Complexity Reduction of MPA Detection

Using Joint IQ Factor Graph in SCMA

Hiroto Hirama and Takahiko Saba
Dept. of Computer Science, Chiba Institute of Technology
2-17-1 Tsudanuma, Narashino, Chiba 275-0016 JAPAN

Abstract—Sparse code multiple access (SCMA) is a non- are updated by propagating the probabilities on the factor
orthogonal multiple access scheme. In a downlink SCMA system, graph. These processes are iterated so that the likelihood
a message passing algorithm (MPA) is employed for multiuser of the SCMA codeword converges sufficiently. Thereafter, a
detection. However, MPA detection has a problem of high
computational complexity in downlink environment. In this study, posteriori probability of the SCMA codeword is obtained to
we propose a method to reduce the computational complex- decode symbols. Although the MPA detection can efficiently
ity of MPA detection while maintaining the transmission rate calculate a posteriori probability of the SCMA codeword,
and the number of simultaneously connectable users. First, we the computational complexity increases exponentially as the
employ a codebook including symbols having only in-phase (I) number of superposed users df and the codebook size M
or quadrature-phase (Q) component. Next, MPA detection is
performed using a joint IQ factor graph in which real and increase. This is because there are M df combinations of
imaginary parts of a symbol are treated separately as different SCMA codewords. In downlink wireless communications, less
subcarrier symbols. As a result, the number of superposed users complex detection is required because the capacity of a battery
in MPA detection is virtually reduced. Therefore, the proposed in a mobile terminal is limited. However, the reduction in
method can reduce the computational complexity. Numerical M and df decreases the transmission rate and the number of
results show that the proposed method does not degrade the bit
error rate performance compared to MPA using the conventional simultaneously connectable users, respectively. In [4], MPA is
factor graph. calculated in the logarithm domain to reduce the computational
complexity. Partially marginalization based MPA is a method
I. I NTRODUCTION to partially update a posteriori probability of the SCMA
Multiple access is a technique for enabling multiple users codeword [5]. In [6], [7], the number of MPA iterations is
to access a base station simultaneously by assigning resources reduced by letting the likelihood of the SCMA codeword
such as time and frequency to the users. In the long term converge rapidly. In these studies, the likelihood of SCMA
evolution (LTE), a multiple access scheme called orthogonal codewords of M df combinations must be calculated.
frequency division multiple access (OFDMA) is used [1]. In
OFDMA, since subcarriers are orthogonal to one another and In this study, we propose a method of virtually reducing
each subcarrier is assigned to one user, multiple users can the number of superposed users df in MPA detection by
connect to the base station simultaneously without causing treating in-phase (I) and quadrature-phase (Q) components
interference between users. On the other hand, sparse code of the received signal (real and imaginary parts of the re-
multiple access (SCMA) is a multiple access scheme in which ceived complex symbol) as signals on different subcarriers.
each subcarrier is simultaneously assigned to multiple users To achieve this, we employ a codebook including symbols
[2]. Therefore, the use of SCMA can increase the number of having only I or Q component. Thus, when the received signal
simultaneously connectable users. is handled separately by the I and Q components, interference
In downlink SCMA, base stations and users have com- from a signal in an orthogonal component which is transmitted
mon codebooks. Based on the codebooks, the base station by another user is negligible. As a result, the number of
selects SCMA codewords corresponding to users’ symbols. An superposed users in MPA detection is virtually reduced.
SCMA codeword contains information on subcarriers assigned
to a user and their signal constellation points, which are non- At a receiver, MPA detection is performed using a factor
zero elements of the SCMA codeword. Since each subcarrier is graph that treats the real part and the imaginary part of
simultaneously assigned to multiple users, each user receives the received symbol as different subcarriers. We call this
a signal on which the signals of other users are superposed. factor graph the joint IQ factor graph. By adopting the joint
Thus, users need to obtain their desired symbols from their IQ factor graph, it is possible to reduce the computational
received signals using multiuser detection. complexity of MPA detection without degrading the detection
In SCMA systems, the message passing algorithm (MPA) performance as compared to the conventional MPA. Computer
is used for multiuser detection [3]. MPA is an algoritm that simulation was conducted to verify the effectiveness of the
efficiently calculates a posteriori probability of the SCMA proposed method, and it is shown that the proposed method
codeword by using a bipartite graph called a factor graph. The can reduce the computational complexity of MPA detection
likelihood and a priori probability of the SCMA codeword without degrading the detection performance.

978-1-7281-5207-3/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 26,2020 at 13:43:18 UTC from IEEE Xplore. Restrictions apply.
 codebook
II. SCMA codebook
for user 1
codebook
for user 2 for user 6

SCMA is a non-orthogonal multiple access scheme in which

signals are not orthogonal among users [8]. In SCMA sys-
tems, although the subcarriers are orthogonal, one subcarrier
···
is simultaneously assigned to multiple users. Therefore, the b1 (00) (01) (10)(11) b2 (00)(01) (10) (11) b6 (00)(01) (10)(11)
orthogonality of signals between users is collapsed. However,
the number of users that can simultaneously connect to a base
station can be increased as compared to OFDMA in which f1
f2
one subcarrier is assigned to a single user. + + · · ·+ = f3
In SCMA systems, symbols to be transmitted are converted f4
to SCMA codewords based on the codebook. The codebook SCMA SCMA
codeword 1 codeword 2
SCMA
codeword 6
transmitted
SCMA codeword
describes SCMA codewords, each of which indicates the sub-
carriers to be assigned to a user and their signal constellation
points. The SCMA codeword is a sparse codeword with many Fig. 1. SCMA encoding.
zero elements, and the positions of the non-zero elements
correspond to the subcarriers assigned to the user. At the where hj is the channel response between the base station and
SCMA base station, SCMA codewords of multiple users are the user j, n is the additive white Gaussian noise, and diag(h)
superposed and transmitted. is a diagonal matrix with elements of h as diagonal elements.
At the receiver, the received signal contains multiple SCMA Furthermore, the received signal in each subcarrier can be
codewords. However, since all users have common codebooks, expressed by Eq. (2) when the set of user indices superposed
each user can detect the desired symbols from the received on the subcarrier k is ξk = {j|xj,m,k = 0}.
signal by multiuser detection. In SCMA, MPA detection is 
used as multiuser detection. Hereinafter, the downlink SCMA yj,k = hj,k xj,m,k + n. (2)
system will be described in more detail. ξk

A. SCMA encoding It can be seen from Eqs. (1) and (2) that signals from users
other than the user j are superposed on the received signal.
Fig. 1 shows the generation of SCMA codewords in the Therefore, each user needs to perform multiuser detection for
case where the number of subcarriers K = 4, the number demodulating the desired symbol.
of users J = 6, the codebook size M = 4, and the
number of subcarriers with non-zero elements N = 2. As III. MPA DETECTION
shown in Fig. 1, a codebook is assigned to each user and For multiuser detection in SCMA, maximum a posteriori
an SCMA codeword corresponding  to a symbol is T selected (MAP) estimation is presented as the maximum likelihood
for each user. The symbol bj = bj,1 , . . . bj,log2 M for the detection method of SCMA codewords [10]. However, the use
user j ∈ {1, . . . , J} is converted into the SCMA codeword of MAP estimation is not practical because the computational
xj = [xj,1 , . . . , xj,K ] based on the codebook. The elements complexity required for detection is large. Therefore, MPA
of the SCMA codeword for the user j on the subcarrier detection, which has less computational complexity than MAP
k ∈ {1, . . . , K} is represented by xj,k = [xj,1,k , . . . , xj,M,k ], estimation, is used.
where xj,m,k (m ∈ {1, . . . , M }) is the constellation point of
the user j on the subcarrier. SCMA codewords of every user A. MPA
are superposed and transmitted. MPA is a suboptimal detection scheme that requires less
computational complexity compared to MAP. MPA calculates
B. Downlink SCMA system model
a posteriori probability on a bipartite graph called a factor
In a downlink SCMA system, the base station converts graph. The factor graph consists of function nodes (FNs) cor-
symbols of users into SCMA codewords. Then, SCMA code- responding to subcarriers, variable nodes (VNs) corresponding
words are converted from the frequency domain to the time to users, and edges, which connect FNs and VNs. In MPA, the
domain signal by the inverse fast Fourier transform (IFFT). probability called the message is propagated through edges on
After inserting the cyclic prefix (CP), the resultant signal is the factor graph, and a posteriori probability of the SCMA
transmitted from the base station [9]. codeword is calculated after iterative processing.
At the receiver, CP is first removed from the received signal. Here, we consider MPA detection in the case where the
Next, FFT is performed to convert the received signal from the number of users J = 6, the number of subcarriers K = 4,
time domain to the frequency domain signal. The frequency the number of non-zero elements N = 2, and the number
domain received signal yj of the user j is expressed by of superposed users df = 3. The factor graph in this case is
shown in Fig. 2. In Fig. 2, square nodes correspond to FNs,

J
circle nodes correspond to VNs, k is the index of the subcarrier
yj = diag(hj ) xj,m + n, (1)
J=1
assigned to the SCMA codeword, and j is the index of the user.

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 26,2020 at 13:43:18 UTC from IEEE Xplore. Restrictions apply.
 FN-1 FN-2 FN-3 FN-4 calculations. However, for simplicity, the number of complex
divisions is counted as that of complex multiplications and the
number of complex subtractions is counted as that of complex
additions.
VN-1 VN-2 VN-3 VN-4 VN-5 VN-6
First, based on Eq. (4), we evaluate the computational
complexity required to update the likelihood of an SCMA
FN-k
Fig. 2. Factor graph. codeword at FN-k. The number of multiplications Nmul
in Eq. (4) can be expressed as Eq. (7), and the number of
FN-k
additions Nadd can be expressed as Eq. (8).
The likelihood of each SCMA codeword is first calculated
FN-k
from the received signal in MPA. The likelihood of the SCMA Nmul = IT df M df (5df + 5), (7)
codeword can be obtained by FN-k
Nadd = IT df M df (4df + 2). (8)
 ⎛ ⎞
 
 Next, based on Eq. (5), we evaluate the computational
−1   
f (xj,m,k ) = 2 yk − ⎝hk xj,m,k + hk xl,m,k ⎠ , complexity required to update a priori probability at VN-
VN-j
σ   j. The number of multiplications Nmul in Eq. (5) can be
l∈ξk \j
2
(3) expressed as
where σ 2 is the noise power.
N VN-j = I M d (d − 1),
mul T v v (9)
Next, we update the likelihood and a priori probability of
the SCMA codeword by propagating the message through where dv is the number of edges from VN-j to FNs. As shown
edges of the factor graph. The likelihood of the SCMA in Eqs. (7) to (9), the computational complexity required for
codeword is updated by propagating the message from FN- MPA detection is mostly dominated by the calculations at FNs.
k to VN-j. The message rk→j (xj,m,k ) from FN-k to VN-j
can be expressed by C. SMPA
Serial MPA (SMPA) is a method to reduce the computa-

M
rk→j (xj,m,k ) = f (xj,m,k ) ql→k (l, m, k), (4) tional complexity of MPA detection [6]. SMPA can reduce
m=1 l∈ξk \j
the number of MPA iterations by letting the likelihood of the
SCMA codeword converge rapidly. In the conventional MPA,
where ql→k (l, m, k) represents the message from VN-l to FN- a priori probability is updated at each VN after updating
k. the likelihood of SCMA codewords at all FNs. In SMPA,
A priori probability is updated by propagating the message on the other hand, the message is sent from FN-k to VN-j
from VN-j to FN-k. The message from VN-j to FN-k can be immediately after the likelihood of the SCMA codeword at
expressed by FN-k is updated, and a priori probability of VN-j is updated.
Therefore, new messages updated by other FNs can be used
qj→k (xj,m,k ) = rl→j (xj,m,l ), (5)
immediately in updating the likelihood of SCMA codewords.
l∈ζj \k
As a result, the likelihood of an SCMA codeword at an FN
where ζj is the set of subcarrier indices on which non-zero is reflected to other FNs earlier than in the case of using the
elements of the user j are transmitted. Eqs. (4) and (5) are conventional MPA, and the likelihood of SCMA codewords
iterated IT times. converges rapidly.
Finally, a posteriori probability of the SCMA codeword In terms of the computational complexity, since SMPA is
is determined based on a priori probability of the SCMA a method to reduce the number of iterations by changing the
codeword determined by the iterative processing. A posteriori order of message propagation, the computational complexities
probability of the SCMA codeword can be expressed by at FN and VN are the same as those in the conventional
MPA. Therefore, it is possible to evaluate the computational
P (xj,m ) = Ap ql→j (xj,m,l ), (6)
complexity of SMPA using Eqs. (7), (8), and (9).
l∈ζj

where Ap is the initial value of a priori probability 1/M . Each IV. R EDUCTION IN COMPUTATIONAL COMPLEXITY OF
user calculates the log-likelihood for each bit from Eq. (6) and MPA DETECTION
determines a symbol through the decoder. From Eqs. (7) and (8), the computational complexity of
MPA detection in SCMA increases as the number of super-
B. Computational complexity of MPA posed users df increases. That is, the computational complex-
There are two types of calculations required for MPA ity can be reduced by decreasing df . However, decreasing of
detection: updating of the likelihood of an SCMA codeword df reduces the number of simultaneously connectable users.
at each FN and updating of a priori probability at each VN. In Therefore, in this study, we employ a codebook including sym-
this study, we evaluate the number of complex multiplications bols having only I or Q component. As a result, when the re-
and the number of complex additions performed in these two ceived signal is handled separately by the I and Q components,

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 26,2020 at 13:43:18 UTC from IEEE Xplore. Restrictions apply.
 Q Real part Imaginary part

FN 1 FN 2 FN 3 FN 4 FN 1 FN 2 FN 3 FN 4

I
VN 1 VN 2 VN 3 VN 4 VN 5 VN 6 VN 1 VN 2 VN 3 VN 4 VN 5 VN 6

user 1
user 2
user 3
Fig. 4. Factor graphs of real part and imaginary part.

Real part Imagaginary part

Fig. 3. Codebook design.
FN-1 FN-2 FN-3 FN-4 FN-5 FN-6 FN-7 FN-8

interference from a signal in an orthogonal component which

is transmitted by another user is negligible. In addition, MPA VN-1 VN-2 VN-3 VN-4 VN-5 VN-6
detection is performed using a factor graph called the joint IQ
factor graph that treats the real part and the imaginary part of
the received symbol as different subcarriers. By employing Fig. 5. Joint IQ factor graph.
the joint IQ factor graph, the number of superposed users
in MPA detection is virtually reduced, and the computational
complexity can be reduced. graph. The joint IQ factor graph is shown in Fig. 5. FN-1 to
FN-4 in Fig. 5 indicate the real part of four subcarriers, and
A. Codebook design FN-5 to FN-8 indicate the imaginary part of four subcarriers.
As a codebook design, a method of rotating the phase of By performing detection using the joint IQ factor graph, it
the constellation point of M -ary amplitude shift keying (ASK) is possible to reflect the likelihood of SCMA codewords
has been proposed [11]. In this method, when the codebook in both real and imaginary parts in detection. Therefore,
size is M = 4 and the number of superposed users is df = 3, the degradation of the detection performance is prevented.
a 3-user constellation pattern is defined on each subcarrier as In addition, MPA detection can be performed with a small
shown in Fig. 3. At the transmission, ASK signals of these number of superposed users by treating real and imaginary
three users are added and transmitted on a subcarrier. parts as different FNs. As a result, it is possible to reduce
In this study, ASK constellation points with a phase angle the computational complexity without degrading the detection
of 0 and π/2 are assigned to two users to be superposed. performance.
Since the signal of these two users have only the real part (I
V. E VALUATION OF BER CHARACTERISTICS AND
component) or the imaginary part (Q component), interference
COMPUTATIONAL COMPLEXITY
from signal of the other user is negligible if the received
signal is handled separately by the I and Q components. In order to show the effectiveness of the proposed method,
Thus, the number of superposed users is virtually reduced. computer simulation was performed for MPA using the joint
However, if MPA is simply performed separately by the I IQ factor graph, MPA using the conventional factor graph,
and Q components, the detection performance is degraded. SMPA using the joint IQ factor graph, and SMPA using the
Therefore, we consider a method to carry out MPA detection conventional factor graph. The codebook shown in Section IV
without degrading the detection performance by making use was used for the conventional factor graph and the joint IQ
of the codebook characteristics. factor graph. Simulation specifications are shown in TABLE I.
The parameters of SCMA are K = 4, J = 6, and M = 4.
B. Joint IQ Factor Graph Furthermore, the number of iterations IT of MPA is six, and
Here, we consider MPA detection that exploits the charac- that of SMPA is two.
teristics of the codebook to mitigate interference from other Fig. 6 shows the bit error rate (BER) performance. The
users’ signals. When we perform MPA detection separately horizontal axis is the ratio of bit energy to noise power spectral
by the real part and the imaginary part as in Fig. 4, the density (Eb /N0 ), and the vertical axis is BER. Fig. 7 shows
number of superposed users in each factor graph is reduced. the packet error rate (PER) performance. The horizontal axis
Thus, the computational complexity is reduced even if two is the carrier to noise power ratio (CNR), and the vertical axis
MPA detections are performed in parallel. However, the de- is PER. From Fig. 6 and Fig. 7, the proposed method exhibits
tection performance is degraded compared to the conventional the BER and the PER performances equivalent to those of
MPA with one factor graph. This is because the detection MPA using the conventional factor graph.
performance of user’s symbols having both real and imaginary Next, we evaluate the computational complexity. The results
components is degraded. of calculating the number of multiplications and the number
In this study, we propose a factor graph that treats real and of additions in MPA detection using Eqs. (7), (8), and (9) are
imaginary parts as different FNs, and call it the joint IQ factor shown in TABLE II. As shown in TABLE II, the numbers of

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 26,2020 at 13:43:18 UTC from IEEE Xplore. Restrictions apply.
 TABLE I MPA using the conventional factor graph (r=1/2)
SMPA using the conventional factor graph (r=1/2)
S IMULATION PARAMETERS . MPA using the joint IQ factor graph (r=1/2)
SMPA using the joint IQ factor graph (r=1/2)

Number of subcarriers 600 MPA using the conventional factor graph (r=1/3)
SMPA using the conventional factor graph (r=1/3)
Number of FFT points 1024 MPA using the joint IQ factor graph (r=1/3)

FFT sampling rate 20 MHz SMPA using the joint IQ factor graph (r=1/3)

Channel coding Turbo Code

Code rate 1/2, 1/3

BER
Turbo decoding Log-MAP
Channel interleaver Block Interleaver
Channel model 4-path Rayleigh fading channel
with exponential decay
Delay between paths 2 dB
Attenuation between paths 2 sample
Normalized Doppler frequency 1.0 × 10−4
Eb /N0 (dB)

TABLE II Fig. 6. BER performance.

C OMPUTATIONAL COMPLEXITY OF MPA DETECTION .
MPA using the conventional factor graph (r=1/2)
Multiplications Additions SMPA using the conventional factor graph (r=1/2)

MPA (IT = 6) 92448 64512 MPA using the joint IQ factor graph (r=1/2)
SMPA using the joint IQ factor graph (r=1/2)

SMPA (IT = 2) 30912 21504 MPA using the conventional factor graph (r=1/3)
SMPA using the conventional factor graph (r=1/3)
Proposed (IT = 6) 23712 15360 MPA using the joint IQ factor graph (r=1/3)
SMPA using the joint IQ factor graph (r=1/3)
Proposed+SMPA (IT = 2) 8320 5120

multiplications and additions in MPA using the joint IQ factor

PER
graph are 25.65 percent and 23.81 percent of those in MPA
using the conventional factor graph, respectively. In addition,
the numbers of multiplications and additions in MPA using the
joint IQ factor graph is even smaller than those in SMPA using
the conventional factor graph. The numbers of multiplications
and additions in MPA using the joint IQ factor graph are 76.71
percent and 71.43 percent of those in the SMPA using the
CNR (dB)
conventional factor graph, respectively. This is because the
proposed method virtually reduces the number of superposed
users in MPA detection. Fig. 7. PER performance.
Since SMPA is a method to reduce the number of MPA
iterations, it can be used with the proposed method. The num- R EFERENCES
bers of multiplications and additions in SMPA using the joint
[1] J. Wannstrom, “LTE-advanced,” 3GPP, Jun. 2013.
IQ factor graph are 26.92 percent and 23.81 percent of those [2] H. Nikopur and H. Baligh, “Sparse code multiple access,” in Proc. IEEE
in SMPA using the conventional factor graph, respectively. PIMRC, pp. 322–336, London, Sept. 2013.
Therefore, the proposed method is effective to reduce the [3] Y. Wu, S. Zhang, and Y. Chen, “Iterative multiuser receiver in sparse
code multiple access system,” in Proc. IEEE ICC, pp. 2918–2923, Lon-
computational complexity of MPA detection. don, Jun. 2015.
[4] S. Zhang, X. Xu, L. Lu, Y. Wu, G. He, and Y. Chen, “Sparse code
VI. C ONCLUSIONS multiple access: An energy efficient uplink approach for 5G wireless sys-
In this study, we have proposed a method to reduce the com- tems,” in Proc. IEEE GLOBECOM, pp. 4782–4787, Austin, Dec. 2014.
[5] H. Mu, Z. Ma, M. Alhaji, P. Fan, and D. Chen, “A fixed low complexity
putational complexity of MPA detection by using a codebook message pass algorithm detector for up-link SCMA system,” IEEE
including symbols having only I or Q component, and the Wireless Commun. Lett., vol. 4, no. 6, pp. 585–588, Dec. 2015.
joint IQ factor graph. The computational complexity of MPA [6] Y. Du, B. Dong, Z. Chen, J. Fang, and L. Yang, “Shuffle multiuser detec-
tion schemes for uplink sparse code multiple access,” IEEE Commum.
increases exponentially as the number of superposed users Lett, vol. 20,no. 6, Jun. 2016.
increases. However, the proposed method can virtually reduce [7] Y. Du, B. Dong, Z. Chen, J. Fang, and X. Wang, “A fast convergence
the number of superposed users while maintaining the trans- multiuser detection scheme for uplink SCMA systems,” IEEE Wireless
Commun. Lett., vol. 5, no. 4, pp. 388–391, Aug. 2016.
mission rate and the number of simultaneously connectable [8] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and K.
users. Thus, the proposed method can reduce the computa- Higuchi, “Non-orthogonal multiple access (NOMA) for cellular future
tional complexity. Numerical results show that MPA using the radio access,” in Proc. IEEE VTC Spring, pp. 1–5, Germany, Jun. 2013.
[9] A. Peled and A. Ruiz, “Frequency domain data transmission using
joint IQ factor graph can reduce the numbers of multiplications reduced computational complexity algorithms,” in Proc. IEEE ICASSP,
and additions to 25.65 percent and 23.81 percent of those pp. 964–967, Denver, Apr. 1980.
of MPA the conventional factor graph, respectively, without [10] F. R. Kschishang, B. J. Frey, and H. A. Loeliger, “Factor graphs and
the sum-product algorithm,” IEEE Trans. Inf. Theory, vol. 47, no. 2, pp.
degrading the BER performance. Therefore, the proposed 498–519, Feb. 2001.
method is effective to reduce the computational complexity [11] M. Taherzadeh, H. Nikopur, A. Bayesteh, and H. Baligh “SCMA
of MPA detection. codebook design,” in Proc. IEEE VTC Spring, Vancouver, Sept. 2014.

Authorized licensed use limited to: Lulea University of Technology. Downloaded on July 26,2020 at 13:43:18 UTC from IEEE Xplore. Restrictions apply.