PROCEEDINGS OF SPIE

SPIEDigitalLibrary.org/conference-proceedings-of-spie

An improved detection algorithm for

massive MIMO system

Du Pan, YiMing Yu, XiangChen Ma, SongTao Gao,

YanHong Jiao, et al.

Du Pan, YiMing Yu, XiangChen Ma, SongTao Gao, YanHong Jiao, HaiTao
Zhang, "An improved detection algorithm for massive MIMO system," Proc.
SPIE 12331, International Conference on Mechanisms and Robotics (ICMAR
2022), 1233105 (10 November 2022); doi: 10.1117/12.2652691

Event: International Conference on Mechanisms and Robotics (ICMAR 2022),

2022, Zhuhai, China

Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 12 Feb 2023 Terms of Use: https://www.spiedigitallibrary.org/terms-of-use

 An Improved Detection Algorithm for Massive MIMO System
Du Pan1,a*, YiMing Yu1,b, XiangChen Ma1,c, SongTao Gao1,d, YanHong Jiao1,e, HaiTao Zhang1,f
1
Radio Research Agency, China Mobile Group Design Institute Co., Ltd., Beijing, China
*a
[email protected], [email protected],
c
[email protected], [email protected],
e
[email protected], [email protected]

Abstract

The massive multiple-input multiple-output (MIMO) system is one of the most important key technologies in 5G era and
will play an important role in the future telecommunication development. In massive MIMO signal detection, a huge
number of matrix calculations need to be completed, especially matrix inversion. This makes the massive MIMO
detection algorithm computationally complex and consumes lots of resources. Firstly, this paper briefly introduces and
analyzes a variety of traditional iterative algorithms. Based on simplifying the inversion of MMSE matrix by solving
linear equations, an improved algorithm is proposed to reduce the number of iterations by optimizing the initial solution
to achieve the same detection performance as the traditional algorithms. The computational complexity of the proposed
algorithm keep consistently at 𝑂(K 2 ), which is one order of magnitude lower than the traditional ones.
Keywords-massive MIMO; Signal detection; Iterative algorithms; Optimized initial solution

1. INTRODUCTION
As one of the core technologies of 5G, massive MIMO technology is equipped with dozens or even hundreds of antenna
arrays at the base station to serve multiple single-antenna user equipment(UE) at the same time, which can not only
significantly improve the spectrum efficiency of the system, but also the energy efficiency [1].
Since massive MIMO has a large number of antennas on the base station side greatly increasing the difficulty of detecting
signals, how to ensure the efficient detection performance of the system, and high accuracy of recover signals while
reducing the computational complexity as much as possible, are key issues that needs to be solved urgently in the fifth
generation mobile communication research [2] .
In recent years, in order to balance the relationship between signal detection performance and computational complexity,
many researchers have successively proposed a variety of different matrix inversion simplification algorithms based on the
traditional linear detection algorithm MMSE. These algorithms can be divided into four categories: Signal detection
algorithm based on approximate matrix inversion [3], signal detection algorithm based on iterative algorithm to solve
linear equation [4], detection algorithm based on gradient search approximate solution [5] and signal detection algorithm
based on approximate information transmission [6].
From the perspective of improving the detection performance of the algorithm, relying on the signal detection algorithm
based on solving linear equations, this paper proposes an optimized initial solution for the iterative algorithm. In order to
verify the improvement of detection performance by the proposed solution, the paper applies it to the Jacobi and Gauss-
Seidel two classical algorithms. The simulation results show that the optimized initial solution can achieve a good balance
between detection performance and computational complexity. In addition, the optimized initial solution proposed in this
paper can be applied to a variety of iterative algorithms, with a wide range of application scenarios.

2. SYSTEM MODEL
For the uplink massive MIMO system, the system model is shown in Fig. 1 [7]. Assuming that the MIMO system consists
of K transmitting antennas and N receiving antennas, K N. If the base station and the single-antenna UE are fully
synchronized, all users in the uplink use the same time-frequency resource to send data to the base station, which is full
frequency multiplexing technology.

International Conference on Mechanisms and Robotics (ICMAR 2022),

edited by Zeguang Pei, Proc. of SPIE Vol. 12331, 1233105
© 2022 SPIE · 0277-786X · doi: 10.1117/12.2652691

Proc. of SPIE Vol. 12331 1233105-1

Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 12 Feb 2023
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Figure 1. System model of Massive MIMO
Then the signal received by the base station can be expressed as:
𝐲 = 𝐇𝐬 + 𝐧 (1)
where 𝐲 is the (𝑁 × 1) dimensional received signal matrix at base station, 𝐬 is the (𝐾 × 1) dimensional transmitted signal
matrix at UE, 𝐧 denotes the (𝑁 × 1) additive white Gaussian noise vector with 0 mean and variance s 2. 𝐇 is (𝑁 × 𝐾)
dimensional the channel matrix, which equals to:
h11 … h1K
H=[ ⋮ ⋱ ⋮ ] (2)
hN1 ⋯ hNK
Where h ji represents the channel fading coefficient between the i-th transmitting antenna and the j-th receiving antenna. In
the case of a flat fading channel and the antenna distance is far enough, h ji is not correlated with each other [8]. In
addition, 𝐲=[y1 , y2 , … yN ]T ,yi is the signal received by the i-th receiving antenna. s=[s1 , s2 , … sK ]T , si represents the signal
sent by the i-th transmitting antenna, namely desired signal sent from the i-th user.
2.1 MMSE Detection Algorithm
In the uplink massive MIMO system, the MMSE detection algorithm can achieve a near-optimal detection performance,
that is, the original transmitted signal s can be recovered from the received signal y at the receiver through a relatively
simple linear operation. In order to recover the transmitted signal, a weighted filter matrix A is used to achieve channel
inversion. Then the estimated value of the transmitted signal can be written as:
𝒔̂ = 𝑨𝒚 = 𝑨(𝑯𝒔 + 𝒏) (3)
The weighted filter matrix of the MMSE detection algorithm is as follows [9]:
𝑁0 −1 𝐻
𝑨𝑀𝑀𝑆𝐸 = (𝑯𝐻 𝑯 + 𝑰) 𝑯 (4)
𝐸𝑠
The detection and restoration signal under the above detection matrix can be written as:
𝑁0 −1 𝐻
𝒔̂𝑀𝑀𝑆𝐸 = 𝑨𝑀𝑀𝑆𝐸 𝒚 = (𝑯𝐻 𝑯 + 𝑰) 𝑯 𝒚 (5)
𝐸𝑠
2
where 𝑁0 represents the variance of independently and identically distributed Gaussian noise , 𝐸𝑠 is the average
transmission power of each symbol. Although the MMSE detection algorithm minimizes the noise interference while
ensures the signal being well received, it involves complex matrix inversion operations, which increases the difficulty of
hardware implementation and limits the application of the algorithm to a certain extent.

Proc. of SPIE Vol. 12331 1233105-2

Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 12 Feb 2023
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 2.2 Low Complexity Detection Algorithm
If the signal detection algorithm formula 𝒔̂ = 𝑾−1 𝒚𝑀𝐹 written in the form of 𝑾𝒔̂ = 𝒚𝑀𝐹 , then the complex matrix
inversion problem is transformed into a linear equation solving problem.
The iterative algorithm based on linear equation solving is using the symmetric positive definite characteristic of the
MMSE filter matrix to directly estimate the transmission vector, hence avoiding the process of inversion of the high-
dimensional matrix. In a massive MIMO system, as the number of base station antennas increases, when the number of
base station antennas is far greater than the number of users (𝑁 >> 𝐾), the column vectors of the uplink channel matrix
H are asymptotically orthogonal and the MMSE filtering matrix is a symmetric positive definite matrix. Therefore, 𝑾can
be decomposed into:
𝑾=𝐃−𝐋−𝐔 (6)
D, L and U are the diagonal element matrix, strictly lower triangular element matrix and strictly upper triangular element
matrix of 𝑾 respectively.
2.3 Jacobi Algorithm
The Jacobi iterative algorithm has the slowest convergence rate among the common iterative algorithms. So its detection
performance is worse than the other iterative algorithms. The estimated value of the received signal after signal detection
and restoration through the Jacobi algorithm is:
𝒔̂(𝑛) = 𝑫−1 ((𝑼 + 𝑳)𝒔̂(𝑛−1) + 𝒚𝑀𝐹 (7)
Jacobi iterative matrix can be expressed as 𝑩𝑗𝑎 = 𝑫−1 (𝑼 + 𝑳)，then the convergence condition of this iterative algorithm
is ρ(𝑩𝑗𝑎 ) < 1. Equation (7) can be written in the form of elements as follows:
𝑖−1 2𝐾
(𝑛) 1 (𝑛−1) (𝑛−1)
𝒔̂𝑖 = ̂𝑖 − ∑ 𝑾𝑖,𝑗 𝒔̂𝑖
(𝒚 − ∑ 𝑾𝑖,𝑗 𝒔̂𝑖 ) (8)
𝑾𝑖,𝑗
𝑗=1 𝑗=2

2.4 Gauss-Seidel Algorithm

(𝑛)
In the Jacobi iterative algorithm, the latest estimated value is not used in the current iteration process. Even 𝑠̂𝑖 has
(𝑛) (𝑛) (𝑛)
already got, 𝑠̂𝑖 is used to estimate ̂𝑠𝑖+1 . That means, if the latest estimated value 𝑠̂𝑖 can be used in iterative process, a
faster convergence speed can be obtained. Based on this concept, Gauss-Seidel iterative algorithm is proposed:
𝑖−1 2𝐾
(𝑛) 1 (𝑛) (𝑛−1)
𝒔̂𝑖 = ̂𝑖 − ∑ 𝑾𝑖,𝑗 𝒔̂𝑖 − ∑ 𝑾𝑖,𝑗 𝒔̂𝑖
(𝒚 ) (9)
𝑾𝑖,𝑗
𝑗=1 𝑗=2

The estimated value of the received signal after signal detection and restoration using the Gauss-Seidel iterative algorithm
is:
𝒔̂(𝑛) = (𝑫 − 𝑳)−1 (𝑼𝒔̂(𝑛+1) + 𝒚𝑀𝐹 ) (10)
Using 𝑩𝐺𝑆 = (𝑫 − 𝑳)−1 𝑼 represents the iterative matrix of Gauss-Seidel, when ρ(𝑩𝐺𝑆 ) < 1, Gauss-Seidel algorithm
converged. The Gauss-Seidel iterative algorithm is widely used in practical applications due to its simplicity. However,
there is an obvious shortcoming of the algorithm: when ρ(𝑩𝐺𝑆 ) getting close to 1, the convergence speed of the
algorithm slows down.

3. THE PROPOSED DETECTION ALGORITHM

In this section, this paper analyzes the effect of the initial solution for iterative algorithm firstly. Then based on a widely
common used diagonal matrix initial solution, the paper proposed an improved initial solution for detection algorithm
which reaches the same detection performance as the previous ones but with fewer iterations.

Proc. of SPIE Vol. 12331 1233105-3

Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 12 Feb 2023
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Step 1. Analyze the Effect of the Initial Solution
For each kind of iterative algorithm, although the initial solution will not affect the convergence result of the algorithm, it
plays an important role in affecting the converge speed of the algorithm. Similarly, when the number of iterations is
limited or small, initial solution will greatly affect the computational complexity and detection accuracy of the algorithm
[10]. Therefore, using a suitable initial solution is very important for the iterative detection algorithm.
In the current massive MIMO detection, the iterative detection algorithm based on solving linear equations usually sets the
initial solution to 2𝐾 × 1 dimension zero vector. This method is simple and easy to implement, but it requires a long
iterative process. 𝒔̂(0) = 𝑫−1 𝒚𝑀𝐹 is another often used diagonal matrix initial solution.
Step 2. Design the Optimized Initial Solution
This paper proposes an innovative initial solution on this basis. The optimized initial solution has a certain improvement in
the detection performance and reaches the same detection performance as the zero vector with fewer iterations. The
optimized initial solution is expressed as:
𝑾𝑖𝑗 , |𝑖 − 𝑗| ≤ 𝛼
𝑻𝑖𝑗 = { (11)
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
where 𝑻𝑎 = (𝑻𝑖𝑗 ), 𝑻𝑎 is a banded matrix in which all non-zero elements fall in a band-area centered on the diagonal line.
In detail, banded matrix 𝑻𝑎 is composed of 𝑾𝑖𝑗 a square matrix’s main diagonal and its upper and lower 𝛼 elements with
2𝛼 + 1 . 𝑾𝑖𝑗 can be decomposed into:
𝑾𝑖𝑗 = 𝑻𝛼 -𝑳𝛼 -𝑼𝛼 (12)
𝑳𝛼 and 𝑼𝛼 are strictly lower and upper triangular element matrix of 𝑾𝑖𝑗 . When 𝛼 =1, the proposed optimized initial
solution is the initial solution of the diagonal matrix. So 𝒔̂(0) = 𝑫−1 𝒚𝑀𝐹 can be regarded as a special case of the optimized
initial solution.
Step 3. Analyze the Computational Complexity
This paper analyzes the computational complexity of the iterative algorithm with optimized initial solution according to
the number of multiplications required in the algorithm. Since all MMSE and iterative based detection algorithms need to
realize the calculation of 𝑾𝑀𝑀𝑆𝐸 and 𝒚𝑀𝐹 . Therefore, iterative algorithm computational complexity analysis using
different initial solution mainly focus on the following two parts:
1) Initial solution computational complexity
𝒚𝑀𝐹 is a 𝐾 × 1dimension column vector. 𝒔̂(0) = 𝑻𝑎 −1 𝒚𝑀𝐹 needs to be calculated 2𝐾(2α − 1) − 2αtimes. Since the
optimal bandwidth parameter is usually small 𝛼𝑜𝑝𝑡 ≤ 3 which is quite different from the number of single-antenna users
in massive MIMO,2α can be ignored. When 𝛼 = 1,computational complexity of diagonal matrix 𝒔̂(0) = 𝑫−1 𝒚𝑀𝐹 is 2𝐾 .
2) Iterative algorithm computational complexity
Jacobi and Gauss-Seidel algorithm have the same computational complexity, equal to 4nK 2, which is the lowest among
all kinds of iterative algorithm.
TABLE I. COMPUTATIONAL COMPLEXITY OF ALGORITHMS

Zero Diagonal Optimized

vector matrix banded matrix

Jacobi / 4nK 2
4nK 2 4nK 2 2K
Gauss-Seidel 2K(2α − 1)

4. SIMULATION RESULTS
This article uses the MATLAB simulation platform to verify the impact of the optimized initial solution on the detection
performance. The wireless transmission channel in the simulation is set to the fast fading Rayleigh channel, the antenna

Proc. of SPIE Vol. 12331 1233105-4

Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 12 Feb 2023
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 size is configured to 128 16 , the base-band signal modulation method is 16QAM, and the noise is Gaussian white Noise.
The bit error rate (BER) is used as the standard to measure the detection performance, and the curve of the MMSE
algorithm is introduced as the reference benchmark.
This paper first applies the optimized initial solution to the most widely used Gauss-Seidel algorithm, and then applies it to
the Jacobi algorithm with relatively poor detection performance, then observes the relationship between the algorithm's bit
error rate and the signal-to-noise ratio after being optimized.
In this paper, "Jacobi /GS" is used to denote the algorithm that uses the zero vector as the initial solution. "Jacobi /GS-
diag" is the algorithm which uses the main diagonal matrix 𝒔̂(0) = 𝑫−1 𝒚𝑀𝐹 as the initial solution. "GSnew =2/3/4"
means the Gauss-Seidel algorithm using the optimized initial solution of the corresponding bandwidth parameter. "Jacobi
=2/3" is the Jacobi algorithm using the optimized initial solution.
In order to determine the best selection of the optimal parameter , the paper simulate three different values of under
different iteration times. According to the simulation results, when =2, the optimized detection algorithm improves the
bit error rate performance the most.

Figure 2. Twice iteration with different

As shown in Fig. 2, when iterates twice, the detection performance of the optimized algorithm with =3 and =4 is
almost same, but slightly worse than =2. Compared with the Gauss-Seidel algorithm that uses the main diagonal matrix
𝒔̂(0) = 𝑫−1 𝒚𝑀𝐹 as the initial solution, the optimized initial solution has a certain improvement in detection performance.
Taking Fig. 3 as an example, when the signal-to-noise ratio is 10dB, the bit error rate of the GS algorithm using the main
diagonal matrix as the initial solution is less than 10-4. However, the bit error rate of the GS algorithm with optimized
initial solution bandwidth parameter =2 has already exceeded 10-4 and is close to the bit error rate performance curve of
the MMSE. Therefore, it can be concluded that under this simulation configuration,𝛼𝑜𝑝𝑡 = 2is the optimal bandwidth
parameter of the GS iterative algorithm to optimize the initial solution.

Proc. of SPIE Vol. 12331 1233105-5

Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 12 Feb 2023
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Figure 3. Three times iteration with different
Fig. 4 shows the comparison of the detection performance among the traditional GS algorithm, the GS algorithm that uses the
main diagonal matrix as the initial solution, and the GS algorithm that uses the proposed optimized initial solution. Set the
bandwidth parameter to 𝛼𝑜𝑝𝑡 = 2. it can be clearly seen from the figure that the curve with the best detection performance is
the GS algorithm that uses the optimized initial solution (when n=3). As the number of iterations increases, the detection
performance of the three GS algorithms is improved. It is worth noticing that the GS algorithm that uses the optimized initial
solution with twice iterations can achieve the same bit error rate performance with the traditional GS algorithm with three
times iterations. It means the proposed optimized GS algorithm reduces the computational complexity of the algorithm to a
certain extent.

Figure 4. BER performance of GS with different initial solution

In the simulation of Jacobi algorithm, referring to the optimal parameter in GS, select =2 and =3 for simulation
verification. Fig. 5 shows the simulation results of bit error rate with twice iterations. It shows that the Jacobi algorithm
with optimized initial solution has a significant improvement in detection performance compared with the traditional
Jacobi algorithm, and even realizes the detection performance close to the Gauss Seidel algorithm. The traditional Jacobi
algorithm has just reached the bit error rate of 10-2 when SNR = 12dB, while the Jacobi algorithm with optimized initial
solution has exceeded 10-3. To reach the bit error rate of 10-2, the Jacobi algorithm with optimized initial solution only
needs 6dB signal-to-noise ratio. It can be seen that the optimal bandwidth parameter of the optimized initial solution of

Proc. of SPIE Vol. 12331 1233105-6

Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 12 Feb 2023
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Jacobi algorithm is =3 not =2. When =3, the detection performance of the optimized Jacobi algorithm with SNR
6dB to 12dB exceeds that of the traditional GS algorithm. Therefore, it can be concluded that for Jacobi algorithm, the
optimal bandwidth parameter is 𝛼𝑜𝑝𝑡 =3. Besides, the optimized initial solution improves the detection performance of
Jacobi algorithm better at low SNR.

Figure 5. BER performance of Jacobi with different when n=2

Fig. 6 shows the simulation result when iteration is 3. When SNR is lower than 8dB, the curve of the optimized Jacobi
algorithm with =3 is almost the same as that of GS. As in the case of twice iterations, the detection performance of
the optimized Jacobi algorithm with =3 is better than that of =2. When SNR<10dB, the two curves of =2/3 are
closer to each other.

Figure 6. BER performance of Jacobi with different when n=3

5. CONCLUSION
This paper proposed an optimized initial solution of the inverse algorithm of MMSE detection matrix based on the
iterative algorithm, which improves the detection performance and is suitable for more than one iterative algorithms. The
iterative algorithm using the optimal initial solution can iterate less times and converge quickly. At the same time, the
optimized initial solution proposed in this paper can set with different optimal bandwidth parameters according to

Proc. of SPIE Vol. 12331 1233105-7

Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 12 Feb 2023
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 different iterative algorithms to achieve a better detection performance. So the work of this paper can help saving
computing resources with great performance in massive MIMO signal detection. In the future, more scenario applications
of the proposed algorithm will be studied.

REFERENCES

[1] T. L. Marzetta, "Noncooperative Cellular Wireless with Unlimited Numbers of Base Station Antennas," IEEE
Transactions on Wireless Communications, 2010, 9(11): 3590-3600.
[2] M. Wu, B. Yin, G. Wang, C. Dick, J. R. Cavallaro and C. Studer, "Large-Scale MIMO Detection for 3GPP LTE:
Algorithms and FPGA Implementations," IEEE Journal of Selected Topics in Signal Processing, 2014, 8(5):
916-929.
[3] Z. Chen, F. Sohrabi and W. Yu, "Multi-Cell Sparse Activity Detection for Massive Random Access: Massive
MIMO Versus Cooperative MIMO," IEEE Transactions on Wireless Communications, 2019, 18(8): 4060-4074.
[4] M. A. Albreem, M. Juntti and S. Shahabuddin, "Massive MIMO Detection Techniques: A Survey," IEEE
Communications Surveys & Tutorials, 2019, 21(4): 3109-3132.
[5] Keke Zu, Jun Zhu and Martin Haardt, “Uplink Multi-user MIMO Detection via Parallel Access,” 2019 IEEE
International Conference on Acoustics, Speech and Signal Processing, 2019, pp. 4365-4369.
[6] S. Yang and L. Hanzo, "Fifty Years of MIMO Detection: The Road to Large-Scale MIMOs," IEEE
Communications Surveys & Tutorials, 2015, 17(4): 1941-1988.
[7] E. G. Larsson, O. Edfors, F. Tufvesson and T. L. Marzetta, "Massive MIMO for next generation wireless
systems," IEEE Communications Magazine, 2014, 52(2): 186-195.
[8] Zanella A, Chiani M, Win M Z. “Analytical evaluation of MIMO systems with unequal power transmission in a
Rayleigh fading environment,” IEEE GLOBECOM, 2003:1837-1841.
[9] W. Song, X. Chen, L. Wang and X. Lu, "Joint conjugate gradient and Jacobi iteration based low complexity
precoding for massive MIMO systems," 2016 IEEE/CIC International Conference on Communications in China
(ICCC), Chengdu, 2016, pp. 1-5.
[10] Lei Liu, Yuhao Chi, Chau Yuen, Yong Liang Guan and Ying Li, “Capacity-Achieving MIMO-NOMA: Iterative
LMMSE Detection,” IEEE Transactions on Signal Processing, 2019, 67(7): 1758-1773.

Proc. of SPIE Vol. 12331 1233105-8

Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 12 Feb 2023
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use