Vous êtes sur la page 1sur 10

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2014.2361337, IEEE Transactions on Communications
1

Lattice-Reduction-Aided Conditional Detection for


MIMO Systems
Hossein Najafi and Mohamed Oussama Damen

AbstractWe introduce a low-complexity detector with nearoptimal performance for transmission over multi-antenna systems. By using lattice basis reduction for generating almost
orthogonal channel submatrices, we enhance the conditional optimization technique to implement a fast yet efficient detector. The
lattice-reduction-aided (LRA) conditional method is presented
as a general detection technique over fading channels to yield
significant saving in computational complexity while achieving
close to Maximum Likelihood (ML) error performance. By
employing the orthogonality defect factor as a universal measure
to select a near-orthogonal channel submatrix for conditional
detection, we implement efficient detectors for MIMO systems.
In particular, an almost optimal decoder with linear complexity
for the Golden code is presented over quasi-static channels.
Index TermsConditional detection, Golden code, lattice reduction, low-complexity detection, maximum likelihood detection,
MIMO systems, space-time codes.

I. I NTRODUCTION
Finding fast and efficient decoding methods for spacetime codes and high-rate MIMO transmission is an important
design problem in wireless communications. Decoders with
low computational complexity but yet, close to the optimal performance are challenging issues for practical implementation
of schemes with multiple transmit and receive antennas such
as the V-BLAST (Vertical Bell Labs Layered Space-Time)
transmission model [1] and the Golden code [2], [3]. The
low-complexity and powerful decoders are of special interest
since they have many applications for widely incorporated
MIMO schemes in the wireless standards such as 3GPP LTE,
IEEE 802.16 WiMAX and IEEE 802.11 WLANs. Motivated
by efficient search algorithms in the lattice theory, the multiantenna detection problem over fading channels can be translated into finding the closest point in the lattice formed by
the channel matrix [4][7]. Lattice basis reduction, such as
the LLL algorithm [8] as an efficient one, is widely used for
the implementation of the search algorithms. In a different
approach, by taking advantage of orthogonal sub-channels,
a low-complexity decoder for multiplexed designs was proposed in [9] where it was shown that optimal decoding can
be implemented by employing the conditional optimization
technique [10]. In [11], using the conditional decoder together
with selecting the best submatrix choice, a fast decoder with
quadratic complexity in the size of symbol constellation with
essentially ML performance was proposed for the Golden
The authors are with the Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, ON, Canada, N2L 3G1 (email:
{hnajafi, mdamen}@uwaterloo.ca).

code. Low-complexity and almost ML performance of conditional decoding has also been studied in [12] for several
space-time block codes (STBC).
In this work, we propose a fast universal detector1 for multiantenna systems by employing lattice reduction as a major
part in the conditional optimization. By taking advantage
of the near-orthogonal matrix at the LLL output, close to
optimal performance is achieved over fading channels with
significant saving in complexity. For the fast low-complexity
implementation of the conditional detector, especially for a
small dimension of the second subchannel, the use of lattice
reduction and effective selection of channel submatrices are
vital for the near-ML performance. To select the best available
channel submatrix, we use the orthogonality defect factor as
a general measure that perfectly fits the orthogonality requirement of the conditional optimization. For a given dimension of
the conditioned symbols, we compare the orthogonality defect
factor of the LLL output of the possible choices and select
the more orthogonal submatrix to implement the conditional
detection on it. We also discuss the use of more powerful but
still simple detection methods such as the decision-feedbackequalizer (DFE) at the first part of decoding, instead of the
zero-forcing (ZF), to improve the overall detection and close
the gap to the ML solution.
We apply the new detectors for three scenarios over quasistatic channels: the Golden code, the diagonal algebraic spacetime (DAST) block codes [13], [14] and the spatial multiplexing MIMO system (or the V-BLAST transmission model
where we send independent symbols over different transmit
antennas). For a fast detection, we aim to implement the
lattice-reduction-aided (LRA) conditional detector with O(N )
complexity, where N is the size of the employed signal
constellation. Therefore, we maximize the likelihood function
for a single constellation point conditioned on the rest of points
which are already estimated by the LRA methods. The fulldiversity near-ML performance is studied and also verified
by the numerical results. Additionally, various detectors with
close to ML error performances are presented for a larger
MIMO system. Finally, we consider the scenario where the
channel is not assumed to be constant during the transmission
of a frame and hence, the fading coefficients have temporal
correlation and change slowly through time. We employ the
adaptive detection method from [15] where it is shown that
significant saving in complexity is achievable with a minimal
performance degradation. Here, to save in the complexity
1 We use the terms detection and decoding interchangeably throughout
the paper.

0090-6778 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2014.2361337, IEEE Transactions on Communications
2

of the LRA conditional decoding, we use the previous LLL


results together with the previous selection of the best nearorthogonal columns of channel matrix and adaptively update
the best reduced submatrix.
This paper is organized as follows: In Section II, the system
model along with lattice reduction and conditional detection are reviewed. In Section III, we present the application
of lattice reduction in conditional optimization over fading
channels. In addition, the optimal achievable diversity of the
proposed method is studied. In Section IV, the fast decoder
for the Golden code and the DAST block codes is discussed.
Section V describes the adaptive detector over correlated fading channels. Simulation results for various implementations
of the fast conditional detectors are presented in Section VI.
This paper is concluded in Section VII.
II. S YSTEM M ODEL
Several wireless communication systems can be represented
by a model where linear combinations of transmitted symbols
are corrupted by additive noise at a receiver node. As an example, assume the spatial multiplexing MIMO system with M
transmit and K receive antennas. Let x denote the transmitted
vector with symbols taken from a QAM constellation, X , with

size N and with average energy M


where is the observed
SNR at any receive antenna. The received signal y takes the
form
y = Hx + n,
(1)
where the channel matrix H = [hk,m ] has independent and
identically distributed (i.i.d) components with zero mean and
unit variance complex Gaussian distribution and known at
the receiver side. The noise vector n has i.i.d components as
zero mean, unit variance complex Gaussian random variables.
Assuming the transmitted symbols are distributed uniformly
on the constellation, the maximum-likelihood (ML) solution
to the MIMO system is given by
= arg min ||y
x
x2X M

Hx||2 ,

(2)

where X M denotes the M -dimensional hyper-cube with components from QAM constellation. Additionally, the model can
be transformed to a real counterpart by using the proper
transformations defined for vectors and matrices. The real
transmitted vector components are then taken from the equivalent PAM constellation. In general, finding the optimal solution
for the optimization problem in (2) has prohibitive complexity
and hence, approximate methods have been presented to offer
significant saving in computational complexity over fading
channels.
A. Lattice Reduction Application in MIMO Detection
Considering the MIMO channel model with an appropriate
translation and scaling, one can map the equivalent PAM
(or QAM) signals to integers where the transmitted symbols
at each signal
1}
p dimension are in ZQ = {0, 1, ..., Q
where Q = N for the assumed square QAM constellation.
Moreover, if the search boundaries of the ML problem are
relaxed to be in the integers Z instead of ZQ , then, a

lattice detection is performed. Therefore, the MIMO detection


problem is converted to finding a lattice point transmitted
over a linear channel distorted by additive white Gaussian
noise. Various optimal and suboptimal solutions have been
presented in the literature for lattice detection (e.g., [4][7]
and references therein). Lattice reduction methods, especially
the LLL algorithm [8], are efficient and powerful tools for
finding the solution to lattice detection problem [7]. The goal
is to find an equivalent channel matrix with almost orthogonal
columns in order to apply a simple decoder such as zeroforcing (ZF) and have near-ML performance (ZF is optimal
for an orthogonal channel matrix). Orthogonality defect factor
is a measure of basis orthogonality with minimum value equal
to one for an orthogonal matrix and it is defined as follows:
(B) ,

||b1 ||2 ||b2 ||2 ||bM ||2


,
det (BH B)

(3)

where bm s are the columns of the basis B. The LLL


algorithm tries to determine a basis with smaller orthogonality
defect factor for the channel lattice. By performing the lattice
reduction, we have
B = HU,
(4)
where B is the reduced channel and U is an M M
unimodular transformation matrix (a square matrix with integer entries and unit determinant). Assuming the scaled and
translated to real channel and using the reduced matrix, one
can equivalently solve the detection problem as
' arg minm ||y
x

Hx||2

= arg minm ||y

HUU

x2Z
x2Z

= U arg min
||y
0
m
x 2Z

(5)
1

x||2

Bx0 ||2 ,

(6)

where x0 = U 1 x. It is shown in [16] that by using LLL,


one can achieve the maximum receive diversity (the number
of receive antennas) with the V-BLAST transmission scheme.
Note that in the LRA detection methods, the lattice reduction is performed during the preprocessing stage in order
to transform the original problem into a simplified one for
the final stage of detection. Assuming slow fading channels,
the preprocessing stage will be shared by many transmitted
symbols. Therefore, when counting the complexity per detected symbol, the preprocessing stage complexity is negligible
compared to that of the tree search or the sphere decoder
which has to be done independently for every symbol (e.g.
[17] and references therein). However, in a time-varying fading
environment, optimizing the preprocessing stage or updating
it according to previous channel realizations is an important
issue. For instance, adaptively updating the LLL is considered
in [15].
B. MIMO Detection with Conditional Optimization
To reduce the complexity of ML decoding for multiplexed
Alamouti space-time codes, an efficient decoder through conditional optimization was applied in [9] where it is shown that
huge saving is possible if the search space is optimized for
the orthogonal structure of the code. In fact, for multiplexed

0090-6778 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2014.2361337, IEEE Transactions on Communications
3

orthogonal designs, the conditional decoding algorithm is


optimal and has low-complexity in decoding the transmitted
symbols. The main idea is to split the optimization parameter
set (transmitted symbols) in (2) and use a low-complexity
decoding method such as zero-forcing for a subset in order to
reduce the computational complexity. This technique enables
the decoder to perform a reduced search conditioned on a lowcomplexity solution for the remaining parameters.
Assume that the transmitted vector x is divided into two
vectors x1 , x2 and the channel matrix H is accordingly divided
into two submatrices H1 , H2 . One can write the channel
model as
y = H1 x1 + H2 x2 + n,
(7)
where H1 is a K P matrix and H2 is a K (M P ) matrix.
If the channel matrix H has P mutually orthogonal columns
in H1 , then the conditional decoder gives us the optimal ML
solution with a complexity of O(N M P ), where N is the
size of the employed QAM constellation. To implement the
conditional decoder, at first, the zero-forcing solution of x1 is
computed by
x
1 (x2 ) =

1 H
(HH
H1 (y
1 H1 )

(8)

H2 x2 ),

(9)

x
1 (x2 ) = Q(
x1 (x2 )),

where Q denotes the QAM symbol quantizer. Using this result,


the solution for x2 is given by searching over all possible
vectors and selecting the one which minimizes the simplified
metric:
x
2 = arg min ||y
x2 2X M

H1 x
1 (x2 )

H2 x2 || .

(10)

Then, x
1 is quantized accordingly:
x
1 = Q(
x1 (
x2 )),

(11)

Note that the conditional detector performs N M P evaluations


of the metric in (10) for all possible symbols in x2 .
III. L ATTICE -R EDUCTION -A IDED C ONDITIONAL
D ETECTOR FOR M ULTI -A NTENNA C HANNELS
The optimal performance of a conditional decoder, for an
arbitrary dimension of the first submatrix, depends mainly
on the orthogonality of the channel submatrix which uses
zero-forcing solution. This orthogonality condition was in fact
translated into multiplexed orthogonal structure in the design
of high rate space-time codes. However, as it is applied for
the Golden code in [11], when there are near orthogonal
submatrices (with defect factors close to one), one can show
that by selecting the more orthogonal submatrix, an almost
ML performance with a significant saving in the complexity
is achievable. Assuming quasi-static channel model, we can
perform some preprocessing steps at the beginning of the block
in order to reduce the overall complexity of detection. The idea
in this work is to design a MIMO conditional detector such that
a near orthogonal submatrix is generated and close to optimal
detection is achieved with low computational complexity.
Considering lattice reduction as an efficient method to provide
a near orthogonal submatrix, we present the LRA conditional
decoder for multi-antenna channels.

Assume that in a system with M transmit antennas, submatrix H1 with P columns is selected according to some orthogonality measures which we discuss later. Moreover, as mentioned in the previous section, before applying lattice reduction
in order to simplify the constellation quantizer, we translate
the channel model such that both in-phase and quadrature
components of symbols belong to ZQ = {0, 1, ..., Q 1}. Let
C M be the M -dimensional hyper-cube with such components.
Using the reduced basis for the channel submatrix, one can
write the channel model in (7) as
y = B1 x01 + H2 x2 + n.

(12)

where B1 = H1 U1 is the LLL-reduced submatrix, U1 is a


P P unimodular matrix and x01 = U1 1 x1 is the transformed
symbol vector. At first, we apply the ZF part of conditional
detection to get the solution for P symbols in x01 as
x
01 (x2 ) = (BH
1 B1 )
x
01 (x2 )

BH
1 (y

H2 x2 ),

d
x01 (x2 )c,

(13)
(14)

where dc denotes the nearest integer function. Next, the


simplified search for the M
P remaining symbols in x2
can be written as
x
2 = arg min||y
x2

2C M

B1 x
01 (x2 )

H2 x2 ||2 .

(15)

Using this solution, the corresponding x


01 is given by
x
01 = d
x01 (
x2 )c.

(16)

Finally, we use the unimodular transformation matrix to find


the original symbols in x1 as
x
1 = U1 x
01 .

(17)

Depending on the dimension of the selected first submatrix,


the lattice reduction-aided conditional decoder provides a
tradeoff in error performance and complexity between the optimal ML detector and the lattice reduction-aided ZF detector.
Note that since we perform the ZF detection on the LLLreduced first submatrix in the proposed scheme, we expect that
the combination with the exhaustive search for the second part
has less noise enhancement compared to the lattice-reductionaided ZF detection of all symbols. The optimal diversity (the
same as the ML decoder) is shown to be achieved with LLLaided linear detector for V-BLAST transmission [16] and with
LLL-aided regularized linear decoder for space-time coded
systems [18]. Considering the uncoded MIMO system, we
show that as a more powerful decoder, the lattice reductionaided conditional decoder achieves the maximum diversity.
This result, for spatial multiplexing scheme and M K, is
formalized in the following theorem:
Theorem 1: The LLL-aided conditional decoder with VBLAST transmission achieves the maximum receive diversity.
Proof: See Appendix for proof.
Assuming space-time coded systems, we can generalize the
above result by employing a regularized version of the LLLaided linear decoder. Consider the equivalent MIMO channel
as in (1) which may correspond to multiple physical channel
uses. By using the minimum mean-square error (MMSE)

0090-6778 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2014.2361337, IEEE Transactions on Communications
4

filtering, one can achieve the same diversity as the ML decoder


[18]. The result for the LLL-aided conditional decoder is
summarized in the following theorem:
Theorem 2: The regularized LLL-aided conditional decoder
for space-time coded systems achieves the same diversity order
as the ML decoder.
Proof: Proof is given in Appendix.
Note that the regularized LLL-aided conditional decoder is in
fact diversity-multiplexing tradeoff (DMT) optimal. However,
we are interested in the diversity order since we are dealing
with fixed rate MIMO transmissions where the bit rates are
independent of SNR.
A. Submatrix Selection
Based on the accuracy of the first solution, one can expect
a performance close to that of the ML decoder. In fact, for
the completely orthogonal submatrix, zero-forcing and ML are
equivalent. In order to compare the reduced subchannel bases,
we employ the orthogonality defect factor which is defined
in (3). The LLL algorithm tries to minimize this factor in an
iterative manner. Basically, a channel basis is said to be better
reduced if the LLL output has a smaller orthogonality defect
factor. Therefore, we use this factor to select a near orthogonal
submatrix at the first part and build the conditional detector
on it.
To perform the selection, we search over the reduced
submatrices and select the one with the smallest orthogonality
defect factor. To be more specific, assume the MIMO channel
model with K M channel matrix H. Let HP denote a
set of submatrices with K rows and P columns and BP the
corresponding LLL-reduced collection (the submatrices may
have common columns). We select the channel submatrix H1
with the LLL-reduced transformation B1 such that
B1 = arg min (B).

(18)

B2BP

Clearly, the matrix H2 consists of the remaining columns. In


the case of full selection, we search over all possible reduced
choices (HP is the collection of all possible submatrices of
size K P ). By applying the full selection, we make sure
that we have the most reliable first solution with conditional
decoding for a given size of the first channel submatrix P .
As mentioned before, the LLL and the subset selection
are performed in the preprocessing stage. To reduce the
complexity of preprocessing stage, small subset selection can
be done (especially for larger MIMO dimensions). Moreover,
depending on the dimension of the first submatrix, P , the
full selection may not be necessary and one may be able
to fill most of the performance gap to the optimal decoder
with a search over a small set of columns. For example, in
the Golden code application [11] for P = 2 with quadratic
complexity in N , one just needs to compare the orthogonality
defect factor for the first two and the last two submatrices
(without even performing the lattice reduction) and achieve an
essentially ML performance. However, for P = 3 (i.e., with
linear complexity), lattice reduction and the full selection is
necessary for the near-optimal performance. In general, if we
choose a larger P to decode more symbols at the first part,

the lattice reduction and an efficient selection is vital for fulldiversity and near-ML performance.
The LRA conditional detection algorithm is summarized as
follows:
(I) Submatrix selection and lattice reduction:
1) HP : a set of K P channel submatrices
BP : the LLL-reduced submatrices of HP
2) B1 = arg min (B)
B2BP

3) H1 = B1 U1 1
4) H2 : the remaining columns of H
(II) Conditional detection: y = B1 x01 + H2 x2 + n
1) x
2 = arg min||y B1 x
01 (x2 ) H2 x2 ||2
2)

x2 2C M P
x
1 = U1 d
x01 (
x2 )c
0
1 H
where x
1 (x2 ) = (BH
B1 (y
1 B1 )
0
0
x
1 (x2 ) = d
x1 (x2 )c

H2 x2 ) and

As a numerical example, the computational complexity of


the proposed detector is presented in Table I in the number of floating point operations (FLOPs). The 8 8 spatial
multiplexing MIMO system with linear complexity detector
(P = M
1) is considered. In the table, the results are
given for preprocessing complexity of various selection set
size (|HP |). We use the full selection with all possible eight
combinations S8, a random submatrix selection set of size
three S3, and no submatrix selection S1 (i.e. the number
of times the LLL is performed equals eight, three and one,
respectively). Assuming six FLOPs per complex multiplication
and two FLOPs per complex addition, we use the average
complexity results for the LLL in [15] and the matrix operations from [19] to calculate the total number of FLOPs. The
major computation steps are summarized in the table. The ML
detection with the exhaustive search and the LRA conditional
detection with zero-forcing are divided into the preprocessing
and the detection stages. To have a fair comparison to the
LRA conditional detection, we assume that the ML detector
calculates and stores all the metrics that do not depend on the
received signal, in the preprocessing stage. Note that in order
to calculate the average complexity per symbol, the complexity
of the preprocessing stage is divided by the blocklength over
which the channel is fixed. The complexity order reduction
of the proposed method can be seen in this example. Note
that although the complexity of ML detection can be reduced
by using a sphere decoder for example, the complexity of
the latter is still exponential in the problem size [20], [21].
The variable complexity of sphere decoder is shown as N M
where the factor 2 (0, 1] is dependent on the SNR [21].
For instance, at SNR = 15dB in the given example of 8 8
MIMO system, the number of FLOPs for the sphere decoder in
[6] are 1.9668e+5, 6.5387e+7 and 4.4183e+9 for the assumed
constellation sizes, respectively.
Remark 1: We can further improve the solution by applying
more powerful but still simple methods such as a better lattice
reduction algorithm (for example, deep insertion LLL [22])
or the decision-feedback-equalizer (DFE) at the first part of
the conditional detection. Additionally, one can use ordering
methods such as V-BLAST greedy ordering [23] in order to

0090-6778 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2014.2361337, IEEE Transactions on Communications
5

TABLE I
C OMPUTATIONAL COMPLEXITY IN TERMS OF FLOP S
Detector

Complexity

ML
(exhaustive)

NM

Example: 8 8 MIMO, P = M 1
64-QAM
256-QAM
1024-QAM

Computation
Preprocessing:
Detection:

Hx
x 2 CM
arg min||y Hx||2

1.3961e+17
2.1955e+16

9.1496e+21
1.4388e+21

5.9963e+26
9.4296e+25

Preprocessing:

LLL(H1 ) H1 2 HP
H 2 x2 x2 2 C M P
1 BH
B1 = (BH
1 B1 )
1

S8: 173,608
S3: 74,083
S1: 34,273

S8: 182,824
S3: 83,299
S1: 43,489

S8: 219,688
S3: 120,163
S1: 80,353

Detection:

B1 (y H2 x2 ) x2 2 C M
arg min ||y B1 x
01 (x2 )

58,496

233,984

935,936

x2C M

LRA
Conditional
Detector

NM

x2 2C M P

H2 x2 ||2

N = |C|: the constellation size, and S8 : |HP | = 8, S3 : |HP | = 3, S1 : |HP | = 1.

maximize the diagonal elements in the upper triangular matrix


and improve the detected point by the DFE.
Remark 2 (DFE conditional detector): As an additional
approach to improve the performance of conditional decoder
when the first submatrix is not exactly orthogonal, one can
lessen the noise amplification effect of ZF solution by employing the DFE and recursively remove the interferences from
previously decoded entries. In other words, using the DFE
solution at the first step of the conditional optimization in (13)
results in a better solution, especially for the case when the
dimension of the near orthogonal submatrix, P is large compared to the dimension of the remaining columns. The DFE
conditional decoding can be performed by replacing the ZF
step in the previous version. Applying the QR decomposition
on B1 we get upper triangular matrix R = [rk,m ] and unitary
matrix Q. The modified receive vector can be written as
y0 = QH (y

H2 x2 ),

(19)

and hence, the new decoder equations for (13) and (14) is
given by
&
%
PP
0
0
y
r
x

k,m
1m
k
m=k+1
x
01k =
, for k = P, P 1, , 1,
rk,k
(20)
where x
01k is the k-th entry of x
01 (x2 ).
Remark 3 (Soft-output detection): In order to achieve close
to capacity performance with low complexity in MIMO systems, a channel code as the outer code is concatenated with
the spatial multiplexing MIMO system [24]. A soft-input, softoutput (SISO) decoder for the outer code exchanges the soft
information in the form of log-likelihood ratio (LLR) values
with the MIMO detector. To compute the LLRs, a list of best
candidates of the transmitted symbols have been generated
around the received point by methods such as list sphere
decoders (LSD) [24], [25]. In [26], for a more stable spherical
list decoding, the ML point is found first and then used as the
center of a modified sphere decoder to gather the lattice points
around it. Using this idea, a suboptimal detected point can be
used as the center of the sphere to form the list. Moreover,
over block fading channels, a reference list can be constructed
at the beginning of each block by finding the lattice points
inside a sphere centered at the origin. This list is then shifted to

the new estimated near-ML point for the subsequent channels


[27]. By employing the same approach, the presented detector
in this work can also be used at the heart of the soft-output
MIMO detection over slow fading channels.
IV. FAST N EAR -ML DECODER FOR SPACE - TIME BLOCK
CODES

In order to see the performance of the LRA conditional


decoder on space-time block coded systems, we apply it as a
fast low-complexity decoder for the Golden code [2], [3] and
the DAST block code [13], [14].
A. The Golden code
This code is a full-rate full-diversity space-time code which
employs two antennas at each transmit and receive side.
Let gkm denote the channel gain between the m-th transmit
antenna and the k-th receive antenna. Assuming a quasistatic channel model and considering the equivalent MIMO
representation as (1), the total channel matrix can be written
as
2
3
g11 i
g21 g11 i
g21
6
7
g21 g11
g21 g11 7
1 6
6
7
(21)
H= p 6
7,
56
g22 g12 i
g22 7
4 g12 i
5

g22

p
1+ 5
2 ,

g12

g22

g12

where =
= 1 = 1 2 5 , = 1 + i and

= 1 + i . By applying the ZF conditional decoder on the


best possible solution on submatrices with two columns and
a reduced search with quadratic complexity in constellation
size, a fast decoder is designed in [11] with essentially ML
performance. To implement a faster decoder and reduce the
decoding complexity from O(N 2 ) to O(N ), we employ the
LRA conditional decoder presented in the previous section.
By applying the LLL on the best available submatrix with
three columns, P = 3, we obtain the almost orthogonal
submatrix B1 and implement the fast decoder which performs
an exhaustive search over just one symbol of QAM. Considering Theorem 2, to ensure full-diversity, one needs to apply
regularized lattice reduction for the low-complexity solution
at the first step of the conditional decoder.

0090-6778 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2014.2361337, IEEE Transactions on Communications
6

B. The DAST block code


Diagonal algebraic space-time (DAST) block codes are constructed by the use of rotated constellations. The DAST block
codes have a rate of one symbol per channel use and achieve
the full diversity over M transmit and K receive antennas. For
M > 2, these space-time block codes outperform the orthogonal designs (i.e. less error probability with the same SNR
and throughput at the cost of increased decoding complexity).
Here, we consider a system with four transmit and one receive
antennas and apply the LRA conditional decoder with linear
complexity. Let gm denote the channel gain between the m-th
transmit antenna and the receive antenna. Following [13] and
[28], the equivalent channel model is given by
H = GM,

(22)

where M is a complex rotation matrix and G is defined by


G , diag(g1 , , g4 ),

(23)

where diag denotes a diagonal matrix. The unitary transformation M can be written as
M = FH diag(1, i1/4 , i2/4 , i3/4 ),

(24)

where F is the 4 4 discrete Fourier transform (DFT) matrix


with entries
1
fkl = p e 2i(l 1)(k 1)/4 , k, l = 1, , 4.
(25)
4
Using QAM constellation and due to the lattice structure of the
DAST block codes, a near-ML detection can be implemented
by the LRA conditional detector. With linear complexity in
size of QAM constellation, the near-ML performance of the
proposed detector is shown in the simulation results.

do the selection and the LLL again and reset the references.
Specifically, we define the update condition as
(H1 (i)U1 (0))

(B1 (0))

(26)

where B1 (0) = H1 (0)U1 (0) is the reference LLL reduction


for the most reduced submatrix and H1 (i) is the new submatrix with the same column set. Moreover,
is selected
for given fading parameters such that the desired tradeoff between the computational complexity resources and the system
performance is met. A larger
gives us less preprocessing
complexity in the expense of degraded performance. Note that
since the condition in (26) for the orthogonality defect factor
of the almost reduced submatrix preserves the fundamental
LLL upper bound [16], we can still show that the adaptive
LLL-aided conditional decoder achieves the optimal diversity
order [15].
VI. S IMULATION R ESULTS
In this section, we compare the performance of various LRA
conditional detection algorithms with the ML solution for two
scenarios. First, we consider the spatial multiplexing MIMO
transmission and apply the proposed detectors with quasistatic channel assumption and also the adaptive scheme over
slowly varying fading channel. Then, we apply the detectors
to two space-time block codes. In addition, we compare the
performance of the proposed conditional detectors with the
ones which are not using the lattice basis reduction. In order
to achieve near-ML performance and fill the gap, we use the
MMSE filtering as part of the preprocessing stage [29] for all
simulations.
A. Spatial Multiplexing MIMO System

V. A DAPTIVE D ETECTION OVER C ORRELATED FADING


C HANNELS
In this section, we consider the scenario where the fading coefficients have temporal correlation and change slowly
through time. In order to reduce the preprocessing complexity
of the LRA conditional detector, we use the previous LLL
results together with previous selection of the best nearorthogonal submatrix and adaptively update the best reduced
submatrix. The idea was applied in [15] for lattice-reductionaided detectors over MIMO channels and it is shown that
significant saving in complexity is achievable for a small
performance degradation. Using the same approach, one can
adaptively track the changes in the orthogonality of the selected submatrices and perform the LLL and the submatrix
selection when it is getting outdated.
By defining a measure of tradeoff between complexity and
performance, we make sure that the orthogonality defect factor
for the best submatrix remains bounded when we use an old
unimodular basis transformation. In other words, we compute
the ratio of the orthogonality defect factor for the near LLLreduced matrix at time i and the one for the reference channel
matrix at time 0. If it is in the defined interval, we keep
the selected submatrix and just compute the reduction by
multiplying the reference transformation matrix; otherwise we

We consider the V-BLAST transmission scheme over


MIMO fading channels and use 4-QAM and 16-QAM constellations to depict the bit error performance of different detection
methods. At first, with four transmit and four receive antennas,
we choose P = 3 and hence, perform the exhaustive search
for the second part over just one symbol of QAM. The submatrix selection is performed over all possible combinations of
submatrices to select the best one in all conditional detectors.
Fig. 1 shows the bit error rate versus Eb /N0 in decibel for
various conditional detectors and the ML detector. Note that
Eb is the average bit energy at each receiver antenna. The
impact of LLL reduction on the best submatrix is depicted in
this figure for the ZF and DFE conditional detectors. The fulldiversity, almost ML performance is achievable with a search
over one QAM symbol after simply solving for three symbols
at the first part.
Next, we consider a larger MIMO system with eight transmit and eight receive antennas. To show the impact of the first
submatrix dimension P on the performance and the tradeoff
with the complexity of submatrix selection size, we apply the
LLL-aided conditional detection for three scenarios:
1) P = 6, with quadratic complexity (in size of the underlying QAM) and three different submatrix selection
sizes. We consider the full selection with all possible 28

0090-6778 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2014.2361337, IEEE Transactions on Communications
7

combinations (N 2 S28) and a random selection set of


size three (N 2 S3). Finally, no submatrix selection is
considered (N 2 S1) where the first six columns form
the first submatrix. Fig. 2 shows the bit error performance
for these detectors.
2) P = 7, (i.e., linear complexity) and three different
submatrix selection size. We apply the full selection
with all possible 8 combinations (N 1 S8), a random
selection set of size three (N 1 S3) and no submatrix
selection (N 1 S1). The result is depicted in Fig. 3.
3) P = 5, 6, 7, with no submatrix selection. We call them
(N 3 S1), (N 2 S1) and (N 1 S1), respectively. The
result is given in Fig. 4.
The ML and the MMSE-LLL-DFE detectors are added for
comparison. The full-diversity (for all values of P ) and close
to ML performance for the conditional detectors can be seen
in these figures. As expected, the performance gap and nonoptimal diversity order for the ones without LLL are can
be seen. Additionally, one observes that the lattice-reduction
aided conditional detector is much less sensitive to small
selection size. Note that with the LRA conditional detection,
a more complex submatrix search and selection provides a
closer performance to the ML detection.
Finally, to investigate the adaptive decoders, we assume
a correlated fading channel with a power spectral density
(PSD) limited by the maximum Doppler frequency. The fading
parameters of the Rayleigh channel for all the simulations are
fd = 100Hz and fs = 270ksps where fd is the maximum
Doppler frequency and fs = 1/Ts is the sampling rate. For a
44 system with 4-QAM, Fig. 5 shows that by employing the
ZF conditional decoder along with the adaptive method, one
can provide a performance/complexity tradeoff by selecting an
appropriate value for . Here, we selected as 1.1 and 1.5. For
the above fading parameters, it was required to update the LLL
and the best submatrix selection, on average, at every 120 and
265 channel realizations, respectively. Comparing the result
with the ZF conditional decoder without LLL which performs
the best submatrix selection at each channel realization, one
can see that using an outdated lattice reduction can still provide
a better solution although it has less average complexity.

technique to improve the performance of the approximate


detection methods at the first part, we studied the LRA detector
for the spatial multiplexing MIMO transmission systems as
well as two space-time block codes over fading channels. We
introduced a low-complexity method and used lattice reduction
to generate the best almost orthogonal channel submatrix.
By employing the orthogonality defect factor as a universal
measure to select the best available submatrix for conditional
detection, we implement an almost optimal detector with
linear complexity in the size of QAM constellation, for the
Golden code, a DAST block code and the four-by-four spatial
multiplexing MIMO system. Additionally, detectors with close
to ML error performance were studied for larger MIMO
systems. We then considered a practical channel assumption
where we transmit symbols over slow varying fading channels.
To save in the complexity of the detector, we exploited the
past lattice reduction and submatrix selection to adaptively
update for the next channel realizations. Note that the proposed
method is a general multi-antenna detection technique and it
can be used for implementing fast and near-optimal detectors
for any space-time block coded linear system.

B. The space-time block codes

Fig. 1. Bit error performance of the ML and the LRA conditional detectors
for a 4 4 MIMO system.

The performance of the proposed decoder for the Golden


code with the LLL on the best available submatrix of three
columns, P = 3, is depicted in Fig. 6 and Fig. 7 for 4QAM and 16-QAM, respectively. It is shown that with linear
complexity in size of the QAM constellation, we obtain nearML performance for the detectors that use lattice reduction.
It can also be seen that without performing LLL, the fulldiversity is not achievable with O(N ). By applying the same
detector, close to ML error performance with various QAM
constellation sizes is depicted in Fig. 8 for the DAST block
coded system.
VII. C ONCLUSION
A fast low-complexity detector was presented for multiantenna systems. By employing the lattice basis reduction

10

10

BER

10

10

10

10

10

ML 4QAM
MMSELLLDFECD 4QAM N1
MMSELLLCD 4QAM N1
MMSECD 4QAM N1
ML 16QAM
MMSELLLDFECD 16QAM N1
MMSELLLCD 16QAM N1
MMSECD 16QAM N1
4

10

12
Eb/N0(dB)

14

16

18

20

22

A PPENDIX
A. Proof of Theorem 1
Assume the conditional decoder with LLL-aided lattice
decoder for the first part and the exhaustive search for the
second part. The decoding error probability can be upper
bounded by
Pe Pe1 + Pe2|c1 ,
(27)

where Pe1 is the error probability of the first part and Pe2|c1
is the error probability of the second part conditioned on
the first correct decision. Consider the system equation in
(12). After performing the LLL on H1 and removing the
interference from all possible x2 symbols, the zero-forcing

0090-6778 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2014.2361337, IEEE Transactions on Communications
8

10

10

10

10

10

BER

BER

10

10

10

10

10

ML
MMSELLLDFECD N2 S28
MMSELLLDFECD N2 S3
MMSELLLDFECD N2 S1
MMSEDFECD N2 S28
MMSEDFECD N2 S3
MMSEDFECD N2 S1
MMSELLLDFE
2

10

10

10

10
Eb/N0(dB)

12

14

16

18

10

20

Fig. 2. Bit error performance of the ML and the LRA conditional detectors
for the first submatrix of size P = 6 and different submatrix selection size
in a 8 8, 4-QAM MIMO system.

ML
MMSELLLDFECD N3 S1
MMSELLLDFECD N2 S1
MMSELLLDFECD N1 S1
MMSEDFECD N3 S1
MMSEDFECD N2 S1
MMSEDFECD N1 S1
MMSELLLDFE
2

10
Eb/N0(dB)

12

14

16

18

20

Fig. 4. Bit error performance of the ML and the LRA conditional detectors
with no submatrix selection (S1) and different first submatrix size in a 8 8,
4-QAM MIMO system.
0

10

10

10

10

10

BER

BER

10
3

10

10
4

10

10

10

ML
MMSELLLDFECD N1 S8
MMSELLLDFECD N1 S3
MMSELLLDFECD N1 S1
MMSEDFECD N1 S8
MMSEDFECD N1 S3
MMSEDFECD N1 S1
MMSELLLDFE
2

10

10
Eb/N0(dB)

12

14

16

18

20

10

12

14

16

Eb/N0 (dB)

Fig. 3. Bit error performance of the ML and the LRA conditional detectors
for the first submatrix of size P = 7 and different submatrix selection size
in a 8 8, 4-QAM MIMO system.

decoder outputs x
1 (x2 ) = U1 x
01 (x2 ) and forms a set of
0
vectors as {(U1 x
1 (x2 ), x2 )}. Let Pe1 be the probability that
the transmitted vector x is not included in this set. Using
Theorem 3 of [16], it can be shown that for this naive lattice
decoder (with finite constellation) with K receive antennas,
Pe1 1

10

ML
MMSELLLCD N1
Adaptive MMSELLLCD N1, =1.1
Adaptive MMSELLLCD N1, =1.5
MMSECD N1

(28)

where 1 is a constant independent of H. At the second


part, the search is performed over all possible x2 symbols to
minimize the new metric as in (15). Assume that Pe2|c1 is the
second decoding error probability given that the first part was
decoded correctly. Since the search in the new closest point

Fig. 5. Bit error performance of the ML and the adaptive LRA conditional
detectors for a 4 4, 4-QAM MIMO system.

decoder is performed over a subset of total possible vectors,


the probability of error can still be upper bounded by the ML
one [30] as
Pe2|c1 2 K ,
(29)
where 2 is another constant number independent of the channel realization. Therefore, the total decoding error probability
is upper bounded by
Pe (1 + 2 )

(30)

and hence, we achieve the full receive diversity as


lim

!1

log Pe
= K.
log

0090-6778 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

(31)

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2014.2361337, IEEE Transactions on Communications
9

10

10

ML 4QAM
MMSELLLDFECD N1 4QAM
MMSEDFECD N1 4QAM
ML 16QAM
MMSELLLDFECD N1 16QAM
MMSEDFECD N1 16QAM
ML 64QAM
MMSELLLDFECD N1 64QAM
MMSEDFECD N1 64QAM

10

10

10

BER

BER

10
3

10

10
4

10

10

10

ML
MMSELLLDFECD N1
MMSELLLCD N1
MMSEDFECD N1
MMSECD N1

10

10
12
Eb/N0 (dB)

14

16

18

20

Fig. 6. Bit error performance of the ML and the LRA conditional detectors
for the Golden code with 4-QAM.

10

11

13

15

17

19

21
23
Eb/N0(dB)

25

27

29

31

33

35

Fig. 8. Bit error performance of the ML and the LRA conditional detectors
for the DAST block code with four transmit and one receive antennas.

10

both parts. Therefore, using (27) we get


Pe (3 + 4 )

10

RK

(32)

where R = min(rank[EE ]), 1 R M , and E is the


matrix of differences between any two distinct codewords
[30]. Thus, the diversity order of the MMSE-LLL conditional
decoder is the same as the optimal ML decoder (which is equal
to M K for full-diversity codes such as the Golden code).
H

BER

10

10

R EFERENCES

10

10

10

ML
MMSELLLDFECDN1
MMSELLLCDN1
MMSEDFECDN1
MMSECDN1
4

10

12

14
16
Eb/N0 (dB)

18

20

22

24

26

28

Fig. 7. Bit error performance of the ML and the LRA conditional detectors
for the Golden code with 16-QAM.

B. Proof of Theorem 2
Consider the conditional detection with efficient MMSELLL lattice decoder (linear decoder or with DFE which are
called Babai lattice decoders [31]) at the first part and the
exhaustive search at the second part. Following the same
approach as the proof of Theorem 1 and using [18], one
can show that this decoder achieve the same diversity as ML
decoder for general space-time coded systems. By removing
the interference from x2 symbols and applying Corollary 2a
from [18] for the first part and then employing the simplified
search for the closest lattice point decoder at the second part,
we obtain the ML upper bound for the probability of error at

[1] G. J. Foschini, Layered space-time architecture for wireless communication in a fading environment when using multi-element antennas,
Bell labs technical journal, vol. 1, no. 2, pp. 4159, 1996.
[2] J.-C. Belfiore, G. Rekaya, and E. Viterbo, The Golden code: a 22
full-rate space-time code with nonvanishing determinants, IEEE Trans.
Inf. Theory, vol. 51, pp. 14321436, Apr. 2005.
[3] P. Dayal and M. K. Varanasi, An optimal two transmit antenna spacetime code and its stacked extensions, IEEE Trans. Inf. Theory, vol. 51,
pp. 43484355, Dec. 2005.
[4] E. Agrell, T. Eriksson, A. Vardy, and K. Zeger, Closest point search in
lattices, IEEE Trans. Inf. Theory, vol. 48, pp. 22012214, July 2002.
[5] E. Viterbo and J. Boutros, A universal lattice code decoder for fading
channels, IEEE Trans. Inf. Theory, vol. 45, pp. 13691642, July 1999.
[6] 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, pp. 23892402, Oct. 2003.
[7] H. Yao and G. Wornel, Lattice-reduction-aided detectors for MIMO
communication systems, in Proc. IEEE Globecom, Taipei, Taiwan,
pp. 424428, Nov. 2002.
[8] A. K. Lenstra, H. W. Lenstra, and L. Lovasz, Factoring polynomials
with rational coefficients, Math. Ann., vol. 261, pp. 515534, 1982.
[9] S. Sirianunpiboon, Y. Wu, A. R. Calderbank, and S. D. Howard, Fast
optimal decoding of multiplexed orthogonal designs by conditional
optimization, IEEE Trans. Inf. Theory, vol. 56, pp. 11061113, Mar.
2010.
[10] A. Hottinen, O. Tirkkonen, and R. Wichman, Multi-antenna transceiver
techniques for 3G and beyond. John Wiley & Sons, 2004.
[11] S. Sirinaunpiboon, A. R. Calderbank, and S. D. Howard, Fast essentially maximum likelihood decoding of the Golden code, IEEE Trans.
Inf. Theory, vol. 57, pp. 35373541, June 2011.
[12] L. P. Natarajan and B. S. Rajan, An adaptive conditional zero-forcing
decoder with full-diversity, least complexity and essentially-ML performance for STBCs, in Proc. 2012 Int. Symp. on Inform. Theory and its
Applications (ISITA), pp. 235239, Oct. 2012.

0090-6778 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCOMM.2014.2361337, IEEE Transactions on Communications
10

[13] M. O. Damen, A. M. Karim, and J.-C. Belfiore, Diagonal algebraic


space-time block codes, IEEE Trans. Inf. Theory, vol. 48, pp. 628
636, Mar. 2002.
[14] M. O. Damen and N. C. Beaulieu, On diagonal algebraic space-time
block codes, IEEE Trans. Commun., vol. 51, pp. 911919, June 2003.
[15] H. Najafi, M. E. D. Jafari, and M. O. Damen, On adaptive lattice
reduction over correlated fading channels, IEEE Trans. Commun.,
vol. 59, pp. 12241227, May 2011.
[16] M. Taherzadeh, A. Mobasher, and A. Khandani, LLL reduction
achieves the receive diversity in MIMO decoding, IEEE Trans. Inf.
Theory, vol. 53, pp. 48014805, Dec. 2007.
[17] A. D. Murugan, H. E. Gamal, M. O. Damen, and G. Caire, A
unified framework for tree search decoding: redisovering the sequential
decoder, IEEE Trans. Inf. Theory, vol. 52, pp. 933953, Mar. 2006.
[18] J. Jalden 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, pp. 47654780, Oct. 2010.
[19] G. H. Golub and C. F. V. Loan, Matrix Computations. The Johns
Hopkins University Press, 1996.
[20] H. Cohen, A course in computational algebraic number theory. Springer,
2000.
[21] J. Jalden and B. Ottersten, On the complexity of sphere decoding
in digital communications, IEEE Trans. Signal Process., vol. 53,
pp. 14741484, Apr. 2005.
[22] C. P. Schnorr and M. Euchner, Lattice basis reduction: Improved practical algorithms and solving subset sum problems, Math. Programming,
vol. 66, pp. 181191, 1994.
[23] J. Benesty, Y. A. Huang, and J. Chen, A fast recursive algorithm for
optimum sequential signal detection in a BLAST system, IEEE Trans.
Signal Process., vol. 51, pp. 17221731, Jul. 2003.
[24] B. Hochwald and S. ten Brink, Achieving near-capacity on a multipleantenna channel, IEEE Trans. Commun., vol. 51, no. 3, pp. 389399,
Mar. 2003.
[25] H. Vikalo, B. Hassibi, and T. Kailath, Iterative decoding for MIMO
channels via modified sphere decoder, IEEE Trans. Wireless Commun.,
vol. 3, pp. 22992311, Nov. 2004.
[26] J. Boutros, N. Gresset, L. Brunel, and M. Fossorier, Soft-input softoutput lattice sphere decoder for linear channels, in Proc. 2003 IEEE
GLOBECOM, CA, USA, pp. 15831587, Dec. 2003.
[27] H. Najafi, M. E. D. Jafari, and M. O. Damen, Adaptive soft-output
detection in MIMO systems, in Proc. 46th Annu. Allerton Conf. on
Commun., Control, and Computing, pp. 13541358, Sep. 2008.
[28] M. O. Damen, H. El Gamal, and N. C. Beaulieu, Systematic construction of full diversity algebraic constellations, IEEE Trans. Inf. Theory,
vol. 49, pp. 33443349, Dec. 2003.
[29] M. O. Damen, H. El Gamal, and G. Caire, MMSE-GDFE lattice
decoding for under-determined linear channels, in Proc. of Conf. on
Inform. Sci. and Syst. (CISS04), Mar. 2004.
[30] V. Tarokh, N. Seshadri, and A. R. Calderbank, Space-time codes for
high data rate wireless communication: Performance criterion and code
construction, IEEE Trans. Inf. Theory, vol. 44, pp. 744765, Mar. 1998.
[31] L. Babai, On lovasz lattice reduction and the nearest lattice point
problem, Combinatorica, vol. 6 no. 1, pp. 113, 1986.

Hossein Najafi received his B.Sc. degree from Amirkabir University of


Technology (Tehran Polytechnic), Tehran, Iran, in 2003, the M.Sc. degree
from University of Tehran, Tehran, Iran, in 2006, and the Ph.D. degree from
University of Waterloo, Ontario, Canada, in 2012, all in Electrical Engineering. He is currently a Postdoctoral Fellow at the Department of Electrical and
Computer Engineering, University of Waterloo. His research interests are in
the areas of wireless communications and multiuser information theory.

Mohamed Oussama Damen received his Ph.D. (in Electronics and Com
munications) from the Ecole
Nationale Superieure des Telecommunications
(ENST) de Paris, France, in October 1999. He has done post-doctoral research
at the ENST, Paris, France, from November 1999 to August 2000, and at
the Electrical and Computer Engineering department of the University of
Minnesota from September 2000 to March 2001. From March 2001 to June
2004, he was with the Electrical and Computer Engineering department of
the University of Alberta, working as a Senior Research Associate of Alberta
Informatics Circle of Research Excellence (ICORE). In June 2004, he joined
the Electrical and Computer Engineering department of the University of
Waterloo, Ontario, where he is now working as an Associate Professor. He
also held a visiting position at Ohio State University in the summer of 2002.
He is a senior member of IEEE. His current research interests are in the
general areas of wireless communications and coding theory with a special
emphasis on coding for multiple-input multiple-output (MIMO) channels and
cooperative diversity.

0090-6778 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Vous aimerez peut-être aussi