This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
Achievable Rate Region for Iterative
Multi-User Detection via Low-Cost Gaussian
Approximation
Xiaojie Wang, Student Member, IEEE, Chulong Liang, Li Ping, Fellow, IEEE,
and Stephan ten Brink, Senior Member, IEEE
Abstract—We establish a multiuser extrinsic information
transfer (EXIT) chart area theorem for the interleave-
division multiple access (IDMA) scheme, a special form of
superposition coding, in multiple access channels (MACs).
A low-cost multi-user detection (MUD) based on the
Gaussian approximation (GA) is assumed. The evolution
of mean-square errors (MSE) of the GA-based MUD
during iterative processing is studied. We show that the
K-dimensional tuples formed by the MSEs of K users
constitute a conservative vector field. The achievable rate
is a potential function of this conservative field, so it is
the integral along any path in the field with value of
the achievable rate solely determined by the two path
terminals. Optimized error correcting codes can be found
given the integration paths in the MSE fields by matching
EXIT type functions. The above findings imply that i) low-
cost GA detection can provide MAC capacity-approaching
performance, ii) the sum-rate capacity can be achieved
independently of the integration path in the MSE fields;
and iii) the integration path determining achievable rate
tuples of all users can be an extra degree of freedom for
code design.
Index Terms—EXIT chart, non-orthogonal multiple ac-
cess, area theorem, MAC capacity, multi-user detection.
I. I NTRODUCTION
Consider a multiple access channel (MAC) with K
users. The MAC capacity region is bounded by 2K − 1
constraints and determined by a tuple of user rates
Rk , 1 ≤ k ≤ K [2], [3]. To achieve arbitrary points
of the capacity region, joint detection and decoding
is required, which has prohibitively high complexity
exponential to K. Theoretically, successive interference
cancellation (SIC) together with time-sharing or rate-
splitting can achieve the entire capacity region [4]. SIC
Part of the results is published in the proceedings of the IEEE
International Symposium on Information Theory 2019, Paris, France
[1].
The work of Chulong Liang and Li Ping was supported by the
University Grants Committee of the Hong Kong Special Administrative
Region, China, under Projects CityU 11216817 and CityU 11216518.
X. J. Wang and S. ten Brink are with Institute of Telecommuni-
cations, Pfaffenwaldring 47, University of Stuttgart, 70569 Stuttgart,
Germany (e-mail: {wang, tenbrink}@inue.uni-stuttgart.de).
C. Liang and L. Ping are with the Department of Electronic
Engineering, City University of Hong Kong, Hong Kong SAR, China
(e-mail: liangchulong@qq.com and eeliping@cityu.edu.hk).
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
1
involves subtraction of successfully detected signals. If
practical forward error control (FEC) codes are used,
each subtraction incurs an overhead in terms of either
power or rate loss relative to an ideal capacity achieving
code [5, Fig. 13.3]. Such overheads accumulate during
SIC steps, moving its performance away from the ca-
pacity as K grows. Also, both time-sharing and rate-
splitting involve segmenting a data frame of a user
into several sub-frames. In practice, the length of a
coding frame is restricted by the latency requirement.
Frame segmentation results in shorter sub-frames and
therefore reduces the coding gain for a practical turbo
or low-density parity-check (LDPC) type code [6], which
further worsens the losses of accumulation.
Iterative detection [7]–[10] can alleviate the loss accu-
mulation problem using soft cancellations instead of hard
subtraction. A turbo or LDPC code involving iterative
detection can be optimized by matching the so-called
extrinsic information transfer (EXIT) functions of two
local processors, i.e., demapper/detector and FEC de-
coder, [11], [12]. In a single-user point-to-point channel,
such matching can offer near capacity performance, as
shown by the area properties [13], [14].
Code-division multiple-access (CDMA) is a conven-
tional approach to the MAC. Iterative multi-user de-
tection (MUD) has been studied for CDMA [15]–[17]
with impressive gains. CDMA is capacity achieving
under ideal coding and SIC. However, with practical
FEC codes, there is still a considerable gap between
CDMA performance and the MAC capacity. This is
because CDMA is not specifically optimized for iterative
detection and decoding, as explained graphically in [18].
Interleave-division multiple-access (IDMA) provides
a simple solution to closing this gap [19]. According
to information theory, mutually random codebooks for
different users are capacity achieving in a MAC [4].
These codebooks can be implemented by assigning each
user with a codebook generated from a master codebook
using a user-specific interleaving, provided that the resul-
tant codebooks are approximately mutually random. This
principle underpins IDMA and its theoretically capacity
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
approaching performance.
IDMA does also have several practical advantages.
As shown in [18], both CDMA and IDMA with LDPC
coding can be represented by sparse graphs, but the
former is “regular” (prone to a large number of short
cycles) while the latter is randomized with reduced
occurrence probability of short cycles. This difference
leads to the performance advantage of IDMA under
iterative processing. More details can be found in [6],
[18]–[20] on iterative processing on random graphs. The
connection between the occurrence probability of short
cycles and graph randomization was first revealed in
[21].
The complexity of an IDMA receiver can be further
reduced by Gaussian approximation (GA) on the residual
cross-user interference during iterative processing (See
Sec. II-A below for details). The per-user complexity of a
GA-based MUD remains roughly the same for all K. For
comparison, the complexity of a standard a posteriori
probability (APP) based MUD is exponential in K [19].
A question naturally arises: At such low cost, what is
the achievable performance of IDMA under GA-based
MUD? Some partial answers to this question are avail-
able. It was shown that IDMA is capacity approaching
when all users see the same channel [22]. In [23], [24],
the capacity optimality of IDMA with arbitrary user
power and antenna number was firstly proved for two-
user case and numerically verified for three-user case.
However, for the general MAC system, the code design
for IDMA becomes difficult when different users see
different channels. In the latter case, to achieve the entire
capacity region, different FEC code rates are generally
required. Previous works on IDMA focused on the
achievability of some special points in the MAC capacity
region [20], [25]–[28] and/or other aspects, e.g., power
control etc. [29]–[34]. To the best of our knowledge,
no previous work has shown that IDMA under the
GA-based MUD can achieve the entire MAC capacity
region. Recent developments on IDMA and multiple
access for 5G and beyond are discussed with practical
implementation techniques in [18], [35], performance
comparison among different NOMA schemes in [36],
[37] and multi-carrier, multiple antenna and relay aspects
in [38]–[40].
This paper analyzes the achievable performance of
IDMA under GA-based MUD. We approach the problem
by multi-dimensional curve matching of EXIT type
functions. Let vk be the mean-square error (MSE) (i.e.,
the variance) of the GA-based MUD for user k, with
vk = 0 indicating perfect decoding. Using the relation-
ship between mutual information (MI) and minimum
MSE (MMSE) derived in [14], [41], we show that
the achievable sum-rate can be evaluated using a line
integral along a valid path in the K-dimensional vector
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
2
T
field v = [v1 , v2 , · · · , vK ] . A main finding of this
paper is that the integral is path-independent and its
value is solely determined by the two terminations. The
path independence property greatly simplifies the code
optimization problem.
The findings and contributions of this paper are sum-
marized as follows.
• A low-cost GA-based MUD can provide near op-
timal performance. In particular, it is provably
capacity-achieving for Gaussian signaling.
• Relative to Gaussian signaling, the loss due to
discrete modulation can be made arbitrarily small
using a superposition coded modulation (SCM)
technique.
• FEC codes optimized for single-user channels may
not be good choices for MACs involving low-
cost GA-based MUD. The FEC codes should be
carefully designed to match MUD, which facilitates
iterative detection. We will provide examples for the
related code design.
• A multi-user area theorem of EXIT chart is es-
tablished for the code design. We show that the
sum-rate capacity is a potential function of the
MSE vector field formed by v, which leads to the
path independence property. All points of the MAC
capacity region are achievable using only one FEC
code per user. This avoids the loss related to the
frame segmentation of SIC as aforementioned.
• The above results can be extended to MIMO MAC
channels straightforwardly. We will provide simu-
lation results to show that properly designed IDMA
can approach the sum-rate MAC capacity for differ-
ent decoding paths in the MSE vector field within
1 dB .
This paper is structured as follows. In Sec. II, we
present the multi-user iterative detection and decoding
scheme in IDMA along with the matching condition.
Then, we derive the achievable rates of IDMA and show
its implication in code design in single antenna setup in
Sec. III. The achievable rate analysis is further extended
to MIMO cases in Sec. IV. Sec. V provides code design
examples and numerical results verifying our theorems.
Finally, Sec. VI concludes the paper.
II. I TERATIVE IDMA R ECEIVER
Consider a general K-user MAC system, character-
ized by
XK p
y= Pk hk xk + n, (1)
k=1
where Pk denotes the received signal strength of the
kth user’s signal, hk denotes the fading coefficients of
the user, xk is the transmitted signal of the kth user
and n is the additive (circularly symmetric complex)
white Gaussian noise (AWGN) with zero mean and unit
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
3

variance, i.e., CN 0, σ 2 = 1 . We consider complex- BPSK signaling, x̂i = tanh L2i with Li being the

valued signals throughout the paper if not otherwise log-likelihood ratio (LLR) after decoding. For higher
stated. Practically, a sequence of symbols y, forming one modulation schemes, the soft symbol estimates can be
or multiple codewords y, is received. obtained by [42, eqn. 3a], [43]. The feedback from
The iterative receiver is depicted in Fig. 1. The el- channel decoders will be discussed later in Sec. II-B.
The term zk in (2b) is comprised of AWGN and residual
ρ1 , y10 multi-user interference. To reduce complexity, we will
adopt Gaussian approximation (GA) assuming that zk is
Decoder 1 Gaussian-distributed with zero mean and variance σz,k
2
,
1
v1 , x01 i.e., CN (0, σz,k ) . From (2b), we obtain
2
..
. K
2
X
ρk , yk0 2
σz,k = Pi |hi | vi + σ 2 , (3a)
i=1,i6=k
y Decoder k
ESE
vk , x0k
where the MSE of the symbol estimates vi =
.. h
2
i
. E |xi − x̂i | is to characterize the quality of the
0 decoder feedback x̂i . The complexity in (2a) can be
ρK , y K reduced by a sum-and-minus trick by noting that z̃k =
K
Decoder K √ X √
Σ − Pk hk x̂k where Σ = Pi hi x̂i . Here Σ is
vK , x0K i=1
common to all users, so its cost can be shared. The per
user cost for (2a) thus does not grow with K. The quality
Fig. 1. The iterative multi-user detection and decoding model in of the ESE output for user k can be measured by the
IDMA.
signal-to-noise ratio (SNR) offered by yk in (2a)
ementary signal estimator (ESE) module has access to 2 2
Pk |hk | Pk |hk |
the channel observation y and feedbacks x0k from all the ρk = 2 = ,
σz,k K
users’ decoders. It performs the so-called soft interfer- X 2
Pi |hi | vi + σ 2
ence cancellation (SoIC) and output signals with reduced
i=1,i6=k
interference. Each decoder (DEC) performs decoding for
a particular user while treating the residual signals of ∀k = 1, 2, . . . , K. (3b)
other users as noise. Through iterative message passing Assume that the average power of xi is normalized to
between the ESE and DECs, the estimates are refined 1. Then vi = 1 in the first iteration, meaning no a prior
and interference suppressed progressively. More details information about xi . During the iterative detection, vi
on the ESE DECs are given below. For the convenience will be updated using decoder output (See the discussion
of discussions, we will assume that xi are modulated in Sec. II-B below). For K users, we express (3b) in a
using binary phase shift keying (BPSK). vector form as
ρ = φ (v) , (3c)
A. ESE functions
T
The function of the ESE (for elementary signal es- where ρ = [ρ1 , ρ2 , · · · , ρK ] and v =
T
timation) module in Fig. 1 is interference cancellation. [v1 , v2 , · · · , vK ] . Due to h the ifact that the MSE
2
The outputs of ESE are a sequence of yk with [19], [29] is bounded by 0 ≤ vi ≤ E |xi | = 1,we obtain that
p the SNR is also bounded by
yk = y − z̃k = Pk hk xk + zk , (2a)
2 2
K Pk |hk | Pk |hk |
X p ρk,min = ≤ ρk ≤ = ρk,max .
zk = Pi hi (xi − x̂i ) + n. (2b) K σ2
2
X
i=1,i6=k Pi |hi | + σ 2
i=1,i6=k
Here, yk is obtained from y in (1) by canceling out (3d)
K
X √ We will view (3) as a transfer function from v to ρ.
the mean of the interference z̃k = Pi hi x̂i
i=1,i6=k 1 The Gaussian assumption is valid for a large number of users with
based on the feedback of the channel decoders. The
arbitrary independently transmitted symbols xi as the consequence of
soft symbol estimates x̂i are generated by feedbacks the central limit theorem or if the transmit signals xi are Gaussian by
from the users’ channel decoders. For instance with themselves.

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
4

B. DEC functions C. Matching condition

The refined signals yk in (2) generated by the ESE
are forwarded to the DECs. The latter consists of K
local decoders (DECs, see Fig. 1) performing extrinsic
decoding based on yk with SNR ρk . To reduce com-
plexity, we will adopt Gaussian approximation (GA) that
zk is Gaussian-distributed with zero mean and variance
2
σz,k . Then the standard decoding operations [6], [12]
can be applied to the local decoders. The outputs of an
APP decoder are extrinsic messages that are assumed to
resemble observations from the AWGN channel, i.e.,
We will say that the ESE and DEC functions are
matched if the following condition is met
ψ (ρ (t)) = φ−1 (ρ (t)) . (6)
Note that the matching condition in (6) is along a K-
dimensional line given by ρ (t). It is not required to
match ψ (ρ) and φ (ρ) in the entire K dimensional
space, i.e., requiring ψ (ρ) = φ−1 (ρ). The line matching
in (6) is much easier. We will show that such line
matching achieves the MAC capacity (see Sec. III-A).

x0k = xk + wk , (4) III. ACHIEVABLE RATES

  The fundamental relation between achievable rate and
where wk follows a Gaussian-distribution CN 0, σw,k2
. MMSE in AWGN channels y = x + n is found in [41]
Let the MSEs for x̂k be vk after decoding. The MSEs as
vk are also the MMSE of the conditional mean estimator Z snr
due to APP decoding. Thus, we define a transfer function R (snr) = mmse (ρ) dρ
for DEC k as 0
h
2
i for any input distribution of x. The above result is
vk = E |xk − E [ xk | x0k ]| = ψk (ρk ) , extended to iterative decoding in [14]. Following [14]
0 ≤ vk ≤ 1, ∀k = 1, 2, . . . , K. (5a) and also [41], [42], the achievable rate for user k using
GA-based MUD is given by
Or in a vector form for the overall DEC Z ∞
f ρk + f −1 (vk ) dρk , ∀k = 1, 2, . . . , K,


Rk =
v = ψ (ρ) . (5b) 0
(7)
where f (ρ) = v is the achievable MMSE for a given
In general, unlike φ (·) in (3c), we do not have an explicit
constellation of x by observing y at the SNR of ρ.
expression for ψ (·) in (5b); but it can be numerically
Intuitively, ρk and f −1 (vk ) give, respectively, the SNRs
measured. The details can be found in [44].
related to the input and extrinsic messages of the kth
In general, ψk (ρk ) can be written as
DEC. Hence ρk + f −1 (vk ) represents the overall SNR

ρ ≤ ρk,min after combining these two messages.
1,

vk = ψk (ρk ) , ρk,max ≤ ρ ≤ ρk,min . (5c) A. Gaussian alphabets

0 ρ ≥ ρk,max .

We first consider the case when xi are Gaussian
Here the first case of vk = 1 is for the boundary distributed. This can be approximated by using, e.g.,
condition ρ ≤ ρk,min in (3d) at the start of the iterative superposition coded modulation (SCM) [43], [45]–[47].
detection. The last case of vk = 0 is for the boundary The MMSE for Gaussian signals is given by f (ρ) =
1+ρ , and with (7) the achievable rates can be expressed
1
condition ρ ≥ ρk,max in (3d) at the end of the itera-
tive detection when all interference has been perfectly as
Z ∞
canceled out and perfect decoding is assumed to be 1
Rk = −1 dρk (t) , ∀k = 1, 2, . . . , K.
achievable at this point. 0 ρk (t) + vk (t)
To track the convergence behavior of the iterations (7a)
between ESE and DECs, we write the SNR ρ and MSE Here vk (t) and ρk (t) are related by the function in (3b)
v vector as functions of an iteration variable t as and (5), and they are expressed as the functions of t,
as introduced in (5d). In Appendix A, we will consider
v = v (t) and ρ = ρ (t) . (5d) (3b) and (5) and rewrite (7a) into the following form:
Z vk =0
gk
Let t0 and t∞ denote the start and end of the iterative Rk = − K
dvk (t)
processing, we require that vk =1 X
gi vi (t) + σ 2

i=1
v (t0 ) = ψ (ρ (t0 )) = 1, (5e) vk =0
gk
Z
v (t∞ ) = ψ (ρ (t∞ )) = 0, (5f) =− T
dvk (t) , ∀k = 1, 2, . . . , K,
vk =1 g v (t) + σ2
since we are interested in the error-free decoding cases. (7b)

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
5

h iT
2 2 2
where gT = P1 |h1 | , P2 |h2 | , · · · , PK |hK | con- with the following assumptions
tains the powers of all users, v = [v1 , v2 , · · · , vK ] and
T 1) The exchanged messages during the iterative pro-
2 cessing of the extrinsic and a priori channel are
gk = Pk |hk | denotes the kth element of vector g. Note
that the achievable rate expression in (7b) depends on observations from AWGN channels, given in (3a)
multiple variables v1 , v2 , · · · and vK , i.e., the evolution and (4).
of the MSEs of all the DECs. This can be intuitively 2) The channel decoder is APP (i.e., MAP) decoder.
explained by the iterative soft interference cancellation 3) The channel encoders and decoders use “matched
of the ESE based on other users’ DEC feedbacks. codes” for a given path in the MSE vector field
Hence, the achievable sum-rate of all users can be given in (6).
written as Remark: The assumptions used in Theorem 1 are
XK Z
g common for turbo-type iterative receivers. It is generally
Rsum = Rk = − T 2
· dv (t) , (8a) accepted that these assumptions are sufficiently accurate
v(t) g v (t) + σ
k=1 for practical systems, based on which turbo and LDPC
where (8a) denotes a line integral defined by v (t) , t ∈ codes are designed in many modern communication
[t0 , t∞ ]. Notice that the line v (t) is determined by the systems [11], [48]. Theorem 1 provides guidelines for the
evolution of the MSE vector v of all DECs. We recall design of FEC codes in the matching condition discussed
that the terminals of the line are given in (5e) and (5f) as in Sec. II-C for multi-user scenarios. Further, various
v (t0 ) = 1 and v (t∞ ) = 0. It can be verified that the in- channel decoders such as BCJR and belief propagation
tegrands in (8a) constitute a gradient of a scalar field (or (BP) are known for achieving APP performance.
g
potential function), i.e., gT v+σ 2 = ∇v log σ + g v .
2 T

Thus, the achievable sum-rate can be written as

Z B. Finite alphabets
∇log σ 2 + gT v v 0 (t) dt


Rsum = − If the symbols xi ∈ Si are taken from finite alphabets
L=v(t)
 2 |Si | < ∞, the capacity formula, in general, can not be
σ + gT v (t0 )

= log expressed in closed-form. Notice that eq. (7) is still valid,
σ 2 + gT v (t∞ ) using the MMSE-formula for the underlying modulation
!
format f (ρ) = v. It is also well known that the loss
PK 2
(5e),(5f) k=1 Pk |hk |
= log 1 + , (8b) incurred by finite alphabets, compared with Gaussian, is
σ2
negligible in the low-SNR regime. Besides, the Gaussian
which is independent of the path taken for the integra- capacity can be approached by higher order modulations
tion. We note that the achievable rate in (8b) coincides with shaping and/or SCM [43], [45]–[47].
with the multi-user Shannon capacity. In other words, We provide in the following an achievable rate analy-
any path with matched DEC functions can achieve sis for IDMA with quadrature phase shift keying (QPSK)
the multi-user Shannon capacity. Therefore, the match- signaling. The achievable sum-rate can be written as
ing condition given in (6) is a sufficient condition Z
for achieving the sum-rate capacity. It can be further
 
−1
Rsum = fQ ρ (t) + fQ (v (t)) dρ (t) , (9)
verified that the matching condition also constitutes a ρ(t)
necessary condition for achieving the multi-user Shan-
non capacity. Consider the where fQ (·) denotes the MMSE of QPSK, which is
Z case v (t) < φ (ρ (t)),
−1
vk =0 given by
gk
then we have Rk < − gT v+σ 2
dvk and thus
v =1 ∞ y2
√ e− 2
PK k 
Z
P |h |2

Rsum < log 1 + k=1σ2k k . On the contrary, for fQ (ρ) = 1 − tanh(ρ − y ρ) √ dy. (10)
−∞ 2π
the case ψ (ρ (t)) > φ−1 (ρ (t)), error-free decoding is
not possible. In AWGN channels, Gaussian signals are the hardest to
This leads to the following theorem. estimate [49], i.e.,
Theorem 1. The achievable sum-rate in IDMA for any 1
fX (ρ) ≤ fG (ρ) = ,
path L (t) : vs = 1 → ve = 0 (starting from vs = 1 to 1+ρ
ve = 0) is given by the multiuser Shannon capacity
for any input distribution X with the same variance.
Hence, the achievable rate with distributions other than
Z
f ρ (t) + f −1 (v (t)) · dρ (t)


Rsum = −
v(t) Gaussian is essentially smaller.
Example: The users are assumed to have the same
PK !
2
k=1 Pk |hk |
= log 1 + power P with hk = 1, ∀k, and the same modulation
σ2
and coding scheme. For simplicity, we define t0 = 0

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
6

7 simplest path is a straight line between the starting point

Gaussian
v (t0 ) = 1 and the stop point v (t∞ ) = 0, denoted by
K = 1, QPSK
6 path 1. It is straightforward to obtain
K = 2, QPSK
gT 1
 
K = 4, QPSK gk
Rk = T log 1 + 2 , ∀k = 1, 2, . . . , K.
Achievable rate [bpcu]

5 K = 8, QPSK g 1 σ
K = 16, QPSK
K = 32, QPSK In this case, the achievable rate of each user is pro-
4 K = 128, QPSK portional to the received signal power strength gk . This
rate tuple coincides with the point where TDMA/FDMA
3 achieves the sum-rate capacity (see green dot in Fig. 3).
In path 1, it satisfies
2 v1 (t) = v2 (t) = · · · = vK (t) = v (t) , ∀ t.

1 The matching code for kth user shall have the following
MSE characteristic function

0 1  ρk ≤ ρk,min

−10 −5 0 5 10 15 20
 
1 1 2
vk = gT 1−gk · ρk − σ ρk,min ≤ ρk ≤ ρk,max .
SNRsum [dB] 
0 ρ ≥ρ

k k,max
Fig. 2. Achievable rates of multi-user IDMA with matching codes
and QPSK modulation; all users are assumed to have the same power, For path 2 and path 3 which are comprised of K
modulation and coding scheme; For fair comparison, the multi-user segments and each segment has merely value change
SNR SNRsum = KP/σ 2 is used as abscissa. (from vl = 1 to vl = 0) in one particular direction vl ,
i.e., within the segment dv dvk
dt 6= 0 and dt = 0, ∀k 6= l.
l

Depending on the order of the segments, there exist K!

and t∞ = 1 and consider the following integration path
such paths, which constitute the K! SIC corner points
ρ (t)
of MAC capacity region. The user rate can be written as
P !
ρ (t) = 1 gk
(K − 1) P + σ 2 Rk = log 1 + , ∀k = 1, 2, . . . , K,
  K
P P P
gl + σ 2
+ − · t · 1, t ∈ [0, 1],
σ2 (K − 1) P + σ 2 l=π(k)+1

and with (3b) where π (k) = k 0 denotes the permutation of user

decoding order with 1 ≤ π (k) ≤ K and π (k) 6=
(1 − t) σ 2
v (t) = · 1, t ∈ [0, 1]. π (k 0 ) , ∀k 6= k 0 . The corresponding decoding functions
σ2 + (K − 1) P · t are given by
Then, the achievable rate can be numerically evaluated (
for the specified path. We compare the achievable rates 1 ρk < ρk,SIC
vk = ψk (ρk ) = ,
using QPSK with different number of users in Fig. 2 by 0 ρk ≥ ρk,SIC
numerically solving the integral (9). Clearly, the loss to where
Gaussian capacity, due to discrete modulation, can be gk
ρk,SIC =
made arbitrarily small by imposing a larger number of K
X
users or date layers (which may belong to a same user) gπ(l) + σ 2
into the system. Although we assumed equal-power and l=π(k)+1

equal-rate for simplicity, the achievable rates analysis
can be extended to other general cases straightforwardly.
We will provide code matching examples for a three
user case based on QPSK signaling in Sec. V. The
achievable rates can also be found by the MSE evolution
method (compute the SNR-MSE transfer functions),
which are very close to the Gaussian capacity.
C. Path vs rate tuples
Consider a simple two-user case, i.e., K = 2. Fig. 3
illustrates some special paths and their corresponding
achievable rate tuple (or rate pair for K = 2). The
are the decoding thresholds. The decoding functions are
step functions with sharp transitions at corresponding
threshold SNRs ρk,SIC . This type of decoding functions
may pose difficulties for practical code designs, com-
pared to that with smooth transitions.
D. Achievable rate region
To achieve an arbitrary point of the MAC capacity
region, other paths shall be found. In the following the-
orem, we show that the entire MAC capacity region can
be achieved by proving the existence of a dedicated path
achieving an arbitrary point within the capacity region.

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
7

path 2
path 2 start iteration

FDMA
TDMA

1 path 1 path 3

path 3
ath
p

0 0
stop iteration

Fig. 3. Illustration (example for two users) of different integration paths achieving different rate pairs (R1 , R2 ); the arrows in the left figure
illustrate the two-dimensional MSE vector field; the achieved rate pairs are marked in the right figure for the corresponding paths; the dashed
lines denote paths achieving rate pairs moving (from the green dot) toward the corresponding SIC corner points.

Examples for constructing a dedicated path achieving a
feasible rate tuple are provided in case 2 of Sec. V-A.
Theorem 2. Under the assumptions in Theorem 1,
IDMA with GA-based MUD achieves every rate tuple in
the K-user MAC capacity region C (K ). Given a feasible
target rate tuple R = [R1 , R2 , · · · , RK ] ∈ C (K ), there
exists at least one path defined by vR (t) : vs = 1 →
ve = 0 which achieves the target rate tuple R.
Proof: See Appendix B.
Remark: It is easy to prove that there exists a unique
path for each of the K! SIC corner points and the
decoding functions shall be step functions. For other
rate tuples, it can be verified that there exist many
different paths achieving that rate tuple. The choice of
the integration path poses varying degrees of difficulty
for the design of matching codes. Thus, the design of an
appropriate integration path could be an extra degree of
freedom for the code design.
E. Gaussian Approximation
We provide in this section numerical evidence show-
ing that GA used in our achievable rate analysis is
accurate enough for addressing the behavior of a practi-
cal iterative IDMA multi-user demodulator and decoder.
The technique we used to track the probability density
function (PDF) of the exchanged messages during the it-
erative processing is discretized density evolution (DDE)
proposed in [50].
GA is arguably true for large number of users (central
limit theorem) and/or noise-limited scenarios (the noise
density rather the multiple access interference governs
the iterative process). We verified GA through DDE for
these cases (results omitted). Instead, we show results
for the following example with a few number of users
operating at relatively high SNR, since GA becomes
skeptical in these cases.
Example: The number of users is set to K = 4, each
with the same power P = 41 and BPSK modulation.
The multi-user SNR is −10 log σ 2 =20 dB. An LDPC
code with the variable node degree profile (from “edge”
perspective) 0.5231λ1 + 0.3187λ2 + 0.1582λ11 and the
check node degree η 2 is used for each user. The PDF
of the log-likelihood ratio (LLR) at the output of the
LDPC variable node decoder (VND) is tracked using
DDE with 10 bits and shown in Fig. 4 for the first
8 iterations. Clearly, the interference plus noise does
not resemble a Gaussian density at the first iteration.
As the consequence of the soft interference cancellation,
the density at VND becomes more Gaussian-like as the
iteration proceeds. Similar trend can be observed also
at the output of ESE and CND (results not shown).
Surprisingly, GA is quite accurate even for a few number
of users operating at high SNR regime.
IV. MU-MIMO C HANNEL
Assume that the kth transmitter has Nt,k antennas
and the receiver has NR antennas respectively; then, the
received signal can be written as
K p
X
y= Pk Hk xk + n, (11)
k=1
where Hk is the channel
 ofthe kth user, n denotes the
uncorrelated noise E nnH = σ 2 I. In this case, the

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
8

·10−2 where P = diag (P1 , P2 , · · · , PK ). Path independence

iteration 1 follows from the condition
1.4
iteration 2 Nt,k
∂  −1 H
 X −1
iteration 3 log det [R] = trace R Hk Hk = hH
k,i R hk,i
1.2 iteration 4 ∂vk i=1
iteration 5
1 iteration 6 with Jacobi’s formula.
prob. density

iteration 7
0.8 V. R ESULTS
The code design for multi-user can be complicated
0.6 [52]–[55]. For simplicity, we consider a single-input
single-output (SISO) setup. We assume that the power
0.4 levels gi = Pi |hi |2 (i = 1, · · · , K) are constant in our
code design. Thus, the multi-user SNR is defined as
0.2 PK
gi
SNRsum = i=1 . (14)
0 σ2
−10 −5 0 5 10 15 20 25 We consider K = 3 users with the power distribution
LLR at VND g = [g1 , g2 , g3 ]T = [ 71 , 27 , 47 ]T and we target the sum-
rate Rsum = R1 + R2 + R3 = 1 bpcu as an example.
Fig. 4. Evolution of the LLR distribution at the VND during the Theoretically, this sum-rate is attainable at the noise
iterative multi-user detection and decoding model with K = 4 users variance σ 2 = 1. Furthermore, the capacity region (more
at the SNR of 20 dB; the symbol x = +1 is assumed to be transmitted
for the user under test.
precisely, the dominant face which maximizes the sum-
rate) with Gaussian alphabets is given by
 g1 
ESE module is replaced by an iterative linear MMSE 0.1069 ≤ R1 ≤ log2 1 + 2 = 0.1926,
σ 
(LMMSE) receiver [42, eqn. (4a)]. Under the LMMSE-
 g2
0.2224 ≤ R2 ≤ log2 1 + 2 = 0.3626,
based ESE, the SNR of user k can be written as [51] σ 
 g3
Nt,k
0.4854 ≤ R3 ≤ log2 1 + 2 = 0.6521,
X σ
−1
hH
 
k,i R hk,i g1 + g2
R1 + R2 ≤ log2 1 + = 0.5145,
ρk = i=1
, (12) σ2
Nt,k  
X g1 + g3
1 − vk hH
k,i R
−1 h
k,i R1 + R3 ≤ log2 1 + = 0.7776,
σ2
i=1  
g2 + g3
where hk,i denotes the ith column of the kth user’s R2 + R3 ≤ log2 1 + = 0.8931.
σ2
channel matrix Hk and
A. ESE functions
R = σn2 I + HVHH ,
According to the matching condition in (6), for the
with V = diag (P1 v1 , P2 v2 , · · · , PK vK ) and H being design of capacity-achieving codes, the ESE transfer
the concatenated channels of all users. Following a functions ρ(t) = φ(v(t)) shall be determined. For this,
similar approach in Appendix A, we obtain with the we specify the K-dimensional decoding path v(t).
matching condition in (6) the user rate Rk as Following the path independence property of The-
vk =0 orem 1, we can constraint v(t) to be a piece-wise
linear path with n segments starting from the point

Z N t,k

v(t = 0) = x0 = 1, crossing the intermediate points

X
−1
Rk = − hH
k,i R hk,i dvk  .
i=1 v(t = i) = xi = [xi,1 , · · · , xi,K ]T , i = 1, 2, · · · , n − 1,
vk =1
and terminating at the point v(t = n) = xn = 0, where
Therefore, the sum-rate can be obtained as xi 6= xj , ∀i 6= j and for practical decoding
K
X Z v=0 1 ≥ x1,k ≥ x2,k ≥ · · · ≥ xn−1,k ≥ 0 ∀k (15)
Rsum = Ri = − ∇log det [R] dv
i=1 v=1 shall apply. Therefore, the path can be expressed in a

1
 vector form as
= log det I + 2 HH PH , (13)
σn v(t) = xi − (xi − xi+1 ) · (t − i), t ∈ [i, i + 1] (16)

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
9

for i = 0, 1, 2, · · · , n − 1. With the specified path, the Substituting (20) into (18), we obtain
ESE transfer function for user k can be computed as g1 (1 − x1,1 ) 2
gk 0.15 = log2
1 − g1 x1,1 − g2 x1,2 g1 x1,1 + g2 x1,2 + 1
ρk = φk (vk ) = T
g v(t) − gk vk (t) + σ 2 g1 x1,1 g1 x1,1 + g2 x1,2 + 1
gk + log2 ,
g1 x1,1 + g2 (x1,2 − x2,2 ) g2 x2,2 + 1
= h i ,
x −vk g2 (1 − x1,2 ) 2
g T xi − (xi − xi+1 ) · xi,ki,k −xi+1,k − gk vk + σ
2
0.3 = log2
1 − g1 x1,1 − g2 x1,2 g1 x1,1 + g2 x1,2 + 1
gk
=K , g2 (x1,2 − x2,2 ) g1 x1,1 + g2 x1,2 + 1

xi,k0 −xi+1,k0 xi,k xi+1,k0 −xi+1,k xi,k0
 + log2
2 g1 x1,1 + g2 (x1,2 − x2,2 ) g2 x2,2 + 1
P
gk0 xi,k −xi+1,k vk + xi,k −xi+1,k +σ
k0 6=k + log2 (g2 x2,2 + 1) ,
vk ∈ [xi+1,k , xi,k ] for i = 0, 1, 2, · · · , n − 1. (17) 0.55 =
g3
log2
2
.
1 − g1 x1,1 − g2 x1,2 g1 x1,1 + g2 x1,2 + 1
Note that when xi,k = xi+1,k , the above function is not (21)
valid. Actually, ρk is a vertical line from gT xi −ggkkxi,k +σ2
to gT xi+1 −ggkkxi+1,k +σ2 with vk = xi,k . Substituting (16) Solving (21), we obtain one feasible solution given
into (7b), we obtain the user rate Rk by x1,1 = 0.2145, x1,2 = 0.2056, x2,2 = 0.0618.
Z vk (t)=0 Substituting the solution into (17), we can obtain the
gk ESE functions. These transfer functions are depicted
Rk = − T v(t) + σ 2
dvk (t)
vk (t)=1 g in the middle sub-figure of Fig. 5 for the three users,
n−1
X gk (xi,k − xi+1,k ) g T xi + σ 2 respectively.
= T
log T . (18) 3) Case 3: We randomly choose the intermediate
g (xi − xi+1 ) g xi+1 + σ 2
i=0 points {xi }. Then, substituting the points into (17),
To verify the path independence property of Theorems we obtain the ESE transfer functions and subsequently
1 and 2, we consider three different paths for (17) and compute the rate for each user using (18).
evaluate the system performance and achievable rates via Here, we consider a piece-wise linear path with 2
density evolution and bit error rate (BER) simulations. segments by specifying an intermediate point arbitrarily,
1) Case 1: We do not specify any intermediate point e.g., x1 = [0.5, 0.2, 0.2]. Substituting x1 into (17), we
{xi }, i.e., n = 1. The path is a straight line between the have the ESE functions as
starting point v(t = 0) = 1 and the stop point v(t =
( g1
2, 0 ≤ v1 ≤ 0.5,
∞) = 0, as discussed in Sec. III-C. The ESE function ρ1 = (g2 +g3 )0.4v1 +σg1
g2 (1.6v1 −0.6)+g3 (1.6v1 −0.6)+σ 2 , 0.5 ≤ v1 ≤ 1,
for user k is given by (
g2
gk 2, 0 ≤ v2 ≤ 0.2,
ρk = T , vk ∈ [0, 1]. (19) ρ2 = g1 2.5v2 +g3 v2g+σ
(g 1 − gk )vk + σ 2 2
g1 (0.625v2 +0.375)+g3 v2 +σ 2 , 0.2 ≤ v2 ≤ 1,
The rate for user k is proportional to its power gk , i.e.,
(
g3
2, 0 ≤ v3 ≤ 0.2,
Rk = ggTk1 Rsum . Thus, the corresponding rate tuple is ρ3 = g1 2.5v3 +g2 v3g+σ 3
g1 (0.625v3 +0.375)+g2 v3 +σ 2 , 0.2 ≤ v3 ≤ 1.
(R1 , R2 , R3 ) = ( 17 , 27 , 47 ). The transfer functions in (19)
are depicted in the left most sub-figure in Fig. 5 for the These transfer functions are depicted in the right most
three users, respectively. sub-figure of Fig. 5 for the three users, respectively. The
2) Case 2: We construct a dedicated path to achieve achievable rates for each user, given in (22), can be
an arbitrarily chosen rate tuple in the MAC region, e.g., obtained by substituting x1 into (18), see Tab. I.
(R1 , R2 , R3 ) = (0.15, 0.3, 0.55). To find a dedicated
path, we search for {xi } by solving K non-linear B. LDPC code optimization
equations given by (18). Then, the ESE transfer functions As the ESE functions are readily available, according
can be obtained by substituting {xi } into (17). If n > 2, to the matching condition, we optimize the degree profile
there are K(n − 1) unknown variables {xi,k }, which is of LDPC codes to match the ESE functions for user k,
larger than K. This potentially lead to multiple solutions. i.e.,
It is noteworthy to mention that the variables {xi,k } are 
bounded in [0, 1] and shall satisfy (15). We may fix some  1, ρk ≤ ρk,min ,
unknown variables {xi,k } and solve the K non-linear vk = ψk (ρk ) = φ−1 (ρ k ), ρk,min ≤ ρk ≤ ρk,max ,
 k
equations given by (18) to obtain remaining unknown 0, ρk ≥ ρk,max ,
(23)
variables. Usually, we can fix K(n − 2) unknown vari-
where φ−1
k (ρk ) is the inverse of the ESE function of user
ables and have feasible solution for the remaining K gk gk
k and ρk,min = gT 1−g 2 , ρk,max = σ 2 . Using the
unknown variables. Here, we consider a 3-segment path k +σ
EXIT chart matching techniques [42], [48], [56], [57],
having intermediate points
the matching LDPC codes can be designed by properly
x1 = [x1,1 , x1,2 , 0]T and x2 = [0, x2,2 , 0]T . (20) choosing the degree distributions.

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
10

ESE, User 1 ESE, User 2 ESE, User 3

case 1 DEC, User 1 case 2 User 2
DEC, DEC, User 3 case 3
100 100 100

10−1 10−1 10−1

10−2 10−2 10−2

v
v
v

10−3 10−3 10−3

10−4 10−4 10−4

10−5 10−5 10−5

−12 −10 −8 −6 −4 −2 −12 −10 −8 −6 −4 −2 −12 −10 −8 −6 −4 −2
ρ [dB] ρ [dB] ρ [dB]

Fig. 5. ESE transfer function and the matching LDPC code transfer function for three different paths; the x-axis denotes the SNR of ESE and the
y-axis denotes the MSE of the feedback from channel decoder; three users with QPSK and the power distribution g = [g1 , g2 , g3 ]T = [ 17 , 72 , 47 ]T
are considered.

Table I
C ODE O PTIMIZATION R ESULTS FOR T HREE C ASES

Case Case 1, Rsum = 1 Case 2, Rsum = 1

User User 1 User 2 User 3 User 1 User 2 User 3
1 2 4 1 2 4
Power 7 7 7 7 7 7
[1, 1, 1] → [0.2145, 0.2056, 0]
Path [1, 1, 1] → [0, 0, 0]
→ [0, 0.0618, 0] → [0, 0, 0]
Target rate 0.1429 0.2857 0.5714 0.15 0.30 0.55
Check edge
η3 = 1 η4 = 1 η5 = 1 η3 = 1 η4 = 1 η5 = 1
distribution
variable degree
{2:1:30, 35:5:50, 60:10:100} {2:1:30, 35:5:50}
set {dv }
λ2 , 0.5239 λ2 , 0.3770 λ2 , 0.3293 λ2 , 0.5234 λ2 , 0.3779 λ2 , 0.3218
Optimized λ3 , 0.2140 λ3 , 0.2168 λ3 , 0.2351 λ3 , 0.2292 λ3 , 0.2290 λ3 , 0.2273
variable λ7 , 0.1627 λ7 , 0.0719 λ8 , 0.2500 λ7 , 0.0896 λ7 , 0.1431 λ7 , 0.0879
edge λ30 , 0.0685 λ8 , 0.1577 λ21 , 0.0654 λ8 , 0.0590 λ8 , 0.0580 λ8 , 0.1654
distribution λ35 , 0.0309 λ40 , 0.1237 λ22 , 0.0014 λ30 , 0.0708 λ50 , 0.1920 λ20 , 0.0501
λ100 , 0.0529 λ45 , 0.0258 λ35 , 0.0280 λ21 , 0.0335
λ50 , 0.0930 λ50 , 0.1140
Optimized
0.1467 0.3014 0.5707 0.1555 0.3154 0.5522
rate
Case Case 3, Rsum = 1 Case 2, Rsum = 2
User User 1 User 2 User 3 User 1 (1 Layer) User 2 (2 Layers) User 3 (4 Layers)
1 2 4 1 2 4
Power 7 7 7 7 7 7
Path [1, 1, 1] → [0.5, 0.2, 0.2] → [0, 0, 0] [1, 1, 1] → [0.9635, 0.7154, 0.0953] → [0, 0, 0]
Target rate 0.157 0.281 0.562 0.4 0.7 0.9
Check edge
η3 = 1 η4 = 1 η5 = 1 η4 = 1 η3 = 1 η3 = 1
distribution
variable degree
{2:1:30, 35:5:50} {2:1:50, 60:10:100}
set {dv }
λ2 , 0.5250 λ2 , 0.3788 λ2 , 0.3289 λ2 , 0.4107 λ2 , 0.6158 λ2 , 0.5703
Optimized λ3 , 0.2138 λ3 , 0.1925 λ3 , 0.2277 λ3 , 0.2398 λ3 , 0.2261 λ3 , 0.1699
variable λ6 , 0.0978 λ6 , 0.0763 λ8 , 0.1189 λ8 , 0.1875 λ6 , 0.0726 λ6 , 0.1130
edge λ7 , 0.0618 λ7 , 0.1589 λ9 , 0.1747 λ9 , 0.0339 λ7 , 0.0413 λ7 , 0.0713
distribution λ30 , 0.0604 λ50 , 0.1935 λ45 , 0.1337 λ10 , 0.0001 λ44 , 0.0002 λ16 , 0.0001
λ35 , 0.0412 λ50 , 0.0161 λ22 , 0.0001 λ50 , 0.0440 λ25 , 0.0277
λ35 , 0.0834 λ26 , 0.0477
λ100 , 0.0445
Optimized
0.1588 0.2926 0.5608 0.4145 0.6844 0.8652
rate

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
11

g1 (1 − 0.5)
 
1+1
R1 = log2
g1 (1 − 0.5) + g2 (1 − 0.2) + g3 (1 − 0.2) g1 · 0.5 + g2 · 0.2 + g3 · 0.2 + 1
g1 (0.5 − 0)
+ log2 (g1 · 0.5 + g2 · 0.2 + g3 · 0.2 + 1) = 0.157,
g1 (0.5 − 0) + g2 (0.2 − 0) + g3 (0.2 − 0)
g2 (1 − 0.2)
 
1+1
R2 = log2
g1 (1 − 0.5) + g2 (1 − 0.2) + g3 (1 − 0.2) g1 · 0.5 + g2 · 0.2 + g3 · 0.2 + 1
g2 (0.2 − 0) g1 · 0.5 + g2 · 0.2 + g3 · 0.2 + 1
 
+ log2 = 0.281,
g1 (0.5 − 0) + g2 (0.2 − 0) + g3 (0.2 − 0) 0+1
g3 (1 − 0.2)
 
1+1
R3 = log2
g1 (1 − 0.5) + g2 (1 − 0.2) + g3 (1 − 0.2) g1 · 0.5 + g2 · 0.2 + g3 · 0.2 + 1
g3 (0.2 − 0) g1 · 0.5 + g2 · 0.2 + g3 · 0.2 + 1
 
+ log2 = 0.562. (22)
g1 (0.5 − 0) + g2 (0.2 − 0) + g3 (0.2 − 0) 0+1

v 
u   2
dv,max u dc,max p 
X X
(27)
u 
λi · J   −1 1 − j − 1 · J −1 (1 − IE,V )  + 4ρ
IE,V =  (i − 1) J ηj · J ,
t
i=1 j=1

We basically follow the method described in [56, (resp., variable node), ρ is the decoder input SNR, and
Appendix 5G] to design irregular LDPC codes given the J(·) is defined by
a target transfer function v = ψ(ρ), where ρ is the
a priori SNR and v, the decoder output, denotes the 2 /2)2
−(y−σch
∞ 2σ 2
e
Z
extrinsic variance. The difference is that we use mutual J(σch ) = 1 −
ch
· log2 1 + e−y dy,
 
information instead of the mean of LLR to track the
p
2
2πσch
−∞
evolution process.
The asymptotic performance of an LDPC code ensem- its inverse function is further denoted by J −1 (·). Sub-
ble can be specified by its variable node and check node stituting (26) into (25), we obtain (27) where the LDPC
edge distribution polynomials, namely code can be characterized by one single variable IE,V .
dv,max dc,max The degree optimization problem can be formulated in
(28). The cost function in (28) is to maximize the code
X X
λ(x) = λi xi−1 and η(x) = ηi xi−1 , (24)
i=1 i=1
rate. Let IE,V,ini (ρ) be the initial extrinsic information
given by the channel, which can be written as
where λi (resp., ηi ) is the fraction of edges in the
bipartite graph of the LDPC code connected to variable dv,max q 
nodes (resp., check nodes) with degree i, and dv,max 2
X
IE,V,ini (ρ) = λi · J (i − 1) [J −1 (0)] + 4ρ
(resp., dc,max ) is the maximum variable node (resp., i=1
check node) degree. Moreover, we use the Gaussian √
= J (2 ρ) . (29)
approximation [58], i.e., (3a) and (4), to optimize the
edge distributions for the sake of simplicity.
Let IE,V,fin (ρ) denote the extrinsic information upon
In [48], it is shown that the decoder characteristic for
convergence, which corresponds to an output extrinsic
an LDPC code can be computed as
variance to the MUD v = ψ(ρ), i.e., given ρ, IE,V,fin (ρ)
dv,max q  should satisfy the following equation
2
X
IE,V = λi · J (i − 1) [J −1 (IE,C )] + 4ρ ,
i=1 dv,max 2 !
i · J −1 (IE,C,fin )

(25)
X
Λ i · fQ = ψ(ρ), (30)
dc,max i=1
4
X p 
IE,C = 1 − ηj · J j − 1 · J −1 (1 − IE,V ) ,
j=1 where fQ is defined in (10), and
(26)
dc,max
where IE,V (resp., IE,C ) is the extrinsic information
X p 
IE,C,fin = 1 − ηj · J j − 1 · J −1 (1 − IE,V,fin )
from variable node (resp., check node) to check node j=1

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
12

dv,max
X λi
max
{λi }
i=1
i
dv,max
X
s.t. λi = 1,
i=1
v 
u   2
dv,max u dc,max p 
X u X 
λi · J t(i − 1) J −1 1 − ηj · J j − 1 · J −1 (1 − I
E,V )  + 4ρ > IE,V
 
i=1 j=1

for ∀0 < ρ < ∞ and IE,V,ini (ρ) ≤ IE,V ≤ IE,V,fin (ρ). (28)

is the converged message from check nodes to variable 100

nodes. Furthermore, we define
λi /i
Λi = d
v,max
P 10−1
λi /i
i=1
as the fraction of variable node of degree i.
10−2
Shannon limit
Algorithm 1 Algorithm for LDPC Code Optimization
BER
simulation case 1
in IDMA evolution case 1
Input: Target decoder transfer functions vk = ψk (ρk ), 10−3 simulation case 2
check edge distribution ηk (x), maximum trial T , evolution case 2
threshold  and maximum variable degree dv,max . simulation case 3
Output: The optimized variable edge distribution 10−4 evolution case 3
λ(T ) (x).
1: Initialize λ(0) (x) = x.
2: for t = 1 to T do 10−5
0 0.5 1 1.5 2 2.5
3: Solve (28) by linear programming to obtain
λ(t) (x), where IE,V,fin (ρ) in (28) is obtained by SNRsum [dB]
solving (30) using λ(t−1) (x). Fig. 6. BER curves and density evolution results for a three user
dv,max
P (t) (t−1) MAC with matched codes; three different paths and QPSK signaling
λi λi
are considered.
4: if 1- v i=1
! ! ≤  then
dv,max dv,max
u
(t) (t−1) 2
u P P
t (λi )2 (λi )
i=1 i=1
ever, the optimized degree sequences are mostly com-
5: λ(T ) (x) = λ(t) (x). prised of a few small degrees. Therefore, we only use
6: return λ(T ) (x). a subset of degrees less than dv,max to run Algorithm 1
7: end if more efficiently. In the algorithm, we set T = 100
8: end for and  = 0.001 for all optimizations. The optimized
9: return λ(T ) (x). results as well as other parameters are summarized in
Table I. Fig. 5 shows the optimized LDPC DEC transfer
The optimization in (28) is a non-convex optimization. functions (denoted by dashed lines). The DEC functions
However, given {Λi } and η(x), the problem in (28) can match enough well with the ESE functions.
be solved using standard linear programming. We use an After matching the degree distribution, we construct
iterative way to optimize the edge distribution λ(x) with parity check matrices for BER simulations. The parity-
fixed η(x) in Algorithm 1. In practice, Algorithm 1 is check matrix of code-word length 105 for each user in
repeated for several check edge distributions η(x) till a each case is randomly generated and subsequently we
matching code is found. remove the cycle-4 loops in the matrix by edge permu-
tation [59]. Furthermore, we consider QPSK signaling in
C. Numerical Results the simulation to verify that our Theorems work also well
With Algorithm 1, if we use all variable degrees less with finite alphabets, not only with Gaussian alphabets.
than dv,max , the optimization can be quite slow. How- Fig. 6 shows the average BER performance for three

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
users with matching codes along different decoding paths
(denoted by solid lines, respectively), where we set a
maximum iteration of 1000 between the ESE detector
and the LDPC decoders. Moreover, the Shannon limit
at the sum-rate Rsum = 1 along with the density
evolution performance with optimized codes considering
QPSK are provided. From the numerical results, we can
conclude that
• the evolution thresholds for three cases with QPSK
are close to the Gaussian capacity. The loss incurred
by finite alphabets is negligible at the target sum-
rate.
• Three cases have BER below 10−4 within 1 dB
to the Shannon limit. The sum-rate capacity can
be achieved for different paths also with QPSK
signaling.
• For case 2, we computed a dedicated path to achieve
an arbitrarily chosen rate tuple (R1 , R2 , R3 ) =
(0.15, 0.3, 0.55). The numerical results also verified
our path construction based on GA.
• Case 1 has the worst performance; this may be
due to its worst girth property of the randomly
constructed parity check matrix because of its low
code-rate; we conjecture that more sophisticated
methods of parity check matrix design, e.g., pro-
gressive edge growth (PEG) [60] and methods in
[61] could be used to improve its performance.
To further verify that the decoding path (or decoding
trajectory L(t)) in the BER simulation is close to the
desired theoretical path, we compare the decoding trajec-
tories obtained via density evolution and BER simulation
for three cases at the SNR SNRsum = 1 dB, where
all users can decode its signal with high probability.
The evolution trajectories differ from the specified paths
(discussed in Sec. V-A) mainly due to the different
SNRs (the specified paths assume SNRsum = 0 dB).
We observe that the simulation trajectories are consistent
with those of density evolution for all three cases. This
further consolidates the path independence theorem, and
provides numerical evidence the validity of Theorem 1
to finite alphabet cases.
D. High Rate
For scenarios where high rates for users are required,
SCM [43], [45] can be applied. We consider the case that
the data layers in SCM are of the same power. Suppose
that we have a K-user system with power allocation
[g1 , g2 , · · · , gK ]T . The decoding path x0 → x1 →
· · · → xn is also specified. Then, we can convert this
unequal power system into an equivalent equal power
system as follows.
• Find a real number g > 0 such that Li = gi /g is
an integer for all i.
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
13
• Change each K-dimension
PK point xi into a point xi
0
with dimension L = i=1 Li as follows.
x0i =
 
 
xi,1 , · · · , xi,1 , xi,2 , · · · , xi,2 , · · · , xi,K , · · · , xi,K  .
| {z } | {z } | {z }
L1 copies L2 copies LK copies
• Finally, we have an equal power system with L
virtual users each with power g and the decoding
path is x00 → x01 → · · · → x0n , where the original
user i is the superposition
Pi−1 of the Pivirtual users with
indices from 1 + j=1 Lj to j=1 Lj .
By substituting the modified power allocation and path
into (18), it is easy to find that the equal power system
satisfies the requirement of the original K-user system.
Table I shows the optimized LDPC code for a targeted
sum-rate of Rsum = 2 for case 2, where the individual
user rate is pre-defined and a dedicated decoding path
is then specified to achieve that rate tuple. For the
simplicity of code design, we apply SCM to each user.
In particular, the to be transmitted data packet for user
2 and user 3 is divided into two and four independent
data layers, respectively. By doing this, the three-user
MAC system is converted to a MAC system with seven
“users”, each with the same transmit power. The main
motivation for applying SCM is that the curve matching
code design becomes difficult at high rate [22]. It may
require rather complicated joint design of modulation
and coding scheme for each user. Simulation and density
evolution results are depicted in Fig. 8 for the most
interesting case 2 (user rate is prescribed). Compared
to the low-rate scenario in Fig. 6, the gap to Shannon-
limit increases both for density evolution and simulation
results at high rate (0.4 dB and 1.5 dB respectively).
However, we conjecture that this gap can be reduced
by imposing more data layers. However, increasing the
number of data layers essentially decreases the code-rate
per layer. It is a rather challenging task to design low-
rate LDPC codes with good performance [62], [63]. To
this end, more complexity-demanding and sophisticated
LDPC code design methods such as density evolution
[12], generalized constraint nodes based on Hadamard
codes [63], serial concatenation with a repetition code
[20] and PEG [60] may be used in future work to verify
this conjecture.
VI. C ONCLUSION
We proved under Gaussian approximation (GA) that
the simple interleave-division multiple-access (IDMA),
relying on a low-cost GA based multi-user detector
(MUD), is capacity-achieving for the general Gaussian
multiple access channels (GMAC) with arbitrary number
of users, power distribution and with single or multiple
antennas. We showed that IDMA with matching codes
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
14

simulation case 1 evolution case 1

simulation case 2 1 evolution case 2
simulation case 3 evolution case 3
1
0.8
v3

0.5
0.6

v3
0.4
0
1
0.2
0.5
0.8 1
v2 0.4 0.6 0
00 0.2 0 0.2 0.4 0.6 0.8 1
v1 v1

Fig. 7. Evolution and simulation trajectories for three cases at SNRsum = 1 dB; Left: 3-D diagram for a trajectory (v1 , v2 , v3 ); Right: Side
view (v1 , v3 ) of a trajectory.

100 A PPENDIX A
Shannon limit P ROOF OF (7b)
simulation case 2
evolution case 2
10−1 Let the SNR of the output of ESE be ρk,min , ρk,max as
defined in (3d). Since the achievable rate formula in (7)
requires the integration for the SNR ρ spanning (0, ∞),
10−2 we explicitly write the transfer function of the DECs as
BER


1,
 ρ ≤ ρk,min
10−3 vk = ψk (ρk ) , ρk,max ≤ ρ ≤ ρk,min .

0 ρ ≥ ρk,max


10−4
We also assume that the matching condition in (6) holds.
Then, the achievable rates can be expressed as
10−5 ρk,max ρk,min
1 1
Z Z
4.5 5 5.5 6 6.5 7
Rk = dρk + dρk .
SNRsum [dB] ρk,min ρk + vk−1 0 ρk + 1

Let ρ0k be the first derivative of ρk with respect to vk ,

Fig. 8. BER and density evolution results for a three-user MAC with we obtain
matched codes at the sum-rate Rsum = 1.964; SCM and QPSK
signaling are considered.
dρ vk =0 Z ρk,min
ρ0k = dvk ρ0k 1
Z
k
Rk = −1 dvk + dρk
vk =1 ρk + vk 0 ρ k +1
is capacity-achieving for arbitrary decoding path in the
ρk − vk−2 + vk−2
Z 0 0
mean-square error (MSE) vector field. This property was = dvk + log (1 + ρk,min )
further used to prove that IDMA achieves not only the 1 ρk + vk−1 | {z }
=w0
sum-rate capacity, but the entire GMAC capacity region. vk =0
The construction of capacity-achieving codes was also vk−2
 Z
= log ρk + vk−1 +


dvk + w0
provided by establishing an area theorem for multi-user ρk + vk−1 vk =1
extrinsic information transfer (EXIT) chart. We provided  Z   vk =0
numerical evidence supporting GA and our achievable 1
= log ρk + vk−1 + vk−1 − −1


dvk + w0 .
rate analysis. ρk + vk vk =1

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
15

2
Let gk = Pk |hk | be the kth element of the vector g [3] A. E. Gamal and Y.-H. Kim, Network Information Theory.
and v = [v1 , v2 , · · · , vK Cambridge, U.K.: Cambridge Univ. Press, 2011.

], we can express (3b) as ρk = [4] D. Tse and P. Viswanath, Fundamentals of Wireless Communi-
gk / gT v − gk vk + σ 2 and obtain cations. Cambridge, U.K.: Cambridge Univ. Press, 2005.
vk =0 [5] Y. Hu and L. Ping, “Interleave-Division Multiple Access
(IDMA),” in Multiple Access Techniques for 5G Wireless Net-

gk
Z
(3a) −1


Rk = log ρk + vk + log vk − dvk works and Beyond, M. Vaezi, Z. Ding, and H. V. Poor, Eds.
gT v + σ 2 vk =1 New York City, U.S.: Springer, 2019, ch. 13, pp. 417–449.
+ w0 [6] T. Richardson and R. Urbanke, Modern Coding Theory, 1st ed.
Cambridge, U.K.: Cambridge Univ. Press, 2008.
 vk =0 [7] C. Schlegel, R. Kempter, and P. Kota, “A novel random wireless
gk
Z
= log (ρk vk + 1) − dvk + w0 packet multiple access method using CDMA,” IEEE Trans.
g v + σ2
T
vk =1 Wireless Commun., vol. 5, no. 6, pp. 1362–1370, Jun. 2006.
0 [8] M. Jiang, J. Akhtman, and L. Hanzo, “Iterative joint channel
gk
Z
=− dvk , estimation and multi-user detection for multiple-antenna aided
1g
Tv + σ2 OFDM systems,” IEEE Trans. Wireless Commun., vol. 6, no. 8,
pp. 2904–2914, Aug. 2007.
where the last equality is due to w0 = log (1 + ρk,min ) = [9] P. A. Hoeher, S. Badri-Hoeher, Wen Xu, and C. Krakowski,
log (1 + ρk (vk = 1)). “Single-antenna co-channel interference cancellation for TDMA
cellular radio systems,” IEEE Trans. Wireless Commun., vol. 12,
no. 2, pp. 30–37, Apr. 2005.
A PPENDIX B [10] M. Zhao, Z. Shi, and M. C. Reed, “Iterative turbo channel esti-
P ROOF OF T HEOREM 2 mation for OFDM system over rapid dispersive fading channel,”
Z vk =0 IEEE Trans. Wireless Commun., vol. 7, no. 8, pp. 3174–3184,
gk Aug. 2008.
The user rate Rk = − gT v+σ 2
dvk is a con- [11] S. ten Brink, “Exploiting the chain rule of mutual information
vk =1
tinuous and monotonically decreasing function of v. If for the design of iterative deoding shemes,” in Proc. 39th Annual
Allerton Conf. Commun., Control and Computing, 2001.
vk are unbounded, then Rk can take on anyvalue with  [12] T. J. Richardson, M. A. Shokrollahi, and R. L. Urbanke, “De-
X T 2
the single sum-rate constraint Rk ≤ log g 1+σ σ2 . sign of capacity-approaching irregular low-density parity-check
codes,” IEEE Trans. Inform. Theory, vol. 47, no. 2, pp. 619–637,
In other words, there exists at least one integration Feb. 2001.
path which allows achievingX an arbitrary point
 T within [13] A. Ashikhmin, G. Kramer, and S. ten Brink, “Extrinsic informa-
tion transfer functions: model and erasure channel properties,”
2

the region determined by Rk ≤ log g 1+σ σ2 . IEEE Trans. Inform. Theory, vol. 50, no. 11, pp. 2657–2673,
However, the value range of Rk is constrained by the fact Nov. 2004.
that 0 ≤ vl ≤ 1, ∀l. Furthermore, the integrand gT v+σ gk [14] K. Bhattad and K. R. Narayanan, “An MSE-based transfer
2
chart for analyzing iterative decoding schemes using a Gaussian
is monotonically decreasing with vl , ∀l 6= k. Therefore, approximation,” IEEE Trans. Inform. Theory, vol. 53, no. 1, pp.
the maximum of the kth user rate Rk is attained when 22–38, Jan. 2007.
vl = 0, ∀l 6= k, i.e., [15] M. C. Reed, C. B. Schlegel, P. D. Alexander, and J. A. Asen-
storfer, “Iterative multiuser detection for CDMA with FEC: near-
single-user performance,” IEEE Trans. Commun., vol. 46, no. 12,
Z vk =0
gk + σ 2
 
gk
Rk ≤ − 2
dv k = log . pp. 1693–1699, Dec. 1998.
vk =1 gk vk + σ σ2 [16] X. Wang and H. V. Poor, “Iterative (turbo) soft interference
cancellation and decoding for coded CDMA,” IEEE Trans. Com-
Similarly, the following constraints can also be obtained mun., vol. 47, no. 7, pp. 1046–1061, Jul. 1999.
[17] F. N. Brannstrom, T. M. Aulin, and L. K. Rasmussen, “Iterative
gk + gl + σ 2
 
Rk + Rl ≤ log , ∀k 6= l multi-user detection of trellis code multiple access using a pos-
σ2 teriori probabilities,” in Proc. IEEE Int. Conf. Commun., vol. 1,
Jun. 2001, pp. 11–15.
gk + gl + gm + σ 2
 
Rk + Rl + Rm ≤ log , ∀k 6= l 6= m[18] Y. Hu, C. Liang, J. Hu, and L. Ping, “Low-cost implementation
σ2 techniques for interleave division multiple access,” IEEE Wireless
Commun. Lett., vol. 7, no. 6, pp. 1026–1029, Dec. 2018.
.... [19] L. Ping, L. Liu, K. Wu, and W. K. Leung, “Interleave division
..
 T multiple-access,” IEEE Trans. Wireless Commun., vol. 5, no. 4,
g 1 + σ2 pp. 938–947, Apr. 2006.
X 
Rk ≤ log , [20] X. Wang, S. Cammerer, and S. ten Brink, “Near-capacity de-
σ2 tection and decoding: Code design for dynamic user loads
and these constraints constitute the MAC capacity re- in Gaussian multiple access channels,” IEEE Trans. Commun.,
2019.
gion. Hence, for any point in rate region determined by [21] S. Benedetto and G. Montorsi, “Unveiling turbo codes: some
the above constraints (or equivalently the K-user MAC results on parallel concatenated coding schemes,” IEEE Trans.
capacity region), there exists at least an integration path Inform. Theory, vol. 42, no. 2, pp. 409–428, Mar. 1996.
[22] Y. Hu, C. Liang, L. Liu, C. Yan, Y. Yuan, and L. Ping, “Interleave-
constrained by vl , ∀l 6= k achieving that point. division multiple access in high rate applications,” IEEE Com-
mun. Lett., 2018.
R EFERENCES [23] L. Liu, C. Yuen, Y. L. Guan, and Y. Li, “Capacity-achieving
iterative LMMSE detection for MIMO-NOMA systems,” in Proc.
[1] X. Wang, C. Liang, L. Ping, and S. ten Brink, “Achievable rate IEEE Int. Conf. Commun., May 2016, pp. 1–6.
region for iterative multi-user detection via low-cost Gaussian [24] L. Liu, Y. Chi, C. Yuen, Y. L. Guan, and Y. Li, “Capacity-
approximation,” in Proc. IEEE Int. Symp. Inform. Theory, 2019. achieving iterative LMMSE detection for MIMO-NOMA sys-
[2] T. M. Cover and J. A. Thomas, Elements of Information Theory, tems,” IEEE Trans. Signal Process., vol. 67, no. 7, pp. 1758–
2nd ed. New Jersey, USA: John Wiley & Sons, 2006. 1773, Apr. 2019.

1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2020.2971999, IEEE
Transactions on Wireless Communications
[25] K. Li and X. Wang, “EXIT chart analysis of turbo multiuser
detection,” IEEE Trans. Wireless Commun., vol. 4, no. 1, pp.
300–311, Jan. 2005.
[26] G. Song and J. Cheng, “Low-complexity coding scheme to
approach multiple-access channel capacity,” in Proc. IEEE Int.
Symp. Inform. Theory, Jun. 2015, pp. 2106–2110.
[27] X. Wang, S. Cammerer, and S. ten Brink, “Near Gaussian
multiple access channel capacity detection and decoding,” in
Proc. IEEE Int. Symp.Turbo Codes Iterative Inform. Process.,
Dec. 2018, pp. 1–5.
[28] R. Zhang, L. Xu, S. Chen, and L. Hanzo, “EXIT-chart-aided
hybrid multiuser detector for multicarrier interleave-division mul-
tiple access,” IEEE Trans. Veh. Technol., vol. 59, no. 3, pp. 1563–
1567, Mar. 2010.
[29] K. Li, X. Wang, and L. Ping, “Analysis and optimization
of interleave-division multiple-access communication systems,”
IEEE Trans. Wireless Commun., vol. 6, no. 5, pp. 1973–1983,
May 2007.
[30] K. Kusume, G. Bauch, and W. Utschick, “IDMA vs. CDMA:
Analysis and comparison of two multiple access schemes,” IEEE
Trans. Wireless Commun., vol. 11, no. 1, pp. 78–87, Jan. 2012.
[31] J. Song and Y. Shu, “On construction of rate-compatible raptor-
like QC-LDPC code for enhanced IDMA in 5G and beyond,” in
Proc. 10th Int. Symp. Turbo Codes, 2018.
[32] P. Hammarberg, F. Rusek, and O. Edfors, “Channel estimation
algorithms for OFDM-IDMA: Complexity and performance,”
IEEE Trans. Wireless Commun., vol. 11, no. 5, pp. 1722–1732,
May 2012.
[33] T. Yang, J. Yuan, and Z. Shi, “Rate optimization for IDMA
systems with iterative joint multi-user decoding,” IEEE Trans.
Wireless Commun., vol. 8, no. 3, pp. 1148–1153, Mar. 2009.
[34] Y. Zhang, K. Peng, and J. Song, “Enhanced IDMA with rate-
compatible raptor-like quasi-cyclic LDPC code for 5G,” in Proc.
IEEE Globecom Workshops (GC Wkshps), Dec. 2017, pp. 1–6.
[35] Transmitter Side Signal Processing Schemes for NOMA, docu-
ment R1-1804396 3GPP TSG RAN WG1 Meeting # 92, Sam-
sung, Sanya, China, Apr. 2018.
[36] LLS Results for RDMA, GOCA, RSMA and IDMA Schemes, docu-
ment R1-167536 3GPP TSG RAN WG1 Meeting # 86, MediaTek
Inc., Gothenburg, Sweden, Aug. 2016.
[37] Y. Chen and T. Wild, “Exploring performance and complexity
of selected NOMA candidates in 5G new radio,” in Proc. IEEE
Globecom Workshops (GC Wkshps), Dec. 2018, pp. 1–6.
[38] C. Novak, G. Matz, and F. Hlawatsch, “IDMA for the multiuser
MIMO-OFDM uplink: A factor graph framework for joint data
detection and channel estimation,” IEEE Trans. Signal Process.,
vol. 61, no. 16, pp. 4051–4066, Aug. 2013.
[39] T. Peng, Y. Xiao, X. He, and S. Li, “Improved detection of
uplink OFDM-IDMA signals with carrier frequency offsets,”
IEEE Commun. Lett., vol. 16, no. 5, pp. 646–649, May 2012.
[40] F. Lenkeit, C. Bockelmann, D. Wübben, and A. Dekorsy, “IRA
code design for IDMA-based multi-pair bidirectional relaying
systems,” in Proc. IEEE Globecom Workshops (GC Wkshps),
Dec. 2013, pp. 241–246.
[41] D. Guo, S. Shamai, and S. Verdu, “Mutual information and
minimum mean-square error in Gaussian channels,” IEEE Trans.
Inform. Theory, vol. 51, no. 4, pp. 1261–1282, Apr. 2005.
[42] X. Yuan, L. Ping, C. Xu, and A. Kavcic, “Achievable rates
of MIMO systems with linear precoding and iterative LMMSE
detection,” IEEE Trans. Inform. Theory, vol. 60, no. 11, pp.
7073–7089, Nov. 2014.
[43] P. A. Hoeher and T. Wo, “Superposition modulation: myths and
facts,” IEEE Commun. Mag., vol. 49, no. 12, pp. 110–116, Dec.
2011.
1536-1276 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Uppsala Universitetsbibliotek. Downloaded on March 11,2020 at 04:07:01 UTC from IEEE Xplore. Restrictions apply.
16
[44] S. ten Brink, “Convergence behavior of iteratively decoded
parallel concatenated codes,” IEEE Trans. Commun., vol. 49,
no. 10, pp. 1727–1737, Oct. 2001.
[45] L. Ping, J. Tong, X. Yuan, and Q. Guo, “Superposition coded
modulation and iterative linear MMSE detection,” IEEE J. Sel.
Areas Commun., vol. 27, no. 6, pp. 995–1004, Aug. 2009.
[46] T. Yang and J. Yuan, “Performance of iterative decoding for
superposition modulation-based cooperative transmission,” IEEE
Trans. Wireless Commun., vol. 9, no. 1, pp. 51–59, Jan. 2010.
[47] P. A. Hoeher, H. Schoeneich, and J. C. Fricke, “Multi-layer
interleave division multiple access: Theory and practice,” Eur.
Trans. Telecommun., vol. 19, no. 5, p. 523–536, Aug. 2008.
[48] S. ten Brink, G. Kramer, and A. Ashikhmin, “Design of low-
density parity-check codes for modulation and detection,” IEEE
Trans. Commun., vol. 52, no. 4, pp. 670–678, Apr. 2004.
[49] D. Guo, Y. Wu, S. S. Shitz, and S. Verdu, “Estimation in Gaussian
noise: Properties of the minimum mean-square error,” IEEE
Trans. Inform. Theory, vol. 57, no. 4, pp. 2371–2385, Apr. 2011.
[50] Sae-Young Chung, G. D. Forney, T. J. Richardson, and R. Ur-
banke, “On the design of low-density parity-check codes within
0.0045 dB of the Shannon limit,” IEEE Commun. Lett., vol. 5,
no. 2, pp. 58–60, Feb. 2001.
[51] X. Yuan, Q. Guo, X. Wang, and L. Ping, “Evolution analysis
of low-cost iterative equalization in coded linear systems with
cyclic prefixes,” IEEE J. Sel. Areas Commun., vol. 26, no. 2, pp.
301–310, Feb. 2008.
[52] S. Sharifi, A. K. Tanc, and T. M. Duman, “LDPC code design
for the two-user Gaussian multiple access channel,” IEEE Trans.
Wireless Commun., vol. 15, no. 4, pp. 2833–2844, Apr. 2016.
[53] A. Balatsoukas-Stimming and A. P. Liavas, “Design of LDPC
codes for the unequal power two-user Gaussian multiple access
channel,” IEEE Wireless Commun. Lett., vol. 7, no. 5, pp. 868–
871, Oct. 2018.
[54] Y. Chi, L. Liu, G. Song, C. Yuen, Y. L. Guan, and Y. Li, “Prac-
tical MIMO-NOMA: Low complexity and capacity-approaching
solution,” IEEE Trans. Wireless Commun., vol. 17, no. 9, pp.
6251–6264, Sep. 2018.
[55] G. Song and J. Cheng, “Distance enumerator analysis for
interleave-division multi-user codes,” IEEE Trans. Inform. The-
ory, vol. 62, no. 7, pp. 4039–4053, Jul. 2016.
[56] X. Yuan, Low-complexity iterative detection in coded linear
systems, Ph.D. dissertation, City University of Hong Kong, Hong
Kong, China, 2008.
[57] R. Y. S. Tee, R. G. Maunder, and L. Hanzo, “EXIT-chart aided
near-capacity irregular bit-interleaved coded modulation design,”
IEEE Trans. Wireless Commun., vol. 8, no. 1, pp. 32–37, Jan.
2009.
[58] S.-Y. Chung, T. Richardson, and R. Urbanke, “Analysis of sum-
product decoding of low-density parity-check codes using a
Gaussian approximation,” IEEE Trans. Inform. Theory, vol. 47,
no. 2, pp. 657–670, Feb. 2001.
[59] J. A. McGowan and R. C. Williamson, “Loop removal from
LDPC codes,” in Proc. IEEE Inform. Theory Workshop (ITW’03),
Apr. 2003, pp. 230–233.
[60] Xiao-Yu Hu, E. Eleftheriou, and D. M. Arnold, “Regular and
irregular progressive edge-growth tanner graphs,” IEEE Trans.
Inform. Theory, vol. 51, no. 1, pp. 386–398, Jan. 2005.
[61] G. Liva, S. Song, L. Lan, Y. Zhang, S. Lin, and W. E. Ryan,
“Design of LDPC codes: A survey and new results,” J. Commun.
Software Syst., vol. 2, no. 3, p. 191–211, Sep. 2006.
[62] I. Andriyanova and J. P. Tillich, “Designing a good low-rate
sparse-graph code,” IEEE Trans. Commun., vol. 60, no. 11, pp.
3181–3190, Nov. 2012.
[63] G. Yue, L. Ping, and X. Wang, “Generalized low-density parity-
check codes based on Hadamard constraints,” IEEE Trans. In-
form. Theory, vol. 53, no. 3, pp. 1058–1079, Mar. 2007.