Improved Symbol-Based Belief Propagation

Detection for Large-Scale MIMO

Junmei Yang1 , Chuan Zhang1 , Xiao Liang1 , Shugong Xu2 , and Xiaohu You1
1
National Mobile Communications Research Laboratory, Southeast University, Nanjing, China
2
Intel Collaborative Research Institutes on Mobile Networking and Computing, Intel Labs, Shanghai, China
Email: 1 {jm.yang, chzhang, xhyu}@seu.edu.cn, 2 shugong.xu@intel.com

Abstract—In this paper, BP detection based on belief propaga- complexity in the scenario of large-scale MIMO systems with
tion in real domain for large-scale MIMO systems is proposed. a high-order constellation. Although several improved low-
Numerical results have shown that, with quadrature phase shift complexity variations, such as lattice reduction (LS) aided
keying (QPSK) modulation, this approach can show 1 dB perfor-
mance improvement at the BER of 10−2 , compared to conven- detection [4] and distributed K-Best approach [5], have been
tional single-edge based BP (SE-BP) in complex domain. Based proposed, the efficient detection which perfectly balances both
on the proposed BP detection, its symbol-based variation is also performance and complexity remains challenging for large-
investigated for applications in large-scale MIMO systems with scale MIMO. To this end, the sub-optimal detection approach,
a high-order modulation. This symbol-based method successfully minimum mean-squared error (MMSE), is proposed recently
reduces computational complexity by avoiding large dimensional
matrix inversion and decomposition. Since the proposed method by [6]. However, its successful application relies on Neumann
can also shrink the constellation size, its complexity can be approximation to deal with large-scale matrix inversion, and
further reduced. Numerical simulation results and complexity the approximate error is proportional to the antenna ratio
comparison have demonstrated that the proposed symbol-based (M 2 /N ), where M and N are the numbers of transmitting
BP detection can show advantages in both performance and and receiving antennas, respectively.
complexity compared to existing ones. Therefore, it is suitable
for large-scale MIMO system applications, especially for those Therefore, the dominating computation complexity lays in
with high-order modulations. ML method, K-Best method, SD method, or MMSE method
Index Terms—Large-scale MIMO, belief propagation (BP), has put the barrier which could not be override for large-
symbol-based, message passing, iterative detection. scale MIMO detection. In order to bypass this bottleneck,
several alternative algorithms, which are less computation-
I. I NTRODUCTION intensive, have been recently proposed, such as likelihood
By transmitting multiple data streams concurrently and ascent search (LAS) algorithm, reactive tabu search (RTS)
in the same frequency band, multiple-input multiple-output algorithm, and belief propagation (BP) algorithm [7–10].
(MIMO) system successfully improves system capacity and Among those candidates, the BP algorithm provides better
data rate over the single-antenna system. The resultant higher performance in general, and is less vulnerable to the local
spectral efficiency and better link reliability have made it minimum problem. Second, BP algorithm is robust and does
increasingly popular for both academia and industry. Nowa- not requires a carefully selected initial solution vector. Last
days, MIMO in combination with spatial multiplexing has but the most important, BP algorithm is matrix-inverse free,
been adopted by the latest standards such as 3GPP LTE- which make it very attractive for large-scale MIMO detec-
Advanced [1] and IEEE 802.11n [2]. In order to fulfill the ever- tion. In [11], Markov rand field (MRF)-based model and
increasing demands of future wireless communication, systems factor graph (FG)-based model with Gaussian approximation
are required to be equipped with an order of higher magnitude of interference (GAI) are incorporated with BP algorithm.
of antenna arrays than conventional MIMO systems. That is Unfortunately, this bit-based BP detection is only suitable for
how the popular concept of large-scale MIMO comes out. large-scale MIMO systems with low modulation order, such as
Without any doubt, large-scale MIMO has ranked one of the binary phase shift keying (BPSK). In [12], a symbol-based BP
key technologies of next generation wireless communication detection with non-binary low-density parity check (LDPC)
systems, such as 5G, with its significant improvement in codes has been shown to outperform the classical bit-based
spectral efficiency, link reliability, and coverage compared to algorithm. However, its iterative detection is defined over high-
conventional MIMO [3]. Unfortunately, its huge size hinders order constellation in complex domain, and therefore results
the use of optimal data detection methods, such as maximum in overwhelming complexity in message passing. Furthermore,
likelihood (ML), since its computational complexity grows its inter- and intra-iteration scheduling will introduce very long
exponentially with the number of transmit antennas, which latency.
becomes prohibitive for large-scale MIMO. Even those near- In this paper, the BP detection defined over real domain is
optimal detections such as soft-output K-Best and sphere proposed for the first time to reduce the constellation size. Fur-
decoding (SD), which may be favorable for conventional thermore, the proposed method is generalized to symbol-based
MIMO, also suffers a lot from their excessive computational BP detection for high-order constellation, such as quadrature

U.S. Government work not protected by U.S. copyright

amplitude modulation (QAM) and so on. Simulation and com- where Ω is the set of in-phase or quadrature parts of points
parison have demonstrated its advantages of lower complexity in the √complex constellation Θ. From Eq. (4), we know that
and better performance over existing approaches. |Ω| = Q, which implies the √ constellation size of transmitted
The remainder of the paper is organized as follows. Section symbols reduce from Q to Q. As a consequence, a complex
II describes the system models over both complex and real N × M MIMO system with a complex constellation Θ can
domains of large-scale MIMO systems. Section III presents be equivalent to a real 2N × 2M MIMO system with a real
the FG representation of MIMO channels and theoretical foun- constellation Ω.
dation of BP detection based on FGs with GAI. In addition, Given r and H, the optimal maximum a posteriori proba-
the proposed low-complexity symbol-based BP detection in bility (MAP) detector selects si that maximizes the a posterior
reduced real constellation is also introduced in the same sec- likelihood:
tion. Numerical simulation results and complexity comparison ∑
p(xi |r, H) = p(x|r, H). (5)
are given in Section IV. Section V concludes the entire paper. x\xi

II. L ARGE -S CALE MIMO S YSTEM In other word, the MAP estimate of si can be given by
A. System Model in Complex Domain x̃i = arg max p(xi |y, H). (6)
xi ∈Ω
We consider the uplink in a large-scale MIMO system,
where the base station (BS) is equipped with N receiving III. BP D ETECTION BASED ON FG
antennas serving M (<N) single-antenna users at the same A. FG for MIMO Channels
time. The complex transmitted vector x is denoted by x̃ =
For LDPC decoders, factor graph is employed to represent
[x̃1 , x̃2 , . . . , x̃M ]T . Each entry of the vector is mapped to one
message passing between variable nodes and check nodes. The
point of a rectangular complex QAM, which is composed of
MIMO channel can also be represented by factor graph in a
Q = ∥Θ∥ = 2Mc distinct points. Therefore, we have x̃ ∈ ΘM .
similar manner. Eq. (2) reveals that the dependency between
According to the complex MIMO model, the N-dimensional
symbol nodes and observation nodes is determined by the
received vector r̃ = [r̃1 , r̃2 , . . . , r̃N ]T can be obtained by the
channel responses hi,j . Therefore, the fully connected FG of
following equation:
the MIMO channel can be illustrated in Figure 1 as follows.
r̃ = H̃x̃ + ñ, (1)
{ } symbol nodes
where H̃ = h̃i,j
i=1,...,N ;j=1,...,M
denotes the complex- x1 x2  x 2M

valued channel matrix, which is assumed to be independent a prior a posteriori

xi  r j r j  xi
and identically distributed (i.i.d) Rayleigh fading composed of information information
complex Gaussian distributed variable of ∼ CN (0, 1). And r1 r2  r2N
ñ = [ñ1 , ñ2 , . . . , ñN ]T represents the additive white Gaussian observation nodes
noise (AWGN) with ñi ∼ CN (0, σ 2 ).
Fig. 1. Factor graph of a MIMO system.
B. Equivalent Model in Real Domain
Based on the aforementioned complex channel model, the The nodes {x1 , x2 , . . . , x2M }, referred as “symbol nodes”,
equivalent real model can be derived using the real value are responsible for a prior information updating of the trans-
decomposition (RVD) scheme [13]. The real model of Eq. (1) mitted information. And the nodes {r1 , r2 , . . . , r2N }, referred
can be expressed as: as “observation nodes”, compute and store the a posteriori
information observed at the receiver side.
r = Hx + n, (2) In FG for MIMO channels, the dependency is determined by
T T
where r = [r1 , r2 , . . . , r2N ] , x = [x1 , x2 , . . . , x2M ] , and the channel response with statistically distribution. Therefore,
the 2N × 2M matrix H are all real-valued. They are derived the number of major connections would substantially not
according to the following mappings: become so large with the increase of antennas [14]. As a result,
the effective number of critical loops is not that large, which
r2k−1 =ℜ{r̃k },r2k =ℑ{r̃k }
r̃∈CN ================⇒ r∈R2N , makes BP detection promising for large-scale MIMO systems.
h2k−1,2m−1 =h2k,2m =ℜ{h̃k,m }
H̃∈CN ×M =====================⇒ H∈R2N ×2M , B. Symbol-Based BP Detection in Real Domain
h2k,2m−1 =−h2k−1,2m =ℑ{h̃k,m } (3)
x2k−1 =ℜ{x̃k },x2k =ℑ{x̃k } The essential of BP iterative detection lies in message
x̃∈CM =================⇒ x∈R2M , updating and passing between symbol nodes and observation
n2k−1 =ℜ{ñk },n2k =ℑ{ñk }
ñ∈CN =================⇒ n∈R2N , nodes, which can be summarized in two steps as follows:
• Step 1: Each symbol node updates its a prior information
where ℜ{∗} and ℑ{∗} denote the real and imaginary parts
according to the a posteriori information obtained from
of complex variable ∗, respectively. Note that for xi ∈ Ω we
observation nodes in each iteration but the first one. Then
have
√ √ the updated information is passed to all the connected
Ω = {(− Q + 1), . . . , −1, +1, . . . , (+ Q − 1)}, (4) observation nodes.
 ∑ ∑
∑ ( ) ( ) p(l) (rj |x, hj ) exp αt,j
(l−1)
(sk )
2M \i 2M \i
p (l)
(rj |x, hj ) p
(l−1)
x (l−1)
/p x0 x:xi =sk T:xt =sk
x:xi =sk t̸=i
(l)
βj,i (sk ) = log ∑ ( ) = log ∑ ∑ . (12)
2M \i (l−1)
p(l) (rj |x, hj ) p(l−1) (x2M \i ) /p(l−1) x0 p(l) (rj |x, hj ) exp αt,j (sk )
x:xi =s0 x:xi =s0 T:xt =s0
t̸=i

• Step 2: Each observation node computes its a posteriori p(l) (rj |s, hj ) can be referred as likelihood probability. Hence,
information based the a prior information obtained from the formula for calculating likelihood probability and a priori
its neighbouring symbol nodes, and then passes it back information is required. In FG-based BP detection approach
to those symbol nodes. with GAI, the received signal rj can be written as follows:
The message passed from the symbol node xi to observation
∑
2N
node rj , is denoted by Fxi →rj . And the message passed from rj = hj,i xi + hj,k xk +nj = hj,i xi + zj,i . (13)
the observation node rj to symbol node xi , is denoted by k=1,k̸=i
Lrj →xi . | {z }
interference
1) A Prior Information at Symbol Nodes: For each node,
soft information is used for message updating and passing. The Here, the interference term zj,i is modeled with Gaussian
a prior log-likelihood ratio (LLR) vector of symbol nodes in approximation of N (µzj,i , σz2j,i ) with
real constellation Ω is defined as follows: ∑
2M ∑
2M

(l) (l) µzj,i = hj,k E {xk }, σz2j,i = h2j,k Var {xk }+σ 2 . (14)
α(l)
xi →rj = [αi,j (s1 ), . . . , αi,j (s√Q−1 )], (7) k=1,k̸=i k=1,k̸=i

which denotes the message passing from Symbol Node i to Now we can calculate the likelihood probability with the
Observation Node j during the l-th iteration. Its entries are following equation:
are given by: (rj − µzj,i − hj,i xi )2
1
p(rj |x, hj )= √ exp{ }. (15)
(l)
(l)
p (xi = sk ) √ 2πσzj,i 2σz2j,i
αi,j (sk ) = ln (l)
, k = 1, . . . , Q − 1, (8)
p (xi = s0 )
Based on Eq. (12) and (15), we can further obtain Eq. (16).
where p(l) (xi = sk ) denotes the a prior probability that Using the approximation log(ex + ey ) ≈ max(x, y), Eq. (16)
the transmitted symbol xi is equal to sk ∈ Ω. Specially, can be approximated as Eq. (17). This algorithm is named as
for quadrature phase shift keying (QPSK) modulation with full edge based BP (FE-BP) algorithm. The message updating
(l) (l)
Ω = {−1, +1}, (7) can be rewritten as αxi →rj = αi,j (+1), and passing procedure of FE-BP algorithm is illustrated in
whose single entry is given by Figure 2.
 x( l )r  f2  r( l1)x , t  j 
 
(l) p(l) (xi = +1)
 
 r( l ) x  f1 px( l 1)r ,  x( l 1)r , t  i
αi,j (+1) = ln . (9)
p(l) (xi = −1) px( li )r j  f3  r(tl1x) i , t  j
i j t i

j i t j t j


, x i
p ( l 1)
,
x1 x1 r1 (l )

r1

(l 

r1
2) A Posteriori Information at Observation Nodes: Similar x1
(l 
(l )

r1  1)

p x i
(l )
rj x1


1) rj x r
1
x1 i

px( l21)r j ,  x( l21)r j  r( l ) x  r( l1)x px( li ) r2 ,  x( li ) r2

to symbol nodes, the message passed from observation node rj

2
x2 rj j 2
x2 r2
i
xi r2
rj to symbol node xi during the l-th iteration, representing the ( l  r j
p (l )
,
1)
(l  x i 

(l 1) x

  ,
i
)


xM rj rN
a posteriori information given y and H, is defined as follows: ( l  r j



1) rN (l
)

x

M x
i
p xM rN

(l) (l)
βr(l)
j →xi
= [βj,i (s1 ), ....βj,i (s√Q−1 )]. (10) x 2M x 2M r2N r2N

Its entries can be calculated as follows: Fig. 2. Message passing between variable nodes and observation nodes.

(l) p (xi = sk |rj , H)

(l)
To reduce the complexity of massage updating, the single-
βj,i (sk ) = log (l)
p (xi = s0 |rj , H) (l)
edge approach for updating βrj →xi is proposed in [15]. For
∑ (l−1) 2M \i (11)
s:xi =sk p (rj |x, hj )p
(l) the resulted single-edge based BP (SE-BP) algorithm, when
(x )
= ln ∑ (l−1) (x2M \i )
, (l)
updating the message βj,i (sk ) at l-th iteration, the a priori
s:xi =s0 p (rj |x, hj )p
(l)
information passing through its neighboring edges is ignored.
where x2M \i = (x1 , . . . , xi−1 , xi+1 , . . . , x2M ), and H =
(l−1)
As a result, αt,j (sk ) with light red color in Fig.2 can be
2M \i
(hT1 , . . . , hT2N )T . We assume x0 = {x1 = s0 , . . . , xi−1 = removed, and (17) can be simplified as follows:
s0 , xi+1 = s0 , . . . , x2M = s0 }. Based on Eq. (9), we can (l−1)
(rj −µzj,i −hj,i s0 )2 (rj −µzj,i −hj,i sk )2
(l−1)
(l)
rewrite Eq. (11) as shown in Eq. (12), where the transmitted βj,i (sk ) = −
2(σz2j,i )(l−1) 2(σz2j,i )(l−1)
symbols are assumed to be independent. (l−1) (18)
3) Message Updating of Observation Nodes: The ob- 2hj,i (rj −µzj,i )(sk −s0 )−h2j,i (s2k −s20 )
= .
servation nodes update βrj →xi according to (12), where 2(σz2j,i )(l−1)
 ( ( )2 )
∑ rj −µ(l−1) −hj,i xi ∑ (l−1)
exp −
zj,i
2(σz2 )(l−1)
exp αt,j (sk )
x:xi =sk j,i
T:xt =sk t̸=i
(l)
βj,i (sk ) = log ( ( )2 ) . (16)
∑ (l−1)
rj −µzj,i −hj,i xi ∑ (l−1)
exp − 2(σz2 )(l−1)
exp αt,j (sk )
x:xi =s0 j,i
T:xt =s0 t̸=i
 ( )2   ( )2 
(l−1) (l−1)
 ∑ (l−1) rj −µzj,i −hj,i xi   ∑ rj −µzj,i −hj,i xi 
βj,i (sk ) ≈ max   − max  .
(l) (l−1)
 αt,j (sk )− −   αt,j (sk )−  (17)
x:xi=sk
T:xt=sk
2
2(σz ) (l 1) x:xi=s0
j,i
2(σz2 )(l−1) T:xt =s0 j,i
t̸=i t̸=i

Specially, for QPSK modulation, Eq. (10) can be rewritten as Algorithm 1 Decision Process for Output
(l) (l)
βrj →xi = βj,i (+1), whose single entry is given by: Require: s = [s1 , ..., s√Q−1 , s0 ], γ = [γ1 , ..., γ√Q−1 , γ0 ].
1: [index, γmax = max{γ}
(l−1)
(l) 2hj,i (rj − µzj,i ) 2: if γmax > 0 then
βj,i (+1) = (l−1)
. (19)
(σz2j,i ) 3: x̂ = s(index)
4: else
5: x̂ = s(end)
4) Message Updating of Symbol Nodes: At symbol node
(l) 6: end if
side, we can update αxi →rj by

(l)
∑
2N
(l−1)
αi,j (sk ) = βt,i (sk ), (20) IV. S IMULATION R ESULTS AND C OMPARISON
t=1,t̸=j
A. Numerical Simulation Results
which is the standard BP massage updating method. Based on
For Rayleigh fading channel, the numerical results of the
Eq. (8) and (20), together with
√
proposed symbol-based BP detection in real domain are given.
∑
Q−1 Here, different antenna configurations and modulation modes
p(l) (xi = sk ) = 1, (21) are considered to offer a fair evaluation of the proposed
k=0 approach. There are a total of four plots here. Each plot is
the a prior probability can be easily obtained as follows: presented for a specific purpose, which will be detailed in
(l) the following. For all simulations, the number of BP detection
(l) exp(αi,j (sk ))
pi,j (xi = sk ) = √ , (22) iterations is set to 7. No channel coding scheme is considered.
∑
Q−1
(l) 1) Numerical Results with QPSK Modulation: Figure 3
1+ exp(αi,j (sm ))
m=1 shows the BER performances of proposed BP detection
√ in real constellation, general SE-BP detection in complex
with k = 1, ..., Q − 1. Considering the a posteriori infor-
mation updated in Eq. (14) and (16), we define the a priori constellation [12], and linear detection based on MMSE in
probability vector as follows: MIMO system, together with BER performance in single
input single output (SISO) channel with AWGN. According
(l) √ (l)
p(l)
xi →rj = [pi,j (s1 ), ....pi,j (s Q−1 )], (23) to Figure 3, we know that the proposed BP detection and
which should be included in a priori information Fxi →rj , general SE-BP detection, which suffer from the complexity of
(l) large dimensional inversion, all outperform MMSE detection
together with αxi →rj . Specially, for QPSK modulation, Eq.
(l) (l) significantly. Furthermore, the proposed BP detection shows
(23) can be simplified as pxi →rj = pi,j (+1), whose single better performance than the general SE-BP detection due to
entry is given by an effect of increased randomness of the connection strength
(l)
(l)
exp(αi,j (+1)) by increasing nodes in FG. For instance, at the BER of 10−2 ,
pi,j (+1) = (l)
. (24) the performance gap is around 1 dB.
1 + exp(αi,j (+1))
In Figure 4, the simulation results of proposed BP detection
in real constellation with difference antenna configurations
are given. The three configuration sets are: M = N = 16,
5) Decision Process for Output: Soft output after L itera-
32, and 64, respectively. It is observed that with the in-
tions is given by
crease of antennas’ number, the BP detection performance
∑
2N
(L)
√ also improves, approaching the SISO AWGN performance.
γ(sk ) = βt,i , k = 1, ..., Q − 1. (25) This perfectly matches the theoretical behavior of large-scale
t=1 MIMO systems. Thus, the proposed BP detection approach
The details of output decision process are listed as follows: is promising for large-scale MIMO system, especially when
 0 0
10 10

−1
10

−1
10 −2
10

−3
10
BER

BER
−2
10
−4
10

−5
−3 10
10
SISO,AWGN M=8,N=64,Cholesky decomposition
−6
M=N=16,MMSE 10 M=8,N=64,Approximate inversion, k=4
M=N=16,General SE−BP M=8,N=128,Cholesky decomposition
−4
M=N=16,Proposed BP −7 M=8,N=128,Approximate inversion, k=3
10 10
0 2 4 6 8 10 0 5 10 15
Average Received SNR(dB) Average Received SNR(dB)

Fig. 3. Comparison of different detection approaches (QPSK). Fig. 5. Exact and approximate MMSE detections (16-QAM).

0
10
in both cases when TSRR equals 16 and 2, respectively. It is
noted that the proposed BP detection is matrix inverse free,
which is very preferable for large-scale MIMO applications.
−1
10

0
10
BER

−2
10

−1
10

−3
10
SISO,AWGN
M=N=16,Proposed BP
BER

−2
10
M=N=32,Proposed BP
−4 M=N=64,Proposed BP
10
0 2 4 6 8 10
Average Received SNR(dB)
−3
10 SISO,AWGN
M=8,N=32,MMSE
Fig. 4. Proposed BP with different antenna configurations (QPSK). M=8,N=32,Proposed
M=32,N=64,MMSE
−4 M=32,N=64,Proposed BP
10
0 5 10 15
2
the antenna ratio M /N relatively large (see the following for Average Received SNR(dB)

more details).
2) Numerical Results with 16-QAM Modulation: It has Fig. 6. Performance of proposed BP with different TSRRs (16-QAM).
been mentioned in [6] that approximate matrix inversion
relying on Neumann series would reduce the complexity In conclusion, for large-scale MIMO systems, no matter
of linear detection based on MMSE for large-scale MIMO in low or high order modulation, no matter what antenna
systems. However, this approximation does not work when the configuration is employed, the proposed symbol-based BP
antenna ratio M 2 /N is high. Here, we define the Transmitting detection in real domain always outperforms MMSE detection.
antennas’ number Squared to Receiving antennas’ number Furthermore, this method is matrix inversion free. Those ad-
Ratio M 2 /N as TSRR. Illustrated in Figure 5, this approx- vantages make it very suitable for large-scale MIMO systems.
imation approach shows similar performance as the exact
one based on Cholesky decomposition, when TSRR is small B. Complexity Comparison
(M = 8, N = 128). However, when TSRR increases (M = 8, Here, the complexities of different BP detections are ana-
N = 64), even more approximation terms are employed, the lyzed. The complexity of the proposed BP detection involves a
performance is not satisfying. Here, the variable k denotes posteriori information updating at observation nodes (Eq. (14)
k-term Neumann approximation. and (18)), denoted by POST.UP. And a priori information up-
Luckily, the proposed BP detection approach is still effective dating at symbol nodes (Eq. (24)), denoted by PRI.UP. On the
when TSRR is relatively large (> 1). As illustrated in Figure 6, average, the complexity comparison of different BP detections
the proposed BP detection outperforms exact MMSE detection is listed in Table I. According to Table I, the complexity of
FE-BP grows exponentially with M , which is same as MAP with very large arrays,” IEEE Signal Process. Mag.,
detection. Intuitively, the proposed BP and the general SE- vol. 30, no. 1, pp. 40–60, Jan 2013.
BP are of the same magnitude of complexity. In fact, the [4] D. Wubben, D. Seethaler, J. Jalden, and G. Matz, “Lattice
general SE-BP is defined in complex domain, whereas the reduction,” IEEE Signal Process. Mag., vol. 28, no. 3, pp.
proposed BP is defined in real domain. Hence, the proposed 70–91, Mar. 2011.
BP is more preferable for hardware implementation due to its [5] C. Ju, J. Ma, C. Tian, and G. He, “VLSI implementation
lower complexity, especially for high-order (Q) constellation. of an 855 Mbps high performance soft-output K-Best
For example, for antenna configuration of M = 8 and N = 32 MIMO detector,” in Proc. of IEEE International Sym-
with 256-QAM, the proposed BP can achieve about 75% posium on Circuits and Systems (ISCAS), Seoul, Korea,
complexity reduction compared to SE-BP. May 2012, pp. 2849–2852.
[6] M. Wu, Y. Bei, A. Vosoughi et al., “Approximate matrix
TABLE I inversion for high-throughput data detection in the large-
C OMPLEXITY C OMPARISON OF D IFFERENT BP D ETECTIONS .
scale MIMO uplink,” in Proc. of IEEE International
BP algorithms POST.UP PRI.UP Symposium on Circuits and Systems (ISCAS), Beijing,
FE-BP O(QM +1 N (M − 1)) O(Q(N − 1)M ) China, May 2013, pp. 2155–2158.
General SE-BP O(QN (M − 1)) O(Q(N − 1)M ) [7] S. K. Mohammed, A. Chockalingam, and B. S. Rajan,
√ √
Proposed BP O(2 QN (2M − 2)) O(2 Q(2N − 2)M ) “A low-complexity near-ML performance achieving al-
gorithm for large MIMO detection,” in Proc. of IEEE
International Symposium on Information Theory (ISIT),
V. C ONCLUSION Toronto, Canada, Jul. 2008, pp. 2012–2016.
[8] N. Srinidhi, S. K. Mohammed, A. Chockalingam, and
In this paper, the symbol-based BP detection in real do- B. S. Rajan, “Low-complexity near-ML decoding of large
main is proposed. Simulation results with different modula- non-orthogonal STBCs using reactive tabu search,” in
tion modes and antenna configurations have demonstrated the Proc. of IEEE International Symposium on Information
performance advantage of the proposed approach compared Theory (ISIT), Seoul, Korea, Jul. 2009, pp. 1993–1997.
with the general BP detection as well as the MMSE approach. [9] M. Hansen, B. Hassibi, A. G. Dimakis, and W. Xu,
Also, the lower constellation size and matrix inverse free “Near-optimal detection in MIMO systems using Gibbs
property makes the proposed method computation-efficiency, sampling,” in Proc. of IEEE International Communica-
and therefore very suitable for large-scale MIMO systems. tions Conference (ICC), Honolulu, Hawaii, Dec. 2009,
Further work will be directed towards BP detection for MIMO pp. 1–6.
systems with more complicated antenna configurations and its [10] P. Som, T. Datta, A. Chockalingam, and S. S. Rajan,
cooperation with channel codes. “Improved large-MIMO detection based on damped be-
lief propagation,” in Proc. of IEEE Information Theory
ACKNOWLEDGEMENT
Workshop (ITW), Cairo, Egypt, Jan. 2010, pp. 1–5.
This work is partially supported by NSFC under grants [11] P. Som, T. Datta, N. Srinidhi, A. Chockalingam, and B. S.
61501116 and 61221002, International Science & Technology Rajan, “Low-complexity detection in large-dimension
Cooperation Program of China under grant 2014DFA11640, MIMO-ISI channels using graphical models,” IEEE Sig-
Huawei HIRP Flagship under grant YB201504, Jiangsu nal Process. Mag., vol. 5, no. 8, pp. 1932–4553, Dec.
Provincial NSF under grant BK20140636, Intel Collabora- 2011.
tive Research Institute for MNC, the Fundamental Research [12] A. Haroun, C. A. Nour, M. Arzel, and C. Jego, “Symbol-
Funds for the Central Universities under grants 3204004202, based BP detection for MIMO systems associated with
3204004102, and 3204005101, the Research Fund of National non-binary LDPC codes,” in Proc. of IEEE Wireless
Mobile Communications Research Laboratory, Southeast Uni- Communications and Networking Conference (WCNC),
versity under grant 2014B02, and the Project Sponsored by Istanbul, Turkey, Apr. 2014, pp. 212–217.
the SRF for the Returned Overseas Chinese Scholars of State [13] S. Haykin, M. Sellathurai, Y. de Jong, and T. Willink,
Education Ministry. “Turbo-MIMO for wireless communications ,” IEEE
Commun. Mag., vol. 42, no. 10, pp. 48–53, Oct. 2004.
R EFERENCES [14] W. Fukuda, T. Abiko, T. Nishimura et al., “Low-
[1] J. L. J.-K. Han and J. Zhang, “MIMO technologies complexity detection based on belief propagation in a
in 3GPP LTE and LTE-advanced,” EURASIP J. Wirel. massive MIMO system,” in Proc. of IEEE Vehicular
Commun. Netw., vol. 2009, no. 3, pp. 1–10, Mar. 2009. Technology Conference Spring (VTC-Spring), Dresden,
[2] I. Pefkianakis, L. Suk-Bok, and L. Songwu, “Towards Germany, Jun. 2013, pp. 1–5.
MIMO-aware 802.11n rate adaptation,” IEEE/ACM Tran- [15] F. Long, T. Lv, R. Cao, and H. Gao, “Single edge based
s. Netw., vol. 21, no. 3, pp. 692–705, Jun. 2013. belief propagation algorithms for MIMO detection,” in
[3] F. Rusek, D. Persson, B. K. Lau, E. Larsson, T. Marzetta Proc. of IEEE Sarnoff Symposium, Princeton, USA, May
et al., “Scaling up MIMO: opportunities and challenges 2011, pp. 1–5.