You are on page 1of 5

An Iterative Channel Estimator for Fast-Varying

Channels Using Successive OFDM Symbols



Youssef El Hajj Shehadeh
(1)(2)
and Serdar Sezginer
(1)

Sequans Communications, Paris, France
(1)

TELECOM PARISTECH/University of Paris VI, Paris, France
(2)

{yelhaaj,serdar}@sequans.com


AbstractThis paper deals with pilot-based channel estimation
for fast varying channels in Orthogonal Frequency Division
Multiplexing (OFDM) systems. Due to the variation of the
channel during one OFDM symbol, one-tap channel estimation
and the corresponding equalization is no longer the optimum
solution and such a variation results in inter-carrier interference
(ICI). One of the possible approaches to predict this interfering
effect is to use basis expansion model (BEM) with which the
variation of the channel can be approximated successfully.
However, as the pure BEM does not solve completely the
problem, we investigate the estimation problem using decision
feedback to enhance the performance. In particular, we propose
a simple algorithm based on using two successive OFDM symbols
to filter channel coefficients and improve not only the
convergence of pure decision feedback based estimation but also
the system performance. Simulation results based on Jakes
channel model with a high Doppler spread and a practical high
data rate system sustain our claims.
Index Termsbasis expansion model (BEM), channel
estimation, decision feedback, fast-variant channel, OFDM.
I. INTRODUCTION
Orthogonal frequency-division multiplexing (OFDM) is
one of the major transmission techniques used to achieve high
data rates over wireless mobile channels. Indeed, OFDM based
multiple access techniques have been included in many
wireless communications system specifications such as IEEE
802.11 standard for local area networks, IEEE 802.16 for
broadband wireless access systems, and 3GPP E-UTRA for
mobile wireless access systems.
In OFDM, the available channel bandwidth is divided into
N overlapping narrow banded subchannels. The serial high-rate
data stream is thus converted into N parallel low-rate
substreams, which are modulated onto the N orthogonal
subcarriers. A cyclic prefix (CP) is inserted before each
transmitted data block. If the length of the cyclic prefix is equal
to, or longer than, the delay spread of the channel, inter symbol
interference (ISI) is completely eliminated. For time-invariant
fading channels, the channel matrix in the frequency domain
would be diagonal such that each subcarrier is simply
attenuated by the corresponding frequency-domain channel
response. In this case, a simple pilot-based channel estimator
using interpolation may be adopted to estimate the channel in
the frequency domain [2]. However, for rapidly varying
channels, the variation of the channel within one OFDM
symbol destroys the orthogonality between the subcarriers and
this introduces inter-carrier interference (ICI). In this case, an
estimation of the channel in the frequency domain becomes
complex due to the large number of unknowns in the channel
matrix. For this reason, it is more appropriate to estimate the
variation of the channel in the time domain [3]. Basis
expansion Modeling (BEM) is a way to approximate the time-
variation of the channel within a certain time window.
Basically, BEM reduces the complexity as the problem is
reduced to estimating the basis coefficients [3]. It has been
recently adopted to estimate the variation of the channel in
OFDM systems and many BEMs have been proposed in this
context. The optimal one in the Mean Squared Error (MSE)
sense is the Discrete Karhuen-Love BEM (DKL-BEM) [4, 5]
as it takes into account the channel statistics to find the best
fitting basis functions. But it is suboptimal in case where the
real channel statistics deviate from the assumed ones. Other
suboptimal approaches, not depending on channel statistics, are
the Complex Exponential BEM (CE-BEM) [7] which leads to a
strictly banded frequency-domain matrix, the Generalized CE-
BEM (GCE-BEM) which is a set of oversampled complex
exponentials [8], and the Polynomial BEM (P-BEM) [9] which
approximates the channel variation by means of a polynomial
function. In all these methods, the number of basis functions
depends on the normalized Doppler frequency [3].
In this paper, we will focus on the BEM approach to
estimate the channel in high-rate OFDM systems based on an
equispaced pilot distribution. In particular, the duration of an
OFDM symbol in such systems is relatively short leading to a
very small normalized Doppler frequency. In this case, a linear
approximation (P-BEM with 2 basis functions) of the channel
variation shows an interesting fitting [12]. Based on this
property, we will propose an algorithm using two consecutive
OFDM symbols. In fact, the relation of channel variation
between two OFDM symbols has been recently introduced in
[10]. It is shown that using piecewise linearity between
consecutive OFDM symbols helps ICI mitigation. Authors first
get the least squares channel estimates at the pilots for
consecutive OFDM symbols and use them to estimate channel
variation. In this paper, we introduce an explicit usage of BEM
in time domain, then an update of channel coefficients is
performed in frequency domain in a simple way using
consecutive OFDM symbols. In addition, we investigate the
effect of DF in a practical pilot distribution.
978-1-4244-5213-4/09/ $26.00 2009 IEEE 2404
The rest of this paper is organized as follows. Section II
presents the general system model and how BEM can be used
to approximate the channel variations. Section III discusses
BEM approach for an equispaced pilot distribution, the effect
of decision feedback, and proposes a new algorithm based on
multiple OFDM symbols. Simulation results are presented in
Section IV and we conclude the paper in Section V.
Notation: We use upper (lower) bold face letters to denote
matrices (column vectors), ()
T
and ()
H
denote transpose and
complex conjugate transpose operators, respectively. E[]
stands for the expected value and represents the Kronecker
product. We denote an NN identity matrix by I
N
.
Furthermore, we use X
i,k
to indicate the (i+1, k+1)
th
entry of the
matrix X and diag(x) to indicate a diagonal matrix with x on its
diagonal.
II. SYSTEM MODEL
A. General System Model
In an OFDM system, the symbols collected in symbol
vector s are first transformed from the time domain to the
frequency domain using an IFFT. Then, a CP consisting of the
last symbols is added before the parallel to serial converter. At
the receiver side, a serial to parallel conversion is applied, CP
is removed and an FFT operation is performed to obtain the
received symbols in the frequency domain.

Figure 1: Simple OFDM system diagram.
The expression of the received vector for the (p+1)
th

OFDM symbol can be expressed as:

) ( ) ( ) (
) ( ) (
) (
) ( ) (
p p p
p p H
p
t p
z s G
z s F FH r
+ =
+ =
(1)
where
) (
) (
p
t
H and
) ( p
G represent the channel matrices
respectively in time and frequency domains for the (p+1)
th

OFDM symbol; F denotes FFT operation, and z is the complex
additive white Gaussian noise vector.
For an FFT size of N, it can be easily shown that
) (
) (
p
t
H
has the entries

) , mod( , ) (
) (
,
) (
N k i i L N p
p
k i
cp
t
h H
+ +
= (2)
where L
cp
is the CP length. In the sequel, the index p will be
dropped for the sake of clarity.
Due to the time variation of the channel during the OFDM
symbol, the frequency domain channel matrix G would not be
diagonal. In fact, it will have the entries
[ ] 1 , 0 , ,
1
1
0
) ( 2
, ,
=

N k i e H
N
G
N
n
N
i k n j
n k k i
(3)
where { }
1 , , 0
,
= N k
n k
H

denote the Fourier transform of the
channel impulse response { }
1 , , 0
,
= L l
n l
h

at the time instant n for
a channel length of L and given by
N k e h
N
H
L
l
N
kl j
n l n k
< =

0 ,
1
1
0
2
, ,
(4)
B. Basis Expansion Modeling (BEM)
In this section, we will show how a BEM can be used to
estimate the variations of the channel. The idea behind BEM is
to express the N channel coefficients as a function of (Q+1)
basis functions approximating the variation of the channel
during a specific period [3].
In this modeling, for every channel tap, we write

l
b
l l
+ = Bh h (5)
where [ ]
T
N l l l l
h h h
1 , 1 , 0 ,
, , ,

= h contains the N channel


samples corresponding to the l
th
tap and
[ ]
T
b
Q l
b
l
b
l
b
l
h h h
, 1 , 0 ,
,..., , = h contains the corresponding (Q+1) basis
coefficients,
l
is the modeling error, and B is the basis
expansion matrix collecting the Q+1 basis functions.
In particular, for CE-BEM we have
) 2 / )( / 2 (
,
Q q N p j
q p
e B

= ,
where p = 0,,N-1, q = 0,,Q. Similarly, for GCE-BEM we
have =
q p
B
,

) 2 / ))( /( 2 ( Q q kN p j
e

where k > 1 is the oversampling
ratio, and finally for P-BEM we use
q
q p
p B =
,
.
If we collect all the channel taps in a single column vector
as [ ]
T
N L N L
h h h h
1 , 1 1 , 0 0 , 1 0 , 0
,..., ,..., ,...,

= h and similarly all basis
coefficients in a single column vector as
[ ]
T
b
Q L
b
Q
b
L
b
h h h h
, 1 , 0 0 , 1 0 , 0
,..., ,..., ,...,

=
b
h , then neglecting the
modeling error we obtain

b
h I B h ) (
L
= . (6)
2405
As a result, after some algebra, the received symbol vector
can be expressed in terms of the BEM as

=
+ =
Q
q
q q
0
z s D r (7)
where
H
q q
diag F b F D ) ( = and
q
= diag(F
L
[ ]
T
b
q L
b
q
h h
, 1 , 0
,...,

)=
diag(F
L
b
q
h ). Here, F
L
collects the first L columns of the matrix
. F N
Equation (7) can also be written as
z h F s D r + =

=
b
q L
Q
q
q
diag ) (
0
. (8)
If we define [ ]
Q
D D D D .....
1 0
= and ) ). ( (
1 L Q
diag F s I S =
+ ,

then we obtain
z DSh r + =
b
. (9)
Using (9), channel estimators can be derived either based
on the whole knowledge of input symbol vector s (i.e., the full
preamble case) or based on only a part of symbol vector s (i.e.,
in the presence of pilot signals). We note clearly that BEM
simplifies the estimation as the problem is reduced to
estimating h
b
of size (Q+1)L rather than estimating all the NL
channel coefficients.
III. ALGORITHM
We first present the pilot distribution that is used in the
simulations. Next, we discuss the effect of applying BEM
directly on such a pilot distribution. Then, we propose an
algorithm based on decision feedback considering the use of
two consecutive OFDM symbols.
A. Pilot Distribution
It has been observed in a number of attempts that the
optimal placement of pilots appears to be equispaced clusters
and more precisely zero-padded ones [3, 6, 11]. But for next
generation mobile wireless systems, a sparse distribution in
both frequency and time is being considered. In the sequel, we
will mainly focus on a pilot distribution which resembles the
ones recently adapted in mobile wireless systems [14, 15]
where pilots are located equispaced on the FFT grid.
B. Direct BEM approach
As an initial step, we apply directly a basis expansion on
the channel and try to estimate the basis coefficients just using
the assumed pilot distribution. We will show that it is not
efficient to calculate the variation of the channel which is
normally deduced from the interference terms.
Separating the input symbol vector s into a vector s
p

containing pilots and a vector s
d
containing data symbols, we
can rewrite (9) as
z h DS h DS r + + =
b
d
b
p
. (10)
Then, considering the channel realizations and the data
symbols as independent stochastic processes, the expression of
the LMMSE estimator can be obtained as [3]

1
0
) (

+ + =
N x
H
hb
H
hb LMMSE
N I R R R C (11)
where [ ]
H
b b
hb
E h h R = is the autocorrelation matrix between
the basis expansion coefficients which can be concluded from
the autocorrelation matrix of the real channel coefficients, and
N
0
denotes the noise spectral density. In (11),
p
DS =
depends on the pilots and [ ]
H H
d
b b
d x
H
E D S h h S D R = is
calculated using the assumed statistical properties of the
channel and data.
C. Decision Feedback
As shown in the sequel, using pure BEM with equispaced
pilots may not be sufficient to have a satisfactory estimation;
one may use an iterative approach based on detected symbols
to feedback the estimator. In order to improve the performance,
we make use of decision based feedback such that a first
estimation just using the pilots can be used for equalization and
then the detected symbols serve as a preamble to re-estimate
the basis coefficients. For the sake of simplicity and without
loss of generality, our study will be based on hard decision
feedback while it is clear that with soft-decision based
estimators one may provide better performance. Successive
iterations can be performed in order to enhance the
performance. However, it will be shown that with the proposed
approach two iterations are enough to converge rapidly to a
performance limit.
D. Linear Filtering Between Successive OFDM Symbols
In this subsection, we propose a linear filter to further
improve the estimation of the interference coefficients (non
diagonal terms) in the frequency domain channel matrix G. It
is mainly based on the initial information from successive
OFDM symbols. As explained below, this improvement is due
to the fact that ICI terms come mainly from the channel
variation during the OFDM symbol and thus can also be
estimated by the variation of the frequency domain channel
matrix through consecutive OFDM symbols.
Let us first investigate the behavior of the coefficients in
the G matrix assuming a satisfactory approximation.
Particularly, we use P-BEM with Q = 1 (linear variation) for
the channel variation during two consecutive OFDM symbols.
In this case, each channel coefficient for a certain tap l can be
expressed as:
N n n h h
l l n l
< + = 0 ,
0 , ,
(12)
where h
l,0
and
l
become the basis coefficients for the l
th
tap,
and can be seen as the initial point and the slope of a straight
2406
line representing the variation of the channel in a certain time
window (i.e., over several OFDM symbols).
It can be easily proved that for the (p+1)
th
OFDM symbol
one obtains

i cp i
p
n i
E L N p n H N H )) ( (
0 ,
) (
,
+ + + = (13)
where E
i
s are simply the Fourier transform of the
l
s defined
in (12), i.e.,

N i e
N
E
L
l
N
il j
l i
< =

0 ,
1
1
0
2
(14)
Accordingly, the coefficients of G can be expressed as

= + + +
=

elsewhere
e
E N
k i L N p E N N E N H N
G
N
i k j
k
cp i i i
p
k i
,
1
1
.
), ( 2 / ) 1 (
) ( 2
0 ,
) (
,


(15)

Comparing the diagonal terms in (15) for the first and second
OFDM symbols we simply obtain
) (
) 0 (
,
) 1 (
, cp i i i i i
L N E N G G + = (16)
From (15), it can be seen that G
i,k
, for all ik, depends
strictly on E
i
. Based on these observations, we can deduce that
a good knowledge of the diagonal terms of two successive
OFDM symbols allows us to estimate all the coefficients of
the G matrix. This simply means that once we have
sufficiently accurate estimates of the diagonal terms belonging
to the channel matrices for the successive symbols, we can use
presented linear filter to estimate the other terms and improve
the estimation performance.
IV. SIMULATION RESULTS
In this section, we analyze the performance of the P-BEM
with Q=1, the effect of hard decision feedback and the
improvement with the proposed filtering based on multiple
OFDM symbols. In the simulations, the number of subcarriers
is chosen to be N = 256 where 180 of them are used for
transmission and the others will be considered as symmetric
guard carriers. The symbol duration is taken to be 66.7sec. A
maximum Doppler spread of 300Hz is considered and it is
supposed to be perfectly known at the receiver. Therefore, for
an inter-subcarrier spacing of 15 kHz, the normalized Doppler
spread considered in the simulations is equal to 0.02. For data
symbols, 64-QAM constellation is adopted while for pilots
QPSK modulation is used assuming equal energy on the
whole FFT grid. The pilot spacing is taken to be 6 (each 6
successive subcarriers carry one pilot) and the length of the
cyclic prefix is 19 samples.
We use an improved Jakes model [1] to generate the
Extended Vehicular A [13] tapped delay channel profile
having tap delays [0 0.03 0.15 0.31 0.37 0.71 1.09 1.73
2.51](sec) with the corresponding tap relative powers [0 -1.5
-1.4 -3.6 -0.6 -9.1 -7 -12 -16.9] (dB). The above mentioned
simulation parameters are compatible with the ones defined
for 4
th
generation wireless mobile communications systems
LTE [14] and WiMAX [15]. We note here that in the proposed
receiver we do not take into account any channel profile
knowledge. However, a further knowledge on it might
improve the performance of the proposed method. For the
equalization, a zero forcing (ZF) equalizer is used.
Performance is investigated as symbol error rate (SER) and
Mean Squared Error (MSE) as a function of received SNR
which is defined as the ratio of received signal energy per
symbol to the noise spectral density N
0
.
We first investigate the effect of BEM using the pilots. As
shown in Figure 2, using only BEM based on separately
distributed pilots is not enough to estimate the variation of the
channel. This simply introduces inter carrier interference
which results in an error floor at the SER of 3x10
-2
. In fact, the
effect of ICI comes mainly from several nearest samples. So a
sparse pilot distribution doesnt seem enough to have a good
estimation performance. For this reason we introduce an
iterative approach based on hard decisions. In Figure 2, we
can see the effect of such iterations on the channel estimation
accuracy. We see a considerable improvement especially in
the first two iterations where a 5 dB gain can be obtained for
an SER of 3x10
-2
. However, we are still far from the best
achievable performance which is depicted as the solid line.


Figure 2: Effect of hard decision feedback.
Figure 3 shows the mean squared error (MSE) curve for
the diagonal terms of the G matrix using direct BEM approach
based on the pilot signals and after decision feedback. It can
be observed that decision feedback improves the estimation of
diagonal terms of the matrix G where an improvement from
5x10
-4
to 8x10
-5
can be obtained after the first iteration at an
2407
SNR value of 40 dB. This motivates the idea of employing
these diagonal coefficients to estimate the non diagonal ones.


Figure 3: Effect of decision feedback, MSE of diagonal terms in G matrix.

Figure 4: SER performance of the proposed algorithm
based on successive OFDM symbols.
As explained in Section III, at each iteration of the
proposed algorithm, we use the estimated diagonal
coefficients in the G matrices corresponding to successive
OFDM symbols in order to recalculate the ICI coefficients.
Comparing the iterations in Figure 4 with those of Figure 2,
one can easily see the remarkable performance improvement
just by using this simple linear filtering after each iteration.
Our algorithm makes use of the BEM approach first to
estimate the channel in the time domain then to filter the
frequency-domain channel matrices estimated corresponding
to multiple OFDM symbols. This leads to a further gain of 5
dB at high SNR compared to the case where only decision

feedback is employed. It is also worth noting that addition of
such a simple filtering allows a faster convergence to a better
performance limit due to better estimates obtained from the
first iteration.
V. CONCLUSION
In this paper, we investigated a linear modeling to estimate
rapidly varying channels using an equispaced pilot
distribution. We further studied the effect of decision feedback
and we proposed a new method to improve the estimation
performance based on two successive OFDM symbol
observations. It has been shown by simulations that the
proposed method provides a considerable performance
enhancement by means of a simple linear filtering.
ACKNOWLEDGMENT
The authors would like to acknowledge the support of the
European Commission through the FP7 project WiMAGIC
(see www.wimagic.eu).
REFERENCES

[1] W. C. Jakes, Microwave Mobile Channels. New York: Wiley, 1974.
[2] P. Hoeher, S. Kaiser and P. Robertson, Two-dimensional pilot-symbol-
aided channel estimation by Wiener filtering, Proc. Int. Conf.
Acoustics, Speech, and Signal Processing, pp. 1845-1848, Munich,
Germany, April 1997.
[3] Z. Tang, R.C. Cannizzaro, G. Leus, and P. Banelli, Pilot-assisted time-
varying channel estimation for OFDM systems, IEEE Trans. Signal
Process., vol. 55, no. 5, May 2007.
[4] M. Visintin, Karhunen-Loeve expansion of a fast Rayleigh fading
Process, IEEE Electron. Lett., vol. 32, no. 8, pp. 17121713, Aug.
1996.
[5] K. D. Teo and S. Ohno, Optimal MMSE finite parameter model for
doubly-selective channels, in Proc. IEEE Global Telecommun. Conf.
(GLOBECOM), pp. 35033507, St. Louis, MO, Dec. 2005.
[6] X. Ma, G. Giannakis, and S. Ohno, Optimal training for block
transmissions over doubly-selective fading channels, IEEE Trans.
Signal Process., vol. 51, no. 5, pp. 13511366, May 2003.
[7] M. K. Tsatsanis and G. B. Giannakis, Modeling and equalization of
rapidly fading channels, Int. J. Adapt. Control Signal Process, vol.10,
pp. 159176, Mar. 1996.
[8] G. Leus, On the estimation of rapidly time-varying channels, in Euro.
Signal Process. Conf. (EUSIPCO), Sep. 2004.
[9] D. K. Borah and B. D. Hart, Frequency-selective fading channel
estimation with a polynomial time-varying channel model, IEEE Trans.
Commun., vol. 47, no. 6, pp. 862873, June 1999.
[10] Y.Mostofi and D.C. Cox, ICI mitigation for pilot-aided OFDM mobile
systems, IEEE Trans. Wireless Commun., Volume 4, Issue 2, Mar.
2005 pp. 765774
[11] M. Dong, L. Tong, and B. Sadler, Optimal insertion of pilot symbols
for transmissions over time-varying flat fading channels, IEEE Trans.
Signal Process., vol. 52, no. 5, pp. 1403-- -1418, May 2004.
[12] S. Chen and T. Yao, Intercarrier interference suppression and channel
estimation for OFDM systems in time-varying frequency-selective
fading channels, IEEE Trans. Consumer Electronics, Vol. 50, No. 2,
May 2004.
[13] 3GPP TS 36.521-1: "E-UTRA UE conformance specification Radio
transmission and reception," V8.0.1, Dec 2008.
[14] 3GPP TS 36.211 Physical Channels and Modulation, V8.6.0, Mar.
2009.
[15] WiMAX Forum
TM
Mobile System Profile, Release 1.0 Approved
Specification, Revision 1.7.1, Nov. 2008.

2408