Anda di halaman 1dari 5

2016 24th European Signal Processing Conference (EUSIPCO)

Massive Overloaded MIMO Signal Detection

via Convex Optimization with Proximal Splitting
Ryo Hayakawa, Kazunori Hayashi Hampei Sasahara Masaaki Nagahara
Graduate School of Informatics, Engineering School, Institute of
Kyoto University, Tokyo Institute of Technology, Environmental Science and Technology,
Yoshida-Hommachi, Sakyo-ku, Ookayama, Meguro-ku, The University of Kitakyushu,
Kyoto, 606-8501, Japan Tokyo, 152-8550, Japan Hibikino, Wakamatsu-ku, Kitakyushu,
Email:, JSPS Research Fellow Fukuoka, 808-0135, Japan Email: Email:

AbstractIn this paper, we propose signal detection schemes weight, cost and/or power consumption of the receiver. Such
for massive overloaded multiple-input multiple-output (MIMO) MIMO systems, where the number of receive antennas are
systems, where the number of receive antennas is less than that less than that of transmitted streams, are known as overloaded
of transmitted streams. Using the idea of the sum-of-absolute-
value (SOAV) optimization, we formulate the signal detection as (or underdetermined) MIMO systems. The slab-sphere decod-
a convex optimization problem, which can be solved via a fast ing [8] is a signal detection algorithm based on maximum
algorithm based on Douglas-Rachford splitting. To improve the likelihood (ML) for overloaded MIMO systems to find the
performance, we also propose an iterative approach to solve the solution with lower complexity than that of exhaustive search.
optimization problem with weighting parameters update in a Some techniques, such as the pre-voting cancellation [9] and
cost function. Simulation results show that the proposed scheme
can achieve much better bit error rate (BER) performance than the virtual channel [10], transform overloaded systems into
conventional schemes, especially in large-scale overloaded MIMO non-overloaded systems to apply conventional MIMO signal
systems. detection. [11] and [12] employ the ideas in [8] and [9] to
Index Termsmassive MIMO, overloaded MIMO, proximal achieve a good performance with lower complexity. For mas-
splitting methods, Douglas-Rachford algorithm, SOAV optimiza- sive overloaded MIMO systems, however, these schemes are
not practical because their complexity is still high, while the
performance of low-complexity detection for massive MIMO
I. I NTRODUCTION systems is considerably degraded in the overloaded scenario.
Massive multiple-input multiple-output (MIMO) systems, To further reduce the complexity, the enhanced reactive tabu
where tens or hundreds of antennas are equipped in each search (ERTS), which is an extension of the random restart re-
transmitter and receiver, are gathering attention as a method active tabu search (RTS) [6], has been recently proposed [13].
to achieve very high spectral efficiency [1], [2]. In massive ERTS employs RTS iteratively while varying the initial point
MIMO systems, low complexity signal detection method is of the search randomly until the estimate by RTS satisfies a
essential because the required computational complexity of certain condition. In [13], it is shown that ERTS can achieve
MIMO signal detection generally increases along with the a comparable performance to the optimal ML detection with
increase of the antennas. Some of the candidates for massive affordable computational complexity for overloaded MIMO
MIMO systems are linear signal detections, such as the zero systems with tens of antennas. With hundreds of antennas,
forcing (ZF) and the minimum mean square error (MMSE) however, ERTS requires prohibitive computational complexity
detection methods. Besides them, some non-linear detection to achieve such performance because the required number of
schemes have also been proposed. The likelihood ascent search RTSs increases with the number of antennas.
(LAS) [3], [4] and the reactive tabu search (RTS) [5], [6] In this paper, we propose a massive overloaded MIMO
employ local neighborhood search of likelihood and achieve signal detection scheme with much lower complexity than that
much better performance than linear detection. The graph- of conventional schemes [8][12]. We formulate the signal
based iterative Gaussian detector (GIGD) [7] is well known detection problem as a convex optimization problem, where
as a low complexity scheme built upon belief propagation the idea is based on the sum-of-absolute-value (SOAV) opti-
techniques. mization [14], which is a technique to reconstruct a discrete-
In MIMO systems, a sufficient number of receive antennas valued vector from its linear measurements. The optimization
may not be available because of the limits on the size, problem can be efficiently solved with proximal splitting
methods [15] even for underdetermined systems. To improve
This work was supported in part by the Grants-in-Aid for Scientific the performance, we extend SOAV optimization to weighted-
Research no. 15K06064, 15H2252, 15H02668, 15K14006, and 26120521
from the Ministry of Education, Science, Sports, and Culture of Japan and SOAV optimization, where the prior information about the
the Telecommunications Advancement Foundation. discrete-valued vector can be used, and propose an iterative

978-0-9928-6265-7/16/$31.00 2016 IEEE 1383

2016 24th European Signal Processing Conference (EUSIPCO)

approach, named iterative weighted-SOAV (IW-SOAV), using Since s

{1 + j, 1 + j, 1 j, 1 j}n , s is a binary vector
the estimate in the previous iteration as the prior information. whose elements are 1 or 1.
Since the weighted-SOAV optimization problem can also
be efficiently solved with proximal splitting methods, IW- III. P ROPOSED S IGNAL D ETECTION S CHEMES
SOAV can detect the transmitted signals with low computa- In this section, we propose signal detection schemes based
tional complexity. Simulation results show that IW-SOAV can on SOAV optimization for massive overloaded MIMO sys-
achieve much better bit error rate (BER) performance than tems. We briefly review SOAV optimization in Sect. III-A
conventional signal detection schemes especially in large-scale and propose a signal detection scheme in Sect. III-B. In
overloaded MIMO systems. Sect. III-C, we also propose an iterative approach to improve
In the rest of the paper, we use the following notations. the performance.
Superscript ()T and ()H denote the transpose and the Hermi-
tian transpose, respectively. Mathematica symbols, j, I, 1, and A. SOAV Optimization
0 represent the imaginary unit, the identity matrix, a vector
whose elements are all 1, and a vector whose elements are all SOAV optimization [14] is a technique to reconstruct an
0. For a vector a = [a1 aN ]T RN , we define the 1 and unknown discrete-valued vector as x = [x1 xN ]T
2 norms of a as {c1 , . . . , cP }N RN from its linear measurements =
" Ax, where A RM N . If we assume Pr(xi = cp ) =
!N #! 1/P (p = 1, . . . , P ) for all xi (i = 1, . . . , N ), each of
a = 1 |a | and a = $
i 2 a2 , i (1) x c1 1, . . . , x cP 1 has approximately N/P zero elements.
i=1 i=1 Based on this property and the idea of 1 optimization in
respectively. Pr(A) denotes the probability of an event A and compressed sensing [16], SOAV optimization solves
E[] stands for the expectation operator.
1 !
II. S YSTEM M ODEL minimize x cp 11
xRN P p=1
Here we consider a MIMO system with n transmit antennas
and m receive antennas. For simplicity, precoding is not con- subject to = Ax (6)
sidered and the number of transmitted streams is assumed to be
equal to that of transmit antennas. In addition, we employ the to reconstruct x from .
quadrature phase shift keying (QPSK) and define the alphabet
= {1 + j, 1 + j, 1 j, 1 j}. B. Proposed Signal Detection via SOAV Optimization
of the transmitted symbol as S
The transmitted signal vector s = [s1 sn ]T Sn is In MIMO systems, the transmitted signal vector s is
composed of signals transmitted from n transmit antennas, commonly discrete and the received signal vector y can be
where sj (j = 1, . . . , n) denotes the symbol sent from the jth regarded as its linear observations if the noise can be ignored.
transmit antenna, E[ s] = 0, and E[ H ] = 2I. The received
ss Since each element of s is 1 or 1 for the case with QPSK,
signal vector y
= [ y1 ym ] C , where yi (i = 1, . . . , m)
m we can formulate the signal detection problem as SOAV
denotes the signal received at the ith receive antenna, is given optimization, i.e.,
by 1 1
s+v minimize z 11 + z + 11
= H
y , (2) zR2n 2 2
where subject to y = Hz. (7)

h 1,n
h Since the received signal vector y contains the additive noise
.. .. .. Cmn (3)
H= . . . as in (4), we modify the optimization problem as follows:
hm,n 1 1
minimize z 11 + z + 11
i,j represents the channel
is a flat fading channel matrix and h 2n
zR 2 2

gain from the jth transmit antenna to the ith receive antenna. + y Hz22 (8)
Cm is the additive white complex Gaussian noise vector
v 2
with zero mean and covariance matrix of v2 I. The signal by using the idea of 1 -2 optimization. Here, > 0 is a
model (2) can be rewritten as given constant. The solution of (8) can be obtained with the
following theorem [15].
y = Hs + v, (4)
Theorem 1. Let 1 , 2 : R2n (, ] be lower semicon-
where tinuous convex functions and (ri dom 1 ) (ri dom 2 ) = .
+ , + ,
Im{H} In addition, 1 (z) + 2 (z) as z2 is assumed.
y= ,H= ,
Im{y} Im{H} Re{H} A sequence zk (k = 0, 1, . . .) converging to the solution of
+ , + ,
Re{ v}
s= ,v= . (5) minimize 1 (z) + 2 (z) (9)
Im{ v}
Im{ zR2n

2016 24th European Signal Processing Conference (EUSIPCO)

[proxf (z)]j [proxfw (z)]j

1 1 wj 1 wj+

1 1 1 1 dj
0 1 1+ zj 0 1 dj 1+ zj

1 1
1 0 1 sj

Fig. 3. wj+ and wj for IW-SOAV

Fig. 1. [proxf (z)]j Fig. 2. [proxfw (z)]j

can be obtained by using the following Douglas-Rachford C. Proposed Iterative Approach, IW-SOAV
algorithm. Here, the proximity operator of a function : Assuming that we have information on prior probabilities
R2n R is defined as of wj+ = Pr(sj = 1) and wj = Pr(sj = 1), we extend the
1 problem of (8) to weighted-SOAV optimization problem as
prox (z) = arg min2n (u) + z u22 . (10)
uR 2 2n
! 1 2
minimize wj+ |zj 1| + wj |zj + 1|
Algorithm 1. (Douglas-Rachford Algorithm) zR 2n
1) Fix (0, 1), > 0, and r0 R2n .
+ y Hz22 . (14)
2) For k = 0, 1, 2, . . ., iterate 2
If there is no prior information about s, i.e., wj+ = wj = 1/2,

zk = prox2 (rk ) the optimization problem (14) is equivalent to (8). If wj+ > wj
k [, 2 ] then arg min fwj (zj ) = 1, where fwj (zj ) = wj+ |zj 1| +

rk+1 = rk + k (prox1 (2zk rk ) zk ). wj |zj + 1|, thus the solution of zj in (14) tends to take the
value close to 1, and vice versa. The optimization problem (14)
We can rewrite (8) as
can also be solved by using the Douglas-Rachford algorithm.
The proximity operator of
f (z) + g(z), (11)
zR 2n
! 1 2
fw (z) = wj+ |zj 1| + wj |zj + 1| (15)
where f (z) = z 11 /2 + z + 11 /2 and g(z) = y
Hz22 /2. The proximity operators of f (z) and g(z) can
be obtained as can be written as

zj + (zj < 1 ) [proxfw (z)]j

1 (1 zj < 1)
zj + (zj < 1 )

[proxf (z)]j = zj (1 zj 1) , (12)
1 (1 zj < 1 dj )

1 (1 zj < 1 + ) = zj + dj (1 dj zj < 1 dj ) (16)

z (1 + z )
1 (1 dj zj < 1 + )
j j

z (1 + zj )
as shown in Fig. 2, where dj = wj+ wj . By solving the
proxg (z) = (I + H T H)1 (z + H T y), (13) optimization problem (14) via the Douglas-Rachford algorithm
with proxfw and proxg , a new estimate of the transmitted
respectively, where [proxf (z)]j (j = 1, . . . , 2n) represents signal vector s can be obtained.
the jth element of proxf (z). Note that [proxf (z)]j is a The prior information on s is not available in a common
function of zj only as shown in Fig. 1. By solving (8) with the scenario, however, assuming iterative approach, the estimate
Douglas-Rachford algorithm, the estimate of the transmitted in the previous iteration can be used to obtain the prior
signal vector s can be obtained. probabilities. Specifically, in the proposed iterative approach

2016 24th European Signal Processing Conference (EUSIPCO)

named IW-SOAV, we iteratively solve the weighted-SOAV 10


optimization problem (14) while updating the parameters wj+

and wj (i = 1, . . . , 2n) as

0 sj < 1)

1 + s
wj+ = (1 sj < 1) (17) 10


1 (1 sj )

1 sj < 1)
+ 1 sj
wj = 1 wj = (1 sj < 1) , (18) ERTS
IW-SOAV (L=10)
0 (1 sj ) 10
0 5 10 15 20 25 30
where sj is the estimate of sj in the previous iteration. Fig. SNR per receive antenna (dB)
3 shows wj+ and wj as a function of sj . wj+ is large when
Fig. 4. BER performance for (n, m) = (25, 16)
sj is large, and wj is large when sj is small. This is because
the estimates close to 1 or 1 will be more reliable than
those close to 0. The proposed algorithm of IW-SOAV is
summarized as follows:
Algorithm 2. (Proposed Signal Detection via IW-SOAV)
1) Let s = 0 and iterate a)c) for L times.
a) Compute wj+ , wj with (17),(18).
b) Fix (0, 1), > 0, K > 0, and r0 R2n . 10-2

c) For k = 0, 1, 2, . . . , K, iterate

zk = proxg (rk )
k [, 2 ]

rk+1 = rk + k (proxfw (2zk rk ) zk ) 10-4 GIGD
= zK .
and let s IW-SOAV (L=10)
2) Obtain sgn(s) as the final estimate of s. 0 5 10 15 20 25 30
SNR per receive antenna (dB)
In this section, we evaluate the BER performance of the Fig. 5. BER performance for (n, m) = (150, 96)
proposed scheme by computer simulation comparing with
that of conventional detection methods. In the simulation, flat
Rayleigh fading channels are assumed and H is composed of performance of IW-SOAV with L = 1 and 10 in the figures,
independent and identically distributed complex Gaussian ran- where L is the number of iterations, we can see that the
dom variables with zero mean and unit variance. The param- BER performance is improved with the proposed iterative
eter in (14) is selected as = 104 , 103 , 102 , 101 , 1, 1, approach. Although the performance of ERTS is much better
and 1 for SNR per receive antenna of 0, 5, 10, 15, 20, 25, and than that of IW-SOAV in Fig. 4, it considerably degrades for
30 (dB), respectively. The other parameters of the proposed larger number of antennas as shown in Fig. 5 and 6. This
schemes are set as K = 50, = 0.1, = 1, k = 1.9 (k = is because, if the number of transmit antennas is large, RTS
0, 1, . . . , K), and r0 = 0. often fails to find the true transmitted signal vector due to the
Figs. 46 shows the BER performance for overloaded huge number of candidates of the transmitted vector. Although
MIMO systems with (n, m) = (25, 16), (150, 96), and we may get better performance with ERTS by increasing
(200, 128), respectively, where the ratio m/n is fixed to be the number of RTSs, the computational complexity could
0.64 for all cases. In the figures, MMSE, GIGD, and ERTS be prohibitive to achieve comparable performance as IW-
represent the linear MMSE detection, the belief propagation- SOAV. Specifically, given that the computational complexity
based detection [7], and the massive overloaded MIMO signal of ERTS is O(n3 ) + O(NRTS n2 ) in the worst case, and
detection proposed in [13], respectively. The parameters of the number of all candidates of the transmit signal vector
ERTS are the same as those in [13], e.g., the maximum increases exponentially with the number of transmit antennas,
number of RTSs is NRTS = 500 and the maximum number the required NRTS to keep good performance will increase
of iterations in RTS is Nitr = 300. IW-SOAV denotes more rapidly than n. On the other hand, the computational
our proposed scheme shown in Algorithm 2. Comparing the complexity of IW-SOAV is O(n3 ), which is dominated by

2016 24th European Signal Processing Conference (EUSIPCO)

In this paper, we have proposed a massive overloaded
MIMO signal detection scheme, namely IW-SOAV, which
iteratively solves the weighted-SOAV optimization problem
while updating its parameters. Simulation results show that
IW-SOAV can achieve much better performance than conven-
tional massive MIMO detection schemes, especially in large-

scale overloaded MIMO systems. Future work includes the

integration of our proposed scheme and soft channel decoding
schemes, such as low density parity check (LDPC) codes and
GIGD turbo codes.
0 5 10 15 20 25 30 [1] A. Chockalingam and B. S. Rajan, Large MIMO systems, Cambridge
SNR per receive antenna (dB) University Press, 2014.
[2] L. Lu, G. Y. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang, An
Fig. 6. BER performance for (n, m) = (200, 128) overview of massive MIMO: Benefits and challenges, IEEE J. Sel.
Topics Signal Process., vol. 8, no. 5, pp. 742758, Oct. 2014.
[3] K. V. Vardhan, S. K. Mohammed, A. Chockalingam, and B. S. Rajan,
A low-complexity detector for large MIMO systems and multicarrier
CDMA systems, IEEE J. Sel. Areas Commun., vol. 26, no. 4, pp. 473
MMSE 485, Apr. 2008.
100 GIGD
ERTS [4] P. Li and R. D. Murch, Multiple output selection-LAS algorithm in
IW-SOAV (L=1) large MIMO systems, IEEE Commun. Lett., vol. 14, pp. 399401, May
IW-SOAV (L=10) 2010.
10-1 [5] N. Srinidhi, S. K. Mohammed, A. Chockalingam, and B. S. Rajan,
Near-ML signal detection in large-dimension linear vector channels
using reactive tabu search, arXiv:0911.4640v1 [cs.IT] 24 Nov. 2009.
10-2 [6] T. Datta, N. Srinidhi, A. Chockalingam, and B. S. Rajan, Random-
restart reactive tabu search algorithm for detection in large-MIMO

systems, IEEE Commun. Lett., vol. 14, no. 12, pp. 11071109, Dec.
-3 2010.
[7] T. Wo and P. A. Hoeher, A simple iterative Gaussian detector for
severely delay-spread MIMO channels, in Proc. IEEE ICC 2007,
Glasgow, U.K., pp. 45984563, Jun. 2007.
10-4 [8] K. K. Wong and A. Paulraj, Efficient high-performance decoding for
overloaded MIMO antenna systems, IEEE Trans. Wireless Commun.,
vol. 6, no. 5, pp. 18331843, May 2007.
10-5 [9] L. Bai, C. Chen, and J. Choi, Lattice reduction aided detection for
80 85 90 95 100 105 110
underdetermined MIMO systems: a pre-voting cancellation approach,
in Proc. VTC Spring 2010, Taipei, Taiwan, pp. 15, May 2010.
[10] S. Denno and A. Kitamoto, An overloaded MIMO receiver with
Fig. 7. BER performance versus m for n = 150 and SNR per receive extended rotation matrices for virtual channels, in Proc. APCC 2015,
antenna of 20 dB Kyoto, Japan, pp. 5256, Oct. 2015.
[11] R. Hayakawa, K. Hayashi, and M. Kaneko, An overloaded MIMO
signal detection scheme with slab decoding and lattice reduction, in
Proc. APCC 2015, Kyoto, Japan, pp. 4246, Oct. 2015.
[12] R. Hayakawa, K. Hayashi, and M. Kaneko, Lattice reduction-aided
the calculation of (I + H T H)1 in (13). Note that the detection for overloaded MIMO using slab decoding, IEICE Trans.
Commun., vol. E99-B, no. 8, Aug. 2016.
calculation of (I + H T H)1 is required only once, and [13] T. Datta, N. Srinidhi, A. Chockalingam, and B. S. Rajan, Low-
thus the corresponding computational cost does not grow complexity near-optimal signal detection in underdetermined large-
with K or L. If the ratio m/n is fixed, the computational MIMO systems, in Proc. NCC 2012, Kharagpur, India, pp. 15, Feb.
complexity of IW-SOAV is the same order as that of MMSE [14] M. Nagahara, Discrete signal reconstruction by sum of absolute values,
detection O(m3 ) + O(mn) for an overloaded scenario. From IEEE Signal Process. Lett., vol. 22, no.10, pp.15751579, Oct. 2015.
the figures, our proposed scheme has much better performance [15] P. Combettes and J. Pesquet, Proximal splitting methods in signal pro-
cessing, in Fixed-Point Algorithms for Inverse Problems in Science and
than the other schemes for the large-scale overloaded MIMO Engineering, ser. Springer Optimization and Its Applications. Springer
systems. New York, vol. 49, pp. 185212, 2011.
[16] K. Hayashi, M. Nagahara, T. Tanaka, A users guide to compressed
Fig. 7 shows the BER performance versus the number sensing for communications systems, IEICE Trans. Commun., vol. E96-
of receive antennas m for n = 150 and SNR per receive B, no. 3, pp. 685712, Mar. 2013.
antenna of 20 dB. We can observe that IW-SOAV with
L = 10 requires less antennas than other schemes to achieve
good BER performance. For BER = 104 , IW-SOAV can
reduce approximately 10 receive antennas compared to the
conventional ERTS.