Professional Documents
Culture Documents
S. M. S. Sadough
1,2
, M. Ichir
2
, P. Duhamel
2
and E. Jaffrot
3
1
UEI, ENSTA, 32 boulevard Victor, 75015 Paris, France
2
Laboratoire des Signaux et Syst` emes, CNRS-Sup elec, Plateau de Moulon, 91190 GifsurYvette, France
3
Universidad Nacional de San Martin, B1650ANQ San Martin, Pcia. de Buenos Aires, Argentina
SUMMARY
Ultra wideband (UWB) communications involve very sparse channels, since the bandwidth increase results
in a better time resolution. This property is used here to propose an efcient algorithm jointly estimating the
channel and the transmitted symbols. More precisely, this paper introduces an expectation-maximisation
(EM) algorithm within a wavelet domain Bayesian framework for semi-blind channel estimation of
multiband orthogonal frequency-division multiplexing (MB-OFDM) based UWB communications. A prior
distribution is chosen for the wavelet coefcients of the unknown channel impulse response (CIR) in order
to model a sparseness property of the wavelet representation. This prior yields, in maximum a posteriori
(MAP) estimation, a thresholding rule within the EM algorithm. We particularly focus on reducing the
number of estimated parameters by iteratively discarding insignicant wavelet coefcients from the
estimation process. Simulation results using UWB channels issued from both models and measurements
show that under sparsity conditions, the proposed algorithm outperforms pilot based channel estimation
in terms of mean square error (MSE) and bit error rate (BER). Moreover, the estimation accuracy is
improved, while the computational complexity is reduced, when compared to traditional semi-blind methods.
Copyright 2008 John Wiley & Sons, Ltd.
1. INTRODUCTION
An ultra wideband (UWB) radio signal is dened as
any signal whose bandwidth is larger than 20% of
its centre frequency or greater than 500 MHz [1]. In
recent years, UWB system design has experienced a
shift from the traditional single-band radio that occupies
the whole 7.5 GHz allocated spectrum to a multiband
design approach [2]. In these multiband UWB systems, the
available spectrum is divided into several subbands, each
one occupying approximately 500 MHz.
Multiband orthogonal frequency-division multiplexing
(MB-OFDM) [3] is an appropriate candidate for multiband
UWB which enables high data rate UWB transmission to
inherit all the strength of OFDM that is widely used in
numerous communication systems. This approach uses a
* Correspondence to: Seyed Mohammad Sajad Sadough, UEI, ENSTA, 32 boulevard Victor, 75015 Paris, France. E-mail: sajad.sadough@lss.supelec.fr.
A previous version of this paper was presented in the 13th European Wireless Conference (EW 2007), Paris, France.
conventional coded OFDM system [4] together with bit
interleaved coded modulation (BICM) [5] and frequency
hopping over different subbands to improve the diversity
and to enable multiple access.
It is well known that an efcient detection of the
transmitted symbols requires a good knowledge of the
channel realisations at the receiver. The required channel
estimation is often based on the transmission of known
training symbols (also called pilots) [6]. However, the
length of the required training sequence may become
large when the channel impulse response (CIR) is highly
resolved, such as in UWB OFDM systems. Besides,
obtaining an accurate channel estimate in highly mobile
environments only through the use of pilots, would require
inserting multiple training symbols per frame, which can
result in a waste of bandwidth and power.
Copyright 2008 John Wiley & Sons, Ltd. Accepted 16 June 2008
762 S. M. S. SADOUGH ET AL.
In order to reduce the number of pilot symbols, the
information carried by the observations corresponding
to the data symbols can also be exploited for channel
estimation. An attractive technique in this area is that based
on the so-called expectation-maximisation (EM) algorithm
[7] which provides a framework to iteratively estimate the
unknown channel thanks to the a posteriori probabilities on
the unknown data. Moreover, by using the EM algorithm,
one can naturally combine the iterative process of channel
estimation with the decoding operation of encoded data.
Several papers have recently reported promising results
on the combination of EM-based channel estimation and
data decoding process. For instance, iterative or turbo
processing that includes the channel estimation into the
iterative process of decoding turbo like codes is addressed
in Reference [810]. In particular, Reference [9] addresses
a turbo channel estimation based on a KarhunenLoeve
expansion of the unknown channel for reducing the
number of estimated parameters. However, the proposed
scheme requires the channel covariance matrix as an a
priori information which can not be available in realistic
situations. In Reference [11], the authors present several
EMbased algorithms for time or frequency domain channel
estimation of an OFDM system that is subject to slow time-
varying frequency-selective fading. In Reference [12], Lu
et al. proposed an iterative receiver for space-time block-
coded OFDM systems based on a reduced complexity EM
algorithm.
Though iterative joint channel estimation and data
detection scheme outperforms receivers using a pilot-only
based channel estimation, it has a higher complexity that
may be a critical concern for its practical implementations.
This complexity is mainly driven by the number of
estimated parameters for channel updating and the decoding
algorithm within each iteration.
In this work, we consider a semi-blind joint channel
estimation and data detection scheme for MB-OFDM
systems based on the EM algorithm, with the objective
of minimising the number of estimated parameters and
enhancing the estimation accuracy. This is achieved by
expressing the unknown UWB CIR in terms of its
discrete wavelet expansion, which has been shown to
provide a parsimonious representation[13, 14]. Our wavelet
domain framework enables us to choose a particular a
priori distribution for the channel wavelet coefcients
that renders the maximum a posteriori (MAP) channel
estimation equivalent to a hard thresholding rule at each
iteration of the EM algorithm. The latter is then exploited
to reduce the estimators computational complexity by
discarding insignicant wavelet coefcients from the
estimation process. Moreover, the adopted wavelet domain
prior model adapts itself to the actual channel scenario
since its parameters are learnt from the observed data
during the channel estimation process. More precisely,
the prior model is not restricted to a specic propagation
environment. We show that when the channel has a sparse
wavelet expansion, the proposed algorithmoutperforms the
classical EMestimators and provides a signicant reduction
in the number of estimated parameters. Hence, the proposed
receiver has not to modify the prior information for different
channel scenarios.
This paper is organised as follows: Section 2 introduces
the MB-OFDM system and its equivalent model for
wavelet domain channel estimation. Section 3, rst presents
the derivation of the MAP channel estimator based on
the EM algorithm and then shows how the number of
estimated parameters can be reduced through the EM
iterations. Section 4 presents the combination of the
channel estimation with the decoding operation, as well
as some implementation issues. Section 5 illustrates, via
simulations, the performance of the proposed receiver in
different realistic UWB channel environments and Section
6 concludes the paper.
Notational conventions are as follows: D
x
is a diagonal
matrix with diagonal elements x = [x
1
, . . . , x
N
]
T
, E
x
[]
refers to expectation with respect to x, I
N
denotes an (N
N) identity matrix; CN(m, ) denotes complex Gaussian
distribution with mean m and covariance matrix ; Card{}
denotes set cardinality; ()
T
and ()
H
denote matrix or vector
transpose and Hermitian transpose, respectively.
2. SYSTEM MODEL AND WAVELET DOMAIN
PROBLEM FORMULATION
2.1. MB-OFDM transmission
We consider the MB-OFDM transmission proposed in
Reference [3]. This scheme divides the spectrum between
3.1 and 10.6 GHz into several non-overlapping subbands
eachone occupying528 MHz of bandwidth[3]. Information
is transmitted using OFDM modulation over one of the
subbands in a particular time slot. As shown in Figure 1,
the transmitter architecture for the MB-OFDM system is
very similar to that of a conventional wireless OFDM
system. The main difference is that MB-OFDMuses a time-
frequency code (TFC) to select the centre frequency of
different subbands. This code is used to provide frequency
diversity as well as to distinguish between multiple users
(see Figure 2). Here, we consider MB-OFDM in its basic
Copyright 2008 John Wiley & Sons, Ltd. Eur. Trans. Telecomms. 2008; 19:761771
DOI: 10.1002/ett
ULTRA WIDEBAND OFDM CHANNEL ESTIMATION 763
Figure 1. TX architecture of the multiband OFDM system.
Figure 2. Example of time-frequency coding for the multiband
OFDM system: TFC = {1, 3, 2, 1, 3, 2, . . .}.
mode, i.e. employing the rst three subbands with N data
subcarriers over each subband.
Let us consider a single-user MB-OFDM transmission.
At the receiver, assuming a cyclic prex (CP) longer than
the channel maximum delay spread and perfect carrier
synchronisation, OFDM converts a frequency-selective
channel into N parallel at fading subchannels [4]. Under
these conditions, the transmission of the nth OFDMsymbol
(inside a frame of size N
sym
) over the ith subband can be
written as [4]
y
i,n
= D
s
i,n
H
i,n
+ z
i,n
(1)
for i {1, 2, 3} and n = 1, . . . , N
sym
; (1 N) vectors
y
i,n
, s
i,n
and
H
i,n
denote respectively the received and
transmitted symbols, and the channel frequency response;
the noise vector z
i,n
is assumed to be a zero-mean
circularly symmetric complex Gaussian randomvector with
distribution CN(0,
2
I
N
).
In what follows, we group the data and the observation
corresponding to three subbands of the MB-OFDM system
in a single vector. This operation can be written in the
frequency domain as
Y
m
= D
S
m
H
m
+ Z
m
m = 1, . . . , M
sym
(2)
where Y
m
= [y
1,n
, y
2,n+1
, y
3,n+2
]
T
, S
m
= [s
1,n
, s
2,n+1
,
s
3,n+2
]
T
, H
m
= [
H
1,n
,
H
2,n+1
,
H
3,n+2
]
T
and Z
m
= [z
1,n
,
z
2,n+1
, z
3,n+2
]
T
are (M 1) vectors, with M = 3N
and M
sym
= N
sym
/3. In the remainder, unless otherwise
mentioned, we will not write the time index mfor notational
convenience.
2.2. A proper model for wavelet domain channel
estimation
In order to take advantage of the wavelet based estimation,
the CIR is expressed in terms of its orthogonal discrete
wavelet coefcients. Let F
M,L
be the truncated fast Fourier
transform(FFT) matrix constructed fromthe (M M) FFT
matrix by keeping the rst Lcolumns where Lis the length
of the CIR over a group of three subbands. We dene W as
the (L L) orthogonal discrete wavelet transform(ODWT)
matrix. The unknown channel frequency response can be
expressed as H = F
M,L
W
H
g, where g is the (L 1) vector
of the CIR wavelet coefcients. The observation model (2)
is rewritten as
Y = D
S
Tg + Z (3)
where T = F
M,L
W
H
. In this model, although the channel
is practically used (by the transmitter) by slices of 528 MHz
bandwidth (corresponding to a single subband), on the
receiver side, three received OFDM symbols are grouped
for estimating the wavelet coefcients of the CIR, taken
over all three subbands (1.584 GHz bandwidth). This is
motivated by the fact that estimating the channel over a
wider bandwidth leads to a sparser representation in the
wavelet domain.
Most of wavelet based estimation algorithms, rely on
an observation model in which the unknown wavelet
coefcients are corrupted by an AWGN [13]. Since the
model (3) does not provide such a framework, our rst
step consists in enforcing this property. In order to do so,
the AWGN in Equation (3) is split into two independent
Gaussian terms as suggested in Reference [15]
Z = D
S
Z
1
+ Z
2
(4)
For the sake of notational brevity in Equation (2), we have assumed that
the TFC is equal to {1, 2, 3, . . .}.
Copyright 2008 John Wiley & Sons, Ltd. Eur. Trans. Telecomms. 2008; 19:761771
DOI: 10.1002/ett
764 S. M. S. SADOUGH ET AL.
where Z
1
and Z
2
are (M 1) independent Gaussian
noise vectors such that p(Z
1
) = CN(0,
2
I
M
) and p(Z
2
) =
CN(0,
2
I
M
2
D
S
D
S
). Since we are using power
normalisedQPSKsymbols, D
S
D
S
= I
M
andthe covariance
matrix
2
of Z
2
reduces to
2
= (
2
2
)I
M
. We dene
the positive design parameter
2
/
2
, (0 < 1) as
the proportion of noise that is assigned to Z
2
. Obviously,
the value assigned to parameter is important and will have
to be tuned. Note that setting = 0 leads to Z
1
= 0 and is
equivalent to working with the initial model (3). However,
for 0 < < 1, the above noise decomposition allows the
introduction of a hidden channel vector
H and a two-stage
observation model dened as
_
H = Tg + Z
1
Y = D
S
H + Z
2
.
(5)
This model introduces a hidden vector
H which provides
us with a direct relation
.
the auxiliary or the Q-function and dened as
Q
_
g, g
(t)
_
= E
S,
H
_
log p(Y, S,
H|g)
Y, g
(t)
_
(6)
Maximisation Step (M-step): The vector of estimated
parameters is updated according to
g
(t+1)
= arg max
g
_
Q
_
g, g
(t)
_
+ log (g)
_
(7)
where (g) a prior distribution for the wavelet coefcients
which ensures a certain percentage of coefcients to be set
to zero (see Subsection 3.2). When applied to Equation (5),
each step can be calculated as follows:
3.1. E-step: computation of the Q-function
The complete likelihood is
p(Y, S,
H|g) = p(Y|S,
H, g)p(S|
H, g)p(
H|g).
According to Equation (5), given
H, Y and g are
independent. Furthermore, S which results fromcoding and
interleavingof bit sequence is independent of
Handg. Since
Z
1
is a complex white Gaussian noise, the complete log-
likelihood can be simplied to
log p(Y, S,
H|g) = log
_
p(Y|S,
H)p(S)p(
H|g)
= log p(
H|g) + cst.1
=
g
Tg 2Re
_
g
H
_
2
+ cst.2
where cst.1 and cst.2 are constant terms that do not depend
on g. According to Equation (6), we have
Q
_
g, g
(t)
_
= E
S,
H
_
Tg 2Re
_
g
H
_
2
+ cst.2
Y, g
(t)
_
=
H
(t)
Tg
2
2
+ cst.3 (8)
where the latter equation is obtained by adding and
subtracting
H
(t)
2
, with
H
(t)
E
S,
H
[
H|Y, g
(t)
] and
cst.3 representing a constant term.
Copyright 2008 John Wiley & Sons, Ltd. Eur. Trans. Telecomms. 2008; 19:761771
DOI: 10.1002/ett
ULTRA WIDEBAND OFDM CHANNEL ESTIMATION 765
From Equation (8), it is obvious that the E-step involves
the computation of
H
(t)
, as follows:
H
(t)
=
(t)
H
(S)p(S|Y, g
(t)
) (9)
where
(t)
H
(S) E
H
_
H|Y, g
(t)
=
_
Hp(
H|Y, g
(t)
, S)d
H.(10)
Equation (9) results from the independence between S
and
H belonging respectively to the sets and .
In order to evaluate
H
(t)
in Equation (9), we rst have to
evaluate the conditional mean
(t)
H
(S) of
Hin Equation (10).
To this end we write p(
H|Y, g
(t)
, S) p(Y|
H, S)p(
H|g
(t)
).
Since both p(Y|
H, S) and p(
H|g
(t)
) are Gaussian densities,
it is well known that their product remains Gaussian. We
have to evaluate the mean of this Gaussian density.
After some algebra we get [16]
(t)
H
= Tg
(t)
+ D
S
_
Y D
S
Tg
(t)
_
. (11)
By introducing Equation (11) in Equation (9) we obtain
H
(t)
= (1 )Tg
(t)
+ D
S
Y (12)
where D
S
=
s
D
S
p(S|Y, g
(t)
). Here, we assume that
a part of the receiver called soft-input soft-output (SISO)
decoder, is able to provide the vector of probabilities
p(S|Y, g
(t)
) required in Equation (12) (see Section 4 for
more details). The E-step is then completed by inserting
H
(t)
into Q(g, g
(t)
) of Equation (8).
3.2. M-step: wavelet based MAP estimation
In this step the estimate of the parameter g is updated as
given in Equation (7) where Q(g, g
(t)
) is given by Equation
(8). We have
g
(t+1)
= arg max
g
_
H
(t)
Tg
2
2
+ log (g)
_
. (13)
It can be easily shown that the expression (13) is
equivalent to
g
(t+1)
= arg max
g
_
g
(t)
g
2
2
+ log (g)
_
(14)
where
g
(t)
= T
H
(t)
= (1 )g
(t)
+ (D
S
T)
Y. (15)
In fact, g
(t+1)
in Equation (14) is not more than the MAP
estimate of g from the observation model:
g
(t)
= g + Z
1
(16)
where Z
1
CN(0,
2
I
L
).
Note that Equation (16) is very important since it shows
that the initial observation model is equivalent to an
observation model which involves a direct relation between
the unknown wavelet coefcient g and g
(t)
, and this direct
relation is corrupted by an AWGN. This is the reason of
using the two level observation model in Equation (5). As
clear from Equation (16), our channel estimation problem
can now be viewed as a standard wavelet domain denoising
problem. In what follows, we derive the update formula of
our wavelet domain channel estimator.
From the Bayes theorem, the posterior distribution of g
is given by
p
_
g
g
(t)
_
p
_
g
(t)
g
_
(g) (17)
where from Equation (16), p( g
(t)
|g) is the Gaussian
likelihood, g
(t)
CN(g,
2
I
L
). In this approach, (g) is
a prior distribution, chosen for the wavelet coefcients g of
the unknown CIR as follows.
Choice of the BernoulliGaussian Prior Model: Here, the
parsimonious characteristic of wavelet basis is imposed
through the following prior model: each wavelet coefcient
is assumed to have a probability to be zero and a
probability 1 to be Gaussian distributed as CN
g
j
(0,
2
).
This corresponds to an independent and identically
distributed (i.i.d.) BernoulliGaussian [13] prior model for
the probability density of the jth wavelet coefcient g
j
as
(g
j
) = (g
j
) + (1 )CN
g
j
(0,
2
) (18)
for j = 1, . . . , L. The parameters and (hyperparameters
in the Bayesian wording) are estimated from the observed
data.
In order to deal with that particular model, we introduce
an additional state variable (or indicator)
j
{0, 1} such
Copyright 2008 John Wiley & Sons, Ltd. Eur. Trans. Telecomms. 2008; 19:761771
DOI: 10.1002/ett
766 S. M. S. SADOUGH ET AL.
that we can express the above prior model conditionally as
_
(g
j
|
j
= 0) (g
j
) with probability ,
(g
j
|
j
= 1) CN
g
j
_
0,
2
_
withprobability 1 .
(19)
This prior model, conditionally on the state variable
j
, leads to a Gaussian posterior for g
j
which makes the
estimationexplicit; fromthe direct observationmodel g
(t)
j
=
g
j
+ Z
1,j
(Equation (16)), we can express these posterior
probabilities of
j
as [16]
p
_
j
= 0| g
(t)
j
_
= N
_
0,
2
_
/c
p
_
j
= 1| g
(t)
j
_
= (1 ) N
_
0,
2
+
2
_
/c
(20)
where the parameter c = N(0,
2
) + (1 )N(0,
2
+
2
). From this set of equations, we notice that the indicator
variable
j
allows us to discriminate between the noise
coefcients (for
j
= 0) and the effective channel wavelet
coefcients (for
j
= 1), eventually corrupted by noise.
The indicator variables
j
are estimated in the MAP
sense by
(t+1)
j
=
_
0, if p
_
j
= 0
g
(t)
j
_
0.5
1, elsewhere.
(21)
Therefore, the MAP estimates of the channel wavelet
coefcients are obtainedbya simple denoising/thresholding
rule as
g
(t+1)
j
=
_
_
0, if
(t+1)
j
= 0
2
+
2
g
(t)
j
, if
(t+1)
j
= 1.
(22)
Updating the Prior Parameters and : The prior
parameters and stand respectively for the signicant-
wavelet coefcients energy and insignicant coefcient
probability. Since is a scale parameter, a Chi-square
prior is chosen for its inverse while a Dirichlet prior is
chosen for the probability . We point out that these two
priors are widely used in Bayesian estimation theory where
scale and discrete probabilities are in concern. The main
advantage for such priors is that they generally result in
relatively simple update algorithms. Here, the update rules
for these two parameters are MAP based rules derived from
assigning conjugate priors to these parameters [17]. After
some algebra we obtain [16]
= (
L 1/2)/(L 1),
2
= /(L
L) (23)
where
L = Card{j |
j
= 0} and =
j
=1
g
(t+1)
j
2
.
Reduction of the Number of Estimated Parameters: The
thresholding procedure derived in this section, provides an
automatic framework for reducing the number of estimated
coefcients. This can be achieved by discarding at each
iteration, the elements of g
(t+1)
that are replaced by zero in
Equation (22). The underlying assumption is as follows:
whenever the estimator assimilates an unknown wavelet
coefcient to noise (i.e. replaces it by zero), this coefcient
will always be considered as noise, hence it will not be
estimated in future iterations. We veried by simulations
that incorporating this scheme into the EM algorithm
reduces the number of estimated parameters without any
signicant performance degradation.
This operation is shown in Figure 3 and can be modelled
as
g
(t+1)
tr
=
_
g
(t+1)
_
, T
tr
=
_
T
_
(24)
where the truncation operator () gathers in g
(t+1)
tr
the
components of g
(t+1)
that must be kept and the operator
() constructs T
tr
from T by keeping the columns
corresponding to the kept indexes. During the rst iteration
(t = 0), the algorithm does not perform any truncation and
the EM algorithm estimates all the coefcients. However,
after each M-step, the number of unknown parameters to
be estimated in the next iteration is reduced according
to Equation (24) by using g
(t+1)
tr
and T
tr
in the updating
formula of the E-step (12).
Figure 3. EM-MAP channel estimation combined with the
decoding process.
Copyright 2008 John Wiley & Sons, Ltd. Eur. Trans. Telecomms. 2008; 19:761771
DOI: 10.1002/ett
ULTRA WIDEBAND OFDM CHANNEL ESTIMATION 767
4. DECODING METHOD AND
IMPLEMENTATION ISSUES
The block diagram of the receiver is shown in Figure 3.
Besides the channel estimation part, the rest of the receiver
principally consists of the combination of two sub-blocks
that exchange soft informations with each other. The rst
sub-block, referedtoas soft demapper (alsocalleddetector),
produces bit metrics (probabilities) from the input symbols
and the second one is a SISOdecoder. Here, SISOdecoding
is performed using the well known forward-backward
algorithm [18].
Let c
k,i
(i = 1, . . . , B) be the ith
j=1
j=i
P
dec
(c
k,j
) (25)
where P
dec
(c
k,j
) are prior probabilities coming from the
SISO decoder. At the rst iteration, where no a priori
information is available on bits c
k,j
, the probabilities
P
dec
(c
k,j
) are set to 0.5.
The main steps of the EM-MAP estimator are
summarised as follows:
r
Initialisation (t = 0)
Set all probabilities of coded bits P
dec
(c
k,i
) to 0.5 and
derive P(S
k
|Y
k
,
H
(0)
k
) for all k according to Equation
(25) and then derive p(S|Y, H
(0)
) = p(S|Y, g
(0)
).
Initialise the unknown vector g by g
(0)
obtained from
pilot symbols.
r
for {t = 1, . . . , t
max
}