A Reduced-Complexity Sphere Decoding Algorithm for MIMO Systems

YING-J UN NI1 , MING-QI LI1 , WEN-QIANG GUO2

1. School of Mathematical Sciences, University of Electronic Science and Technology of Ch ina, Chengdu 611731, China
2. School of Co mputer Science and Engineering, Xin jiang University of Finance & Economics , Uru mqi 830012 , China
E-MAIL: handsome219@163.co m

Abstract: 0.5 N  0.5 N 1

S phere decoding (S D), as an algorithm searching for the 0 (0.5N )
e  d   1   (2)
closest lattice point within a certain search radius, is a
corrected maximum likelihood (ML) algorithm for multiple-
input multiple-output (MIMO) systems, with low complexity. where N is the problem dimension and  2 is the noise
The sphere radius and the number of search floating points variance, () is gamma function, 1   is a probability
visited throughout the tree traversal searching are the decisive close to 1, 1    0.99 . If the closest point is not found, we
factors for the complexity of the algorithm. In this paper, a
new reduce-complexity detection algorithm based on the can increase the probability 1   , adjust the radius, and
complex S D for MIMO systems is presented. The algorithm search again. Although its average complexity is
performs a fixed number of total search floating points to polynomial [3][4], the co mplexity depends on the channel
detect the p-QAM modulation symbols independently on the condition and the noise level.
noise level. Simulation results show that with 4  4 system via
4-QAM and 16-QAM, the new SD algorithm has a lower How to decrease the complexity of sphere decoder is
complexity than the conventional SD at low signal -noise ratio always a hot point for researchers . The increased radius
(SNR) especially. search (IRS) algorithm was proposed in [1]. An imp roved
increasing radius reach (IIRS) algorith m was suggested in
Keywords: [5], wh ich improved SD co mp lexity efficiency by making a
MIMO systems; p-QAM modulation; sphere decoding; good choice of the sphere radius, and tried to reduce the
ML decoding; computational complexity
number of candidates in the search space. In [6], a new
lattice representation was proposed which enabled the SD
1. Introduction
algorith m to decode the real and imag inary parts of each
The use of multip le-input mult iple-output technology complex symbol independently. But it decomposed the
has become the new frontier of wireless communication. It symbols real and imaginary two parts. In [7], by co mb ining
uses mult ipath propagations and space-time coding to get a novel channel matrix preprocessing with a small subset
high rate data transmission and improve overall bit error search, a fixed number of operations was achieved, which
ratio (BER) in co mmunication link. The optimu m receiver decreased the complexity greatly. We find that all the search
for MIMO systems is ML detector, but it is exponentially radiuses of these complexity reduction techniques for SD
complex. Especially, it is unsuitable in practical systems algorith m depended on the noise level greatly. At high noise
with large number of antennas and higher order level, the search radius must be large to ensure no empty
constellations. Sphere decoding has been proposed, search in the sphere, which will increase the complexity.
providing ML performance with reduced comp lexity [1][2]. There are other methods, such as a combination of the SD
A sphere decoder searches for the closest lattice point and the K-best decoder in [8] was applied to MIMO
within a certain search radius. To ensure there is a systems, while the co mplexity reduction is at the cost of
non-empty sphere, the search radius should not be too small, performance. And in [9] the modificat ions of the SD
and it should not be too large to ensure not increase the algorith m reduced greatly the co mplexity wh ile required
number o f search floating points to a great extent. There additional operations. In [10], a new defin ition for sphere
exists a experiential fo rmula presented in [3] for the radius, radius was proposed with changing the symbol search
which is strategy, which reduced computational co mplexity, wh ile
required more hardware resources.
d 2   N 2 , (1)
In [11], a method for evaluating the comp lexity of SD
algorith ms not only in terms of nu mber of visited nodes, but

978-1-4244-8026-5/ 10/$26.00 ©2010 IEEE 431

Authorized licensed use limited to: Ingrid Nurtanio. Downloaded on February 27,2021 at 15:02:15 UTC from IEEE Xplore. Restrictions apply.
also in terms of total number of operations , has been xˆ  arg{ min
2
y  Rx  d 2 } , y   Q H y
presented. In this paper, we imp rove the SD co mp lexity x p QAM

efficiency by reducing the number of search floating points. (6)

The method, based on the conventional complex SD
where H is the conjugate transpose representation.
algorith m, has a fixed nu mber of point operations and
ensures there is no empty point. However, there is a trade
off between the performance and the complexity. It is close
According to the formulat ion
to the performance of conventional sphere decoding
algorith m, with lo wer co mp lexity. N N

 | y   r x
i 1
i
j i
ij j |2  d 2 , (7)
2. Conventional S phere Decoding Algorithm
we can search the closest point xi ( i  1, 2,..., N ) in the
The MIMO system in this paper has N transmit and M
receive antennas. The transmitted symbols are taken hyper-sphere of radius d around the received signal y . The
independently fro m QAM constellation of p points solution of (6) or (7) can be obtained recursively, starting
( p  QAM ). The received signal, using matrix notation, is fro m i  N to i  1 . The algorith m is shown as follow.
given by Step 1: i  N , fix on xN that satisfied
y  Hx  n (3)
| yN  rNN xN |2  d 2 ,
where x  ( x1 , x2 ,..., xN ) denotes the vector of transmitted
T
using  N represents the number of xN .
symbols. n  (n1 , n2 ,..., nM )T is the vector of independent and
identically distributed (i.i.d.) co mplex Gaussian noise, Step 2: i  N  1 to 1, fix on xi that satisfied
which samples with variance 2 , and N
y  ( y1 , y2 ,..., yM ) is the vector of received symbols. H
T
| yi|i 1  rii xi |2  d i2 , yi|i 1  yi  rx
j  i 1
ij j ,
denotes the M  N channel matrix, where hij is the comp lex
transfer function fro m transmitter j to receiver i. The entries di2  di21  | yi1|i  2  ri 1,i 1 xi 1 |2 ,
of H are modeled as i.i.d. Rayleigh fad ing and estimated
perfectly at the receiver. and using  i represents the number of xi .

2.1. Conventi onal S phere Decodi ng Algorithm If there is not xi that satisfied (6), increase d 2 and
restart from i  N to i  1 until each level having response.
The ML detection is carried out as Then search the closest point in {x1 , x2 ,..., xN } , where xi is
2 the subset of the p  QAM constellation, as the result.
xˆ  arg min y  Hx (4)
x p QAM N
Denote the total search floating points by    i .
It must search the whole p  QAM constellation i 1

space. To reduce the complexity of the ML, in [1] the

formula (4) is replaced by 3. A Deduced-Complexity S phere Decodi ng Algori thm
2
xˆ  arg{ min y  Hx  d } . 2
(5)
x p QAM For the conventional SD, the search radius gets from
the formulas (1) and (2). Obviously it has large rad ius at
Then the searching points only lie in a hyper-sphere of high noise level, with increasing the computational
radius d around the received signal y . complexity. A new method, wh ich can get the search subset
at each dimension independently on the noise level, is
Considering N  M in the system, H  QR , where
proposed here. Analyzing (7), we can write as
R  (rij ) N  N is an upper triangular mat rix and Q is unitary
matrix. Then (5) can be written as

432

Authorized licensed use limited to: Ingrid Nurtanio. Downloaded on February 27,2021 at 15:02:15 UTC from IEEE Xplore. Restrictions apply.
 N reflects on the computational co mplexity of the algorith m.
| y1   r1 j x j |2 ...
j 1
(8) The most important part of the proposed algorith m is
N the search radius is independent on the noise level, and the
 | yN 1  
j  N 1
rN 1, j x j |2  | yN  rNN xN |2  d 2 Nth dimension xN determines the i  N  1 to 1 dimension
xi of the transmitted signals, and there is no worried about
Step 1: for j  1 to p gaining empty to restart fro m N to 1 or getting much points
to increase the computational complexity.
xN  p  QAM ( j) , | yN  rNN xN |2  d(2j ) ,
next j 4. Si mulation
p
d 2   d (2j ) p , In the simulat ions, consider the MIMO systems
j 1
with N M 2 and N M 4 ,
save xN , which satisfied using p  4 and p  16 QAM modulat ion schemes. The
| yN  rNN xN |2  d 2 and cN2 | yN  rNN xN |2 . average symbol energy Es is 2. The entries of channel
matrix H are modeled as i.i.d. Rayleigh fading. The
Because of the arith metic mean value d 2 , we can get
variance of co mplex Gaussian noise is  2  4 Es10  SNR 10 .
corresponding xN without empty or whole constellation. We
The transmit symbols are processed via space-time coding,
denote the number of survivors at the end of this step by  N .
liking [12]. For ML decoding, we use the formu la (4), and
for the conventional SD algorithm, we get the initial search
Step 2: for every survivor xN , calcu late xN 1 ,..., x2 , x1
radius d fro m (1) and (2), with 1    0.99 .
recursively as shown below
for i  N  1 to 1
j  1 to p
N
xi  p  QAM ( j ) , ci2 | yi  r
k  i 1
ik xk  rii xi |2 ,

next j
save the minimal ci2 , and corresponding xi , and

ci2  ci21  ci2

next i .

Because the number of xi is determined by the number

of xN , there is no worried about empty xi . Then we denote Figure 1. the BER performance of M L, conventional SD
the number of xi by i . The search subsets {xi } are obtained and new SD with N  M  2 using 4-QAM
by the above two steps.

Step 3: choose c12  min{c12 } as an init ial value, and

search in the set of xN , xN 1 ,...,x2 ,x1 , to make the least value
to be the detected message (If c12 is the least value, the
corresponding x1 , x2 ,..., xN are the detected).

According to the above algorithm, the total number of

search floating points of candidates x1 , x2 ,..., xN , whose
N
Euclidean distance is calculated, is    i , which
i 1

433

Authorized licensed use limited to: Ingrid Nurtanio. Downloaded on February 27,2021 at 15:02:15 UTC from IEEE Xplore. Restrictions apply.
Figure 2. the BER performance of conventional SD and 5. Conclusion
new SD with N  M  4 using 4-QAM
Sphere decoding algorith m is an effect ive algorithm for
decoding transmitted signals in MIMO co mmunications. In
this paper, we proposed a new algorithm of SD. According
to [11], noting the total nu mber of search floating points
evaluates the complexity of new SD algorith m. The
performance and complexity of the proposed algorithm has
been obtained via simulat ions for different modulation
schemes and MIMO systems. The main aim is to show its
low co mplexity and its suitability as a receive detector.
Simu lation results show that the new algorith m has lower
complexity than conventional sphere decoding and the
performance is closed to the conventional SD’s.

Acknowledgements
Figure 3. the search floating points of M L, conventional SD
and new SD with N  M  4 using 4-QAM This paper is partly supported by the NSFC(60902074)
and the Fundamental Research Funds for the Central
Universities (ZYGX2009J101)

References

[1] E. Viterbo and J. Boutros, “A Universal Lattice Code

Decoder for Fading Channels ”. IEEE Trans. Inform.
Theory, vol. 45(5), pp. 1639-1642, Ju ly 1999.
[2] M. O. Damen, H. E. Gamal, and G. Caire, “on
Maximu m Likelihood Detection and the Search for the
Closest Lattice Point”. IEEE Trans. Inform. Theory, vol.
49(10), pp. 2389-2402, Oct. 2003.
[3] B. Hassibi, and H. Vikalo. “on the Sphere-Decoding
Figure 4. the search floating points of M L, conventional SD Algorith m: I. Expected Co mplexity”. IEEE Trans.
and new SD with N  M  4 using 16-QAM Signal Processing, vol. 53(8), pp. 2806-2818, August
2005.
According to Figure 1 and Figure 2, we can see the [4] B. Hassibi and H. Vikalo, “on the Expected Co mp lexity
new SD algorith m’s performance is the same as of Sphere Decoding.” Proc.35th Asilo mar Conference
conventional SD and ML for MIM O system N  M  2 , on Signals, Systems and Co mputers, Monterey, CA, pp.
however, there is a performance loss for system N  M  4 , 1051- 1055, Nov. 2001
using 4-QAM. Figure 3 and Figure 4 show that the total [5] W. Zhao and G. Giannakis, “Sphere Decoding Algori-
number of search floating points  of new SD algorithm is thms with Imp roved Radius Search”. IEEE Trans.
fixed, and in low SNR, it is far small than  which is the Co mmun. vol. 53(7), pp. 1104-1109, Ju ly 2005.
conventional SD’s, for N M 4 , [6] L. Azzam and E. Ayanoglu, “Reduced Co mp lexity
using p  4 and p  16 QAM. Sphere Decoding fo r Square QAM v ia a New Lattice
Representation.” IEEE Global Telecommun.
We can see that a trade-off exists between the Conference, pp. 4242–4246, Nov. 2007.
complexity and the performance of the proposed SD at low [7] L. G. Barbero and J. S. Tho mpson, “A
2 Fixed-Co mplexity M IMO Detector Based on the
SNR. If we make d in step 1 bigger, more candidates are
Co mplex Sphere Decoder.” IEEE Workshop on Signal
searched and the performance will be closer to that of the Processing Advances in Wireless Communications, 2006
conventional SD, but the computational complexity will
[8] J. Tang, A. Tewfik, and K. K. Parh i, “Reduced Co m-
increase.
plexity Sphere Decoding and Application to Interfering

434

Authorized licensed use limited to: Ingrid Nurtanio. Downloaded on February 27,2021 at 15:02:15 UTC from IEEE Xplore. Restrictions apply.
 IEEE 802.15.3a piconets.” IEEE Inter. Conference. on Sphere Decoding Algorith ms.” Proceedings of
Co mmunicat ion, Minnesota, vol. 5, pp. 2864–2868, ISWCS’08, pp. 438-442, 2008.
June 2004. [11] J. Fink, S. Roger and A. Gon zalez, “Co mp lexity
[9] L. Brunel, “Multiuser Detection Techniques Using Assessment of Sphere Decoding Methods for MIMO
Maximu m-Likelihood Sphere Decoding in Mult icarrier Detection.” IEEE ISSPIT2009, pp. 9-14, 2009.
CDMA systems.” IEEE Trans. Wireless Commun. vol. [12] M. O. Damen, A. Ch keif, and J. C. Belfiore, “Lattice
3(3), pp. 949-957, May 2004. code decoder for space-time codes.” IEEE Co mm.
[10] R. S. Yazdi and T. Kwasniewski, “Lo w Co mp lexity Letters, vol. 4(5)pp. 161-163, May 2000.

435

Authorized licensed use limited to: Ingrid Nurtanio. Downloaded on February 27,2021 at 15:02:15 UTC from IEEE Xplore. Restrictions apply.