You are on page 1of 12

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO.

6, JULY 2009 2803


Optimizing Pilot Locations Using
Feedback in OFDM Systems
Ali Y. Panah, Student Member, IEEE, Rodney G. Vaughan, Fellow, IEEE, and
Robert W. Heath, Jr., Senior Member, IEEE
AbstractPilot-aided orthogonal frequency-division multiplex-
ing (OFDM) channel estimation has, so far, been optimized only
for open-loop systems for which the uniform spacing of pilots is
optimal. In this paper, we examine closed-loop OFDM systems
and show how uniform pilots may no longer be optimal. In doing
so, we propose a feedback technique in which the pilot allocation
mechanism is adapted to the mobile channel. After deriving a
general expression for the symbol error rate as a function of
pilot allocation, we optimize the pilot locations and propose two
practical solutions using bounds on the objective function. We
proceed to formulate pilot allocation based on the maximal av-
erage OFDM channel capacity. We examine the feedback over-
head that is associated with our method and make comparisons
with the conventional uniform framework. Finally, we suggest
the use of vector quantization in the context of the generalized
Lloyd algorithm to reduce feedback and extend our method to a
multiple-inputmultiple-output (MIMO)-OFDM system based on
the Alamouti spacetime block code. Monte-Carlo simulations il-
lustrate the error-rate/capacity performance gains via nonuniform
pilots and illustrate the effects of Doppler frequency and carrier
frequency offset.
Index TermsFeedback communications, least squares es-
timation, orthogonal frequency-division multiplexing (OFDM),
pilot-symbol-aided modulation (PSAM).
NOMENCLATURE
:= Equation by denition.
()

Conjugate.
()
T
Transpose of a vector or matrix.
()
H
Conjugate-transpose.
()
1
Square matrix inverse.
()

MoorePenrose (pseudo) inverse.


z (((,
2
) Complex Gaussian distributed random vari-
able with mean and variance
2
.
tr and E Trace and expectation operators.
|(|) Ceiling (oor) operation.
a
mn
(m, n)th element of the matrix A.
I
N
N N identity matrix.
Manuscript received February 29, 2008; revised August 16, 2008. First
published October 31, 2008; current version published May 29, 2009. The
work of A. Y. Panah and R. W. Heath, Jr. was supported in part by Samsung
Electronics. The review of this paper was coordinated by Dr. H.-C. Wu.
A. Y. Panah and R. W. Heath, Jr. are with the Wireless Networking and
Communications Group, Department of Electrical and Computer Engineering,
University of Texas at Austin, Austin, TX 78712 USA (e-mail: ayp@mail.
utexas.edu; rheath@ece.utexas.edu).
R. G. Vaughan is with the School of Engineering Science (ENSC), Simon
Fraser University, Burnaby, BC V5A 1S6, Canada (e-mail: rvaughan@sfu.ca).
Color versions of one or more of the gures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TVT.2008.2008655
F N N Fourier matrix with f
mn
=
e
j2mn/N
for 0 m, n N 1.
rank() Rank of a matrix.
diagx Diagonal matrix formed from the elements of
the vector x.
I. INTRODUCTION
T
HEORETICAL research on orthogonal frequency-
division multiplexing (OFDM) as an efcient broadband
wireless modulation scheme has come a long way since its in-
troduction in the late 1960s, with the pivotal work of Gibby [1],
Weinstein [2], and Bingham [3]. At present, the desire for
accurate coherent reception and the attraction of packet-
switched mobile links has encouraged research toward more
efcient and accurate yet simple and plausible channel-
estimation processes for OFDM systems [4].
One widely employed means of acquiring channel-state in-
formation (CSI) at the receiver is to multiplex known pilot sym-
bols into the transmission data streama technique referred
to as pilot-symbol-aided modulation (PSAM). From a system
design point of view, PSAM-OFDM can be very exible,
yielding an array of techniques with various performances, con-
siderations, implementations, and assumptions. The estimate-
then-interpolate approach is, perhaps, the most common of
these approaches. Many classes of this technique, such as
linear interpolation, second-order interpolation, timedomain
interpolation, etc., have been simulated [5][14], analyzed
in terms of error-rate performances [15][19], and demon-
strated in more recent ad hoc solutions such as [20] and [21].
Optimization of PSAM-OFDM can be found in [22][25];
however, the Wiener lter, as a form of optimal PSAM, was
rst derived and analyzed in [26]. Computational complexity
has hindered the widespread applicability of this approach,
even in modern OFDM systems. Suboptimal PSAM for OFDM
has subsequently been extensively investigated in the literature.
More recently, particular attention has been paid to optimizing
and analyzing the effects of the location of the pilot symbols
within the frequency band, the total number of pilot symbols
needed in each packet, and the power dedicated to the pilot sym-
bols relative to the data symbols. It has been shown in [27] that
the channel mean square error (MSE) is minimized for PSAM-
OFDM if and only if the pilots are equispaced in frequency.
This result has been substantiated by many other authors,
including further quantication of the MSE for maximum-
likelihood (ML) estimators (see also [28]), as well as for the
Bayesian minimum MSE (MMSE) estimator approach [29].
0018-9545/$25.00 2009 IEEE
2804 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 6, JULY 2009
Although theoretically attractive and, at times, mathemati-
cally tractable, channel MSE minimization does not necessarily
lead to bit-error-rate (BER) minimization, which is often a
more desirable objective in packet-driven OFDM links [30].
Many authors have tackled the problem of directly minimizing
the BER of PSAM-OFDM, concluding that equispaced pilots
are, in fact, not only MSE-minimizing [30] but also error-
minimizing pilots [31]. Moreover, when MMSE is utilized,
such pilots also maximize the average capacity over Rayleigh
fading channels [32].
The subject of dedicated feedback utilization in improving
various performance measures of wireless links has been an
area of intensive research for more than a decade (see, e.g.,
[33][35] and references therein). Given the optimality of
uniformly spaced pilots, the topic of nonuniform pilots for
PSAM has received scarce scientic attention, at best (see, e.g.,
[36]). To the best of our knowledge, no previous work has
attempted to incorporate feedback in improving PSAM over
an OFDM link. This paper is based on the observation that
conventional equispaced pilots are only optimal in the sense
that the transmitter has no a priori knowledge regarding the
Rayleigh fading channel through which it is transmitting. Using
CSI, however, we propose that the degrees of freedom afforded
by PSAM (e.g., the pilot locations) be gainfully exploited at the
transmitter. Such an approach was partially investigated in our
previous work, where pilot symbol locations were obtained to
maximize the average signal-to-noise ratio (SNR) per OFDM
symbol (see [4] and results therein).
In this paper, we directly tackle the optimization of the SER
and the average OFDM channel capacity. We show how, by
utilizing a feedback link, an OFDM transmitter may adopt a
more intelligent pilot allocation strategy to improve the symbol
error rate (SER) and the average end-to-end capacity of the
link. Our feedback strategy is unique, because the optimization
process is conducted at the receiver, and the feedback need not
include the entire CSI. The proposed system only requires feed-
back of indexes that represent the current best pilot allocation.
We also extend our solution to more general congurations.
For example, to illustrate exibility, we extend from single-
antenna OFDM to MIMO-OFDM in the context of Alamouti
spacetime block-coded (STBC) systems. Our simulations will
also show considerable improvements for these systems. In
addition, we propose a clustering algorithm based on vector
quantization to reduce feedback overhead. Our simulations here
will show that, when properly designed, codebooks of relatively
short length (e.g., 4 bits) exhibit marginal loss compared with
the optimum solution. We also compare our results with those
in [4] to see the improvements that were gained fromdirect SER
optimization as opposed to SNR optimization.
A. Contributions of This Paper
The contributions of this paper are threefold: 1) the math-
ematical formulation of nonuniformly spaced pilots and opti-
mization in terms of the average SER and the average capacity;
2) the formulation of an extension to MIMO-OFDM using
Alamouti-STBC; and 3) a clustering algorithm for reducing
feedback overhead.
The rest of this paper is organized as follows. An overview
of the OFDM multipath fading channel and signal models is
presented in Section II. Parameterized least squares channel
estimation is formulated in Section III, followed by an analysis
of pilot location optimization in Section IV. Feedback reduction
is considered in Section V, and a simple extension to MIMO-
OFDM is presented in Section VI. Numerical results are pre-
sented in Section VII, and concluding remarks are presented in
Section VIII.
II. SYSTEM MODEL
A baseband single-user OFDM system model is illustrated
in Fig. 1. In this section, we review the mechanics and
assumptions of our system model. We begin with a descriptive
overview of the wireless channel model, which will also serve
to introduce our notations.
A. Channel Model
We assume a frequency selective channel, which is modeled
as a tapped delay line. The channel is also quasistatic, which
means that the channel remains constant for each OFDM
symbol and varies in time according to the Doppler frequency
and the Jakes model (also known as Clarks model) [37].
Let h() denote the channel impulse (time-delay) response
that encompasses the baseband effects of both pulse shap-
ing and the multipath physical channel between the trans-
mitter and the receiver. The corresponding physical situation
is the baseband channel of an omnidirectional antenna that
moves within a uniform planar omnidirectional dense scattering
environment.
Sampling the channel at the symbol rate T
s
, assuming that
there is no timing error, leads to the tapped delay line model
with channel taps given by h(lT
s
). Normalizing T
s
to unity
and assuming that the channel taps are signicant only up
to the Lth sample value, the sample-spaced channel may be
represented by an (L + 1)-length vector of sample, which
is denoted by h = [h(0), h(1), . . . , h(L)]
T
. This vector will
constitute the channel impulse response (CIR) throughout this
paper. With sufcient scattering present in the physical medium
and a suitably normalized power delay prole, each channel
tap may be modeled as an independent identically distributed
Gaussian random variable such that h(l) (((0,
2
h(l)
) and

L
l=0

2
h(l)
= 1, l. Note that, for the simulations, we use
an exponential power delay prole, but the results basically
depend only on the delay spread rather than on the details of
the prole.
Assuming a total of N orthogonal modulating fre-
quencies (subcarriers), the frequency response on the nth
subcarrier is H(n) =

L
l=0
h(l)e
j(2/N)ln
, 0 n N
1, which, when collected in an N-length vector H =
[H(0), H(1), . . . , H(N 1)]
T
, constitutes the channel fre-
quency response (CFR). The CFR may directly be derived from
the CIR vector as
H = F
L
h (1)
PANAH et al.: OPTIMIZING PILOT LOCATIONS USING FEEDBACK IN OFDM SYSTEMS 2805
Fig. 1. OFDM baseband system model.
where F
L
is a matrix that consists of the rst L + 1 columns of
the N N Fourier matrix F.
Given the zero-mean CIR, the elements of the CFR are
correlated with a covariance matrix cov(H) = EHH
H
=
F
L
cov(h)F
H
L
, where cov(h) = Ehh
H
. Hence, the CFR
may be modeled as a complex Gaussian random vector H
(((0
N
, cov(H)), where 0
N
is an N-length vector of zeros.
Moreover, the channel taps are assumed to be uncorrelated;
thus, cov(h) is an (L + 1) (L + 1) diagonal matrix, with the
ith diagonal element being equal to
2
h(i)
for 0 i L. Then,
it is easy to verify that the channel variance on each subcarrier is

2
H(n)
=
L

l=0

2
h(l)
= 1. (2)
B. Signal Model
At any signaling interval, the OFDM transmitter in Fig. 1
maps a buffered block of the input binary stream into N
d
baseband symbols, each with a normalized time duration of
T
s
= 1 unit. The symbols are drawn from the unit energy
multiple phase-shift keying (MPSK) constellation space /.
Next, the pilot allocation module inserts a total of N
p
com-
mon pilot symbols, simply denoted by s
p
/, at prede-
termined locations between the data symbols to construct a
length N = N
p
+N
d
OFDM symbol, which is denoted by S =
[S(0), S(1), . . . , S(N 1)]
T
. The pilot symbols are known a
priori to the transmitter and receiver in amplitude, phase, and
their location within each OFDM symbol. We assume equal
power pilot and data symbols, which leads to a power constraint
of T =

N1
n=0
E[S(n)[
2
= N for each OFDM symbol. A
time-domain symbol is constructed next by applying an inverse
discrete (fast) Fourier transform (IFFT) to the OFDM symbol.
Finally, to eliminate intersymbol interference (ISI) caused by
the L-path channel and, hence, the need for an equalizer at
the receiver, a guard interval (or cyclic prex) of length N
g
is
appended to this symbol prior to transmission. The demodulator
at the receiver ignores the rst N
g
symbols (ISI portion) of
the received signal and gathers the next N samples. Taking
the N-point discrete (fast) Fourier transform (FFT) of these
samples yields, what we call here, the N-length received vector
R. Assuming perfect timing and carrier synchronization, a
sufciently large N so that each subcarrier experiences at-
fading, and a cyclic prex length that exceeds the channel order,
i.e., N
g
L, the received signal on the nth subcarrier can be
written as
R(n) = S(n)H(n) +V (n) (3)
where H(n) is the nth element of the CFR, V (n) is an additive
white Gaussian noise (AWGN) with V (n) (((0, N
0
), and
N
0
is the AWGN noise variance per subcarrier. Using matrix
notations and (1), we have
R =S
D
H+V
=S
D
F
L
h +V (4)
where S
d
= diagS is the N N diagonal matrix of
transmitted symbols, including pilot symbols. R = [R(0),
R(1), . . . , R(N 1)]
T
is the received vector, and h =
[h(0), h(1), . . . , h(L)]
T
and H = [H(0), H(1), . . . , H(N
1)]
T
are the CIR and CFR vectors, respectively. The AWGN
vector is V (((0
N
, N
0
I
N
) and is assumed independent of
all other variables.
III. LEAST SQUARES CHANNEL ESTIMATION
The pilot locations in each OFDM symbol may conveniently
be represented by an N N diagonal pilot allocation matrix X
such that
x
ii
=
_
1 if ith subcarrier is pilot
0 if ith subcarrier is data
x
ij
= 0 i ,= j.
(5)
2806 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 6, JULY 2009
Denote the set of all possible pilot allocation matrices as
1
p
= X
1
, X
2
, . . . , X
[1
p
[
, where [1
p
[ =
_
N
N
p
_
is the cardinal-
ity of the set. In addition to the aforementioned denition, each
matrix is also constrained by the total number of pilots in each
OFDM symbol N
p
, which may be translated to the following
constraint:
trX = N
p
X 1
p
. (6)
Using this matrix, any received vector R may be decou-
pled into pilot and data-bearing subcarriers as R = R
(pilot)
+
R
(data)
, where R
(pilot)
= XR. Substituting from (4)
R
(pilot)
=S
D
XF
L
h +XV
=s
p
XF
L
h +XV. (7)
Premultiplying both sides of (7) by the common pilot symbol
s

p
, we arrive at an observation vector

R
(pilot)
= s

p
R
(pilot)
,
where

R
(pilot)
=[s
p
[
2
XF
L
h +X(s

p
V)
=XF
L
h +X

V (8)
and the nonzero elements of

Vare statistically equivalent to V.
From

R
(pilot)
, a least squares estimate (LSE) [38] of the CIR
may be obtained via the MoorePenrose matrix inverse as [39]

h =(XF
L
)

R
(pilot)
=
_
F
H
L
XF
L
_
1
F
H
L
X

R
(pilot)
. (9)
Note that rank(X) = N
p
and rank(F
L
) = L + 1; hence,
rank(F
H
L
XF
L
) = L + 1, and the inversion in (9) is possible
only if N
p
> L, which is intuitive, because, to estimate the CIR,
we require as many pilots as the channel length needs. Using
(1), (9), and the invariance property of the LSE [38], the CFR
estimate is

H = F
L

h = F
L
_
F
H
L
XF
L
_
1
F
H
L
X

R
(pilot)
. (10)
Substituting for

R
(pilot)
from (8) and simplifying using the fact
that XX = X, we have

H =H+F
L
_
F
H
L
XF
L
_
1
F
H
L
X
. .
J

V
=H+E (11)
where we have dened the N N projection matrix as
J = F
L
(F
H
L
XF
L
)
1
F
H
L
X, with rank(J) = L + 1. The corre-
sponding N-length error vector is E = J

V. Note how the error


vector E becomes a function of the pilot allocation matrix X
through the projection matrix J. On a per-subcarrier basis, (11)
becomes

H(n) = H(n) +e(n) (12)


where e(n) is the nth element of E for 0 n N 1.
The error vector in (11) is zero mean and independent of
H; thus, the covariance of the channel estimate

H can be
decomposed as
cov(

H) =cov(H) + cov(E)
=cov(H) +N
0
F
L
_
F
H
L
XF
L
_
1
F
H
L
:=cov(H) +N
0
(13)
where := F
L
(F
H
L
XF
L
)
1
F
H
L
is an N N matrix.
Equation (13) indicates that the variance of the channel
estimate for the nth subcarrier can be written as

2
H(n)
=
2
H(n)
+N
0

n
= 1 +N
0

n
(14)
where
n
denotes the nth diagonal element of .
Note that each element of the error vector e(n), although
independent of H(n), is correlated with the corresponding
element of the channel estimate

H(n). As in [31], to decompose
the error, let e
|
(n) denote the MMSE estimate of e(n) in (12),
which minimizes E[e(n) e
|
(n)[
2
, where e
|
(n) = wH(n),
and dene e

(n) := e(n) e
|
(n) as the residual. Then, the
optimum value of w satises the orthogonality condition that
Ee

(n)

H(n)

= 0, and hence, the MMSE estimate of e(n)


is obtained as
e
|
(n)
MMSE
:=
E
_
e(n)

(n)
_
E
_

H(n)

(n)
_

H(n) =
N
0

n
1 +N
0

H(n)
(15)
where the second equality comes, because e(n) and H(n) are
independent of each other. Thus, substituting (15) into (11) and
rearranging, we get
H(n) =
1
1 +N
0

H(n) e

(n) (16)
where Ee

(n) = 0, and E[e

(n)[
2
= N
0

n
/(1 +N
0

n
).
IV. OPTIMUM PILOT ALLOCATION
With the formulated LS channel estimation, we proceed
to derive the instantaneous SNR as a function of the pilot
allocation. We will then maximize the average SER and, later,
the OFDM system capacity in terms of the pilot allocation.
A. Optimizing for SER
The matched lter (MF) at the receiver uses the received
samples of Rand the estimated channel gains of

Hto construct
the decision variables on each subcarrier as
y(n)
MF
:=

H

(n)R(n)
=
S(n)

H(n)

2
1 +N
0

n
S(n)

(n)e

(n) +

H

(n)V (n) (17)


where we have substituted for H(n) from (16) into (3).
PANAH et al.: OPTIMIZING PILOT LOCATIONS USING FEEDBACK IN OFDM SYSTEMS 2807
e

(n) and

H

(n) are independent [31]; thus, the instanta-


neous SNR of the decision variable on each subcarrier is
(n) =
[S(n)[
2

H(n)

2
(1 +N
0

n
)
_
N
0

n
[S(n)[
2
+N
0
(1 +N
0

n
)
_
=

H(n)

2
(1 +
n
/)(1 +
n
+
n
/)
. .
loss due to channel estimation
(18)
where we used [S(n)[
2
= 1 for MPSK modulation, and :=
1/N
0
may be viewed as the average SNR without channel
estimation error. Note that
n
was dened in (14) as a function
of the pilot allocation matrix.
Given (n), the SER with MPSK modulation on the nth
subcarrier is approximately
P
e
(n) 2Q
_
_
2(n) sin(/M)
_
(19)
where the Q-function is dened as Q(z) = (2)
1/2

z
e

2
/2
d [40, eq. (4.1)]. The SER, however, has practical
meaning only on the nonpilot subcarriers, i.e., the data-bearing
subcarriers. Using the denition of (5), these subcarriers may
readily be identied by indicators of the form 1 x
nn
, n.
The average SER on the data-bearing subcarriers is then, as a
function of X, given by
P
e
(X)=
2
NN
p
N1

n=0
(1x
nn
)Q
_
_
2(n) sin(/M)
_
(20)
where x
nn
is dened in (5). Hence, we have the following
optimization problem in X:
optimize X
opt
= arg min
X1
p
P
e
(X)
subject to trX = N
p
(21)
where the constraint is from (6).
Given the high computational complexity of the Q-function,
we may also resort to suboptimal solutions. Using the Chernoff
bound
1
[40, pp. 85], for example, will lead to an upper bound
on (20) as
P
e
(X)
N1

n=0
(1 x
nn
)e
(n) sin
2
(/M)

N1

n=0
(1 x
nn
)(n), for (n) 0 (22)
where we also used the fact that e
ax
< x for 0 < a 1 and
x 0 to get the second expression.
1
Chernoff bound: Q(x) (1/2)e
x
2
/2
, which is tight for x 0.
Thus, we have the sum-power optimization problem of
optimize X
sum
= arg max
X1
p
_
N1

n=0
(1 x
nn
)(n)
_
subject to trX = N
p
. (23)
Otherwise, because at a high SNR, the error rate of (20) is
prominently dominated by the fading subcarrier(s), an alterna-
tive could be the max-min problem of
optimize X
maxmin
= arg max
X1
p
_
min
0nN1
(n)
_
subject to trX = N
p
. (24)
B. Optimizing for Capacity
A lower bound on the OFDM channel capacity with channel
estimation error has been derived in [32] and [41] for equal
error covariance on all the subcarriers, i.e., uniform pilot al-
location. Such expressions may readily be extended to general
location pilots. Using the indicators 1 x
nn
, n, (2) and [32,
eq. (24)(26)], we can get a lower bound expression for the
average capacity on the data subcarriers as a function of the
pilot allocation matrix in (5). We have
C(X)
1
N
N1

n=0
E
_
log
2
_
1 + (1 x
nn
)
n
[H(n)[
2
__
(25)
where
n
= (1
2
e(n)
)/(
2
e(n)
+N
0
) = (
n
)/(1 +
n
)
may be viewed as an effective SNR (see [32, App. II] for details
of derivation). Note that, as N
0
0,
2
e(n)
0, and
n

, corresponding to perfect channel estimation. The capacity
optimizing problem is then simply
optimize X
cap
= arg max
X1
p
C(X)
subject to trX = N
p
. (26)
To summarize, the task of the proposed OFDM receiver at
timestamp t is given as follows.
1) Estimate the current channel

H(t) using the current pilot
allocation matrix X(t) using LS via the expression in
(10).
2) Using

H(t), perform data detection on each subcarrier
using an (imperfect) MF via (17).
3) Update X(t) X(t + 1) using the minimal SER criteria
via the integer program of (21), (23), or (24) or using the
maximal capacity criteria via the integer program in (26).
4) Feed the updated pilot allocation matrix X(t + 1) back to
the transmitter, which will be used in the next signaling
interval.
This concludes the nonuniformpilot allocation strategies. We
elaborate more on the mismatch that was caused by using X(t)
for the (t + 1)th interval in Section VII.
2808 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 6, JULY 2009
V. FEEDBACK REDUCTION
A fundamental issue in feedback-oriented designs is the
amount of overhead that is associated with the feedback infor-
mation. Ultimately, the feedback reduces the capacity, and the
benet of the overhead must outweigh its cost in capacity; thus,
the goal is to reduce the amount of feedback as much as possible
while retaining most of its benets. Examining this tradeoff for
our proposed system is the topic of this section.
The feedback information for nonuniform pilots is fully
characterized by the pilot allocation matrix X. For each chan-
nel realization H, a unique X matrix is selected from 1
p
=
X
1
, X
2
, . . . , X
[1
p
[
using one of the aforementioned opti-
mization criteria. Hence, each Xconveys log
2
[1
p
[ = log
2
_
N
N
p
_
bits of information that must be relayed to the transmitter,
where, for simplicity, we assume that the feedback link is error
free and instantaneous. The factorial nature of the amount of
feedback implies very high overhead, even for moderate values
of N and N
p
. To reduce feedback, we propose using subcarrier
clustering via the generalized Lloyd algorithm (GLA), which
will be explained as follows.
Assuming that the size of the pilot allocation codebook 1
p
is Q = [1
p
[, the idea in codebook clustering is to partition the
original codebook into a total of Q
/
subsets where Q
/
Q and
to construct a new lower order codebook such as
1
/
p
=
_

X
(1)
,

X
(2)
, . . . ,

X
(Q

)
_
,

1
/
p

= Q
/
. (27)
One method of creating the new codebook is to use vector
quantization. Vector quantization is a technique often used in
lossy data compression, in which the basic idea is to code values
from a multidimensional vector space into values from a dis-
crete subspace of lower dimension. Vector quantization can ef-
ciently be implemented using the well-known GLA [42][44].
GLA is basically an extension of the theory of nonuniform
quantization to vectors and matrices. It is an iterative quan-
tization method that converges to an optimum quantization
while minimizing some user-dened distortion function. The
distortion function is design specic, and here, we choose the
following metric:
T(1
/
) = E
_
min
X1

_
_
_XX
(opt)
_
_
_
2
F
_
. (28)
The objective of an optimal quantizer is to seek the codebook
that minimizes the average distortion over all possible code-
books. It can easily be shown that the optimal quantizer must
satisfy the following two conditions: 1) It must be a nearest
neighbor quantizer, i.e., it assigns to an arbitrary vector the
codeword that is closest to it, and 2) for a given partition of
the feature space, it must satisfy the centroid condition, i.e.,
each codeword must be the centroid of the vectors that are
mapped to it. We will design our vector quantizer to satisfy both
of these conditions. The clustering algorithm is summarized as
follows:
1) Training. Randomly generate M OFDM CFRs and con-
struct the codebook H = H
(1)
, H
(2)
, . . . , H
(M)
ac-
cording to the complex Gaussian channel model of (1).
2) For each channel within this codebook, determine the
corresponding optimum (or suboptimum) pilot allocation
matrix, and construct ( = X
(1)
, X
(2)
, . . . , X
(M)
.
3) Of the M pilot allocation matrices of (, randomly choose
Q
/
to construct the initial pilot index codebook (
/
0
=

X
(1)
0
,

X
(2)
0
, . . . ,

X
(Q

)
0
.
4) Set i = 1 as the iteration counter.
5) Clustering. Partition the vectors of ( into a total of Q
/
Voronoi regions (quantization regions) using a minimum
distance criteria such that the qth region is dened as
Q
q
=
_
X (

_
_
_X

X
(q)
i1
_
_
_
2
F

_
_
_X

X
(l)
i1
_
_
_
2
F
, l ,=q
_
.
6) Construct a new codebook (
/
i
, with the qth matrix of the
qth Voronoi region being given as

X
(q)
i
= arg min
X
E
_
_
_
XX
_
_
2
F
_
, X Q
q

=
1
[Q
q
[

XQ
X (29)
where the last step is an approximation by choosing the
center of mass of the Voronoi region.
7) If T((
/
i
) T((
/
i1
) or i I, terminate the algo-
rithm. Otherwise, set i = i + 1, and go to Step 5. The
distortion function T() is dened in (28) and can be
approximated with the centroid as in step 6, [see (29)].
8) The approximation of (29) leads to

X
(q)
i
matrices with
entries from R or the set of real numbers; thus, the pilot
allocation matrix of the nal codebook will violate the
binary structure for all X matrices as required by (5).
Hence, as a nal step, we map the codebook to binary
matrices that conform to (5). One method is to consider
the rst N
p
largest values on the diagonal of each

X
(q)
i
and substitute them with unity while substituting zero for
all other entries.
2
VI. EXTENDING TO MIMO STBC
We extend the notion of nonuniform pilot allocation to a
2 1 Alamouti-based STBC OFDM (A-STBC-OFDM) system
[45]. Consider two successive OFDM symbols S
k
and S
k+1
in
time (i.e., N N diagonal matrices). Denote the rst symbol
as the odd symbol S
o
and the second as the even symbol S
e
.
For the rst antenna, S
o
is transmitted during the rst time
slot, followed by S

e
in the second time slot. For the second
antenna, S
e
is transmitted rst, followed by S

o
. At each time
instance, two OFDM symbols are transmitted into the channel;
thus, to make future comparisons with single antenna systems
fair in terms of power, we reduce the power of each symbol by
half. Equivalently, these symbols are encoded in space and time
to form the transmission matrix, i.e.,
(
2
=
1

2
_
S
o
S

e
S
e
S

o
_
(30)
2
Assuming here that the GLA has iterated a total of I times.
PANAH et al.: OPTIMIZING PILOT LOCATIONS USING FEEDBACK IN OFDM SYSTEMS 2809
where the columns and rows of (
2
represent temporal and
spatial dimensions, respectively. Assume that H
1
and H
2
are
the CFRs between the single receiver antenna and the rst and
second transmitting antennas, respectively. In addition, assume
that these channels are constant and mutually independent. The
received signals in the corresponding time slots are
R
1
=
1

2
(S
o
H
1
+S
e
H
2
) +V
1
(31)
R
2
=
1

2
(S

e
H
1
+S

o
H
2
) +V
2
(32)
where V
1
and V
2
are uncorrelated AWGN with identical
correlation matrices of
2
n
I
N
. In matrix form, we have

R =
1

2
_
S
o
S
e
S

e
S

o
_

H+

V (33)
where

R = [R
T
1
R
T
2
]
T
,

H = [H
T
1
H
T
2
]
T
, and

V = [V
T
1
V
T
2
]
T
are (2N 1)-length vectors. Formulation for the LS channel
estimation procedure using (33) and the pilot allocation matrix
is given in the Appendix.
VII. NUMERICAL RESULTS
A. Setup
For each channel realization, the corresponding pilot alloca-
tion matrix conveys log
2
[1
p
[ = log
2
_
N
N
p
_
bits of information
that must be relayed to the transmitter. For simplicity, we
assume that the feedback link is error free and instantaneous.
However, we acknowledge that the algorithmic structure of the
receiver (as explained at the end of Section IV) yields a pilot
allocation matrix that is outdated by one OFDM symbol time
T
ofdm
= NT
s
. To account for this mismatch, we incorporate,
in our simulations, a time-varying frequency-selective channel
model with normalized Doppler frequency F
d
= f
d
NT
s
, where
f
d
is the channel Doppler frequency (relating to the receiver
velocity and the scenario, i.e., the antenna pattern and the
distribution of incoming waves at the receiver), and T
s
is the
symbol-rate (normalized here to unity, i.e., T
s
= 1). Clearly,
the mismatch will be reected in the choice F
d
, with the
number of subcarriers, i.e., N, being held constant. The channel
gains are modeled according to Section II-A and are simulated
using the method for generation of correlated Rayleigh random
variates by the IFFT formulated in [46]. This method yields
accurate discrete channel realizations obeying the time auto-
correlation function EH
t
(n)H

t+t
(n) = J(2t
/
F
d
), where
H
t
(n) is the nth element of the CFR at time t, and J() is the
zero-order Bessel function of the rst kind.
B. Error-Rate Performance
To compare the proposed feedback schemes, we begin by
simulating the SER in Fig. 1 over a multipath channel of length
L = 4 with a normalized Doppler frequency of F
d
= 10
4
, i.e.,
slow fading. A total of N = 16 subcarriers is used for modula-
tion, of which N
p
= 4 are reserved for equipower pilot symbols
over a channel with two multipath components, i.e., L = 2.
Fig. 2 shows the SER performance with QPSK modulation over
Fig. 2. Average SER performances: N = 16, N
p
= 4, and L = 2.
a wide range of SNRs for the uniform (benchmark) PSAM and
the nonuniform PSAM proposed here, i.e., the optimum of (21),
the max-min of (24), the sum power of (23), and the SNR-
maximizing method in [4]. Clearly, as shown in Fig. 2, there is
signicant improvement in using feedback to optimize PSAM-
OFDM, particularly at high SNRs, where fading is a more
prominent factor. To achieve this gain, the required feedback
payload is log
2
_
16
4
_
| = 11 bits (per channel realization). This
payload, plus protective coding bits and any protocol overhead,
is the capacity cost.
Given a specic channel response, it is difcult to make
generalizations on how any of the proposed optimizations al-
locate pilot locations, even at asymptotic SNR values. Fig. 4,
however, sheds some light into the matter by illustrating the
resulting (nonuniform) pilot allocations for a single realization
of a length L = 4 CFR operating over N = 32 subcarriers at an
SNR of = 20 dB. The total number of pilots in each method
is N
p
= 4, which is the minimum number required for proper
ML estimation. In addition, Fig. 4 shows that there are the four
trivial uniform pilot locations: 1) 1; 2) 9; 3) 17; and 4) 25. This
particular channel realization suffers a deep fade in the vicinity
of the 24th subcarrier and a marginal fade around the 16th
subcarrier. Note here that, given a reasonably high SNR, the
deep-fading subcarrier is of particular interest, because it should
account for the primary contribution to the (instantaneous)
SER. It is interesting to see how both the optimum and the
max-min solutions combat fading by allocating pilots to the
24th subcarrier. By doing so, more reliable, i.e., higher gain,
subcarriers may be made available for data transmission. The
sum-power solution, on the other hand, strives to reach a more
conservative solution, i.e., allocating both high- and low-gain
pilot locations. In other words, similar to uniformallocation, the
sum-power strategy aims at leveraging data power output, i.e.,
high data SNR via the allocation of pilots to low channel gains,
on one hand, and acceptably low channel MSE, i.e., decreasing
the interpolation noise via the allocation of pilots to high
channel gain locations, on the other. However, as Fig. 3 clearly
shows, in the low SNR regime (< 10 dB), where AWGN is
dominant, such conservative equalizing strategies serve to be
benecial, resulting in lower SER for the sum-power solution,
2810 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 6, JULY 2009
Fig. 3. Low SNR regime.
Fig. 4. Example. Uniform and nonuniform pilot allocation over a length
L = 4 channel with N = 32 subcarriers and N
p
= 4 pilots per symbol.
as compared with the max-min solution. In fact, at SNRs that
are lower than 3 dB, even the uniform allocation marginally
outperforms the max-min solution, altogether questioning the
use of feedback in this range of SNRs. In addition, these
gures show the performance of the (optimum) average SNR-
maximizing solution in [4]. The method in [4] optimizes the
overall SNR, which is dened over the entire OFDM band,
hence ignoring the per-channel uctuations that have a greater
inuence on the SER. The methods here (i.e., optimum and
suboptimum), on the other hand, account for these uctuations
and, hence, exhibit better performances in terms of SER over
the entire range of SNRs.
C. Capacity Performance
Fig. 5 shows the average capacity as derived in (26) com-
pared with the uniform pilot allocation. In addition, the hy-
pothetical case of no channel estimation error (i.e., uniform
allocation with
n
= 0, n) is included. It is worthwhile to
point out that the authors in [32] consider the benets in
terms of capacity of optimizing the relative power between the
Fig. 5. Comparison of average capacity: N = 16, and N
p
= 4.
Fig. 6. Average SER performance of codebooks: N = 16, and N
p
= 4.
data and pilot symbols for uniform pilot allocation. Here, we
consider equal power pilots and consider the benets in terms of
capacity to optimize the location of the pilots. Interestingly, by
comparing [32, Fig. 1] and our plot in Fig. 5, we see that we get
almost identical results, which is correct by intuition, because
by relocating the pilots, we are, in effect, changing their power
(relative to data) as they experience different channel gains at
the different frequencies.
D. Limited Feedback Performance
We begin by creating various-bit codebooks for N = 16,
N
p
= 4, and L = 2 using the GLA in Section V, with 2 10
4
initial channel instances and I = 5 iterations. We use the op-
timum average SER criteria in (21) to obtain X
opt
s at each
iteration and for all Voronoi regions. Note that the objective
function is a function of the SNR; hence, to take a conservative
approach, we design our codebook at = 30 dB (high SNR).
The distortions, as calculated from (28), and the respective SER
performances are shown in Figs. 6 and 7, respectively. Note
how, despite a 7-bit reduction in feedback, the 4-bit codebook
is within 1 dB of the optimal performance at an SER of 10
2
.
PANAH et al.: OPTIMIZING PILOT LOCATIONS USING FEEDBACK IN OFDM SYSTEMS 2811
Fig. 7. Average distortion of codebooks.
Fig. 8. Average SER versus normalized Doppler frequency. N = 16, and
N
p
= 4.
E. Performance Under Imperfections
We simulate the performances under two prominent OFDM
channel impairments: 1) Doppler frequency and 2) carrier
frequency offset. To evaluate the effects of the mismatch in
Section IV, we rst simulate the average SER performances
of the proposed solutions versus the normalized Doppler
frequency F
d
at an SNR of 15 dB. As expected, due to the
increase in the mismatch of the two subsequent OFDM channel
responses, the performance of our schemes deteriorate as
F
d
increases. The uniform pilot spacing is, of course, not
affected by F
d
. However, Fig. 8 shows that the nonuniform
solutions are superior for a wide range of Doppler frequencies.
To account for frequency offset, we dene [[ 1/2 as
the frequency offset (normalized to the subcarrier spacing)
and rewrite (3) as R(n) = A(n)S(n)H(n) +I(n) +V (n),
where A(n) = N
1
e
j(N1)/N
(sin /(sin /N)) is the
symbol rotation in frequency, and I(n) = N
1

N1
k=0

n,=m
S(n)H(n)e
j2k(nm+)/N
is the intercarrier-
interference term [47]. The rest of the proposed optimization
procedure in Section IV then follows with regard to the rotated
OFDM symbol, and Fig. 9 shows the SER performances at
an SNR of 15 dB. It is interesting to note here how the high-
Fig. 9. Average SER versus normalized frequency offset. N = 16, and
N
p
= 4.
Fig. 10. A-STBC-OFDM. N = 16, and N
p
= 4.
performances are quite similar to the low-SNR performances
in Fig. 3, which means that the max-min solution performs
worse than the sum-power solution. In fact, the optimum
and sum-power solutions are quite robust to , whereas the
max-min solution is only effective for [[ < 0.2.
F. OSTBC Performance
Fig. 10 shows the average SER performance of the A-STBC-
OFDM system using uniform and nonuniform (optimized) pilot
allocation. In addition, the previous single-inputsingle-output
(SISO) results in Fig. 2 are shown. Again, there is improvement
in using nonuniformpilots with feedback. When compared with
the SISO case, however, this improvement is less signicant,
because the pilot allocation must leverage between two un-
correlated channel matrices. A more interesting observation,
however, lies in the diversity order that each method obtained.
The 21 A-STBC-OFDM is known to have a diversity order
of 2, because it does not utilize any portion of the frequency
diversity (equal to the channel length). Looking at the different
curves in Fig. 10, we see that the optimum allocation for the
SISO case achieves the same slope (diversity) as the uniform
2812 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 6, JULY 2009
Alamouti case. Hence, it appears that optimum nonuniform
pilots for a SISO-OFDM system can obtain a signicant, if
not complete, frequency dimensionality that is inherent in the
system.
VIII. CONCLUSION
We have introduced closed-loop PSAM for OFDM systems
in wireless channels. We have derived the average error rate
and average capacity of PSAM-OFDM, and we have shown
how the pilot symbols may be adapted to the current chan-
nel conditions to improve performance. By considering upper
bound(s) on the SER, we have arrived at practical suboptimum
solutions with acceptable performance losses. At an average
SER of 10
2
, our simulations have shown an improvement of
more than 5 dB relative to conventional uniform PSAM for an
OFDM system operating over a slow-fading channel. We have
suggested a codebook-reduction framework based on GLA to
reduce feedback and extended our method to a MIMO-OFDM
structure based on the 21 Alamouti STBC. An interesting
direction for future research could be extensions to multiuser
MIMO-OFDMA systems.
APPENDIX
NONUNIFORM LS CHANNEL ESTIMATION FOR
A-STBC-OFDM
Similar to the single-antenna case, we are left with the task of
determining the pilot locations using the pilot allocation matrix
X. Having two distinct channels, i.e., H
1
and H
2
, however,
begs the question: Do we need two allocation matrices, i.e.,
X
1
and X
2
? The answer is no. This can easily be justied by
viewing (30) on a per-carrier basis. If, for example, S
o
has been
allocated a pilot on a certain subcarrier, then S

o
is automatically
allocated a pilot on the same subcarrier. Note that S
o
and S

o
are symbols that were transmitted through the channels H
1
and
H
2
, respectively. We can, therefore, conclude that a single pilot
allocation matrix must jointly be designed as function of both
channel responses. The solution that we consider here is the
pilot allocation matrix that minimizes the average of SER over
both symbols, i.e.,
optimize X
opt
STBC
= arg min
X1
p
P
e(S
e
)
(X) +P
e(S
o
)
(X)
subject to trX = N
p
. (34)
The pilot allocation matrix X may be used to extract the pilots
from the received signals in (33) by dening the extended pilot
allocation matrix, i.e.,

X =
_
X 0
NN
0
NN
X
_
= I
2
X (35)
and multiplying both sides of (33) by

Xto get

R
(pilot)
=

X

R.
Based on (33)

R
(pilot)
=
1

2
_
S
o
X S
e
X
S

e
X S

o
X
_

H+ w
=
1

2
_
s
p
X s
p
X
(s
p
X)

(s
p
X)

_

H+ w
=S
p
_
X 0
NN
0
NN
X
_

H+ w
=S
p
(I
2
X)

H+ w (36)
where w := [(XV
1
)
T
(XV
2
)
T
]
T
, s
p
is a known pilot symbol,
and we have dened
S
p
=
1

2
_
s
p
I
N
s
p
I
N
s

p
I
N
s

p
I
N
_
, S
H
p
S
p
= I
2N
.
Multiplying both sides of (36) by S
H
p
and considering the CIR
vector

h := [h
T
1
h
T
2
]
T
on the right hand side, we get S
H
p
(I
2

X)

R
(pilot)
= (I
2
XF
L
)

h + w and the LS solution of

H = (I
2
F
L
)(I
2
XF
L
)

S
H
p
(I
2
X)

R
(pilot)
(37)
where

H := [

H
T
1

H
T
2
]
T
. With the channel estimates

H
1
and

H
2
, the derivation of the average SER as a function of X is
similar to the procedure in Section IV and is omitted here.
REFERENCES
[1] R. Chang and R. Gibby, A theoretical study of performance of an or-
thogonal multiplexing data transmission scheme, IEEE Trans. Commun.,
vol. COM-16, no. 4, pp. 529540, Aug. 1968.
[2] S. Weinstein and P. Ebert, Data transmission by frequency-division mul-
tiplexing using the discrete Fourier transform, IEEE Trans. Commun.,
vol. COM-19, no. 5, pp. 628634, Oct. 1971.
[3] J. A. C. Bingham, Multicarrier modulation for data transmission: An idea
whose time has come, IEEE Commun. Mag., vol. 28, no. 5, pp. 514,
May 1990.
[4] A. Y. Panah, B. Nosrat-Makouei, and R. G. Vaughan, Non-uniform
pilot-symbol allocation for closed-loop OFDM, IEEE Trans. Wireless
Commun., vol. 7, no. 7, pp. 27232731, Jul. 2008.
[5] L. Cimini, Jr., Analysis and simulation of a digital mobile channel us-
ing orthogonal frequency-division multiplexing, IEEE Trans. Commun.,
vol. COM-33, no. 7, pp. 665675, Jul. 1985.
[6] S. Coleri, M. Ergen, A. Puri, and A. Bahai, Channel estimation tech-
niques based on pilot arrangement in OFDM systems, IEEE Trans.
Broadcast., vol. 48, no. 3, pp. 223229, Sep. 2002.
[7] J. Kim, J. Park, and D. Hong, Performance analysis of channel estimation
in OFDM systems, IEEE Signal Process. Lett., vol. 12, no. 1, pp. 6062,
Jan. 2005.
[8] H. Cheon and D. Hong, Effect of channel estimation error in OFDM-
based WLAN, IEEE Commun. Lett., vol. 6, no. 5, pp. 190192,
May 2002.
[9] J. Rinne and M. Renfors, Pilot spacing in orthogonal frequency-division
multiplexing systems on practical channels, IEEE Trans. Consum.
Electron., vol. 42, no. 4, pp. 959962, Nov. 1996.
[10] M. Dong, L. Tong, and B. M. Sadler, Optimal pilot placement for channel
tracking in OFDM, in Proc. MILCOM, Oct. 2002, vol. 1, pp. 602606.
[11] Y. Li, L. J. Cimini, Jr., and N. R. Sollenberger, Robust channel estimation
for OFDM systems with rapid dispersive fading channels, IEEE Trans.
Commun., vol. 46, no. 7, pp. 902915, Jul. 1998.
[12] J. J. van de Beek, O. Edfors, M. Sandell, S. K. Wilson, and P. O. Borjesson,
On channel estimation in OFDM systems, in Proc. IEEE 45th Veh.
Technol. Conf., Chicago, IL, Jul. 1995, vol. 2, pp. 815819.
[13] A. Y. Panah, B. N. Makouei, and R. G. Vaughan, An expectation-
maximization solution to interpolated OFDM systems, in Proc. IEEE
CCECE, Vancouver, BC, Canada, Apr. 2226, 2007, pp. 5255.
[14] W. G. Song and J. T. Lim, Pilot-symbol-aided channel estimation for
OFDM with fast fading channels, IEEE Trans. Broadcast., vol. 49, no. 4,
pp. 398402, Dec. 2003.
[15] M.-X. Chang and Y. T. Su, Performance analysis of equalized OFDM
systems in Rayleigh fading, IEEE Trans. Wireless Commun., vol. 1, no. 4,
pp. 721732, Oct. 2002.
[16] S. Chennakeshu and J. B. Anderson, Error rates for Rayleigh fading
multichannel reception of MPSK signals, IEEE Trans. Commun., vol. 43,
no. 234, pp. 338346, Feb.Apr. 1995.
PANAH et al.: OPTIMIZING PILOT LOCATIONS USING FEEDBACK IN OFDM SYSTEMS 2813
[17] X. Tang, M.-S. Alouini, and A. J. Goldsmith, Effect of channel estima-
tion error on M-QAMBERperformance in Rayleigh fading, IEEE Trans.
Commun., vol. 47, no. 12, pp. 18561864, Dec. 1999.
[18] I. Gaspard, Impact of the channel estimation onto the BER-performance
of PSAM-OFDM systems in mobile radio channels, in Proc. IEEE Veh.
Technol. Conf., 2001, vol. 1, pp. 673677.
[19] W. Zhang, X.-G. Xia, and P. Ching, Optimal training and pilot pattern
design for OFDM systems in Rayleigh fading, IEEE Trans. Broadcast.,
vol. 52, no. 4, pp. 505514, Dec. 2006.
[20] M. J. F.-G. Garca, J. L. Rojo-lvarez, F. Alonso-Atienza, and
M. Martnez-Ramn, Support vector machines for robust channel estima-
tion in OFDM, IEEE Signal Process. Lett., vol. 13, no. 7, pp. 397400,
Jul. 2006.
[21] X. Zhou and X. Wang, Channel estimation for OFDM systems using
adaptive radial basis function networks, IEEE Trans. Veh. Technol.,
vol. 52, no. 1, pp. 4859, Jan. 2003.
[22] J. H. Manton, Optimal training sequences and pilot tones for
OFDM systems, IEEE Commun. Lett., vol. 5, no. 4, pp. 151153,
Apr. 2001.
[23] J.-W. Choi and Y.-H. Lee, Optimum pilot pattern for channel estima-
tion in OFDM systems, IEEE Trans. Wireless Commun., vol. 4, no. 5,
pp. 20832088, Sep. 2005.
[24] S. Ohno and G. B. Giannakis, Average-rate optimal PSAM transmissions
over time-selective fading channels, IEEE Trans. Wireless Commun.,
vol. 1, no. 4, pp. 712720, Oct. 2002.
[25] D. Hu, L. He, and L. Yang, Joint pilot tone and channel estimator design
for OFDM systems, IEEE Commun. Lett., vol. 10, no. 10, pp. 695697,
Oct. 2006.
[26] J. K. Cavers, An analysis of pilot-symbol-assisted modulation for
Rayleigh fading channels, IEEE Trans. Veh. Technol., vol. 40, no. 4,
pp. 686693, Nov. 1991.
[27] R. Negi and J. Ciof, Pilot tone selection for channel estimation in a
mobile OFDM system, IEEE Trans. Consum. Electron., vol. 44, no. 3,
pp. 11221128, Aug. 1998.
[28] M.-X. Chang, A new derivation of least-squares-tting principle for
OFDM channel estimation, IEEE Trans. Wireless Commun., vol. 5, no. 4,
pp. 726731, Apr. 2006.
[29] M. Morelli and U. Mengali, A comparison of pilot-aided channel estima-
tion methods for OFDM systems, IEEE Trans. Signal Process., vol. 49,
no. 12, pp. 30653073, Dec. 2001.
[30] S. Ohno and G. B. Giannakis, Optimal training and redundant precoding
for block transmissions with application to wireless OFDM, IEEE Trans.
Commun., vol. 50, no. 12, pp. 21132123, Dec. 2002.
[31] X. Cai and G. B. Giannakis, Error probability minimizing pilots for
OFDM with MPSK modulation over Rayleigh-fading channels, IEEE
Trans. Veh. Technol., vol. 53, no. 1, pp. 146155, Jan. 2004.
[32] S. Ohno and G. B. Giannakis, Capacity maximizing MMSE-optimal
pilots for wireless OFDM over frequency-selective block Rayleigh-fading
channels, IEEE Trans. Inf. Theory, vol. 50, no. 9, pp. 21382145,
Sep. 2004.
[33] S. Sanayei, A. Nosratinia, and N. Aldhahir, Opportunistic dynamic sub-
channel allocation in multiuser OFDM networks with limited feedback,
in Proc. IEEE Inf. Theory Workshop, 2004, pp. 182186.
[34] J. Choi and R. W. Heath, Jr., Interpolation based transmit beamforming
for MIMO-OFDM with limited feedback, IEEE Trans. Signal Process.,
vol. 53, no. 11, pp. 41254135, Nov. 2005.
[35] D. J. Love, R. W. Heath, Jr., W. Santipach, and M. L. Honig, What is the
value of limited feedback for MIMO channels? IEEE Commun. Mag.,
vol. 42, no. 10, pp. 5459, Oct. 2004.
[36] H. Lo, D. Lee, and J. A. Gansman, A study of non-uniform pilot spacing
for PSAM, in Proc. IEEE Int. Conf. Commun., 2000, vol. 1, pp. 322325.
[37] W. Jakes and D. Cox, Microwave Mobile Communications. New York:
Wiley-IEEE Press, 1994.
[38] S. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Upper Saddle River, NJ: PrenticeHall, 1993.
[39] R. Bhatia, Matrix Analysis. New York: Springer-Verlag, 1997.
[40] M. Simon and M. Alouini, Digital Communication Over Fading Chan-
nels: A Unied Approach to Performance Analysis. Hoboken, NJ: Wiley,
2000.
[41] M. Torabi, S. Aissa, and M. R. Soleymani, MIMO-OFDM systems with
imperfect channel information: Capacity, outage and BER performance,
in Proc. IEEE Int. Conf. Commun., Jun. 2006, vol. 12, pp. 53425347.
[42] A. Gersho and R. Gray, Vector Quantization and Signal Compression.
Norwell, MA: Kluwer, 1992.
[43] Y. Linde, A. Buzo, and R. Gray, An algorithm for vector quan-
tizer design, IEEE Trans. Commun., vol. COM-28, no. 1, pp. 8495,
Jan. 1980.
[44] I. Katsavounidis, C.-C. J. Kuo, and Z. Zhang, A new initialization tech-
nique for generalized Lloyd iteration, IEEE Signal Process. Lett., vol. 1,
no. 10, pp. 144146, Oct. 1994.
[45] S. M. Alamouti, A simple transmit diversity technique for wireless com-
munications, IEEE J. Sel. Areas Commun., vol. 16, no. 8, pp. 14511458,
Oct. 1998.
[46] D. J. Young and N. C. Beaulieu, The generation of correlated Rayleigh
random variates by inverse discrete Fourier transform, IEEE Trans.
Commun., vol. 48, no. 7, pp. 11141127, Jul. 2000.
[47] X. Ma, H. Kobayashi, and S. Schwartz, Joint frequency offset and
channel estimation for OFDM, in Proc. IEEE GLOBECOM, 2003, vol. 1,
pp. 1519.
Ali Y. Panah (S06) was born in Tehran, Iran,
on August 1, 1980. He received the B.Sc. degree
in electrical engineering from Sharif University of
Technology, Tehran, in 2004 and the M.A.Sc. de-
gree in electrical engineering from the Simon Fraser
University, Burnaby, B.C., Canada, in 2007. He is
currently working toward the Ph.D. degree with the
Department of Electrical and Computer Engineering,
University of Texas at Austin.
He is currently with the Wireless Networking and
Communications Group, Department of Electrical
and Computer Engineering, University of Texas at Austin. His research interests
include signal processing for mobile communications, in particular multiple-
inputmultiple-output with limited feedback, orthogonal frequency division
multiplexing channel estimation, and cooperative communications via relays.
Rodney G. Vaughan (M84SM89F07) received
the Bachelor and Masters degrees in electrical
engineering from the University of Canterbury,
Christchurch, New Zealand, in 1975 and 1976, re-
spectively, and the Ph.D. degree in electrical engi-
neering fromAalborg University, Aalborg, Denmark,
in 1985.
He was with the New Zealand Post Ofce (now
Telecom NZ Ltd.), the NZ Department of Scientic
and Industrial Research, and Industrial Research Ltd.
He has worked on several mechanical and electrical
projects, including network analysis and trafc forecasting. He has also de-
veloped microprocessor and digital signal processing (DSP) technology for
equipment that range from abattoir hardware to communications networks. In
1982 and 1983, he was an International Union of Radio Science (URSI) Young
Scientist for Fields and Waves and for Electromagnetic Theory, respectively. He
developed research programs and personnel working in communications tech-
nology for Industrial Research Limited, in particular multipath communications
theory (e.g., electromagnetic, line, and acoustic media), diversity design, signal
and sampling theory, signal processing, and DSP. His industrial projects include
the design and development of specialist antennas for personal, cellular, and
satellite communications, multiple-inputmultiple output (MIMO) test bed and
systems design, capacity theory, and spatial eld theory. In 2003, he became
a Professor of electrical engineering and the Sierra Wireless Chair in Commu-
nications with the School of Engineering Science, Simon Fraser University,
Burnaby, BC, Canada. His current research is focused on communications
signal processing, antennas, and multipath propagation. His recent projects
include compact mammalian bio-implantable antennas; multielement antenna
design and evaluation; multifaceted circularly polarized large array systems,
MIMO capacity realization, and precoding techniques.
Dr. Vaughan is a Fellow of the BC Advanced System Institute, an URSI
Correspondent, and a New Zealand URSI Commission B (Fields and Waves)
Representative. He has been a Guest Editor of several special issues for several
IEEE TRANSACTIONS, including the IEEE TRANSACTIONS ON ANTENNAS
AND PROPAGATION Special Issue on Wireless Communications.
2814 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 6, JULY 2009
Robert W. Heath Jr. (S96M01SM06) received
the B.S. and M.S. degrees in electrical engineering
from the University of Virginia, Charlottesville, in
1996 and 1997, respectively, and the Ph.D. degree
in electrical engineering from Stanford University,
Stanford, CA, in 2002.
From 1998 to 2001, he was a Senior Member of
Technical Staff and then a Senior Consultant with
Iospan Wireless Inc, San Jose, CA, where he worked
on the design and implementation of the physical
and link layers of the rst commercial multiple-
inputmultiple-output (MIMO)orthogonal frequency division multiplexing
(OFDM) communication system. In 2003, he founded MIMO Wireless Inc.: a
consulting company that is dedicated to the advancement of MIMO technology.
Since January 2002, he has been with the Department of Electrical and
Computer Engineering, University of Texas at Austin, where he is currently
an Associate Professor and the Associate Director of the Wireless Networking
and Communications Group. His research interests include several aspects of
MIMO communication, in particular limited feedback techniques, multihop
networking, multiuser MIMO, antenna design, and scheduling algorithms, as
well as 60-GHz communication techniques and multimedia signal processing.
Dr. Heath is a member of the Signal Processing for Communications
Technical Committee, at the IEEE Signal Processing Society. He was a
Technical Cochair of the 2007 Vehicular Technology ConferenceFall, the
General Chair of the 2008 Communication Theory Workshop, and a General
Cochair and Coorganizer of the 2009 Signal Processing for Wireless Com-
munications Workshop. He is the recipient of the David and Doris Lybarger
Endowed Faculty Fellowship in Engineering and is a registered Professional
Engineer in Texas. He has been an Editor for the IEEE TRANSACTIONS ON
COMMUNICATION and an Associate Editor for the IEEE TRANSACTIONS ON
VEHICULAR TECHNOLOGY.

You might also like