Professional Documents
Culture Documents
Advisor(s)
Author(s)
Shao, Ziyun.; .
Citation
Issued Date
URL
Rights
2011
http://hdl.handle.net/10722/174460
SHAO, Ziyun
B.Sc.(Eng), M.Sc.(Telecom)
at the
Department of Electrical and Electronic Engineering
The University of Hong Kong
in
October 2011
DECLARATION
I hereby declare that this thesis represents my own work, except where due
acknowledge is made, and that it has not been previously included in a thesis,
dissertation or report submitted to this University or to any other institution for a
degree, diploma or other qualifications.
Signed
SHAO, Ziyun
October 2011
The increasing demand for high-mobility and high data rate in wireless
communications results in constraints and problems in the limited radio spectrum,
multipath fading, and delay spread.
The multiple-input multiple-output (MIMO) system has been generally
considered as one of the key technologies for the next generation wireless
communication systems. MIMO systems which utilize multiple antennas in both
the transmit side and the receive side can overcome the abovementioned
challenges since they are able to increase the channel capacity and the spectrum
usage efficiency without the need for additional channel bandwidth.
The detection algorithm is a big bottleneck in MIMO systems. Generally, it is
expected to fulfill two main goals simultaneously: low computational complexity
and good error rate performance. However, the existing detection algorithms are
either too complicated or suffering from very bad error-rate performance.
The purpose of this thesis is to comprehensively investigate the detection
algorithms of MIMO systems, and based on that, to develop new methods which
can reduce the computational complexity while retain good system performance.
Firstly, the background and the principle of MIMO systems and the previous work
on the MIMO decoding algorithms conducted by other researchers are thoroughly
reviewed. Secondly, the geometrical analysis of the signal detection is
investigated, and a geometric decoding algorithm which can offer the optimum
BLER performance is proposed. Thirdly, the semidefinite relaxation (SDR)
detection algorithms are extended to high-order modulation MIMO systems, and a
novel SDR detector for 256-QAM constellations is proposed. The theoretical
analysis on the tightness and the complexity are conducted. It demonstrates that
the proposed SDR detector can offer better BLER performance, while its
complexity is in between those of its two counterparts. Fourthly, we combine the
SDR detection algorithms with the sphere decoding. This is helpful for reducing
the computational complexity of the traditional sphere decoding since shorter
initial radius of the hyper sphere can be obtained. Finally, the novel
lattice-reduction-aided SDR detectors are proposed. They can provide
near-optimum error rate performance and achieve the full diversity gain with very
little computational complexity added compared with the stand-alone SDR
detectors.
II
ACKNOWLEDGEMENT
First of all, I would like to express my sincere gratitude to my supervisors, Dr.
S. W. Cheung and Dr. T. I. Yuk for their precious guidance and persistent
encouragement throughout my entire PhD study. They taught me academic
knowledge and research skills and enlightened my passion to explore the
unknown scientific world. The thesis would not have been completed without
their supports.
I would also like to thank Mr. Eric W.L. Ng, Ms. Julie Hung and Ms. Lily Lo
in the Department of Electrical and Electronic Engineering for their kind help
during the past few years.
I truly appreciate the friendship of all teammates and friends in HKU for their
kind help, advice, guidance and encouragement, most notably Dr. Z. Zhang, Dr.
M. X. Xiao, Dr. F. Mai, Dr. X. G. Dai, Dr. W. Zhou, Dr. Z. Kong, Dr. K. C.
Leung, Mr. Y. F. Weng, Mr. Y. Y. Sun, Miss M. J. Mao, Miss L. Li, Mr. H. L.
Xiahou, and Mr. Z. B. Ni. In particular, many thanks to Dr. Dai who always had
taken seriously every question I asked him.
Finally, I am most grateful to my parents and my husband. Their selfless love,
continuous supports and encouragements throughout all these years are the most
precious thing to me. Without these, I could never get my work done well.
III
CONTENTS
ABSTRACT
ACKNOWLEDGMENTS
III
CONTENTS
IV
CHAPTER 1
INTRODUCTION
1.1
1.2
Systems..........................................................................................8
1.3
Literature Review................................................................9
1.4
1.5
CHAPTER 2
ALGORITHMS
2.1
Introduction .......................................................................17
2.2
2.3
2.4
2.5
2.6
Summary ...........................................................................34
CHAPTER 3
GEOMETRIC DETECTION
ALGORITHMS
IV
3.1
Introduction .......................................................................35
3.2
Channels ......................................................................................35
3.3
3.4
3.5
Summary ...........................................................................52
CHAPTER 4
4.1
Introduction .......................................................................54
4.2
4.3
Semidefinite Relaxation....................................................59
4.4
Semidefinite
Relaxation
Detection
Algorithms
for
Semidefinite
Relaxation
Detection
Algorithms
for
4.7
Summary ...........................................................................93
CHAPTER 5
LATTICE-REDUCTION-AIDED
SEMIDEFINITE RELAXATION
DETECTION ALGORITHMS
V
5.1
Introduction .......................................................................94
5.2
5.3
5.4
5.5
CHAPTER 6
CONCLUSIONS AND
RECOMMENDATIONS
6.1
6.2
LIST OF FIGURES
121
LIST OF TABLES
124
ABBREVIATIONS
125
REFERENCES
127
PUBLICATIONS
138
VI
Chapter 1 Introduction
CHAPTER 1
INTRODUCTION
of
the
wireless
communications
technologies.
From
the
Chapter 1 Introduction
The MIMO technology has gone through a long history. It was firstly
proposed for application in wireless communication systems in the 1970s. In 1993,
Indian scientists Paulraj and Kailath introduced the idea of using spatial
multiplexing (SM) in MIMO system [1]. Since 1990s, the researchers in AT&T
Bell Lab have given a huge boost on MIMO technology. In 1995, Telatar showed
that the capacity of the MIMO systems in the fading channel conditions increases
linearly with the number of the transmit antennas and the receive antennas [2]. In
1996, Foschini proposed a diagonally-bell laboratories layered space-time
(D-BLAST) architecture for MIMO systems [3]. In 1998, Golden and other
researchers [4] built the laboratorial platform of MIMO system by using
vertical-bell laboratories layered space-time (V-BLAST) algorithm, where the
spectral efficiencies could reach 20-40bit/s/Hz at the indoor fading rates.
Up till now, MIMO technology has been widely considered as one of the key
technologies of the next generation wireless communication systems [5]. Some
mobile communications standards such as the 3G syetems, long-term-evolution
(LTE) and 4G have included the MIMO technology. The standard of wireless local
area network (WLAN) 802.11n recommends MIMO combined with orthogonal
frequency-division multiplexing (MIMO-OFDM). In many other wireless
communication research fields, such as ultra-wide-band (UWB) system and
cognitive radio (CR), researchers are considering to take MIMO technology into
consideration. Therefore, MIMO system is a promising solution to future wireless
communications and has become a very hot issue in both the academic and the
industrial fields.
Chapter 1 Introduction
Chapter 1 Introduction
x2*
x1*
(1.1)
The column vectors of the matrix are orthogonal to each other, and sent by
different antennas during each time slot. At the receiver, the received signal is
separated by linear transformation and then decoded by maximum likelihood
decoding.
Chapter 1 Introduction
without in need of additional bandwidth and power. In the SM-MIMO system, the
data stream is split into several sub-streams, then modulated and transmitted by
different antennas simultaneously. With no doubt, this process could result in a
gain in data rate. The D-BLAST [3] is the first structure using MIMO
multiplexing. The data streams are multiplexed diagonally and transmitted in the
period of transmitting a block of signal. Then another V-BLAST [4, 7] structure is
proposed, which is more effective. The data streams are transmitted in parallel,
that is, the i th data symbol is transmitted by the i th antenna directly. The
BLAST structure can provide high multiplexing gain, and the capacity of such
system is increased linearly with the number of the transmit antennas and the
receive antennas [2]. In this thesis, we will focus on the SM-MIMO system.
wireless channel. At the transmit side, the user data stream is partitioned into NT
sub-streams and then sent by different transmit antennas. At the receive side, each
receive antenna receives signal vectors from all the transmit antennas.
Chapter 1 Introduction
h12
h22
hN R 2
h1NT x1 n1
h2 NT x2 n2
+
hN R NT xNT nNT
(1.2)
or
r = Hx + n
(1.3)
matrix, with elements hij representing the transfer function from the j-th transmit
antenna to i-th receive antenna. n is the N R -dimensional additive noise. In this
thesis, the signal vector x is assumed to be a statistically independent variable
with zero mean and unit variance x 2 = 1 . Perfect channel knowledge is assumed
to be known to the receiver. In addition, the channel matrix H is assumed to be a
flat fading channel and all the entries in H are complex Gaussian and
independent. The noise is an independently and identically distributed (i.i.d.)
Chapter 1 Introduction
=
+
Im ( r ) Im ( H ) Re ( H ) Im ( x ) Im ( n )
(1.4)
with Re() and Im() being the real and imaginary parts of (), respectively. The
real-valued representation is written as [5]:
r = Hx + n
(1.5)
(1.6)
HH H , N R < NT
Q= H
H H , NT < N R
(1.7)
Chapter 1 Introduction
(1.8)
It implies that ML calculates the Euclidean distance between the possible transmit
signal vectors and the received signal vector, and then choose the one which is
closest to the received vector as the solution. However, since all the possible
signal vector in the lattice space should be considered, and their Euclidean
Chapter 1 Introduction
distances away from the received vector have to be calculated, the complexity of
ML decoding increases exponentially with the number of transmit antennas and
polynomially with the size of the constellation [5]. Thus, in an SM-MIMO system,
the complexity of ML decoding is non-deterministic polynomial time hard
(NP-hard). This disadvantage makes the system impractical to be implemented.
Thus, detection has become one of the major challenges in MIMO system.
The main objective of designing a detection algorithm is that it provides good
error rate performance with low computational complexity. The key point is how
to balance the performance and complexity.
Chapter 1 Introduction
10
Chapter 1 Introduction
branch until a node is encountered. For all other branches, the same process is
conducted until all nodes within the hyper sphere are discovered.
It has been shown that the S-E enumeration is more computational efficient
than the F-P enumeration [20]. The S-E strategy performs traversing the tree
depth-first too. But it calculates the branch weights and searches them in
increasing order. And after it obtains a lattice point, the radius of the hyper sphere
is reduced to be the distance between the received signal point and the lattice
point. Then the search process is restarted again with the new radius. As a result,
the lattice points visited is less and the searching process becomes faster.
Figure 1.4
Sphere decoding.
There are two major aspects to be improved for the sphere decoding. The first
one is to determine the initial radius of the hyper sphere. If the radius is too large,
there will be too many node weights need to be calculated. In the contrast, if it is
too small, it would be possible that no point is inside the hyper sphere, and the
search should be restarted again with a larger radius. Thus, a proper initial radius
11
Chapter 1 Introduction
of the hyper sphere can reduce the complexity of SD. In [14] and [21], it uses the
noise variance and probability equation to define a proper initial radius. The
MMSE equalizer is also applied to set the initial radius [22]. Another aspect is the
searching strategy. In [23], a statistical pruning method that uses a set of bounds
based on the minimum metric of the current solution is proposed for S-E sphere
decoding. A preprocessing stage and a new ordering are engaged in the searching
method in [24]. With the new ordering, the nodes are expanded according to the
level and offset coefficients. These searching strategies provide higher
computational efficiency. Although SD is able to provide the BLER performance
of ML detection with less complexity, it has been proven that its expected
complexity is still exponential [25]. Thus it becomes impractical when the system
order is high and the SNR is low.
The successive interference cancellation detection [88] has the error rate
performance gain by sacrificing a certain complexity compared with the linear
detections. It detects the signal from the first transmit antenna instead of that of all
the transmit antennas, and then subtract its impact from the received signal vector.
After that, it detects the signal from the second transmit antenna. At this time, the
row of the channel matrix that is corresponding to the first antenna is deleted.
Thus, the dimension of channel matrix becomes M R ( M T 1) and the
dimension of the transmitted signal vector is reduced to M T 1 . The process
continues until all the elements of the signal vector are detected. Because the
detected signal has effect on the later signal to be detected, we need to firstly
detect the reliable signals in order to reduce the error propagations. Usually,
ordering is adopted to improve the performance. The authors in [10] proposed an
12
Chapter 1 Introduction
13
Chapter 1 Introduction
introduced in [33]. It can reduce the average complexity by almost half of the
conventional LLL and achieve full diversity in LR detection.
Recently, MIMO decoders using semidefinite relaxation approach have
attracted great attention. They are able to provide acceptable BLER performance
and feature polynomial worst-case complexity [34]. Since the ML decoding
problem has the optimum error rate performance, the SDR detection applies the
convex optimization toolbox such as SEDUMI [35] to solve the convex relaxation
of ML optimization problem, which is called semidefinite programming (SDP).
The SDR approach was firstly applied to detect binary phase-shift keying
(BPSK) and four quadrature amplitude modulation (4-QAM) signals [36]-[38]. It
has been shown that the SDR detector for BPSK can achieve full receive diversity
[39]. Then the extensions to different SDR techniques for 16-QAM signals had
been
proposed,
such
as
polynomial-inspired
SDR
(PI-SDR)
[40],
14
Chapter 1 Introduction
which can achieve the optimum BLER performance while retain acceptable
computational complexity.
is able to achieve the full diversity gain and improve the error rate performance
with a little complexity added.
15
Chapter 1 Introduction
16
CHAPTER 2
STATE-OF-THE-ART MIMO DETECTION
ALGORITHMS
2.1 Introduction
This chapter will be devoted to review the state-of-the-art of the MIMO
detection algorithms. Firstly, the linear decoders which include the zero-forcing
decoder and the minimum mean-square-error decoder are introduced. Linear
decoder has the advantage of extremely low computational complexity. However,
they suffer from the unsatisfactory block error rate (BLER) performance.
Secondly, the sphere decoding is elaborated. Similar to the ML decoding, sphere
decoding is another searching-based detector, and it can offer optimum BLER
performance. Nevertheless, its expected complexity is still very high, although it
has been dramatically reduced compared with the ML decoder. Thirdly, the
successive interference cancellation and the lattice reduction detection are
introduced. They both can be combined with the linear decoders, so as to improve
their BLER performance with certain complexity added.
17
decoders are called linear decoders. The well-known linear decoders include the
zero-forcing (ZF) decoder and the minimum-mean-square-error (MMSE) decoder.
2.2.1 Zero-Forcing
Zero forcing (ZF) detection [8] is the simplest linear detection algorithm. It
forces the impact of the channel matrix to be zero and is given by:
1
x ZF = Q ( H*r ) = Q ( H H H ) H H r
(2.1)
x MMSE = Q ( H H H + n 2 I ) H H r
1
(2.2)
18
where n 2 denotes the noise power and I is the identity matrix. Compared with
the equation (2.1) of ZF, the difference lies in the term of n 2 I . ZF decoding
separates the co-channel signals and cancels all the inter-symbol-interference (ISI).
However, this inevitably leads to noise enhancement. On the other hand, MMSE
detection attempts to minimize the overall errors which are caused by the noise,
and make balance between the ISI mitigation and noise enhancement. Assuming
that the noise n 2 is zero, the MMSE becomes the same with ZF. Generally,
MMSE decoding tends to provide better error rate performance than ZF decoding.
Both ZF and MMSE are classified as linear detection algorithm. Although
they have very low complexity, their error rate performances are still far away
from being satisfactory.
19
searched exhaustively, the calculations are based on the branches in a tree which
are possible to lead to the final result.
There are two main problems to be solved in SD. One is how to determine the
initial radius. If the radius is too large, there will be a large number of points
inside the hyper sphere and the complexity will be too large. If the radius is too
small, there may be no point inside the sphere and the search process has to be
restarted. Usually, the ZF equalized result is taken to calculate the initial radius.
The other problem is how to tell which lattice points are located inside the sphere.
It is very difficult to identify whether a lattice point is located inside a hyper
sphere or not, but it is very easy to do so for a two-dimensional sphere by simply
checking whether the integer values of the lattice points lie in the interval of the
sphere. Inspired by this, for an N-dimensional sphere the points can be determined
from one dimension to the other successively. It means that for a - dimensional
point, if its ( 1) dimension values lie in the ( 1) dimensional sphere of a
certain radius, there will be a new interval for its th value to determine if it
lies in the - dimensional sphere.
SD is aimed at finding out the solution x SD which is the same with x ML that
has the minimum Euclidean distance from the received signal r . It searches the
lattice points x within a M T dimensional hyper sphere, which can be given by:
r Hx d 2
(2.3)
where d is the initial radius of the hyper sphere. The searching of the lattice
points is a kind of iterative algorithm. For simplicity, we have to separate the
original problem into several sub-problems. Thus, the channel matrix H is firstly
reduced into an upper triangular matrix by using the QR decomposition:
20
R
H =Q
0
(2.4)
r1,1
0
0
= [Q1 , Q 2 ] 0
where Q
M R M R
r1,2
r2,2
0
0
r1, MT
r2,M T
rM T , M T
0
MT MT
is an upper triangular
M R M T
and Q 2
M R ( M R M T )
r Hx
R
= r [Q1 , Q 2 ] r
0
Q T
R
= 1T r x
0
Q 2
(2.5)
= Q1T r Rx + Q 2T r
r Rx + Q 2T r
r Rx
(2.6)
2
is equivalent to minimize
r Rx d 2
(2.7)
21
r1 r1,1
r 0
2
rM T 0
r1,M T x1
r2, M T x2
d 2
0 rM T , MT xM T
r1,1
r2,2
0
2
2
ri ri , j x j d
i =1
j =i
MT
(2.8)
(2.9)
It can be seen from the inequality equation (2.9) that there is only one unknown
quality xMT in the first term (rM T rM T , MT xM T ) 2 . Similarly, there are two unknown
qualities xMT and xMT -1 in the second term (rM T 1 rM T 1, M T xM T rM T 1, M T 1 xM T 1 ) 2
and so on. The necessary condition of inequality equation (2.9) is:
(rM T rM T , M T xM T ) 2 d 2
(2.10)
r d
r + d
MT
xM M T
T
rM T , M T
rM T , M T
(2.11)
It is easy to find out the possible values of xM by using (2.11). For example,
T
they are the odd numbers in the interval for QAM constellations. Usually, there
may be more than one possible value which are saved in the memory.
Secondly, one of the possible values of xM T is selected to solve xM T 1 by the
following inequality equation:
(rM T rM T , M T xM T ) 2 + (rM T 1 rMT 1, M T xMT rM T 1, MT 1 xM T 1 ) 2 d 2
(2.12)
22
The boundary of xM
is
2
d 2 ( r r
MT
M T , M T xM T ) + rM T 1 rM T 1, M T xM T
rM T 1, MT 1
xM 1
T
2
d 2 ( r r
MT
M T , M T xM T ) + rM T 1 rM T 1, M T xM T
xM T 1
rMT 1, M T 1
(2.13)
It can be seen from (2.13) that the signals for the previously detected
dimensions have been subtracted from the received signal.
If there is no possible value satisfying the inequality equation, the searching
process goes back to the previous step and then selects another possible value.
When all the element values of one lattice point are obtained, its Euclidean
distance away from the received point is calculated. The new distance is of course
smaller than the initial radius d , so we replace it by the new distance in (2.9) and
restart the searching process.
The searching process works in an iterative way to update the radius until no
more vectors satisfy the inequality equation (2.9). The last vector is then taken as
the decoding solution.
In conclusion, the searching process is done from the M T th dimension to
the 1st dimension. The intervals of x are calculated and the value of the each
element of x are then determined. When one lattice point is obtained, the radius
of the hyper sphere is reduced and the searching is restarted again. The searching
process goes on until only one branch is left, which is then taken as the final result.
23
Figure 2.1
shown in Fig. 2.1. Each level of the tree represents each dimension of the signal
vector x. The nodes in the figure are the values of the element of x. Take 16-QAM
constellation for example: a node emits four sub-nodes which are the values of -3,
-1, 1, 3 from the left to right. The black nodes denote the possible values that
have been visited. From the figure, we know that in the M T th dimension, the
possible values of xM T are -1 and 1. Then we select -1 to continue the searching
process down to the next dimension. It can be seen that there is one branch
emitted from the node xM T =-1 that is visited, which means all the element values
of one lattice point are obtained. Then we calculate the new radius and restart the
searching process. The search continues until no other branch can be found, the
lattice point corresponding to the last branch is selected as the final solution.
Although lots of new methods have been proposed to determine a proper
initial radius [14, 21], and many new searching strategies [22-24, 62-66] have
24
been developed to improve the searching efficiency, sphere decoding still suffers
from some drawbacks. The expected complexity of SD increases with the number
of antennas and the size of the constellation [15]. So SD is not suitable for large
size MIMO system. In addition, the complexity of SD is not fixed, but changing
with the number of the nodes visited in the searching strategies, which makes it
impractical for hardware implementation. As a result, the sub-optimum successive
interference cancellation and lattice reduction aided detection are developed to
improve the performance of linear detection with a comparable low complexity.
H = QR
(2.14)
r = QRx + n
QT r = Rx + QT n
(2.15)
25
ri = Ri , j x j + ni
(2.16)
j =i
(2.17)
rM T
RM T , MT
(2.18)
Then the ( M T 1) th symbol xNt 1 will be detected next, where the interfering
symbol xMT can be canceled by subtracting its impact:
xM T 1 =
rM T 1 RMT , MT xMT
RMT 1, MT 1
(2.19)
ri Ri +1,i +1 xi +1
Ri ,i
(2.20)
The SIC works from the M T th symbol xMT upwards to the 1st symbol x1 till all
the symbols are detected in the end.
The drawback of the SIC detection is the error propagation caused when there
is wrong decision in one symbol. If the former detected symbol is wrong, the
latter ones have very large probability to suffer from errors. In order to solve this
problem, the methods of detection ordering are proposed [7, 10, 11]. The symbol
detection ordering is in accordance with the power of the symbols, which can be
determined by the SNR or the signal to interference plus noise ratio (SINR). That
means that the symbol which has large SNR or SINR will be detected first.
26
Similarly, the detection ordering will improve the error rate performance but
increase the complexity.
many lattice bases, and the lattice basis composed of near orthogonal vectors
which are also with short lengths is the desirable one.
The procedure of determining a lattice basis with short and near orthogonal
vectors is termed as Lattice Basis Reduction [19]. To realize the reduction goal is
an N-P hard problem, whose running time varies exponentially with the
dimension of the lattice. A popular suboptimum reduction algorithm which is
called LLL algorithm [30] is proposed to reduce the complexity of Lattice Basis
Reduction.
In MIMO systems, for ill conditioned channel, the noise enhancement is large.
However, for orthogonal channel, there will be no noise enhancement. The lattice
reduction (LR) algorithm is employed to transform the original channel matrix
into a new channel matrix with much better channel condition [47, 76-81]. The
new channel matrix is composed of vectors with the shortest lengths or roughly
orthogonal to each other. The LR algorithm can be combined with the existing
27
linear detectors such as ZF and MMSE detectors. The linear detectors with a
better channel conditioned matrix will achieve better BLER performances.
Figure 2.2
With the distortion resulted from the channel matrix, the received signal
lattice without Gaussian noise is indicated in Fig 2.3. The vectors [h1 h 2 ] is the
basis of the channel matrix H [65].
28
Figure 2.3
Fig. 2.4 and Fig 2.5 show the decision regions of the ZF decoding and SIC
detection, respectively.
Figure 2.4
29
Figure 2.5
It can be seen that the ZF decoding transforms the received signal to the
transmitted signal space to make the decision, where the decision region becomes
parallelogram. The decision region of SIC detection is rectangular which is
different from that of the ZF decoding.
The ML decoding decision region as indicated in Fig. 2.6 is the best one,
which is formed by two roughly orthogonal bases. Any point in a decision region
is closer to the lattice point in its decision region than to other lattice points in
other regions. It thus inspires that if we make the decision of the signal in an
orthogonal or roughly orthogonal basis, then transform the decision value back to
the original signal space, the detection result will be much better than that in the
original basis.
30
Figure 2.6
written as:
r = Hx + n =HTT1x + n = Hz + n
(2.21)
= H* Hz + n =z + H*n
=T1x ZF
(2.22)
31
The new pseudo-inverse matrix H* tends to generate less noise enhancement than
the original pseudo-inverse matrix H* . The solution z LR ZF is then quantitized
and multiplied by T, so as to recover the solution in the transmit signal lattice:
z LR ZF =TQ ( z LR ZF )
(2.23)
Figure 2.7
z LR MMSE = H H H + n 2 IT H T
=T1x MMSE
HH r
(2.24)
(2.25)
32
Define a
(2.26)
x ext :
x
x ext =
0
(2.27)
H ext H rext
=Text 1x ext
(2.28)
33
2.6 Summary
Several existing detection algorithms for MIMO systems have been reviewed
thoroughly. As far as the comprehensive performance is concerned, they either
suffer from very high computational complexity, or rather bad error rate
performance. Thus, a lot more attentions should be paid on developing new
detection algorithms for MIMO systems, especially for the cases in which large
number of antennas and high level modulation are involved.
34
CHAPTER 3
GEOMETRIC DETECTION ALGORITHMS
3.1 Introduction
Since the minimum Euclidean distance principle of ML decoding could result
in the optimum error rate performance, the purpose of this chapter is to introduce
another perspective of reconsidering this principle. In the hyper space spanned by
the transmit lattice points, the Euclidean distance involved in the ML decoding is
found to be related to a series of concentric hyper ellipsoids. Searching the lattice
point with the minimum Euclidean distance away from the received signal point is
equivalent to searching for the lattice point that lies on the surface of the smallest
possible hyper ellipsoid. Decoding algorithms following this perspective are often
termed as geometrical detection. In this chapter, the geometrical analysis of signal
decoding for MIMO channels is presented. Then, the proposed ellipsoid searching
decoding algorithm (ESA) [69] are elaborated. It is an add-on to the standard
suboptimal detection schemes, such as ZF or MMSE. Simulation results
demonstrate that it offers the optimum error rate performance and higher diversity
gains than the standard suboptimal detection schemes.
f ( x ) = r Hx
= ( x x c ) 1 ( x x c )
T
(3.1)
35
= x + ( HT H ) HT n
1
(3.2)
= x+n
where n = ( HT H ) HT n .
1
(3.3)
It can be seen from (3.2) and (3.3) that in the absence of noise, i.e., the
transformed Gaussian noise term n = ( HT H ) HT n , both ZF equalization and
1
ML decoding result in the same correct solution. The reason why ML decoding
can offer much better performance than ZF equalization lies in the fact that the
transformed Gaussian noise has been minimized by the exhaustive search used in
ML decoding. However, the results of ZF equalization are directly distorted by the
transformed Gaussian noise n .
By using eigenvalue decomposition [72-73], the matrix M can be
decomposed into:
= ( HT H ) =VV T
1
where
arranged
MT M T
order,
and
= diag 1 , 2 ,..., MT
in
descending
(3.4)
, represents the
V = V1 , V2 ,
MT
, VNT
eigenvalues
MT M T
is
the
( H ) =1 / M
(3.5)
It has been proven [68] that the channel condition number ( H ) has a profound
36
lines of the ZF decision regions and the dash lines are the boundary lines of the
ML decision regions. The channel condition number in Fig. 3.1 (a) is 1.5 which
can be considered as a good channel, while the channel condition number in Fig.
3.1 (b) is 7.5 which can be thought as a bad channel. In the case with good
channel condition, the boundary lines of ZF decision region is very likely close to
those of the boundary lines of ML decision region. So the performance of ZF
detection is good in this case. In the case with bad channel condition, the ML
decision regions are able to match to the PDF of the received signal vector, but the
ZF decision regions can not. The boundary lines of the ML decision region are
approximately orthogonal to the dominant principal axis which is corresponding
to the vector V1 . In general, the decision regions of ZF detector cannot have this
property since its boundary lines always go through the origin point.
37
(a)
(b)
Figure 3.1 Probability density function of the received signal vector of ZF
detections in 2 2 MIMO systems.
(a) Case for good channel condition. (b) Case for bad channel condition.
f ( x ) given in (3.1)
38
(3.6)
(3.7)
where a 2 represents the height of the cross section above the M T dimensional
space as shown in Fig. 3.2. The length and the direction of the i-th semiaxis of the
hyper ellipsoid are given as a i and Vi , respectively. With different value of
39
Fig. 3.3 shows a two dimensional lattice point space ( x1 x2 plane) with
three lattice points Point 1, Point 2, and Point 3 as shown in the figure. With
different a 2 , a group of similar hyper ellipsoids can be obtained, and their
projection onto the x1 x2 plane are ellipses which are all centered at the point
40
xc = H 1r . For each lattice point, there exists an ellipse that passes through it. The
corresponding ellipse of the ML solution is the one that has the minimum area. As
shown in Fig. 3.3, Point 1 is taken to be the ML solution while Point 2 and Point 3
are not, since Point 1 is located on the inner-most ellipse which has the minimum
area.
However, finding the smallest hyper ellipsoid containing the solution signal
vector is not an easy task. If we use the largest hyper ellipsoid which contains all
the signal vectors, then the complexity will be the same as ML decoding. Here we
propose an ellipsoid-searching decoding algorithm that uses a small hyper
ellipsoid containing the solution symbol vector to start the search and then
identify all the symbol vectors inside. The ESA consists of the following 4 steps:
(3.11)
After determining the hyper ellipsoid, the next key task is to identify whether
41
there are any lattice points located inside this hyper ellipsoid. The axes of the
M T -dimensional rectangular coordinate system for the lattice point space are
denoted as i - axes. Since the directions of the hyper ellipsoids semiaxes are not
in parallel with the axes of the coordinate system of the lattice point space, it is
rather complicated to directly use the surface equation (3.11) of the hyper
ellipsoid. Here we propose to use a circumscribed hyper rectangle as follows.
We set up a new M T -dimensional rectangular coordinate system with i axes ( i = 1, 2,3,..., M T ) being coincided with the i th semiaxis of the hyper
ellipsoid and the origin coincided with the global minimum point xc . We use the
superscript prime to denote the variables in the new coordinate system. The
coordinates of the 2M T apexes of the circumscribed hyper rectangle in this new
coordinate system are given by
k p = xp1 , xp 2 ,...xpMT
where p = 1, 2,3,...2 MT ,
(3.12)
(3.13)
VT = V1 , V2 ,
v11
v12
, VM T = v13
v1M T
v21
v31
v22
v32
v23
v33
v2 M T
v3 MT
vM T 1
vM T 2
vM T 3
vM T M T
(3.14)
42
(3.15)
q =1
where xci is the i th component of xc . Since xpq = azf q , the maximum and
minimum boundaries of the values of the each component in k p in the i - axes
can be expressed as:
MT
(3.16a)
q =1
MT
(3.16b)
q =1
Since the circumscribed hyper rectangle encloses the hyper ellipsoid, any
lattice point s = s1
i = 1, 2,3,..., M T
(3.17)
It should be noted that this is not a sufficient condition for identifying the
lattice points lying inside the hyper ellipsoid.
From (3.17), we can obtain the possible value set i = { i1 , i 2 , i 3 ,
of the
i th element for the lattice points located inside the hyper ellipsoid. So the
search set becomes a larger hyper rectangle that encloses the circumscribed hyper
rectangle. For PAM and QAM, the elements of j are the odd numbers between
xi _ max and xi _ min , and it can be easily shown that the number of elements is:
MT
(3.18)
43
i =1,i l
then it means that there is no lattice point located inside the hyper ellipsoid. The
searching process will terminate and the zero forcing point chosen before is
considered as the solution.
Otherwise, assuming the possible value set has the largest number of
elements among all the possible value sets, we form the combinations from the
other M T 1 possible value sets, and then substitute each of these combinations
into (3.11), to determine the lattice point elements of the possible value set
that are located inside the hyper ellipsoid. In doing so, the number of
combinations that need to be considered is smaller and hence lesser computation
complexity. Denoting the k th combination by:
1,k , +1,k
, M T , k
(3.19)
MT
k = 1, 2,...,
j =1, j
Num j
1,k , ,d ,k , +1,k
, MT ,k
(3.20)
d = 1, 2,..., nk
where nk is the number of the elements of
,k for Com k .
44
d +1,k = d ,k + d ,k
(3.21)
where x d +1,k = x d ,k + 2 .
d ,k = 4 hi 4 ( r Hx d ,k ) hi
2
(3.22)
After all the Euclidean distances are calculated, the signal vector with the
minimum distance is then selected as the solution.
3.3.5 Examples
The following subsections will give two examples of the ESA in two
dimensional space and three dimensional space.
45
B = azf 1 , + azf 2
),
C = azf 1 , azf 2
) , and
D = azf 1 , + azf 2
, respectively.
Substituting these vectors into (3.15) yields the corresponding coordinates in the
lattice point space. From (3.16) the x1 coordinates of points A and D are chosen
as
x1_ min
possible set of values along each axis, i.e., two values {1, 3} along the x1 -axis
and one value {1} along the x 2 -axis. Since the number of values along the
x1 -axis is larger than that along the x 2 -axis, we substitute 2,1 = 1 into the hyper
ellipsoid equation (3.11). As shown in Fig. 3.5, the possible value along the
x1 -axis is 1,1,1 = 3 , so the point x1,1 = [3 1]
46
rectangle which has been set up by the method introduced in section 3.3.2. xc is
the center of the ellipsoid, whose semiaxes are aligned along vectors V1 , V2 ,
V3 , with their lengths being azf 1 , azf 2 and azf 3 , respectively. By
substituting the coordinates of the eight points A to H to (3.15) and (3.16), x1_ min
and x1_ max , x2 _ min
and x2 _ max , x3_ min and x3_ max , which are all marked as
dots are obtained. The possible set of values along x1 -axis is {1, 3, 5}, and the
possible set of values along the x 2 -axis is {1, 3}. Along x3 -axis, the possible set
of value is {-1}. Since the number of possible values along the x1 -axis is the
largest
compared
to
those
and
along
the
other
axes,
we
into
substitute
(3.11)
to
possible value set 1,1 along the x1 -axis is {1} for Com1 and 1,2 is {5} for
Com 2 , so the point x1,1 = [1 1 1]
are
47
48
points while the ESA algorithm only searches a small subset. The ESA algorithm
is assessed by means of the simulation results of the error rate performance. In the
simulations, we used 4-QAM, 16-QAM, 64-QAM in Rayleigh flat fading
channels with i.i.d. complex zero-mean Guassian noise. Fig. 3.6 illustrates the
BLER performance of ESA compared with ML decoding and ZF decoding using
4-QAM. Fig. 3.7 shows the BLER performance of ESA compared with ML
decoding ZF decoding using 16-QAM. And Fig. 3.8 shows the BLER
performance of ESA compared with ML decoding and ZF decoding using
64-QAM. It can be seen that the performances of ESA can achieve that of ML
decoding and are much better than ZF decoding.
Table 3.1 compares the complexity of ML decoding and ESA. The numbers
of lattice points visited by ML decoding and ESA for transmitting 16-QAM and
64-QAM constellations in 2 2 to 4 4 MIMO systems are indicated. It can be
observed that compared with the ML decoding, the number of lattice points
visited by the ESA is substantially reduced from 95.7% to 99.8%. The more
number of antennas and the higher level of modulation the system applies, the
greater complexity reduction the ESA can achieve.
49
(a)
(b)
Figure 3.6 Comparison of BLER performance of ESA, ML decoding and ZF
using 4-QAM.
(a) 4 4 MIMO systems. (b) 6 6 MIMO systems.
50
(a)
(b)
Figure 3.7 Comparison of BLER performance of ESA, ML decoding and ZF
using 16-QAM.
(a) 4 4 MIMO systems. (b) 8 8 MIMO systems.
51
22
33
44
16-QAM
256/11
95.7%
4096/20
99.5%
4096/69
98.3%
262144/647
99.7%
65536/265
99.6%
16777216/39361
99.8%
64-QAM
3.5 Summary
In this chapter, the geometrical analysis of signal decoding for MIMO
channels is presented. The ellipsoid searching decoding algorithm is introduced in
detail. It is an add-on to the standard suboptimal detection schemes. Simulation
52
results demonstrate that it can provide the optimum performance and higher
diversity gain compared to the standard suboptimal detection schemes, and has
considerable reduction in complexity compared with ML decoding.
53
CHAPTER 4
MIMO DETECTION ALGORITHMS BASED ON
SEMIDEFINITE RELAXATION
4.1 Introduction
In this chapter, the background of the convex optimization problems and the
concept of the semidefinite relaxation programming will be introduced. Then the
semidefinite relaxation methods for MIMO detection for both low-order
modulation system and high-order modulation systems will be presented in details.
The tightness of the semidefinite relaxation detectors will be compared and their
performances will be illustrated. Finally, the SDR-initiated sphere detectors will
be proposed together with the simulation results.
54
(4.1)
fi ( x )
(4.2)
xR
st.
bi , i = 1, , m
represents = or or . Denoting
bi , i = 1, , m
being satisfied.
The optimization problems can be classified into three types according to the
objective function and constraint functions: the linear optimization problems, the
nonlinear optimization problems and the convex optimization problems, wherein
the objective function and constraint functions are linear, nonlinear and convex
respectively. This section will focus on the convex optimization problems.
So, what is convex, and what is a convex function? Firstly, a set S R n is
convex if it contains the straight line segment between any two distinct points in S:
x+ yS
(4.3)
55
(a)
Figure 4.1
(b)
(a) Convex set. (b) Non-convex set.
(4.4)
( x, f ( x ) )
and
(4.5)
56
(a) f ( x )
Figure 4.2
(b) - f ( x )
(a) Convex function. (b) Concave function.
Figure 4.2 shows the examples of convex function and concave function [49].
It can be seen that the line segment joining ( x, f ( x ) ) and ( y, f ( y ) ) lies above
the graph of the function f ( x ) , so f ( x ) is convex and of course - f ( x ) is
concave.
Thirdly, a generalized convex optimization problem can be formed as [49]:
minn f 0 ( x )
xR
st.
fi ( x ) 0, i = 1, , m1
(4.6)
hi ( x ) = 0, i = 1, , m2
xS
are affine functions; the constraint set S is a convex set. If any one of the above
conditions is not satisfied, the problem is a non-convex problem.
The optimization variable x S is feasible if it satisfies the constraints
functions fi ( x ) 0 and hi ( x ) =0 . If there exists a feasible variable x for the
problem (4.6), it is called a feasible problem. Otherwise, the problem is an
infeasible one. The set which contains all the feasible variables x is called the
57
58
0 , S 0 , S 0 , the
matrix S is called positive definite matrix, negative definite matrix and negative
semidefinite matrix, respectively.
The characteristics of the positive semidefinite matrix S are as follows:
1) S is positive semidefinite: S
0.
,n.
,n.
mn
S =A T A .
A convex cone is a kind of convex set which is closed under positive scaling.
Let S +n denote a set of n n positive semidefinite matrices, then the set S +n of
59
particular dimension is called the convex semidefinite cone. The cone defines a
partial order for X, Y S n by X
Y.
min cT x
(4.7)
n
st.
F ( x ) = F0 + xi Fi 0
(4.8)
AT x = a
(4.9)
i =1
where Fi S +n , i = 1, 2, , n , A
jn
60
linear function and the constraint functions (4.8-4.9) are convex functions. That
means that if F ( x ) 0 , F ( y ) 0 , then:
F ( x + y ) = F ( x ) + F ( y ) 0
(4.10)
where 0 , 1 and + =1 .
The semidefinite programming problems also have another form which is
give by:
st.
min cT Xc = Tr ( CX )
(4.11)
Tr ( A i X ) = ai , i = 1, 2, , n
(4.12)
(4.13)
61
(4.14)
where r
MR
, x
MT
, H
M R M T
In the following subsections, the SDR detectors for MIMO systems with
different level of modulations will be elaborated.
62
min r - Hx
st.
(4.15)
xi = 1 , i = 1, 2, , M T
HT H HT r
By introducing the matrix Q = T
, the problem can be
T
r
H
r
r
expressed as:
min x
Define
by: X= xT
HT H HT r x
1 T
T
r H r r 1
st.
xi 2 = 1, i = 1, 2, , M T
rank-one
1 xT
semidefinite
matrix,
(4.16)
which
is
given
H T H HT r x
1 T
=min Tr ( QX )
T
r H r r 1
(4.17)
Since xi equals either 1 or -1, and the matrix X is a positive semidefinite matrix,
the rank of X equals one and its diagonal entries are equal to 1. Assuming the
matrix X satisfies the above three characterising properties, so that the problem
becomes:
min Tr ( QX )
st.
diag ( X ) = e MT +1
(4.18)
rank ( X ) = 1
63
where the diag ( i ) means the diagonal elements, and e represents all ones vector
of M T dimension. The problem is a non-convex optimization problem because of
the rank-one constraint function, so we eliminate it to yield the basic semidefinite
relaxation [50] as:
min Tr ( QX )
st.
diag ( X ) = e MT +1
(4.19)
1)
64
Let X = i qi qiT
1 2 M
+1
corresponding eigenvectors.
The feasible solution is :
x = Q qM T +1 q1 , q2 , , qM T
(4.20)
2)
Randomization
Firstly, the optimal solution X is Cholesky factorized to be X =FT F . Then,
, N rand ,
x i = Q ( FT ui ) , i = 1, 2,
l = arg max
i =1,2, , N rand
, N rand ,
and
choose
x = x l ,
where
x iT Qx i .
3)
(4.21)
Rank-one approximation
Firstly, the optimal solution X is partitioned into
X
X = 1,1T
X 1,2
X1,2
X2,2
(4.22)
65
x = Q ( X1,2 / X2,2 )
(4.23)
1]T [xT
(4.24)
1]
X1,2
X 2,2
(4.25)
(4.26)
(4.27)
X1,2 =x {1
3}
MT
X 2,2 = 1
(4.28)
(4.29)
st.
T
T
H H H r
min Tr X T
T
r H r r
(4.30)
X =X T
(4.31)
(4.32)
(4.33)
X 2,2 = 1
(4.34)
However, this problem is still not a convex problem due to the non-convex
constraints diag ( X1,1 ) {1,9} and X1,1 = X1,2 X1,2T . Thus, relaxation of these
66
diag ( X1,1 ) {1,9} . In the following, three methods: PI-SDR [40], BC-SDR [41]
and VA-SDR [42] for dealing with this constraint will be introduced.
, MT
(4.35)
(x
1)( xi 2 9 ) = 0 , i = 1, 2,
, MT
(4.36)
ti 2 10ti + 9 = 0 , i = 1, 2,
, MT
(4.37)
tT
(4.38)
V1,1
V = V1,2T
V1,3T
V1,2
V2,2
V
T
2,3
V1,3
V2,3
V3,3
(4.39)
where V1,1 , V1,2 and V2,2 are all M T M T dimensional matrixes, V1,3 and V2,3 are
i = 1, 2,
the equation ti = xi 2 is the same as diag ( V1,1 ) V2,3 = 0 . Then, (4.31) along with
(4.32) can be relaxed into V
0 and rank ( V ) = 1 .
st.
HT H 0 HT y
min Tr V 0
0
0
yT H 0 yT y
(4.40)
(4.41)
(4.42)
V3,3 = 1
(4.43)
(4.44)
rank ( V ) = 1
(4.45)
Dropping the non convex rank-one constraint can yield the SDR problem:
st.
HT H 0 HT y
min Tr V 0
0
0
yT H 0 yT y
(4.46)
(4.47)
(4.48)
V3,3 = 1
(4.49)
(4.50)
Then the optimal solution V can be solved by the IPM and the approximation
is the final
procedure. The first 2M T elements of the last row in the solution V
solution x .
The PI-SDR precisely transforms the alphabet constraint (4.33) into linear
constraints by a higher dimensional matrix, so this is an increased-dimension
relaxation.
68
st.
HT H HT r
min Tr Y T
T
r H r r
(4.51)
(4.52)
e diag{Y1,1} 9e
(4.53)
Y2,2 = 1
(4.54)
rank ( Y ) = 1
(4.55)
st.
T
T
H H H r
min Tr Y T
T
r H r r
(4.56)
(4.57)
e diag{Y1,1} 9e
Y2,2 = 1
(4.58)
(4.59)
This BC-SDR problem can then be solved by any of the SDP solvers, such as
Sedumi, based on interior-point methods. Although (4.57) could be deduced from
(4.31) and (4.32), and (4.58) can also be deduced from (4.33), the BC-SDR
problem is however not exactly equivalent to the ML detection problem. Thus, the
69
solution obtained from solving BC-SDR has more errors than ML solution. Once
the solution is found, the approximation procedure approach is applied to quantize
till the constraint (4.33) is satisfied.
the resulting Y
1,2
This SDR method transfers alphabet constraint (4.33) into the linear
constraints in the same dimensional space, so it is a fixed-dimension relaxation.
x {1
3}
MT
where U = [ I 2I ] , p = [p1 p 2 ] , I
(4.60)
MT M T
and pi {1}
MT
, i = 1, 2,3, 4 .
Substituting (4.60) into (4.24), and defining a matrix Z , which is given by:
Z = [p 1]T [p 1]
(4.61)
the objective function (4.30) can be equivalently transformed into (4.62), and the
constraints (4.33) and (4.34) are relaxed into (4.64). Similarly, (4.31) along with
(4.32) can be relaxed into (4.63). Thus, we obtain the VA-SDR problem given by:
st.
T
T
T
T
U H HU U H r
min Tr Z
T
r Tr
r HU
(4.62)
(4.63)
diag ( Z ) = 1e
(4.64)
rank ( Z ) = 1
(4.65)
st.
T
T
T
T
U H HU U H r
min Tr Z
T
r Tr
r HU
(4.66)
(4.67)
70
diag ( Z ) = 1e
(4.68)
(4.69)
The VA-SDR method transfers the alphabet constraint (4.33) into the linear
constraints in the same dimensional space. It is also a fixed-dimension relaxation.
71
Figure 4.3
72
Figure 4.4
Comparison of BLER performance of SDR detectors, ML
decoding, ZF using 4 4 16-QAM MIMO systems.
(b)
Figure 4.5
Comparison of BLER performance of SDR detectors, ML
decoding, ZF using 8 8 16-QAM MIMO systems.
73
To enable high data rate in the future wireless communication systems, the
ability to utilize the high-order modulation is in great need. The sphere decoding
(SD) can provide optimum BLER performance at the low-order signal
constellations or small system dimensions. However, SD cannot efficiently handle
high-order symbol constellations or high problem dimensions, especially at low
SNR. In contrast, the decoding algorithms based on semidefinite relaxation (SDR)
approach have become more and more attractive simply because of the fact that
SDR problems can be solved very efficiently even for high-order symbol
constellations or high problem dimensions.
The SDR approach was firstly applied to detect binary phase-shift keying
(BPSK) and quadrature amplitude modulation (4-QAM) signals. Then the
extensions to different SDR techniques for 16-QAM signals had been proposed,
all exhibit acceptable BLER performance and relatively low complexity. In [43],
it has been proved that there exists equivalence among PI-SDR, BC-SDR and
VA-SDR for 16-QAM. However, due to its high complexity, the PI-SDR is not
suitable for extension to 256-QAM system. In the following subsections, the
extension of BC-SDR and VA-SDR to 256-QAM has been investigated, a new
SDR detector is proposed for 256-QAM system. Then, a comprehensive
comparison between the proposed method and the previous SDR detectors is
made.
74
st.
HT H HT r
min Tr Y T
T
r H r r
(4.70)
Y = YT
(4.71)
(4.72)
MT
Y2,2 = 1
(4.73)
(4.74)
It can be observed that the high complexity of the ML detection is due to the
presence of the two non-convex constraints (4.72) and (4.73). Thus, relaxation of
these constraints will be engaged to transform the original problem into a SDR
problem, which can then be efficiently solved in polynomial time. Firstly, the
constraint (4.73) implies 1 xi2 225 , where xi denotes the i-th component of x .
Please be noted that the diagonal elements of Y1,1 are xi2 , for i = 1, 2, , M T , thus,
it satisfies that e diag{Y1,1} 225e . Secondly, the constraints (4.71) and (4.72) can
be relaxed into Y 0 . Consequently, the BC-SDR problem is obtained as:
st.
T
T
H H H r
min Tr Y T
T
r H r r
(4.75)
(4.76)
e diag{Y1,1} 225e
(4.77)
Y2,2 = 1
(4.78)
75
The BC-SDR problem can then be solved by any of the SDP solvers. Once
the solution is found, the randomization approach is applied to quantize the
till the constraint (4.73) is satisfied.
resulting Y
1,2
where U = [ I 2I 4I 8I ] , p = [p1 p 2 p3 p 4 ] , I
(4.79)
M T M T
and pi {1}
MT
i = 1, 2,3, 4 .
(4.80)
By substituting (4.79) and (4.80) into (4.30), the objective function becomes
T
T
T
T
U H HU U H r
min Tr Z
. Finally, we obtain the VA-SDR problem
T
r Tr
r HU
given by:
st.
U T H T HU U T H Tr
min Tr Z
T
r Tr
r HU
(4.81)
(4.82)
diag ( Z ) = 1e
(4.83)
76
x = Uz
(4.84)
As introduced in section 4.5.2, one of the key issue for deducing the
corresponding SDR detector lies in how to deal with the non-convex constraint of
the alphabet set xi {15, 13, 11, 9, 7, 5, 3, 1} . In the extended BC-SDR,
the alphabet set is directly relaxed into 1 xi2 225 . This is reasonable for the
low-order constellations, such as 16-QAM. However, for 256-QAM system, it
will inevitably cause considerable errors. In the extended VA-SDR, the alphabet
set is reformed by expanding the dimension of the problem into four times of its
original size. This will result in the significant increase of computational
complexity. In order to keep the relaxation as tight as possible, while the
complexity as low as possible, a novel method to deal with the alphabet set is
proposed.
Actually, the signal xi {15, 13, 11, 9, 7, 5, 3, 1} in 256-QAM
constellations could also be expressed as:
x = VqT
where V = [ I 4I ] ,
q = [q1 q 2 ] , I
M T M T
(4.85)
77
Table 4.1 gives the values of xi for the possible combinations of q1i and q2i ,
where xi , q1i and q2i denote the j-th element of x , q1 and q 2 , respectively.
By substituting (4.85) into (4.24), and defining a matrix W , which is given by:
W = [q 1]T [q 1]
(4.86)
it
can
be
known
from
(4.85)
that
the
constraint
essentially the well known indices used to characterize the 16-QAM constellation.
Herein, the set operation method [44] is engaged to formulate the alphabet
constraint (4.73) into (4.88) and (4.89). Similarly, (4.71) along with (4.72) can be
relaxed into (4.90). Also (4.74) can be reformulated as (4.91). Thus, we obtain the
proposed SDR problem given by:
(4.87)
1e diag {W1,1} 9e
(4.88)
st.
V T H T HV V T H Tr
T
r Tr
r HV
min Tr W
W2,2 = 1
0,
(4.89)
(4.90)
(4.91)
is then
The problem can be solved by using IPM. The optimal solution W
is reconstructed by:
x = Vw
(4.92)
78
As mentioned above, these three SDR problems are all relaxed from the
original ML problem, and their objective functions are equivalent, which are
actually calculating the Euclidean distance given by (4.14). Thus, the tightness of
the constraints of each SDR algorithm implies how close it is to the ML decoding.
In what follows, we will compare their tightness.
( 4 M T +1)( 4 M T +1)
( M T +1)( M T +1)
(4.76)-(4.78).
Proof: For any matrix Z which satisfies the constraints (4.82)-(4.83), it has
the following form:
Z
Z = 1,1
Z 2,1
4 M T 4 M T
1 4 M T
Z1,2
1
4 M T 1
(4.93)
M T , MT +1 , MT + 2 ,
2 MT , 2 M T +1 , 2 M T + 2 ,
, 3 M T , 3 M T +1 , 3 M T + 2 ,
4 M T , 4 M T +1
(4.94)
that satisfies:
Z = T
where m
( 4 M T +1)1
(4.95)
, m = 1, 2, , 4 M T + 1 .
79
( G )
I
where G =
=
0 1 0
U 0
MT MT
1 M T
2I
0
MT MT
4I
1 M T
(4.96)
MT MT
8I
1 M T
MT MT
M T 1
1 M T
1/2
= 1 , n = 1, 2,
, 4M T
(4.97)
T4 M T +1 4 M T +1 = 1
(4.98)
Y2,1 Y2,2
( 1 + 22 + 43 + 84 )T ( 1 + 22 + 43 + 84 )
=
T
T4 MT +1 ( 1 + 22 + 43 + 84 )
( 1 + 22 + 43 + 84 )
T4MT +1 4MT +1
4 MT +1
(4.99)
( 4 M T +1) M T
where 1 = 1 , 2 ,
Nt
2 = MT +1 , MT + 2 ,
2 MT
( 4 M T +1) M T
3 = 2 Nt +1 , 2 Nt + 2 ,
3 Nt
( 4 M T +1) M T
4 = 3 Nt +1 , 3 Nt + 2 ,
4 Nt
( 4 M T +1) M T
,
,
,
.
From (4.98) and (4.99) it can be known that Y2,2 = 1 , and (4.78) is satisfied.
Moreover, we have:
Y1,1 = ( 1 + 2 2 + 43 + 8 4 )
( 1 + 22 + 43 + 8 4 )
(4.100)
and the element located in the j-th row and the j-th column of Y1,1 is:
y jj = j + 2 M T + j + 4 2 MT + j + 8 3 M T + j
(4.101)
80
(4.103)
1e diag{Y1,1} 225e
(4.104)
( M T +1)( M T +1)
( 4 M T +1)( 4 M T +1)
(4.83).
Proof: For any matrix Y that satisfies the constraints (4.76)-(4.78), it has the
following form:
Y
Y = 1,1
Y2,1
MT M T
1 M T
Y1,2
1
M T 1
(4.105)
Nt , MT +1
that satisfies:
Y = T
where l
( M T +1)1
(4.106)
, l = 1, 2, , M T + 1 .
M T +1 = TM T +1 M T +1
1 j = ( Tj j )
1/ 2
1/ 2
=1
15 , j = 1, 2,
(4.107)
, MT
(4.108)
81
From the perspective of geometry, it is easy to know that there should be vectors
j1 , j 2 , j 3 , j 4
( M T +1)1
which satisfy:
j = j1 + 2 j 2 + 4 j 3 + 8 j 4
(4.109)
j1 = j 2 = j 3 = j 4 = 1
(4.110)
Now we construct another matrix in the basis of (4.106) and (4.109). which is
given by:
= 11 , 12 , 13 , 14 , 21 , 22 , 23 , 24 ,
Nt 1 , Nt 2 , Nt 3 , Nt 4 , Nt +1
( M T +1)( 4 M T +1)
(4.111)
( 4 M T +1)( 4 M T +1)
Substituting (4.111) into (4.112), the element located in the ( j i ) -th row and the
=1
(4.113)
z( 4 M T +1)( 4 M T +1) = TM T +1 M T +1 = 1
(4.114)
(4.115)
82
( Nt +1)( Nt +1)
Proof: For any matrix W that satisfies the constraints (4.88)-(4.91), it has the
following form:
W
W = 1,1
W2,1
2 M T 2 M T
W1,2
1 2 M T
2 M T 1
(4.116)
MT , MT +1 , MT + 2 ,
2 MT , 2 MT +1
(4.117)
that satisfies:
W = T
where s
( 2 M T +1)1
(4.118)
, s = 1, 2, , 2 M T + 1 .
1/2
3 , k = 1, 2,
, 2M T
T2 M T +1 2 M T +1 = 1
(4.119)
(4.120)
( M T +1)( M T +1)
( E )
T
83
I
where E =
=
0 1 0
V 0
MT MT
MT MT
4I
1 M T
1 M T
M T 1
0
1
Y1,1
Y=
Y2,1
( 1 + 4 2 )
2 MT +1
2 MT +1 2 M T +1
T
(4.122)
where
1 = 1 , 2 ,
M T
( 2 M T +1) M T
, 2 = M +1 , M
T
+2
2 MT
( 2 M T +1)M T
From (4.120) and (4.122) it can be known that Y2,2 = 1 , and (4.78) is satisfied.
Moreover, we have:
Y1,1 = ( 1 + 4 2 )
( 1 + 4 2 )
(4.123)
and the element located in the j-th row and the j-th column of Y1,1 is:
2
y jj = j + 4 M T + j , j = 1, 2,
, MT
(4.124)
(4.125)
Thus we have:
1 y jj 225
(4.126)
1e diag{Y1,1} 225e
(4.127)
( M T +1)( M T +1)
( 2 M T +1)( 2 M T +1)
(4.90) and (4.91). However, this W may not satisfy the constraint (4.89).
84
Through some deduction process similar to that given in the step 2 of section
( 2 M T +1)( 2 M T +1)
( M T +1)( M T +1)
. Next, we
2 MT
(4.128)
It is easy to know that W 0 , diag {W1,1} = 2.25e and diag {W1,2 } = 1.5e , thus we
have:
diag {W1,1} 4diag {W1,2 } + 3e
= 2.25e 6e + 3e = 0.75e
(4.129)
Obviously, (4.129) being negative can not satisfy the constraint (4.89).
From both step 1 and step 2, it can be concluded that the constraints of the
proposed SDR are tighter than those of the BC-SDR, and also tighter than those of
the VA-SDR due to the aforementioned equivalence.
[41].
Secondly,
the
VA-SDR
( 4 M T + 1) ( 4 M T + 1) matrix variable
given
in
(4.81)-(4.83)
involves
constraint (4.83) are also separable, thus the complexity of the VA-SDR detector
( 2 M T + 1) ( 2 M T + 1) matrix variable
85
According to [44] and [52], it can be known that the complexity of the proposed
86
demonstrate that the BC-SDR has the lowest computational complexity and the
VA-SDR has the highest computational complexity while the proposed SDR is in
between the two detection methods. It is also found that the complexity of SDR
detectors is independent of SNR. This is a distinct advantage over the SD whose
complexity varies as a function of SNR, making SD difficult to be implemented in
practice.
Figure 4.6
87
Figure 4.7
Figure 4.8
Figure 4.9
89
hyper sphere and the searching has to be restarted with a new initial radius [14,
53]. A traditional way to attain the initial radius is the ZF equalization. The radius
is the distance between the received point and the ZF equalized point. For the
MIMO systems with small constellation size, such as 8-QAM and 16-QAM, the
ZF-initialed SD has acceptable complexity. However, due to the poor BLER
performance of ZF or MMSE for cases with large constellation size, such as 256QAM, the complicated searching process becomes unacceptable.
The decoding algorithms based on SDR have become more and more
attractive simply because of the fact that SDR problems can be conveniently
solved in polynomial time. Although the SDR detectors can offer significantly
low computational complexity, as well as better BLER performance than ZF and
MMSE, their BLER performance is still worse than SD detector because of the
relaxation process.
The purpose of this section is to propose a new detection algorithm for 256QAM signals which combines the SDR with the SD. In this method, the SDR
algorithms are engaged to obtain a primary result. The feasible solution of the
SDR problem is proposed to be the initial point of SD. The radius becomes the
distance between the received point and the feasible point. Then, a hyper sphere is
constructed which is centered at the received signal and has its radius equals to the
Euclidean distance between the primary result and the received signal. Finally, the
SD searching strategy is employed to determine the final result which satisfies the
principle of maximum likelihood. This method can offer optimum BLER
performance as well as lower computational complexity than the traditional SD
detectors. Since the SDR detections have much better BLER performance than ZF
90
for 256-QAM MIMO systems, the radius given by SDR tends to be smaller. Thus,
the number of lattice points to be visited inside the sphere is smaller, which means
it can reduce the complexity of SD.
Matlab simulation has been used to assess the performances of the proposed
SDR-SD used in the 4 4 MIMO systems transmitting 256-QAM in a fading
channel. The BC-SDR, VA-SDR and the newly proposed SDR are separately
combined with SD. Their BLER performances are compared with the stand-alone
SD detector and SDR detectors and the results are shown in Figure 4.10. It can be
seen that the three SDR-SD detectors can offer the same BLER performance as
the SD detector, and have much better BLER performance than the stand-alone
SDR detectors. Figure 4.11 shows the comparison of the complexity of the
different detection algorithms. The complexity of a detection algorithm is
measured by the average computational time required. It can be seen that all three
combined SDR-SD detectors are faster than the stand-alone SD detector.
Furthermore, the BC-SDR-SD has the lowest complexity.
91
Figure 4.10
Figure 4.11
92
4.7 Summary
The semidefinite relaxation approach usually consists of following four steps:
1) Convert the objective and constraint functions into convex or affine functions.
2) Relax the non convex constraints.
3) Solve the SDR problem by using interior-point methods.
4) Convert the optimal solution to a feasible solution of the original problem.
In this chapter, we firstly reviewed the existing SDR detectors for 16-QAM
constellations, which include PI-SDR, BC-SDR and VA-SDR. Then, the BC-SDR
and VA-SDR were extended into high order modulation MIMO system, such as
256-QAM constellations. After that, we proposed a new SDR detector which can
transform the 256-QAM constellation into an equivalent 16-QAM constellation.
Next, we analyzed the tightness and the complexity of these three SDR detectors
for 256-QAM. Both theoretical analysis and simulation results demonstrate that
the proposed SDR can provide the best BLER performance among these three
detectors, while the BC-SDR detector and the VA-SDR detector provide exactly
the same BLER performance. Moreover, the BC-SDR offers the lowest
computational complexity and the VA-SDR is with the highest computational
complexity, while the complexity of the proposed SDR is higher than that of the
BC-SDR and lower than that of the VA-SDR. Finally, the principle of the
combined SDR-SD detection algorithms was proposed. The simulation results
show that the SDR-SD detectors have much lower complexity compared with the
stand-alone SD detector while maintaining the optimum BLER performance.
93
CHAPTER 5
LATTICE-REDUCTION-AIDED SEMIDEFINITE
RELAXATION DETECTION ALGORITHMS
5.1 Introduction
The semidefinite relaxation (SDR) method for MIMO detection is an
attractive alternative to the ML decoding because it is very computationally
efficient. In the case of the BPSK and 4-QAM constellation, the BLER
performance of SDR detection is fairly close to the optimum BLER performance.
It has been proven that the receive diversity order of SDR detector is M R / 2 in
real valued cases, which is equal to that of the optimal ML decoding. However,
for higher level modulation, the SDR detection could not achieve the full diversity
which can be observed from the simulation results presented in Chapter 4.
For BPSK and 4-QAM, the real elements of the constellation belong to the set
{-1, 1}. When deducing the corresponding SDR problem, the constraints that can
represent the characteristic of the elements of the constellation are given by:
diag ( X ) = 1e
(5.1)
rank ( X ) = 1
Although there is no explicit integer constraint, these three conditions together are
exactly the same as x {1,1} . In order to obtain the SDR problem, the nonconvex constraint rank ( X ) = 1 is removed so that it can be efficiently solved by
the SDP solvers. But due to the elimination of this rank-one constraint, the
.
extracts the feasible solution x from solution X
But in the cases of higher level modulation such as 16-QAM (24-QAM), 32-
1e diag ( X1,1 ) ( 2q /2 1) e
2
(5.2)
X 2,2 = 1
rank ( X ) = 1
It should be noted that (5.2) can cover all the elements of the constellation of 2qQAM.
{ ( 2
q 2
However,
1) ,
, 1,1,
it
is
not
exactly
equivalent
to
the
alphabet
set
alphabet set can not be strictly satisfied besides the rank-one constraint being
omitted. For PI-SDR and VA-SDR, similar problems will inevitably be
encountered when deducing their corresponding SDR problems. Thus, in higher
level modulation MIMO systems, the BLER performance of SDR detection is
much worse than that of the ML detection. In what follows, we will apply the
lattice reduction (LR) algorithm to SDR detection, so that the novel LR-aided
SDR detectors could offer much improved BLER performance.
95
H = UDV H
(5.3)
min ( M
,M R )
(H) =
max ( i )
i
min ( i )
(5.4)
96
MT
, the possible received signals which are not disturbed by the Gaussian
MT
L ( H ) = L h1 , , h MT = hl
For
any
lattice
= h1 h 2
L,
there
are
many
(5.5)
l =1
possible
lattice
bases.
Herein,
( )
L ( H ) = L = HT
(5.6)
The matrix T is unimodular means that T and T1 has only integer entries and
the determinant of T equals 1 or -1. Apparently, we have Hx = ( HT ) ( T-1x ) = z ,
where z = T1x . Therefore, a point which is denoted by x in the basis H can also
be denoted by z in the basis .
By using a unimodular matrix T, LR transforms the matrix H into a new
matrix . The newly derived basis vectors are more orthogonal to each other and
have shorter lengths, so that a better decision domain can be obtained. It is worth
noting that although the decision domain is changed with the basis vector, the
lattice still remains itself unchanged.
Although the LR technology may not always lead to the optimum solution, it
indeed could improve the BLER performance of the detectors. The more
correlated the columns of the original channel matrix H are, the more significant
improvement of the BLER performance could be obtained. Consequently, to find
the optimum lattice basis is the so-called lattice reduction problem.
97
shown in orange solid lines is obtained from the basis ( h1 , h 2 ) indicated in blue
dash lines by using LLL reduction algorithm. It can be seen that the basis vectors
h1 and h 2 are more orthogonal to each other and their lengths are shorter.
Figure 5.1
H = QR
(5.7)
MT MT
98
parameter where 1/ 4 < < 1 , the elements in the matrix R should fulfill the
following conditions:
rl ,k < 1/ 2 rl ,l
for 1 l < k M T
for k = 2, M T .
(5.8)
(5.9)
Initialization step
1: Q := Q , R := R , T := I MT
2. k=2
3. while k M T
99
4.
for l = k 1,
,1
5.
= round ( R ( l , k ) / R ( l , l ) )
6.
if 0
7.
8.
end
9.
10.
end
if R ( k 1, k 1) > R ( k , k ) + R ( k 1, k )
2
12.
13.
14.
G=
R ( k , k 1)
R ( k 1: k , k 1)
where
R ( k 1, k 1)
R ( k 1: k , k 1)
15.
R ( k 1: k , k 1: M T ) := GR ( k 1: k , k 1: M T )
16.
Q (:, k 1: k ) = Q (:, k 1: k ) G T
17.
k := max {k 1, 2}
18.
else
19.
k := k + 1
20.
end
and
21. end
100
Figure 5.2
101
Figure 5.3
r = Hx + n = ( HT ) ( T1x ) + n
= Hz + n
(5.10)
It can be considered that the channel matrix is H and the transmitted signal is z.
Thus, the ML detection problem given in (1.8) becomes:
z ML = arg min r Hz
(5.11)
x=T1z
z = T-1 x
x {3, 1,1,3}
(5.12)
(5.13)
In order to derive the SDR problem, the optimization problem (5.12-5.13) should
be formulated in a higher dimension. Thus, we define a rank-one semidefinite
matrix Z, which is given by:
102
Z = z T 1 z T 1
T1x ( T1x )T
=
T
1
( T x )
T1x
T1xxT ( T1 )T
=
T 1 T
x ( T )
T 1x
(5.14)
(5.15)
HT H HT r
= min Tr Z T
r H rT r
Z1,2
Z 2,2
(5.16)
st.
HT H HT r
min Tr Z T
r H rT r
(5.17)
(5.18)
1e diag ( TZ1,1TT ) 9e
(5.19)
Z 2,2 = 1
(5.20)
rank ( Z ) = 1
(5.21)
103
It is well known that the extremely high complexity of the ML detection is caused
by the presence of the non-convex rank-one constraint (5.21). Thus, relaxation of
this constraint will be engaged to transform the non-convex problem into a
semidefinite relaxation problem:
st.
HT H HT r
min Tr Z T
r H rT r
(5.22)
(5.23)
1e diag ( TZ1,1TT ) 9e
Z 2,2 = 1
(5.24)
(5.25)
The SDR detector can be solved by any of the SDP solvers, such as Sedumi
[35] which is based on the interior-point methods [52].
The problem (5.22-5.25) is not exactly equivalent to the problem (5.12-5.13)
since the rank-one constraint (5.21) has been relaxed, and the constraint (5.24) is a
sufficient but non-necessary condition of the constellation of 16-QAM given in
(5.13). Therefore, the solution obtained from solving the problem (5.22-5.25) may
have more errors than ML decoding. Once the solution of problem (5.22-5.25) is
to
solved by the SDP solver, the randomization approach is applied to Z
2,1
recover z .
After the solution z is obtained, the estimate of x is given by x = TQ {z } .
Since the lattice reduction is based on the integer lattice, Q {i} is the quantization
operation to the integer lattice. However, the quantization does not fit the
boundary region of the constellation of x, and the final solution x may be outside
the boundary of the constellation and will result in errors. Thus, the solution
104
( )
( )
infinite integer space , the lattice L ( H ) and the reduced lattice L H represent
the same lattice. However, the transmitted signal of QAM modulation is limited in
a constellation space and does not fulfill the latter assumption. Thus should
be considered as an amendment version of a subset of the infinite integer
space [79-81]:
= a + 1MT
2
M
M
where =
,
+ 1,
2
2
(5.26)
the QAM energy normalization parameter. Therefore, the transmitted signal can
be expressed as:
1
x = a + 1MT
2
(5.27)
= aT1 + 1MT
2
= a z + T11M T
2
(5.28)
In order to obtain the final result x , z is shifted and quantized in integer region:
1
1
1
z sq = a Q z T11MT + T11MT
2
2
a
(5.29)
1
x = Tz sq = aT Q z T11M T + T11MT
2
2
a
(5.30)
105
(5.31)
z = T-1 x
st.
x 2
+ 1, 2
+ 3,
,2
(5.32)
st.
HT H HT r
min Tr Z T
r H rT r
(5.33)
(5.34)
0
2
q
1e diag ( TZ1,1TT ) 2 2 1 e
(5.35)
Z 2,2 = 1
(5.36)
rank ( Z ) = 1
(5.37)
st.
HT H HT r
min Tr Z T
r H rT r
(5.38)
(5.39)
0
2
q
1e diag ( TZ1,1TT ) 2 2 1 e
(5.40)
Z 2,2 = 1
(5.41)
106
(5.42)
= 1 5 3 2 4
where 1 5 3 = [ 3, 1] [1,3] , 2 = [ , 3] [ 1, + ] ,
4 = [ ,1] [3, + ] .
Fig. 5.4 shows the interval of these three sets: 1 5 3 , 2 and 4 .
Apparently, it can be observed that the intersection set of these three sets is
exactly the alphabet set of 16-QAM: {3, 1,1,3} .
These sets can also be expressed by two quadratic inequalities equations:
107
Figure 5.4
(5.43)
1 xi 2 9
(5.44)
xi 2 + 4 xi + 3 0
(5.45)
xi 2 4 xi + 3 0
(5.46)
st.
z = T-1 x
(5.47)
(5.48)
1 ( Tz )i 9
(5.49)
( Tz )i
+ 4 ( Tz )i + 3 0
(5.50)
( Tz )i
4 ( Tz )i + 3 0
(5.51)
108
st.
(5.52)
(5.53)
1e diag ( TZ1,1TT ) 9e
(5.54)
(5.55)
(5.56)
Z 2,2 = 1
(5.57)
rank ( Z ) = 1
(5.58)
But it is not convex due to the rank-one constraint, so we eliminate the constraint:
HT H HT r
min Tr Z T
r H rT r
st.
(5.59)
(5.60)
1e diag ( TZ1,1TT ) 9e
(5.61)
(5.62)
(5.63)
Z 2,2 = 1
(5.64)
At this point, the SDP solvers could be used to solve the problem and obtain
, then the randomization procedure is applied to Z
the solution Z
2,1 to recover z .
Consequently, the final solution x is obtained by equation (5.30).
For higher level QAM MIMO system, the LR-SDR problem is formulated as:
HT H HT r
min Tr Z T
r H rT r
(5.65)
109
st. Z
(5.66)
1e diag ( TZ1,1TT ) 9e
q +1
diag ( TZ1,1TT ) 2 2 + 4i + 4 TZ1,2
, i = 0,1,
q
q
2
2
+ 2 + 2i + 1 2 + 2i + 3 0
(5.67)
, 2
(5.68)
Z 2,2 = 1
(5.69)
The problem can be solved by common SDP solver Sedumi, then the
110
111
(a)
(b)
Figure 5.5
(a)
(b)
Figure 5.6
113
(a)
Figure 5.7
(b)
The BLER performance of the LR-SDR detectors, SDR detector,
SD and LR-ZF detector using 256-QAM.
(a) 4 4 MIMO systems. (b) 6 6 MIMO systems.
114
Computational time
BC- SDR
0.32s
LR-SDR1
0.37s
LR-SDR2
0.40s
Computational time
BC- SDR
0.36s
LR-SDR1
0.44s
LR-SDR2
0.48s
Computational time
BC- SDR
0.32s
LR-SDR1
0.37s
LR-SDR2
0.49s
Computational time
BC- SDR
0.36s
LR-SDR1
0.44s
LR-SDR2
0.60s
115
Computational time
BC- SDR
0.32s
LR-SDR1
0.37s
LR-SDR2
0.67s
Computational time
BC- SDR
0.36s
LR-SDR1
0.44s
LR-SDR2
0.83s
5.4 Discussion
The LR-SDR detection algorithms presented in this chapter are inspired by
the LR technology combined with the linear detectors. It has been introduced in
Chapter 2 that the decision region of ZF decoder can be dramatically improved by
combining LR technology and the LR-ZF decoder can offer much better BLER
performance than ZF decoder. The proposed LR-SDR detection algorithms are
also able to provide significantly improvement in BLER performance compared
with the sole SDR detector, and with only a little complexity added. However,
since the SDR MIMO detection method is an empirical method, it is quite difficult
to theoretically analyze how its BLER performance is affected and by what
factors. Nevertheless, it is for sure that the performance of SDR decoders is
influenced by the condition of the channel. By using LR to improve the channel
116
condition, the SDR decoder could offer better BLER performance. This has been
fully verified by the simulation results presented in section 5.3. For the
completeness of the theory, it would be better if this improvement can be
supported by theoretical analysis. In our future work, we will try to investigate
this problem.
117
CHAPTER 6
CONCLUSIONS AND RECOMMENDATIONS
6.1 Conclusions
The background, system model, working principle and detection design
details of MIMO systems have been comprehensively analyzed and simulated in
this thesis. The main achievements of the thesis can be summarized as follows:
1.
2.
118
3.
4.
The
lattice-reduction-aided
semidefinite
relaxation
detection
algorithms are proposed. This kind of detectors can reach the full
diversity order and provide near optimum BLER performances. The
LR-SDR detectors have very low polynomial-time complexity. Its
computational time keeps unchanged with different SNR conditions
and is only a little longer than that of the SDR detector.
6.2 Recommendations
Some suggestions for future research are:
1.
2.
The software solver of the SDR detectors used in this thesis is based
on the interior-point method. New improved interior-point methods to
reduce the iteration complexity can be investigated.
3.
119
120
List of Figures
LIST OF FIGURES
Figure 1.1 Structure of MIMO system....................................................................3
Figure 1.2 MIMO space-time block code system. ..................................................4
Figure 1.3 MIMO spatial multiplexing system.......................................................6
Figure 1.4 Sphere decoding. .................................................................................11
Figure 2.1 The tree search structure of sphere decoding ....................................24
Figure 2.2 The original transmit signal lattice. .....................................................28
Figure 2.3 The received signal lattice. ................................................................29
Figure 2.4 ZF decoding decision region. ..............................................................29
Figure 2.5 SIC detection decision region..............................................................30
Figure 2.6 ML decoding decision region. .............................................................31
Figure 2.7 LR-ZF decision region.........................................................................32
Figure 3.1 Probability density function of the received signal vector of ZF
detections in 2 2 MIMO systems.
(a) Case for good channel condition. (b) Case for bad channel condition. ..........38
Figure 3.2 Elliptic paraboloid with axis perpendicular to a subspace spanned by
lattice points...........................................................................................................39
Figure 3.3 Elliptic paraboloid in 3-dimensional space. ........................................40
Figure 3.4 2-D lattice space example....................................................................46
Figure 3.5 3-D lattice space example....................................................................48
Figure 3.6 Comparison of BLER performance of ESA, ML decoding and ZF
using 4-QAM.
(a) 4 4 MIMO systems. (b) 8 8 MIMO systems. .........................................50
Figure 3.8 Comparison of BLER performance of ESA, ML decoding and ZF
121
List of Figures
using 16-QAM.
(a) 4 4 MIMO systems. (b) 8 8 MIMO systems. .........................................51
Figure 3.9 Comparison of BLER performance of ESA, ML decoding and ZF
SD. .........................................................................................................................92
Figure 5.1 Example of LLL algorithm..................................................................98
122
List of Figures
123
List of Tables
LIST OF TABLES
Table 3.1 Number of lattice points visited using ML decoding/proposed GD
(reduction is indicated as percentage)....................................................................52
Table 4.1 The values of x. .....................................................................................77
Table 5.1 LLL Lattice Reduction Algorithm.........................................................99
Table 5.2 The computational time of 4 4 16-QAM MIMO systems. ............115
Table 5.3 The computational time of 6 6 16-QAM MIMO systems.............. 115
Table 5.4 The computational time of 4 4 64-QAM MIMO systems.............. 115
Table 5.5 The computational time of 6 6 64-QAM MIMO systems. ............ 115
Table 5.6 The computational time of 4 4 256-QAM MIMO systems. .......... 116
Table 5.7 The computational time of 6 6 256-QAM MIMO systems. .......... 116
124
Abbreviations
Abbreviations
2G
second-generation
3G
third-generation
4G
fourth-generation
MIMO
multiple-input multiple-output
SM
spatial multiplexing
DBLAST
VBLAST
LTE
long-term-evolution
WLAN
UWB
ultra-wide-band
CR
cognitive radio
STC
space-time coding
STTC
STBC
SNR
ML
maximum likelihood
NP-hard
BER
bit-error rate
EUB
LDC
linear-dispersion code
SD
sphere decoding
FSD
AWGN
PEP
pair-wise-error probability
OSTBC
ML
maximum likelihood
ZF
zero-forcing
MMSE
Abbreviations
SD
sphere decoding
SIC
LRD
SDR
semidefinite relaxation
F-P
Fincke-Pohst
S-E
Schnorr-Euchner
LLL
LenstraLenstraLovsz
SDP
semidefinite programming
BPSK
QAM
PI-SDR
polynomial-inspired SDR
BC-SDR
bound-constrained SDR
VA-SDR
virtually-antipodal SDR
BLER
block-error rate
MSE
mean-square-error
ISI
inter-symbol-interference
SINR
LR
lattice reduction
ESA
QCQP
IPM
interior-point method
LR-SDR
lattice-reduction-aided SDR
126
References
REFERENCES
127
References
128
References
129
References
2004.
[26]. C. P. Schnorr and M. Euchner, Lattice basis reduction: Improved practical
algorithms and solving subset sum problems, Math. Program., vol. 66, no.
13, pp. 181199, Aug. 1994.
[27]. C.
Windpassinger,
L.
Lampe,
and
R.
F.H.
Fischer,
From
130
References
131
References
132
References
semidefinite programs and cutting planes, Math. Program., vol. 82, no. 3,
pp. 219315, 1998.
[51]. W.-K. Ma, P.-C. Chung and Z. Ding, Semidefinite relaxation based
multiuser detection for M-Ary PSK multiuser systems, IEEE Trans. on
Signal Processing, vol. 52, No. 10, pp. 2862- 2872, Oct. 2004.
[52]. C. Helmberg, F. Rendl, R. J. Vanderbei, H. Wolkowicz, An interior-Point
method for semidefinite programming, SIAM J. Optim., vol. 6, no. 2,
342-361, 1996.
[53]. B. M. Hochwald and S. ten Brink, Achieving near-capacity on a
multiple-antenna channel, IEEE Trans. on Communications, vol. 51, no. 3,
pp. 389399, Mar. 2003.
[54]. N. Karmarkar, A new polynomial-time algorithm for linear programming,
Combinatorica, 4 (1984), pp. 373-395.
[55]. M. Kojima, N. Megiddo, T. Noma, and A. Yoshise, A united approach to
interior point algorithms for linear complementarity problems, Lecture
Notes in Computer Science, Springer-Verlag, 1991.
[56]. Y. Nesterov and A. Nemirovsky, A general approach to polynomial-time
algorithms design for convex programming, Tech. report, Centr. Econ. &
Math. Inst., USSR Acad. Sci., Moscow, USSR, 1988.
[57]. Optimization over positive semidefinite matrices: Mathematical background
and user's SEMIDEFINITE PROGRAMMING 49 manual, USSR Acad. Sci.
Centr. Econ. & Math. Inst., 32 Krasikova St., Moscow 117418 USSR, 1990.
[58]. Self-concordant functions and polynomial time methods in convex
programming, Tech. Rep., Centr. Econ. & Math. Inst., USSR Acad. Sci.,
133
References
134
References
135
References
Windpassinger
and
R.
Fischer,
Low-complexity
Windpassinger,
R.
F.
H.
Fischer,
and
J.
B.
Huber,
136
References
137
Publications
PUBLICATIONS
Book Chapter:
[1]. Z. Y. Shao, S. W. Cheung and T. I. Yuk, Geometrical detection algorithm for
MIMO systems, MIMO Systems, Theory and Applications, InTech, April
2011.
Articles:
[1]. Z. Y. Shao, S. W. Cheung and T. I. Yuk, A Simple and Optimum Geometric
Decoding Algorithm for MIMO Systems, The 4th International
Symposium on Wireless Pervasive Computing (ISWPC), Melbourne,
Australia, pp. 1-5, Feb. 2009.
[2]. Z. Y. Shao, S. W. Cheung and T. I. Yuk, A fast geometric decoding
algorithm for MIMO systems, The 11th International Conference on
Advanced Communication Technology (ICACT), Korea, vol. 3, pp.
2108-2111, Feb. 2009.
[3]. Z. Y. Shao, S. W. Cheung and T. I. Yuk, Semi-definite relaxation decoder
for 256-QAM MIMO system, IET Electronics Letters, vol. 46, no. 11, pp.
796-797, 2010.
[4]. Z. Y. Shao, S. W. Cheung and T. I. Yuk, Combined semi-definite relaxation
and sphere decoding method for multiple antennas systems, 2011
International ITG Workshop on Smart Antennas (WSA), Aachen, Germany,
pp. 1-4, Feb. 2011.
[5]. Z. Y. Shao, S. W. Cheung and T. I. Yuk, Comparison of semidefinite
relaxation detection for 256-QAM MIMO system, submitted to Signal
138
Publications
Processing.
[6]. Z. Y. Shao, S. W. Cheung and T. I. Yuk, Lattice-Reduction-Aided MIMO
Detections Based on Semidefinite Relaxation, submitted to IEEE Signal
Processing Letter.
139