IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO.

3, MARCH 2010 1269

Overspread Digital Transmission Over Wireless
Linear Time-Varying MIMO Systems
Pedro M. Marques and Slvio A. Abrantes, Member, IEEE
AbstractAccompanying the journey for higher spectral ef-
ciency in mobile wireless communications, this paper studies
multiple-inputmultiple-output (MIMO) transmission in a unied
framework that consolidates the phenomena responsible for fre-
quency selectivity and time selectivity, i.e., delay overspreading
and Doppler overspreading, respectively. Starting with a baseband
description of the linear time-varying MIMO system, a novel
approach using continuous, discrete, and hybrid linear operators
leverages the process of channel orthonormalization and model
discretization. Notably, this process leads to an optimal time-
varying semiorthonormal matrix matched lter (the ORTHO-TS-
MMF). Monte Carlo simulations test the error performance for
typical wireless MIMO realizations. In particular, they unveil that
combining the ORTHO-TS-MMF with optimal linear detection
takes advantage of both delay and Doppler diversities, signi-
cantly reducing the symbol error probability.
Index TermsDelay- and Doppler-dispersive wireless channels,
linear operator orthogonalization, maximum-likelihood sequence
estimation (MLSE), multiple-inputmultiple-output (MIMO) sys-
tems, noise whitening, nonstationary multivariate decompositions,
spatial and delay-Doppler diversity, time-varying matrix matched
lters (MMFs).
I. INTRODUCTION
T
HE TYPICAL delay spread of the wireless mobile com-
munications channel in common urban or suburban prop-
agation environments is much smaller than its coherence time.
This means that the channel may be viewed as almost time
invariant during consecutive time frames spanning less than
the coherence time. Since the coherence time is on the order
of milliseconds, one approach for designing a digital commu-
nications receiver is to rst consider the channel as interval-
wise invariant, devise an equalizer to the received signal (e.g., a
linear or decision-feedback equalizer), and then use an adaptive
technique (e.g., least mean square, recursive least squares, or
Kalman) to track the channels variations and update the equal-
izer parameters [1][5]. This procedure nds its most important
Manuscript received April 14, 2009; revised August 13, 2009 and
November 7, 2009. First published December 4, 2009; current version pub-
lished March 19, 2010. This work was supported in part by the Ministrio
da Cincia e Ensino Superior, Fundao para a Cincia e Tecnologia, Lisbon,
Portugal, under Grant SFRH/BD/17131/2004. The review of this paper was
coordinated by Dr. C. Ling.
P. M. Marques was with the Department of Electrical and Computer Engi-
neering, Faculty of Engineering, University of Porto, 4200-465 Porto, Portugal.
He is now with EFACEC Sistemas de Electrnica, S.A., 4471-907 Maia,
Portugal (e-mail: pmarq@fe.up.pt).
S. A. Abrantes is with the Department of Electrical and Computer Engi-
neering, Faculty of Engineering, University of Porto, 4200-465 Porto, Portugal
(e-mail: sam@fe.up.pt).
Digital Object Identier 10.1109/TVT.2009.2037640
application in frequency-selective slowly fading channels and,
under mild conditions, may be replaced by blind or decision-
directed techniques [6]. There are, however, several drawbacks
in the adaptive process: 1) It is not a smooth process, as abrupt
changes between consecutive intervals may occur [2]. 2) No
pilot-aided transmission is exploited. 3) It may be unstable un-
der certain response variations [7]. 4) It does not take advantage
of the time variations in the channel to improve the receiver
performance. Introducing the vertical Bell Labs layered space-
time (V-BLAST) [8] architecture into the picture also has its
own problems [9], [10].
It has been shown that the RAKE receiver is inefcient
when the channel is time varying [11], and attempts have been
made to correct the problem using techniques such as time-
frequency (canonical) representations and the basis-expansion
model [12][15]. In [16], maximum-likelihood sequence esti-
mation (MLSE) and maximum a posteriori probability equal-
ization approaches are researched in the context of doubly
selective channels. These developments encourage the study
of wireless channels in a more general setting. Instead of
devising a receiver that matches the current channel conditions
and then using an adaptive algorithm to track the variations
in an interval-by-interval basis, one that matches both channel
and variability conditions across adaptation intervals should be
obtained. The primary barrier for such a study is the analysis of
linear time-varying (LTV) lters in a system-wide perspective,
which is much more involved than analyzing conventional
linear time-invariant systems. This paper will try to shed some
light onto this subject by dwelling on practical aspects of
multiple-inputmultiple-output (MIMO) transmission and re-
ception. No constrictive assumptions will be presumed from
the start, except for that of accurate channel estimation at the
receiver [17][20].
The wireless channel between each pair of antennas will be
considered linear and time varying, with maximum Doppler
shift f
D
and RMS delay spread

. In addition, depending on
the coherence time (t)
c
(2f
D
)
1
, coherence bandwidth
(f)
c
(2

)
1
, signal bandwidth W, and signaling inter-
val T, it can be of two types, i.e., underspread or overspread.
The typical wideband channel (e.g., CDMA2000 or W-CDMA)
is overspread in delay (

1/W W (f)
c
) and,
hence, frequency selective yet slowly fading (T (t)
c
).
On the contrary, to capitalize on capacity, some multicar-
rier systems (e.g., orthogonal frequency-division multiplex-
ing) with very low carrier separations (e.g., not uncommonly
lower than 10 kHz, viz., Third-Generation Partnership Project
(3GPP) long-term evolution (LTE) evolved Multicast Broadcast
0018-9545/$26.00 2010 IEEE
1270 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010
Multimedia Service and 1 kHz in Digital Audio Broadcasting
may be frequency at (W < (f)
c

< T/2) and, hence,

delay underspread; however, their signaling rate may be too low
to permit a slow-fading behavior. These narrowband subchan-
nels are possibly fast fading; hence, Doppler overspread (T >
(t)
c
f
D
> W/2), which means that, since intercarrier in-
terference could be introduced, the induced Doppler spread
might not be negligible from the receiver perspective. Channels
such as those invoked self-dictate the need for a thorough
analysis of transmission overspreading in the context of MIMO
technology.
This paper is organized as follows: Section II introduces the
continuous baseband input/output model of the time-varying
MIMO channel in both time and frequency domains and de-
velops some considerations about the design of the input and
output of a digital MIMO system. The method for discretiz-
ing the baseband input/output model by means of channel-
extracted basis functions is presented in Section III, where an
orthonormalized version of the matrix matched lter (MMF) for
the time-varying MIMO channel is also proposed. Section IV
discusses noise whitening under time-varying conditions, and
Section V derives three MMSE-based linear detectors for the
equivalent discrete input/output model. Section VI resorts to
Monte Carlo simulations to assess the error performance of the
semiorthonormal MMF with linear detection, and Section VII
summarizes this paper.
II. INPUT/OUTPUT MODEL
For a multipath wireless MIMO channel, the LTV im-
pulse response between the jth input and the ith output is
given by [21]
c
ij
(, t) =
N
ij
(t)

k=1

(ij)
k
(t)
_

(ij)
k
(t)
_
(1)
accounting for the attenuations
(ij)
k
(t) and delays
(ij)
k
(t)
of N
ij
(t) multipath replicas. The time-varying parameters of
(1) must be interpreted as sample functions of statistically
independent stochastic processes.
A. Baseband MIMO Channel Model
In a typical wireless application, the input will be limited
to a passband [f
c
B/2, f
c
+B/2] of bandwidth B = 2W
and center frequency f
c
so that one may dene the base-
band response as c
ij
(, t) = 2c
ij
(, t)e
j2f
c

. Denoting the
number of inputs by n
T
, the baseband signal from the jth
input by s
j
(t), and the additive noise at the ith output by
n
i
(t), the baseband signal at the ith output is given by the
superposition
r
i
(t) =
1
2
n
T

j=1
_

c
ij
(, t)s
j
(t )d +n
i
(t). (2)
In a more compact vector-matrix notation
r(t) =
1
2
_

C(, t)s(t ) d +n(t)

=
1
2
C(, t)

s(t) +n(t) (3)

where one denes the column vector-valued signals r(t) =
[r
i
(t)]
n
R
1
, n(t) = [n
i
(t)]
n
R
1
, and s(t) = [s
j
(t)]
n
T
1
, and
the matrix-valued channel response C(, t) = [c
ij
(, t)]
n
R
n
T
(with n
R
being the number of outputs). The notation
x
refers to a MIMO convolution in the x variable. The noise
term n(t) = 2n(t)e
j2f
c
t
is modeled as a zero-mean wide-
sense stationary vector random process with autocorrelation
R
nn
() = N
0
()I and power density spectrum S
nn
(f) =
2S
nn
(f +f
c
) = N
0
I.
Dening the 2-D Fourier transform of C(, t) as C(f, ),
the noiseless baseband model in the frequency domain will be
given by
u(f) =
1
2
C(f , )

s(f) (4)
which, due to the convolution in the variable , explains the
channels Doppler spread.
B. Baseband MIMO Input Design
A vector-valued sequence a
k
at a symbol rate f
s
and
signaling interval T
s
= 1/f
s
usually lies at the input of a
MIMO system. It can be conveyed to the receiver by a vector-
valued signal x(t) with bandwidth f
s
/2 and satisfying
x
s
(t) = x(t)

k
(t kT
s
) =

k
(t kT
s
)a
k
. (5)
Accordingly, a low-pass pulse shaping lter G(f) may be
applied before transmission without the risk of losing any
information required to reconstruct the samples. The MIMO
output will be given by
r(t) =

k
H(t kT
s
, t)a
k
+n(t) (6)
where H(, t) is the full channel response, i.e.,
H(, t) =
1
2
_

C(, t)G( ) d. (7)

C. Baseband MIMO Output Design
Given the discrete input of transmission model (6), it makes
sense to search for a model with a discrete output that lacks no
relevant information for the decision process. For the most part,
ones interest is a discrete input/output model after matched
ltering of the received signal, and therefore, presampling is not
a requirement. Moreover, avoiding MIMO output presampling
will also avoid working at a higher sampling rate than the
MARQUES AND ABRANTES: OVERSPREAD DIGITAL TRANSMISSION OVER WIRELESS LTV MIMO SYSTEM 1271
signaling rate, which will simplify derivations. Starting with the
continuous model in (6), the discrete model will be found by
resorting to 1) the theory of general LTV operators (see [22]
[24] for some insight) and 2) the classical method of MLSE
extended to a time-varying environment.
III. DISCRETIZING THE
MULTIPLE-INPUTMULTIPLE-OUTPUT
INPUT/OUTPUT MODEL
It was found that the discrete-input model of the MIMO
channel is given by
r(t) =u(t) +n(t) = H(, t)

a
l
+n(t)
=

l
H(t lT
s
, t)a
l
+n(t) (8)
where the matrix-valued functions H(t lT
s
, t) are withdrawn
from parallel lines of equation t = +lT
s
(henceforward to be
known as the domains of transmission) in support of the graph
of H(, t). When the channel is time varying and particularly
under fast-fading conditions, a signicantly different H(t
lT
s
, t) is used every lT
s
seconds.
To simplify the treatment of time-varying MIMO lters, it
will be important to circumvent the intricate convolution nota-
tion. (In LTI systems, one simply takes the Fourier transform,
which does not help in the LTV case.) One possibility is to use
operator theory to describe the entire MIMO system. In fact, the
expression for u(t) can be rewritten as
u(t) = H(, t)

x
s
(t) =
_

H(, t)x
s
(t ) d
=
=tt

_
t

H(t, t
/
)x
s
(t
/
) dt
/
(9)
where

H(t, t
/
) = H(t t
/
, t) H(t, t
/
) =

H(t
/
, t
/
t) (10)
and

H denotes a two-sided-continuous time-varying matrix
operator (or a matrix of operators).

H can be interpreted as
a matrix of continuous-time matrices (the coefcients), and
x
s
(t
/
) can be interpreted as a vector of continuous-time vectors
(functions). In fact, standard matrix notation and algebraic
conventions may be used without limitations; one just has to
remind (10) and acknowledge that continuous operator multi-
plication is an innite sum of innitesimal parts (the integral).
The product of operators

Aand

B is dened as

C =

A

B

=

C(t, t
/
) =
_
t

A(t, t
//
)

B(t
//
, t
/
) dt
//
(11)
and the Hermitian adjoint product is given by

C =

A
H

B

=

C(t, t
/
) =
_
t

A
H
(t
//
, t)

B(t
//
, t
/
) dt
//
. (12)
The continuous input/output model is thus rewritten as
r = u +n =

Hx
s
+n. (13)
Moreover, one can also introduce a one-sided-discrete (hy-
brid) matrix operator and reassemble (8) as
r(t) =

l
H(t lT
s
, t)a
l
+n(t)

H(t, lT
s
)a
l
+n(t)

= r =

H a +n (14)
where the dot above the operator indicates that it is discrete on
one of its sides (implicit from the construction) and continuous
on the other, and the dot above the vector indicates that it acts
as a vector of discrete-time vectors. Similarly, two upper dots
will denote a two-sided-discrete matrix operator.
A. Dening a Basis for MIMO Response Expansion
The foremost objective is to nd an approximation u(t) for
u(t) as a function of generic orthonormal matrix-valued basis
functions (t lT
s
, t) and vector-valued coefcients b
l
u=

b

= u(t)=

l
(t lT
s
, t)b
l
=(, t)

b
l
(15)
such that the following semiunitarity condition is satised:

=

I

=
_
t

H
(t kT
s
, t)(t lT
s
, t) dt =
kl
I.
(16)
A nite-energy requirement on u(t) is set as
|u|
2
= u
H
u

= |u(t)|
2
=
_
t
u
H
(t)u(t) dt < . (17)
By minimizing the energy in the difference between u and
u, the standard minimum squared error criterion determines the
optimal vector

b as follows:

b
opt
= arg min

b
|u u|
2
= arg min

b
|u

b|
2
. (18)
Let u
o
be the orthogonal projection of u onto the range of

,
(

), and write

b
opt
= arg min

b
|u u
o
+ u
o

b|
2
[u u
o
]

(

) = ^(

H
); [ u
o

b] (

) (19)
where ^(

H
) is the null space of the Hermitian adjoint

H
,
and denotes the orthogonal complement. Hence, applying the
1272 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010
Pythagorean theorem, it follows that

b
opt
= arg min

b
| u
o

b|
2
(20)
which leads to the conclusion that setting u =

b = u
o
min-
imizes the energy in the vector-valued error function. Letting
u = u
o
+ [u u
o
], it is clear that
[u u
o
] ^(

H
)

H
[u u
o
] = 0

=
_
t

H
(t lT
s
, t) [u(t) u
o
(t)] dt = 0 (21)
and consequently

b
opt
=

H
u

b
opt
=

H
u = u

=b
(opt)
l
=
_
t

H
(t lT
s
, t)u(t) dt. (22)
As expected, the optimal coefcients are obtained by orthog-
onally projecting u(t) onto the basis functions (t lT
s
, t)
or, equivalently, onto (

). The optimal coefcients are least-
square solutions of

b = u.
The additional sufcient restriction

H
=

I

=

l
(t lT
s
, t)
H
(t
/
lT
s
, t
/
) = (t
/
t)I
(23)
renders

as a full-orthonormal (i.e., full-unitary) matrix oper-
ator and guarantees an approximation almost everywhere (a.e.)
between u(t) and u(t)
|u u|
2

=
_
t
(u(t) u(t))
H
(u(t) u(t)) dt = 0
u(t) = u(t) a.e. (24)
but it will nevertheless be shown in Section III-D that (23) is
unnecessary for optimal detection.
The optimal coefcients

b
opt
can be rewritten in terms of the
original coefcients a (the sequence of transmitted symbols) as
follows:

b
opt
=

H
u =

H a =

H

=b
(opt)
k
=

H

k,l
a
l
=

l
H
l,k
a
kl
(25)
where

H

H denotes the projection of

H onto

or
their correlation

H

k,l
=
_
t

H
(t, kT
s
)

H(t, lT
s
) dt. (26)
In turn, using (23),

H can be expressed in terms of

and

as

H =

=

H(t, lT
s
) =

(t, kT
s
)

H

k,l
. (27)
The rst equality in (27) resembles the familiar QR fac-
torization from linear algebra with matrices replaced by linear
operators, except that (to be causal)

H

is necessarily a lower
triangular operator, i.e.,

H

k,l
= 0 k < l. To determine

,
it shall be assumed (with no loss of generality) that the rst
and last symbols to be transmitted are a
1
and a
L
, respectively,
such that

H(t, lT
s
) =
L

kl

(t, kT
s
)

H

k,l
. (28)
The concept behind the factorization is a backward block-
wise GramSchmidt semiorthonormalization procedure. After
a
L
is received and

H(t, LT
s
) is known, the Lth normal basis
matrix is computed as

H(t, LT
s
) =

(t, LT
s
)

H

L,L
,

H

L,L
=
_
_
_

H(t, LT
s
)
_
_
_

(t, LT
s
) =

H(t, LT
s
)
_
_
_

H(t, LT
s
)
_
_
_
1
(29)
where |

H(t, LT
s
)| = [

H
H

H]
1/2
LL
is given by
_
_
_

H(t, LT
s
)
_
_
_ =
_
_
_
t

H
H
(t, LT
s
)

H(t, LT
s
) dt
_
_
1/2
. (30)
The subsequent basis matrices

(t, (L m)T
s
), 1 m <
L are found by projecting onto the previous space, computing
the difference, and normalizing the outcome, i.e.,

(t, (L m)T
s
) =
_
_
H(t, (L m)T
s
)
L

i=Lm+1

(t, iT
s
)

_
t

H
(t, iT
s
)

H(t, (L m)T
s
) dt
_
_
H
1

Lm,Lm
(31)
where

H

Lm,Lm
=
_
_
_
_
_
_

H(t, (L m)T
s
)
L

i=Lm+1

(t, iT
s
)

_
t

H
(t, iT
s
)

H(t, (L m)T
s
) dt
_
_
_
_
_
_
(32)
and the remaining (below diagonal) entries are obtained from

H

i,Lm
=
_
t

H
(t, iT
s
)

H(t, (L m)T
s
) dt (33)
for L m < i L.
This process builds an orthonormal basis for the received
signal and guarantees a causal

H

. The main drawback for

achieving causality is that the orthonormalization runs back-
ward from

H(t, LT
s
), which is something that may not be
feasible if L is too large. An alternative approach is to keep L
MARQUES AND ABRANTES: OVERSPREAD DIGITAL TRANSMISSION OVER WIRELESS LTV MIMO SYSTEM 1273
Fig. 1. Discrete MIMO model after projecting the output onto R(

).
sufciently low and perform the orthonormalization for every
L vector symbols.
Fromthe channel decomposition, the MIMOnoiseless output
is given by
u =

H a =

a =

b
opt
= u
o
(34)
indicating that the system

b = u is, in fact, consistent. (

has
u(t) in its range.) Expressing the input/output model as
r =u +n =

b
opt
+n

=r(t) =

l
(t LT
s
, t)b
(opt)
l
+n(t) (35)
and applying the adjoint operator on the left yield the discrete
model
r

H
r=

H
u+

H
n=

b
opt
+

H
n

=r
k
=
_
t

H
(tkT
s
, t)r(t)dt =b
(opt)
k
+n
k
. (36)
One should notice that applying the adjoint operator is
equivalent to ltering with
H
(, t ) and then sampling
at t = kT
s
because
r
k
=
_

H
(, kT
s
)r(kT
s
) d
=
H
(, t )

r(t)
_
t=kT
s
. (37)
As shown in Fig. 1, after cascading the channel with

H
,
the discrete output is determined from the discrete input by
applying the two-sided-discrete operator

H

(or a discrete
matrix lter H
l,k
).
The operator

does not span the entire noise space of the
MIMO output, because the condition

H
=

I is not met. As
a result, some information is lost after the projection of r(t)
onto (

). It will now be determined whether this information
is relevant from an optimal estimation perspective.
B. Time-Varying MLSE (TV-MLSE)
Let

be an ideal full-orthonormal matrix operator that satis-
es both conditions

=

I and

H
=

I. As previously
discussed, this operator spans the entire space of the MIMO
noisy output r(t). Since its estimate is r =

r

, where r

H
r, then it is equal to r(t) at least a.e., that is, r =

H
r
a.e.
=
r. Developing this, an equivalent discrete representation that
losslessly describes r(t) is found by orthogonally projecting
r(t) onto the range of

. That is
r

H
r = u

+ n

= r
k
= u
k
+n
k
(38)
where, generically, x
k
=
_
t

H
(t kT
s
, t)x(t) dt. The dis-
cretized noise is space-time white

R
n

=
1
2
E
_
n

n
H

_
= N
0

=R
n
k
n
l
(k l) = N
0

kl
I (39)
and normally distributed
n
k


^
_
E[n
k
], R
n
k
n
k
_

^(0, N
0
I)
f
n

( n

) =
_

k
(2N
0
)
n
R
_
e
(2N
0
)
1
n
H

. (40)
The probability density function in (40) lies at the root of
the maximum-likelihood (ML) detection criterion of the input
vector-valued sequence of symbols. If the symbols of the input
set are equally likely to be transmitted, then the classical Bayes
theorem establishes that
a
ML
= arg max
a
f( r

[ u

) = arg min
a
| r

|
2
. (41)
C. TS-MMF
Expanding the norm in (41) and using r

H
(r
u), one arrives at
a
ML
= arg min
a
_
2Re a
H
y +|

H a|
2
_
(42)
1274 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010
Fig. 2. Discrete MIMO model after matched ltering with the TS-MMF.
where y =

H
H
r is the output information required for ML
decision. Again
y =

H
H
r

= y(kT
s
) = H
H
(, t )

r(t)
_
t=kT
s
(43)
yielding H
H
(, t ) as the MMF for the time-varying
MIMO channel. This is the optimal predetection lter. It is
noncausal and nonpredictive and, in effect, translates to the
adjoint operator

H
H
(t
/
, t) in the operator domain; therefore, the
optimal receive lter operator is always the Hermitian adjoint
operator of the channel operator. H
H
(, t ) is merely
required to satisfy H
H
(, kT
s
) =

H
H
(kT
s
, kT
s
) so
that it can be matched to the domains of transmission. In
fact, the receiver is correlating with each matrix-valued func-
tion H(t lT
s
, t) as
y(kT
s
) =

l
_
_
_
t
H
H
(t kT
s
, t)H(t lT
s
, t)dt
_
_
a
l
+
_
t
H
H
(t kT
s
, t)n(t) dt (44)
or, more simply, as
y(kT
s
) =

l
S((k l)T
s
, kT
s
) a
l
+n
Hk
(45)
i.e., y =

S a + n

H
, where

S =

H

H
=

H
H

H is the global
system operator, and

S(kT
s
, lT
s
) =
_
t

H
H
(t, kT
s
)

H(t, lT
s
) dt. (46)
Contrary to the LTI scenario,

S(kT
s
, lT
s
) is nonstationary, be-
cause it depends on both the difference k l and each particular
k; hence, it is not Toeplitz. The same is true for the noise
autocorrelation, i.e.,

R
n

H
n

H
=N
0

S

= R
n
Hk
n
Hl
(k l, k)
=N
0
S((k l)T
s
, kT
s
) . (47)
One may also check that |u|
2
= a
H

S a 0, i.e.,
_
t
u
H
(t)u(t) dt =

k
a
H
k

S(kT
s
, lT
s
)a
l
0 (48)
meaning that

S is nonnegative denite Hermitian and, hence, a
valid covariance operator. In addition, since r

H
is a sufcient
statistic, the ML criterion is given by
a
ML
= arg min
a
( r

H

S a)
H

S
1
( r

H

S a). (49)
The presence of the middle factor

S
1
is caused by the lack
of semiorthonormality of the channel operator

H; therefore,
removing the rst is an extra motivation for orthogonalizing the
TS-MMF.
Writing MMF(, t) = H
H
(, t ), it is clear that there
is an implicit linear transformation
(, t) (, t )
_

t
_
=
_
1 0
1 1
__

/
t
/
_
(50)
where (
/
, t
/
) are the old coordinates. This reveals that the
matched lter has the form of a /4-skewed mirroring of
the full channel response H(, t). In other words, H(, t)
experiences a vertical (time) shear of unit slope, followed by
a horizontal (delay) reection. In light of this, the matched
lter will be distinguished with the name time-shear MMF (TS-
MMF). The system diagram after ltering with the TS-MMF is
shown in Fig. 2.
The TS-MMF is always implementable, except in a very
atypical scenario, where (t)
c
< T
s
<
max
, because H(, t)
will effectively change before it can be measured. In all other
situations, the TS-MMF may be determined by channel estima-
tion or sending pilots. Whenever the channel is fast fading such
that
max
< (t)
c
< T
s
, a candidate for lling estimation gaps
is interpolation [25], [26].
D. Optimality of the Discrete MIMO Model in (

)
At this point, one no longer needs the operator

, which
is introduced in Section III-B. The next step is to return to
the semiorthonormal channel operator

from Section III-A
and try to express the ML criterion within (

), checking
for its equivalence. The projection of the noise n(t) onto the
space (

) is denoted by n

H
n so that the MIMO
MARQUES AND ABRANTES: OVERSPREAD DIGITAL TRANSMISSION OVER WIRELESS LTV MIMO SYSTEM 1275
Fig. 3. Decomposition of the TS-MMF and global system response into discrete and semiorthonormal lters.
input/output relation after the channel decomposition

H =

is equal to r =

H a +n =

(

H

a + n

) + , where

= (t) is the noise part out of (

). In addition
=n

n

= r

r

H
=0 ^(

H
) =

(

). (51)
From these considerations, the ML criterion may be re-
written as
a
ML
= arg min
a
_
_
_(r

r

) +

( r

a)
_
_
_
2
= arg min
a
| r

a|
2
(52)
because ( r

H

a
)
H

H
= 0, and |r

r

| is indepen-
dent of a. The ML estimates a
ML
are entirely based on the
discretized MIMO model of Fig. 1, and one should notice
the exclusive use of the semiorthonormality condition

I. All the information required for optimal detection is still

available after projecting the MIMO output by means of the
operator

H
extracted from the full channel response H(, t);
hence, one can condently rely on the time-varying causal
discrete model
r

=

H

a + n

= r
k
=
k

l=kD

H

k,l
a
l
+n
k
. (53)
Not only is the model attractively discrete and causal, but the
noise is also space-time white:

R
n

= N
0

I. Capitalizing
on this, (52) has the form
a
ML
= arg min
a

k
_
_
_
_
_
r
k

l=kD

H

k,l
a
l
_
_
_
_
_
2
(54)
from which one may easily derive a detector based on the
Viterbi algorithm. This would not be possible without a semi-
orthonormal MMF.
E. Decomposing the TS-MMF Into Discrete and
Semiorthonormal Factors
In addition to decomposing the channel operator as in (27), it
is also relevant to seek an equivalent expression that factorizes
the TS-MMF H
H
(, t ). With operator conventions

H
H
=

H
H

H
=

H
H
(t, lT
s
) =

H
H

k,l

H
(t, kT
s
).
(55)
Manipulating, (37) and (43) yield
y(kT
s
) =

H
H

l,k
r

l
= H
H
l,kl

l
r
k
=H
H
l,rl

H
(, t )

r(t)
_
t=kT
s
(56)
revealing a convolution between a discrete MMF H
H
l,rl
and
a continuous semiorthonormalized MMF
H
(, t ) (from
this point onward to be designated ORTHO-TS-MMF).
Moreover, it is straightforward to show that the global system
response can also be decomposed into discrete and orthogonal
terms. From (27) and (46), one nds

S =

H
H

=

S(kT
s
, lT
s
) =

H
H

r,k

H

r,l
(57)
which, after some manipulations, yields
S((k l)T
s
, kT
s
) =

r
H
H
rk,r
H
rl,r
lkl
S(lT
s
, kT
s
) =H
H
r,kr

r
H
l,k
(58)
which is an expression that resembles the continuous correla-
tion function S(, t) = H
H
(, t )

H(, t). The role

of the discrete MMF H
H
r,kr
is to match the receiver to
H
l,k
. The system diagram pertaining to these decompositions
is shown in Fig. 3.
1276 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010
IV. NOISE WHITENING
The uncorrelated noise samples at the ORTHO-TS-MMF
output represent a clear advantage, as they induce a reduction
in detection complexity. In the following, the possibility of
orthonormalizing the TS-MMF using a fully discrete time-
varying matrix lter will be addressed.
Since the noise autocorrelation at the output of the TS-MMF
is given by

R
n

H
n

H
= N
0

S, one may question the existence of

a lter

that, when applied at the TS-MMF output, renders the
noise white, i.e.,

R
n

H
n

H
= N
0

H
H

H
= N
0

I. (59)
If such a fully discrete lter indeed exists, it should satisfy the
condition
(

H

H
)
H

H

H
= (

H

H
)
H

H

H
=

=

I
where

H
=

H
H

H
. It is known, nonetheless, that the
solution of the minimum Frobenius norm of the system

H
H
=

H
H

H
is obtained from the MoorePenrose pseudoinverse of

, i.e.,

H

; therefore, one may test the choice of a lter that

satises

H
=

H
H
=

H
H

H
H

H
H
=

H
H

H
H

H
. (60)
To make things clear,

H

is expanded using its singular

value decomposition (SVD)

H

=

U

V
H
, where

U
H

U =

V
H

V =

I, and (

U) = (

H

); (

V) = (

H
H

). The
pseudoinverse is decomposed as

H

=

V

D
1

U
H
, and (60)
is rewritten in the form

H
=

U

D
H

V
H

D
H

U
H

H
=

U

U
H

H
. (61)
Consequently, a sufcient condition for whitening is

U

U
H
=

I, which is satised if

Uis square and nonsingular; hence,

H

should be a full row-rank operator. The full row-rank formula

=

H
H

H
H

_
1

=

I
shows that the pseudoinverse is a right inverse of

H

. In other
words,

=

H
H

whitens the noise at the TS-MMF output, i.e.,

R
n

=

R
n

= N
0

I.
It should be kept in mind that

H

is the band-based sparse

L factor resulting from a block-wise QL operator factorization.
Its dimensions are n
T
N
L
n
T
L, where L is the number of
transmitted vector-valued symbols, and N
L
is the number of in-
dependent matrix-valued functions H(t lT
s
, t). N
L
is equal
to L with high probability. From the statistical independence of
a set of N
L
domain functions H(t lT
s
, t), one assesses the
statistical independence of the matrix coefcients within each
column of

H

, and hence, whitening can be deemed possible

with high probability.
V. LINEAR DETECTION FOR OVERSPREAD
MULTIPLE-INPUTMULTIPLE-OUTPUT CHANNEL
Optimal detection of the sequence of vector samples a is
accomplished by employing the Viterbi algorithm extended to
a MIMO setup, as it yields an ML estimate of a. As usual,
the penalty incurred by optimal detection is computational
complexity, and in the MIMO case, it may become overwhelm-
ing. When high detection complexity becomes an issue, time-
varying linear ltering is an alternative method that warrants
linear complexity scaling, and therefore, it may be used instead.
This section presents the optimal linear detector for the over-
spread MIMO channel: the unconstrained linear detector. The
idea is to extend standard linear estimation (see [28][31]) to
time-varying channel operators.
A. Unconstrained Linear Detection
The optimal linear detector can be derived by processing
an entire sequence of vector samples r
0
r
L
(i.e., r

).
This means that all the information available in

H

will be
used for detection, improving the error performance. The linear
estimator of a is

a =

D r

, and the error induced by the

estimator is the sumof two independent entities, i.e., the symbol
distortion and the residual noise, i.e.,
e = a

a = a

D

a
. .
a
distortion

D n

. .
n
residual
. (62)
The error is minimized when it is orthogonal to the matched
lters output, i.e.,
E
_
e r
H

_
= 0

D
opt
= arg min

D
E
_
e
H
e

. (63)
Expanding the error leads to E[ a r
H

] =

D
opt
E[ r

r
H

],
which means that two statistical expectations need to be
determined: 1) E[ a r
H

] = 2

R
a a

H
H

, and 2) E[ r

r
H

] =
2

H

R
a a

H
H

+ 2

R
n

, which, since

R
n

= N
0

I,
simplies to

R
r

=

H

R
a a

H
H

+N
0

I. Equating, one
obtains

R
a a

H
H

=

D
opt
_

R
a a

H
H

+N
0

I
_
which, after applying the generalized inverse (or Gaussian
elimination), yields the general expression

D
opt
=

R
a a

H
H

R
a a

H
H

+N
0

I
_

. (64)
For independent zero-mean symbols, (64) simplies to

R
a a
= (P
a
/2)

I

D
opt
=

H
H

H
H

+
1

I
_

(65)
MARQUES AND ABRANTES: OVERSPREAD DIGITAL TRANSMISSION OVER WIRELESS LTV MIMO SYSTEM 1277
where, again,
t
= P
a
/
2
n

. The minimum average error en-

ergy is given by
E[ e
H
e]
min
=2tr
_
(

I

D
opt

H

R
a a
_
=2tr
_
N
0

D
opt

H
H

_
(66)
where

H
H

is the MoorePenrose pseudoinverse of

H
H

.
VI. ERROR PERFORMANCE OF OPTIMAL TIME-VARYING
SEMIORTHONORMAL MATRIX MATCHED FILTER
WITH LINEAR DETECTION
This section will study the error performance of the ORTHO-
TS-MMF when combined with the unconstrained linear detec-
tor. The goal is to infer how well the receiver compensates (and
possibly benets from) the delay-Doppler overspreading of the
MIMO channel. To accomplish this, one shall engage in Monte-
Carlo-based numerical simulation for several wireless MIMO
channel realizations.
A. Continuous MIMO Channel Model
The continuous passband channel model was given in (1). Its
baseband version is
c
ij
(, t) = 2
N
ij
(t)

n=1

(ij)
n
(t)e
j
(ij)
n
(t)

_

(ij)
n
(t)
_
(67)
where the multipath delays and associated phases are
(ij)
n
(t) =

(ij)
n
(
(ij)
n
/
c
)t, and
(ij)
n
(t) = 2f
c

(ij)
n
(t) =
(ij)
n
+

(ij)
n
t, respectively. The carrier frequency is
c
, and the an-
gular Doppler shift is given by
(ij)
n
= 2(/) cos
(ij)
n
=
2f
D
cos
(ij)
n
, where is the relative velocity between the
transmitter and receiver, is the wavelength,
(ij)
n
are the
different angles of arrival relative to the direction of motion,
and f
D
is the maximum Doppler shift.
Since the dominant cause of mobile channel variability is the
multipath phase variation due to the Doppler effect, the fol-
lowing approximations were considered for model simulation:
1) f
c
f
D

(ij)
n
(t) =
(ij)
n
,
(ij)
n
(t) =
(ij)
n
, and N
ij
(t) =
N
ij
, for the small-scale motion of the receiver unit. 2) Each
multipath component arrives at different receiver antennas
by parallel paths, i.e.,
(ij)
n
=
(1j)
n
; moreover,
(ij)
n
=
(1j)
n
.
3) N
ij
= N. 4) Multipath components from different transmit-
ter antennas are independent. 5) The receiver moves in parallel
to the axis of the receiver array, as shown in Fig. 4; thus

(ij)
n
= 2f
c
_

(1j)
n
(d/c)(i 1) cos
(1j)
n
_
(68)
where = 2/, d is the antenna separation, and c is the speed
of light. The antennas will be either equispacedly t within
d
max
= 0.5 or freely scale, i.e., d
max
= dn
R
= 0.5n
R
.
The multipath delays follow the exponential model:
(1j)
n
=

ln(1 u
(1j)
n
), where f

(
(1j)
n
) = (1/

)e

(1j)
n
/

, and
u
(1j)
n
|(0, 1). The other delays are given by
(ij)
n
=
(1j)
n

Fig. 4. Different multipath phases of a moving array with collinear axis.
(d/c)(i 1) cos
(1j)
n
. Equivalently, the attenuations are ex-
ponentially decaying, i.e.,
(1j)
n
= e

(1j)
n
/2

, and the angles

of arrival
(1j)
n
|(0, 2) simulate the isotropic scattering
model. The RMS delay spread will be the rule of thumb

= 3 s, and the number of multipath components will be

N = 8. Since it is a natural candidate for fourth-generation
LTE licensing, the frequency of operation f
c
= 2.6 GHz will
be chosen for the simulation, and the maximum Doppler shift
will be f
D
200 Hz.
The pulse-shaping lter is the one-third rolloff square-root-
raised-cosine lter (SRRC) with impulse response
srrc() = (4roll
_
f
s
)
_

_
1 (4rollf
s
)
2
_
1

_
cos ((1 + roll)f
s
) +
sin ((1 roll)f
s
)
(4rollf
s
)
_
(69)
delayed an integer of symbols and truncated on both sides
g() =
_
srrc( /f
s
), 0 2/f
s
0, otherwise
(70)
which yields the full pairwise channel response
h
ij
(, t) =
N

n=1

(1j)
n
e
j
_

(ij)
n
+
(1j)
n
t
_
g
_

(ij)
n
_
. (71)
B. Discrete MIMO Channel Model for Simulation
To avoid aliasing, the noiseless MIMO output u(t) will be
sampled at twice the signaling rate, i.e.,
u(kT
s
/2) =

H(kT
s
/2, lT
s
)a
l

=
2
u =
2

H a (72)
where the operator
2

H has two different discrete rates on each
of its sides. If the noise is conned to a bandwidth B
N
/2 = f
s
,
then the global noisy discrete channel model is given by
r(kT
s
/2) =

H(kT
s
/2, lT
s
)a
l
+n(kT
s
/2)

=
2
r =
2

H a +
2
n. (73)
The entries of

H(kT
s
/2, lT
s
),

h
ij
(kT
s
/2, lT
s
) are 1 1
operators that dene the response between the jth input and the
1278 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010
Fig. 5. Average 16-QAM error rates for two receiver array sizes, increasing
number of transmitter antennas, and receiver antennas equispacedly t within
d
max
= 0.5 (T
s
= 100 s).
ith output of the MIMO channel. They can be related to (71) as
follows:

h
ij
(kT
s
/2, lT
s
)
= h
ij
(kT
s
/2 lT
s
, kT
s
/2)
=
N

n=1

(1j)
n
e
j
_

(ij)
n
+
(1j)
n
kT
s
/2
_
g
_
kT
s
/2 lT
s

(ij)
n
_
.
(74)
The discrete ORTHO-TS-MMF
2

may be computed from
2

H using a discretized version of the semiorthonormalization
procedure described in Section III-A. Unfortunately, imple-
mentation showed that it was rather slow to complete and
had poor numerical stability. To address this issue, a novel
method based on backward block-wise Householder reections
was conceived, proving to be remarkably fast and stable. It is
presented in Appendix A.
C. Monte Carlo Simulation and Numerical Results
Simulation starts by randomly generating and modulating a
message of 64 complex symbols from a 16-state quadrature
amplitude modulation (16-QAM) constellation. The random
message is applied to
2

H, and noise is added, simulating the
MIMO transmission. The output is then ltered by
2

H
, and
the unconstrained linear detector derived in Section V-A. A
slicer makes the nal decisions on the transmitted constellation
symbols. To yield reliable error rates, this process is repeated,
so that at least 1 10
5
symbol transmissions are simulated.
Several plots were obtained for randomly chosen channel re-
alizations, and each plot assesses the error rate as a function of
the SNR (stepped at 2 dB) at the output of the MIMO channel.
The plots in Fig. 5 reveal that increasing the number of trans-
mitter antennas has the effect of increasing the error probability,
which is legitimate, because the symbols transmitted from
different transmitter antennas will interfere with one another
Fig. 6. Average 16-QAM error rates for one transmitter antenna and several
receiver array sizes (receiver antennas equispacedly t within d
max
= 0.5
and free scaling considered) for T
s
= 100 s.
Fig. 7. Average 16-QAM error rates for several signaling rates, a MIMO
conguration of n
T
= 2 and n
R
= 4, and receiver antennas equispacedly t
within d
max
= 0.5.
and escalate confusion at the receiver. The penalty for higher
data rates (using the same bandwidth and thus improving the
spectral efciency) is an increase in the probability of symbol
error. Nevertheless, by decreasing the signaling interval (at the
expense of a larger bandwidth), much better error performances
are possible.
A different situation appears in Fig. 6: a single transmitter
antenna and an increasing number of receiver antennas. One
readily gathers that increasing the size of the receiver array
has the consequence of decreasing the error probability, which
means that the receiver is able to use reception diversity to im-
prove the error performance. This improvement is even larger
when free scaling of the receiver array is permitted, which is
natural, given the reduction in cross correlation.
To assess the variation of the error performance with the
signaling rate, plots for several T
s
were obtained as shown in
Fig. 7. The shortest signaling interval is responsible for the
MARQUES AND ABRANTES: OVERSPREAD DIGITAL TRANSMISSION OVER WIRELESS LTV MIMO SYSTEM 1279
Fig. 8. Average 16-QAM error rates for several maximum Doppler shifts,
a MIMO conguration of n
T
= 2 and several n
R
, and receiver antennas
equispacedly t within d
max
= 0.5 (T
s
= 1 s).
lowest error rates. As the signaling interval increases below

= 3 s, the error rate also steadily increases, which is

explained by an added difculty of resolving multipath com-
ponents. The error performance severely degrades when T
s
becomes somewhat higher than

, as demonstrated by the
error rate curve of T
s
= 0.1 ms. In this situation, there is little
multipath diversity available for resolution, and the receiver
cannot compensate for the intersymbol interference (ISI) intro-
duced. However, as the signaling rates get lower and lower, the
ISI will become negligible, and the error rates will signicantly
be reduced (despite the eminent fast-fading situation).
The plots in Fig. 8 conrm that, when the channel is fast
fading (which is the case for T
s
= 1 s), the receiver is able
to use the extra variability introduced by the Doppler shift to
improve the detection, irrespective of the number of receiver
antennas. One realizes that a cardinal advantage of the ORTHO-
TS-MMF is its channel variability awareness, which exploits
the available Doppler diversity.
Plots for the error rate variation with the maximum antenna
separation are shown in Fig. 9. They display an error per-
formance deterioration as the maximum separation decreases.
Bringing the antennas increasingly closer to one another has the
effect of increasing correlation to the extent of not permitting
the compensation of both intrasample ISI and time selectivity.
The plots corroborate ones intuition.
VII. CONCLUSION
This paper has proposed a new framework for the study
of digital communications over delay-Doppler dispersive LTV
MIMO systems. Using a baseband setup, it started by dening
a continuous input/output model of digital transmission that
did not ignore the time-varying nature of the wireless channel.
Then, it aimed at model discretization using a channel-extracted
orthogonal basis. This was accomplished by orchestrating a
theory of continuous, discrete, and hybrid matrix operators that
impressively simplied all MIMO calculations. An orthonor-
Fig. 9. Average 16-QAM error rates for several maximum receiver antenna
separations, a MIMO conguration of n
T
= 2 and n
R
= 4, and receiver
antennas equispacedly t within d
max
(T
s
= 1 s).
malized MMF, which maintained the white nature of the input
noise and was optimal from an ML detection perspective, has
been derived.
The optimal linear detector for the new model has been
derived and tested with success using Monte Carlo simulations.
They have purveyed considerable insight into the error per-
formance of the time-varying (i.e., Doppler aware) semiortho-
normal MIMO matched lter, conrming that time-frequency
selectivity can be put to ones favor as it is the result of
delay-Doppler diversity. They have also shown that the error
performance is boosted when the number of receiver antennas
increases and that the penalty for achieving higher spectral
efciencies with more transmitter antennas is an increase in
error probability.
APPENDIX
BUILDING OPTIMAL TIME-VARYING SEMIORTHONORMAL
MATRIX MATCHED FILTER WITH BACKWARD
BLOCK-WISE HOUSEHOLDER REFLECTIONS
The goal is to perform the decomposition
2

H =
2

=

H(kT
s
/2, lT
s
) =

(kT
s
/2, mT
s
)

H

m,l
(75)
where

H

is block lower triangular with block size n

T
n
T
,
and
2

is block orthogonal, with each block consisting of n
T
columns.
The idea is to sequentially eliminate the entries above each
block in the lower right diagonal of
2

H, where the lower right
diagonal is interpreted as the diagonal that starts with the utmost
lower right block of
2

H. The entries above the utmost lower
right block can be eliminated by nding the QR factorization of
the rightmost n
T
columns of
2

H, i.e.,

H(:, LT
s
) = Q, where R
is n
T
n
T
lower triangular. Next, the utmost lower right block
of Qis retrieved, and P is computed as follows:
B = Q(end, :) P = B(B
H
B)
1/2
. (76)
1280 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 3, MARCH 2010
Dene v = QeP(Q
H
Q)
1/2
= QeP, where e is the
same size of Q and is all zeros, except on the lower right
diagonal, where it is all ones. The block v can be used to build
the rst reector
H
r
= I 2v(v
H
v)
1
v
H
. (77)
Moreover, from (76), it is known that P
H
P = I and P
H
B =
B
H
P = (B
H
B)
1/2
. Expanding the products v
H
v and v
H
Q
reveals v
H
v = 2I 2(B
H
B)
1/2
= 2v
H
Q, which means that
H
r
Q = Q2v(v
H
v)
1
v
H
Q = Qv = eP. (78)
The entries above the utmost lower right block of Q have been
eliminated.
Now, since the goal is to perform the elimination in
2

H,
(78) is multiplied by R, which yields the same elimination
H
r
QR = H
r

H(:, LT
s
) = ePR. H
r
is applied to the entire
operator
2

H
2

H
set
= H
r2

H =
2

H2v(v
H
v)
1
v
H
2

H (79)
and then, the bottom block row is normalized with
E = (R
H
R)
1/2
R
H
P
H

H(end, :)
set
= E

H(end, :) (80)
which ends the rst step of the decomposition. Further steps
are required to eliminate the rest of the entries above the
block diagonal, except that they are applied to
2

H, excluding
the block columns (and corresponding block rows) that have
already been eliminated. The process ends when the rst block
column of
2

H undergoes elimination, and at this point, the
lower Ln
T
Ln
T
entries of
2

H are, in fact,

H

in its block
lower triangular form. To avoid the inherent loss of numer-
ical information in computing the product B
H
B, the SVD
B = USV
H
is performed, and P = B(B
H
B)
1/2
= UV
H
is computed, which replaces (76).
The preceding description explains how to obtain the opera-
tor

H

for simulation, but the ORTHO-TS-MMF

2

H
has yet
to be determined. One obvious option is to apply the procedure
just described to the identity matrix, instead of
2

H, which, after
normalization with (80), yields
2

H
. What this approach has
in obviousness, it also has in cost. It is slow. Experimentation
revealed that the swiftest approach was to perform Gaussian
elimination on the system

H
H

H
=
2

H
H
.
REFERENCES
[1] C. Komninakis, C. Fragouli, A. H. Sayed, and R. D. Wesel, Multi-
input multi-output fading channel tracking and equalization using Kalman
estimation, IEEE Trans. Signal Process., vol. 50, no. 5, pp. 10651076,
May 2002.
[2] K.-J. Lin and Y. Lee, Adaptive minimum variance detection for spatial
multiplexing over time-varying MIMO channels, in Proc. IEEE Veh.
Technol. Conf., May 1114, 2008, pp. 13601364.
[3] S. Kun and Z. Xudong, Equalization of time-varying MIMO systems
using modied Kalman lter algorithm, in Proc. 7th Int. Conf. Signal
Process., Aug. 31Sep. 4, 2004, vol. 2, pp. 16701673.
[4] G. T. Zhou, Y. Kim, and G. B. Giannakis, Estimation and equalization of
time-selective channels using precoding, in Proc. 33rd Asilomar Conf.
Signals, Syst., Comput., Pacic Grove, CA, Oct. 2427, 1999, vol. 1,
pp. 248252.
[5] C. Komninakis, C. Fragouli, A. H. Sayed, and R. D. Wesel, Channel
estimation and equalization in fading, in Proc. 33rd Asilomar Conf.
Signals, Syst. Comput., Pacic Grove, CA, Oct. 2427, 1999, vol. 2,
pp. 11591163.
[6] E. Karami and M. Shiva, Blind multi-inputmulti-output channel
tracking using decision-directed maximum-likelihood estimation, IEEE
Trans. Veh. Technol., vol. 56, no. 3, pp. 14471454, May 2007.
[7] S. He, X. Shan, and Y. Ren, A new adaptive equalization scheme for
MIMO channels, in Proc. 3rd Int. Symp. Image Signal Process. Anal.,
Sep. 1820, 2003, vol. 2, pp. 10281033.
[8] P. W. Wolniansky, G. Foschini, G. Golden, and R. A. Valenzuela,
V-BLAST: An architecture for realizing very high data rates over
the rich-scattering wireless channel, in Proc. URSI Int. Symp. Signals,
Syst., Electron., Sep. 29Oct. 2, 1998, pp. 295300.
[9] J. Choi and H. Y. Lee, Adaptive MIMO decision feedback equalization
for receivers in time-varying channels, in Proc. 57th IEEE Semiannual
Veh. Technol. Conf., Apr. 2225, 2003, vol. 3, pp. 18511856.
[10] D. K. C. So and R. S. Cheng, Layered maximum likelihood
detection for MIMO systems in frequency selective fading
channels, IEEE Trans. Wireless Commun., vol. 5, no. 4, pp. 752762,
Apr. 2006.
[11] A. M. Sayeed and B. Aazhang, Joint multipath-Doppler diversity in
mobile wireless communications, IEEE Trans. Commun., vol. 47, no. 1,
pp. 123132, Jan. 1999.
[12] S. Bhashyam, A. M. Sayeed, and B. Aazhang, Time-selective signaling
and reception for communication over multipath fading channels, IEEE
Trans. Commun., vol. 48, no. 1, pp. 8394, Jan. 2000.
[13] X. Ma and G. B. Giannakis, Maximum-diversity transmissions over dou-
bly selective wireless channels, IEEE Trans. Inf. Theory, vol. 49, no. 7,
pp. 18321840, Jul. 2003.
[14] R. Boudreau, J.-Y. Chouinard, and A. Yongacoglu, Exploiting Doppler-
diversity in at, fast fading channels, in Proc. Can. Conf. Elect. Comput.
Eng., 2000, vol. 1, pp. 270274.
[15] B. D. Hart and D. P. Taylor, Extended MLSE diversity receiver for the
time- and frequency-selective channel, IEEE Trans. Commun., vol. 45,
no. 3, pp. 322333, Mar. 1997.
[16] I. Barhumi and M. Moonen, MLSE and MAP equalization for transmis-
sion over doubly selective channels, IEEE Trans. Veh. Technol., vol. 58,
no. 8, pp. 41204128, Oct. 2009.
[17] H. Zhu, B. Farhang-Boroujeny, and C. Schlegel, Pilot embedding
for joint channel estimation and data detection in MIMO commu-
nication systems, IEEE Commun. Lett., vol. 7, no. 1, pp. 3032,
Jan. 2003.
[18] J. Gao and H. Liu, Decision-directed estimation of MIMO time-varying
Rayleigh fading channels, IEEE Trans. Wireless Commun., vol. 4, no. 4,
pp. 14121417, Jul. 2005.
[19] L. Tong, B. M. Sadler, and M. Dong, Pilot-assisted wireless
transmissions, IEEE Signal Process. Mag., vol. 21, no. 6, pp. 1225,
Nov. 2004.
[20] D. Samardzija and N. Mandayam, Pilot assisted estimation of MIMO
fading channel response and achievable data rates, IEEE Trans. Signal
Process., vol. 51, no. 11, pp. 28822890, Nov. 2003.
[21] J. G. Proakis, Digital Communications., 4th ed. New York: McGraw-
Hill, 2001.
[22] A. W. Naylor and G. R. Sell, Linear Operator Theory in Engineering and
Science. New York: Springer-Verlag, 2000.
[23] N. I. Akhiezer and I. M. Glazman, Theory of Linear Operators in Hilbert
Space. New York: Dover, 1993.
[24] K. Jrgens, Linear integral operators, in Surveys and Reference Works
in Mathematics, vol. 7. Boston, MA: Pitman, 1982.
[25] H.-N. Lee and G. J. Pottie, Fast adaptive equalization/diversity combin-
ing for time-varying dispersive channels, IEEE Trans. Commun., vol. 46,
no. 9, pp. 11461162, Sep. 1998.
[26] W. H. Chin, D. B. Ward, and A. G. Constantinides, Algorithm for
exploiting channel time selectivity in pilot-aided MIMO systems, IET
Commun., vol. 1, no. 6, pp. 12671273, Dec. 2007.
[27] S. Barbarossa and A. Scaglione, On the capacity of linear time-varying
channels, in Proc. IEEE Int. Conf. Acoust., Speech Signal Process.,
Mar. 1519, 1999, vol. 5, pp. 26272630.
[28] T. Kailath, A. H. Sayed, and B. Hassibi, Linear Estimation. Upper
Saddle River, NJ: Prentice-Hall, 2000.
[29] H. L. Van Trees, Detection, Estimation, and Modulation Theory.
New York: Wiley, 2001, pt. I.
[30] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Upper Saddle River, NJ: Prentice-Hall, 1993.
[31] S. M. Kay, Fundamentals of Statistical Signal Processing: Detection
Theory. Upper Saddle River, NJ: Prentice-Hall, 1998.
MARQUES AND ABRANTES: OVERSPREAD DIGITAL TRANSMISSION OVER WIRELESS LTV MIMO SYSTEM 1281
Pedro M. Marques received the Licenciatura and
Ph.D. degrees in electrical and computer engineering
fromthe University of Porto, Porto, Portugal, in 2003
and 2009, respectively.
In 2003 and for an entire semester, he was a
Research Collaborator with Instituto de Engenharia
de Sistemas e Computadores Porto, working in
novel areas of next-generation digital communica-
tion networks. In particular, the conception and im-
plementation of a state-of-the-art monitoring system
for Mobile-IPv6-based wireless access networks re-
sulted from that collaboration. He is currently with EFACEC Sistemas de
Electrnica, S.A., Maia, Portugal, which is the largest Portuguese company in
the eld of electronics. His current research interests include spectrally efcient
wireless communications, multiple-inputmultiple-output systems, orthogonal
frequency-division multiplexing, cooperative communication networks, and
radio-over-free space optics. He is also a teacher at Universidade Lusfona do
Porto, Porto, Portugal.
Slvio A. Abrantes (M90) received the Licenciatura
degree in electrical engineering and the Ph.D. degree
in electrical engineering and computers from the
University of Porto, Porto, Portugal, in 1976 and
1990, respectively.
He is currently an Assistant Professor with the
Department of Electrical and Computer Engineering,
Faculty of Engineering, University of Porto, and a
Senior Researcher with INESC Porto, which is an
R&D institute afliated with the same university.
In 2004, he was a Visiting Professor with the In-
formation and Telecommunication Technology Center, University of Kansas,
Lawrence. He is the author of three books in Portuguese: Adaptive Signal
Processing (Lisbon: Fundao Gulbenkian, 2000), Source Coding: Two Short
Visits (Porto: FEUP Edioes, 2000), and Error-Correcting Codes for Digital
Communications (Porto: FEUP Edies, to be published). His past research
interests include high-speed digital subscriber loops, simulation of communica-
tion systems, adaptive signal processing, information theory, and error control
coding. His current research interests are iterative decoding with turbo and low-
density parity check codes.
Prof. Abrantes is a member of the Ordem dos Engenheiros (Portugal).