IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO.
3, MAY 2008
systems, in this paper, we consider the NGMN to limit its dimen-
sioning process to the core network. A dimensioning framework is
proposed that derives a bounded initial estimate of core network
elements (the number of BSs/APs per AR, the number of ARs per
MAP, and the number of MAPs per GW). The proposed algorithm
facilitates NGMN deployment by enabling the providers/operators to
select the core network estimate in accordance with their preferred cell
loading bounds, future traffic projection, and data rate variability. The
cell loading bounds are known a priori, whereas the upper limit on
variable data rates assumes pedestrian and vehicular terminal mobility.
To accommodate the increased number of users and the bandwidth-
intensive applications of the future, fiber optic links with high data
transmission rate are considered in the core network.
R EFERENCES
[1] K. Knightson, N. Morita, and T. Towle, NGN architecture: Generic
principles, functional architecture, and implementation, IEEE Commun.
Mag., vol. 43, no. 10, pp. 4956, Oct. 2005.
[2] A. R. Rofougaran, M. Rofougaran, and M. Behzad, Radios for
next-generation wireless networks, IEEE Microw. Mag., vol. 6, no. 1,
pp. 3843, Mar. 2005.
[3] K. Zheng, L. Huang, W. Wang, and G. Yang, TD-CDM-OFDM: Evo-
lution of TD-SCDMA toward 4G, IEEE Commun. Mag., vol. 43, no. 1,
pp. 4552, Jan. 2005.
[4] D. Garg and F. Adachi, DS-CDMA with frequency-domain equalization
for high speed downlink packet access, in Proc. IEEE VTC, Sep. 2004,
vol. 1, pp. 689693.
[5] A. Jamalipour, T. Wada, and T. Yamazato, A tutorial on multiple access
technologies for beyond 3G mobile networks, IEEE Commun. Mag.,
vol. 43, no. 2, pp. 110117, Feb. 2005.
[6] C. N. Konstantinopoulou, K. A. Koutsopoulos, G. L. Lyberopoulos,
and M. E. Theologou, Core network planning, optimization and fore-
casting in GSM/GPRS networks, in Proc. IEEE SCVT, Oct. 2000,
pp. 5561.
[7] B. Jaumard, C. Meyer, R. Pooyania, Y. Solari, and C. Voisin, Causal
and anticipative models for the dimensioning of 3G multiservice
networks, in Proc. IEEE GLOBECOM Conf., Dec. 2003, vol. 6,
pp. 34173422.
[8] G. Giambene and C. Tommasi, 3G wireless core network dimensioning
in the presence of self-similar traffics, in Proc. IEEE Int. Symp. Wireless
Commun. Syst., Sep. 2005, pp. 223227.
[9] B. C. Sowden and K. W. Sowerby, Variable-bit-rate traffic modelling for
dimensioning 3G CDMA systems, in Proc. IEEE VTC, Apr. 2003, vol. 1,
pp. 670674.
[10] A. Jamalipour, V. Mirchandani, and M. R. Kibria, Dimensioning of an
enhanced 4G/B3G infrastructure for voice traffic, in Proc. IEEE PIMRC,
Berlin, Germany, Sep. 2005, vol. 3, pp. 20032007.
[11] M. R. Kibria, V. Mirchandani, and A. Jamalipour, A consolidated archi-
tecture for 4G/B3G networks, in Proc. IEEE WCNC, Mar. 2005, vol. 4,
pp. 24062411.
[12] General Overview of NGN, ITU-T Rec. Y.2001, 2004.
[13] J. Zhang, L. Guo, and J. Y. Wu, An integrated approach for UTRAN
planning and optimization, in Proc. IEEE VTC, May 2004, vol. 4,
pp. 23602364.
[14] V. K. Garg, IS-95 CDMA and CDMA2000. Englewood Cliffs, NJ:
Prentice-Hall PTR, 2000.
[15] Digital HierarchyOptical Interface Rates and Formats Specication
(SONET), ANSI T1.105-1991, 1991.
[16] R. T. Fernandez and J. E. P. Velasco, From GSM toward UMTS: A
radio interface for multimedia services transmission, in Proc. IEEE Int.
Telecommun. Symp., 1998, vol. 2, pp. 388393.
[17] D. Lister, S. Dehghan, R. Owen, and P. Jones, UMTS capacity and
planning issues, in Proc. Int. Conf. 3G Mobile Commun. Technol.,
Mar. 2000, pp. 218223.
[18] H. B. Pinto, J. G. da Silva, and A. Rodrigues, Uplink capacity of the
WCDMA FDD mode in UMTS networks with mixed services, in Proc.
IEEE VTC, 2000, vol. 6, pp. 26172624.
[19] Y. C. Tay and K. C. Chua, A capacity analysis for the IEEE 802.11 MAC
protocol, Wirel. Netw., vol. 7, no. 2, pp. 159171, Mar. 2001.
[20] RF System Scenarios, 3GPP TS RAN 25.942, 1999.
0018-9545/$25.00 2008 IEEE
1963
MAP-Based Channel Estimation for MIMOOFDM
Over Fast Rayleigh Fading Channels
Jin-Goog Kim and Jong-Tae Lim
AbstractThis paper presents a channel estimation scheme for
multiple-input multiple-output with orthogonal frequency-division mul-
tiplexing (MIMO-OFDM) in fast varying Rayleigh channels. To handle
rapid variation of channels within a transmission block, we propose a
novel maximum a posteriori probability-based channel estimation scheme
using pilot symbols. With the estimate of the channel matrix for the
current symbol interval, a zero-forcing (ZF) receiver is applied to detect
the spatially multiplexed data. In simulation results, the effectiveness of
the proposed method is shown, as compared with the polynomial and the
perfect channel estimates.
Index TermsFast varying Rayleigh channels, maximum a posteriori
probability (MAP)-based channel estimation, multiple-input multiple-
output with orthogonal frequency-division multiplexing (MIMOOFDM).
I. I NTRODUCTION
In recent years, multiple-inputmultiple-output (MIMO) antennas
combined with orthogonal frequency-division multiplexing (OFDM)
have been widely studied in wireless communications because they
can provide high data rates and are robust to multipath delay. Channel
parameters are needed to coherently decode the transmitted signal
and to combine the diversity branches. Channel estimation has been
extensively studied for single-antenna systems [3], [4], [7], [8]. For
MIMOOFDM systems, most of the channel estimation schemes have
focused on pilot-assisted approaches [3], [8] based on a quasi-static
fading model that allows the channel to be constant for a block of
OFDM symbols. However, in fast fading channels, the time variation
of a fading channel over an OFDM symbol period results in a loss of
subchannel orthogonality, which leads to intercarrier interference (ICI)
[12][14]. To support high-speed mobile channels, the time variation
of a fading channel over an OFDM block must be considered. There
are more channel parameters in fast fading channels than in quasi-static
fading channels. Hence, if we use a channel estimation method that
only estimates a few coefficients corresponding to different multipath
delays, then we can improve the estimation performance. In [1] and
[4], a channel estimation algorithm based on a Qth-order polynomial
function has been proposed, but the Qth-order polynomial approxima-
tion does not reflect the channel characteristic.
In this paper, we derive a novel pilot-symbol-aided maximum
a posteriori probability (MAP)-based channel estimation scheme for
a MIMOOFDM system over fast Rayleigh fading channels. Since
the proposed method uses the covariance matrix of the channel, the
channel characteristic is considered by the estimation scheme. Then, a
zero-forcing (ZF) receiver is applied to detect the spatially multiplexed
data transmitted in the current symbol interval.
Manuscript received September 27, 2006; revised April 27, 2007, June 23,
2007, and August 9, 2007. This work was supported by the Ministry of Science
and Technology through Korea Aerospace Research Institute. The review of
this paper was coordinated by Prof. J. Choi.
The authors are with the Department of Electrical Engineering and Com-
puter Science, Korea Advanced Institute of Science and Technology, Daejeon,
305 701, Korea (e-mail: genuinekjg@kaist.ac.kr; jtlim@stcon.kaist.ac.kr).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2007.907315
1964 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 3, MAY 2008

II. S YSTEM M ODEL
Consider a MIMOOFDM system with NT transmit antennas,
NR receive antennas, and N subcarriers, which employs quadra-
ture amplitude modulation (QAM). The transmitted symbols are
denoted by X = [X(0)T , X(1)T , . . . , X(N 1)T ]T with X(k) =
[x1 (k), x2 (k), . . . , xNT (k)]T , where xi (k) is the transmitted signal
It is observed that the channel matrix H in (3) is no longer a block cir-
culant matrix as in the case of slowly fading channels. Consequently,
G = FNR HFH NT is no longer a block diagonal matrix. This shows
that time-selective fading causes ICI, which is represented by the off-
diagonal blocks of G. Note that (1) can be rewritten as

N 1

by the ith transmit antenna on subcarrier k, and the received sym-
bols are denoted by Y = [Y(0)T , Y(1)T , . . . , Y(N 1)T ]T with
Y(k) = [y1 (k), y2 (k), . . . , yNR (k)]T , where yj (k) is the received
signal by the jth receive antenna on subcarrier k. The transmitted
symbols are fed to an inverse discrete Fourier transform (IDFT) to
produce the OFDM signal, and a guard interval is inserted, which
is a cyclic extension of the IDFT output sequence to eliminate the
intersymbol interference (ISI). The guard interval is chosen to be
longer than the maximum delay spread of the channel. This signal
is transmitted over the multipath channel, and the received samples
are demodulated by taking the discrete Fourier transform (DFT) after
removal of the guard interval.
As shown in [2] and [7], it is well known that the general form
of MIMOOFDM over slowly fading channels and time-invariant
channels over several OFDM symbol periods can be expressed as
Y(p) = G(p, p)X(p) + G(p, q)X(q) + Z(p). (6)
q=0
q
=p
In a time-selective channel, the first term in the right-hand side of (6) is
the desired term without ICI in the frequency domain, which denotes
the contribution from the same symbol, and the second term in the
right-hand side of (6) is the ICI term in the frequency domain, which
denotes the contribution from other symbols.
The fading channel process h(t) is modeled as a normalized zero-
mean complex wide-sense stationary Gaussian process with a correla-
tion function rh (t) = E{h(t)h (t + t)} [6], where E{} denotes
the statistical mean, and () represents the complex conjugate of
(). As shown in [5] and [15], the spacetime correlation function of
the channel in a typical mobile communication environment can be
modeled as

Y(k) = (k)X(k) + Z(k) (1) rh (t) = Jo (2fd t) (7)

where X(k) and Y(k) represent the transmitted and received signals, where fd is the maximum Doppler shift, and J0 () is the zeroth-order
respectively, for all antennas on subcarrier k; (k) is an NR NT Bessel function of the first kind.
matrix with {(k)}ij as the channel frequency response between
transmit antenna j and receive antenna i; and Z(k) is an NR 1 III. C HANNEL E STIMATION
vector denoting the zero-mean additive white Gaussian noise (AWGN)
with variance z2 INR for all antennas on subcarrier k. In this section, we consider the problem of channel estimation
In fast fading channels, since the channel within an OFDM block in a MIMOOFDM system over time-selective channels. To esti-
is time selective, MIMOOFDM systems are no longer of the form in mate parameters, we change the equation Y = GX + Z to the form
F is the standard N -dimensional DFT matrix defined by
(1). Note that of Y = Wh + Z. Let h = [hT (0), . . . , hT (N 1)]T and h(n) =
Fp,q = (1/ N )ej2pq/N for 0 p, q N 1, and Fn = F In , [vec(h(n, 0))T , . . . , vec(h(n, L 1))T ]T , where vec(x) is the vector
where is the Kronecker product. Now, the overall system can be obtained by stacking the columns of x on top of each other. Here,
expressed in matrix notation as h(n, l) is an NR NT matrix, and the entry of h(n, l) is defined by

Y = FNR HFH h1,1 (n, l) h1,NT (n, l)
NT X + Z (2)
h(n, l) = ..
.
..
.
..
.

where Z CN NR 1 is the AWGN vector, and ()H denotes the
hNR ,1 (n, l) hNR ,NT (n, l)
complex conjugate transpose of (). The entry of the channel matrix
0 n N 1; 0 l L 1. (8)
H is defined by [2]
The pth received symbol at the rth receive antenna, i.e., Yr (p), is
H0 (0) HL1 (0) H1 (0)
expressed as follows:
H= .
. . . .
.. (3)
. .
1   
N 1 L1 NT N 1
0 HL1 (N 1) H0 (N 1)
Yr (p) = hr,t (n, l)xt (q)D(p, q, n, l)
N
where L is the number of resolvable paths, and 0 is an NR NT n=0 l=0 t=1 q=0
zero matrix. Each nonzero block of H contains the NR NT chan- + Zr (p)
nel matrix Hl (n) for path l at time nTs (Ts is the data symbol  
N 1 L1 NT
period), and the (p, q)th entry of Hl (n) is hp,q (n, l). Let G = = W (p, n, l, t)hr,t (n, l) + Zr (p)
FNR HFH NT . Then n=0 l=0 t=1
N1
G(0, 0) G(0, N 1)
where W(p, n, l, t) = (1/N ) q=0 xt (q)D(p, q, n, l), and D(p, q, n, l) =
G= ..
.
..
.
..
.

ej(2/N )(pq)n ej(2/N )lq .
G(N 1, 0) G(N 1, N 1) Letting a set of NP pilot tones be P = {P (1), . . . , P (NP )},
Y = GX + Z (4) W (p, n, l, t) can be split into two matrices as follows:
1 
where G(p, q) is an NR NT matrix, and the (r, t)th entry gr,t (p, q) W (p, n, l, t) = xt (q)D(p, q, n, l)
N
of matrix G(p, q) is as follows: qP
1 
+ xt (q)D(p, q, n, l)
1 
N 1 L1
N
gr,t (p, q) = hr,t (n, l)ej2n(pq)/N ej2ql/N . (5) qP
N P D
n=0 l=0 =: W(p,n,l,t) + W(p,n,l,t) . (9)
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 3, MAY 2008 1965

Thus, we obtain The covariance matrix of the zero-mean vector vr,t (n, l) can be
written as
Y(p) = WpP h + WpD h + Z(p) (10)  
H

Rh11 rh12
Rh = E vr,t (n, l)vr,t (n, l) = (15)
with WpP = [[W(p,0,0,1)
P P
INR ], . . . , [W(p,N rh21 rh22
1,L1,NT ) INR ]],
D D D
and Wp = [[W(p,0,0,1) INR ], . . . , [W(p,N 1,L1,NT ) INR ]].
where
Since we do not know xt (q) for q  P, WpD has unknown elements.

Rh11 = E ur,t (n, l)uH

r,t (n, l) (16)

A. Channel Estimation Scheme rh12 = E ur,t (n, l)hH

r,t (n, l) (17)

When we consider a system where ICI is not negligible, WpD rh21 = E hr,t (n, l)uH H
r,t (n, l) = rh12 (18)
H

cannot be simply removed. Hence, we should account for the ICI in rh22 = E hr,t (n, l)hr,t (n, l) . (19)
the channel estimation process, as shown in [16] and [17]. For the
pilot-based channel estimation scheme over fast fading channels, the If we know ur,t (n, l), the estimate of hr,t (n, l) can be obtained by
unknown WpD contributes to the approximation error. By defining maximizing its conditional probability density function p[hr,t (n, l)|
e(p) := WpD h + Z(p), (10) can be rewritten as ur,t (n, l)] as

Y(p) = WpP h + e(p) hr,t (n, l) := max p [hr,t (n, l)|ur,t (n, l)] . (20)
hr,t (n,l)

where e(p) is composed of the ICI contribution from nonpilot tones For the Rayleigh channels being considered, the (M + 1) 1 vector
and the noise contribution. Let Y = [Y(P (1)), . . . , Y(P (NP ))]T vr,t (n, l) is a complex Gaussian. As shown in [10], we write the
P T P T T
and W = [[WP (1) ] , . . . , [WP (NP ) ] ] . Then, we can form the conditional probability density function as
NP NR N LNT NR system of linear equations as
p [hr,t (n, l)|ur,t (n, l)]
Y = Wh + e (11) p [ur,t (n, l)hr,t (n, l)]
=
p [ur,t (n, l)]
where Y, e, and W are defined as shown at the bottom of the page.  
1
exp vr,t
H
(n, l)R1
h vr,t (n, l)
However, W is an NP NR N LNT NR matrix, and NP =
M +1 |Rh |
  (21)
1
N LNT for each receive antenna. To obtain h as the least square 1
M |rh11 |
exp uH
r,t (n, l)rh11 ur,t (n, l)
solution, the number of rows of W has to be larger than the number
of columns of W. Thus, we need to reduce the number of channel where |A| denotes the determinant of matrix A. By using the matrix
parameters. To reduce the number of parameters needed for channel inversion lemma, R1
h is expressed as
estimation, we use the sliding-window approach in which hr,t (n, l)  1
is derived from the previous channel parameters hr,t (m, l), m = n Rh11 rh12
R1
h =
M, . . . , n 1, where M is the window size, and M Np /LNT , rh21 rh22
where A/B represents the integer part of A/B. That is, we construct  1
1 1

Rh11 rh12 rh22 rh21 Rhi11 rh12 rh22
hr,t (n, l) as follows: = 1
rhi22 rh21 R1
h11 rh22 rh21 R1
h11 rh12
hr,t (n, l) = a1 hr,t (n M, l) + + aM hr,t (n 1, l) (22)
= a [hr,t (n M, l), . . . , hr,t (n 1, l)] .
T T
(12) where Rhi11 = (Rh11 rh12 rh22 1
rh21 )1 , and rhi22 = (rh22
rh21 R1 r
h11 h12 )1
.
Note that we have used the same M 1 tap weight vector a for all of Maximizing the conditional probability density function is equiva-
the L channel taps of hr,t (n, l). lent to minimizing the following objective function [9]:
Let us define an M 1 vector ur,t (n, l) and (M + 1) 1 vector
vr,t (n, l) as H
f = vr,t (n, l)R1 H 1
h vr,t (n, l) ur,t (n, l)Rh11 ur,t (n, l). (23)

ur,t (n, l) = [hr,t (n M, l), . . . , hr,t (n 1, l)]T (13) By letting the conjugate derivative of f with respect to hr,t (n, l) be
  equal to zero, we obtain the estimate of hr,t (n, l) as
ur,t (n, l)
vr,t (n, l) = . (14)
hr,t (n, l) hr,t (n, l) = aT ur,t (n, l) (24)

Y (P (1)) e (P (1))
Y = .. , e = ..
. .
Y (P (NP )) e (P (NP ))
P P

W(P (1),0,0,1) INR W(P (1),N 1,L1,NT ) INR
.. .. ..
W = . . .
P P
W(P (NP ),0,0,1) INR W(P (NP ),N 1,L1,NT ) I NR
1966 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 3, MAY 2008

where aT = rh21 R1 h11 is the M 1 tap weight vector, and rh21 and
Rh11 are obtained using (7). Since all elements of H are identi-
cally distributed, tap weight vector a is common for all coefficients
[14]. Then, we can reduce the N LNT NR channel parameters to the
M LNT NR channel parameters by using the tap weight vector.
Consider a linear system of N equations using (12) as follows:

hr,t (0, l) = a1 hr,t (s1 , l)+ +aM hr,t (sM , l)

hr,t (1, l) = a1 hr,t (s2 , l)+ +aM hr,t (0, l)
..
.
hr,t (N 2, l) = a1 hr,t (N 2M, l)+ +aM hr,t (N 3, l)
hr,t (N 1, l) = a1 hr,t (N 1M, l)+ +aM hr,t (N 2, l).

We reconstruct the linear system as follows:

hr,t (0, l) hr,t (s1 , l)
hr,t (1, l) hr,t (s2 , l)
=Q Fig. 1. BER performance of channel estimation schemes (100-Hz Doppler
.. ..
. . frequency).
hr,t (N 1, l) hr,t (sM , l)
detection, the minimum mean square error detection, the ZF detection,
= Qhr,t (l) (25) and the Bell Laboratories layered spacetime architecture scheme. For
where the lowest computational complexity, the ZF scheme will be adopted
a in this paper for symbol detection.
1 a2 aM In the ZF scheme [11], the transmitted symbol vector is written as
1 0 a1 aM 1
1 0 0
.. .. X = G Y.
aM 1 0 ..
.
(28)
a a
. .

Q= M 1 M 0 0 0 a1
.. .. . We will further discuss the pilot placement problem in the next section.
.
0 0 0
. ..
0 aM 1 .. ..
.

. . IV. S IMULATION
0 0 0
In the simulation, we consider a system with a 2 2 antenna
Note that matrix A is defined as A = Q IL INT INR , and structure. To construct an OFDM signal, assume that the entire channel
vector h is defined as h = vec([h(0), . . . , h(L 1)]T ), where h(l) = bandwidth, i.e., 800 kHz, is divided into 128 tones. The symbol
[h1 (l), . . . , hNT (l)], and ht (l) = [h1,t (l), . . . , hNR ,t (l)]. Using our duration is 160 s, and an additional 40-s guard interval is used
proposed method, we conclude that h = Ah, and we need to estimate to avoid ISI due to channel multipath delay spread. Thus, the total
M LNT NR parameters. Then, we rewrite the system model in (11) as block length T = 200 s, and the subchannel symbol rate rb = 5 kBd.
follows: For its simplicity with coherent demodulation, quadrature phase-
shift keying modulation is chosen to demonstrate the performance of
Y = Wh + e = WAh + e (26) channel estimation scheme, and bit error rates (BERs) are computed
over the symbol duration. To gain the average behavior of the chan-
where WA has a full column rank. The estimated channel vector is
= A(WA) Y, where () denotes the pseudoinverse of (). nel estimator, we have averaged the performance over 1000 OFDM
h = Ah
blocks. The pilot symbols are inserted in the system with NP = 16.
Fading channels among all transmit and receive antennas are assumed
B. Pilot Tone Selection and Detection to be independent. Then, we consider channels with fd = 100 Hz
Pilot tone placement is very important for the performance of (fd T = 2 102 ) and fd = 200 Hz (fd T = 4 102 ), where fd
channel estimation. For a time-invariant frequency-selective channel, is the Doppler frequency, and the channels corresponding to different
pilot tones should minimize the effects of frequency selectivity and receivers have the same statistics. In Figs. 15, Q denotes the polyno-
be equispaced on the fast Fourier transform (FFT) grid. However, for mial order, and M denotes the window size.
a time-varying frequency-selective channel, both frequency and time Figs. 1 and 2 show the BER performance of channel estimation
selectivity should be considered. For a time-selective channel, pilot schemes using pilot set 4 of Table I. For comparison purposes, the
tones should be placed as close as possible. From the aforementioned error rate curve with the perfect channel estimate is also shown. In
results, in a time-invariant frequency-selective channel and a time- Figs. 1 and 2, we show that the proposed method with M = 1, 2
varying channel, we choose the grouped and equispaced pilot tones has better BER performance than the polynomial approximation with
on the FFT grid. Q = 0, 1 at low Eb /N0 , respectively. In both schemes, the error
The estimation error by the least square is performance significantly improves when the window size M = 3 and
the polynomial order Q = 2, whereas the computational complexity
 2
Ee = I (WA)(WA)  Y 2 . (27) increases as O(n3 ), where n = (Q + 1)LNT NR for the polynomial
approximation and n = M LNT NR for the proposed method. The
Given the channel parameters, the transmitted symbols can be computational complexity is critical for the high transmission system,
detected using several algorithms, such as the maximum-likelihood and the high computational burden is undesirable, as shown in [1],
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 3, MAY 2008 1967

Fig. 2. BER performance of channel estimation schemes (200-Hz Doppler

frequency).

TABLE I
PILOT SETS

[4], [16], and [17]. Hence, Q = 2 and M = 3 might not be practical

choices due to the intensive computational burden. Then, the proposed
method shows better performance than the polynomial approximation
at low Eb /N0 under the constraint of low computational complexity.
In Fig. 3, we show the BER performance of estimation schemes with
various Doppler frequencies using pilot set 4 of Table I. Since it uses
the result of the least square estimation, it is sensitive to the Doppler
frequency. However, for all Doppler frequency ranges, the proposed
scheme has the lower error rate.
In Figs. 4 and 5, we show the performance for the pilot placement
schemes. We consider four pilot sets. The first pilot set is the equi-
Fig. 3. BER performance for various Doppler frequencies. (a) Various
spaced pilot tones, and the second pilot set is the grouped pilot tones. Eb /N0 . (b) Identical Eb /N0 .
The third and fourth pilot sets are the grouped and equispaced pilot
tones. For a time-invariant frequency-selective channel, the pilot tones
should be equispaced on the FFT grid. However, for a time-varying
channel, the pilot tones should be placed as close as possible. Hence,
it seems that the grouped and equispaced pilot tones should be chosen.
These pilot sets are described in Table I. As expected, the grouped and
equispaced pilot sets result in good performance. Between pilot set 3
and pilot set 4, pilot set 4 shows better performance, which results from
the ability of estimating the time selectivity of the channel.

V. C ONCLUSION
In this paper, we have proposed MAP-based channel estima-
tion using pilot symbols over fast Rayleigh fading channels in a
MIMOOFDM system. We used the sliding-window approach to
reduce the number of channel parameters needed for channel estima-
tion within an OFDM block. Considering the computational burden,
the proposed estimation scheme showed better performance than the
polynomial approximation. Through simulations, we showed that our
proposed scheme has a lower error rate at low Eb /N0 than the scheme
based on the polynomial approximation, under the constraint of low
computational complexity. Fig. 4. Effect of the pilot placement with M = 2 (100 Hz Doppler).
1968
Fig. 5. Effect of the pilot placement with M = 2 (200-Hz Doppler
frequency).
R EFERENCES
[1] K. W. Park and Y. S. Cho, An MIMOOFDM technique for high-
speed mobile channels, IEEE Commun. Lett., vol. 9, no. 7, pp. 604606,
Jul. 2005.
[2] Y. Zhang and H. Liu, Frequency-domain correlative coding for
MIMOOFDM systems over fast fading channels, IEEE Commun. Lett.,
vol. 10, no. 5, pp. 347349, May 2006.
[3] Y. Li, Pilot-symbol-aided channel estimation for OFDM in wireless sys-
tems, IEEE Trans. Veh. Technol., vol. 49, no. 4, pp. 12071215, Jul. 2000.
[4] D. K. Borah and B. D. Hart, Frequency-selective fading channel es-
timation with a polynomial time-varying channel model, IEEE Trans.
Commun., vol. 47, no. 6, pp. 862873, Jun. 1999.
[5] W. C. Jakes, Microwave Mobile Communications. New York: Wiley,
1974.
[6] J. Du and Y. Li, MIMOOFDM channel estimation based on subspace
tracking, in Proc. IEEE VTC, 2003, pp. 10841088.
[7] Y. Li, Simplified channel estimation for OFDM systems with mul-
tiple transmit antennas, IEEE Trans. Wireless Commun., vol. 1, no. 1,
pp. 6775, Jan. 2002.
[8] 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.
[9] 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.
[10] A. Leon-Garcia, Probability and Random Processes for Electrical
Engineering. Reading, MA: Addison-Wesley, 1993.
[11] D. A. Gore, R. W. Heath, Jr., and A. J. Paulraj, Transmit selection
in spatial multiplexing systems, IEEE Commun. Lett., vol. 6, no. 11,
pp. 491493, Nov. 2002.
[12] J. Li and M. Kavehrad, Effects of time selective multipath fading on
OFDM systems for broadband mobile applications, IEEE Commun.
Lett., vol. 3, no. 12, pp. 332334, Dec. 1999.
[13] G. L. Stuber, J. R. Barry, S. W. Mclaughlin, Y. Li, M. A. Ingram, and
T. G. Pratt, Broadband MIMOOFDM wireless communications, Proc.
IEEE, vol. 92, no. 2, pp. 271294, Feb. 2004.
[14] Y. Zhang and H. Liu, Impact of time-selective fading on the performance
of quasi-orthogonal spacetime coded OFDM systems, IEEE Trans.
Commun., vol. 54, no. 2, pp. 251260, Feb. 2006.
[15] Y. Li, L. J. Cimini, 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.
[16] A. Stamoulis, S. N. Diggavi, and N. Al-Dhahir, Intercarrier interfer-
ence in MIMO OFDM, IEEE Trans. Signal Process., vol. 50, no. 10,
pp. 24512464, Oct. 2002.
[17] S. Chen and T. Yao, Intercarrier interference suppression and channel
estimation for OFDM systems in time-varying frequency-selective fading
channels, IEEE Trans. Consum. Electron., vol. 50, no. 2, pp. 429435,
May 2004.
0018-9545/$25.00 2008 IEEE
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 3, MAY 2008
On the Transmitter-Based Preprocessing for 2-D
OFDM-CDMA Forward-Link Systems
Over Time-Varying Rayleigh Fading Channels
Chih-Cheng Kuo, Student Member, IEEE,
Wern-Ho Sheen, Member, IEEE, Chung-Ju Chang, Fellow, IEEE,
and Chang Lung Hsiao, Member, IEEE
AbstractTransmitter-based preprocessing is investigated for 2-D or-
thogonal frequency-division multiplexing code-division multiple-access
(OFDM-CDMA) forward-link systems for improving performance and
shifting signal processing complexity from a mobile unit to a base station.
Preprocessing schemes that are based on zero forcing (ZF) with power
normalization, minimum mean square error (MMSE), and ZF with mul-
tiuser water filling (ZF-MWF) criteria are jointly investigated with 2-D
spreading pattern optimization and multiuser scheduling from an informa-
tion-theoretic viewpoint. Numerical results show that 1) the performance
of preprocessing is quite sensitive to the 2-D spreading pattern for SNRs of
interest, for example, 20% degradation on the sum data rate is observed
for MMSE preprocessing if the spreading pattern is not properly selected;
2) ZF-MWF may substantially outperform the other two criteria de-
pending on the SNRs; and 3) multiuser scheduling provides a significant
performance improvement on the system sum data rate.
Index TermsForward-link systems, sum data rate, transmitter pre-
processing, two-dimensional (2-D) orthogonal frequency-division multi-
plexing code-division multiple access (OFDM-CDMA).
I. I NTRODUCTION
Orthogonal frequency-division multiplexing code-division multiple
access (OFDM-CDMA) is a promising radio access technology for the
next-generation mobile communication systems thanks to its ability
to overcome intersymbol interference that is incurred in high-data-
rate transmission, its ability to provide universal frequency reuse in a
multicell environment, and its ability to achieve high-order diversity
gain by spreading data over frequency and time domains [1][5].
Traditionally, OFDM and CDMA are combined in a 1-D fashion, that
is, a data symbol is spread either in frequency or time domain (see
[1] and references therein). Recently, 2-D OFDM-CDMA, where data
are spread over time and frequency domains, has been proposed to
improve the performance of the 1-D one by simultaneously exploiting
the temporal and spectral characteristics of the fading channels [2][5].
Transmitter-based preprocessing, on the other hand, has been pro-
posed for improving performance and shifting signal processing com-
plexity from a mobile unit to a base station in mobile communication
systems [6][13]. Zero forcing (ZF) [6], [7] and prerake [8], [9]
preprocessing methods with a different degree of receiver complexity
were proposed for direct-sequence CDMA (DS-CDMA) systems. In
addition, transmitter preprocessing schemes that are based on the
MMSE [11], [13] or ZF [10][12] were investigated for broadcast
multiple-inputmultiple-output (MIMO) systems.
This paper aims to design 2-D OFDM-CDMA forward-link systems
with transmitter preprocessing to improve system performance. Three
preprocessing methods, including ZF with power normalization (ZF-
PN), the MMSE, and ZF with multiuser water filling (ZF-MWF),
Manuscript received August 7, 2006; revised June 11, 2007, July 10, 2007,
and August 13, 2007. The review of this paper was coordinated by Prof. J. Choi.
The authors are with the Department of Communication Engineering,
National Chiao Tung University, Hsinchu 300, Taiwan, R.O.C. (e-mail: cckuo.
cm90g@nctu.edu.tw; whsheen@cm.nctu.edu.tw; cjchang@cc.nctu.edu.tw;
clhsiao@itri.org.tw).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2007.909256