You are on page 1of 5

Comparison of Interference Cancellation Schemes

for Non-Orthogonal Multiple Access System


Guanghui Song and Xianbin Wang
Western University, London, Canada
Email: ghsong2008@hotmail.com, wang@eng.uwo.ca

Abstract—Three potential interference cancellation schemes


are compared for the application to a non-orthogonal multiple
access communication system. One is the conventional successive
h1
interference cancellation (SIC) scheme based on independent h2
single-user decodings. The other two, proposed in this paper, are
a soft-in soft-out parallel interference cancellation (SISO-PIC)
and a hybrid interference cancellation (HIC). The SISO-PIC is
an improved joint iterative multi-user detection scheme, which
has lower complexity than the prevalent multi-user detection. hK
The HIC combines the above two schemes to permit users to
be successively processed by a SISO-PIC window in the order
of their power levels. A comprehensive comparison is given for
these three schemes in aspects of error propagation, detection
delay, and complexity when a practical channel code, repeat-
accumulate code, is employed. Numerical results show that HIC Fig. 1. K-user NOMA communication model.
is a trade-off scheme of the three aspects.

I. Introduction yt = p1 h1,t x1,t + γ 1,t x1,t


F ul l Decoding

Non-orthogonal multiple access (NOMA) is considered as of User 1

a potential technique for application to the 5th generation x2,t


wireless system (5G) [1] [2]. By allowing users physically F ul l Decoding
p2 h2,t x2,t + γ 2,t
utilizing the same time and frequency resource, NOMA can of User 2

provide higher capacity load and increase the flexibility of het-


erogeneous services than the conventional orthogonal multiple
access. xK ,t
A new problem of NOMA is multi-user interference since pK hK ,t xK ,t + γ K ,t F ull Decodi ng

signals of different users cannot be separated in the con- of User K


ventional mean of time or frequency. Works [1]–[6] use a
Fig. 2. Conventional K-user SIC receiver model.
very simple successive interference cancellation (SIC) scheme
based on independent single-user decodings to deal with this
problem. The receiver first decodes the messages for strong
users with high power level, then remove their effects from the illustrated in Fig. 1. The received signal at the base station at
receive to decode weak users. SIC has no any sum rate loss time t is
in information point of view, but if a nonideal channel code 
K

is employed for each user there will be an error propagation yt = pk hk xk,t + zt , t = 1, ..., n
problem, i.e., decoding errors of strong users will effect the k=1
decodings for weak users. Another problem in SIC is “user
delay” since weak users should wait for the decodings of where pk , hk , xk,t , and zt are transmission power, channel gain,
strong users. These two problems become more serious as transmitted signal of user k, and Gaussian noise, respectively.
the users for SIC increase. Here hk is complex Gaussian with mean 0 and variance 1/2
for each dimension. We consider QPSK modulation xk,t ∈
II. Conventional SIC and Power Allocation { √12 + i √12 }, and the noise is complexity Gaussian with mean
In this section, we simply introduce the conventional SIC 0, variance N0 /2 for each dimension.
approach and power allocation. Due to the duality between An SIC receiver decodes the K users’ messages successive-
uplink and downlink channels, we simply consider the uplink ly. Suppose the decoding is in order of increasing user index.
communication scenario. Suppose there are K active mobile After recovering the messages of users 1, ..., k − 1, the receiver
users simultaneously transmitted messages to a base station subtracts these k−1 users’ interferences from the receive signal

978-1-5090-1698-3/16/$31.00 ©2016 IEEE


and decodes user-k’s message based on pk h1,t x1,t + ζ 1,t Processing
xˆ1,t

√ 
K
√ √
of User 1

pk hk xk,t + pk hk xk,t + zt = pk hk xk,t + γk,t (1)


yt = p2 h2,t x2,t + ζ 2,t xˆ2,t
i=k+1 Processing


K
pk hk ,t xk ,t + zt
Δ K √ of User 2

where γk,t = i=k+1 pk hk xk,t + zt is the equivalent noise for k =1 SISO

-PIC
user-k’s decoding. Figure 2 gives a diagram of the K-user SIC
scheme. pk hK ,t xK ,t + ζ K ,t
Processing
xˆ K ,t
To successfully recover the K users’ messages, the trans-
mission rates should satisfy of User K
⎛ ⎞
⎜⎜⎜ pk | hk |2 ⎟⎟⎟

Rk ≤ C ⎝ K ⎟⎠ , k = 1, ..., K (2)
i=k+1 pi | hi |
2 Fig. 3. K-user iterative receiver model based on SISO-PIC.

where Rk is the rate of user k and C(x) is the maximum


achievable rate at SNR x. Note that C(x) is determined by the effect is eliminated from the receive. The iteration terminates
employed channel codes. If optimal code is employed, C(x) until all the users’ messages are correctly recovered or a
becomes the capacity of AWGN channel. maximum iteration time is achieved. The estimation of the
For given transmission rates R1 , ..., RK , we have the optimal message of each user is noisy initially, but will be accurate as
power allocation derived from (2) the iteration time increases.

K Based on the receive signal yt and estimations xk,t output by
N0C −1 (Rk ) i=k+1 (C −1 (Rk ) + 1)
pk = , k = 1, ..., K (3) the single-user processings, the SISO-PIC performs interfer-
| hk |2
ence cancelation for each user and re-estimate user-k’s signal
where C −1 (x) is the inverse function of C(x). as
One issue that may be concerned for the application of √ √ √
pk hk,t xk,t + pk hk,t (xk,t −
xk,t ) + zt = pk hk,t xk,t + ζk,t
SIC to practical wireless system is “user delay” since user-k’s
ik
decoding should wait for the decodings of users 1, ..., k − 1.
 √ Δ
User delay becomes more serious when the user number is where ζk,t = ik pk hk,t (xk,t −
xk,t ) + zt is the residual
large. interference and Gaussian noise. Here xk,t = E[xk,t
Re
] + iE[xk,t
Im
]
Another problem is that SIC suffers a serious error propaga- Re
is estimated by single-user processing [7], where xk,t and xk,t Im
tion. In practical communication, there is always a probability are real and imaginary parts of xk,t . The variances of xk,t Re
and
that decoding error occurs. Since each user performs a full Im
xk,t output from the single-user processing are [7]
decoding as a one-off, hard decision errors of user k perma-
nently exists as residual noise in the subsequent decodings for 1 1
Re
D[xk,t ]= − E[xk,t
Re 2
] , D[xk,t
Im
] = − E[xk,t ].
Im 2
users k, k + 1, ..., K. The decoding errors are accumulated and 2 2
propagated as the decoding proceeds, so the problem becomes There are low-complexity near-optimal algorithms for the
more significant when the user number is large. For this single-user processing such as the belief-propagation (BP)
reason, the average decoding BER increases as K increases. algorithm for the decoding of turbo and LDPC codes [8], [9].
To proceed single-user processing, we should evaluate the
III. A Low-Complexity SISO-PIC equivalent noise variance of both the real and imaginary parts
The original SISO-PIC framework is proposed in [7] for of ζk,t . In the following, rather than obtaining the accurate
iterative multi-user detection. For an AWGN MAC the com- variances for both with high complexity O(K 2 ) as in [7], we
putation complexity is O(K), while for a general fading MAC, first compute their total variance as
its complexity is O(K 2 ). We modify the original SISO-PIC to 
give a simpler detection scheme for fading MAC with linear
Re
D[ζk,t ] + D[ζk,t
Im
]= pi | hi,t |2 (D[xk,t
Re
] + D[xk,t
Im
]) + N0 .
ik
complexity.
Figure 3 illustrates a K-user iterative SISO-PIC receiver Then we use (D[ζk,tRe
] + D[ζk,t
Im
])/2 as an approximate variance
model. The decoding is realized through global iterations for the real and imaginary parts of ζk,t . The complexity of this
between SISO-PIC and K single-user processings. Note that processing is just O(K).
the single-user processing is not a full decoding, or we can The SISO-PIC based iterative multi-user detection can sig-
simply consider that the full decoding consists of several inner nificantly mitigate the error propagation problem since the K
iterations each of which is a single-user processing [8]. At the users’ decodings are performed jointly. Moreover, since the
beginning, the SISO-PIC output an initial estimation for the single-user processings are performed in parallel, user delay
signals of each user, which are further refined by the single- problem is also mitigated.
user processings. The K single-user processings are performed One disadvantage of this SISO-PIC is that it does not utilize
in parallel and the outputs are used for interference cancelation the power difference between users for user detection, so a
for the next iteration. If one user message is fully recovered, its lot of computation is wasted for joint multi-user processing.
1
from the receive and a new user join the window. We give
0.9
the step-by-step procedure in Algorithm 1 to describe the HIC.
0.8

0.7
Algorithm 1 Hybrid interference cancellation for NOMA
1: Empty the decoding window
2: for k = 1, 2, ..., w do
0.6

0.5 3: Add user k to the window;


0.4
4: Perform SISO-PIC for users in the window;
5: Processing once for users in the window;
0.3
user1 2 3 4 5 6 7 8 6: end for
0.2
7: k = w + 1, t = 1; // t denotes the iteration time
8: while k ≤ K or t ≤ T max do
0.1
9: Perform SISO-PIC for users in the window;
0 10: Processing once for users in the window;
0 5 10 15 20 25 30 35 40

Iteration Time 11: if strongest user is recovered then


12: Hard decision for the strongest user;
Fig. 4. MSEs of user 1, ..., 8 under 8-user SISO-PIC. The channel code 13: if k = K then
employed by each user is a rate-1/4 regular RA code.
14: Break;
15: else
If we employ the optimal power allocation in (3), the SINR 16: Remove the effect of strongest user;
of some users is very low at the beginning of iteration, little 17: Add user k + 1 to the window;
benefit could be obtained from their processings. 18: k = k + 1;
19: end if
IV. An Hybrid Interference Cancellation Scheme 20: end if
Under the optimal power allocation in (3), we give an HIC 21: t = t + 1;
22: end while
scheme for NOMA based on the power difference of each
user. We successively activate the w, w < K, strongest users
for joint decoding. So the complexity is significantly reduced
V. Error Propagation
if w is small.
Our motivation is from Fig. 4, where the mean square error In this section, we give some numerical results to show
(MSE) of each user for 8-user SISO-PIC under the optimal advantages of HIC scheme in the aspects of error propagation,
power allocation is illustrated as a function of iteration time. user delay, and decoding complexity. We still consider the
For user k, uplink communication and assume that the base station and
Δ 1
n
mobile users know the accurate channel gain of the link so that
MSE = | xk,t −
xk,t |2 an optimal power allocation is used to fascinate interference
n t=1
cancelation. We simulate the BER performances of SIC, SISO-
which reflects the reliability of the estimated message. Small PIC and HIC over block Rayleigh fading channel in Fig. 5.
MSE (near 0) means high reliability while high MSE (near 1) We employed a rate-1/4 regular RA code [9] with information
means low reliability. MSE= 0 implies near 0 decision error length 1024 for each user. The maximum decoding iteration
rate. The channel code employed by each user is a rate-1/4 times for each user in SIC is 50 and for joint decoding in
regular repeat-accumulate (RA) code [9]. Even using the 8- SISO-PIC and HIC are 60. The target SNRs of the power
user joint decoding, due to the power difference, their MSEs allocation for each user are fixed at −2.03 and −1.78 dB which
converge to 0 in decreasing order of the power. MSE of user correspond to the target BERs of about 1E − 4 and 1E − 5,
1 converges fastest in about 8 iterations but MSE of user k respectively, for the employed RA code. We obtain the average
keeps near 1 for iteration time < 30, so participating iteration decoding error rate of K users with K = 1, 2, 4, 8. The bit-
at the beginning for user K is waste of computation. error-rate (BER) of SIC increases with the number of users due
We give a low-complexity HIC scheme by utilizing the to error propagation. In fact the BER of SIC at least linearly
power difference of each user. Our scheme activates users increases with K. BERs of SISO-PIC decrease as the number
to participate the decoding iteration successively according of users increases due to joint decoding. Our HIC is much
to their power level. We imagine a virtual decoding window better than SIC and slightly worse than SISO-PIC since we
with size w which is aligned with the present w strongest just considered joint processing for w users.
users for joint processing. The activated users in the decoding
window are processed in parallel as the procedure given VI. User Delay and Detection Complexity
in Sec. III. If the message of the strongest user in the Figure 6 illustrates the cumulative distribution function
window is fully recovered, its effect would be eliminated (CDF) of total iteration times of SIC, SISO-PIC, and HIC
10-2 180
SIC (-2.03)
SIC
SISO-PIC (-2.03)
SISO-PIC

160 w
SIC (-1.78)
HIC, =2
SISO-PIC (-1.78)

HIC (-2.03, w =2)


HIC, w =3

HIC (-2.03, w =3)

HIC (-1.78, w =2) 140


10-3 HIC (-1.78, w =3) sg
ni
sse120
co
rP
re100
RE S
-4 U
-e
B10 lg
ni 80
Sf
roe
b 60
m
u
10-5 N
40

20

10-6
1 2 3 4 5 6 7 8 0
User Number K 1 2 3 4 5 6 7 8
User Number K
Fig. 5. BERs of K-user SIC, SISO-PIC, and HIC. The channel code employed
Fig. 7. Number of single-user processings.
by each user is a rate-1/4 regular RA code.

1
TABLE I
Comparison of three interference cancellation schemes.
0.9

Schemes Error propagation User delay Complexity


0.8
SIC Yes High Low
SISO-PIC No Low High
0.7
HIC No Low Low
0.6

F
D
C0.5

Since the complexity for interference cancelation is very small


0.4
compared with that of the single-user processing, we ignore
0.3 Single-user K =1 this computation complexity and only consider the complexity
SIC, K =2

SIC, K =4
of K single-user decodings. The complexity is mainly de-
0.2 K
SIC,

SISO-PIC,
=8

K =2
termined by the average number of single-user processings
SISO-PIC, K =4
required for successful decoding. Since the required iteration
0.1 SISO-PIC, K =8

HIC, K =4, w =2
number of the K-user SIC linearly increases with the user
0
HIC, K =8, w =2

0 20 40 60 80 100 120 number K. The decoding complexity of SIC also linearly


Total Iteration Time increases with K as shown in Fig. 7. SISO-PIC has high
complexity due to joint K-user decoding. HIC scheme with
Fig. 6. CDF of total iteration time.
w = 2 has the similar complexity with SIC. Table I gives a
summary of the comparison between the three schemes.
under the optimal power allocation with the target SNR −1.78 References
dB for each user. Note that the total iteration times of SIC is [1] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and K.
the sum of iteration times of K users’ decodings. For a given Higuchi, “Non-orthogonal multiple access (NOMA) for cellular future
iteration times x, CDF(x) is the probability of that the K users’ radio access,” in Proc. IEEE VTC2013-Spring, pp. 1–5.
[2] Y. Saito, A. Benjebbour, Y. Kishiyama,, and T. Nakamura, “Systemlev-
messages are recovered correctly with fewer than x iterations. el performance evaluation of downlink non-orthogonal multiple access
The required iteration time for SIC linearly increases with (NOMA),” in Proc. IEEE PIMRC2013, pp. 611–615.
number of user K. Let T be the iteration time required for [3] K. Higuchi and Y. Kishiyama, “Non-orthogonal access with random
beamforming and intra-beam SIC for cellular MIMO downlink,” in Proc.
single-user (RA) decoding K = 1. The required iteration time IEEE VTC 2013-Fall, pp. 1–5.
for K-user SIC is about KT . As K is large, there will be a [4] Z. Ding, Z. Yang, P. Fan, and V. Poor, “On the performance of non-
significant user delay. SISO-PIC reduce the user delay and orthogonal multiple access in 5G systems with randomly deployed users,”
IEEE Signal Process. Lett., vol. 21, no. 12, pp. 1501–1505, Dec. 2014.
requires fewer iterations to converge observed in Fig. 6 due to [5] S. Timotheou and I. Krikidis, “Fairness for non-orthogonal multiple
parallel processing for K users. HIC even with w = 2 has the access in 5G systems,” IEEE Signal Process. Lett., vol. 22, no. 10, pp.
similar user delay as joint K-user detection based on SISO- 1647–1651, Oct. 2015.
[6] P. Wang, J. Xiao, and P. Li, “ Comparison of orthogonal and nonorthog-
HIC. onal approaches to future wireless cellular systems.” IEEE Veh. Technol.
Let’s evaluate the decoding complexity of the three schemes. Mag., vol. 1, no. 3, pp. 4–11, Sept. 2006.
[7] P. Li, L. H. Liu, K. Y. Wu, and W. K. Leung, “Interleaving-division
multiple-access,” IEEE Trans. Wireless Commun., vol. 5, no. 4, pp. 938–
947, Apr. 2006.
[8] G. Song, J. Cheng, and Y. Watanabe, “Maximum sum rate of repeat-
accumulate interleave-division system by fixed-point analysis,” IEEE
Trans. Commun., vol. 60, no. 10, pp. 3011–3022, Oct. 2012.
[9] W. E. Ryan and S. Lin, Channel Codes: Classical and Modern, Cam-
bridge,

You might also like