1

Short Block-length Codes for Ultra-Reliable Low

Latency Communications
Mahyar Shirvanimoghaddam, Mohamad Sadegh Mohamadi, Rana Abbas, Aleksandar Minja, Chentao Yue, Balazs
Matuz, Guojun Han, Zihuai Lin, Yonghui Li, Sarah Johnson, and Branka Vucetic

Abstract—This paper reviews the state of the art channel rigorous latency and reliability levels. For example, power
coding techniques for ultra-reliable low latency communication electronics based industrial control needs the overall network
(URLLC). The stringent requirements of URLLC services, such latency to be less than 0.1msec and reliability of (1 − 10−9 )
as ultra-high reliability and low latency, have made it the
most challenging feature of the fifth generation (5G) of mobile [2].
arXiv:1802.09166v4 [cs.IT] 5 Sep 2018

networks. The problem is even more challenging for the ser-

vices beyond the 5G promise, such as tele-surgery and factory Physical layer design of URLLC is very challenging be-
automation, which require latencies less than 1ms and packet cause URLLC should satisfy two conflicting requirements:
error rates as low as 10−9 . This paper provides an overview ultra-low latency and ultra-high reliability. One could use short
on channel coding techniques for URLLC and compares them in packets to reduce latency which in turn causes a severe loss
terms of performance and complexity. Several important research in coding gain. Alternatively, the system bandwidth should be
directions are identified and discussed in more detail.
widened, which is not always possible especially for some
Index Terms—Mission critical communication, short block- URLLC applications in industrial control that might operate
length channel codes, URLLC. over unlicensed spectrum [2]. On the other hand, for enhancing
reliability, we need to use strong channel codes eventually
I. I NTRODUCTION paired with retransmission techniques which indeed increase
the latency.
T HE third generation partnership project (3GPP) has de-
fined three main service categories in 5G. Enhanced
mobile broadband (eMBB) is the service category designed
For high reliability transmissions of URLLC data, a channel
code with low code rates is generally used [3]. Several
for services which have high requirements for bandwidth
candidate channel codes such as low density parity check
such as virtual reality, augmented reality, and high-resolution
(LDPC), Polar, tail-biting convolutional code (TB-CC), and
video streaming. The second category is massive machine-
Turbo codes, were considered for both eMBB and URLLC
type communication (mMTC) which promises to support the
data channels. While it was recognized that LDPC , Turbo
massive number of machine-type devices and ultra-low power
and Polar codes have similar performance at large block sizes,
consumption to increase the device lifetime.
they could have big performance differences at small block
The third service category is ultra-reliable and low latency
sizes. LDPC has been already selected for eMBB data channel
communication (URLLC) which focuses on delay sensitive
and Polar codes were selected for the eMBB control channel.
applications and services (see Fig. 1). Factory automation
However, recent investigations showed that there exists some
and tele-surgery have the strictest reliability requirement of
error floors for LDPC codes constructed using base graph (BG)
(1 − 10−9 ) with an end-to-end latency of less than 1ms.
2, which has been considered for short block low rate scenarios
Other services such as smart grids, tactile internet, intelligent
[3]. Moreover, as shown in [3], Polar codes outperform LDPC
transportation systems, and process automation have more
codes without any sign of error floor. This means that it is not
relaxed reliability requirements of (1 − 10−3) ∼ (1 − 10−6) at
straight-forward to extend/modify eMBB channel coding for
latencies between 1ms to 100ms [1]. As shown in Fig. 1, 5G
URLLC with very diverse latency and data rate requirements.
may not be able to achieve the requirements for some industrial
It has been clearly specified that channel coding for URLLC
and medical applications with a very strict latency requirement
should be further studied, especially for information blocks of
of less than 1ms and block error rate (BLER) of 10−9 . These
less than 1000 bits [3].
systems might need to have their own standards with more
In this paper, we compare the main contenders channel
M. Shirvanimoghaddam, R. Abbas, C. Yue, Z. Lin, Y. Li, and B. Vucetic
are with the School of Electrical and Information Engineering, The University codes for URLLC, with the aim to achieve a favorable trade-
of Sydney, NSW, Australia. off between the latency and reliability. We mainly focus on
M. S. Mohammadi is with Silicon Laboratories. short blocks, i.e., in the order of a few hundreds bits for
A. Minja is with the Faculty of Engineering (aka Faculty of Technical
Sciences), University of Novi Sad, 21000 Serbia. URLLC. We review existing short channel codes and compare
B. Matuz is with the the Institute of Communications and Navigation of them in terms of rate efficiency at the reliability of interest for
the German Aerospace Center (DLR), Munchner Strasse 20, 82234 Wessling, some URLLC applications. We show that existing candidate
Germany.
G. Han is with the School of Information Engineering, Guangdong Univer- channel codes for URLLC still show a considerable gap to
sity of Technology, Guangzhou 510006, China. the normal approximation [4] benchmark, therefore there are
S. Johnson is with the School of Electrical Engineering and Computing, still rooms for further improvements. We highlight several
The University of Newcastle, NSW, Australia.
Corresponding Author: M. Shirvanimoghaddam (email: mah- important research directions to improve the performance of
[email protected]) URLLC channel codes.
 2

5G eMBB Target
10 −2
5G URLLC Target 5G Promise
10 −3
ITS
10 −4 traffic efficiency
road safety
Block Error rate

Tactile
Internet automated guided vehicles
10 −5 urban intersection
remote Remote Control
10−6
consulta- Tele-surgery Smart Grids
tion
haptic Process
10−7 Automation
feedback

10−8 remote
Factory Automation
surgery Packaging machines
10−9
Machine tools Printing machines Manufactuing cell

1msec 4msec 10msec 100msec

End-to-End Latency

Fig. 1. Latency and reliability requirements for different URLLC services.

II. K EY M ETRICS , R EQUIREMENTS , AND P ERFORMANCE 4) Performance Benchmark: There are two effects which
B ENCHMARK should be distinguished here to better understand the code
design problem for short blocks. The first one is the gap to
1) Latency: In the physical layer, we mainly focus on user the Shannon’s limit, that is if we decrease the block length, the
plane latency, which is defined as the time to successfully coding gain will be reduced and the gap to Shannon’s limit will
deliver a data block from the transmitter to the receiver via the increase. This is not a problem of code design but is mainly
radio interface in both uplink and downlink directions. User due to the reduction in channel observations that comes with
plane latency consists of four major components: the time-to- finite block lengths. We will use the normal approximation
transmit latency, the propagation delay, the processing latency, (NA) [4], that incorporates the reduction in channel observa-
e.g., for channel estimation and encoding/decoding, and finally tions, as the performance benchmark for comparison. For a
the re-transmission time. Propagation delay is typically defined coding block of length N , the normal approximation is given
as the delay of propagation through the transmission medium, by [4]:
and it depends on the distance between the transmitter and
receiver. The time-to-transmit latency is required to be in the r
order of a hundred microseconds, which is much less than the V −1 1
R=C− Q (ǫ) + log2 (N ),
1ms currently considered in 4G [1]. N 2N
2) Reliability: Reliability is defined as the success probabil-
ity of transmitting K information bits within the desired user where R is the code rate, C is the channel capacity, V is the
plane latency at a certain channel quality. Sources of failure channel dispersion, ǫ is the average block error rate (BLER),
from a higher layer perspective are when the packet is lost, and Q(.) is the cumulative distribution function of the standard
or it is received late, or it has residual errors. It is essential to normal distribution. Fig. 2 shows the normal approximation for
maximize the reliability of every packet in order to minimize different code rates and information block lengths. As can be
the error rate, so as to minimize the number of retransmissions. seen, when the block length increases, the gap to the Shannon’s
In this paper, we use block error rate (BLER) as a metric to limit [4] decreases1 .
compare different channel codes in terms of reliability. The second effect is the gap to the finite length bounds,
3) Flexibility: The flexibility of the channel coding scheme that is if we decrease the block length, modern codes, such as
is an important aspect along with the evaluation of the coding LDPC or Turbo codes, show a gap to finite length bounds. This
performance. Bit-level granularity of the codeword size and is often due to the suboptimal decoding algorithms. As can be
code operating rate is desired for URLLC [5]. The actual seen in Fig. 2, long term evolution (LTE) Turbo and TB-CC
coding rate used in transmission could not be restricted and codes show a considerable gap to the normal approximation
optimized for specified ranges [5]. The channel codes therefore at short blocks. However, when the block length of the Turbo
need to be flexible to enable hybrid automatic repeat request code increases, the gap to the normal approximation and the
(HARQ). The number of retransmissions however needs to be Shannon’s limit decreases.
kept as low as possible to minimize the latency.
The general URLLC requirement according to 3GPP is that 1 It is important to note that Shannon’s theoretical model breaks down for

the reliability of a transmission of one packet of 32 bytes short codes, as the channel capacity, defined as the maximum possible rate
at which reliable communications is possible, is only valid for infinite block
should be (1 − 10−5 ), within a user plane latency of 1ms length. The normal approximation was shown to be tight for moderate block
(with or without HARQ) [5]. lengths (>100 bits) [4].
 3

10−1
LTE TB-CC
LTE Turbo
K = 40
K = 1000
10−2 K = 8000
B LOCK E RROR R ATE

NA [4], K = 40
NA [4], K = 1000

10−3
1
R= 3
1
R= 12
Shannon’s limit

Shannon’s limit
10−4

10−5
−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5
SNR [dB]
Fig. 2. Comparison of error performance of LTE channel codes with different information block lengths, K.

III. C ANDIDATE S HORT B LOCK L ENGTH C HANNEL condition is met and the codeword with the smallest Euclidean
C ODES FOR URLLC distance from the ordered channel output is kept as the best
Here we briefly discuss several fixed-rate channel codes, decision. Recent advances in OSD design [7] significantly
which might be suitable for URLLC. Throughout the paper, the reduce the decoding complexity which makes OSD a good
information block length and codeword length are respectively choice for the decoding of short bock length codes.
denoted by K and N . For convolutional codes, we use n, k, BCH codes have large minimum distances which avoid
and m to denote the bit input and bit output per time instant flooring the performance at low BLER. However, BCH codes
and memory order, respectively. are not flexible as the block length and information length
cannot be selected arbitrary.

A. BCH Codes B. Convolutional Codes

Bose, Chaudhuri, and Hocquenghem (BCH) codes are a Convolutional codes (CC) were first introduced by Elias in
class of powerful cyclic error-correcting codes that are con- 1955 [6]. They differ from block codes as the encoder contains
structed using polynomials over finite fields [6]. The main memory. Generally, a rate R = k/n convolutional encoder
feature of BCH codes is that the number of guaranteed with memory order m can be realized as a linear sequential
correctable symbols, t, is defined during the code design circuit with input memory m, k inputs, and n outputs, where
process. The minimum distance dmin of BCH codes is at inputs remain in the encoder for m time units after entering.
least 2t + 1 [6]. The decoding of BCH codes is usually done Large minimum distances and low error probabilities for
using a bounded distance decoder, like the Berlekamp-Massey convolutional codes are achieved by not only increasing k
algorithm, that can correct any combination of up to t symbol and n, but also by increasing the memory order. The decoding
errors. In order to increase the coding gain, in particular on complexity however scales in general exponentially with the
noisy channels, one may use a soft-input decoder, such as memory order in both Viterbi and Bahl, Cocke, Jelinek, and
ordered statistics decoder (OSD). Raviv (BCJR) algorithms [6].
OSD is a near maximum likelihood (ML) soft decision When short packets have to be transmitted, terminated
decoding algorithm for an (N, K) binary linear block code convolutional codes represent a promising candidate solution,
with a given generator matrix. The decoding process consists although the rate loss due to a zero tail termination at short
of three steps. The first step is to reorder the channel output block lengths may be unacceptable. A tail-biting approach
in decreasing reliability order which yields a permutation. [8] eliminates the rate loss and hence it deserves particular
The same permutation is applied to the generator matrix. The attention when comparing channel codes for short blocks.
reordered generator matrix is transformed into a systematic For these reasons, tail-biting convolutional codes (TB-CCs)
form via Gaussian elimination. The hard decision of the K are currently considered within the 5G standardization for
most reliable values of the channel output are encoded into a URLLC. It is worth mentioning that the decoders for TB-
codeword via the permuted generator matrix. The reprocessing CCs are more complex than those for convolutional codes.
step consists of generating test error patterns of increasing TB-CC was used in LTE for the broadcast channel and
Hamming weight. This step is repeated until a predefined downlink/uplink control information.
 4

C. Turbo Codes identical binary-input discrete memoryless channels into N

In 1993, Berrou, Glavieux, and Thitimajshima, introduced synthesized channels with capacities either (close to) zero or
Turbo coding, which combines a parallel concatenation of one. The message is only sent over the set of near-perfect
two convolutional encoder and iterative maximum a-posteriori channels, and the unreliable channels are unused; In practice
probability (MAP) decoding [6]. Turbo codes have been ex- they are assigned constant inputs a priori known for both the
tensively used for the data channel in LTE. For large blocks, encoder and decoder (frozen symbols).
Turbo codes are capable of performing within a few tenths Under successive cancellation (SC) decoding which requires
of dB from the Shannon’s limit. Unfortunately, Turbo codes a complexity of O(N log N ), and for sufficiently large code-
with iterative decoding in short and moderate block lengths word lengths, the block error probability decays exponentially
show a gap of more than 1 dB to the finite-length performance in the square root of the code length. The recursive nature of
benchmark. LTE Turbo code is known to be well designed for the SC decoding may impose a large latency depending on the
medium block length and code rate ≥ 1/3. When the code rate implementation.
and block length are small, LTE-Turbo code performance is A major improvement in the decoding performance is
degraded. For Turbo codes, 1-bit granularity is feasible for all achieved by using successive cancellation list (SCL) decoding
coding rates and for full range of block size, and the ability of which keeps a list of most likely decoding paths at all times,
Turbo codes to support both chase combining and incremental unlike the SC decoder which keeps only one decoding path,
redundancy HARQ is well known [9]. i.e., it performs a symbol-wise hard decision at each decoding
stage. A significant improvement to SCL is cyclic redundancy
check (CRC)-aided SCL (CA-SCL), where the message is
D. Low Density Parity Check (LDPC) Codes encoded with a CRC error detection code, and the result is
Low-density parity-check (LDPC) codes were originally polar coded [11]. In this way, the CRC checksum is used at
proposed by Gallager in the early 1960s and later rediscovered the decoder side to pick the right decoding path in the list,
in the 1990s, when researchers began to investigate codes-on- even if it is not the most probable path.
graph based on Tanner’s work in 1981 and iterative decoding Polar codes have been selected for short blocks for control
[6]. LDPC codes with iterative belief propagation (BP) de- channels in eMBB [9]. Recent investigations and proposals
coding have been shown to perform very close to Shannon’s submitted to 3GPP also demonstrated that Polar codes outper-
limit with only a fraction of a decibel gap. Binary LDPC form LDPC codes in short block lengths and low code rates
codes with iterative BP decoding however do not perform without any sign of error floor; therefore suitable for URLLC
well at short to moderate block-lengths which is mainly due use cases [9]. In Polar codes, 1-bit granularity can be achieved
to the existence of many short cycles in the code’s bipartite for all coding rates and for full range of block size. However,
graph. Recently, protograph-based LDPC codes have been the implementation complexity of the list decoder increases
shown to perform well under belief propagation decoding at with increasing list size, especially with larger block sizes [9].
short-to-moderate block length, but their performance is not
comparable with BCH codes under OSD or TB-CCs with IV. C OMPARISON BETWEEN C HANNEL C ODES FOR
large memory. However, they favor the very low decoding URLLC
complexity under iterative decoding algorithms. Non-binary In this part, different channel codes for URLLC are com-
LDPC codes are also shown to perform very close to the finite pared in terms of reliability, rate performance, and algorith-
length performance bound where the decoding complexity is mic complexity. We consider a binary input additive white
the major drawback. Gaussian noise (BI-AWGN) channel, where unit power binary
There are some other advantages for LDPC codes, e.g. in antipodal signals are sent over a channel which are subject to
parallelization of the decoding algorithm. LDPC codes have the additive white Gaussian noise of variance σ 2 . The signal
been adopted for eMBB data channel and it is a natural to noise ratio (SNR) is then defined as σ12 . For each SNR point
extension to apply LDPC codes for URLLC. Two base graphs and code rate, the simulation is run to obtain 100 codeword
are considered for LDPC in eMBB. BG 1 is used for high data errors at the decoder output.
rate and long block lengths. BG 2 is used for low code rates
and short block lengths. Recent investigations demonstrated
A. Reliability
that there exist some error floors for LDPC codes at certain
rates and block lengths [9]. Moreover, the complexity of LDPC Fig. 3 shows the BLER versus the SNR for different
increases with increasing flexibility [9]. candidate channel codes at rate R = 12 and block length
N = 128 under the maximum likelihood decoding (MLD)
[12]. By using an optimal decoder in the (ML sense), the plot
E. Polar Codes gives insights on the code performance itself. As can be seen
Polar codes as introduced in [10], are binary linear codes in this figure, the extended BCH code closely approaches the
that can provably achieve the capacity of a binary-input normal approximation benchmark over the whole SNR region
discrete memoryless channel using low-complexity encoding and can provide a very low BLER as small as 10−7 with only
and decoding as the code length tends to infinity. Channel 0.1dB gap to the normal approximation. Another competitive
polarization is a central technique in the construction of these code is the TB-CC code with m = 14, which can provide a
codes, in which the block code translates N independent and BLER of 10−5 with only 0.1dB gap to the NA benchmark
 5

100

10−1

10−2
B LOCK E RROR R ATE

10−3

10−4
Polar Code [12] Polar Code + CRC-7 [12]
10−5 Reed-Muller Code [12] F16 LDPC Code [12]
F256 LDPC Code [12] Binary LDPC [12]
TB-CC, m = 8 [8] TB-CC, m = 11 [8]
10−6 TB-CC, m = 14 [8] eBCH Code [12]
Normal Approximation [4]
10−7
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5
SNR [dB]
Fig. 3. Comparison of error performance of different rate R = 1/2 channel codes with codeword length of N = 128 under MLD [12]. For TB-CC, we
used the circular Viterbi algorithm (CVA).

0.8
R ATE [b/s/Hz]

0.6

0.4

eBCH, OSD (order 5) CA-Polar code, SCL, L= 32

CA-Polar code, SCL, L= 4 eMBB LDPC, BP Decoder, 50 iterations
0.2 Normal Approximation [4] BI-AWGN Capacity

−2 −1 0 1 2 3 4 5 6 7 8 9 10
SNR [dB]
Fig. 4. Comparison of different channel codes with codeword length of N = 128 with different rates at BLER= 10−4 .

[4], however, when it goes to a lower BLER of 10−7 the between complexity and performance in more details in later
gap increases to 0.3dB. Decreasing the memory to 11, TB-CC sections.
still gives a performance within 0.1dB gap from the normal
approximation at a BLER of 10−5 . Other competitive codes
are LDPC codes designed over a large Galois field (here F256 ), B. Rate Performance
which have almost the same performance as the TB-CC code Fig. 4 shows the rate performance of different candidate
with m = 14. The circular Viterbi algorithm (CVA) algorithm codes at a BLER of 10−4 when the codeword length is N =
has been used for decoding of TB-CCs [8]. 128. As can be seen, BCH codes perform very close to the
The BCH code outperforms all other existing codes owing to normal approximation and outperform other existing codes at
its better distance spectrum. Other codes are mainly designed all SNRs. The generator polynomials of the used BCH codes
to provide good performance while maintaining the decoding were taken from [6] and OSD was used with a maximum
complexity at a reasonable order. We will discuss the tradeoff re-factoring order of 5. As in [7], we used the probabilistic
 6

100

10−1
B LOCK E RROR R ATE

10−2

10−3

10−4

10−5
−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5
SNR [dB]
CA-Polar, L=32 LTE-TB-CC, List-1 Viterbi LDPC, min-sum, 25 iterations LTE-Turbo, Max-log-MAP
K = 40 K = 1000 NA [4]

Fig. 5. Comparison of different channel codes with different rates and information block lengths. Rates R = 1/3, 1/6, and 1/12, are respectively shown
with blue, red and black colors. For the deatil of the channel codes refer to [13].

necessary condition and the probabilistic sufficient condition in terms of the block length. However, for short block lengths
to reduce the complexity of the OSD decoder. other code parameters have significant impact on the decoding
Two sets of Polar codes with SA-SCL with list sizes 4 and complexity. Here we only focus on the algorithmic complexity
32 are also shown in Fig. 4. Although, the decoder with the list which can be represented in terms of the number of binary
size of 32 significantly outperforms the other decoder with list operations. For example, the decoding complexity of a TB-CC
size 4, this comes with a significant increase in the decoding code using the Viterbi decoder is mainly dominated by the
complexity, as the decoder needs to store and list additional memory order, in short block lengths, as the memory order
candidates. should be usually large to guarantee the performance. Fig.
For the sake of comparison, we also show in Fig. 4 the 6 shows the complexity versus the performance of different
performance of the short block length LDPC codes specifically channel codes. As can be seen, Polar codes with the SCL
designed for eMBB. These codes have the advantage of very decoder achieve the error rate of 10−4 at only 0.5dB gap to
low complexity iterative BP decoding and have slightly better the normal approximation benchmark with the complexity in
performance than the CA-Polar code with SCL decoding of the order of 103 operations per bit. The complexity can be
list size 4. reduced by reducing the list size, which however degrades
It is important to note that although BCH codes perform the performance. TB-CC codes have huge complexity which
very close to the normal approximation benchmark, their significantly increases with the memory order. The original
decoding is very complex. For lower code rates, usually an OSD decoder has the complexity in the order of K ℓ , for ℓ
OSD with much higher reprocessing order should be used being the order number.
to guarantee the performance. This however significantly in-
creases the complexity. On the other hand, as can be seen in V. R ECOMMENDATIONS AND R ESEARCH D IRECTIONS
Fig. 5, Polar and LDPC codes can still offer good performance
at low rates with considerably lower decoding complexities. To achieve ultra-high reliability and low latency, we identify
three main directions where major improvements are essential.
We also provide some recommendations in each direction.
C. Complexity vs. Performance
Turbo and LDPC codes have shown to provide near capacity
A. Developing Low Complexity OSD Decoders
performance at large block-lengths with reasonable complexity
due to the iterative nature of the decoders and the fact that most As we show in Fig. 4, under the OSD, BCH codes out-
of the calculations can be done in parallel. The complexity of perform other existing channel codes, including Polar codes,
such decoders, for example belief propagation, scale linearly Turbo codes, and LDPC codes. The complexity of OSD can
with the block-length. In fact, in most of the complexity be significantly reduced using several approaches such as suf-
analysis of such codes, the complexity is usually characterized ficient conditioning and segmentation [7]. Recent work [7] has
 7

1010
N UMBER OF B INARY O PERATIONS P ER I NF. B IT

109 TB-CC m = 14

108
F256 LDPC Code (log-BP)
7 TB-CC m = 11
10

106 F256 LDPC Code (FFT-BP)

TB-CC m = 8
105

104 Turbo Code, m=4 (Max-log-MAP)

eBCH (OSD order 5 with bounded complexity)
eMBB LDPC BP (50 Iterations)
103 Polar Code, SCL, L = 32

102 Polar Code, SCL, L = 4

Polar Code, SC
101
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
G AP TO THE N ORMAL A PPROXIMATION [dB]
Fig. 6. Algorithmic complexity versus performance for different rate-1/2 channel codes with block length N = 128 at BLER= 10−4 . The algorithmic
complexity for different decoders are obtained from [14].

shown significant complexity reduction for OSD while main- codes, the code rate is determined on the fly and automatically
taining rate efficiency similar to the original OSD decoder. adapts to channel conditions, without having any CQI at the
Moreover, as shown in Fig. 6 using an OSD with bounded transmitter side. PBRL-LDPC codes offer a fine granularity
complexity, a tradeoff between performance and complexity over information block size and rate. They have been demon-
can be achieved. We identify that to further improve the strated to achieve more than 90% of the NA benchmark and
performance of channel codes for URLLC more sophisticated BLERs as low as 10−6 . However, recent investigations show
ML-like decoders should be designed to allow low complexity that the performance of PBRL-LDPC codes constructed from
decoding of fundamentally better codes in short block lengths, a base matrix degrades gradually when more coded symbols
like BCH codes. are transmitted. For self-adaptive codes, we can identify two
main research directions.
B. Self-Adaptive Joint Coding and Modulation Schemes 1) The design of a self-adaptive joint coding and mod-
Current cellular networks adopt a rate-adaptive scheme, ulation scheme: PBRL LDPC codes are binary codes and
that is prior to every transmission, the transmitter sends pilot the joint design of the code and modulation can significantly
signals to the receiver which enables the latter to estimate improve the performance. As the rateless property of PBRL-
the channel state. Then, the receiver feeds back a channel LDPC is inspired by Raptor Codes, we propose to replace the
quality indicator (CQI). Based on this CQI, the transmitter Raptor part of the code with rateless codes over real domain,
selects the best combination of fixed-rate code and modulation like Analog Fountain Codes (AFC) [15]. The joint design
scheme from a predetermined set. In case the receiver cannot of PBRL-LDPC and AFC can offer significant performance
recover the information (which can be verified with a CRC), it improvements.
will request a retransmission based on the well-known HARQ 2) Reducing pilot sequence length for channel estimation at
protocol. the receiver side: For any rateless code, the receiver still needs
This rate adaptive scheme suffers from two main drawbacks. to know the channel to decode the information. This implies
The first drawback is due to the choice of channel codes that that the transmitter needs to insert pilot symbols into each
targets BLERs of 10−2 making the scheme heavily dependent transmitted block which can incur significant performance
on retransmissions. In the current LTE standard, each retrans- losses. These performance losses become more noticeable
mission takes about 7∼8ms, introducing significant delays when the information block size is small. Assuming a powerful
unacceptable for URLLC. To tackle this, one needs to consider (128,64,22) extended BCH (eBCH) code over a BI-AWGN
new channel codes more suitable for short blocks and with channel with SNR 3.5dB and using only 7 symbols for pilot
lower BLERs. Unfortunately, even so, the rate adaptive scheme transmission, our initial results see 5% loss in the spectral
still suffers from the channel estimation overhead which takes efficiency at BLER 10−5 . The loss increases to 9.8% and
about 5∼8ms in the current LTE standard. This latency can 18% when a QPSK and 16QAM modulations are considered,
be very costly for mission critical URLLC applications. respectively. A solution would be to design near ML decoding
Self-adaptive [15] channel coding is a promising approach using OSD without requiring the accurate CSI at the receiver
to ensure ultra-low latency transmissions. With self-adaptive using iterative approaches.
 8

C. Space-Frequency Channel Coding [9] “R1-1611081 - Final Report of 3GPP TSG RAN WG1 #86bis v1.0.0,”
3GPP, Final Minutes Report, Nov. 2016.
In current cellular systems, the spatial domain was mainly [10] E. Arikan, “Channel polarization: A method for constructing capacity-
used to provide multiplexing gain rather than diversity as the achieving codes for symmetric binary-input memoryless channels,” IEEE
main objective was to improve the throughput for moderate Transactions on Information Theory, vol. 55, no. 7, pp. 3051–3073, July
2009.
reliability order, i.e., BLER of 10−2 as in LTE. In order to [11] K. Niu and K. Chen, “CRC-aided decoding of Polar codes,” IEEE
increase reliability, one can use the available transmit and Communications Letters, vol. 16, no. 10, pp. 1668–1671, 2012.
receive antennas to provide spatial diversity rather than spatial [12] M. Helmling, S. Scholl, F. Gensheimer, T. Dietz, K. Kraft, S. Ruzika,
and N. Wehn, “Database of Channel Codes and ML Simulation Results,”
multiplexing. In 4G and 5G, orthogonal frequency division www.uni-kl.de/channel-codes, 2017.
multiple access (OFDMA) is the major multiple access tech- [13] “R1-1611108- Evaluation on Channel Coding Candidates for uRLLC
nology, in which each resource block consists of a number and mMTC,” 3rd Generation Partnership Project, 3GPP TSG RAN WG1
Meeting #87- Discussion, Nov. 2016.
of OFDM symbols. Reducing the number of OFDM symbols [14] M. Sybis, K. Wesolowski, K. Jayasinghe, V. Venkatasubramanian, and
per resource block has been identified as an effective approach V. Vukadinovic, “Channel coding for ultra-reliable low-latency commu-
to reduce the latency. The design of the channel code in nication in 5G systems,” in Vehicular Technology Conference (VTC-
Fall), 2016 IEEE 84th. IEEE, 2016, pp. 1–5.
these systems is then challenging when considering different [15] M. Shirvanimoghaddam, Y. Li, and B. Vucetic, “Near-capacity adaptive
diversity sources, including space or frequency. So far, there analog fountain codes for wireless channels,” IEEE Communications
is no universal framework to design space-frequency channel Letters, vol. 17, no. 12, pp. 2241–2244, Dec. 2013.
codes to provide different level of reliability for low latency

VI. C ONCLUSIONS
This paper reviewed the most recent progresses in the design
and implementation of short block length channel codes for
ultra-reliable low latency communications (URLLC). Several
candidate channel codes, including Polar codes, Turbo codes,
LDPC codes, Convolutional codes, and BCH codes, were
considered and compared in terms of block error rate under
optimal decoder, rate performance under practical decoders,
and algorithmic complexity of decoding algorithms. BCH
codes provide the highest reliability under optimal decoding,
since they have the highest minimum Hamming distance.
Polar codes with successive cancellation list decoding provide
a reliability of (1 − 10−4 ) with only 0.5dB gap to the
normal approximation with reasonable complexity, however
better results might be achieved with reduced complexity OSD
and BCH codes with the same level of complexity. We also
identified several major research directions for channel coding
for URLLC.

R EFERENCES
[1] “3GPP TS 22.261 Service requirements for the 5G system; Stage 1 (Re-
lease 16),” 3rd Generation Partnership Project, Technical Specification
Group Services and System Aspects, Mar. 2018.
[2] V. K. L. Huang, Z. Pang, C. J. A. Chen, and K. F. Tsang, “New trends
in the practical deployment of industrial wireless: From noncritical to
critical use cases,” IEEE Industrial Electronics Magazine, vol. 12, no. 2,
pp. 50–58, June 2018.
[3] “R1-1804849- Remaining Issues on URLLC Data Channel Coding,” 3rd
Generation Partnership Project, 3GPP TSG RAN WG1 Meeting #92bis-
Discussion and Decision, Apr. 2018.
[4] T. Erseghe, “Coding in the finite-blocklength regime: Bounds based on
laplace integrals and their asymptotic approximations,” IEEE Trans. Inf.
Theory, vol. 62, no. 12, pp. 6854–6883, 2016.
[5] “R1-1608770- Flexibility evaluation of channel coding schemes for NR,”
3rd Generation Partnership Project, 3GPP TSG RAN WG1 Meeting #86-
Discussion and Decision, Oct. 2016.
[6] S. Lin and D. Costello, Error Control Coding: Fundamentals and
Applications. Pearson-Prentice Hall, 2004.
[7] J. V. Wonterghem, A. Alloumf, J. J. Boutros, and M. Moeneclaey, “On
performance and complexity of OSD for short error correcting codes in
5G-NR,” in BalkanCom 2017, June 2017, pp. 1–5.
[8] L. Gaudio, T. Ninacs, T. Jerkovits, and G. Liva, “On the performance of
short tail-biting convolutional codes for ultra-reliable communications,”
in SCC 2017; 11th International ITG Conference on Systems, Commu-
nications and Coding; Proceedings of. VDE, 2017, pp. 1–6.