (2), r is equally divided into m’s
RS code and d min ≥ D . subvectors of length N, i.e.,
(1) (m)
For convenience, let Gp1 and Gp2 denote the following r = (r1 , r2 , " , rmN ) = ( r , ", r ) and
submatrices r ( j)
= (r( j −1) N +1 ,", r jN ) , for 1 ≤ j ≤ m . A subvector
 G bch
(1 )
0 " 0 
  reliability of r (i ) is defined as follows:
 0 (2)
G bch " 0   0 
  iN
G p1 =  # 0 " 0  and G =  #  (5)
 p 2 δ (i) = ∑| r j |. (9)
 # # ( m −1 )
G bch   G (m)  j = ( i −1) N +1
 (1)   gv 
G gv( 2 ) " G gv( m −1 )  After m’s subvector reliability values have been
 G gv calculated, a pattern A = (1,2, ", i − 1, i + 1,", m, i) is
formed, for which δ
(i )
(b ) is the least subvector reliability
Then the codeword c ∈ C becomes value among m subvector reliability values. Based the
 G p1 G p 2 
c = (u p1 , u p 2 )  = (c p1 , c p 2 ) ,
( m) 
(6) order in pattern A, a permutation λi is formed, such that
 0 Gbch  λi [r ] = (r (1) , ", r ( i −1) , r (i +1) , " , r ( m) , r (i ) ) = ( rp(1i ) , rp(2i ) ),
(1) ( m −1) (g)
where u p1 = (u , ," , u ,u ) and u p 2 = u (m)
(10)
(i ) (1) ( i −1) ( i +1) ( m)
and where r p1 = ( r , ", r ,r , ", r ) and
c p1 = u p1G p1 (7)
rp(2i ) = r ( i ) = (r( i −1) N +1 ," , riN ) . According to the
(m) (g) (m) (m) (m)
c p 2 = u p1G p 2 + u p 2 G =u G +u G
permutation λi , the submatrices G p( i1) and G p(i2) are
bch gv bch
(m)
=c gv + cbch formed as follows:
(8) G(b) , for 1≤ j ≤ m , j ≠ i
 j
 
In the encoding process, the codeword c is a
 Gbch "
(1)
0 0 " 0 
concatenation of code sequences c p1 and c p2 . Code  
 0 % # # 
sequences c p1 and c p2 are formed in (7) and (8).  ( i −1) 
 # G bch # # 
Equations (7) and (8) have a common part u p1 . As a G (pi1) =  # ( i +1)
Gbch # 
 
result, for a proposed two-stage decoding algorithm, once  # % 0 
the subcode sequence c p1 is correctly decoded, the  0 0 ( m) 
0 G bch
 (1) ( i −1) ( i +1)

( m) 
corresponding message subsequence u p1 can be  G gv " G gv G gv " G gv 
calculated. The term c
( m)
in (8) is then computed. After  0 
gv
 
cbch has been successfully decoded. Finally, a decoded and G (pi2) =  #  . (11)
 G (i) 
codeword c is found with a pair of (c p1 , c p 2 ) .  gv 
 (i)
The submatrix G p1 of a certain permutation λi is (d) To find the subsequence c p(i1) , which holds the
formed with a concatenation of (m−1) submatrices shortest Euclidean distance to rp(1i ) with MLD[8]
G (bj ) in (3), for 1 ≤ j ≤ m, j ≠ i, and then all-zero row based on submatrix G (pi1) .
vectors is deleted. In other words, for a certain (i )
(e) To obtain information bits u p(i1) from c p1 , then
permutation λi , the submatrix G (pi1) is obtained: deleting
calculate the code sequence c gv(i ) = u p(i1) G (pi2) .
a certain submatrix Gi(b ) in a generator matrix G ( b ) in (2)
and then eliminating all-zero row vectors. (f) To find a BCH codeword cbch , such that
In the first phase decoding, MLD[8] based on G p(i1) is cbch + c gv(i ) holds the shortest Euclidean distance
used to obtain a sub-sequence c p(i1) , which holds the to rp(2i ) , with MLD[8] based on a subfield
shortest Euclidean distance with respect to rp(1i ) . After the subcode BCH generator matrix G bch .
information bit vector u (i )
has been obtained from c , (i )
ci
p1 p1 (g) To output decoded codeword , which
(i )
c gv
is then computed by performing the corresponding inverse
c (i )
=u G (i ) (i )
=u G . (g) (i )
(12) permutation λi−1 on c (i) and
gv p1 p2 gv
c ( i ) = (c p(i1) , c gv(i ) + cbch ) = (c p(i1) , c p( i2) ) .
A BCH codeword cbch is then decoded by using
MLD[8] based on Gbch , which cbch + c gv(i ) holds the In Figure 1, the steps (a) and (b) in above proposed
(i ) partition decoding algorithm are realized in the device of
shortest Euclidean distance with respect to r p2 . A
m subvector reliability estimates, while steps (c), (d) and
candidate codeword ci is finally obtained by using the (e) are implemented in the first phase decoder and steps (f)
and (g) are accomplished in the second phase decoder.
corresponding inverse permutation λi
−1
on c (i ) , i.e.,
IV. PERFORMANCE ANALYSIS
ci = λi−1 [c (i ) ] , (13)
Consider that a binary image (mN, mK, dmin)RS(b) of an
where
(N, K, D)RS code over GF(2m) is with BPSK
c (i )
 p2

 transmission over an AWGN channel with zero mean and
c (i ) = (c p1 , c gv + cbch ) = (c p( i1) , c p( i2) ) . (14)
(i ) (i )
variance
N0
. On [11, p398], the error probability for an
2
Hence, based on m subvector reliability values δ
(i )
MLD denoted as PMLD, is
of a received sequence r , a corresponding (15)
PMLD ≈ N dmin Q( 2d min REb / N o )
permutation λi , generator submatrices G p(i1) and G (pi2)
where N d min denotes the error coefficient, and R is
in (11) and a subfield subcode BCH generator
matrix G bch , a proposed partition decoding algorithm is the code rate, and Eb is the received energy per
presented below: information bit and Q (z ) is the Q-function, which is
(a) To partition equally a received sequence r 1 ∞ −x2 / 2 .
into m subvectors, and calculate m subvector
Q( z ) =
2π z
∫ e dx (16)

reliability values δ
( j)
, 1≤ j ≤ m. Since for a binary image (mN, mK, dmin)RS(b) , there
(b) To form a pattern exist m permutations, λi and 1 ≤ i ≤ m , one of them
A = (1,2," , i − 1, i + 1," , m, i) , for which is equal likely employed in such a proposed partition
δ is the least subvector reliability value.
(i ) decoding algorithm at certain time. Therefore, the
(c) To permute a received sequence r with a average error performance of proposed partition decoding
is approached by
permutation λi based on the order in A, m m
P = 1 / m × P ( i ) ≈ 1 / m × N Q ( 2d (i ) RE / N ) (17)
∑ ∑
i.e. λi [r ] = (rp(1i ) , rp(2i ) ) , and err
i =1
err
i =1
(i)
d min min b o

rp(1i ) = (r (1) , ", r ( i −1) , r (i +1) ," , r ( m ) ) and

rp(2i ) = r (i ) = (r(i −1) N +1 ,", riN ) .
 r rp(2i)
c gv(i )
r output ci
c p(i1) Gbch
λ i , G (pi1) , G (pi2) λ−1
i
λ−i 1

Figure 1: Configuration of the decoder adopts proposed partition decoding, in which each decoder employs MLD[8]
based on a corresponding generator submatrix G p(i1) or G bch .
V. SIMULATION AND DISCUSSION
REFERENCE
Simulations of proposed partition decoding and [1] G. D. Forney, Jr., ‘Generalized Minimum Distance Decoding,’
MLD[8] and hybrid decoding[10] are performed on IEEE Trans. Inform. Theory, vol. IT-12, no. 1, pp. 125-131, Apr,
binary images of (N, K, D) RS codes. Without loss of 1966
[2] D. Chase, ‘A New Class for Decoding Block Codes with Channel
generality, an all-zero codeword was transmitted over easurement Information,’ IEEE Trans. Inform. Theory, vol. IT-18,
AWGN channels. Figures 2 show the simulation results no. 2, pp. 170-182, Jan.1972
of decoding. [3] J. K. Wolf, ‘Efficient Maximum Likelihood Decoding of Linear
Block Codes,’ IEEE Trans. Inform. Theory, vol. IT-24, no. 1, pp.
For (155,125,7)RS(b), binary images of (31,25,7)RS a 76-80, Jan. 1978
codes respectively, with BPSK signaling over AWGN [4] A. Vardy and Y. Be'ery, ‘Bit-Level Soft Decision Decoding of
channels, the performances of proposed partition Reed-Solomon Codes,’ IEEE Trans. Commun., vol. 39, no. 3, pp.
440-444, March 1991.
decoding are a little poorer than those of MLD[8] by 1.4 [5] M. P. C. Fossorier and S. Lin, ‘Soft-Decision decoding of Linear
dB at BER 10-5, but better than those of GMD decoding[1] Block codes based on Ordered Statistics Algorithm,’ IEEE Trans.
by 0.8 dB. For SNR of 5dB, proposed partition decoding Inform. Theory, vol. IT-41, no. 5, pp. 1379-1396, Sept. 1995.
[6] M. P. C. Fossorier and S. Lin, ‘Computationally Efficient
only takes 50% to 60% amount of bit operations of the soft-Decision Decoding of Linear Block Codes Based on Ordered
MLD. Statistics,’ IEEE Trans. Inform. Theory, vol. IT-42, no. 3, pp.
For a long code, such as the binary image of an RS 738-750, May 1996.
[7] M. P. C. Fossorier and S. Lin, ‘Computationally Reliability- Based
code, an MLD algorithm is almost impossible to be Decodings of Binary Linear Block Codes,’ IEEE Trans. Inform.
implemented in limit decoding complexity. A feasible Theory, vol. IT-43, no. 5, pp. 1667-1672, Sept. 1997.
decoding algorithm is then required under the constraint [8] D. Gazelle and J. Synders, ‘Reliability-Based Code-search
Algorithms for Maximum-Likelihood Decoding of Block Codes,’
of decoding time delay or the complexity in decoder IEEE Trans. Inform. Theory, vol. IT-43, no. 1, pp. 239-249,
hardware or software. This study presents a proposed January 1997.
partition decoding algorithm for a binary image of an RS [9] F. Hemmati, ‘Closet Coset Decoding of |u|u+v| Codes,’ IEEE
Journal Selec. Areas Commun., vol. 7, no. 6, pp. 982-988 , Aug.
code based on its binary generator matrix and received bit 1989.
reliability. This study has indicated that the error [10] T. H. Hu, and S. Lin, ‘An Efficient Hybrid Decoding Algorithm of
performance of proposed partition decoding with low Binary Images of Reed Solomon Codes Based on Bit Reliability,’
IEEE Trans. Commun., vol. 51, no.7, pp. 1073-1081, July 2003.
complexity is a little worse than that of an MLD, and [11] S. Lin and D. J. Costello, Jr, Error Control Coding, 2nd edition,
better than that of the GMD decoding. Prentice Hall, 2004.

Figure 2: The performance of the proposed partition decoding

algorithm of (155,125,7)RS(b), a binary image of (31,25,7)RS code,
with BPSK transmission over AWGN channels.