Design and Implementation of Variable Radius Sphere Decoding Algorithm

Jan Zizka et al. (Eds) : CCSEIT, MoWiN, IT, AIAP, ICBB - 2015
pp. 31–36, 2015. © CS & IT-CSCP 2015 DOI : 10.5121/csit.2015.51104
DESIGN AND IMPLEMENTATION OF
VARIABLE RADIUS SPHERE DECODING
ALGORITHM
Wu Di, Li Dezhi and Wang Zhenyong
School of Electronics and Information Engineering,
Harbin Institute of Technology, Harbin, China
wudi45_2012@163.com, lidezhi@hit.edu.cn, ZYWang@hit.edu.cn
ABSTRACT
Sphere Decoding (SD) algorithm is an implement decoding algorithm based on Zero Forcing
(ZF) algorithm in the real number field. The classical SD algorithm is famous for its
outstanding Bit Error Rate (BER) performance and decoding strategy. The algorithm gets its
maximum likelihood solution by recursive shrinking the searching radius gradually. However, it
is too complicated to use the method of shrinking the searching radius in ground
communication system. This paper proposed a Variable Radius Sphere Decoding (VR-SD)
algorithm based on ZF algorithm in order to simplify the complex searching steps. We prove the
advantages of VR-SD algorithm by analyzing from the derivation of mathematical formulas and
the simulation of the BER performance between SD and VR-SD algorithm.
KEYWORDS
MIMO, Signal Detecting, Maximum Likelihood Solution, Sphere Decoding, Variable Radius
1. INTRODUCTION
As the communication demand growing constantly, researchers have a stricter requirement for
transform rate and BER performance of signal detection. Maximum Likelihood (ML) algorithm
has the best BER performance among the MIMO signal detecting algorithms. The ML algorithm
searches the feasible signal location in all possible areas, so it is seldom used in practical project.
SD algorithm is an optimization algorithm of ML algorithm. The main strategy of SD algorithm
is searching the feasible signal location within the super-sphere volume, when the radius is R.
After finished one circulation, change the numerical value of R and do the same thing with the
new super-sphere, whose radius is new R. Finally find the signal located nearer than any other
signals. The classical SD algorithm has two categories. One is called the SD algorithm based on
breadth first search, the other is based on depth first search. The results of both search algorithms
make no differences.
The main deficiency of classical SD algorithm is that the complexity is related to the original
value of the convergence radius. If the original value of R is too small, there will be no signal
location in the super-sphere area, and the system will have to re-choose the original value of R
again. If we try to increase R in some algorithms, it is necessary to take the step size into
consideration. If the original value of R is too large, the system will take all the possible signal
location into consideration, which makes the algorithm complexity comparable with ML’s. In

32 Computer Science & Information Technology (CS & IT)
addition, if the convergence algorithm of R does not suit the situation, the computational
complexity will make no difference with ML’s. In that way, the SD algorithm may lose its
significance of existence. In conclusion, whether the original value of the convergence radius is
too large or too small, it will waste a lot of operands before it gets a suitable radius.
Considering all the analysis above, it demonstrates that the original value of the convergence
radius has a great influence on the performance of classical SD algorithm. An excellent
convergence strategy can greatly decrease the computational complexity of classical SD
algorithm. An algorithm is proposed in this paper to calculate the original value of convergence
radius to avoid a too large or small original value of the convergence radius. Besides, this paper
proposed the VR-SD algorithm to find the target signal location after calculating the original
convergence radius.
2. AN IMPROVED STRATEGY OF SD ALGORITHM
In this paper, the Multiple Input Multiple Output (MIMO) system deploys M antennas to transmit
the signal, and N antennas to receive the signal (M<N). Firstly, preprocess the channel state
information (CSI) matrix H to a square matrix. If M=N, the CSI matrix H will be a square matrix
without pre-processing. This paper only discuss the M=N system.
In the assumption, the CSI matrix H can be transformed into real number by the functions below:
Re( ), Im( )
Im( ),Re( )
H H
H
H H
− 
=  
 
The size of H is 2N×2N.
If X and Y are respectively transmit and receive matrix in complex number field, the transmit and
receive matrix in real number field can be transformed in the following functions:
Re( )
Im( )
X
X
X
 
=  
 
Re( )
Im( )
Y
Y
Y
 
=  
 
If Z is the noise matrix, and Z is the noise matrix in real number field, the equation Y HX Z= +
still exists. The target of SD algorithm is to find the maximum likelihood solution in the
convergence sphere, which is expressed as follows.
( ) ( )
2 2ˆ ˆˆ arg min
T
T
SD
x
x y Hx x x H H x x R
∈Λ
= − = − − ≤ (1)
SDR is the convergence radius. Let the center of the super-sphere be ˆx :
1ˆ ( )H H
x H H H y−
= (2)

Computer Science & Information Technology (CS & IT) 33
QR decomposes the matrix H , and the R matrix is an upper triangular matrix in real number field.
The equation can be transformed as following:
( ) ( ) ( )
2
2
1 11,1 1,2 1,3 1,2
2,2 2 2
3,2 3 3
2 ,2
2 2
2 2
2 ,2 2 2 2 1,2 1 2 1 2 1 2 1,2 2 2
ˆ ˆ ˆ
ˆ
ˆ0
ˆ0 0
0 0 0
0 0 0 0 ˆ
ˆ ˆ ˆ( ) ( ) ( )
T
T
N
N
N
N N
N N
N N N N N N N N N N N N
x x H H x x R x x
x xr r r r
r x x
r x x
r x x
r x x r x x r x x− − − − −
− − = −
 − 
  
−  
  = −  
  
  
   − 
= − + − + − + ≤
L
O L L
O L g
O M M
L 2
SDR
(3)
In the functions below, 0SDR stands for the original radius of convergence. In VR-SD algorithm,
0SDR is not set by empirical value but a serious of calculation. The function of 2 2 1 1( , , , )N Nf x x x− L is
set to be:
2 2
2 2 1 1 2 ,2 2 2 2 1,2 1 2 1 2 1 2 1,2 2 2
ˆ ˆ ˆ( , , , ) ( ) ( ) ( )N N N N N N N N N N N N N Nf x x x r x x r x x r x x− − − − − −= − + − + − +L L (4)

3. THE EXPERIMENTS OF VR-SD ALGORITHM
This paper compares the performance between classical SD algorithm and VR-SD algorithm in
2×2 MIMO system using QPSK modulation. Knowing from the aforementioned analysis, if the
convergence circulation is free from loop times, the performance of classical SD algorithm will
be the same as ML algorithm’s. The circulation of classical SD algorithm may not keep doing
until it finds the final target signal, because the CSI keeps changing during the process of
implementation. If the time overhead of signal detection algorithm is too large, its CSI may
change before it finishes the calculation. Furthermore, since the signal is being transferred all the
time, if the system wastes too much time in signal detection, the subsequent signal will jam
together. Considering all the situation above, the times of convergence circulation of classical SD
algorithm will be constant during the simulation in this paper. After the times of convergence
circulation reach the setting number, the system will choose a signal in the super-sphere of the
last circulation as the target signal randomly, and force the algorithm to exit the circulation.
To get the convergence time of different convergence circulations of classical SD algorithm, this
paper did the simulation of classical SD algorithm within the limited convergence times of 4, 5, 6
and 7. On the condition of the same hardware and the same version of MATLAB software,
compare the convergence time between the limited times classical SD algorithm and the unknown
convergence times VR-SD algorithm. The simulation is done in 2×2 MIMO system using QPSK
modulation with the 220
code length. As the variance of each kind of simulation is small, we do
each kind of simulation ten times and calculate the average time. The results are shown in table 1:
Table 1. Calculation time of VR-SD and classical SD
Algorithm Calculation time (s) Convergence times
VR-SD 136.2 N/A
Classical SD
120.7 4
142.1 5
148.5 6
150.3 7
By comparing the convergence time of classical SD algorithm, it is concluded that the calculation
time rises with the rising convergence times. The calculating time difference between 4 and 5
times is nearly 22 seconds, but the time difference between 5 and 6 times sharply decreases to just
over 6 seconds. The time difference between 6 and 7 times is even less. This shows that with the
increasing convergence times, the percent of classical SD algorithm finishes their calculation
before reaching the limit convergence times is increasing. From Table 1, the calculating time of
VR-SD algorithm is close to the classical SD algorithm when the convergence times is 5. In the
following simulation, we compare the BER between VR-SD and the classical SD algorithm with
5 convergence times.

Computer Science & Information Technology (CS & IT) 35
-5 0 5 10
10
-4
10
-3
10
-2
10
-1
10
0
SNR(dB)
BER
VR-SD
ZF
Classical SD
Figure 1. Comparison of BER between VR-SD, classical SD and ZF
Figure 1 shows the BER performance of VR-SD compared with classical SD algorithm and ZF
algorithm. The computational complexity of ZF is much lighter than VR-SD algorithm, while the
BER of VR-SD and classical SD is much better than ZF’s. Because of the limitation of
convergence times, the BER of classical SD algorithm doesn’t have the same performance as the
one which without limitation. Though the limitation of convergence times decreases the
computational complexity, it still has some bad influence on its BER. Considering this, it is not
strange that the BER of classical SD is not as well as the VR-SD’s. We still cannot ignore the fact
that if the times of convergence circulation keep rising, the BER of classical SD algorithm will be
better than VR-SD.
4. CONCLUSIONS
This paper solve the problem of the algorithm complexity of convergence radius by designing a
SD algorithm strategy—VR-SD algorithm. The VR-SD algorithm can avoid the unnecessary
computational complexity caused by an unsuitable value of original convergence radius.
Subjected to the same BER, both the classical SD and the VR-SD algorithm have a nearly 3 dB
gain from the ZF algorithm. While the gain between the classical SD algorithm and the VR-SD
algorithm is about 0.5dB. Considering the time overhead of classical SD algorithm, which
convergence times is 5, is 4% more than the VR-SD algorithm’s. The gain of 0.5dB from
classical SD to VR-SD is a really great improvement. It demonstrates that the use of variable
radius strategy raises the BER of SD algorithm in a considerable degree.
ACKNOWLEDGEMENTS
The paper is sponsored by “the Fundamental Research Funds for the Central University” (Grant
No.HIT.NSRIF.2015024).
REFERENCES
[1] Vikalo H, Hassibi B & Kailath T, (2004) “Iterative decoding for MIMO channels via modified sphere
decoding”, IEEE Transactions on Wireless Communications, Vol. 3, No. 6, pp2299-2311.
[2] Lampe L, Schober R & Pauli V, (2004) “Multiple-symbol differential sphere decoding”, IEEE
Transactions on Communications, Vol. 53, No. 12, pp1981-1985.
[3] Ding Y, Wang Y R & Li N, (2014) “Widely linear sphere decoding by exploiting the hidden
properties of PSK signals”, IEEE Globecom 2014- Signal Processing for Communications
Symposium (GC14 SPC).

[4] Soltani. M.D., Alimadadi. M & Amindavar, H, (2014) “A hybrid method to reduce the complexity of
k-best sphere decoding algorithm”, 2014 22nd Iranian Conference on Electrical Engineering (ICEE),
pp1765-1770.
[5] El-Khamy. M, Vikalo. H., Hassibi, B & McEliece, R.J, (2009) “Performance of sphere decoding of
block codes”, IEEE Transactions on Communications, Vol. 57, No. 10, pp2940-2950.
[6] Viet-Hoa Nguyen, Berder. O. & Scalart, P, (2013) “On the efficiency of sphere decoding for linearly
precoded MIMO systems”, Wireless Communications and Networking Conference (WCNC), 2013
IEEE, pp4021-4025.
[7] Li Shi-ping, Wang Long & Chen Fang-chao, (2012) “Ordered sphere decoding detection algorithm
for MIMO systems”, 2012 24th Chinese Control and Decision Conference (CCDC), pp3322-3325.
[8] Xu Ren & Huahua Wang, (2012) “Research of Sphere Decoding Detection Algorithm in LTE-A
System”, 2012 8th International Conference on Wireless Communications, Networking and Mobile
Computing (WiCOM), pp1-4.
[9] Singh. A, Elia. P & Jalden. J, (2013) “Rate-reliability-complexity tradeoff for ML and lattice
decoding of full-rate codes”, 2013 IEEE International Symposium on Information Theory
Proceedings (ISIT), pp1267-1271.
AUTHORS
Wu Di was born in 1990, and get his B.S degree in Harbin Institute of Technology in
2013.He is now getting his M.S degree in the same school. His studying field is about the
wireless communication and the signal detecting of MIMO.
Li Dezhi was born in 1981. He has received his PhD degree in Harbin Institute of
Technology. He is the lecturer in school of electronics and information Engineering. His
main research fields are satellite communication, wireless multimedia transmission and
wireless communication.
Wang Zhenyong received his B.S. and M.S. degrees in communication and information
system from Harbin Institute of Technology in 2000 and 2002, respectively. He is the
lecturer in school of electronics and information Engineering. His main research interests
are cross-layer design, satellite communication, wireless multimedia transmission, and
wireless heterogeneous network

Design and Implementation of Variable Radius Sphere Decoding Algorithm

More Related Content

What's hot (18)

Viewers also liked (16)

Similar to Design and Implementation of Variable Radius Sphere Decoding Algorithm (20)

Recently uploaded (20)

Design and Implementation of Variable Radius Sphere Decoding Algorithm