IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO.
10, OCTOBER 2008
A Robust Asynchronous Multiuser
STBC-OFDM Transmission Scheme for Frequency-Selective Channels
Dandan Wang, Student Member, IEEE, Hlaing Minn, Senior Member, IEEE, and Naofal Al-Dhahir, Fellow, IEEE
Abstract—In this paper, we propose a robust STBC trans-
mission scheme to combat timing synchronization errors over
frequency-selective multiple-access channels. We start by deriv-
ing the equivalent channel model in the presence of timing syn-
chronization errors. Based on this correlated equivalent channel
model, the statistical channel power gain profile is analyzed and
shown to exhibit unequal channel power gains on different sub-
carriers. Furthermore, a robust statistical bit loading algorithm
is proposed to optimize the BER performance in scenarios where
link adaptation based on instantaneous channel information
is infeasible or undesirable. Simulation results show that our
proposed scheme is robust to timing synchronization errors,
providing around 7.5dB SNR advantage at the BER of 10−4
over the conventional scheme in various channel environments.
Index Terms—STBC, OFDM, bit loading, asynchronous,
multiple-access, ISI.
I. I NTRODUCTION
M ULTIPLE-INPUT-MULTIPLE-OUTPUT (MIMO) sys-
tems can utilize the spatial diversity and multiplexing
gains using spatially-separated antennas in a multipath scat-
tering environment. In broadband systems, space-time block
codes (STBC) have been integrated with orthogonal frequency
division multiplexing (OFDM) by implementing STBC at a
block level in a structure commonly known as ST-OFDM
[1]. Since wireless communications is challenged by limited
spectral resources, multi-user spacial multiplexing has recently
received considerable attention. Multi-user MIMO systems
can significantly improve system throughput via transceiver
signal processing if the number of transmit antennas is much
larger than the number of receive antennas [2]. In this context,
a space-time transceiver was investigated in [3] over flat-
fading channels assuming two perfectly-time-synchronized
users. However, in practical systems, it is very difficult, if not
impossible, to achieve perfect synchronization among all up-
link users. In [4], the authors considered an equivalent channel
model for a flat-fading system with synchronization errors in a
cooperative network. Since the channels are usually frequency-
selective in broadband systems, this motivates us to investigate
the multi-user MIMO transceiver with synchronization errors
Manuscript received May 31, 2007; revised October 30, 2007 and March
26, 2008; accepted June 10, 2008. The associate editor coordinating the
review of this letter and approving it for publication was A. Chockalingam.
This work was supported in parts by the Erik Jonsson School Research
Excellence Initiative, the University of Texas at Dallas, USA, the Texas
Advanced Technology Program (ATP) under Contract 009741-0023-2003, the
National Science Foundation (NSF) under Contracts CCF 04-30654 and DMS
05-28010, and a gift from Texas Instruments Inc. The material in this paper
was presented in part at the IEEE Sarnoff Symposium, Princeton, NJ, USA,
May 2007.
The authors are with the Dept. of Electrical Engineering, University of
Texas at Dallas, Richardson, TX 75083 (e-mail: {dxw053000, hlaing.minn,
aldhahir}@utdallas.edu).
Digital Object Identifier 10.1109/T-WC.2008.070577
1536-1276/08$25.00 
c 2008 IEEE
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:11 from IEEE Xplore. Restrictions apply.
3725
over frequency-selective channels in spacially-multiplexed ST-
OFDM systems.
By exploiting the special structure of the statistical channel
profile of the equivalent channel model in the presence of tim-
ing synchronization errors, we propose a robust transmission
scheme to combat the synchronization errors. The proposed
scheme is appealing to scenarios where link adaptation based
on instantaneous channel information is infeasible or undesir-
able.
This paper is organized as follows. In Section II, the
system set-up and the equivalent channel model with syn-
chronization errors are presented. The statistical power profile
of the equivalent channel model and a robust statistical bit
loading algorithm are presented in Section III. In Section IV,
simulation results are presented, and conclusions are drawn in
Section V.
II. S YSTEM M ODEL
A. System Setup
In this paper, we consider an uplink system in which
two users access the base station at the same time within
the same bandwidth using spatial division multiple access
(SDMA). Each user has Nt transmit antennas, and there are
Nr receive antennas at the base station. Each user implements
orthogonal space-time block codes (OSTBC) [5] on each sub-
carrier over adjacent OFDM symbols each consisting of N
sub-carriers. At the base station, the FFT is first applied
and then an interference cancelation scheme is employed
in the frequency domain on each sub-carrier to decode the
independent information streams from different users. In this
paper, we assume that all the receive antennas at the base
station operate with the same clock and transmissions on all
the transmit antennas at each user are also synchronous. In
addition, the arrival delays from different transmit antennas
of the same user are assumed to be the same. However, the
arrival delays can vary from user to user. Then, the base station
has to decide on a single sampling time reference point for
the two users with different arrival delays, which induces a
timing synchronization error for each user.
B. Equivalent Channel Model
In this subsection, we will derive the equivalent multipath
channel model with synchronization errors. The distributions
of the channel taps between different transmit and receive
antenna pairs are assumed to be the same. Without loss of
generality, in the following, we analyze the equivalent channel
between the jth antenna of the ith user and the first receive
 3726 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 10, OCTOBER 2008

antenna. The original channel between the jth antenna of the where
ith user and the first receive antenna is ⎛ ⎞
p0 0 0 0 ... 0 0
Li,j ⎜ p1 p0 0 0 ... 0 0 ⎟
 ⎜ ⎟
gi,j,l δ(t − ti,j,l ), (1) ⎜ p2 p1 p0 0 ... 0 0 ⎟
gi,j (t) = ⎜ ⎟
⎜ p3 p2 p1 p0 ... 0 0 ⎟
l=1 ⎜ ⎟
⎜ 0 p3 p2 p1 ... 0 0 ⎟
where Li,j is the number of channel taps, gi,j,l and ti,j,l are P =⎜

⎜
⎟
⎟ (8)
⎜ . . . . ... . . ⎟
the channel tap gain and the delay of the lth channel tap, ⎜ 0 0 0 0 . . . p1 p0 ⎟
respectively. For notation simplicity, we neglect the subscripts ⎜ ⎟
⎜ 0 0 0 0 . . . p2 p1 ⎟
{i, j} in the following. Denote the combined transmit and ⎜ ⎟
⎝ 0 0 0 0 . . . p3 P2 ⎠
receive pulse shaping filter as p(t). Let τ denote the timing 0 0 0 0 ... 0 p3
error with respect to the ideal samples for the ni,j th symbol of
the j-th antenna of the i-th user. Then, the equivalent channel
III. ROBUST S TATISTICAL B IT L OADING S CHEME
is given by
In this section, we first analyze the statistical power profile
L
 of the equivalent channel model and then design a statistical
ha (t) = g(t) ⊗ p(t + τ ) = gl p(t + τ − tl ), (2) bit loading algorithm which utilizes this channel power profile
l=1 knowledge.
where ⊗ denotes the linear convolution operation.
The equivalent discrete-time channel is A. Statistical Channel Power Profile
L
 Let P (F ) and G(F ) denote the Fourier transforms of p(t)
ha (nTs ) = gl p(nTs + τ − tl ), (3) and g(t), respectively. Let H = [H(1), H(2), ..., H(N )]T
l=1 denote the frequency-domain sub-carrier channel gains of the
equivalent correlated channel tap vector h. Then, we have the
where T1s is N times the sub-carrier spacing. following proposition:
The equivalent channel vector is defined as Proposition 1: If |P (F )| ≥ 0 only for |F | ≤ KFs (K is an
h = [· · · , ha (−2Ts ) ha (−Ts ) ha (0) ha (Ts ) ha (2Ts ) · · · ]T . integer and Fs = T1s ) and E{gl gl∗ } = 0 for l = l , then the
Special Case: average channel power gain on the nth sub-carrier, denoted as
Let p(t) be the raised cosine pulse shape function given as Cn,n , is
K
L 
sin(πt/Ts ) cos(απt/Ts )  n 

p(t) = , (4) Cn,n = Fs 2
|P (( − k1 )Fs )| 2
E |gl |2
πt/Ts 1 − 4α2 t2 /Ts2 N
k1 =−K l=1
where α is the roll-off factor. For the truncated pulse shape  K−k
K−1 1 n n
function with the first-order side lobes only, p(t) is nonzero + 2Fs2 − k1 )Fs )||P (( − k1 − k2 )Fs )|
|P ((
only within −2Ts ≤ t ≤ 2Ts . N N
k1 =−K k2 =1
For 0 < τ < Ts , and tl = (l − 1)Ts , where l ≥ 1 is an L 


2
integer, define p0 = p(τ − Ts ), p1 = p(τ ), p2 = p(τ + Ts ) × E |gl | cos(2πk2 (τ + tl )Fs − θ1 + θ2 ) ,
and p3 = p(τ − 2Ts ). Then, the equivalent channel vector can l=1
be constructed from the original channel taps as (9)
n
where θ1 is the phase of − k1 )Fs ) and θ2 is the phase
P (( N
h = P g, (5) n
of P (( N − k1 − k2 )Fs ). Note that if p(t) is symmetric as
where typically used in practice, θ1 = θ2 = 0.
⎛ ⎞ Proof: The proof is given in Appendix I.
p3 0 0 0 ... 0 0 If tl = (l − 1)Ts , (9) becomes
⎜ p0 p3 0 0 ... 0 0 ⎟ ⎛ ⎞ L  K
⎜ p1 p0 p3 0 ... 0 0 ⎟ g1
⎜ ⎟   n
⎜ p2 p1 p0 p3 ... 0 0 ⎟ ⎜ g2 ⎟ 2
E(|gl | ) Fs 2
|P (( − k1 )Fs )|2
⎜ ⎟ ⎜ g ⎟ Cn,n =
⎜ 0 p2 p1 p0 ... 0 0 ⎟ ⎜ 3 ⎟ N
P =⎜
⎜
⎟ and g = ⎜ g4
⎟ ⎜
⎟.
⎟
l=1 k1 =−K
⎜ . . . . ... . . ⎟ ⎜ . ⎟
⎜ 0 0 0 0 ... p0 p3 ⎟ ⎝ .. ⎠  K−k
K−1 1 n n
⎜ ⎟ +2 |P (( − k1 )Fs )||P (( − k1 − k2 )Fs )|
⎜ 0 0 0 0 ... p1 p0 ⎟ gL N N
⎝ 0 0 0 0 ... p2 p1 ⎠ k1 =−K k2 =1
0 0 0 0 ... 0 p2 ×cos(2πk2 τ Fs − θ1 + θ2 )} .
(6)
(10)
For −Ts < τ < 0, define p0 = p(τ − Ts ), p1 = p(τ ), It can be seen from (10) that the average power does not
p2 = p(τ + Ts ) and p3 = p(τ + 2Ts ). Then, the equivalent change with the exact value of  gl and only varies with the
channel vector is given by total energy of all the taps, i.e., L 2
l=1 E(|gl | ).
The early-late gate synchronizer is a simple and practical
h = P  g, (7) method that is widely used for symbol synchronization and

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:11 from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 10, OCTOBER 2008 3727

of the raised cosine function with the sinc function, which

is very complicated to evaluate. Thus, we analyze the power
1
profile from the time-domain equivalent channel vector. From
Average channel power gain

(5), the average channel power gain on the nth sub-carrier is

0.8
given by
average
0.6 τ=0  L
L 
τ=0.1Ts ∗
Cn,n = Rm,k Vn,m Vn,k
τ=0.2Ts
k=1 m=1
0.4
τ=0.3Ts L
(13)
  j2π(n−1)(k−m)
τ=0.4Ts = Rm,m + Rm,k e N ,
0.2
τ=0.5Ts m=1 k−m=0

j2π(n−1)(m−1)
0
0 10 20 30 40 50 60 where Vn,m = e− N , Rm,k is the (m, k)-th element
Sub−carrier index of R, and R = E(hhH ) = P E(gg H )P . Then, we have the
Fig. 1. Average channel power gain profile on the sub-carriers for the
following proposition:
equivalent channel. Proposition 2: The average power gain profile for the equiv-
alent channel vector given in (5) for 0 < τ < Ts and
−Ts < τ < 0 are respectively given as
sampling at the receiver [6]. Define τ1 and τ2 as the arrival
3
L 
delays of Users 1 and 2, respectively. It is shown in [4] that the 2 2π(n − 1)
Cn,n = ( E(|gl | )) p2l + 2cos( )
sampling time decided by the early-late gate synchronizer is 
N
l=1 l =0
about τ1 −τ
2
2
ahead of the ideal time for one signal and τ1 −τ2
2
4π(n − 1)
behind the ideal time for the other, i.e., τ = τ1 −τ2
2
for User × (p0 p1 + p1 p2 + p3 p0 ) + 2cos( )(p0 p2 + p3 p1 )
1 and τ = − τ1 −τ 2
for User 2. Without loss of generality,  N
2 6π(n − 1)
we now assume 0 ≤ |τ1 − τ2 | ≤ Ts . In practical systems, +2cos( )(p3 p2 ) ,
τ is unknown at the receiver. Assume that τ is a random N
variable with probability density function (pdf) f (τ ). Then, (14)
given gl and tl , the average channel power profile is obtained and
by averaging (9) over τ as follows

 ∞ L 3

 2 2π(n − 1)
C̄n,n = Cn,n f (τ )dτ. (11) Cn,n =( E(|gl |2 )) p l + 2cos( )
N
−∞ l=1 l =0
2
Note that if tl is not fixed, we have to average (11) over tl to
obtain the average power profile.
Special Case Revisited:
If the pulse shaping filter is the raised cosine function given
in (4), then
⎧ (15)
⎪
⎨ Ts , 0 ≤ |F | < (1−α)
2Ts
P (F ) = Ts
[1 + cos( π(|F |2T2α
s −(1−α))
)], (1−α)
≤ |F | < (1+α)
⎪
⎩
2 2Ts
0, |F | ≥ (1+α)
2Ts
and K = 1 in the Proposition 1, and hence the average
channel power gain on the nth sub-carrier can be easily
calculated from (9). For tl = (l − 1)Ts , the numerical results
of the average channel power gain obtained from (10) are
shown in Fig. 1 with the assumption of τ being uniformly
distributed over [0, T2s ] and N = 64. From Fig. 1, we can
see that the larger the timing synchronization error is, the
smaller is the lowest channel power gain. Thus, larger timing
synchronization errors cause more serious power unbalance
among all the sub-carriers. Note that if a long pulse shape
truncation length is used, it can be well approximated by the
untruncated pulse for which (9) yields an easy computation
of the channel power gains. For the equivalent channel vector
given in (5) which only considers the first-order side lobes
of the pulse shape (i.e., a short truncation length), the average
power profile with tl = (l−1)Ts can be obtained from (10) by
evaluating the Fourier transform of the time-domain truncated
pulse shaping function. However, this involves the convolution
 4π(n − 1)  
× p l p l +1 )(p 0 p 2 + p 3 p 1 )
+ 2cos(
N
l =0

6π(n − 1)  
+2cos( )(p 3 p 0 ) .
N
Proof: See Appendix II.
2Ts
Regarding the above derivation, the following two remarks
(12)
are in order:
• We have assumed that the timing error is 0 < |τ1 − τ2 | <
Ts . When the timing error is D = dT +τ1 −τ2 , where d is
an integer and 0 < τ1 − τ2 < Ts , the frequency-domain-
based results in (9) and (10) still hold. The time-domain-
based approach for a short pulse shape truncation length
can also be applied with appropriate modifications.
• When there are more than two users, our derivation for
the channel power gain profile given a specific timing
error τ still holds. However, the value of τ may depend
on the exact synchronization algorithm used in the system
and the arrival delays of different users.
In Figs. 2 and 3, (14) and (15) are evaluated numerically
for N = 64 with the assumption of τ being uniformly
distributed over [0, T2s ] and [− T2s , 0], respectively. We observe
that the truncated pulse shaping function results in a slightly
larger imbalance of channel power gains across the subcarriers
compared with the non-truncated case shown in Fig. 1.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:11 from IEEE Xplore. Restrictions apply.
 3728 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 10, OCTOBER 2008

1.2
1) Approximate BER Expression for Each Sub-carrier :
Here, we will derive the BER performance for each sub-
1
carrier, where a maximum likehood (ML) receiver is used at
Average channel power gain

the base station as an interference cancelation detector. With

0.8
average ML receiver, the diversity order of each user in the multi-
τ=0 user case in SDMA is the same as the diversity order in
τ=0.1T
0.6 s the single-user case, and for analytical tractability, we neglect
τ=0.2T
s interference among users in the BER expression used in our
τ=0.3T
0.4 s
bit loading algorithm.
τ=0.4Ts
τ=0.5Ts
Since the channel response on each sub-carrier is a flat-
0.2
Rayleigh-fading channel with variance Cn,n , it is equivalent
to derive the BER performance for a single user using OSTBC
0
0 10 20 30 40 50 60 with Nt transmit antennas and Nr receive antennas under
Sub−carrier index Rayleigh flat-fading channels. A MIMO fading channel with
OSTBC transmission can be transformed into an equivalent
Fig. 2. Average channel power gain profile on the sub-carriers for the
equivalent channel with truncated pulse shape function and 0 < τ < T2s . Gaussian SISO channel [9]. Then, on the nth sub-carrier,
the pdf and the moment generation function (MGF) of the
1.4
instantaneous signal-noise-ratio (SNR) γn of the equivalent
SISO model are given by
γnNt Nr −1
1.2
γ
e− γ̄n (16)
Average channel power gain

pγn (γn ) = Nt Nr
1 Γ(Nt Nr )γ̄n
and
0.8 average  ∞
τ=0
0.6 τ=−0.1Ts φγn (s)  e−sγn pγn (γn )dγn = (1 + sγ̄n )−Nt Nr , (17)
0
τ=−0.2T
s
S C
0.4 τ=−0.3Ts respectively, where γ̄n  Nnt ξN
n,n
0
. The quantities Sn , ξ and
τ=−0.4Ts N0 are the total transmit power (over all transmit antennas) on
0.2 τ=−0.5T
s the nth sub-carrier, the OSTBC rate and the noise variance,
0
respectively. The authors in [9] also gave the exact symbol
0 10 20 30 40 50 60 error probability of OSTBC for flat-Rayleigh-fading channels.
Sub−carrier index
However, the expression is too complicated to use in the bit
Fig. 3. Average channel power gain profile on the sub-carriers for the loading scheme. Here, we will derive an approximation BER
equivalent channel with truncated pulse shape function and − T2s < τ < 0. expression.
A simple approximate BER expression for square MQAM
1
with Gray bit mapping in additive white Gaussian noise
B. Statistical Bit Loading Algorithm (AWGN) channels is given in [10] as
−1.6γn
In the previous subsection, the statistical channel power Pen |γn ∼
= 0.2 e M −1 . (18)
gain profile was derived. In this subsection, we will present a
statistical bit loading algorithm to fully exploit it. A bit loading Let 2kn denote the constellation size on the nth sub-carrier.
scheme, typically used in multicarrier wireline systems, as- Then, the BER on the nth sub-carrier is given by
 ∞  ∞
signs a high data rate to sub-carriers with high SNR and a low −1.6γn

data rate to sub-carriers with low SNR [7]. In wireless systems, Pen = Pen |γn pγn (γn )dγn = 0.2e 2kn −1 pγn (γn )dγn
0   0
it is difficult to obtain the instantaneous channel information 1.6 0.2
at the transmitter for fast time-varying channels and thus it is = 0.2 φγn
2kn − 1
= 1.6 Sn Cn,n Nt Nr
.
(1 + 2kn −1 Nt ξN0 )
infeasible to adopt an adaptive bit loading algorithm. However,
as shown in Fig. 1, the average channel power gain profile (19)
has a very special structure. Although we do not know the 2) Statistical Bit Loading Algorithm: Our statistical bit
instantaneous channel information, it can be seen that there are loading algorithm solves the following optimization problem:
some sub-carriers which have less average channel power gain 0.2
than the other sub-carriers. With this average power profile, we arg min Pe =  
Sn Cn,n Nt Nr
= Pen , ∀n
Rn ,Sn 1+ k1.6
propose a statistical bit loading scheme. In [8], a statistical bit ⎧ N2 n −1 Nt ξN0

⎨ (a) n=1 kn = RT
loading scheme was proposed to maximize the overall spectral N
subject to (b) n=1 Sn = ST
efficiency and minimize the transmission power requirement ⎩
while maintaining the target BER level. However, in this paper, (c) Pen is the same for all n = 1, . . . , N.
(20)
our objective is to develop a transmission scheme where all
sub-carriers have the same smallest bit error rate (BER) given 1 For analytical tractability in the bit loading algorithm, we use this
total power and total data rate constraints. expression for non-square MQAM as well.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:11 from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 10, OCTOBER 2008 3729

a) 3−tap exponential power delay profile

From (20), we know that minimizing Pe is equivalent to 10
0

maximizing with statistical bit loading

without timing errors
Sn Cn,n −2
y = kn . (21) 10 without statistical bit loading
2 −1

BER
−4
From (21), we obtain 10

y(2kn − 1) −6
Sn = . (22) 10
0 5 10 15 20 25
Cn,n Eb/No (dB)

Substituting (22) into Constraint (b) of (20), we obtain 10

0
b) 3−tap uniform power delay profile

with statistical bit loading

ST ST
y = N 2kn −1
= N 2kn
N 1
. (23) −2
without timing errors

n=1 Cn,n n=1 Cn,n − n=1 Cn,n

10 without statistical bit loading

BER
N kn −4
Thus, maximizing y is equivalent to minimizing n=1 C2n,n . 10

Performing the Lagrange optimization under Constraint (a) of −6

10
(20), we obtain 0 5 10 15 20 25
E /N (dB)
b o
N

RT 1
kn = − log2 ( (Cm,m )) + log2 (Cn,n ). (24) Fig. 4. BER performance comparison with/without statistical bit loading for
N N m=1 3-tap Rayleigh fading channels with exponential/uniform power delay profile.
We summarize the proposed statistical bit loading algorithm
in the following five steps:
IV. S IMULATION R ESULTS
• Step 1: Select active sub-carriers

|I| In our simulations, the channel statistics between different

RT 1  transmit antennas and different receive antennas are assumed
kn = − log ( (Cm,m )) + log2 (Cn,n ). (25)
|I| |I| 2 m=1 to be the same, each of which is modeled as the following
four types of channels:
Initially, I = {1, 2, ..., N } where |I| denotes the size of
the set I. If kn < 0, then remove sub-carrier n from I. • 3-tap Rayleigh fading channel with an exponential power
This is done iteratively until all rates of the remaining delay profile (the tap variance decays at 3dB/tap) and
sub-carriers are positive. tl = (l − 1)Ts .
• Step 2: kn is quantized to Qn as • 3-tap Rayleigh fading channel with a uniform power
⎧ delay profile and tl = (l − 1)Ts .
⎨ Rmax : kn >= Rmax − 0.5 • 7-tap Rayleigh fading channel with an exponential power
Qn = Floor(kn + 0.5) : 0.5 < kn < Rmax − 0.5 , delay profile (the tap variance decays at 3dB/tap) and
⎩
0 : kn < 0.5, tl = (l − 1)Ts .
(26) • 3-tap Rician fading channel with an exponential power
where Rmax is the predefined maximum signal constel- delay profile (the tap variance decays at 3dB/tap) and
lation size on a sub-carrier. tl = (l − 1)Ts . The first tap is Rice distributed, where
√
• Step 3: If n∈I Qn , < (>)RT , the rate of the channel the real part has a Gaussian distribution of mean 0.4
with the largest (smallest) Δkn = kn − Qn , is incre- and the imaginary part has a Gaussian distribution of
mented
 (decremented). The initial bit allocation stops if zero mean. Since the complicated BER expression for
n∈I Q n = RT . Rician fading channel introduces an analytical obstacle in
• Step 4: If the number of sub-carriers with Qn = 1 is odd, the statistical bit loading algorithm, we use the algorithm
to adjust it to be even, we reduce the constellation size of derived based on the Rayleigh fading channel in Section
the sub-carrier with the smallest average channel power III for the Rician fading channel as well. This will
gain among those with the largest constellation size by reflect performance sensitivity of the proposed scheme
‘1’ bit. Based on the results from Step 3, adjust the bit to channel model mismatches.
allocations on the sub-carriers with Qn = 1 such that
The equivalent channels impaired by the timing synchro-
half of these sub-carriers with the higher (lower) channel
nization errors are assumed to be perfectly known at the
power
 gain profile are assigned Qn = 2 (Qn = 0) so that
receiver2 . Note that the transmitters do not have knowledge of
n∈I Q n = RT .
the instantaneous channel gains. The difference of the arrival
• Step 5: Transmit power is assigned to the sub-carriers
delays from Users 1 and 2 are assumed to be uniformly
recorded in the set I such that all sub-carriers achieve
distributed over [−Ts , Ts ]. We consider two transmit antennas
the same error probability, i.e.,
at each user and two receive antennas at the base station. In
ST (2Qn −1) both cases, there are N = 64 sub-carriers. The ML detector
Cn,n
Sn =  , (27) is implemented on each sub-carrier to decode the source
2Ql −1
l∈I Cl,l information.
where ST is the fixed total transmit power. 2 In practice, they are obtained using channel estimation techniques.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:11 from IEEE Xplore. Restrictions apply.
 3730 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 10, OCTOBER 2008

0 −1
10 10
with statistical bit loading with statistical bit loading
without timing errors without timing errors
−1 without statistical bit loading without statistical bit loading
10 −2
10

−2
10
−3
10
BER

−3

BER
10

−4
−4
10
10

−5 −5
10 10

−6
10 −6
0 5 10 15 20 25 10
E /N (dB) 0 5 10 15 20 25
b o
E /N (dB)
b o

Fig. 5. BER performance comparison with/without statistical bit loading for

7-tap Rayleigh fading channels with exponential power delay profile. Fig. 6. BER performance comparison with/without statistical bit loading for
3-tap Rician channels with exponential power delay profile

Figs. 4, 5, and 6 present the BER performance compar-

ison of our proposed scheme and the conventional scheme If |P (F )| ≥ 0 only for |F | ≤ KFs , then
without statistical bit loading assuming two transmit antennas K
for different channel types. Also, the performance without 
H(f ) = Fs P ((f − k)Fs )G((f − k)Fs )e−j2π(f −k)Fs τ .
synchronization errors is given as a benchmark, in which k=−K
the average channel power gains on all sub-carriers are the (30)
same and no statistical bit loading scheme is needed. The Since there are N sub-carriers, the frequency response on the
results show that the proposed scheme is not sensitive to the nth sub-carrier is
channel type. Our proposed algorithm improves the robustness K
to timing synchronization errors, yielding an SNR advantage n  n n n
H( ) = Fs P (( −k)Fs )G(( −k)Fs )e−j2π( N −k)Fs τ .
of around 7.5 dB at BER = 10−4 in the channels considered. N
k=−K
N N
(31)
n
V. C ONCLUSIONS Let x(k) = P (( N n n
− k)Fs )G(( N − k)Fs )e−j2π( N −k)Fs τ .
A novel STBC-OFDM transmission scheme robust to tim- Then, the average channel power gain on the nth sub-carrier
ing synchronization errors over frequency-selective multiple- is given as
access channels is proposed in this paper. The analysis of n 2
the statistical power gain profile of the correlated equivalent Cn,n = E(|H( )| )
N
channel model shows that the average channel power gains  K−k
K−1 1
on some sub-carriers are always less than the others. A robust = Fs2 2 {E [x(k1 )x∗ (k1 + k2 )]}
statistical bit loading algorithm is proposed to exploit this k1 =−K k2 =1
(32)
average channel power profile resulting in an appreciable K

performance improvement. This algorithm is appealing to + Fs2 E(|x(k1 )|2 ).
systems where link adaption based on instantaneous channel k1 =−K
information is infeasible or undesirable. L
Let g(t) = l=1 gl δ(t − tl ). Then
A PPENDIX A  L
∞ 
P ROOF OF P ROPOSITION 1 G(F ) = g(t)e−j2πF t dt = gl e−j2πF tl . (33)
From (2), we obtain −∞ l=1

Ha (F ) = P (F )G(F )e−j2πF τ . (28) Since E{gl gl∗ } = 0 for l = l , we have

 
Let h(n) = ha (nTs ). Then, its frequency domain representa- n n
E G(( − k1 )Fs )G∗ (( − k1 − k2 )Fs )
tion is given by [11]  NL N 
 L

∞
 n n
−j2π( N −k1 )Fs tl1 ∗ j2π( N −k1 −k2 )Fs tl2
H(f ) = Fs Ha [(f − k)Fs ], (29) =E g l1 e g l2 e
l1 =1 l2 =1
k=−∞  L  L
 

where f = FFs , representing the relationship between the =E |gl |2 e−j2πk2 Fs tl = E |gl |2 e−j2πk2 Fs tl .
frequency variables F and f of the corresponding analog and l=1 l=1
discrete-time signals, respectively. (34)

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:11 from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 10, OCTOBER 2008 3731

Therefore, (i, j)-th element of P , we have

∗  L
L  L

E[x(k1 )x (k1 + k2 )] ∗
n n Rm,k = Pm,r Rg (r, l)Pk,l = Pm,l E(|gl |2 )Pk,l
∗
.
= |P (( − k1 )Fs )||P (( − k1 − k2 )Fs )|ejθ1 e−jθ2 l=1 r=1 l=1
L N N  (35) (41)
  
× E |gl |2 e−j2πk2 (τ +tl )Fs , Thus,
M
 L
 L

l=1
Rm,m = E(|gl |2 ) |Pm,l |2 . (42)
n
where θ1 is the phase of − k1 )Fs ) and θ2 is the phase
P (( N m=1 l=1 m=1
n
of P (( N − k1 − k2 )Fs ). Then, From (6), we get
L

{E[x(k1 )x∗ (k1 + k2 )]}
n n |Pm,l |2 = p20 + p21 + p22 + p23 , l = 1, 2, ..., L, (43)
= |P (( − k1 )Fs )||P (( − k1 − k2 )Fs )|× m=1
L N N  (36)

 and (42) becomes
2
E |gl | cos(2πk2 (τ + tl )Fs − θ1 + θ2 ) . L L
 
l=1 Rm,m = (p20 + p21 + p22 + p23 ) E(|gl |2 ). (44)
Substituting (36) into (32) gives (9). m=1 l=1

From (41), we get

A PPENDIX B L−p
 L
 L−p

P ROOF OF P ROPOSITION 2 (Rm,m+p ) = E(|gl |2 ) ∗
Pm,l Pm+p,l . (45)
m=1 l=1 m=1
Denote p = |k − m| ≥ 1. From (13), we have
From (6), we know that
L
 ⎧
⎪
⎪ p0 p1 + p1 p2 + p3 p0 ,p = 1
Cn,n = Rm,m L−p
 ⎨
m=1 ∗ p0 p2 + p3 p1 ,p = 2
 L−p  Pm,l Pm+p,l =
L−1 L−p ⎪
⎪ p3 p2 ,p = 3
  j2π(n−1)p  j2π(n−1)(−p) m=1 ⎩
+ Rm,m+p e N + Rm+p,m e N . 0 , otherwise
p=1 m=1 m=1
(46)
(37) .
Thus, (40) becomes (14). Similarly, we can obtain Cn,n for
∗
Since R is a Hermitian matrix, Rm+p,m = Rm,m+p . We have −Ts < τ < 0 from (8) as in (15).
L−p
 L−p

Rm,m+p e
j2π(n−1)p
N + Rm+p,m e
j2π(n−1)(−p)
N
R EFERENCES
m=1 m=1 [1] N. Al-Dhahir, “Overview and comparison of equalization schemes
L−p
  for space-time-coded signals with application to EDGE,” IEEE Trans.
2π(n − 1)p (38) Signal Processing, vol. 50, pp. 2477–2488, Oct. 2002.
= 2 (Rm,m+p ) cos( ) [2] R. L. Choi, M. T. Ivrlac, R. D. Murch, and W. Utschick, “On strategies
m=1
N of multiuser MIMO transmit signal processing,” IEEE Trans. Wireless
2π(n − 1)p Commun., vol. 3, no. 6, pp. 1936–1941, Nov. 2004.
+2
(Rm,m+p ) sin( ) [3] A. F. Naguib, N. Seshadri, and A. R. Calderbank, “Applications of
N space-time block codes and interference suppression for high capacity
and high data rate wireless systems,” in Proc. 32nd Asilomar Conf.
Thus, (37) becomes Signals, Systems, and Computers, Nov. 1998, vol. 2, pp. 1803–1810.
[4] Y. Mei, Y. Hua, A. Swami, and B. Daneshrad, “Combating synchroniza-
L
 L−1
 L−p
2π(n − 1)p  tion errors in cooperative relays,” in Proc. IEEE ICASSP, Mar. 2005,
Cn,n = Rm,m + 2 cos( ) (Rm,m+p ) vol. 3, pp. 18–23.
m=1 p=1
N m=1
[5] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, “Space-time block
coding for wireless communications: performance results,” IEEE J.
L−1
 L−p
2π(n − 1)p  Select. Areas Commun., vol. 17, no. 3, pp. 451–460, 1999.
+2 sin( )
(Rm,m+p ). [6] J. G. Proakis, Digital Communications. McGraw-Hill, 4th edition, 2001.
p=1
N m=1
[7] R. F. H. Fisher and J. B. Huber, “A new loading algorithm for discrete
multitone transmission,” in Proc. IEEE Globecom, Nov. 1996, vol. 1,
(39) pp. 724–728.
[8] Z. Song, K. Zhang, and Y. Guan, “Joint bit-loading and power-allocation
If all the elements of R are real, e.g., a channel with for OFDM systems based on statistical frequency-domain fading model,”
uncorrelated zero-mean channel taps, then (39) becomes in Proc. IEEE VTC, Fall 2002, vol. 2, pp. 724–728.
[9] H. Shin and J. H. Lee, “Exact symbol error probability of orthogonal
L
 L−1
 L−p space-time block codes,” in Proc. IEEE Global Telecomm. Conf., Nov.
2π(n − 1)p  2002, vol. 2, pp. 1197–1201.
Cn,n = Rm,m +2 cos( ) (Rm,m+p ).
N [10] S. T. Chung and A. J. Goldsmith, “Degrees of freedom in adaptive
m=1 p=1 m=1 modulation: a unified view,” IEEE Trans. Commun., vol. 49, pp. 1561–
(40) 1571, Sept. 2001.
Since R = P Rg P H and Rg = [11] J. G. Proakis and D. G. Manolakis, Digital Signal Processing: Princi-
diag{E(|g1 |2 ), E(|g2 |2 ), ..., E(|gL |2 )} with Pi,j denoting the ples, Algorithms, and Applications. Prentice Hall, 3rd edition, 1996.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 10:11 from IEEE Xplore. Restrictions apply.