Sunteți pe pagina 1din 6

A Comparison of Channel Coding Schemes for 5G

Short Message Transmission


Onurcan İşcan, Diego Lentner, Wen Xu
Huawei Technologies Duesseldorf GmbH
European Research Center, Riesstr. 25, 80992 Munich, Germany
Email: {onurcan.iscan, diego.lentner, wen.dr.xu}@huawei.com

Abstract—Different coding schemes (Turbo, Polar, binary and and compare their error correction performances, as well as
non-binary LDPC and tail-biting convolutional codes) which can their decoding complexity. The paper is structured as follows:
be potentially used in the next generation mobile communication In Section II, we briefly introduce the channel coding schemes
systems for the short message length regime (message length
𝑘 < 512) are discussed. Their error correction performances are considered in this work, and discuss their decoding complexity.
compared with finite length bounds and the complexities of the In Section III, we compare the block error rate performance
corresponding decoders are evaluated in terms of number of basic on the AWGN channel, together with theoretical bounds. We
operations. It is shown that significant performance improvement also compare the decoding complexity of the discussed codes.
can be obtained in the short message length regime by replacing Section IV presents the discussions and concludes the work.
the LTE Turbo code with other modern coding schemes, however,
at a cost of increased decoder complexity. II. C HANNEL C ODES
I. I NTRODUCTION We briefly introduce the considered channel coding schemes
In communication systems, there exists a delay between the and show their decoding complexity. As the decoder com-
generation of the message at the source and the reconstruction plexity is highly related to the used hardware, we will only
of that message at the receiver. Depending on the application, consider the number of basic operations (such as additions and
the delay can have a critical importance. For example, a multiplications) and assume a naive implementation without
delay in orders of seconds might be disturbing for a voice considering the highly optimized decoder structures.
call, whereas for downloading a file from a server, delays We use 𝑘 to denote the message length, 𝑛 to codeword
of several seconds are mostly tolerable. The fifth generation length, and 𝐼 to the number of decoder iterations.
mobile communication systems (5G) will support a multitude
of services where a low latency is required, such as machine A. Convolutional Codes
type communication (MTC) and tactile internet. Design of Convolutional codes are error correction codes that are
the physical layer components becomes critical for supporting widely adopted to many standards due to their many advan-
these highly challenging requirements. tages, like low-complexity encoding, possibility for simple
Channel coding is an important part of the physical layer rate-adaptation by puncturing, hardware-friendly decoding al-
and plays a crucial role in latency and reliability. In gen- gorithms, etc. In contrast to many modern coding schemes,
eral, channel encoders and decoders work block-wise, i.e., their bit-error performance does not improve with increasing
in order the encoder or the decoder to work, the whole message length and hence they are mostly not suitable for
sequence has to be available at the input, which inherently transmitting long messages. However, they still show compa-
causes a delay. One way to reduce this delay is to divide rable performance for short message lengths with maximum-
the messages into smaller parts and encode them separately. likelihood (ML) decoding.
However, the choice of the channel code becomes critical A convolutional code can be defined by the parameters 𝐾,
as the error correction performance degrades with decreased 𝑁 and 𝑚, which denote the input length, output length and the
message lengths. Moreover, for short message length channel number of memory elements, respectively. The convolutional
coding, many existing design and analysis tools (such as EXIT encoder can be defined with a generator polynomial, that
and Density Evolution analysis) become less accurate, as they is mostly given in octal form. Convolutional codes can be
mostly rely on asymptotic results. Also, Shannon’s channel decoded by using the Viterbi algorithm that estimates the
coding theorem becomes a not-suitable metric to evaluate the maximum likelihood sequence using the 𝑆 = 2𝑚 state trellis
performance, since it assumes infinite block lengths (hence representation of the code.
an infinite delay) and a vanishing error probability, which is The convolutional codes used in LTE standard to trans-
mostly not of practical interest. mit control information are tail-biting convolutional codes
In this work, we consider channel coding schemes that have (TBCC). TBCCs do not require a termination which may
the potential to be used in the 5G for short message communi- cause a significant rate loss for short lengths. In this work,
cations. In particular, we consider convolutional codes, Turbo we will consider only TBCC with 𝐾 = 1. In contrast to the
codes, binary and non-binary LDPC codes and Polar codes, terminated convolutional codes, TBCC can start and end at any

978-1-5090-2482-7/16/$31.00 ©2016 IEEE


state which is unknown to the decoder. A common approach passing algorithm. We use 𝑑¯𝑣 and 𝑑¯𝑐 to denote the average
to solve this problem is to use circular decoding by extending variable node degree and the average check node degree,
the decoding trellis by a decoding depth 𝑡 [1], which can be respectively.
chosen as an integer multiple of the constraint length. The LDPC codes are decoded by deploying belief propagation
number of basic operations for a TBCC decoder is given in on a bipartite graph given by their parity check matrix.
Table I. Since the check node operation involves multiple non-linear
functions, one typically relies on sub-optimal approximations
B. Turbo Codes like the min-sum decoder. In this work, we use the scaled
Turbo codes are powerful error correction codes invented min-sum decoder from [7] with the constant scaling factor 0.8.
in the 1990s that are built by parallel concatenating two Similar to the scaled-max-log-MAP decoder for Turbo codes,
convolutional codes, connected via an interleaver, and decoded the scaled-min-sum decoder can reduce the approximation
iteratively by using two soft-in-soft-out (SISO) decoders that error due to min-sum decoding by scaling the outgoing check
exchange extrinsic information during each iteration. Their node messages by a constant factor.
favorable properties such as low complexity encoders, suit- Non-binary LDPC codes can be regarded as an extension
ability to simple rate adaptation, and low complexity iterative of the binary LDPC codes, where the parity check matrix is
decoding schemes made them the channel coding schemes for not binary and has the size 𝑀𝑞 × 𝑁𝑞 , where 𝑁𝑞 = 𝑛/𝑝 and
the 3G and 4G mobile communication standards. Besides, their 𝑀𝑞 = (𝑛 − 𝑘)/𝑝 if defined over the Galois field GF(2𝑝 ).
close-to-capacity error correction performance makes them a The degrees of the variable nodes and the check nodes are
very good choice for many applications. described as 𝑑𝑣 and 𝑑𝑐 .
The iterative decoders of Turbo codes rely on exchanging Non-binary LDPC codes can be decoded using a message
extrinsic information between two constituent decoders that passing decoder, however with increased complexity. In [8],
work well with long messages. Therefore, the performance it was shown that LDPC codes over non-binary Galois fields
of Turbo codes especially for long codewords can be close can perform better than their binary counterparts, especially
to the theoretical limits. Although the iterative algorithm in the short message length regime. Thus, non-binary LDPC
loses its effectiveness with decreased message lengths, their codes have recently attracted an increasing amount of research
performance for short lengths is still comparable with other efforts. In this work, we will also consider short LDPC codes
modern coding schemes. defined over higher order Galois fields decoded with a message
The computationally most challenging part of the Turbo passing decoder. The number of different operations of a non-
decoder can be regarded as the max∗ (𝑥, 𝑦) = max(𝑥, 𝑦) + binary message passing decoder (working on the probability
𝑓𝑐 (𝑥, 𝑦) function in the BCJR algorithm, where 𝑓𝑐 (𝑥, 𝑦) is domain [9]) can be seen in Table I.
a non-linear correction function. To reduce the complexity,
many different methods have been proposed. In this work, we D. Polar Codes
consider a BCJR decoder with scaled-max-log-MAP decoding
[2] with scaling factor 0.75, which gives a good compromise Polar codes are a new class of channel codes introduced by
between complexity and performance. With this method, one Arıkan in 2009 [10]. They have attracted great interest because
neglects the correction function 𝑓𝑐 (𝑥, 𝑦) and scales the ex- they can provably achieve the symmetric capacity of binary-
trinsic L-values at the input of each decoding iteration by the input discrete memoryless channels under low complexity
scaling factor to compensate the loss due to the approximation. successive cancellation (SC) decoding. Polar codes rely on
The number of different operations of a Turbo decoder by the phenomenon called channel polarization which converts
using the scaled-max-log-MAP BCJR algorithm with two the transmission channel into virtual bit channels of different
convolutional codes can be seen in Table I. capacities. As the polarization level grows to infinity, the input
channels are transformed into bit channels which are either
C. LDPC Codes completely noiseless or completely noisy under SC decoding.
LDPC codes were first introduced by R. Gallager in the The fraction of completely noiseless channels approaches the
early 1960s [3] as a class of sparse linear block codes symmetric capacity of the underlying channel [10]. Polar
that are able to approach the channel capacity (in the long encoding is done by assigning input bits to the noiseless
codeword length regime). They were considered impractical bit channels, and by freezing all other bits to known values.
at that time and were neglected for about 30 years. Soon after Extensive research has been done on optimizing polar codes
the introduction of Turbo codes in 1990s, LDPC codes were using different techniques [11], [12], [13].
rediscovered [4], [5] and it is shown that a similar performance Although the asymptotic performance of Polar codes is
to Turbo codes can be obtained [6]. Nowadays, LDPC codes optimal, the finite length performance under SC decoding is
have been adopted by several standards such as WiMAX, not comparable with other modern codes like Turbo and LDPC
DVB-S2, etc. codes. However, by employing successive cancellation list
A binary LDPC code can be defined by a parity check (SCL) decoding and CRC-aided SCL decoding [14], [15] with
matrix 𝐻 of size (𝑛 − 𝑘) × 𝑛. The sparsity of the parity list size 𝐿 and CRC-size 𝑧, the performance can be improved
check matrix allows a low complexity decoder using a message significantly.
TABLE I
D ECODING C OMPLEXITY FOR C HANNEL C ODES IN T ERMS OF BASIC O PERATIONS

Channel Code Additions / Subtractions Multiplications Divisions XOR


TBCC (𝑘 + 2𝑡)𝑆(2𝑁 + 2) - - -
(2𝐼 − 1)
Turbo [(𝑘 + 𝑚)(2𝑆)(2 + 𝑁 ) 2(𝐼 − 1)𝑘 - -
+𝑘(4𝑆 + 1)] + 4(𝐼 − 1)𝑘
LDPC 𝐼𝑛𝑑¯2 𝑣 𝐼(𝑛 − 𝑘)(2𝑑¯𝑐 + 3) - -
𝐼 [𝑁𝑞 𝑑𝑣 (𝑑𝑣 − 1)𝑞
NB-LDPC GF(𝑞 = 2𝑝 ) 𝐼 [𝑁𝑞 𝑑𝑣 (𝑞 − 1) + 𝑀𝑞 𝑑𝑐 𝑞𝑝] +3𝑀𝑞 (𝑑𝑐 − 2)𝑞] 𝐼𝑁𝑞 𝑑𝑣 𝑞 -
+𝑁𝑞 𝑑𝑣 𝑞
3𝐿𝑛 log 𝑛
Polar + CRC 𝐿𝑛/2 log 𝑛 + 𝐿(𝑘 + 𝑛) 𝐿𝑛 log 𝑛 - +𝐿(𝑘 + 𝑛) + 2𝐿𝑘 log 2𝐿
+(𝑘 + 𝑧)𝐿

In this work, we consider Polar codes concatenated with LDPC Codes: We use protograph based AR3A codes [18]
a CRC code, which are decoded with a SCL decoder and which are lifted by using progressive edge growth algorithm to
the CRC is used to select the correct codeword from the list obtain the desired lengths. We decode with a message passing
of codewords at the output of the decoder. The number of decoder using scaled min-sum algorithm where the outgoing
different operations for the CRC-aided SCL decoder can be messages from each check node are scaled by the factor 0.8.
found in Table I. Note that the protograph contains punctured variable nodes
and hence decoding occurs over a larger parity check matrix,
III. E RROR C ORRECTION P ERFORMANCE AND which is also considered during the calculation of the decoder
C OMPUTATIONAL C OMPLEXITY E VALUATION complexity.
Non-Binary LDPC Codes: We use regular Non-binary
In this section, we evaluate the performances of different
LDPC codes in GF(256) with variable node degree 2 and check
coding schemes for short message lengths. In order to simplify
node degree 4, where the non-binary coefficients of the parity
the evaluation, we set the code rate to 𝑅 = 1/2 and evaluate
check matrix is selected randomly from the set of optimized
the block error rate (BLER) performance of different codes
coefficients in [8]. We decode with a message passing decoder
with message length 𝑘 ≤ 512 on an AWGN channel with
in probability domain [9] with 100 iterations.
BPSK modulation by means of Monte Carlo simulation at
different SNR levels, where the simulation is stopped after at Polar Codes: We use Polar+CRC codes where the set of
least 100 block errors. frozen bits are obtained numerically by using density evolution
Since the codes investigated here are of short length, Shan- [11]. CRC of length 8 is used for 𝑘 < 128 bits and CRC of
non’s capacity is not a suitable theoretical bound, because it length 16 is used for 𝑘 ≥ 128 bits. We decode with a SCL
relies on infinite length codewords. Instead, we compare our decoder with 𝐿 = 32 and 𝐿 = 1024.
results with the recently developed finite length performance
bounds of Polyanskiy et al [16], where tight bounds on the A. Error Correction Performance
block error rate performance depending on the message length Fig. 1 to 4 show the block error rate (BLER) performance
are given. Specifically, we use the Gaussian approximation of of the studied codes for message lengths 𝑘 = 64, 128, 256 and
the bounds (also given in [16]) for BPSK modulated AWGN 512, together with the finite length bounds. One can observe
channel. The details of the studied codes are as follows: that except for the TBCC 1, all the other coding schemes
Convolutional Codes: We use two different tail-biting con- outperform LTE Turbo codes and LDPC codes in terms of
volutional codes. TBCC 1 has memory 𝑚 = 6 and has the BLER performance for all lengths. TBCC 2 outperforms LTE
generators 133 and 171, both in octal as in the LTE standard. Turbo code for lengths 𝑘 ≤ 128, and Polar codes with 𝐿 =
TBCC 2 has memory 𝑚 = 10 and has the generators 4672 1024 and NB-LDPC codes perform close to the finite length
and 7542 in octal representation. Both TBCC 1 and TBCC 2 bounds.
are decoded with a decoding depth 𝑡 = 6 ⋅ (𝑚 + 1). Fig. 5 shows the required 𝐸𝑏 /𝑁0 to achieve a target BLER
Turbo Codes: We use the Turbo codes from the LTE of 0.001 for different choices of message lengths and coding
standard with two identical constituent convolutional codes schemes. One observes that performance improvement of the
with the feedforward and feedback generators 15 and 13 in modern codes compared to LTE Turbo code reduces as the
octal, respectively. We puncture the code to obtain the code- message length increases. One can conclude that especially for
rate 𝑅 = 1/2 according to the rate matching method specified very short message lengths (𝑘 ≤ 128), significant improvement
in the LTE standard [17] and decode with 10 iterations of can be obtained by replacing the LTE Turbo code by other
scaled-max-log-map decoding with scaling factor 0.75. modern codes.
100 100

10−1 10−1

10−2 10−2
BLER

BLER
TBCC 1 TBCC 1
TBCC 2 TBCC 2
10−3 LTE Turbo 10−3 LTE Turbo
Polar+CRC 𝐿 = 32 Polar+CRC 𝐿 = 32
Polar+CRC 𝐿 = 1024 Polar+CRC 𝐿 = 1024
10−4 LDPC 10−4 LDPC
NB-LDPC GF(256) NB-LDPC GF(256)
Polyanskiy Approx. Polyanskiy Approx.
10−5 10−5
−2 0 2 4 −2 0 2 4
𝐸𝑏 /𝑁0 [dB] 𝐸𝑏 /𝑁0 [dB]

Fig. 1. Block error rate for different channel codes of rate 𝑅 = 1/2 and Fig. 3. Block error rate for different channel codes of rate 𝑅 = 1/2 and with
with 𝑘 = 64 on AWGN channel with BPSK modulation. message length 𝑘 = 256 bits on AWGN channel with BPSK modulation.

100 100

10−1 10−1

10−2 10−2
BLER

BLER

TBCC 1 TBCC 1
TBCC 2 TBCC 2
10−3 LTE Turbo 10−3 LTE Turbo
Polar+CRC 𝐿 = 32 Polar+CRC 𝐿=32
Polar+CRC 𝐿 = 1024 Polar+CRC 𝐿=1024
10−4 LDPC 10−4 LDPC
NB-LDPC GF(256) NB-LDPC GF(256)
Polyanskiy Approx. Polyanskiy Approx.
10−5 10−5
−2 0 2 4 −2 0 2 4
𝐸𝑏 /𝑁0 [dB] 𝐸𝑏 /𝑁0 [dB]

Fig. 2. Block error rate for different channel codes of rate 𝑅 = 1/2 and with Fig. 4. Block error rate for different channel codes of rate 𝑅 = 1/2 and with
message length 𝑘 = 128 bits on AWGN channel with BPSK modulation. message length 𝑘 = 512 bits on AWGN channel with BPSK modulation.

B. Complexity Comparison terms of memory requirements or any parallelization aspects.


The decoder complexity for different classes of codes de- Except for division, all other basic operations can be real-
pends on many factors, such as the used hardware technol- ized within one cycle, and therefore they can be summed up
ogy and implementation details. Moreover, if implemented in to give a rough complexity figure of merit (FOM). For a small
hardware optimizations can be done individually for each of number of divisions, this operation can usually be replaced by
the decoding methods, which would make a fair comparison look-up tables and multiplications. For this reason, we also
difficult. Note that lots of work has been done in reducing sum up the divisions and include them into the complexity
the decoder complexity of the discussed codes. In this work FOM.
we assume a naive implementation of the decoders and use Fig. 6 shows the number of basic operations for different de-
the number of basic operations (additions, subtractions, mul- coders and message lengths. One can observe that NB-LDPC
tiplications, divisions and binary additions) as our complexity codes in GF(256) and Polar+CRC codes with 𝐿 = 1024 show
metric, as given in Table I. We do not consider complexity in the highest decoder complexity, with significant performance
BPSK Constr. Cap. 108
5 Poly. Apr. for BPSK
TBCC 1
TBCC 2 107
𝐸𝑏 /𝑁0 [dB] @ BLER = 0.001

4 LTE Turbo
Polar+CRC 𝐿=32

Number of Operations
Polar+CRC 𝐿=1024 106
LDPC
3
NB-LDPC GF(256)
105
TBCC 1
2 TBCC 2
104 LTE Turbo
Polar+CRC 𝐿 = 32
1 Polar+CRC 𝐿 = 1024
103
LDPC
NB-LDPC GF(256)
0 102
0 100 200 300 400 500 100 200 300 400 500
Message Length 𝑘 [bits] Message Length 𝑘 [bits]

Fig. 5. Required 𝐸𝑏 /𝑁0 to achieve a target BLER with channel codes of rate Fig. 6. Decoder complexity in terms of number of basic operations for
𝑅 = 1/2 on AWGN channel with BPSK modulation for different message different message lengths.
lengths 𝑘 ≤ 512, together with Shannon’s BPSK constrained capacity, and
the Gaussian approximation of Polynaskiy’s finite length bound [16].

NB-LDPC codes. Therefore, TBCC can also be consid-


gains in terms of BLER, whereas LTE Turbo codes, LDPC ered as a good alternative for the very short messages.
codes, and Polar+CRC with 𝐿 = 32 show similar decoding ∙ Polar+CRC codes with list decoding performs especially
complexity. The complexity of the tail-biting convolutional well with large list sizes, which usually also have a high
codes scales with the number of states. The complexity of decoding complexity. One advantage of the list decoding
TBCC 1 with 𝑚 = 6 is lower than other candidates, whereas is that the receiver can adjust the list size according to its
the TBCC 2 with 𝑚 = 10 has complexity higher than the LTE computational power, which makes SCL flexible in terms
Turbo code. of decoding complexity.
∙ NB-LDPC codes also show promising performance.
However, similar to Polar codes with large list sizes, their
IV. D ISCUSSIONS AND C ONCLUSION
decoding complexity is very high.
According to our simulation results, one can make the Note that the decoding performance of the discussed codes
following interpretations: mainly depend on two aspects: code properties and decoder
∙ LTE Turbo codes and LDPC codes show a similar error- properties. As one increases the list size of a Polar decoder,
correction performance at comparable decoder complex- one obtains a close-to-ML decoder. Therefore for very large
ities. Across all considered message lengths, both codes list sizes, code properties (such as the minimum distance)
have a gap of 1 ∼ 2.5dB to the finite length bounds which have an important role. Similarly, NB-LDPC decoders process
can be especially important for short message lengths. block of bits jointly (instead of bitwise like the binary LDPC
By using other modern channel coding schemes such codes). As the field order grows, more bits are processed
as NB-LDPC codes over higher order Galois fields or jointly, which improves the decoding performance.
Polar+CRC codes with SCL decoding, one can reduce The relatively poor performance of LTE Turbo codes and
this gap and improve the error correction performance binary LDPC codes in the short message length regime is
by up to 1dB. However, this performance improvement partly due to the suboptimal iterative decoding method. By
would come at a cost of increased decoder complexity. decoding the Turbo or LDPC code with an ML decoder,
∙ Tail-biting convolutional codes with medium size mem- one can also improve the performance, at a cost of higher
ory elements perform well for very short message lengths computational complexity. Some simulation results on ML
(such as 64 bits), and have decoders with lower complex- decoding of Turbo codes and LPDC codes can be found in
ity. However, their error correction performance is not [19], [20].
well suited for longer messages. For very short message To sum up, the short message length channel coding per-
lengths, one can obtain further improvement by utilizing formance of the next generation systems can be significantly
codes with higher memory such that the performance improved by using better codes (e.g. with better distance prop-
becomes comparable to other modern channel codes. The erties), or more importantly by allowing higher computational
complexity can still be kept lower than Polar codes and complexity at the receivers. How to enhance existing codes to
have better code properties (e.g. minimum distance) and how
to lower the decoding complexities of modern codes remain
as interesting future work.
ACKNOWLEDGMENT
This work has partly been performed in the framework
of the Horizon 2020 project FANTASTIC-5G (ICT-671660)
receiving funds from the European Union. The authors would
like to acknowledge the contributions of their colleagues in the
project, although the views expressed in this work are those
of the authors and do not necessarily represent the project.
The authors also thank Giuliano Garrammone for his help on
designing NB-LDPC codes.
R EFERENCES
[1] W. Sung, “Minimum decoding trellis lengths for tail-biting convolutional
codes,” Elect. Let., vol. 36, no. 7, pp. 643–645, 2000.
[2] J. Vogt and A. Finger, “Improving the max-log-MAP turbo decoder,”
Elect. Let., vol. 36, no. 23, p. 1, 2000.
[3] R. Gallager, “Low-density parity-check codes,” IRE Trans. on Inf.
Theory, vol. 8, no. 1, pp. 21–28, 1962.
[4] D. J. MacKay and R. M. Neal, “Near shannon limit performance of low
density parity check codes,” Elect. Let., vol. 32, no. 18, pp. 1645–1646,
1996.
[5] D. J. MacKay, “Good error-correcting codes based on very sparse
matrices,” IEEE Trans. Inf. Theory, vol. 45, no. 2, pp. 399–431, 1999.
[6] T. J. Richardson, M. A. Shokrollahi, and R. L. Urbanke, “Design of
capacity-approaching irregular low-density parity-check codes,” IEEE
Trans. Inf. Theory, vol. 47, no. 2, pp. 619–637, 2001.
[7] J. Chen, A. Dholakia, E. Eleftheriou, M. P. C. Fossorier, and X.-Y. Hu,
“Reduced-complexity decoding of LDPC codes,” IEEE Tran. on Comm.,
vol. 53, no. 8, pp. 1288–1299, August 2005.
[8] C. Poulliat, M. Fossorier, and D. Declercq, “Design of regular (2, 𝑑𝑐 )-
LDPC codes over GF(q) using their binary images,” IEEE Trans. on
Comm., vol. 56, no. 10, pp. 1626–1635, 2008.
[9] L. Costantini, B. Matuz, G. Liva, E. Paolini, and M. Chiani, “Non-binary
protograph low-density parity-check codes for space communications,”
Int. Journal of Sat. Commu. and Net., vol. 30, no. 2, pp. 43–51, 2012.
[10] E. Arikan, “Channel polarization: A method for constructing capacity-
achieving codes for symmetric binary-input memoryless channels,” IEEE
Trans. Inf. Theory, vol. 55, no. 7, pp. 3051–3073, July 2009.
[11] R. Mori and T. Tanaka, “Performance and construction of polar codes
on symmetric binary-input memoryless channels,” in Proceedings of the
2009 IEEE Int. Symp. on Inf. Theory (ISIT), June 2009, pp. 1496–1500.
[12] P. Trifonov, “Efficient design and decoding of polar codes,” IEEE Trans.
on Comm., vol. 60, no. 11, pp. 3221–3227, November 2012.
[13] I. Tal and A. Vardy, “How to construct polar codes,” IEEE Trans. Inf.
Theory, vol. 59, no. 10, pp. 6562–6582, October 2013.
[14] ——, “List decoding of polar codes,” IEEE Trans. Inf. Theory, vol. 61,
no. 5, pp. 2213–2226, May 2015.
[15] A. Balatsoukas-Stimming, M. B. Parizi, and A. Burg, “LLR-based
successive cancellation list decoding of polar codes,” IEEE Trans. on
Sig. Proc., vol. 63, no. 19, pp. 5165–5179, October 2015.
[16] Y. Polyanskiy, H. V. Poor, and S. Verdú, “Channel coding rate in the
finite blocklength regime,” IEEE Trans. Inf. Theory, vol. 56, no. 5, pp.
2307–2359, 2010.
[17] T. ETSI, “136 212 LTE; Evolved universal terrestrial radio access (E-
UTRA); Multiplexing and channel coding,” 2013.
[18] D. Divsalar, S. Dolinar, and C. Jones, “Low-rate LDPC codes with
simple protograph structure,” in Proceedings. Int. Symp. on Inf. Theory,
2005. ISIT 2005. IEEE, 2005, pp. 1622–1626.
[19] M. Helmling and S. Scholl, “Database of channel codes and ML simula-
tion results,” University of Kaiserslautern URL: www.uni-kl.de/channel-
codes, 2016.
[20] S. Scholl, F. Kienle, M. Helmling, and S. Ruzika, “ML vs. BP decoding
of binary and non-binary LDPC codes,” in 7th Int. Symp. on Turbo
Codes and It. Inf. Proc. (ISTC), August 2012, pp. 71–75.