5G LDPC Coded Modulation With Hybrid Shaping

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

C HINA C OMMUNICATIONS

5G LDPC Coded Modulation with Hybrid Shaping

Jiayi Yang1 , Qianfan Wang1 , Congduan Li2,* , Xiao Ma*1

1
School of Computer Science and Engineering, and Guangdong Key Laboratory of Information Security Technology, Sun Yat-sen
University, Guangzhou 510006, China
2
School of Electronics and Communication Engineering, Sun Yat-sen University, Shenzhen 518107, China
*
The corresponding author, email: licongd@mail.sysu.edu.cn, maxiao@mail.sysu.edu.cn

Abstract: In this paper, we integrate the 5G low- demapping); 3) the HS-MLC has a significantly
density parity-check (LDPC) coded modulation better performance compared to the time-division
systems with hybrid shaping, where the centroid- HS-BICM-ID over fast fading channels.
based geometric shaping is implemented to remedy Keywords: Bit-interleaved coded modula-
the performance loss of the many-to-one probabilistic tion (BICM); coded modulation; geometric shaping;
shaping. Taking into account the fact that the 5G multilevel coding (MLC); probabilistic shaping
parity-check matrices have an uneven density in differ-
ent parts, we elaborately design a simple row-column I. INTRODUCTION
interleaver for the bit-interleaved coded modulation
with iterative decoding (BICM-ID) system to allocate With the increasing need of faster and more reli-
the ambiguous bits caused by the many-to-one map- able data transmissions, efficient coded modulation
ping to the sparser parity part, resulting in the hybrid schemes have drawn significant interest. The concept
shaping for BICM-ID (HS-BICM-ID) system. For of combining coding and modulation was first intro-
the multi-layer coding (MLC) system, we propose a duced by Massey in 1974 [1]. Since then, numerous
cyclic shift mapping scheme to allocate evenly the coded modulation schemes have been proposed, in-
ambiguous bits to each layer, resulting in the hybrid cluding multilevel coding (MLC) [2], bit-interleaved
shaping for MLC (HS-MLC) system. By treating the coded modulation (BICM) [3] and trellis coded modu-
multiple binary codes of the HS-MLC system as a lation (TCM) [4]. It has been demonstrated in [5] that
non-binary LDPC code, we present a joint decoding MLC with multistage decoding (MSD) approaches
algorithm, which does not require iterative processing the constellation constrained channel capacity if the
between demapping and decoding. Numerical results component code rates are properly chosen. How-
show that: 1) the HS-BICM-ID can obtain shaping ever, the MLC system requires the joint design of
gains of about 0.2 dB and 1.0 dB compared to the multiple codes. Different from the MLC system, the
constant composition distribution matching (CCDM) BICM system adopts a single code [6] and can be
shaping and the geometric shaping, respectively, while decoded iteratively, known as the BICM with itera-
it can obtain a shaping gain of about 1.2 dB compared tive decoding (BICM-ID) [7] system. In fading chan-
to the scheme with uniform input; 2) for the HS-MLC, nels, BICM is favored over TCM as its simplicity
the performance of the nonbinary joint decoding and flexibility [8]. As an emerging technology, the
algorithm (without iterative demapping) is slightly delayed BICM (DBICM) [9] [10] is also capable of
better than the iterative MSD algorithm (with iterative achieving constellation constrained capacity by uti-
lizing the interdependency among signals at different
Received: Nov. 21, 2023 time slots [11], while only requiring a single channel
Revised: Mar. 05, 2024 code.
Editor: Tao Jiang
The channel inputs of the aforementioned coded

China Communications 1
modulation schemes are usually uniformly distributed MLC coded modulation systems. The main contribu-
and regularly equidistant, leading to an asymptotic tions of this paper are listed as follows.
loss of about 1.53 dB, referred to as the shaping • We present a hybrid shaping scheme for pulse am-
gain [12]. Constellation shaping is a commonly used plitude modulation (PAM) constellations, where
technique to optimize the distribution of the channel the centroid-based geometric shaping is imple-
input, which aims to achieve the shaping gain. This mented to remedy the performance loss of the
technique can be broadly classified into two types: dyadic many-to-one probabilistic shaping. Distin-
probabilistic shaping [13–16] and geometric shap- guished from the shaping scheme in [23], the pre-
ing [17–20]. sented hybrid shaping can avoid the complicated
The basic idea of probabilistic shaping is to em- implementation of the distribution matcher, and
ploy a non-uniform distribution on equidistant signal the main complexity of our scheme lies in the con-
points. This can be achieved by many-to-one map- stant complexity introduced during the demap-
ping [21], named as Gallager mapping in [13], which ping. Numerical results show that, in terms of
can be implemented by superposition or by table-look- capacity gain, the presented hybrid shaping out-
up, resulting in the dyadic probability distribution. performs the conventional probabilistic and geo-
When dyadic shaping is combined with the binary for- metric shaping schemes.
ward error correction (FEC), the receiver typically em- • Taking into account the fact that the 5G parity-
ploys an iterative demapping scheme to reverse the im- check matrices have an uneven density in differ-
pact of the many-to-one mapping, say, [22]. The prob- ent parts, we elaborately design a simple row-
abilistic amplitude shaping (PAS) with constant com- column interleaver for the BICM-ID system to al-
position distribution matching (CCDM) was proposed locate the ambiguous bits caused by the many-to-
in [23], which can efficiently approximate the optimal one mapping to the sparser parity part, resulting in
distribution. However, the complexity of the shaping the hybrid shaping for BICM-ID (HS-BICM-ID)
code for CCDM is a challenge in cost-sensitive trans- system. Simulation results demonstrate that the
missions. HS-BICM-ID can achieve shaping gains of ap-
The basic idea of geometric shaping is to employ a proximately 0.2 dB and 1.0 dB when compared
uniform distribution on non-equidistant constellation to the CCDM shaping and the geometric shaping,
points. In [17], it is shown that Gaussian-like signal respectively. It can obtain a shaping gain of about
sets achieve the capacity of additive white gaussian 1.2 dB when compared to the scheme with uni-
noise (AWGN) channel if the order of constellation form input.
sets goes to infinity. Recently, an optimized geomet-
ric shaping scheme was implemented to the quadrature • The hybrid shaping scheme is applied to the
amplitude modulation (QAM) constellation in [24]. MLC system, resulting in the hybrid shaping for
Compared to probabilistic shaping, geometric shaping MLC (HS-MLC) system, where a cyclic shift
usually has lower complexity for coded modulation mapping is proposed to allocate the ambiguous
systems, but the shaping gain is limited by the order bits evenly to each layer. By treating the mul-
of constellation sets. tiple binary codes of the MLC system as a non-
Typically, the existing work separately adopts either binary LDPC code, we present a joint decoding
the probabilistic or geometric shaping design, while algorithm, which does not require iterative pro-
we consider a joint optimization design of the con- cessing between demapping and decoding. Sim-
stellation set and the probability distribution. Taking ulation results show that the performance of the
into account the pros and cons of probabilistic shap- nonbinary joint decoding algorithm (without iter-
ing and geometric shaping, we present in this paper ative demapping) is slightly better than the iter-
a hybrid shaping scheme by using centroid-based geo- ative MSD decoding (with unavoidable iterative
metric shaping to compensate for the performance loss demapping).
of dyadic probabilistic shaping. Furthermore, the hy- • We compare the performance of time-division
brid shaping is tailored to the 5G low-density parity- HS-BICM-ID and HS-MLC over Rayleigh fading
check (LDPC) code and is applied to BICM-ID and channels. Due to the adoption of multi-layer joint

2 China Communications
modulation, the HS-MLC scheme is more suit- where
able for fading channels compared to the time-
division HS-BICM-ID. Simulation results show 1 (y − x)2
pY |X (y|x) = √ exp(− ) (3)
that the HS-MLC system has a significantly better 2πσ 2σ 2
performance compared to the time-division HS-
BICM-ID over fast fading channels. is the channel transition probability density func-
The rest of this paper is organized as follows. In tion (PDF) and
Section II, we review the probabilistic shaping with
X
many-to-one mapping and the geometric shaping. The pY (y) = PX (x)pY |X (y|x) (4)
hybrid shaping scheme is also presented in this sec- x∈A
tion. In Section III, we present the HS-BICM-ID and
HS-MLC systems. Simulation results are presented in with the input distribution PX (x).
Section IV. Section V concludes this paper.

2.2 Probabilistic Shaping


II. HYBRID SHAPING WITH MANY-TO-
ONE MAPPING To achieve the shaping gain with probabilistic shaping,
we need to assign an optimized probability distribution
2.1 AWGN Channel Model over a given signal constellation. It has been proved
Consider the transmission over the AWGN channel that the Maxwell-Boltzmann (M-B) distribution [26]
Yt = Xt + Zt , where Xt ∈ R and Yt ∈ R are the is nearly optimal, which can be expressed by
channel input and output at time t, respectively. The 2
sequence Z = (Z1 , · · · , Zt , · · · ) is assumed to be e−βx
PX (x) = P −βx′2
, x ∈ A, (5)
an independent and identically distributed (i.i.d) ad- x′ ∈A e
ditive noise sequence, where Zt is a Gaussian ran-
dom variable with mean 0 and variance σ 2 , denoted where β is a non-negative scaling factor. For β = 0,
by Zt ∼ N (0, σ 2 ). The channel input is subject to the distribution is uniform, while for β > 0, the distri-
the power constraint, i,e., E(Xt2 ) ≤ Es , where E(.) bution is Gaussian-like. By selecting β properly, the
denotes the expectation. We know that the channel minimum average power can be achieved for a given
capacity equals the maximum mutual information and mutual information.
Xt ∼ N (0, Es ) is the optimal distribution. The capac- In this paper, the dyadic distribution is used to quan-
ity in bits per channel use (bits/channel use) is a func- tize the M-B distribution. By using the many-to-one
tion of the signal-to-noise ratio (SNR) and is given by mapping, multiple labels may be mapped to the same
Shannon in [25] signal point, resulting in a non-uniform distribution of
channel input. Let m denote the number of bits per
1 constellation point. A signal mapping is defined by
C= log2 (1 + SNR), (1)
2 ϕ(v) = x, where v = (v (0) , · · · , v (m−1) ) ∈ Fm 2 and
x ∈ A. The dyadic distribution for constellation set A
where SNR = Es /σ 2 .
is
In practice, the distribution of channel input Xt X δ(ϕ(v) = x)
is usually discrete, which is constrained by the con- PX′ (x) = , x ∈ A, (6)
2m
stellation set A of finite cardinality. In this pa- v∈F2 m

per, we only consider one-dimensional signaling for


the AWGN channel, where we assume that A = where the indication function δ(P) = 1 if the propo-
{aj ∈ R, 0 ≤ j ≤ M − 1}. Then the mutual informa- sition P is true and δ(P) = 0 otherwise. From (6),
tion can be written as it is not difficult to see that the probabilities of the
channel input are integer multiples of 2−m . To obtain
pY |X (Y |X) the sub-optimal quantized distribution PX∗ of given m,
I(X; Y ) = E(log2 ), (2) the criterion of minimum Kullback-Leibler (KL) dis-
pY (Y )

China Communications 3
tance [27] is used, which can be expressed by 2.4 Hybrid Shaping
The centroid-based geometric shaping is easy to op-
PX∗ = arg min D(PX′ ||PX ). (7) erate without increasing the complexity, but the gain

PX
is limited when M is small. The dyadic probabilis-
tic shaping with many-to-one mapping is also easy to
Example 1.: Consider a transmission of 16-PAM implement without a complicated design of the shap-
constellation with 3 bits/channel use. The constel- ing code, but there is a performance loss between the
lation set A consists of the one-dimensional signals achievable capacity with the nearly optimal distribu-
{−15, −13, · · · , −1, 1, · · · , 13, 15}. From (5), the tion.
nearly optimal distribution PX of 3 bits/channel use To remedy the loss from the capacity of nearly opti-
can be obtained. Then, we use the criterion of mini- mal distribution, a hybrid shaping scheme is presented.
mum KL distance to obtain the sub-optimal distribu- Different from the typical shaping schemes, the pre-
tion for given m. For m = 5, the sub-optimal distribu- sented hybrid shaping can be viewed as a joint opti-
tion PX∗ of the constellation set can be obtained, where mization of the constellation probability and the con-
the probabilities
 −3of −3
the positive constellation points stellation set. Due to the limited approximation ef-
are PX = 2 , 2 , 2−4 , 2−4 , 2−5 , 2−5 , 2−5 , 2−5 ,
half
fectiveness of dyadic many-to-one probabilistic shap-
while the negative constellation points have symmet- ing for the target distributiomn, we integrate it with a
ric probabilities. The KL distance between the sub- universal centroid-based geometric shaping algorithm.
optimal distribution and nearly optimal distribution This allows us to adjust the positions of constellation
is 0.0616 for m = 5. For m = 6, the sub- points based on the dyadic distribution, further ap-
optimal distribution
 −3 −3of positive constellation points proximating a Gaussian distribution input and achiev-
half −4 −4 −4 −5 −6 −6
are PX = 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , ing better performance.
where the KL distance from the optimal distribution is • Optimization for the input distribution by us-
0.0282. ing the dyadic probabilistic shaping with many-
to-one mapping: The process can be viewed
as an optimization of probability for equidistant
2.3 Geometric Shaping constellations. Firstly, for given spectral effi-
ciency, the nearly optimal distribution PX can
It has been shown that the channel capacity can also be be obtained under the equidistant constellations
approached by equiprobable signaling with Gaussian- from (5). Secondly, based on the idea of dyadic
like signal sets [17]. Specifically, we divide the stan- many-to-one mapping, the criterion of minimum
dard Gaussian PDF, N (0, 1), into M disjoint regions KL distance in (7) can be used to approach the
such that nearly optimal distribution for given m, thus ob-
taining the sub-optimal distribution PX∗ .
ξk+1
−x2
Z
1 1 • Optimization for the constellation set by using the
√ exp( )dx = , k = 0, 1, · · · , M −1,
ξk 2π 2 M geometric shaping: The process can be viewed as
(8) an optimization of the constellation set for a given
where the k-th interval is given by (ξk , ξk+1 ) with ξ0 = sub-optimal distribution of equidistant constella-
−∞ and ξM = ∞. The centroid of each interval can tions. We consider using a centroid-based geo-
be calculated by metric shaping scheme [17] to optimize the signal
set A under the sub-optimal distribution. Firstly,
Z ξk+1
M −x2 the standard Gaussian PDF is divided into M in-
xk = √ xe 2 dx, k = 0, 1, · · · , M − 1.
2π ξk tervals by choosing the boundaries ξk such that
(9)
ξk+1
Then, we select a scaling factor α > 0 such that the −x2
Z
1
√ exp( )dx = PX∗ (xk ), k = 0, 1, · · · , M −1,
constellation set A = {aj = αxj , 0 ≤ j ≤ M − 1} ξk 2π 2
has an average power E(X 2 ) = Es . (10)

4 China Communications
where PX∗ (xk ) denotes the sub-optimal distribu- 5
3
tion probability of the k-th signal point obtained 4.5
2.98

from (6) and (7) for a given m. Then the signal set

Mutual Information (bits/channel use)


4 2.96

A′ of hybrid shaping can be obtained by scaling


2.94
3.5 2.92

the centroids, 3
2.9

2.88

2.5 2.86
17.8 18 18.2 18.4
Z ξk+1
1 −x2
x′k = √
2
xe 2 dx, k = 0, 1, · · · , M −1,
2πPX∗ (xk ) ξk 1.5
(11) 1
such that E(X 2 ) = Es .
0.5

Given the dyadic distribution and the optimized con- 0


0 5 10 15 20 25 30
stellation set, we may use a binary Huffman tree to SNR(dB)
define a bit-to-symbol Gray mapping [28]. This is il-
Figure 1. Mutual information versus SNR curves for differ-
lustrated by the following example.
ent shaping schemes.
Example 2.: Given m = 6, Es = 1. Firstly,
the probability of constellation is optimized by using
2.5 Numerical Analysis
dyadic probabilistic shaping with many-to-one map-
ping, resulting in the sub-optimal distribution PX∗ . The mutual information I(X; Y ) defined in (2) can be
Then, the constellation set is optimized by using the evaluated by the Monte Carlo method or other numer-
geometric shaping, resulting in the signal set A′ . Fur- ical integration methods. In Fig. 1, we plot the mutual
ther, we can get the signal mapping ϕ from a binary information of the presented hybrid shaping, the geo-
Huffman tree, as shown in Table 1, where the sym- metric shaping, the probabilistic shaping with many-
bol [×] is used for the so-called ambiguous bit mean- to-one mapping and the uniform for 16-PAM. The nu-
ing that its value can be either 0 or 1. For exam- merical results are tabulated in Table 2 at SNR = 18
ple, the signal point a1 = −2.01 has one label with dB. It can be seen that hybrid shaping with m = 6 is
ϕ(000001) = −2.01, and a2 = −1.69 has two labels nearly optimal in the low and moderate SNR regions,
with ϕ(000010) = ϕ(000011) = −1.69. outperforming both probabilistic shaping with many-
to-one mapping and geometric shaping, and achieving
a better capacity gain. We also present the signal posi-
Table 1. Mapping Scheme for 16-PAM Hybrid Shaping m = 6
tions and the probability distributions of the schemes
x = ϕ(v) v (0) v (1) v (2) v (3) v (4) v (5) mentioned above for 16-PAM in Fig. 2. Obviously,
-2.52 0 0 0 0 0 0 the proposed hybrid shaping scheme provides a better
-2.01 0 0 0 0 0 1 approximation to the Gaussian-like channel inputs.
-1.69 0 0 0 0 1 ×
-1.33 0 0 0 1 × ×
-1.02 0 0 1 1 × × III. LDPC CODED MODULATION
-0.78 0 0 1 0 × ×
-0.49 0 1 1 × × × 3.1 Review of 5G LDPC Codes
-0.16 0 1 0 × × ×
+0.16 1 1 0 × × × The 5G LDPC codes are a class of rate-compatible
+0.49 1 1 1 × × × quasi-cyclic (QC) LDPC codes with raptor-like con-
-0.78 1 0 1 0 × × struction. Two base matrices, known as base graphs
-1.02 1 0 1 1 × × BG1 and BG2, have been specifically designed for
-1.33 1 0 0 1 × × different ranges of block lengths and code rates [29],
+1.69 1 0 0 0 1 ×
from which the parity-check matrices can be con-
+2.01 1 0 0 0 0 1
structed by the lifting operator with a factor Z. The
+2.52 1 0 0 0 0 0
structure diagram of the parity-check matrix is illus-
trated in Fig. 3, where the sub-matrices A and E con-

China Communications 5
Table 2. Information-Theoretic Analysis of Different Shaping Schemes SNR = 18 dB

Shaping Scheme m Geometric/Probabilistic Properties I(X; Y )


M-B Distribution 4 equidistant, non-equiprobable 2.99
Hybrid Shaping 6 non-equidistant, non-equiprobable 2.97
Many-to-One Mapping 6 equidistant, non-equiprobable 2.95
Geometric Shaping 4 non-equidistant, equiprobable 2.88
Uniform 4 equidistant, equiprobable 2.82

(a) (b)

(c) (d)

Figure 2. Four different constellations for 16-PAM when Es = 1. (a) Uniform for equidistant 16-PAM. (b) Geometric
shaping for equiprobable 16-PAM. (c) Dyadic probabilistic shaping for equidistant 16-PAM. (d) Hybrid shaping for 16-
PAM.

sist of circulant permutation matrices and zero matri-


ces, the sub-matrix I is an identity matrix, O is a zero A D
A O
matrix and D is a dual-diagonal matrix. A codeword
can be partitioned into two parts, where the left part
corresponding to [A, D] is referred to as the kernel
part and the remaining right part is referred to as the E I
extension parity part.

3.2 HS-BICM-ID System With 5G LDPC Punctured information bits


Codes
Figure 3. A diagram of the parity-check matrix for 5G
The HS-BICM-ID system is depicted in Fig. 4. At LDPC codes, where the sub-matrices A and E consist of
circulant permutation matrices and zero matrices, the sub-
the transmitter, the input sequence of information bits
matrix I is an identity matrix, O is a zero matrix and D is a
u ∈ Fk2 is firstly encoded by the LDPC encoder into a dual-diagonal matrix. The part [A, D] is referred to as the
codeword c ∈ Fn2 , and the codeword c is interleaved kernel part and corresponds to a high rate codeword.
by π to obtain ce. Then, the interleaved codeword ce is

6 China Communications
u c c x sign of the interleaver. This issue becomes more crit-
5G LDPC
encoder π HS Mapper
ical for the 5G LDPC codes since their parity-check
matrices have an uneven density in different parts.
AWGN
channel
Recalling that the many-to-one mapping introduces
û LA,Dec LE ,Dem y ambiguous bits, which can be viewed as punctured
π 1
5G LDPC
HS Demapper
decoder
bits for the LDPC codes, an intuitive idea is to as-
LE ,Dec LA,Dem
sociate the ambiguous bits with those “unimportant”
π coded bits. To be precise, we associate the ambigu-
ous bits with the extension parity part of a 5G LDPC
Figure 4. The diagram of the HS-BICM-ID system, where
codeword, which involves less number of parity-check
π is a row-column interleaver. At the receiver, an iterative
decoding/demapping is performed, which exchanges extrin- equations and hence has less effect on performance
sic messages between the demapper and the decoder. even punctured. This can be implemented by using
a row-column interleaver π between the encoder and
the mapper, and correspondingly, a deinterleaver π −1
mapped to the transmitted signal sequence x accord- between the decoder and demapper at the receiver for
ing to the HS mapper. At the receiver, an iterative HS-BICM-ID. A codeword of length n is written row
decoding/demapping is performed, which exchanges by row into an m × m n
array from top to bottom and
extrinsic messages between the demapper and the de- read out column by column from left to right. The po-
coder. Formally, following [30], we denote by LA,Dem sitions in the array where ambiguous bits may occur
(initialized properly) the input a priori information to are marked with the symbol ×. An example of row-
the demapper and by LE,Dem the output extrinsic in- column interleaver with n = 42 and m = 6 is shown
formation from the demapper. Similarly, LA,Dec and in Fig. 5, from which we see that, without interleaving,
LE,Dec are used for the decoder. Given a noisy sig- many ambiguous bits are associated with the kernel
nal yτ at time τ = 0, 1, · · · , n/m − 1, the extrinsic part (the left part), while, with interleaving, most am-
information in terms of log-likelihood-ratio (LLR) as- biguous bits are associated with the extension parity
sociated with the i-th bit of c̃τ from the demmaper can part of the codeword. Also notice that, with the row-
be calculated as column interleaver, one column of m bits is mapped
)to a signal point.
P
c̃τ ∈L0i p(yτ |x = ϕ(c̃τ ))p(c̃τ |LA,Dem
LE,Dem (τ, i) = ln P
c̃τ ∈L1i p(yτ |x = ϕ(c̃τ ))p(c̃τ |LA,Dem ) We can also optimize the design of the interleaver

− LA,Dem (τ, i), by considering the degrees of the variable nodes, fol-
(12) lowing [23], as the variable and check node degrees
influence the decoding threshold of the LDPC code.
n
(i)
o In Table 3, we present the variable node degree dis-
where Lbi = c̃τ ∈ Fm 2 |c̃τ = b for i =
tributions of the rate-1/2 5G LDPC codes, which are
0, 1, · · · , m − 1. Given the LLR sequence LA,Dec , the constructed based on BG2 with lifting factors 96, 192
extrinsic information of the t-th bit of a codeword from and 384. All codes have six different variable node
the decoder can be calculated as, degrees, which decrease from left to right. A bit inter-
leaver π sort the coded bits by bit-levels, i.e.,
p(ct = 0|LA,Dec )
LE,Dec (t) = ln − LA,Dec (t), (13)
p(ct = 1|LA,Dec )
π
c0 , c1 , · · · , cn−1 −→ c̃(0) , c̃(1) , · · ·, c̃(m−1) , (14)
where t = 0, 1, · · · , n − 1.
In the presented HS-BICM-ID system, we need to
specialize the structure of the interleaver. Different where c̃(i) denotes a string of n/m bits of level i.
from conventional BICM systems, in which a uniform There are (m−1)! choices of π, from which we choose
interleaver typically performs well, the BICM system the one with the best error performance, which is the
with many-to-one mapping requires an elaborate de- row-column interleaver.

China Communications 7
Codeword 100
kernel part extension parity part

Mapping without 10-1


interleaving
kernel part extension parity part

Row-input 10-2

FER
Column-output

Row-column interleaving 10-3



       extension parity part

10-4

Mapping with row-column


interleaving kernel part extension parity part
18.5 19 19.5 20 20.5 21 21.5 22
ambiguous bits

Figure 5. An example of row-column interleaver with Figure 6. Performance comparison of the designed row-
n = 42 and m = 6, from which we see that, without inter- column interleaver and the random interleaver for the HS-
leaving, many ambiguous bits are associated with the kernel BICM-ID system. The code is a rate-1/2 5G LDPC code of
part (the left part), while, with interleaving, most ambigu- length 1920 constructed based on BG2.
ous bits are associated with the extension parity part of the
codeword.
3.3 HS-MLC System With 5G LDPC Codes
The HS-MLC
n system is depicted
o in Fig. 7, where
(i) (i)
Table 3. Variable Node Degree Distribution of 5G LDPC codes c = ct , 0 ≤ t ≤ n − 1 denotes the codeword
of the i-th layer, ϕ is a cyclic shift mapper, and πi is
variable node degrees an interleaver at the i-th layer to be designed in the
k 7 5 4 3 2 1 following.
960 192 192 384 288 96 768 Again, due to the existence of the ambiguous bits,
1920 384 384 768 576 192 1536 we need to specialize the signal mapping ϕ. Let ψ :
3840 768 768 1536 1152 384 3072 Fm2 7→ A be a conventional signal mapper which takes
as input m bits, one from each layer, and delivers as
output a signal point for transmission, denoted by

(0) (1) (m−1)


xt = ψ(c̃t , c̃t , · · · , c̃t
), (0 ≤ t ≤ n − 1).
Example 3.(Performance Comparison of Different (15)
Interleavers): We have elaborately designed a row- A cyclic shift signal mapper ϕ is then defined by
column interleaver to place the ambiguous bits at the
extension parity part as much as possible. In this ex- (0) (1) (m−1)
xt = ϕ(c̃t , c̃t , · · · , c̃t )
ample, we compare the performance of the designed (16)
((t+0) mod m) ((t+m−1) mod m)
row-column interleaver and the random interleaver for = ψ(c̃t , · · · , c̃t ).
HS-BICM-ID. The code is a rate-1/2 5G LDPC code
of length 1920 which is constructed based on BG2 Evidently, the signal mapper ϕ is time-varying in the
with a lifting factor Z = 96. The average number sense that the input at time t to the signal mapper ψ is
((t+0) mod m) ((t+1) mod m) ((t+m−1) mod m)
of ambiguous bits is about 740 for the hybrid shap- (c̃t , c̃t , · · · , c̃t ),
(0) (1) (m−1)
ing scheme. Simulation results are shown in Fig. 6, i.e., a cyclic shift of (c̃t , c̃t , · · · , c̃t ), implying
where it can be seen that the adopted row-column in- that all layers are treated equally. In this way, the
terleaver (allocating about 645 ambiguous bits to the power or rate allocation strategy [31] can be avoided
extension parity part) can achieve a better performance and the same LDPC code can be applied to all layers,
compared to the random interleaver (allocating about suggesting a simple design. It is not difficult to
295 ambiguous bits to the extension parity part). imagine that, even with a unified labeling strategy

8 China Communications
u(0) 5G LDPC Encoder
c (0) π0 c (0)

u (1) c (1) π1 c (1)


5G LDPC Encoder
u Cyclic shift x
Demux mapping

u( m 1) c ( m 1) π c
( m 1)

5G LDPC Encoder m 1

Figure 7. The diagram of the HS-MLC system, where Demux denotes a demultiplexer that divides the input information bits
into m parts, πi denotes the elaborately designed interleaver of the i-th layer, ϕ denotes the cyclic shift mapper which is
used for allocating the ambiguous bits evenly to each layer.

from [32], if there is no cyclic shift mapping, all t  0,1, ,6

ambiguous bits will be allocated to the lower layers,


   c (0)
resulting in an unacceptable performance.
With the cyclic shift mapping, the coordinates at
   c (1)

m  0,1,
each layer are distinguished as two parts, one of which    c (2)

is with ambiguous bits. Then the i-th interleaver πi is ,5     c (3)

designed to arrange as many as possible extension par-     c (4)


ities in reverse order to the positions with the ambigu-     c (5)
ous bits. See the following example for illustrations.
Example 4.(Design of the interleavers for HS- Figure 8. A 6 × 7 array is used to illustrate the design of the
MLC): In this example, we provide the design method interleaver for each layer in HS-MLC. The i-th row of the
of the interleaver πi for HS-MLC, where m = 6, array corresponds to the codeword sequence c̃(i) , and each
n = 7 and three ambiguous bits might appear in each column is mapped to a signal point. The i-th interleaver πi
mapping. We use a 6 × 7 array as shown in Fig. 8 is designed to arrange the extension parity part in reverse
order to the positions marked by the symbol × as many as
to illustrate the design approach. The i-th row of the
possible while arranging the kernel part in positive order
array corresponds to the codeword sequence c̃(i) , and for the remaining positions. For example, the interleaver for
each column is mapped to a signal point. The i-th in- the 0-th layer can be expressed by π0 = (0, 1, 2, 6, 5, 4, 3).
terleaver πi is designed to arrange the extension par-
ity part in reverse order to the positions marked by the
symbol × as many as possible while arranging the ker-
nel part in positive order for the remaining positions. as confirmed in Fig. 9.
For example, the interleaver for the 0-th layer can be
expressed by π0 = (0, 1, 2, 6, 5, 4, 3). In the above example, we have implemented the it-
Example 5.(Performance of the Cyclic Shift Map- erative MSD algorithm for the HS-MLC system, as
ping): The performance of the cyclic shift mapping illustrated in Fig. 10 (a), where the soft-input-soft-
and the conventional signal mapping for HS-MLC are output (SISO) decoders and SISO demapper exchange
shown in Fig. 9, where the code of each layer is a rate- soft messages iteratively and the decoders are executed
1/2 5G LDPC code of length 1920 constructed based sequentially layer-by-layer. We can also decode the
on BG2. At the receiver, an iterative MSD algorithm is m layers in parallel by a nonbinary LDPC decoding
adopted. For the conventional signal mapping accord- algorithm, which does not require iterations between
ing to (14), due to the fact that all the ambiguous bits the decoders and the demapper. To do so, we need to
are allocated to Layer-3, Layer-4 and Layer-5, they treat the MLC system as a nonbinary LDPC code, as
have high error rates. For the presented cyclic shift presented in [33]. This is feasible since all the compo-
mapping, the ambiguous bits are evenly distributed to nent codes of the HS-MLC system are  with the same 
(i) (i) (i)
different layers, definitely improving the performance, parity-check matrix H. Let c = c0 , · · · , cn−1

China Communications 9
100 IV. PERFORMANCE COMPARISONS

10-1
In this section, the 5G LDPC codes with sum-product
algorithm (SPA) decoding are adopted in all simula-
tions. Unless otherwise stated, the signal constellation
10-2
for hybrid shaping is 16-PAM.
BER

10-3
4.1 Performance of HS-BICM-ID
10-4 For the HS-BICM-ID system, the information length
is set as k = 960 and the order of the dyadic mapping
is set as m = 5 or 6. To achieve the same spectral effi-
18.5 19 19.5 20 20.5 21
ciency SE = 3 bits/channel use, we employ 5G LDPC
codes with different code rates for different schemes.
Figure 9. Performance of the designed cyclic shift mapping For the uniform input scheme, a rate-3/4 5G LDPC
according to (16) and the conventional signal mapping ac- code of length 1301 is adopted. For the CCDM shap-
cording to (15) for the HS-MLC system. The code of each
ing scheme, a rate-5/6 5G LDPC code (without in-
layer is a rate-1/2 5G LDPC code of length 1920.
formation bits puncture) of length 1152 is used. For
the presented hybrid shaping scheme, a rate-1/2 and
a rate-3/5 5G LDPC code are employed for 6-bit and
be a codeword at the i-th layer. Define 5-bit hybrid shaping, respectively. The SPA with max-
imum local iteration number Imax = 50 is performed
c(0)
 
at the LDPC decoder and the maximum global itera-
C =  ...  tion (between the LDPC decoder and the demapper)
 
number Tmax = 10. The simulation results are shown
c(m−1)
 (0) in Fig. 11, from which we observe that
(0) (0) 
c0 c1 · · · cn−1 (17) • the proposed hybrid shaping can obtain a shaping
=  ... .. .. ..  gain of about 0.85 dB compared to the scheme

. . . 
(m−1)
c0
(m−1)
c1
(m−1)
· · · cn−1 with uniform input;

= (c0 , c1 , · · · , cn−1 ), • the proposed hybrid shaping can obtain an ex-


tra gain of about 0.2 dB compared to the CCDM
shaping;
(0) (m−1)
where ct = (ct , · · · , ct )T can be viewed as • the proposed hybrid shaping with m = 6 bits has
m
an element from F2 . It is not difficult to see that a better performance in comparison to that with
HCT = H(c0 , · · · , cn−1 )T = 0 and hence C can be m = 5 bits.
viewed as a codeword from a nonbinary LDPC code
Example 6.(Performance Comparison with latest
specified by the parity-check matrix H. Therefore,
Schemes): We compare the performance of hybrid
the HS-MLC system can be decoded by the nonbinary
shaping with the latest proposed probabilistic shap-
LDPC decoding algorithm, as illustrated in Fig. 10 (b).
ing scheme in [34] and the geometric shaping scheme
Remark: In terms of computational complexity, the in [35], respectively. The spectral efficiency is kept as
complexity of nonbinary joint decoding is higher than SE = 2.63 bits/channel use, and the length of the 5G
that of MSD decoding. As for the time complexity, LDPC code is n ≈ 7800. The simulation results of
MSD decoding is serial, leading to higher decoding la- different shaping schemes are shown in Fig. 12, from
tency, while joint decoding is parallel, allowing for si- which we observe that the proposed hybrid shaping
multaneous decoding of multiple layers. Furthermore, can obtain extra gains of 0.2 and 1.1 dB compared to
for the joint decoding, the iterations between the de- the shaping schemes in [34] and [35], respectively.
coders and the demapper can be avoided, effectively We have also simulated the performance of the hy-
reducing the consumption of hardware resources. brid shaping with information length k = 8448 and

10 China Communications
L(0) L(0) û (0)
π 01
E ,Dem A ,Dec
5G LDPC Decoder

L(1) L(1) û (1)


π11
E ,Dem A ,Dec
y HS 5G LDPC Decoder

Cyclic Shift Mux
Demmaper
1) 1)
L(Em,Dem
π m11
L(Am,Dec
5G LDPC Decoder
uˆ ( m1)

L(0) L(0)
A ,Dem
π0 E ,Dec

L(1) L(1)
E ,Dec
A ,Dem π1
1) 1)
L(Am,Dem L(Em,Dec
π m1
(a)

π 01

y HS π11 Bit-to- Nonbinary û


Cyclic Shift
Symbol Decoder
Demmaper

π m11

(b)

Figure 10. Two decoding schemes of the HS-MLC system. (a) The iterative multistage decoding, which requires iterations
between the decoders and the demapper. (b) The nonbinary joint decoding, which does not require iterations between the
demapper and the decoder.

SE = 3 bits/channel use
100
SE = 2.63 bits/channel use
100

10-1

10-1
FER

10-2
FER

10-2
0.85 dB
10-3
1.6 dB
10-3

10-4
18.5 19 19.5 20 20.5 21 21.5 22
10-4
15 15.5 16 16.5 17 17.5 18 18.5 19 19.5
Figure 11. Performance of the HS-BICM-ID scheme, where
the information length is set as k = 960 and the order of
Figure 12. Performance comparison with different shaping
the dyadic mapping is set as m = 5 or 6. A rate-5/6 5G
schemes, where the 5G LDPC codes of length n ≈ 7800 are
LDPC code (without information bits puncture) of length
adopted and m = 6. The spectral efficiency is kept as SE =
1152 constructed based on BG1 is adopted for CCDM shap-
2.63 bits/channel use.
ing. A rate-1/2 5G LDPC code of length 1920 is adopted
for 6-bit hybrid shaping.

China Communications 11
SE = 3 bits/channel use SE = 2 bits/channel use
100 100

10-1 10-1
FER

FER
10-2 10-2

1.2 dB 1.7 dB
10-3 10-3

10-4 10-4
17.5 18 18.5 19 19.5 20 20.5 21 11 11.5 12 12.5 13 13.5 14 14.5 15 15.5 16

Figure 13. Performance of the HS-BICM-ID scheme, where Figure 14. Performance of the HS-BICM-ID scheme, where
the information length k = 8448 and m = 6. A rate-5/6 5G the information length k = 8448 and m = 6. A rate-5/6 5G
LDPC code of length 11136 is adopted for CCDM shaping. LDPC code of length 11136 is adopted for CCDM shaping.
A rate-1/2 5G LDPC code of length 16896 is adopted for A rate-1/3 5G LDPC code of length 25344 is adopted for
the presented 6-bit hybrid shaping scheme. the presented 6-bit hybrid shaping scheme.

m = 6. A rate-3/4 5G LDPC code of length 11264 is 4.2 Performance Analysis of Different Spec-
adopted for the uniform input scheme and the geomet-
tral Efficiency
ric shaping scheme, and a rate-5/6 5G LDPC code of
length 11136 is adopted for CCDM shaping. A rate-
1/2 5G LDPC code of length 16896 is used for our The simulation results in Fig. 11 and Fig. 14 have also
presented 6-bit hybrid shaping scheme. The simula- shown that our hybrid shaping can more effectively
tion results are shown in Fig. 13, from which we ob- remedy the performance loss of many-to-one mapping
serve that at 2 bits/channel use compared to that at 3 bits/channel
• the proposed hybrid shaping scheme can obtain use. This can be veirified by the KL distance between
a shaping gain of about 1.2 dB compared to the the dyadic sub-optimal distribution PX∗ and the M-B
uniform input scheme; distribution PX , as shown in Table 4. It can be ob-
served that, at low and moderate spectral efficiency,
the KL distance between the dyadic distribution and
• the proposed hybrid shaping scheme can obtain the M-B distribution is relatively large, requiring our
extra gains of 0.2 and 1.0 dB compared to the hybrid shaping to further remedy the loss.
CCDM shaping and the conventional geometric
shaping, respectively;
Table 4. KL Distance Analysis of Different Spectral Efficiency for
• the proposed hybrid shaping scheme can achieve a Many-to-One Mapping m = 6
better performance compared to the dyadic prob-
abilistic shaping with many-to-one mapping. SE (bits/channel use) Code Rate D(PX∗ ||PX )
0.5 1/12 0.179
We also present the performance of hybrid shaping at
1.0 1/6 0.156
different spectral efficiency (SE = 2 bits/channel use),
1.5 1/4 0.153
which is shown in Fig. 14. From the simulation results
2.0 1/3 0.105
we can obtain that the proposed hybrid shaping can
2.5 5/12 0.058
obtain a shaping gain of about 1.7 dB compared to the
3.0 1/2 0.028
uniform scheme.

12 China Communications
SE = 3 bits/channel use ... ... ...

-1
10 Time 0 Time 1 Time 2
(a)

... ... ...

10-2 (b)
BER

Figure 16. The channel inputs for the time-division HS-


BICM-ID and the HS-MLC schemes. (a) The channel inputs
10-3 of the time-division HS-BICM-ID system. (b) The channel
inputs of the HS-MLC system.

10-4
19 19.5 20 20.5 21 21.5 22
SNR(dB)
4.4 Comparison of the Proposed Two Schemes
Figure 15. Performance of the HS-MLC scheme with non-
binary joint decoding and iterative MSD, where m = 6 and
SE = 3 bits/channel use. The code of each layer is a rate- In this subsection, we show by simulation that the
1/2 5G LDPC code of length 1920 for the presented hy-
HS-MLC can outperform the HS-BICM-ID when ap-
brid shaping scheme. A rate-3/4 5G LDPC code of length
1301 is used for the uniform input scheme and the geomet- plied to a fading broadcast channel characterized by
ric shaping scheme. Yt = Ht Xt + Zt , where Ht is a fading coefficient
that follows the Rayleigh distribution. We assume
that a transmitter attempts to broadcast in total Ns
4.3 Performance of HS-MLC symbols to m users, Ns /m for each user. One ap-
proach is to transmit Ns /m symbols in each of m time
slots (corresponding to m users), which is referred to
In this subsection, we compare the performance be-
as the time-division HS-BICM-ID scheme. Another
tween the nonbinary decoding and the iterative MSD
approach is to employ the HS-MLC scheme, jointly
for the HS-MLC, where m = 6. The code of each
processing the information of multi-layers to obtain
layer is a rate-1/2 5G LDPC code of length 1920 con-
Ns channel inputs. Figure. 16 shows the channel in-
structed based on BG2 for the 6-bit hybrid shaping
puts for the time-division HS-BICM-ID scheme and
scheme. To achieve the same spectral efficiency, a
the HS-MLC scheme, where different colors represent
rate-3/4 5G LDPC code of length 1301 is adopted for
different users. It can be observed that, for the HS-
the uniform input scheme and the geometric shaping
MLC system, an input signal contains coded bits from
scheme. For the iterative MSD decoding, the max-
all users, resulting in better tolerance to burst errors
imum global iteration number is Tmax = 5 and the
compared to the time-division HS-BICM-ID system.
maximum local iteration number is Imax = 10. The
maximum iterations of the nonbinary joint decoding Example 7.(Performance Comparison in Fading
is Imax = 50. The simulation results are shown in Channels): We present the performance of the time-
Fig. 15, from which we observe that division HS-BICM-ID scheme and the HS-MLC
• the performance of the hybrid shaping is signifi- scheme over fast fading channels, where the fading
cantly better than the uniform and the geometric coefficients are known at the receiver and change in-
shaping, both using the nonbinary joint decoding dependently from symbol to symbol. The signal con-
algorithm; stellation is 16-PAM and m = 6. In both cases, all
users use the same rate-1/2 5G LDPC code of length
• the performance of the nonbinary joint decod- 1920 constructed based on BG2. The simulation re-
ing (without iterative demapping) is slightly better sults are shown in Fig. 17. It can be seen that the HS-
than the iterative MSD (with unavoidable itera- MLC scheme has a significantly better performance
tive demapping) for the presented hybrid shaping compared to the time-division HS-BICM-ID scheme
scheme. over fading channels.

China Communications 13
and the Science, Technology and Innova-
100
tion Commission of Shenzhen Municipal-
ity with grant no. JCYJ20210324120002007,
10-1
and ZDSYS20210623091807023.
10-2
BER

10-3
References

[1] MASSEY J L. Coding and modulation in digi-


10-4
tal communication[C]//Zurich Seminar on Digi-
tal Communications, 1974: Vol. 2. 1974.
22 23 24 25 26 27 [2] IMAI H, HIRAKAWA S. A new multilevel
coding method using error-correcting codes[J].
Figure 17. Performance of the time-division HS-BICM-ID IEEE Transactions on Information Theory, 1977,
scheme and the HS-MLC scheme over fast fading channels. 23(3): 371-377.
In both cases, all users use the same rate-1/2 5G LDPC [3] ZEHAVI E. 8-PSK trellis codes for a Rayleigh
code of length 1920 constructed based on BG2. channel[J]. IEEE Transactions on Communica-
tions, 1992, 40(5): 873-884.
V. CONCLUSIONS [4] UNGERBOECK G. Channel coding with multi-
level/phase signals[J]. IEEE Transactions on In-
In this paper, we have integrated the 5G LDPC coded formation Theory, 1982, 28(1): 55-67.
modulation with hybrid shaping, resulting in the HS- [5] WACHSMANN U, FISCHER R, HUBER J.
BICM-ID and HS-MLC systems. We have achieved Multilevel codes: Theoretical concepts and prac-
significant shaping gains by elaborate designs without tical design rules[J]. IEEE Transactions on Infor-
introducing extra complexity compared to the dyadic mation Theory, 1999, 45(5): 1361-1391.
many-to-one probabilistic shaping. For HS-BICM-ID, [6] CAIRE G, TARICCO G, BIGLIERI E. Bit-
it can obtain shaping gains of about 0.2 dB and 1.0 dB interleaved coded modulation[J]. IEEE Transac-
compared to the CCDM shaping and the geometric tions on Information Theory, 1998, 44(3): 927-
shaping, respectively, while obtaining a shaping gain 946.
of about 1.2 dB compared to the scheme with uniform [7] LI X, CHINDAPOL A, RITCEY J. Bit-
input. For HS-MLC, the performance of the nonbi- interleaved coded modulation with iterative de-
nary joint decoding (without iterative demapping) is coding and 8PSK signaling[J]. IEEE Transac-
slightly better than the iterative MSD algorithm. For tions on Communications, 2002, 50(8): 1250-
the proposed two schemes, the HS-MLC system has a 1257.
significantly better performance compared to the time- [8] I FABREGAS A G, MARTINEZ A, CAIRE G,
division HS-BICM-ID system over fast fading chan- et al. Bit-interleaved coded modulation[J]. Foun-
nels. This work has also shown that, with many-to- dations and Trends® in Communications and In-
one mapping, the design of interleaver is critical, es- formation Theory, 2008, 5(1–2): 1-153.
pecially for unbalanced structured codes such as the [9] LIAO Y, QIU M, YUAN J. Design and analysis
5G LDPC codes. of delayed bit-interleaved coded modulation with
LDPC codes[J]. IEEE Transactions on Commu-
ACKNOWLEDGEMENT nications, 2021, 69(6): 3556-3571.
[10] MA H, LEUNG W K, YAN X, et al. Delayed bit
This work was supported by the National Key R&D interleaved coded modulation[C]//2016 9th In-
Program of China (No. 2021YFA1000500), the NSF ternational Symposium on Turbo Codes and It-
of China (No. 62301617 and No. 62271514), erative Information Processing. 2016: 86-90.
the Guangdong Basic and Applied Basic Re- [11] WANG L, CAI S, MA H, et al. Bit-labeling for
search Foundation (No. 2023A1515011056) delayed BICM with iterative decoding[C]//2018

14 China Communications
IEEE International Symposium on Information Transactions on Communications, 2015, 63(12):
Theory (ISIT). 2018: 1311-1315. 4651-4665.
[12] FORNEY G, GALLAGER R, LANG G, et al. [24] KURIHARA E, OCHIAI H. A systematic con-
Efficient modulation for band-limited channels stellation design for BICM systems with geo-
[J]. IEEE Journal on Selected Areas in Commu- metric shaping[C]//2023 IEEE 20th Consumer
nications, 1984, 2(5): 632-647. Communications Networking Conference. 2023:
[13] MA X, PING L. Coded modulation using super- 1048-1053.
imposed binary codes[J]. IEEE Transactions on [25] SHANNON C E. A mathematical theory of com-
Information Theory, 2004, 50(12): 3331-3343. munication[J]. The Bell system technical jour-
[14] VALENTI M C, XIANG X. Constellation shap- nal, 1948, 27(3): 379-423.
ing for bit-interleaved LDPC coded APSK[J]. [26] KSCHISCHANG F, PASUPATHY S. Optimal
IEEE Transactions on Communications, 2012, nonuniform signaling for Gaussian channels[J].
60(10): 2960-2970. IEEE Transactions on Information Theory, 1993,
[15] BOCHERER G. Probabilistic signal shaping for 39(3): 913-929.
bit-metric decoding[C]//2014 IEEE International [27] BOECHERER G. Optimal non-uniform map-
Symposium on Information Theory. 2014: 431- ping for probabilistic shaping[C]//International
435. ITG Conference on Systems, Communication
[16] YUAN R, BAI B, FANG J, et al. Two- and Coding. 2013: 1-6.
level LDPC-coded modulation with probabilis- [28] ZHOU D, NIU K, DONG C. Constellation shap-
tic shaping[J]. IEEE Communications Letters, ing for bit-interleaved polar coded-modulation
2023: 1-1. [C]//2016 IEEE 27th Annual International Sym-
[17] SUN F W, VAN TILBORG H. Approaching ca- posium on Personal, Indoor, and Mobile Radio
pacity by equiprobable signaling on the Gaussian Communications. 2016: 1-5.
channel[J]. IEEE Transactions on Information [29] 3GPP. [Z]. 3GPP TS 38.212 V15.1.1, 2018.
Theory, 1993, 39(5): 1714-1716. [30] FENG D, LI Q, BAI B, et al. Gallager map-
[18] LE GOFF S. Signal constellations for bit- ping based constellation shaping for LDPC-
interleaved coded modulation[J]. IEEE Transac- coded modulation systems[C]//2015 Interna-
tions on Information Theory, 2003, 49(1): 307- tional Workshop on High Mobility Wireless
313. Communications. 2015: 116-120.
[19] BARSOUM M F, JONES C, FITZ M. Constella- [31] MA X, PING L. Power allocations for multilevel
tion design via capacity maximization[C]//2007 coding with sigma mapping[J]. Electronics Let-
IEEE International Symposium on Information ters, 2004, 40(10): 1.
Theory. 2007: 1821-1825. [32] XIAO L, ZHAI X, LIU Y, et al. A unified bit-
[20] CRONIE H S. Signal shaping for bit-interleaved to-symbol mapping for generalized constellation
coded modulation on the AWGN channel[J]. modulation[J]. China Communications, 2023, 20
IEEE Transactions on Communications, 2010, (6): 229-239.
58(12): 3428-3435. [33] ZHAO S, MA X, BAI B. Decoding algorithms of
[21] GALLAGER R G. Information Theory and Re- LDPC coded superposition modulation[J]. IEEE
liable Communication: Vol. 588[M]. Springer, Communications Letters, 2014, 18(3): 487-490.
1968. [34] CORLAY V, DEMBELE H. A modified prob-
[22] YANKOV M, FORCHHAMMER S, LARSEN abilistic amplitude shaping scheme to use sign-
K J, et al. Rate-adaptive constellation shaping for bit-like shaping with a BICM[C]//2023 IEEE In-
near-capacity achieving turbo coded BICM[C]// formation Theory Workshop. 2023: 435-439.
2014 IEEE International Conference on Commu- [35] BOUTROS J J, EREZ U, WONTERGHEM J V,
nications. 2014: 2112-2117. et al. Geometric shaping: low-density coding of
[23] BOCHERER G, STEINER F, SCHULTE P. gaussian-like constellations[C]//2018 IEEE In-
Bandwidth efficient and rate-matched low- formation Theory Workshop. 2018: 1-5.
density parity-check coded modulation[J]. IEEE

China Communications 15
Jiayi Yang He is currently pursuing the
Ph.D. degree with Sun Yat-sen University,
Guangzhou, China. His research interests
include information theory, channel cod-
ing theory, and their applications to com-
munication systems.

Qianfan Wang received the Ph.D. de-


grees from Sun Yat-sen University, China,
in 2022. He is currently a Post-Doctoral
Fellow with Sun Yat-sen University. His
research interests include information the-
ory, channel coding theory, and their ap-
plications to communication systems.

Congduan Li received the B.S. degree


from the University of Science and Tech-
nology Beijing, China, in 2008, the M.S.
degree from Northern Arizona University,
AZ, USA in 2011, and Ph.D. degree from
Drexel University, PA, USA in 2015, re-
spectively, all in Electrical Engineering.
From Oct 2015 to Aug 2018, he was a
postdoctoral research fellow in the Institute of Network Coding
at the Chinese University of Hong Kong and in the Department of
Computer Science at City University of Hong Kong, China. He
is currently an Associate Professor with the School of Electronics
and Communication Engineering, Sun Yat-sen University, China.
His research interests lie in the broad areas related with networks,
such as coding, security, wireless, storage, and caching.

Xiao Ma received the Ph.D. degree in


communication and information systems
from Xidian University, China, in 2000.
He is a Professor of the School of Com-
puter Science and Engineering, Sun Yat-
sen University, Guangzhou, China. From
2000 to 2002, he was a Post-Doctoral Fel-
low with Harvard University, Cambridge,
MA. From 2002 to 2004, he was a Research Fellow with City Uni-
versity of Hong Kong. His research interests include information
theory, channel coding theory and their applications to communi-
cation systems and digital recording systems. He is a co-recipient,
with A. Kavčić and N. Varnica, of the 2005 IEEE Best Paper
Award in Signal Processing and Coding for Data Storage.

16 China Communications

You might also like