You are on page 1of 6

MMSE-Based Lattice-Reduction-Aided

Fixed-Complexity Sphere Decoder


for Low-Complexity Near-ML MIMO Detection
Hyunsub Kim, Student Member, Hyukyeon Lee, Student Member, and Jaeseok Kim, Senior Member, IEEE
Department of Electrical and Electronic Engineering,
Yonsei University, Seoul 120-749, Korea
E-mail: {sarapio, hyukyeon, jaekim}@yonsei.ac.kr

AbstractIn this paper, we propose a minimum-meansquared-error (MMSE)-based lattice-reduction (LR)-aided fixedcomplexity sphere decoder (FSD) for low-complexity nearmaximum-likelihood (near-ML) multiple-input multiple-output
(MIMO) detection. In order for the FSD to achieve optimal
performance, the number of full expansion (FE) stages should
be sufficient, which is the major cause of the increase in the
computational complexity when either a large signal constellation
or a large number of antennas are adopted. However, the
proposed algorithm maintains the near-ML performance with
the aid of the MMSE-based LR algorithm while reducing the
number of FE stages. Although there exists the increase in the
computational complexity for the application of the additional
processing elements, the decrease in the number of FE stages
results in the lower computational complexity of the overall
algorithm. The numerical analysis demonstrates that there is a
considerable decrease in the computational complexity while the
performance degradation is negligible, compared to the optimal
FSD.
Index Termslattice reduction (LR), fixed-complexity sphere
decoder (FSD), multiple-input multiple-output (MIMO)

I. I NTRODUCTION
After the adoption of the IEEE 802.11n standard, multipleinput multiple-output (MIMO) communication system has
become the solution to the considerable demands for higher
data rate and more reliable wireless local area networks
(WLANs). However, the high computational complexity of
its receiver is a major obstacle to the implementation of an
enhanced MIMO system. Although the maximum-likelihood
(ML) decoder (MLD) is known as the optimal solution to the
MIMO symbol detection problem, the exponential increase in
its complexity makes it infeasible when either a large signal
constellation or a large number of antennas are employed.
Hence, it is one of the most challenging issues in MIMO
technique to reduce the complexity of the MIMO reciever
while achieving comparable bit-error-rate (BER) performance
to the MLD.
Some decoders such as sphere decoder (SD) [1],[2] and
semi-definite programming (SDP) [3][5] have been proposed
to achieve the (near-)ML performance with reduced complexity. However, the variable complexity of these receivers is a
major drawback for practical systems where the data should
978-1-4673-6762-2/15/$31.00 2015 IEEE

be processed at a constant rate. In this regards, the K-best


decoder [6] and the fixed-complexity SD (FSD) [7] have been
developed to achieve the constant throughput by removing the
feedback in the data flow. In particular, the FSD can achieve
the optimal performance in a fixed number of operations, by
performing the full expansion (FE) for some levels and single
expansion (SE) for the remaining levels of the tree with a
specific ordering of the channel matrix in the pre-process.
Nevertheless, the complexity of the FSD is still high when
a large signal constellation or a large number of antennas are
employed.
Recently, the lattice reduction (LR) algorithm has emerged
as a promising pre-processing technique that can enhance
the diversity order with reduced complexity by transforming
the system model to the lattice-reduced one whose channel
matrices are near-orthogonal [8]. LR-aided (LRA) linear decoders have been proposed as an alternative, offering as many
diversity order as MLD [9][11]. Moreover, their complexities
are close to those of the linear decoders when the channel
remains constant for a frame. However, considerable gap still
exists between the performance of the optimal decoder and
those of the LRA linear decoders as the number of transmit
antennas increases or a large signal constellation is employed.
In this paper, we propose a minimum-mean-squared-error
(MMSE)-based LRA FSD that achieves near-ML BER performance while a large number of antennas or a large signal
constellation are employed. While the complexity is reduced
with the lower number of tree levels in the FE stage, the
proposed algorithm maintains the near-ML performance with
the aid of the MMSE-based LRA successive interference
cancellation (SIC) in the SE stage. As the major obstacle to the
application of the LR is that the bases of the channel matrix are
modified after the LR, an efficient way to apply the MMSEbased LR to the FSD is proposed. The numerical results show
that the computational complexity of the proposed method is
much lower than that of the corresponding FSD, while the
performance degradation is negligible.
The remainder of this paper is organized as follows. In
Section II, the MIMO system model and the MIMO decoders
involved with the proposed algorithm are briefly introduced.
In Section III, the proposed MIMO detection algorithm is

root

i=2

i=3

i=4

i=8

Find Minimum Path Metric

In this subsection, we briefly introduce the MIMO system model. Consider a flat-fading MIMO system with NT
transmit and NR receive antennas, where NT NR . When
sc = [sc,1 , sc,2 , , sc,NT ]T denotes the transmitted symbol
vector, the NR 1 received symbol vector at one sample time
can be expressed as
yc = Hc sc + nc ,

(1)

where nc ( CNR 1 ) denotes the additive white Gaussian


noise (AWGN) vector with a zero mean and a covariance
matrix E[nc nc H ] = n2 INR , and Hc ( CNR NT ) represents
the channel matrix whose elements are independent and identically distributed (i.i.d) complex Gaussian coefficients with a
zero mean and a unit variance. We assume that the total power
of every antenna is normalized to one, i.e., E[sc H sc ] = 1, and
the channel matrix Hc is known at the receiver.
For simplicity, the complex MIMO system is converted to
the real MIMO system model by performing the real value
decomposition (RVD) to the system model in (1) as







R(H) I(H) R(s)
R(n)
R(y)
=
+
. (2)
I(H) R(H)
I(s)
I(n)
I(y)
|
| {z }
{z
} | {z }
| {z }
H(R2NR 2NT )

s(2NT )

n(R2NR )

Then,
sM L , the MLD solution for (2), is given by

sM L =

A. MIMO System Model

i=1

SE
stage

II. P RELIMINARIES

y(R2NR )

FE
stage

described, and the numerical results are discussed in Section


IV. Finally, the conclusions are drawn in Section V.
Notations: The superscript ()T denotes the transpose of
a matrix. Upper (lower) boldface letters are used for matrices
(vectors). C and R denote the set of complex and real numbers.
Also, denotes the set of real parts of the signal constellation
points. IN denotes the N N identity matrix. || is the
absolute value of scalar or cardinality of if is a set.
k k and dc represent the 2-norm of a vector and the rounding
operation, respectively.

min ky Hsk2 .

arg

Fig. 1. The tree structure of the real FSD in a 4 4 MIMO system with
||2 -QAM and p = 2. The tree node which is colored by gray is the one
that contains the FSD solution.

where GZF is the ZF filter matrix, Q() denotes the slicing


(quantization to a constellation point) operation in the real
:= H n denotes the noise amplified after linear
domain, n
equalization and H is the Moore-Penrose pseudo-inverse of
the channel matrix which can be written as
H = (HT H)1 HT .

(7)

However, the noise amplification in (5) is the major cause that


degrades the BER performance of the linear decoders.
The MMSE decoder minimizes the mean squared error
(MSE) between the transmitted symbols and the output of the
linear decoder. The MMSE filter matrix is given by [12]
GM M SE = (HT H + n2 INT )1 HT .

(8)

The MMSE decoding is performed in the same manner as in


(5) and (6) with the MMSE filter matrix instead of the ZF
filter matrix.
C. FSD

(3)

s 2NT

While this MLD is optimal, the search space is proportional to


||NT . This exponential complexity makes the MLD infeasible
for practical systems.
B. Linear Decoders
In this subsection, we recall the linear detection algorithms
with respect to the zero-forcing (ZF) and MMSE criterion. The
ZF decoder is the simplest linear decoder whose complexity
is far lower than that of the MLD. The ZF decoding can be
formulated as
GZF = H ,

(4)

sZF = GZF y = s + H n = s + n
,

(5)

The channel matrix H can be decomposed as H = QR,


where Q ( R2NR 2NT ) is a unitary matrix and R (
R2NT 2NT ) is an upper triangular matrix. By multiplying both
sides of (2) by QT , the system model can be rewritten as
q = Rs + v,

(9)

where q = QT y and v = QT n. As it is illustrated in Fig. 1,


the FSD performs a constrained tree-searching on (9) which
consists of the FE and SE stages. For the first p levels, FE
is performed, where all possible || branches are expanded.
Then, for the remaining 2NT p levels, SE is performed,
where only one branch is expanded on each node in the manner
of SIC. In other words, the FSD solution is given by
sF SD = arg min kq Rsk2 ,

(10)

sL

where L is the candidate list which is generated as


sZF = Q(sZF ) =

arg
s 2NT

min |sZF s|,

(6)

L = [s1 , s2 , , s||p ],

(11)

where sl = [
sl,1 , , sl,NT ]T and

, i = 2NT ,
, 2NT p + 1

2N
T
X
sl,i =
(12)
Q (qi
ri,j sl,j )/ri,i ,
else,

Constellation Points

j=i+1

where qi and ri,j are the i-th element in q and the (i, j)-th
element in R. In order to achieve the optimal performance, the
channel matrix should be ordered prior to the tree-searching so
that the signals with maximum and minimum post-processing
noise amplifications should be detected at the FE and SE
stages, respectively [7].

-1

-2

-2

(a)

-1

-2

-2

(b)

Fig. 2. An example of the possible integer values in (a) the original domain
and (b) the lattice-reduced domain with 16-QAM.

D. Lattice-Reduced System Model


To employ the lattice-reduced system model, the received
signal is first scaled and shifted to map the received symbol
to the consecutive integer lattice as

y = z + H
n = z + w,
LRZF = H
z

(16)

sLRZF = (Td
zLRZF c 1).

(17)

Note that w := H n in (16) is the noise amplified by


. With the aid of the
the lattice-reduced channel matrix H
near-orthogonal nature of the lattice-reduced matrix, the noise
amplification in (16) is much less than (5) so that the BER
performance of the LRA linear decoders has the same diversity
order as that of the MLD.
III. P ROPOSED MIMO D ETECTION A LGORITHM
In this section, we introduce the proposed low-complexity
MIMO detection algorithm with MMSE-based LRA FSD.
As all the constellation points are expanded at the FE stage
of the FSD, the number of the parent nodes is too large,
which is the main cause of the increase in the complexity
of FSD. Thus, in order to reduce the complexity, the proposed
algorithm lowers the number of tree levels of the FE stage

QR Decomposition

i=2

FE Signal Cancellation

SE
Stage
(LRAMMSESIC)

i=3

i=4

Then, the LRA ZF decoder can be formulated as

i=1

(15)

FE
stage

+ n.

y = HTT1 x + n = Hz

LRA-MMSE
Matrix

1
1
1
(14)
y + H1 = H( s + 1) + n = Hx + n,

where n = 1 n.
Through the lattice basis reduction on H, a near-orthogonal
= HT can be obtained, where H
spans
basis matrix H
the same lattice as H and T is an integer unimodular matrix. Throughout this paper, we consider the Lenstra-LenstraLovasz (LLL) algorithm [13] as the LR algorithm which
Let
is used to acquire the near-orthogonal basis matrix H.
1
z = T x, then the lattice-reduced system model can be
represented as
y ,

Lattice
Reduction

root

1
s + 1,
(13)

where 1 = [1, ..., 1]T , is the minimum distance between


QAM constellation points. The scaled and shifted received
symbol vector can be written now as
x=

Channel Ordering

i=8

Post-processing & ML Test

Fig. 3. The tree structure of the proposed MIMO detection algorithm in a


4 4 MIMO system with ||2 -QAM and p = 2. The tree node which is
colored by gray is the one that contains the solution of the proposed algorithm.

to reduce the number of the parent nodes generated at the


FE stage. Nevertheless, the proposed algorithm maintains the
near-optimal BER performance by adopting the LRA MMSESIC at the SE stage.
However, it is not straightforward to apply the LR algorithm
to the MIMO detection algorithm based on the tree-searching
where the child nodes should be expanded from the predetermined constellation points. As it is illustrated in Fig. 2,
when the LR algorithm is applied, the bases of channel matrix
are transformed from the original constellation domain to the
lattice-reduced domain, which means that the predetermined
constellation points can not be obtained. As the calculation of
all possible values in the lattice-reduced domain has almost
the same complexity as the MLD [14], a simple solution is
proposed in this algorithm.
In Fig. 3, the tree structure of the proposed algorithm is
illustrated. The proposed algorithm consists of the four steps
including i) channel ordering, ii) FE signal cancellation, iii)
SE stage and iv) post-processing & ML test. The detailed
explanation of each step is as follows:
Step 1 Channel Ordering: First, the channel matrix H
is ordered in the same manner as the ordering in the FSD [7].

l=2NT p+1

0
y(k)

10

10

10

10

10

10

12

14

16

18

20
22
Eb/No [dB]

24

26

28

30

Fig. 4. Comparisons of the uncoded BER performance in the 4 4 MIMO


system with 64-QAM. p = 2 for optimal FSD, p = 1 for non-optimal FSD,
and p = 1 for the proposed algorithm.

2NR 1

0
y(k)
, H0 s0(k) + n,

s0(k)

10

10

(2NT p)1

0T y 0 = R
0 z(k) n Q
0T z(k) +Q
0T n
0(k) , Q
q
(k)
1
2
1
| {z }

(21)

interference

SD (MLD)
FSD (optimal)
FSD (nonoptimal)
LRAMMSE
LRAMMSESIC
Proposed

( R2NR 2NT ) and Q


( R2NR 2NT ) are unitary
where Q
1
2
0
2N
2N
T
( R T
matrices, and R
) is an upper triangular matrix.
0
Note that y (k)
( R2NR 1 ) is the nulled received signal
vector which is scaled and shifted to map the received symbol
to the consecutive integer lattice as in (14). Also, although
there exists the interference which remains in the MMSE 0T z(k) , the decrease in the noise
SIC system model, n Q
2
amplification has more influence so that the BER performance
is enhanced.
Finally, the SE candidate signals for the remaining 2NT p
levels are generated in the manner of SIC as
(2) , , z
(||2p) ],
L = [
z(1) , z
(22)

BER

2NR (2NT p)

10

(19)

where H ( R
) and
(
)
are the channel matrix and the transmitted signal whose lth (l = 2NT p + 1, , 2NT ) column vectors and elements
are nulled.
Step 3 SE stage (LRA-MMSE-SIC): In this step, LRAMMSE-SIC is performed to the system model in (19) to complete the candidate list. First, the LR algorithm is performed
to the channel matrix, H0 , to obtain the unimodular channel
for lattice-reduced system model, T0 . Then, the system model
to carry out the LRA-MMSE-SIC is formulated as
" 0#
 0 0

Q
0
0 0
HT

0,

1 R
=QR =
(20)
H =
0

n T0
Q2

SD (MLD)
FSD (optimal)
FSD (nonoptimal)
LRAMMSE
LRAMMSESIC
Proposed

10

where
( R
) is the received signal in the revised
system model, hl ( R2NR 1 ) is the l-th column vector of the
channel matrix H, and sl(k) ( ) is the k-th FE candidate
signal transmitted from the l-th transmit antenna. Finally, the
system model is transformed to a nulled system model as

BER

In this way, the signals with maximum post-processing noise


amplifications are detected at the first p levels, and those with
the minimum post-processing noise amplifications are detected
first at the remaining 2NT p levels. Because the channel
ordering step is common to all the detection schemes based
on the tree searching algorithm, for convenience, it is assumed
that the optimal ordering is [1, 2, , 2NT ].
Step 2 FE Signal Cancellation: As it is mentioned above,
the main obstacle in the application of the LR algorithm to
the FSD is the lack of pre-determined candidate nodes in the
lattice-reduced domain. That is why, we generate the candidate
signals in the FE stage, and cancel the FE signals in the
original constellation domain before transforming the system
model to the lattice-reduced one.
In other words, the FE candidate signals are canceled and
nulled as
2N
XT
0
y(k) = y
hl sl(k) ,
(18)

10

10

10

16

18

20

22

24

26
28
Eb/No [dB]

30

32

34

36

Fig. 5. Comparisons of the uncoded BER performance in the 4 4 MIMO


system with 256-QAM. p = 2 for optimal FSD, p = 1 for non-optimal FSD,
and p = 1 for the proposed algorithm.

T
(l) = [
where z
z(l),1 , , z(l),(2NT p)
] and

P2NT p 0
q0(l),i j=i+1
ri,j z(l),j

, i = 2NT p, (23)
z(l),i =
1,
0

ri,i

0(l) and the (i, j)where q0(l),i and r0i,j are the i-th element in q
0.
th element in R
Step 4 Post-processing & ML test: The post-processing
step is performed to the SE candidate signals, which involves
i) the conversion of the candidate vectors to the original
constellation, ii) re-ordering of the detected vector and iii)
slicing. First, the candidate vectors in L are converted back to
the original constellation as in (17), and the candidate vectors
are re-ordered with the order in Step 1 reversed. Then, the

TABLE I
C OMPUTATIONAL COMPLEXITY IN TERMS OF THE NUMBER OF FLOP S .

Algorithm

Operation

No. of FLOPs

FSD

FSD ordering of H
QR decomposition of H
q = QT y
SIC
ML test

2 + 14N + 8N N + 4N + 9)
NR (NT 1)(2NR
R
T R
T
2 + 1 N3 )
4 (NT2 NR + NT NR
3 R
8NT NR
4 (NT p)(NT + p + 1)
8 (NT NR + NT ) ||p

FSD ordering of H
FE signal cancellation
LR of H0
0
QR decomposition of H
0T
y 0
0 = Q
q

2 + 14N + 8N N + 4N + 9)
NR (NT 1)(2NR
R
T R
T
8NT p||p
2 + 1 N3 )
6.4 ((NT p)2 NR + (NT p)NR
3 R
2
4 ((NT p) (2NR ) + (NT p)(2NR )2 + 13 (2NR )3 )

Proposed

(k)

(k)

SIC
ML test

8NT NR ||p
4 (NT p)(NT + p + 1)
8 (NT NR + NT ) ||p

post-processing is completed by slicing the estimated symbols


to the original constellation points.
This slicing operation is essential for achieving the optimal
performance because of the infinite characteristics of the lattice
space where the SE is performed. When these symbols are
converted to the original constellation, some symbols exist
outside the finite constellation set owing to the AWGN, which
amplifies the error in finding the candidate vector with the
minimum Euclidean distance (ED). Therefore, to compensate
for this error, the symbol vectors are sliced to the original
constellation prior to the ML test.
These SE candidate signals, along with the FE candidate
signals, formulate the candidate list as
L = [s0(1) , s0(2) , , s0(||2p) ],

(24)

where s0(l) = [
s0(l),1 , , s0(l),(2NT ) ]T is the l-th candidate
signal.
Finally, the detection algorithm is completed by the ML test
where the symbol vector with the minimum ED is detected as
the solution as

sLRAF SD = arg min ky Hs0 k2 .

(25)

s0 L

IV. N UMERICAL R ESULTS


A. BER performance
In this subsection, we verified the performance of the
proposed algorithm through Monte Carlo simulations. The
simulation was conducted on an uncoded 4 4 MIMO system
with 64-QAM and 256-QAM. We assumed that the channel
varies at every sample time, and all the lattice-reduced channel
matrices were obtained by the LLL algorithm [13].
In Fig. 4 and Fig. 5, the BER performances of the
MIMO detection algorithms are compared with 64-QAM and
256-QAM modulations, respectively. As the computational
complexity of the MLD is too large to conduct computer
simulations, the optimal performance was obtained by the
optimal SD instead. We can see that the LRA linear decoders
have considerable performance degradations with high order
modulations, compared to the optimal decoder.

Example for
4 4 MIMO system
with p = 2, p = 1
13, 929 (64-QAM)
44, 649 (256-QAM)

7, 937 (64-QAM)
10, 497 (256-QAM)

The optimal FSD has almost the same performance as that


of the SD (MLD). However, the diversity order (the absolute
number of the slope of the graph) of the non-optimal FSD is
much lower, because the number of FE stages is not sufficient.
On contrary, in spite of the insufficient number of tree levels
in the FE stage, the proposed scheme achieves the optimal
diversity order owing to the better-conditioned channel matrix
of the MMSE-based lattice-reduced system model in the SE
stage. Also, the performance degradation of the proposed LRA
FSD algorithm is less than 0.5dB at the BER of 105 with
both 64-QAM and 256-QAM.
B. Complexity Analysis
For fair comparisons, the number of floating point operations (FLOPs) is considered in this subsection. The number of
FLOPs is summarized in Table I, which is obtained according
to the following rules [15], [16]:
Multiplication of l m and m n real matrices requires
2 lmn FLOPs.
Moore-Penrose pseudo-inverse of an m n real matrix
requires 2m3 2m2 + m + 16mn FLOPs.
QR decomposition of an m n (m n) real matrix
requires 4 (mn2 + m2 n + 13 m3 ) FLOPs
The complexity of the complex LR is about 1.6 times
that of the QR decomposition.
In the proposed scheme, we can see that there are some
additional processes for the LRA detection including the FE
signal cancellation and the LR. However, as a large signal constellation is employed, the overall computational complexity
is highly dependent on the number of tree levels in the FE
stage (p or p) which is dominant in the complexity of the ML
test.
After all, as it is shown at the rightmost column of Table
I, the computational complexity of the proposed algorithm is
reduced to 56.98% and 23.51% with 64-QAM and 256-QAM
in terms of the number of FLOPs, respectively. This reduction
results from the decrease in the number of tree levels in the
FE stage, and will be increased as the number of transmit
antennas increases.

V. C ONCLUSION
In this work, we proposed a MMSE-based LRA FSD
algorithm which achieves near-ML performance with reduced
complexity. As the number of tree levels in the FE stage
is decreased, the proposed scheme has far lower computational complexity due to the reduced complexity in the ML
test which dominates the overall computational complexity.
Nevertheless, the proposed scheme achieves the near-ML
performance by performing the LRA-MMSE-SIC in the SE
stage owing to the better-conditioned channel matrix of the
MMSE-based lattice-reduced system model.
ACKNOWLEDGMENT
This work was supported by System LSI Division on
Samsung Electronics Co., LTD. This work was also supported
by the Industrial Core Technology Development Program
(10049009, Development of Main IPs for IoT and Image
Based Security Low-Power SoC) funded by the Ministry of
Trade, Industry & Energy.
R EFERENCES
[1] E. Viterbo and J. Boutros, A universal lattice code decoder for fading
channels, IEEE Trans. Inf. Theory, vol. 45, no. 5, pp. 16391642, Jul.
1999.
[2] M. O. Damen, H. E. Gamal, and G. Caire, On maximum-likelihood
detection and the search for the closest lattice point, IEEE Trans. Inf.
Theory, vol. 49, no. 10, pp. 23892402, Oct. 2003.
[3] P. H. Tan and L. K. Rasmussen, The application of semidefinite
programming for detection in CDMA, IEEE J. Sel. Areas in Commun.,
vol. 19, no. 4, pp. 14421449, Apr. 2001.
[4] W.-K. Ma, T. N. Davidson, K. M. Wong, Z.-Q. Luo, and P. C. Ching,
Quasi-maximum-likelihood multiuser detection using semi-definite relaxation with application to synchronous CDMA, IEEE Trans. Signal
Process., vol. 50, no. 4, pp. 912922, Apr. 2002.
[5] Z.-Q. Luo, W.-K. Ma, A. So, Y. Ye, and S. Zhang, Semidefinite
relaxation of quadratic optimization problems, IEEE Signal Process.
Mag., vol. 27, no. 3, pp. 2034, May 2010.
[6] K. wai Wong, C.-Y. Tsui, R.-K. Cheng, and W.-H. Mow, A VLSI
architecture of a K-Best lattice decoding algorithm for MIMO channels,
in IEEE Int. Symp. Circuits and Syst., 2002, vol. 3, pp. 273276.
[7] L. Barbero and J. Thompson, Fixing the complexity of the sphere
decoder for MIMO detection, IEEE Trans. Wireless Commun., vol. 7,
no. 6, pp. 21312142, Jun. 2008.
[8] H. Yao and G. W. Wornell, Lattice-reduction-aided detectors for MIMO
communication systems, in IEEE Global Telecommun. Conf., Nov. 2002,
vol. 1, pp. 424428.
[9] D. W
ubben, R. B
ohnke, V. K
uhn, and K. D. Kammeyer, Nearmaximum-likelihood detection of MIMO systems using MMSE-based
lattice reduction, in IEEE Int. Conf. Commun., Jun. 2004, vol. 2, pp.
798802.
[10] M. Taherzadeh, A. Mobasher, and A. K. Khandani, LLL reduction
achieves the receive diversity in MIMO decoding, IEEE Trans. Inf.
Theory, vol. 53, no. 12, pp. 48014805, Dec. 2007.
[11] J. Jald
en and P. Elia, DMT optimality of LR-aided linear decoders for
a general class of channels, lattice designs, and system models, IEEE
Trans. Inf. Theory, vol. 56, no. 10, pp. 47654780, Oct. 2010.
[12] D. W
ubben, R. B
ohnke, V. K
uhn, and K. D. Kammeyer, MMSE
extension of V-BLAST based on sorted QR decomposition, in Proc.
IEEE Veh. Technol. Conf., Oct. 2003, vol. 1, pp. 508512.
[13] A. K. Lenstra, H. W. Lenstra, and L. Lov
asz, Factoring polynomials
with rational coefficients, Math. Ann., vol. 261, pp. 515534, 1982.
[14] M. Shabany and P. Gulak, The application of lattice-reduction to the
K-Best algorithm for near-optimal MIMO detection, in IEEE Int. Symp.
Circuits and Syst., May 2008, pp. 316319.
[15] G. H. Golub and C. F. Van Loan, Matrix computations. The Johns
Hopkins University Press, 2013.

[16] K. Zu and R. de Lamare, Low-complexity lattice-reduction-aided regularized block diagonalization for MU-MIMO systems, IEEE Commun.
Lett., vol. 16, no. 6, pp. 925928, Jun. 2012.
[17] X. Ma and W. Zhang, Performance analysis for MIMO systems
with lattice-reduction aided linear equalization, IEEE Trans. Commun.,
vol. 56, no. 2, pp. 309318, Feb. 2008.
[18] E. Perahia, IEEE 802.11n development: history, process, and technology, IEEE Commun. Mag., vol. 46, no. 7, pp. 4855, Jul. 2008.

You might also like