476
2, APRIL 2019
Interleave-Division Multiple Access in High Rate Applications
Yang Hu , Chulong Liang , Lei Liu , Chunlin Yan, Yifei Yuan, and Li Ping, Fellow, IEEE
Abstract—Interleave-division multiple access (IDMA) is a
multiple access scheme that has been considered in several recent
proposals for the 5th generation cellular system. In this letter,
basing on evolution analysis, we show that the performance of
IDMA can be enhanced using the transfer function matching
principle. Such matching can be realized by superposition coded
modulation, power control, repetition coding, and zero padding.
Zero padding together with cyclic shifting also leads to reduced
implementation complexity. Our analysis is based on additive
white Gaussian noise channels and we show by simulations that
the matching techniques can also provide impressive performance
in fading channels.
Index Terms—IDMA, evolution technique, system design.
I. I NTRODUCTION
NTERLEAVE-DIVISION multiple access (IDMA) [1] is
I inspired by the success of low-density parity-check (LDPC)
codes [2]. Recently, IDMA has been discussed for the 5th
generation (5G) cellular system [3]–[5].
For LDPC codes, decoding performance can be optimized
by matching the transfer functions of local decoders [2]. This
matching principle was later extended to different iterative
systems [6]–[8]. An IDMA receiver also involves two local
processors named as, respectively, an elementary signal esti-
mator (ESE) and a decoder (DEC). (See Section II.) As
shown in [9], the performance of IDMA can be improved
by tuning an underlying LDPC code for better matching
between ESE and DEC. There are, however, some obstacles
for this strategy. First, in 5G, the LDPC code used has already
been specified [10] so other alternatives, instead of alter-
ing code structure, should be used for system optimization.
Second, there is also a lack of efficient matching method when
high order modulation is involved for high rate applications.
Third, matching for multiuser systems is generally a difficult
problem. Very limited progress is made in this direction.
In this letter, we consider IDMA system design in high
sum-rate situations. We first derive the achievable rate for
IDMA using the matching principle. We show that, with
perfect matching, IDMA is potentially capacity approaching in
Manuscript received July 19, 2018; revised October 5, 2018; accepted
October 9, 2018. Date of publication October 17, 2018; date of current version
April 9, 2019. This work was supported in part by the Research Fund of ZTE
Corporation, and in part by the University Grants Committee of the Hong
Kong Special Administrative Region, China, under Project CityU 11280216
and Project CityU 11216817. The associate editor coordinating the review of
this paper and approving it for publication was W. Zhang. (Corresponding
author: Lei Liu.)
Y. Hu, C. Liang, L. Liu, and L. Ping are with the Department of
Electronic Engineering, City University of Hong Kong, Hong Kong (e-mail:
yhu228-c@my.cityu.edu.hk; chuliang@cityu.edu.hk; leiliu@cityu.edu.hk;
eeliping@cityu.edu.hk).
C. Yan and Y. Yuan are with the Algorithm Department, ZTE
Corporation, Shenzhen 518057, China (e-mail: yan.chunlin@zte.com.cn;
yifei.yuan@ztetx.com).
Digital Object Identifier 10.1109/LWC.2018.2876538
2162-2345 c 2018 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.
IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO.
nonfading additive white Gaussian noise (AWGN) channels.
We outline several practical matching techniques including
modulation, power control, repetition coding and zero padding.
Incidentally, we also show that zero padding together with
cyclic shifting can reduce the implementation cost related to
user-specific interleaving in IDMA. Our analysis is based on
AWGN channels and we will provide experimental results for
fading channels. We will show that the proposed techniques
can provide noticeable performance enhancement.
II. S YSTEM M ODEL AND E VOLUTION A NALYSIS
A. Transmitter Principles
Consider a K-user up-link multiple access system with
received symbols:
y(j ) =
K

hk xk (j ) + η(j ), j = 1, 2, . . . , J , (1)
k =1
where hk is the channel coefficient of user k, xk (j ) a transmit-
ted symbol, η(j ) a complex AWGN sample with mean zero
and variance σ 2 , and J the frame length. We assume an under-
lying orthogonal frequency division multiplexing (OFDM)
layer that resolves the intersymbol interference problem and a
quasi-static channel that remains unchanged over a frame.
The principle of IDMA is illustrated graphically in Fig. 1.
The graph is randomized with user-specific interleaving, which
is illustrated in Fig. 1 by the shuffled edge connections
between {ck (j )} and {xk (j )}. Fig. 1 can be seen as a graphic
extension of a single-user LDPC code to a multiuser system.
The randomness resulting from interleaving reduces short
cycles in the graph, which facilitates low-cost message passing
decoding. More details can be found in [1] and [11].
Fig. 1 involves a user-specific interleaver for each user. This
interleaver can be combined with the inherent interleaver in
the LDPC code involved. This is equivalent to the scheme
in [12], in which each user employs a unique interleaver for
its code. Later we will show that such an interleaver can be
realized by cyclic shifting (see Fig. 5 below), which further
reduces the hardware implementation cost for IDMA.
B. Receiver Principles
We divide an iterative detector for the system in Fig. 1 into
two local processors: an ESE and a DEC. The iterative process
is outlined below.
Initialization: Assume that the modulation constellation of
{xk (j )} is with zero mean and unit average power. Then
E(xk (j )) and Var(xk (j )) are respectively initialized to 0 and
1, ∀k , j .
ESE Operations: We rewrite (1) as
y(j ) = hk xk (j ) + ζk (j ), (2a)
HU et al.: IDMA IN HIGH RATE APPLICATIONS 477

Fig. 1. A factor graph of a 2-user IDMA system with LDPC coding. J = 8. {ck (j ), j = 1, 2, . . . , J } is a codeword that is interleaved and modulated to
produce {xk (j )}. Circles represent variables and squares constraints. Three types of constraints are presented: a white square for an LDPC coding constraint,
a square with “×” for a modulation constraint and a square with “+” for a multiple access constraint defined in (1).

where If the system is symmetric for all k, we can drop index k

K
and use two common functions to characterize the behaviors
 of all users as:
ζk (j ) = y(j ) − hk xk (j ) = hk  xk  (j ) + η(j ) (2b)
k  =1,k  =k ESE: snr = φ(v ), (6a)
includes the interference and noise seen by user k. We model DEC: v = ψ(snr ). (6b)
ζk (j ) ∼ CN (μk (j ), vζ,k (j )) using Gaussian approximation In this case, the final performance is determined by the first
(GA). Given a priori {E(xk (j ))} and {Var(xk (j ))}, we can fixed point of φ and ψ as shown in Fig. 2.
evaluate μk (j ) and vζ,k (j ) and then estimate xk (j ) based
on (2a). For example, for binary phase shift keying (BPSK) III. IDMA D ESIGN T ECHNIQUES BASED ON E VOLUTION
modulation, each xk (j ) ∈ {−1, +1} and the estimation A NALYSIS
outputs are log-likelihood ratios (LLRs):
A. Matching Principle
Pr(xk (j ) = +1) Re(2hk∗ (y(j ) − µk (j ))) For an LDPC or a turbo code, it is known that the over-
LLR(xk (j )) = = . (3)
Pr(xk (j ) = −1) vζ,k (j )/2 all performance can be optimized by matching the transfer
Similar results can be obtained for quadrature phase shift functions of the two local decoders [2], [14]. Following
keying (QPSK) and other modulations. this principle, a symmetric IDMA system can be optimized
DEC Operations: The DEC is further divided into K con- by matching φ and ψ in (6). The matching condition is
stituent decoders {DEC 1, DEC 2, · · · DEC K}, one for given by
each user. The LLRs in (3) are used as the inputs to the ψ(z ) = φ−1 (z ), (7)
DEC. Assume soft-output decoding, by which {E(xk (j ))} and
{Var(xk (j ))} are updated. For an LDPC or a turbo code, such where φ−1 is the reverse function of φ. The analytical treat-
decoding follows the standard procedures in [2], [13], and [14]. ments on general situations are beyond the scope of this letter.
Iterative Process: The DEC outputs {E(xk (j ))} and We will only provide derivation for AWGN channels and rely
{Var(xk (j ))} are fed back to the ESE. Then (3) is re-evaluated on simulation results for fading channels below.
and the iterative process continues.
B. Optimality of IDMA in AWGN Channels
C. Evolution Analysis Consider an AWGN channel in which hk = 1 in (1) for all
users. Assume the same code for all users. In this case, (6a)
Let vk be the average of Var(xk (j )) over j and snrk the
is given by
average signal-to-noise-ratio (SNR) related to {LLR(xk (j ))}
in (3). From [1], snrk can be expressed as a function: snr = φ(v ) = 1/((K − 1)v + σ 2 ). (8)
 2
 Following the minimum mean square error (MMSE)-SNR
snrk = E |xk (j )| /E(vζ,k (j )) ≡ φk (v1 , . . . .vk −1 , vk +1 , . . . , vK ). (4)
relationship developed in [6], [8], and [15], the achievable rate
Now let the input-output relationship of DEC k be charac- of each user is given by
terized by a function vk = ψk (snrk ) that can be generated  +∞
numerically. The behavior of the iterative process can be char- R= mmse(snr )dsnr , (9)
acterized by the following recursions (initialized to vk = 1): 0
where mmse is the MMSE at the output of DEC k with snr at
its input. Then, following [8], we have
ESE: snrk = φk (v1 , . . . .vk −1 , vk +1 , . . . , vK ), (5a)  −1
DEC: vk = ψk (snrk ). (5b) mmse(snr ) = snr + (ψ(snr ))−1 , (10)
478
Fig. 2. An example of evolution trajectory.
which includes the contributions of both a priori information
(related to snr) and extrinsic one (related to ψ(snr )). From (7)
and (8), we have

ψ(snr ) = 1, 0 ≤ snr < 1/(K − 1 + σ 2 ),
ψ(snr ) = (K1−σ snr , 1/(K − 1 + σ 2 ) ≤ snr ≤ 1/σ 2 . (11)
2
−1)snr
Here the first equation in (11) results from the constraint
Var(xk (j )) ≤ 1 so that ψ(snr ) ≤ 1.
Substituting (10) and (11) into (9) and with some straight-
forward manipulations, we have the achievable rate per user
R = K −1 log(1 + K /σ 2 ). (12)
For K users, the sum-rate is
2
Rsum = KR = log(1 + K /σ ), (13)
which achieves the K-user channel capacity. This demonstrates
the optimality of IDMA under perfect matching.
Equation (13) extends the conclusions in [8] from single-
user systems to multiuser ones in AWGN channels. With
practical coding constraints, it can be difficult to design per-
fectly matched φ and ψ. Also, the situation is much more
complicated if the received powers are random variables due
to fading. Detailed analytical discussions on the matching prin-
ciple for fading channels are beyond the scope of this letter.
Nevertheless, the matching condition in (7) suggests a way
for performance optimization. Below, we will discuss some
empirical results along this direction.
C. Design Strategies Based on the Matching Principle
The followings are some strategies to shape either φ or ψ,
so as to improve matching.
1) A standard approach is to shape ψ by adjusting the degree
polynomials for an LDPC code [2], [9]. This method
requires different code structures for different situations,
which may cause difficulty in hardware implementation.
In this letter, we will examine options with fixed codes.
2) Properly designed modulation and labeling methods can
be used to shape ψ in the high rate region. In partic-
ular, with superposition coded modulation (SCM), two
or more layers of modulated codes (with, say, QPSK
modulation) are scaled to different power levels and then
linearly superimposed, as demonstrated in [16].
3) Other simple options include zero padding (ZP) and rep-
etition coding. With ZP, some users are silent on some
IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Fig. 3. Evolution functions for the 3GPP NR LDPC coded IDMA system.
σ 2 is set to −9dB. Power levels of the two SCM layers are 0.2 and 0.8 for
ψ (2) .
symbols. Then (8) becomes
snr = φ(v ) = 1/((K  − 1)v + σ 2 ), (14)
where K is the nonzero symbols transmitted simultane-
ously (assume that ratio of zero symbols is uniform over
all resources). ZP can shape φ to a certain extend and
can also lead to reduced multiuser detection complexity.
Repetition coding does not provide coding gain. However,
it can be useful in multiuser systems to shape ψ.
IV. N UMERICAL R ESULTS
We follow the 3GPP NR requirement [17] of frame length
of 864 symbols with information bits Jinfo = 616 per user. We
consider AWGN first. Fig. 3 shows the φ and ψ functions in
AWGN channels. φ is the ESE function of K = 4. ψ (1) is the
DEC function of the 3GPP NR LDPC code and QPSK modu-
lation. We can clearly see poor matching between ψ (1) and φ.
ψ (2) is the DEC function of a two-layer SCM scheme with the
3GPP NR LDPC code followed by a rate-1/4 repetition cod-
ing. (The power levels of the two SCM layers are obtained by
exhaustive searching.) We can see a staircase shape of ψ (2) ,
which is due to the different convergence behaviors of differ-
ent SCM layers with different powers: the higher-power layer
leads the convergence in the first stage (see Fig. 3) and then the
lower-power layer follows in the second stage. Clearly, such
staircase shape improves the matching between ψ (2) and φ.
Fig. 4 shows the frame error rates (FERs) of different
schemes in a 4-user AWGN system. Schemes 1 and 2 cor-
respond to ψ (1) and ψ (2) in Fig. 3 respectively. We can see
that scheme 1 does not work at all, while scheme 2 works well
due to better matching. Let M be the number of SCM layers.
Although increasing M may improve matching in theory, there
is a limit in practice since the value of M will influence the
coding gain of each layer under a given Jinfo . We observed
that, for Jinfo = 616, M = 2 represents a best compromise.
We now consider fading channels, in which the fluctua-
tions of channel gains provide more diversity to support more
users. ZP can be used to reduce receiver complexity. A cyclic
shifted ZP (CSZP) scheme is shown in Fig. 5. Each user is
active only in half of a frame (i.e., with 50% ZP) and the
active half of a frame for user k is cyclically shifted by an
amount of (k −1)J /K  where J is the frame length. Assume
HU et al.: IDMA IN HIGH RATE APPLICATIONS
Fig. 4. IDMA in AWGN channels. K = 4. SNR per user = 1/σ 2 . Sum-rate
= KJinfo /864 = 2.85.
Fig. 5. An example of CSZP. K = 4. Two SCM layers are used.
Fig. 6. IDMA vs. SCMA in TDL-C channels. 3GPP NR LDPC code is used
for both IDMA and SCMA. IDMA involves SCM (with power levels of 0.865
and 0.135 for both cases) and CSZP. performances of SCMA are from [17].
that a random interleaver is used in the underlying LDPC
code. The user-specific shifting in Fig. 5 provides an equiv-
alent realization for the user-specific interleaving in Fig. 1,
since a shifted version of a random interleaver can be seen
as a different random interleaver [11]. Fig. 5 also involves a
two-layer SCM scheme. To distinguish the two layers of each
user, each layer is further independently and cyclically shifted
within its active frame duration. Such shifting technique, both
user-specific as well layer-specific, significantly reduces the
hardware implementing cost for interleavers in IDMA.
Fig. 6 shows the FERs with SCM and CSZP in the TDL-
C fading channels [17]. Two antennas are assumed at the
receiver. We consider the following two settings:
• Case 1: Jinfo = 616, K = 10 and sum-rate = 7.13;
• Case 2: Jinfo = 496, K = 12 and sum-rate = 6.89.
The power levels of the two SCM layers given in the caption
of Fig. 6 are obtained by exhaustive searching. Iterative linear
MMSE detection [11] is used at the receiver. As reference,
479
Fig. 6 also includes the performances of sparse code multiple
access (SCMA) [17]. The sparse codebook in SCMA is an
equivalent form of ZP in IDMA. The performance advantage
of IDMA over SCMA mainly attributes to the freedom of
curve shaping using SCM.
In conclusion, in this letter we examined several design
techniques for IDMA following the transfer function match-
ing principle. We showed that IDMA is capacity approach-
ing in AWGN channels under perfect matching. We also
showed by simulations that the matching techniques for
IDMA can provide impressive performance in fading chan-
nels. Some of the softwares used in this letter are available
in the following site: http://www.ee.cityu.edu.hk/%7Eliping/
Research/Simulationpackage/.
R EFERENCES
[1] L. Ping, L. Liu, K. Wu, and W. K. Leung, “Interleave division multiple-
access,” IEEE Trans. Wireless Commun., vol. 5, no. 4, pp. 938–947,
Apr. 2006.
[2] T. Richardson and R. Urbanke, Modern Coding Theory. Cambridge,
U.K.: Cambridge Univ. Press, 2008.
[3] 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
[4] On the Performance of IDMA and RSMA Multiple Access Schemes,
document R1-167561, 3GPP TSG RAN WG1 Meeting #86, InterDigit.
Commun., Gothenburg, Sweden, Aug. 2016.
[5] C. Yan, Z. Yuan, W. Li, and Y. Yuan, “Non-orthogonal multiple
access schemes for 5G,” ZTE Commun., vol. 14, no. 4, pp. 11–16,
Oct. 2016.
[6] K. Bhattad and K. R. Narayanan, “An MSE-based transfer chart
for analyzing iterative decoding schemes using a Gaussian approx-
imation,” IEEE Trans. Inf. Theory, vol. 53, no. 1, pp. 22–38,
Jan. 2007.
[7] K. Wu, K. Anwar, and T. Matsumoto, “BICM-ID-based IDMA using
extended mapping,” IEICE Trans. Commun., vols. E97–B, no. 7,
pp. 1483–1492, Jul. 2014.
[8] X. Yuan, L. Ping, C. Xu, and A. Kavcic, “Achievable rates of
MIMO systems with linear precoding and iterative LMMSE detec-
tion,” IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7073–7089,
Nov. 2014.
[9] Y. Zhang, K. Peng, and J. Song, “Enhanced IDMA with rate-compatible
raptor-like quasi-cyclic LDPC code for 5G,” in Proc. IEEE GC Wkshps,
Singapore, Dec. 2017, pp. 1–6.
[10] Technical Specification Group Radio Access Network; NR; Multiplexing
and Channel Coding (Release 15), V2.0.0, 3GPP Standard TS 38.212,
Dec. 2017.
[11] Y. Hu, C. Liang, J. Hu, and L. Ping, “Low-cost implementation tech-
niques for interleave division multiple access,” IEEE Wirelee Commun.
Lett., to be published. [Online]. Available: https://ieeexplore.ieee.org/
abstract/document/8398463
[12] R. Zhang, L. Xu, S. Chen, and L. Hanzo, “Repeat accumulate code
division multiple access and its hybrid detection,” in Proc. IEEE ICC,
Beijing, China, May 2008, pp. 4790–4794.
[13] C. Berrou, A. Glavieux, and P. Thitimajshima, “Near Shannon
limit error-correcting coding and decoding: Turbo-codes. 1,”
in Proc. IEEE ICC, vol. 2, Geneva, Switzerland, May 1993,
pp. 1064–1070.
[14] S. T. Brink, “Convergence behavior of iteratively decoded parallel con-
catenated codes,” IEEE Trans. Commun, vol. 49, no. 10, pp. 1727–1737,
Oct. 2001.
[15] D. Guo, S. Shamai, and S. Verd, “Mutual information and minimum
mean-square error in Gaussian channels,” IEEE Trans. Inf. Theory,
vol. 51, no. 4, pp. 1261–1282, Apr. 2005.
[16] 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.
[17] Preliminary LLS Evaluation Results for mMTC Scenario, docu-
ment R1-1803665, 3GPP TSG RAN WG1 Meeting #92bis, Huawei,
HiSilicon, Shenzhen, China, Apr. 2018.