See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/333913937

Multi-Stage Bit-Flipping Decoding Algorithms for LDPC Codes

Article in IEEE Communications Letters · June 2019

DOI: 10.1109/LCOMM.2019.2924210

CITATIONS READS
9 192

3 authors:

Tofar C.-Y. Chang Pin-Han Wang

National Taipei University of Technology National Chiao Tung University
14 PUBLICATIONS 68 CITATIONS 3 PUBLICATIONS 18 CITATIONS

SEE PROFILE SEE PROFILE

Yu T. Su
National Chiao Tung University
142 PUBLICATIONS 1,670 CITATIONS

SEE PROFILE

All content following this page was uploaded by Yu T. Su on 06 July 2020.

The user has requested enhancement of the downloaded file.

 1

Multi-Stage Bit-Flipping Decoding Algorithms for

LDPC Codes
Tofar C.-Y. Chang, Member, IEEE, Pin-Han Wang, and Yu T. Su, Senior Member, IEEE

Abstract—We present two general multi-stage (MS) bit-flipping and Elsabrouty [9] presented another two-stage decoder using
(BF) decoding algorithms for low-density parity-check (LDPC) the BP algorithm for the first stage and an SD BF algorithm in
codes. Both algorithms consist of soft-decision (SD) and hard- the second stage. In [10], Li el at. used an SD BF algorithm to
decision BF decoding parts. In comparison with known MS
LDPC decoders, our approach is much simpler as all stages share start decoding and when this SD BF decoder fails after a fixed
the same BF structure. The only complexity increase is due to the number of iterations, a sophisticated OSD takes over using the
use of an adaptive stage-switching (SS) mechanism which gives accumulated bit reliability and the associated syndrome.
near-optimal SS timing. A new design issue we address is that the Sequential multi-stage (MS) decoding seems to be a simple
first-stage algorithm’s parameter has to be re-tuned to achieve and straightforward idea to enhance decoder performance. But
the optimal overall performance. Numerical results demonstrate
that the proposed decoding methods can significantly improve there also arise the problems of implementation complexity
the error-rate performance of conventional SD BF decoders. and the timing of stage-switching (SS), which, unfortunately,
are seldom addressed in the MS literature. For the latter
Index Terms—LDPC codes, bit-flipping decoding, loop detec-
tion. issue, the above-mentioned approaches all set a fixed iteration
number threshold for the first-stage algorithm and switch to the
second stage when the iteration counter reaches this threshold.
I. I NTRODUCTION The predetermined threshold can only be found by simulations
The bit-flipping (BF) algorithm, originally introduced by and is a function of the signal-to-noise ratio (SNR, Eb /N0 ) and
Gallager, is a low-complexity method for decoding low- the code and the decoding algorithm used. The implementation
density parity-check (LDPC) codes [1]. The performance of complexity (including storage) has to do with the algorithms
the BF algorithm, however, is much inferior to that of the used and, as evidenced by the above description, these known
sum-product algorithm, which is also known as the belief MS approaches need substantial complexity increases when
propagation (BP) algorithm. To reduce the performance gap, compared with their single-stage counterparts.
many variants which use soft-decision (SD) information to To solve the SS problem, we first observe that an SD
improve the reliability estimate of the bit decisions, e.g., the BF decoding algorithm is likely to exhibit an oscillatory bit
gradient descent BF (GDBF) algorithm [2] and the family of decision pattern after a certain number of iterations [12], i.e.,
weighted BF (WBF) algorithms [3]-[6], have been introduced. an illegitimate decoded vector reappears periodically, and the
The improvements brought about by these variations, however, decoding process is trapped in an infinite loop that eventually
are rather limited. So, one is tempted to ask if there is any low- leads to a decoding failure. Thus a reasonable SS time is
complexity approach to further improve these SD BF methods. when such a loop is detected and so the MS decoder is to
This question leads us to review proposals that combine be accompanied by a proper loop detector. Loop detection has
two or more existing LDPC decoding algorithms by acti- been used by some SD BF decoders to escape from the loop
vating them sequentially [7]-[10] to improve the error-rate by inserting a perturbation iteration once a loop is detected
or convergence rate performance. Jiang et al. [7] proposed [12]. In the perturbation iteration, the BF rule is changed and
a two-stage decoder in which a conventional BP algorithm the decoder returns to the original BF rule afterward. We will
is used in the first stage and, after I iterations, an ordered- explain in Section III-A that re-activating the previous SD
statistic decoder (OSD) takes over the decoding task if a valid algorithm or other SD BF rules runs the risk of falling into
codeword is not found. Unlike previous BP-OSD decoders, the another loop. Thus we take a different approach: having the
OSD in [7] uses a weighted accumulated log-likelihood ratio Gallager’s hard-decision (HD) BF algorithm takes over the
(LLR) metric that involves past I LLR. The optimal weighting decoding task. As Gallager’s HD BF decoder uses a flipping
(damping) factor depends on I and the code used and has to be function (FF) much simpler than those of SD BF decoders and
found by simulations. Li and Zhang [8] suggested a two-stage requires only binary operation, our approach not only takes the
approach that uses a so-called fast WBF algorithm for the first complexity issue mentioned above into consideration but also
I iterations, followed by a conventional BP decoder. Zeidan enjoys the advantage that the Gallager decoder can be used in
conjunction with any SD BF algorithm.
Manuscript received May 13, 2019; accepted June 10, 2019. This work was
supported in part by Taiwans Ministry of Science and Technology under Grant Besides simplicity, our use of Gallager’s HD BF decoder as
MOST 107-2221-E-009-129-. The associate editor coordinating the review the second-stage decoder is also motivated by the following
of this letter and approving it for publication was C. Feng. (Corresponding argument. We first observe that it is very likely that the residual
author: Yu T. Su.)
The authors are with National Chiao Tung University, Hsinchu 30010, Taiwan error count is small when a decoding failure (loop) occurs.
(e-mail: [email protected]). Hence, a desired property of the second-stage decoder is that
2

it is capable of correcting the small number of errors left by The FF in a BF decoder is used to estimate the reliability
the first-stage decoder with high probability. We will show in of decoded decisions. In [1], the following FF
Section III-A that Gallager’s BF decoder not only guarantees X
certain error correcting capability [11] but can correct far more En = − (1 − 2sm ) (1)
m∈M(n)
errors beyond the theoretical bound with a reasonably high
probability, i.e., it does posses the above desired property. is proposed to estimate the reliability of the decisions ûn ,
Combining an SD BF algorithm with Gallager’s BF decoder where s = (s0 , s1 , · · · , sM−1 ) = û · H T (mod 2) is the syn-
in cascading and employing a loop detector to determine the drome (checksum) vector. In each iteration, the decoder com-
SS time is therefore a low-complexity solution for enhancing putes the FF and makes bit decisions on those VNs with the
the performance of an SD BF decoder. To further improve maximum FF value, i.e., all ûi ’s where Ei = max0≤n<N En ,
the performance, we have to consider the case when a loop are selected for flipping (ûi = ûi + 1 (mod 2)). Then, s is
does occur again in the second stage. To get out of the second renewed based on the updated û. If s = 0 or the iteration
loop with minimum additional complexity, we use a simple counter, l, reaches the maximum allowed iteration number
variation of the Gallager’s HD BF decoder, to be described in lmax , the decoding process stops.
Section III-B, in the third stage as the loop-breaking decoder. By adopting the soft-valued channel information and as-
As the time the HD BF decoder takes over coincides with signing checksum weights, the WBF algorithm and its variants
the time a loop occurs and the probability for the decoder give improved error-rate performance with respect to the above
to successful decode depends on the number of residual Gallager’s BF decoder. The IMWBF [4] and the reliability
erroneous bits, the first-stage algorithm should try to minimize ratio based WBF (RRWBF) [5] algorithms are perhaps the
the (conditional) average residual erroneous bit number in two most well-known WBF algorithms. The FF of the IMWBF
a codeword given a decoding failure has occurred. Such a algorithm is given by
design goal may not be consistent with that of minimizing X
(I)
either the frame (codeword) error rate (FER) or the bit error En = −α|yn | − wmn (1 − 2sm ), α > 0 (2)
m∈M(n)
rate (BER). Hence, we introduce a new metric measuring the
conditional per-frame residual erroneous bit number. For some where (I)
wmn = minn′ ∈N (m)\n |yn′ |. If we define wmn (R)
=
P
SD BF algorithms, it is the scaling factor in the FF that affects |y |/y , the FF of the RRWBF algorithm can be
n ∈N (m) n
′ ′ n
the performance metric of interest, e.g., FER, BER or the written as
conditional residual error count defined in (5). As an example, X
(R)
we find the optimal scaling factor for the improved modified En = − wmn (1 − 2sm ), (3)
WBF (IMWBF) algorithm to enhance the overall performance; m∈M(n)

similar optimization for other SD BF algorithms can bring In [6], the mixed modified WBF (MMWBF) algorithm uses
about improved performance as well. (2) and (3) as the main and auxiliary FFs and selects one or
The numerical results presented in Section IV demonstrate two bits for flipping in each iteration. The GDBF algorithm
that our methods bring about 1 − 1.5 dB gain no matter which [2] also adopts the soft-valued information in its FF:
SD BF decoder is used in the first stage. Optimizing the scaling X
factors further improve the performance of the MSBF (multi- En = −x̂n yn − (1 − 2sm ), (4)
stage bit-flipping) decoders. m∈M(n)

The rest of this letter is organized as follows. In Section where x̂n = 1 − 2ûn . The GDBF algorithm can outperform
II, we briefly review the BF algorithms and the FFs used. In WBF algorithms in decoding some LDPC codes.
Section III, we introduce two MSBF decoding algorithms–a A quick review of the above FFs reveals that Gallager’s
two-stage algorithm and a three-stage one–and explain their FF (1) is a special case of all the rest. Hence, implementing
underlying design considerations. Numerical results and dis- Gallager’s algorithm in the same decoder incurs no complexity
cussion on the scaling factor are given in Section IV. Finally, increase except for the need to set some other parameters to
conclusion remarks are drawn in Section V. one or zero.

II. BF D ECODING FOR LDPC C ODES III. M ULTI -S TAGE BF D ECODING A LGORITHMS
We denote by (N, K)(dv , dc ) a regular binary LDPC code A. The choice of the second-stage decoder
C of rate R = K/N having an M × N parity check As mentioned in Section I, the SD BF decoders may fall
matrix H = [Hmn ] with variable node (VN) degree dv into an infinite decoding loop in which a tentative decoded
and check node (CN) degree dc . A binary codeword u = vector repeatedly reappears and the decoder flips the same set
(u0 , u1 , · · · , uN −1 ) is BPSK-modulated and transmitted over of bits periodically, leading to a decoding failure.
an AWGN channel with noise variance σ 2 . The corresponding Fig. 1 provides the simulated cumulated density function
received and decoded binary sequences are denoted by y = (CDF) of the residual errors in an uncorrected frame when
(y0 , y1 , · · · , yN −1 ) and û = (û0 , û1 , · · · , ûN −1 ), respectively. different SD BF decoders are used to decode the (1008, 504)
We denote the nth VN by vn and the mth CN by cm ; we define regular PEG (denoted by PEGReg 504x1008 ) code [13] and
N (m) as the index set of VNs linked to cm and M(n) as that a decoding failure occurs. We observe that there remains only
of CNs connected to vn in the code graph. few uncorrected bits especially if the SNR is high.
 3

1
1

5 dB 0.975

6 dB

Probability of Correct Decoding

0.95

0.95
4 dB
CDF

SNR=4 dB
0.9 0.925
IMWBF 4dB RRWBF

GDBF MMWBF

SNR=5 dB 0.9

IMWBF RRWBF
0.85
GDBF MMWBF

SNR=6 dB 0.875

IMWBF RRWBF

GDBF MMWBF
0.8 0.85

0 5 10 15 20 25 30 35 40 45 0 10 20 30 40

Number of Incorrect Bits in an Input Frame

Number of Residual Erroneous Bits

Fig. 1: The CDF of uncorrected bit numbers after 200 iterations or a loop is Fig. 2: Correct decoding probability as a function of input error bit number
detected; PEGReg 504x1008 code. for Gallager’s HD BF decoder; PEGReg 504x1008 code (dv = 3, g = 8).

The above observation and the concern of implementation process after saying L1 iterations have passed and check if
complexity, as mentioned before, motivate us to limit our there exists an identical û only in the previous L2 iterations.
choice of the next stage decoder to BF decoders. Another Both parameters L1 and L2 can be easily found empirically,
critical observation is that during the SD BF decoding process, and the storage requirement can be greatly reduced with proper
a very large yn may dominate the FF and a very small yn choices of these two parameters. Note that this loop detector
could make the corresponding weight(s) negligible and the is independent of the BF algorithm used in the first stage
associated checksum sm useless. Oftentimes, these extreme and there is no detector parameter which may depend on the
channel values are the major causes for SD BF decoders to be operating SNR. The first-stage algorithm and the operating
trapped. Therefore, we need to choose a decoder that has little SNR will automatically determine the time a decoding loop
or nothing to do with the channel values, which then rules out occurs.
the use of most SD BF algorithms.
On the other hand, Fig. 2 indicates that the Gallager’s BF Algorithm 1 ABCDE
decoder can correct not only all error patterns of weight less 1: Set the iteration counter l = 1.
than (dv + 1)/2 if the code’s girth g is 6 and up to dv − 1 2: for i = 1, ..., M
when g = 8 [11] but many more error patterns having much 3: for j = 1, ...N (i)
larger weights (< 30 in this case) with high probability. 4: Generate and propagate...
5: end for
B. A Two-Stage Decoding Algorithm 6: end for
The above discussions convince us that using Gallager’s HD
BF algorithm as the second-stage decoder when the first-stage C. A Three-Stage Decoding Algorithm
SD BF decoder falls into a loop yields a high probability of Using Gallager’s BF decoder as the second-stage decoder
eliminating the residual errors left by the first-stage decoder. In in Algorithm ?? still has a chance of being trapped especially
other words, Gallager’s BF decoder is a near-optimal candidate when the weights of residual errors in the first stage are
loop breaker when taking simplicity and decoding capability not small. When a second loop occurs, one may be tempted
and the ability of getting out of a loop into consideration. to reapply the same concept using another HD BF decoder
The resulting two-stage BF (2SBF) algorithm is summarized to break the loop. However, there is no HD BF decoder
in Algorithm ??. simpler than the Gallager’s. Another choice is to switch back
The above algorithm requires that a loop detector be avail- to the first-stage SD BF algorithm. But the FFs of the SD
able. Let û(l) denote the decoded result û corresponding to the BF decoders, (2), (3), are all functions of channel values.
lth iteration. Suppose the current iteration index is i. In gen- As pointed out in Section III-A, the extreme channel values
eral, a loop detector needs to check if û(i) is the same as any either dominate the FF or diminish the usage of checksums,
of previous iterations’ decoding result û(j) , 1 ≤ j < i. Hence thus if we reapply the same first-stage algorithm, similar error
we sequentially check if û(i) = û(j) , j = i − 1, i − 2, . . . , 1. patterns tend to reappear and the decoder just falls into a
In case ℓ = min{k|k > 0, û(i) = û(i−k) }, i.e., ℓ is the loop again. Recall that when a loop occurs, the decoder gives
smallest positive integer such that û(i) = û(i−k) , we then say oscillating FF values or stays at a local plateaus. A simple
a loop of length ℓ is detected. To keep all û(i) will necessitate method to escape the local valley or plateaus is by performing
some additional storage space. We note that the loop usually random moves.
occurs only after a certain iterations and the loop length is To minimize the decoder complexity, a natural choice is to
often not large. Hence we can activate the loop detection randomize the second-stage decoder’s decision which leads to
4

1 1
50 Iterations
IMWBF-2SBF
IMWBF IMWBF-2SBF
switch at the 20th ite.
IMWBF-3SBF
switch at the 100th ite.
0.1 RRWBF RRWBF-2SBF
switch as a loop is detected
RRWBF-3SBF 0.1
IMWBF-3SBF
MMWBF MMWBF-2SBF

MMWBF-3SBF
switch at the 20th and 40th ite.
0.01
GDBF GDBF-2SBF switch at the 100th and 200th ite.

GDBF-3SBF switch as a loop is detected

0.01

FER
FER

m-GDBF m-GDBF-2SBF
1E-3
m-GDBF-3SBF

1E-4 1E-3

1E-5

1E-4

4 5 6 7 8 0 50 100 150 200 250 300

SNR (dB)
Iteration

Fig. 3: FER performance of various SD BF decoding algorithms and their Fig. 5: Decoding performance of the IMWBF-based MSBF algorithms with
MSBF versions; 50 iterations in total, PEGReg 504x1008 code. different SS conditions; SNR=5.5 dB, PEGReg 504x1008 code.

1
200 Iterations

IMWBF IMWBF-2SBF

IMWBF-3SBF
the multi-bit flipping version of the GDBF decoder [2]. For
0.1 RRWBF RRWBF-2SBF the IMWBF, MMWBF, IMWBF-2SBF/3SBF, and MMWBF-
RRWBF-3SBF

MMWBF MMWBF-2SBF 2SBF/2SBF algorithms, we set α = 0.2. It can be seen that

0.01
MMWBF-3SBF

GDBF GDBF-2SBF
the GDBF-2SBF and m-GDBF-2SBF algorithms yield about
GDBF-3SBF 1 dB gain against the GDBF decoder at FER ≈ 10−5 , and the
WBF-based 2SBF algorithms offer at least 1.5 dB gain against
FER

m-GDBF m-GDBF-2SBF
1E-3
m-GDBF-3SBF

the corresponding original SD BF decoders for the same FER.

1E-4 We also found that, compared with the 2SBF algorithm, our
3SBF algorithm provides better converged FER at low and
1E-5 moderate SNRs and gives approximately 0.25 dB gain at FER
≈ 10−5 .
4 5 6 7 8 To demonstrate the efficacy of the proposed loop-detection
SNR (dB) based SS criterion, in Fig. 5, we compare the performance
Fig. 4: FER performance of various SD BF decoding algorithms and their
of our 2SBF and 3SBF algorithms and the IMWBF-2SBF
MSBF versions; 200 iterations in total, PEGReg 504x1008 code. algorithm with a fixed first-stage iteration number (20 or 100),
and the IMWBF-3SBF algorithm with SS times at either the
20th/40th or the 100th/200th iterations. Those curves clearly
show that an early switching (at the 20th/40th iterations)
a random BF (RBF) decoder. The RBF decoder uses (1) as
results in significant performance loss while a late switching
the FF and randomly selects q (0 < q ≤ p) bits from F to
(at the 100th/200th iterations) spends redundant computing
flip where F is the set of those bits having the maximum FF
effort in the first or the second stage and causes much slower
values and p is the cardinality of F .
convergence. Our loop-detection based method does give near-
This three-stage algorithm which appends a RBF stage to
optimal (if not the optimal) switching times and better FER
Algorithm ?? is called three-stage BF (3SBF) algorithm. The
and convergence performance. The benefit of having the third-
modified loop detection scheme mentioned in the last section
stage decoder can be found in the same figure as well.
is also used to decide when the decoder should switch from the
For the FF (2), which is used by the IMWBF and MMWBF
second stage (Gallager’s BF decoder) to the third stage (RBF
decoders, the value of the scaling factor α is usually selected
decoder). However, the additional decoding iteration number
to minimize the converged error-rate performance. However,
in the second stage is often relative small, so L1 should be set
these values may not give the optimal performance for the
to zero. Note that only those û’s in the same decoding stage
2SBF and 3SBF algorithms since both SD and HD BF
are to be monitored by the loop detector.
decoders are involved in our design. As a larger number of
uncorrected bits left by the SD BF decoder will degrade the
IV. N UMERICAL R ESULTS Gallager’s HD decoder’s performance, the relation between α
Figs. 3 and 4 show the FER performance of various SD and the residual error numbers should be considered in select-
BF decoders and the corresponding 2SBF and 3SBF decoders ing α. Actually, less residual errors from the SD BF decoder
for decoding the PEGReg 504x1008 code with L2 = 20 and ensures that the HD BF decoder has a higher probability of
lmax = 50 or 200. The 2SBF (3SBF) algorithm is labeled successfully correcting them.
by χ-2SBF (χ-3SBF) in the figures when an SD BF decoder Fig. 6 depicts the converged BER (denoted by PB (α))
χ is employed in the first stage. m-GDBF in Fig. 3 denotes and FER performance (denoted by PF (α)) of the IMWBF
 5

properly concatenating SD and HD BF decoders, the proposed

10
2SBF algorithm provides significant performance gain against
( )
known SD BF algorithms. With an RBF decoder in the third
1
stage, the 3SBF algorithm gives further performance gain.
( )

SNR=4 dB
Rationales for choosing the second- and third-stage algorithms
0.1
PF( ) SNR=5 dB are given and a simple loop detector is employed to determine
SNR=6 dB

PF( )
near-optimal SS time. The resulting dynamic SS timing helps
0.01

SNR=4 dB to improve both FER and convergence performance. We also

SNR=5 dB

SNR=6 dB
find that it is necessary to modify the first-stage SD BF
1E-3
PB( ) algorithm and suggest a new metric for selecting a proper
PB( )

PB( )
SNR=4 dB

SNR=5 dB
scaling parameter value to optimize the overall performance
1E-4
SNR=6 dB using the IMWBF based 2SBF algorithm as an example.
0.0 0.5 1.0 1.5 2.0 We show that although the optimal scaling parameter α is
a function of SNR, a near-optimal α which is robust to SNR
Fig. 6: BER (PB (α)), FER (PF (α)), and ρ(α) of the IMWBF algorithm can be found.
with different α’s; PEGReg 504x1008 code.
R EFERENCES
1 [1] R. G. Gallager, Low-density parity-check codes, Cambridge, MA: MIT
Press, 1963.
[2] T. Wadayamaet et al., “Gradient descent bit flipping algorithms for
4 dB
0.1 decoding LDPC codes,” IEEE Trans. Commun., vol. 58, no. 6, pp. 1610
1614, June 2010.
[3] Y. Kou, S. Lin, and M. P. C. Fossorier “Low-density parity check codes
0.01 based on finite geometries: a rediscovery and new results,” IEEE Trans.
Inform. Theory, vol. 77, pp. 2711-2736, Nov. 2001.
5 dB [4] M. Jiang, C. Zhao, Z. Shi, and Y. Chen, “An improvement on the
FER

IMWBF-2SBF, SNR=4 dB

1E-3 =0.2 =0.5

modified weighted bit flipping decoding algorithm for LDPC codes,”
=1.0
IEEE Commun. Lett., vol. 9, pp. 814-816, Sep. 2005.
IMWBF-2SBF, SNR=5 dB
[5] F. Guo and L. Hanzo, “Reliability ratio based weighted bit-flipping
1E-4
=0.2 =0.5 decoding for low-density parity-check codes,” IEEE Electron. Lett., vol.
=1.0
6 dB
40, no. 21, pp. 1356-1358, Oct. 2004.
IMWBF-2SBF, SNR=6 dB [6] H. Huang, Y. Wang, and G. Wei, “Mixed modified weighted bit-flipping
=0.2 =0.5 decoding of low-density parity-check codes,” IET Commun., vol. 9, no.
1E-5
=1.0 2, pp. 283-290, Jan. 2015.
20 40 60 80 100
[7] M. Jiang, C. Zaho, E. Xu, and L. Zhang, “Reliability-based iterative de-
Iteration coding of LDPC codes using likelihood accumulation,” IEEE Commun.
Lett., vol. 11, no. 8, pp. 677-679, Aug. 2007.
Fig. 7: FER performance of IMWBF-2SBF algorithms with different α’s; [8] J. Li and X. Zhang, “Hybrid iterative decoding for low-density parity-
PEGReg 504x1008 code. check codes based on finite geometries,” IEEE Commun. Lett., vol. 12,
no. 1, pp. 29-31, Jan. 2008.
[9] H. R. Zeidan and M. M. Elsabrouty, “Modified iterative two-stage hybrid
decoding algorithm for low-density parity-check (LDPC) codes,” in
algorithm when different α values are used in the FF. Also Proc. IEEE Veh. Technol. Conf., pp. 15, Apr. 2009.
[10] G. Li, X. Zou, and X. Wang, “Hybrid decoding for one class of low-
shown in the figure is the conditional average residual error density parity-check codes,” in Proc. 3rd Int. Conf. Commun. and Netw.
numbers (in an unsuccessfully decoded frame) using different in China, Hangzhou, 2008, pp. 340-343.
α’s, which is denoted by ρ(α) and to be computed by [11] W.-Y. Chen and C.-C. Lu, “On error correction capability of bit-flipping
algorithm for LDPC codes,” in Proc. IEEE Int. Symp. Inform. Theory,
N St. Petersburg, Russia, Jul./Aug. 2011, pp. 1283-1286.
ρ(α) = PB (α) · . (5) [12] Z. Liu and D. A. Pados, “A decoding algorithm for finite-geometry
PF (α) LDPC codes,” IEEE Trans. Commun., vol. 53, pp. 415-421, Mar. 2005.
[13] D. J. C. MacKay, Encyclopedia of sparse graph codes [Online]. Availi-
We observe that the best BER and FER performance of the ble: http://www.inference.phy.cam.ac.uk/mackay/codes/data.html
IMWBF decoder is achieved at α ≈ 0.2 while the minimum
ρ(α) is a function of SNR and lies within α ∈ [0.5, 1]. In Fig.
7, we compare the decoding performance of the IMWBF-2SBF
algorithm with different α’s. It can be seen that the IMWBF-
2SBF algorithm yields better converged FER performance
when α = 0.5 or 1. Since a lower ρ(α) indicates fewer residual
bit errors, the HD BF decoder thus has a higher probability to
correct these errors.

V. C ONCLUSION
In this letter, we present two general MSBF decoding
algorithms for LDPC codes. Our algorithms are general for any
SD BF algorithm that can be used as the first-stage decoder. By

View publication stats