Date post: | 30-May-2018 |
Category: |
Documents |
Upload: | khajarasoolsk |
View: | 214 times |
Download: | 0 times |
of 12
8/14/2019 Improved Iterative Detection And
1/12
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008 5039
Improved Iterative Detection andAchieved Throughputs of OFDM Systems under
Imperfect Channel Estimation
Seyed Mohammad Sajad Sadough, Member, IEEE, and Pierre Duhamel, Fellow, IEEE
AbstractAssuming imperfect channel estimation, we pro-pose an improved detector for orthogonal frequency-divisionmultiplexing (OFDM) systems over a frequency-selective fadingchannel. By adopting a Bayesian approach involving the statisticsof the channel estimation errors, we formulate an improvedmaximum-likelihood (ML) detection metric taking into accountthe characteristics of the channel estimates. As a first step, wepropose a modified iterative detector based on a maximum aposteriori receiver formulation which reduces the impact of chan-nel uncertainty on the decoder performance, by an appropriateuse of this metric. The results are compared to those obtainedby using a detector based on a mismatched ML metric, whichuses the channel estimate as if it was the perfect channel. Ina second step, we calculate the information rates achieved byboth the improved and mismatched ML detectors, in terms ofachievable outage rates. These outage rates are compared tothose provided by a theoretical (not practical) decoder definedas the best decoder in the presence of channel estimationerrors. Numerical results over both uncorrelated Rayleigh fadingchannels and realistic ultra wideband channels show that theimproved detector outperforms the classically-used mismatchedapproach in terms of bit error rate and achievable outage rates,without any increase in the receiver complexity.
Index TermsOFDM, BICM, channel estimation errors, mis-matched ML detection, improved iterative BICM detection,achievable outage rates.
I. INTRODUCTION
ORTHOGONAL frequency-division multiplexing
(OFDM) is a spectrally efficient technique for achieving
high data-rate wireless transmission over frequency-selective
fading channels. OFDM offers a simple way to divide the
entire channel into many narrowband subchannels withparallel data transmission, thereby increasing the symbol
duration and reducing the intersymbol interference [1], [2].
Several OFDM based wireless systems employ bit-interleaving combined with a convolutional code as chan-
nel coding. This scheme, referred in the literature as bit-
interleaved coded modulation (BICM) [3], can provide high
Manuscript received August 25, 2007; revised December 7, 2007; acceptedJanuary 26, 2008. The associate editor coordinating the review of this paperand approving it for publication was M. Valenti. The material in this paperwas published in part at the 40th Asilomar Conference on Signals, Systems,and Computers.
S. M. Sajad Sadough is with the Faculty of Electrical and Com-puter Engineering, Shahid Beheshti University (SBU), Tehran, Iran (e-mail:[email protected]).
P. Duhamel is with CNRS/LSS, Suplec - 3 rue Joliot-Curie, 91192 Gif-sur-Yvette cedex, France (e-mail: [email protected]).
Digital Object Identifier 10.1109/T-WC.2008.070944
diversity order for transmission over Rayleigh fading channels
with coherent detection and perfectknowledge of the channel.
It is well known that reliable coherent data detection is
not possible unless an accurate channel state information is
available at the receiver. A typical scenario for wireless com-
munication systems assumes the channel changes so slowly
that it can be considered time invariant during the transmission
of an entire frame. In such situations, pilot symbol assisted
modulation (PSAM) [4] has been shown to be an effectivesolution for obtaining channel state information at the receiver
(CSIR). It consists of sending some known training (alsocalled pilot) symbols, based on which the receiver estimates
the channel before proceeding to the detection of data sym-bols. Obviously, in practical systems, due to the finite number
of pilot symbols and to noise, the receiver can only obtain an
imperfect(and possibly very poor) estimate of the channel. In
this situation, one may resort to blind detectors which estimate
the transmitted symbols without using any pilot. However,
most of these schemes require a differential modulation whichentails restrictions on the type of the constellation and leads to
about 3 dB loss over Gaussian channels compared to coherentdetection [5].
The performance of BICM over fading channels was stud-
ied, for instance, in [3], [6], [7] under prefect channel knowl-
edge at the receiver. In [8], the authors proposed a blind
iterative receiver for differentially encoded BICM OFDM
systems with unknown frequency-selective fading channels.
A rich literature exists on the impact of imperfect channel
estimation. In [9], the authors showed that in low signal-
to-noise ratio (SNR) situations, the poor quality of channel
estimates may prevent iterative decoding of BICM to be used
at the decoder. In [10], Garg et al. considered a training-based
MIMO system and showed that for compensating the perfor-mance degradation due to imperfect channel estimation, the
number of receive antennas should be increased. Obviously,this may not be always possible in mobile applications.
In order to deal with imperfect channel estimation, the clas-
sical technique, referred to as mismatched detection, consists
in replacing the exact channel by its estimate in the receiver
metric. It is shown, for instance, in [11][13] that this scheme
greatly degrades the detection performance in the presence
of channel estimation errors for a multi-carrier system. Fur-
thermore in [14], authors show that under imperfect channel
estimation, the rates achieved by the mismatched detector are
significantly lower than the channel capacity.
1536-1276/08$25.00 c 2008 IEEE
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
2/12
5040 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008
Fig. 1. Transmitter architecture of the BICM OFDM system.
As an alternative to the aforementioned mismatched ap-
proach, Tarokh et al. in [15] and recently Taricco and Biglieriin [16], proposed an improved maximum likelihood (ML)
detection metric that mitigates the impact of imperfect CSIRand applied it to a space-time coded MIMO system. A similar
investigation was carried out in [17] for trellis-coded mod-
ulation in scalar channels. Very recently, in [18], the authors
derived the maximum possible outage capacity induced by the
channel estimation errors, in the context of Rayleigh channels
and a theoretical decoder. This was further developed in [19],
[20] to address the MIMO situation.
A practical concern for the design of a communication
system is to guarantee a prescribed target rate with small error
probability, irrespective of the accuracy achieved during chan-nel estimation. However, adopting a mismatched approach in
the described scenario raises two important questions : i) What
are the maximal achievable information rates of mismatched
decoding and how close does it perform with respect to therates provided by the best possible decoder in the presence of
channel estimation errors ? ii) What type of practical decoder
can improve the overall system performance under imperfect
CSIR ?
The objective of this paper is to address the above questions
for OFDM channels, estimated by a few number of trainingsequences. To this end, we propose a Bayesian framework
based on the a posteriori probability density function (pdf) ofthe perfect channel conditioned on its estimate. This general
framework enables us to formulate any detection problem by
considering the average over the channel estimations errors,
of the detectors cost function, that would be used if thechannel was perfectly known. In particular, we can recover
the metric of [16] by using our general framework in thecase of ML detection. This ML metric is first used for
deriving an improved maximum a posteriori (MAP) iterative
detector, a technique which is known to be very efficientwhen channel coding and bit interleaving are jointly used.
By modifying properly the soft-values at the output of the
MAP detector, we reduce the impact of channel uncertainty
on the decoder performance. In a second step, by using some
tools from information theory, we provide the expression of
the achievable information rates associated to the improved
and mismatched ML metric. Then, we characterize the limits
of reliable information rates achieved by the mismatched andimproved detectors in terms of outage rates, which is more
appropriate for quasi-static channels [21].
Our results may serve in the evaluation of the trade-off
between the required quality of service (in terms of BER and
achieved throughputs) and system parameters (e.g., training
power, transmission power, period of training, outage proba-bility) in the presence of channel estimation errors.
The remainder of this paper is organized as follows. Section
II describes the OFDM system under study and our main
assumptions concerning data transmission and channel esti-
mation. Section III introduces a general Bayesian frameworkfor improved detection in the presence of channel estimation
errors and then formulates the improved ML detection metric.In section IV, we propose the improved MAP detector in
the case of imperfect channel estimation. In section V, we
calculate the achievable outage rates of receivers based on the
improved and mismatched ML metric. Section VI illustrates
via simulations, the performance of the proposed receiver over
both realistic UWB and Rayleigh fading channel environ-
ments. Finally, section VII concludes the paper.
Notational conventions are as follows. IN represents an
(N N) identity matrix; CN denotes a complex Gaussian
distribution; Ex[.] refers to expectation with respect to therandom vector x; det{.}, |.|, and . denote matrix deter-minant, absolute value and vector norm, respectively; (.)T,(.) and (.) denote vector transpose, Hermitian transpose andconjugation, respectively.
I I . SYSTEM DESCRIPTION AND CHANNEL ESTIMATION
A. OFDM System Model
We consider a coded OFDM system with M subcarriersthrough a frequency-selective multipath fading channel, de-
scribed in discrete-time baseband equivalent form by the taps{hl}
L1l=0 . As depicted in Fig. 1, the binary data sequence
is encoded by a non-recursive non-systematic convolutional(NRNSC) code, before being interleaved by a pseudo-random
interleaver. The interleaved bits are gathered in subsequences
of B bits d1k, . . . , dBk and mapped to complex Mc-QAM
(Mc = 2B) symbols sk. At the receiver, after removing
the cyclic prefix (CP) and performing fast Fourier transform
(FFT), the the n-th received OFDM symbol (for brevity weomit n) can be written as [2]
y = Hd s + z, (1)
where the vectors y = [y0, . . . , yM1]T and s =
[s0, . . . , sM1]T respectively denote the received and trans-
mitted symbols, the noise vector z = [z0, . . . , zM1]T is
assumed to be zero-mean circularly symmetric complex Gaus-
sian (ZMCSCG) with distribution z CN(0, 2z IM), andHd diag(H) is the (M M) diagonal channel matrix withdiagonal elements given by vector H = [H0, . . . , H M1]
T,
where Hk =L1
l=0 hl ej2kl/M. The channel coefficients
are assumed to be constant during a frame of OFDM symbols
and change to new independent values from one frame to
another.
For convenience, we also define the diagonal matrix Sd
diag(s). In the following, when there is no confusion, we useH instead ofHd.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
3/12
SADOUGH and DUHAMEL: IMPROVED ITERAT IVE DETECTION AND ACHIEVED THROUGHPUTS OF OFDM SYSTEMS 5041
B. Pilot Based Channel Estimation
Consider the estimation of the k-th channel frequency coef-ficient Hk with N pilot symbols si gathered in the row vectorsk = [s0, . . . , sN1], (k = 0,...,Mdata 1) where Mdata isthe number of data subcarriers in the OFDM symbol. In a
practical OFDM system, Mdata is generally smaller than Mdue to the presence of some unmodulated (also called virtual)
subcarriers. In the following, without loss of generality, weassume that Mdata is equal to M.
According to the observation model (1), during a givenchannel training interval, we receive
yk = Hk sk + zk for k = 0,...,M 1, (2)
where the entries of the noise vector zk have the samedistribution as those of z in (1). Moreover, the definition of
yk and zk follow that of sk.The average power ET of the k-th training vector sk is
ET 1
Nsk
2. (3)
Here, we assume equi-powered training vectors for all sub-
carriers. The least-squares estimate of Hk is obtained byminimizing ykHk sk2 with respect to Hk which coincideshere with the ML estimate. This yieldsHMLk = yk sk sksk1
= Hk + Ek for k = 0,...,M 1 (4)
where Ek = zksk
sks
k
1is the channel estimation error
term. For convenience, the samples in (4) can be written in
vector form as
HML = H+ E (5)
where E is a zero-mean Gaussian vector with a covariance
matrix given by
E = EEE
= 2E IM , where 2E =
2zN ET
(6)
Thus, from (5), the conditional pdf of HML given H readsp(HML|H) = CNH,E. (7)
Consider an uncorrelated i.i.d. Rayleigh channel with a prior
distribution H CN(0,H), where H = 2hIM. By usingthe latter pdf and the pdf of(H
ML|H) from (7), we can derive
the posterior distribution of the perfect channel, conditionedon its ML estimate as follows (see Appendix I.A):
p(H|HML) = CN(HML, E), (8)where
= H(E +H)1 = IM and =
2h(2h +
2E)
.
(9)
The availability of the estimation error distribution constitutes
an interesting feature of pilot assisted channel estimation
which is exploited in the next section, resulting in an improved
detector under imperfect channel estimation.For the sake of simplicity, we will not specify hereafter the
superscript ML for H.
III. DETECTOR DESIGN IN THE PRESENCE OF CHANNEL
ESTIMATION ERRORS
We now provide the formulation of a detection rule thattakes into account the available imperfect CSIR and refer to it
as the improved detector. To this end, we consider the model
(1) and denote by J(y, s,H) the quantity (cost function) thatwould let us to decide in favor of a particular s at the receiver
if the channel was perfectly known. Note that dependingon the detection criteria, the quantity J(y, s,H) can bethe posterior pdf p(s|y,H), the logarithm of the likelihoodfunction p(y|H, s), the mean square error (as in [22]), etc.Under a pilot-based channel estimation characterized by the
posterior pdf of the channel (8), we propose a detector based
on the minimization of a new cost function defined as
J(y, s, H) = H
J(y, s,H)p(H|H) dH= E
H|HJ(y, s,H)H (10)where we have averaged the cost function J over all re-
alizations of the unknown channel H conditioned on its
available estimate H by using the posterior distribution (8).Note that this approach is an alternative to the sub-optimal
mismatched detector, which is based on the minimization of
the cost function J(y, s, H). The latter is obtained by usingthe estimated channel H in the same metric that would beused if the channel was perfectly known, i.e., J(y, s,H). Theapproach in (10) differs from the mismatched detection on
the conditional expectation EH|H[.] which provides a robust
design by averaging the cost function J(y, s,H) over all(true) channel realizations which could correspond to the
available estimate.Consider the problem of detecting symbol s from the
observation model (1) in the ML sense, i.e., so as to maximizethe likelihood function p(y|H, s). The connection of theconsidered detection problem with the MAP detection of
BICM OFDM is discussed in the next section.
It is well known that under perfect channel knowledge and
i.i.d. Gaussian noise, detecting s by maximizing the likelihood
p(y|s,H) is equivalent to minimizing the Euclidean distanceDML as
sML(H) = arg m ins0,...,sM1C
DML(s,y,H)
, (11)
with DML(s,y,H) logp(y|H, s) y Hd s2,where means is proportional to and C denotes the setof constellation symbols.
The detection rule (11) requires the knowledge of the
perfect channel vector H. The sub-optimal mismatched ML
detector consists in replacing the exact channel by its estimateH in the receiver metric assMM(H) = arg min
s0,...,sM1C
DMM(s,y, H)
= arg m ins0,...,sM1C
y Hd s2, (12)
whereDMM(s,y,H) DML(s,y,H)
H=H. (13)
Obviously, the sub-optimality of this detection technique is
due to the mismatch introduced by the channel estimation
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
4/12
5042 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008
Fig. 2. Block diagram of the improved receiver under imperfect channel estimation.
errors; while the decision metric is derived from the likelihood
function p(y|H, s) conditioned on the perfect channelH, thereceiver uses an estimate H different from H in the detectionprocess.
As an alternative to this mismatched detection, an improved
ML detection metric is proposed in [15], [16]. This metric is
based on modified likelihood p(y|H, s) which is conditionedon the imperfect channel H. The pdfp(y|H, s) can be derivedas follows:
p(y|H, s) = HCM1
p(y,H|H, s) dH=
HCM1
p(y|H, s) p(H|H) dH= E
H|Hp(y|H, s) H, (14)where p(H|H) is the channel posterior distribution of equation(8) and C denotes the set of complex numbers. Actually,
equation (14) shows that p(y|H, s) can be simply derivedfrom the general formulation in (10) by replacing the cost
function J(y, s,H) by the likelihood p(y|s,H) that wouldbe used if the channel H was perfectly known.
By applying Theorem 1.2 of Appendix I.B, the averaged
likelihood in (14) is shown to be a complex Gaussian dis-
tributed vector given by
p(y|H, s) = CNmM,M, (15)where m
M= Hd s, and M = z + 2E SdSd.
Finally, the estimate of the symbol s is
sM
(H) = arg m ins0,...,sM1C
DM
(s,y, H), (16)where
DM
(s,y, H) log p(y|s, H)=
M1k=0
log
2z + 2E |sk|
2
+
yk Hk sk22z +
2E |sk|
2(17)
is referred to as the improved ML decision metric under
imperfect CSIR.
Note that when CSIR tends to the exact value, which is
obtained when the number of pilot symbols tends to infinity,
we have 1, 2E 0 and thus
limN
DM
s,y, HDMM
s,y, H = 1. (18)
Consequently, the improved metric of (17) tends to the mis-
matched metric, when the estimation errors tend to zero.
IV. ITERATIVE DETECTION OF BICM OFDM UNDER
IMPERFECT CHANNEL ESTIMATION
At the receiver, we perform MAP symbol detection and
channel decoding in an iterative manner as proposed for
instance in [23]. As shown in Fig. 2, the receiver mainlyconsists in the combination of two sub-blocks that exchange
soft information with each other. The first sub-block, referred
to as soft detector (also called demapper), produces softinformation in the form of extrinsic probabilities from the
input symbols and send it to the second sub-block which is a
soft-input soft-output (SISO) decoder. Each sub-block can take
advantage of the quantities provided by the other sub-block
as an a priori information. Here, SISO decoding is performedusing the well known forward-backward algorithm [24]. In the
following, we make use of the improved ML metric derived
in the previous section to modify the MAP detector part for
the case of imperfect CSIR.
Let dj,mk be the m-th (m = 1,...,B) coded and interleavedbit corresponding to the constellation symbol sk transmittedat the j-th time slot over the k-th subcarrier. We denote byL(dj,mk ) the coded log-likelihood ratio (LLR) of the bit d
j,mk
at the output of the detector. Conditioned on the imperfect
CSIR Hk, L(dj,mk ) is given byL(dj,mk ) = log
Pdem
dj,mk = 1 yk, Hk
Pdem
dj,mk = 0 yk, Hk , (19)
where Pdem(dj,mk
yk,
Hk) is the probability of transmission of
dj,m
k
at the detector output. We partition the set C that containsall possibly-transmitted symbol sk into two sets Cm0 and Cm1 ,for which the m-th bit of sk equals 0 or 1, respectively.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
5/12
SADOUGH and DUHAMEL: IMPROVED ITERAT IVE DETECTION AND ACHIEVED THROUGHPUTS OF OFDM SYSTEMS 5043
We have
L(dj,mk ) = log
sk Cm1
eDM(sk, yk,Hk) B
n=1
n=m
P1dec
dj,nk
sk Cm0
eDM(sk, yk, Hk) Bn=1
n=m
P0dec
dj,nk ,(20)
where P1dec(dj,nk ) and P0dec(dj,nk ) are prior probabilities com-ing from the SISO decoder.
Note that using the metric DM(sk, yk, Hk) for the eval-uation of the LLRs in (20) is an alternative to using the
mismatched ML metric DMM(sk, yk, Hk) which replaces ateach iteration, the exact channel Hk by its estimate Hk inDML(sk, yk, Hk). By doing so, the LLRs are adapted tothe imperfect channel knowledge available at the receiver
and consequently the impact of channel uncertainty on the
SISO decoder performance is reduced. We refer to the latter
approach as improved MAP detector.
The summations in (20) are taken over the product of the
likelihood p(yk|sk, Hk) = eDM(sk,yk,Hk) given a symbolsk and the estimated channel coefficient Hk, and of the a
priori probability on sk (the term
Pdec), fed back from theSISO decoder at the previous iteration. In this latter term, thea priori probability of the bit dj,mk itself has been excluded,so as to let the exchange of extrinsic informations between
the channel decoder and the soft detector. Also, note that this
term assumes independent coded bits dj,nk , which is a validapproximation for random interleaving of large size. At the
first iteration, no a priori information is available on bits dj,nk ,therefore the probabilities P0dec(d
j,nk ) and P
1dec(d
j,nk ) are set
to 1/2. The decoder accepts the LLRs of all coded bits and
computes the LLRs of information bits, which are used fordecision, at the last iteration.
V. ACHIEVABLE INFORMATION RATES ASSOCIATED TOIMPROVED AND MISMATCHED DETECTION METRICS
A. Instantaneous Achievable Information Rates Of OFDM
In [25], the authors derive the mismatched capacity of achannel, i.e., the highest rate at which reliable communication
is possible over the channel with a given sub-optimal decodingmetric. Obviously, these achievable rates are lower than the
channel capacity defined by Shannon which assumes a perfect
channel knowldege and optimal decoders at the receiver.Here, we consider an OFDM transmission over the frequency
selective channel H of which the receiver knows an imperfect
estimate H obtained by some training symbols. Our aim is tocalculate the achievable rates CM and CMM associated to anOFDM receiver based respectively on the improved and the
mismatched detection rules of (17) and (13). The following
theorem follows from the direct application of Theorem 1 in
[25] to an OFDM transmission where the receiver uses a givendetection rule D. This transmission is characterized by thelikelihood p(y|s,H) =
M1k=0 CN(Hksk,
2z) which defines
a mapping from the input symbols s S with distributionp(s), to the set of probability measures on the output alphabety Y.
Theorem 1: The highest achievable rate of the transmission
characterized by p(y|s,H) with a receiver using the sub-optimal detection metric D(s,y, H) is given by
CD = maxp(s)
minf
Ip(s), f
, (21)
where the maximization is performed over all probability
distributions p(s) on S, and
Ip(s), f = p(s)f(y|s, )log2
f(y|s, )p(s)f(y|s, )ds
dy ds, (22)
denotes the mutual information functional [26]; in (21),the minimization is over all channels f(y|s, ) =M1
k=0 CN(ksk, 2k) on S Y satisfying the following con-
straints
(c1) : Es
f(y|s, )
= Esp(y|s,H)
, (23)
(c2,k) : Esk
E{yk:f}
D(sk, yk,
Hk)
Esk E{yk:p}D(sk, yk, Hk), (24)for every k = 0, . . . , M 1, where = [0,...,M1]
T;
E{yk :f}[.] and E{yk:p}[.] refer to conditional expectation withrespect to yk by using the pdff(yk|sk, k) and p(yk|sk, Hk),respectively.
1) Case of Improved ML Metric: In the following, we aim
at solving the above constrained minimization problem for
our specific OFDM channel and metric D = DM of (17). Tothis end, we assume that the transmitter has no information
about the channel estimate and consequently uniform power
allocation is done across subcarriers. Furthermore, we assume
a Gaussian i.i.d. input distribution p(s) = CN(0, PIM).Under these conditions, the mutual information (22) of an
OFDM transmission, averaged over all subcarriers is given
by [27]
Ip(s), f
=
1
Mlog2 det
IM + Pd
d
1
=1
M
M1k=0
log2
1 +
P|k|2
2k
, (25)
where = diag([20 ,...,2M1]) and d = diag(). Accord-
ing to Theorem 1, we have to find the covariance matrix
and the optimal channel vector opt (these two characterizethe pdf f(y|s, )), so as to minimize the mutual information(25) and to satisfy constrains (23) and (24).
The unknown diagonal covariance matrix is obtained
from the constraint (c1) of (23). It is easily seen that thisconstraint leads to the equality f(y|) = p(y|H). Moreover,from the likelihoods f(y|s, ) and p(y|s,H), it is clearthat the conditional random vectors (y|) and (y|H) aredistributed as CN(0,+Pd
d) and CN(0,z+PHdH
d),
respectively. Thus the diagonal covariance matrix is ob-
tained as
= P(HdHd d
d) + z, (26)
with its k-th (k = 0,...,M 1) diagonal element equal to2k = P(|Hk|
2 |k|2) + 2z . (27)
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
6/12
5044 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008
Now, in order to find the optimal vector opt, we have tospecify the set of inequality constraints (c2,k) associated to
the metric DM(sk, yk, Hk) of (17). As shown in AppendixII.A, the k-th inequality constraint set in (24) is equivalent tothe set
VM,k =
k : |k a
Hk|
2 |Hk a
Hk|
2
, (28)
where a = ( 2
z
P 2
E )/( 2
z
P 2
E(1 ) ); =exp
2zP 2
E
E1 2zP 2
E
and E1(x)
+x
exp{u}u du is
defined as the exponential integral.
By using (25), (27) and (28) in Theorem 1, the achievable
rates CM(H, H) associated to the improved metric DM, canbe obtained by solving the following minimization, which is
equivalent to the initial formulation (21)
CM(H, H) =
min
1
M
M1k=0
log2
1 +
P|k|2
2k(|k|)
subject to |k a Hk|2 |Hk a Hk|
2
(29)
for k = 0,...,M 1. Note that the k-th term in (29)is an increasing function of |k|2 and the constraints areconvex. Consequently, the minimum in (29) is obtained at the
border of the sets [28]. Thus, the problem simplifies to thesearch of the optimal channel coefficients k = opt,k fromthe minimization of 2 under the constraint sets resultingfrom the equality in (28). This becomes a classical convex
minimization problem that can be solved by using Lagrange
multipliers. After some algebra provided in Appendix II.B, weget opt = [opt,0,...,opt,M1]
T, where
opt,k = M,kHk, k = 0, . . . , M 1, (30)and
M,k
=
a
|Hk a Hk||Hk| , if a 0
a +|Hk a Hk|
|Hk| , if a < 0.(31)
Finally, the achievable rates associated to metric DM are givenas follows.
CM(H
, H) = 1MM1
k=0 log21 +P 2
M,k|Hk|
2
2z + P(|Hk|2 2M,k |Hk|2) .(32)
2) Case of Mismatched ML Metric: Under the same as-
sumptions as above, we can compute the achievable rates
CMM associated to the mismatched metric DMM of (13) in thesame way. The equality constraint (23) leads to the same result
as equation (27). By following similar steps as in Appendix
II, we can easily derive the k-th inequality constraint setassociated to the metric DMM as
VMM,k
=
k : Re
k
Hk
Re
Hk
Hk
(33)
for k = 0,...,M 1, where Re(.) takes the real part of itsargument. Now, by considering the minimization problem of
equation (29) under the constraints of (33), and using the
Lagrange method, we obtain
CMM(H, H) =1
M
M1k=0
log2
1 +
P 2MM,k
|Hk|22z + P(|Hk|
2 2MM,k
|
Hk|2)
, (34)
where MM,k is given by
MM,k
=Re(Hk
Hk)|Hk|2 . (35)
In order to get insight about the expressions (32) and (34),
assume that perfect CSIR is available at the receiver, i.e.,
Hk = Hk for k = 0,...,M 1. In this situation, 2E = 0, = 1, and we get
MM,k=
M,k= 1 for all k. Consequently,
both detectors provide the instantaneous rate
C(H) =1
M
M1
k=0 log21 +P |Hk|2
2z . (36)
In this case, assuming that there is no delay constraint and the
transmission time is longer than the channel coherence time,
both of the detectors achieve the ergodic capacity [29] which
is defined as
Cerg = EH
C(H)
. (37)
B. Evaluation Of Outage Rates Under Imperfect Channel
Estimation
In several applications in wireless communications, the
channel is chosen randomly at the beginning of the transmis-sion and is held fixed during the whole transmission session.In fact, in this case the transmission time is not longer than
the channel coherence time and consequently the mean of the
maximum mutual information in (37) is in general not equal
to the channel capacity. On the other hand, the capacity in
the Shannon sense does not exist since there is a non-zero
probability that the realized H and its estimate H are notcapable of supporting even a very small rate. In this case, the
capacity is a random entity, as it depends on the instantaneous
channel H. Thus, the concept of capacity-verus-outage [21]
has to be invoked. That is, with any given rate R and channelrealization H, we associate a set of channels (R,H). This setis the largest possible set for which C(H), i.e., the achievedrate for a realization ofH (R,H) satisfies C(H) < R.Formally, we can write (R,H) {H : C(H) < R}.The outage (or failure) probability Pout is then determinedby Pout = Prob
H (R, H)
. This is the definition
of the classical notion of outage capacity, which implicitly
assumes the true channel to be available at the decoder. In
contrast, when this knowledge is not available at the receiver,
but only the channel estimate, we have to invoke the notion
of estimation-induced outage capacity, which has different
properties, as explained in [18].
Following this approach, we make use of the posteriorchannel distribution of (8) that characterizes our channel
estimation process, in order to define the outage probability
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
7/12
SADOUGH and DUHAMEL: IMPROVED ITERAT IVE DETECTION AND ACHIEVED THROUGHPUTS OF OFDM SYSTEMS 5045
PoutM
(associated to the metric DM
), for an outage rate R 0
and an estimated channel H as [18]PoutM
(R, H) = {H
M(R,H)}p(H|H) dH , (38)
where M
(R,
H)
H : CM(H,
H) < R
. Using this, the
outage rate for an outage probability equal to is given by
CoutM (, H) = supRR 0 : PoutM (R, H) . (39)
The outage rates (39) depend on the random channel estimateH. Thus, we consider as performance measure, the averagedoutage rates over all channel estimates as follows.
CoutM () = EHCoutM (, H). (40)
By using (34) and similar steps as above, we can derive the
averaged outage rates CoutMM() associated to the mismatched
receiver. The outage rates associated to the best possible (but
not practical) decoder under imperfect channel estimation is
derived in [18]. The next section evaluates the outage rates
of the improved and mismatched receivers compared to thoseprovided by this decoder, to which we refer as theoreticaldecoder. In our case, the averaged outage rate of the theoretical
decoder is given by
Cout
T() = EHCoutT (, H), (41)
where the outage rates CoutT
(, H) are computed by using theachievable rates
CT
(H, H) = 1M
M1k=0
log2
1 +
P |Hk|2
2z
(42)
in expressions similar to (38) and (39).
VI . SIMULATION RESULTS
In this section, we provide numerical results to evaluate the
performance provided by the proposed detector in the presence
of channel estimation errors, in comparison with more classi-
cal approaches. We focus on the impact of imperfect channel
estimation on receiver bit error rate (BER) and the achievable
outage rates associated to improved and mismatched detectors.
We consider the multiband OFDM (MB-OFDM) scenario
proposed in the IEEE 802.15.3a WPAN proposal [30] for
UWB applications operating in the range of 3.1 to 10.6 GHz.
In this scheme, the available spectrum is divided into severalsubbands with a conventional OFDM modulation within eachsubband. The only difference is that a frequency hopping
pattern selects the subband over which each OFDM symbol
is sent. Throughout the simulations, one OFDM symbol is
composed of M = 100 data subcarriers and the subbandchannel bandwidth is 528 MHz according to the parametersdefined in [30]. An MB-OFDM system employing only the
first three subbands starting at 3.1 GHz has been considered.For channel coding in our BICM scheme, we consider the
rate 1/2 NRNSC code of constraint length 3 defined in octalform by (5, 7)8. The interleaver is pseudo-random and oper-ates over the entire frame. Data symbols belong to 16-QAMconstellation with the Gray or the set-partition (SP) labeling
specified in [31], and the impact of the labeling is discussed
0 2 4 6 8 10 12 1410
6
105
104
103
102
101
100
Eb
/ N0
(dB)
BER
Mismatched N = 1
Improved N = 1
Mismatched N = 2
Improved N = 2
Mismatched N = 8
Improved N = 8
Perfect CSIR
Fig. 3. BER performance of improved and mismatched MAP detectorover the CM1 channel, training sequence lengths N {1, 2, 8}, 16-QAMmodulation with Gray labeling.
below, since it may enlarge or reduce the improvement brought
by the improved metric. However, note that the considered
labelings are not necessarily the optimal choices that may exist
for a 16-QAM constellation (see [32] and references therein
for the optimization of square QAM labeling).
The performance evaluation is performed over two chan-
nels: i) the realistic UWB channel model CM1 specified in
[33] and ii) the uncorrelated Rayleigh fading channel. For
each frame that contains 100 OFDM symbols, a differentrealization of the channel has been drawn and remains constant
during the whole frame. We use training sequences for channel
estimation and the average pilot-symbol power is equal tothe average data-symbol power. Moreover, the number of
decoding iterations is set to 4.
A. Bit Error Rate Analysis
First, the BER performance of the improved and mis-
matched detectors are compared. Let us first address the
case of BICM iterative decoding with 16-QAM and Gray
labeling. It can be observed from Fig. 3 that for N = 1(the shortest possible training sequence), the improvement in
terms of required Eb/N0 in order to attain a given BER isabout 2 dB, compared to the mismatched situation, whilethe perfect channel knowledge would even be 3 dB better.Obviously, it is also observed that these quantities are reduced
(the performance of both the mismatched decoder and the
improved one get closer to the perfect channel one) when
increasing the length of the training sequence. Note that the
performance of the improved receiver with 1 pilot is veryclose to that of the mismatched receiver with 2 pilots. Forcomparison, results obtained with uncorrelated Rayleigh block
fading channel are illustrated in Fig. 4. It can be observed
that for N = 2, the SNR necessary to obtain a BER of 104
is reduced by about 1 dB if the improved MAP detector isused instead of the mismatched detector. The conclusions are
otherwise quite similar on the Rayleigh channel.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
8/12
5046 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008
0 2 4 6 8 10 12 1410
6
105
104
103
102
101
100
Eb
/ N0
BER
Mismatched N = 1Improved N = 1
Mismatched N = 2
Improved N = 2
Mismatched N = 8
Improved N = 8
Perfect CSIR
Fig. 4. BER performance of the improved and mismatched MAP detectorover i.i.d. Rayleigh fading channel, training sequence lengths N {1, 2, 8},16-QAM modulation with Gray labeling.
0 2 4 6 8 10 12 1410
6
105
104
103
102
101
100
Eb
/ N0
(dB)
BER
Mismatched N = 1
Improved N = 1
Mismatched N = 2
Improved N = 2
Mismatched N = 8
Improved N = 8
Perfect CSIR
Fig. 5. BER performance of the improved and mismatched MAP detectorover the CM1 channel, training sequence lengths N {1, 2, 8}, 16-QAMmodulation with set-partition labeling.
Similar plots are shown in Fig. 5 for the case of 16-
QAM and SP labeling on the CM1 channel. These show the
behavior of the detectors with respect to the type of labeling.A first observation is that the distance between the mismatcheddetector and the perfect channel knowledge is even larger
(about 6 dB for N = 1 and a BER of 7 104). Anotherobservation is that, even if the global performance is largely
improved by using the SP labeling when perfect channel
knowledge is available, the difference between SP and Gray
labeling is not very large with the mismatched decoder, i.e., the
sensitivity of the iterative decoder to the channel knowledge
seems to be larger for SP labeling. However, the use of the
improved metric allows to recover most of the improvement,
since, even with a single training sample (N = 1), the BER isimproved with SP compared to Gray labeling by about 0.9 dBfor a BER of5104. In other words, iterative decoding withSP labeling benefits more from the improved metric than the
1 2 4 6 8 10 12 14 1610
5
104
103
102
101
N (number of pilot symbols per frame)
BER
Mismatched
Improved
Perfect CSI
Eb
/ N0
= 12 dB
Fig. 6. Reduction of the number of training sequence at Eb/N0 = 12 dBover the CM1 channel, 16-QAM modulation with Gray labeling .
1 2 4 6 8 10 12 1410
6
105
104
103
102
N (number of pilot symbols per frame)
BER
Mismatched
Improved
Perfect CSIR
Eb
/ N0
= 12 dB
Fig. 7. Reduction of the number of training sequence at Eb/N0 = 12 dBover the CM1 channel, 16-QAM modulation with set-partition labeling.
one with Gray labeling. Otherwise, similar conclusions hold
between the SP-labeling curves.
Figures 6 and 7 show the BER performance versus the
number of pilot symbols N at a fixed Eb/N0 of12 dB for 16-
QAM with Gray and set-partition labeling, respectively. Thisallows to evaluate the length of training sequence necessaryto achieve a certain BER. From Fig. 6 we observe that
the improved detector requires 10 pilot symbols per frameto achieve a BER of 104 at Eb/N0 = 12 dB whilethe mismatched detector attains this BER for 12 trainingsymbols. From Fig. 7 we notice that the performance loss
due to the mismatched receiver with respect to the improved
receiver becomes insignificant for N 12 (about 11 % ofthe overall frame of pilot and data symbols). Actually, our
results illustrate that the improved detector outperforms the
mismatched detector especially when a few number of pilot
symbols is dedicated to channel estimation, which is in perfectagreement with equation (18). This makes the improved metric
particularly useful, since training sequences in multicarrier
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
9/12
SADOUGH and DUHAMEL: IMPROVED ITERAT IVE DETECTION AND ACHIEVED THROUGHPUTS OF OFDM SYSTEMS 5047
0 5 10 15 20 250
1
2
3
4
5
6
7
8
SNR (dB)
ExpectedOutage
Rate(bits/channeluse)
Ergodic Capacity
Theoretical
Improved
Mismatched
Fig. 8. Expected outage rates of MB-OFDM transmission versus SNR forN = 1 pilot per frame and different detection approaches, outage probability= 0.01, i.i.d. Rayleigh fading channel.
0 5 10 15 20 250
1
2
3
4
5
6
7
SNR (dB)
Exp
ectedOutageRates(bits/channeluse)
Ergodic Capacity
Theoretical
Improved
Mismatched
Fig. 9. Expected outage rates of MB-OFDM transmission versus SNR forN = 2 pilots per frame and different detection approaches, outage probability= 0.01, i.i.d. Rayleigh fading channel.
systems are usually of length 1 or 2. Note also that thisimprovement is obtained almost at no computational cost.
B. Achievable Outage Rates
We now analyze the achievable outage rates provided by
a receiver based on either the improved or the mismatched
detection technique for MB-OFDM under imperfect CSIR.
The channel is an uncorrelated i.i.d. Rayleigh fading and
the data symbols are assumed to be distributed as CN(0, 1).Figure 8 shows the expected outage rates (in bits per channel-
use) versus the SNR (in dB), obtained by adopting mismatched
and improved detection approaches. The outage probability
has been fixed to = 0.01 and the channel is estimated bysending N = 1 pilot per frame. For comparison, we alsodisplay the upper bound on the expected outage rates provided
by the theoretical decoder and also the ergodic capacity, given
1 4 8 12 16 201
1.5
2
2.5
3
3.5
4
4.5
5
N (number of pilot symbols per frame)
ExpectedOutage
Rates(bits/channeluse)
Ergodic Capacity
Theoretical
Improved
Mismatched
SNR = 15 dB
SNR = 10 dB
Fig. 10. Expected outage rates of MB-OFDM transmission versus the numberof training sequence for different detection approaches at SNR of 10 and 15dB, outage probability = 0.01, i.i.d. Rayleigh fading channel.
by equations (41) and (37), respectively. The figure clearly
shows the sub-optimality of mismatched detection in terms of
expected outage rates compared to the rates provided by the
theoretical decoder. It can be observed that the mismatchedoutage rate is about 4.8 dB (at a mean outage rates of 5 bits)of SNR far from the rates achieved by the theoretical decoder.
We note that by adopting the improved receiver, the above
SNR gap is reduced by about 1.8 dB.Similarly, Fig. 9 shows the outage rates obtained by the
improved and mismatched receivers when the number of
channel-uses for pilot transmission is N = 2 per frame. As
observed, due to a more accurate channel estimation, boththe detectors achieve a higher outage rate as compared to that
obtained with N = 1. However, we note that at a mean outagerate of5 bits, the rate of the mismatched receiver is still about2.8 dB of SNR far from the rate of the theoretical receiver.The figure shows that the improved detector achieves higher
rates and decreases the required SNR by about 1 dB at a meanoutage rate of 5 bits. An open question is still how the gap
to the optimal estimation induced outage capacities could be
filled.
Finally, Fig. 10 compares the achievable rates of the im-proved and mismatched receivers with respect to the number
of training symbols at fixed SNR values of 10 and 15 dB. Ineach case, we also display the corresponding upper bound
on the achievable outage rates provided by the theoretical
decoder and the ergodic capacity. As can be seen, the improved
detector requires fewer pilot symbols in order to provide a
prescribed mean outage rate. For instance, the gain in the
number of channel-uses for pilot transmission is 2 pilots ata mean outage rates of 3.5 bits. These simulation resultsalso confirm the obvious expectation that under near perfect
channel estimation (here for N 16), the performance ofmismatched and improved receivers is almost the same.
VII. CONCLUSION
The problem of signal detection in a practical coded OFDM
communication system where the receiver has only access to
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
10/12
5048 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008
a noisy estimate of the channel provided by pilot symbols
was investigated. Based on a characterization of the channel
estimation process, we proposed a general detector designthat takes into account the imperfect channel available at the
receiver. In the case of ML detection, this approach led to an
improved ML metric that we used in the derivation of a mod-
ified iterative MAP detector. We also derived the expressions
of the achievable outage rates associated to the improved and
mismatched ML metrics. Our numerical results indicated thatthe mismatched detector is sub-optimal in terms of BER and
achievable outage rates, especially for short training sequence
lengths. They also confirmed the adequacy of the improved
detector under imperfect channel estimation. The impact of
the improved metric was shown to be dependent on the typeof labeling used in the construction of the symbols, and it
was shown that this new metric could save the improvementsbrought by set-partitioning labeling in the case of poor channel
estimates (obtained by very short training sequences). This
performance improvement was obtained without requiring
additional complexity in the receiver. Although the improved
detector was able to reduce the SNR gap between the mis-matched and the theoretical receiver at fixed outage rates, thederivation of other practical detectors providing closer rates to
the theoretical capacity limit is still an open problem.
APPENDIX I
A. Derivation of the A Posteriori Probability (8)
The following theorem is derived in [34].
Theorem 1.1: Let x1 and x2 be circularly symmetric com-
plex Gaussian random vectors with zero means and full-rankcovariance matrices ij = E{xix
j}. Then the conditional
random vector x1
|x2
CN(,) is a circularly symmetriccomplex Gaussian random vector with mean = 12
122 x2
and covariance matrix = 12122 21.
We set x1 = H and x2 = H. From equation (7) and theassumption H CN(0,H), we have 11 = 12 = Hand 22 = H + E in Theorem 1.1. According to thistheorem, the conditional random vector H|H has a circularlysymmetric complex Gaussian distribution with
mean = H where H(H+E)1, and (43)covariance matrix = HH(H+E)
1H = E .
(44)
The equivalence in (44) can be seen by left multiplying both
sides of (E +H) H = E by H(E +H)
1.
As a result, we obtain the a posteriori pdf (8).
B. Evaluation of the Likelihood Function (14)
In order to evaluate the expectation in (14), we use to the
following theorem derived in [35].
Theorem 1.2: For a circularly symmetric complex random
vector u CN(m,) with mean m = E[u] and covariancematrix = E[uu] mm, and a Hermitian matrix A suchthat I+A > 0, we have
Eu exp uAu = expmA(I +A)1mdet{I +A} .(45)
Let us define u = y Hds. Using the a posteriori distri-bution of (8) and after some algebra, we can derive the condi-
tional pdf ofu given s and Hd as u|(s, Hd) CN(mu,u),where mu = yHds and u = ESdSd. We furtherlet A = 1z . By applying Theorem 1.2, we get (14) as
p(y|
H, s) = E
Hd|Hd
exp (y Hds)
z
1(y Hds)
det
z
= 1det
z
I +ESdS
d
1z
exp
y Hds1z
I +ESdSd
1z
1y Hds (46)
Since z, and E are diagonal matrices, the latter
equation is rewritten as
p(y|H, s) = 1det
z + 2ESdS
d
exp y Hdsz + 2ESdSd1y Hds
(47)
= CN
Hds , z + 2ESdSd.
APPENDIX II
A. Details On The Derivation Of The Inequality Constraint
(28)
Using the expression of the modified metric DM from(17), the left-hand side of the k-th constraint in (24) can beexpanded as follows (the index k is omitted for notational
brevity).Es
E{y:f}
DM(s,y, H) = Es log (2z + 2E |s|2)
K
+
Es E{y:f}
|y|2 2Re
y Hs+ 2|H|2|s|2
2z + 2E |s|
2
= K+ Es
E{y:f}[|y|
2] 2ReE{y:f}[y
] Hs+ 2|H|2|s|22z +
2E |s|
2
= K+ Es
2 + ||2|s|2 2Re
H|s|2 + 2|H|2|s|2
2z + 2E |s|2
= K+ Es
2 + |
H|2|s|2
2z + 2E |s|
2 . (48)
Similarly, the right-hand side of (24) can be obtained as
Es
E{y:p}
DM(s,y, H) = K+Es2z + |H H|2|s|2
2z + 2E |s|
2
.
(49)
Using (48) and (49) in (24), the inequality constraint is written
as
Es
2 + | H|2|s|2
2z + 2E |s|
2
Es
2z + |H H|2|s|2
2z + 2E |s|
2
.
(50)
Now, in order to evaluate the expectation in (50), we introduce
the following lemma.
Lemma 1: Assume s C N(0, ) and x = |s|2 be a centeredChi-squared random variable with two degrees of freedom
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
11/12
SADOUGH and DUHAMEL: IMPROVED ITERAT IVE DETECTION AND ACHIEVED THROUGHPUTS OF OFDM SYSTEMS 5049
with pdf p(x) = 1 ex/ , and let a, b, c and d be real and
positive scalars. We have
Ex
a + b x
c + d x
=
b
d+
a
d
bc
d2
exp
c
d
E1
c
d
,
(51)
where E1()
exp{u}u du is the exponential integral.
Proof: The proof is easy and thus omitted for brevity.
By using Lemma 1 and replacing a,b,c and d by theirrespective values from (50), we get
| H|22E
+
2
(2E)P
| H|22z(1 2)P
|H H|22E
+
2z
2E P
|H H|22z(1 2)P
(52)
where exp 2z2EP
E1 2z2EP
.
After some algebraic manipulations, (28) can be derived from
(52).
B. Derivation Of The Optimal Solutions in (30)
We have to find the optimal vector opt =[opt,0, . . . , opt,M1]
T that minimizes 2 =M1k=0 |k|
2 subject to M constraints given by
|k a Hk|2 |Hk a Hk|2 = 0, for k = 0, . . . , M 1.Obviously, this problem can be split into M independentconstraint minimization problems for finding each opt,kfrom the minimization of |k|2. According to the Lagrangemultipliers method, we define the Lagrangian Lk as:
Lk = |k|2 + lk
|k a
Hk|
2 |Hk a
Hk|
2
= 0 (53)
for k = 0, . . . , M 1, where lk is an unknown constant. Theoptimal value opt,k is obtained by solving
Lkk
= k + lk (k a Hk) = 0. (54)We get
k =lk
lk + 1a Hk (55)
where lk is fixed by introducing k from (55) into the initialconstraint as follows
|
lk
lk + 1 a Hk a Hk|2 = |Hk a Hk|2. (56)It can be easily verified that (56) leads to two solutions lk,1and lk,2 for lk such that
lk,1
lk,1 + 1= 1
|Hk a Hk||a| |Hk| and lk,2lk,2 + 1 = 1 + |Hk a
Hk||a| |Hk| .
Clearly, the acceptable solution is lk = lk,1 since it minimizes|k|2 in (55).As a result, from (55) we obtain
opt,k = aa |Hk a
Hk|
|a| |Hk| Hk for k = 0, . . . , M 1,(57)which is the expression (30).
ACKNOWLEDGMENT
The authors would like to acknowledge many useful discus-
sions with P. Piantanida in which he provided insightful and
accurate comments.
REFERENCES
[1] S. B. Weinstein and P. M. Ebert, Data transmission by frequency-division multiplexing using the discrete Fourier transform, IEEE Trans.
Commun., vol. 19, pp. 628634, Oct. 1971.[2] R. Prasad, OFDM for Wireless Communications Systems. Artech House
Publishers, 2004.[3] G. Caire, G. Taricco, and E. Biglieri, Bit-interleaved coded modula-
tion, IEEE Trans. Inform. Theory, vol. 44, pp. 927945, May 1998.[4] J. K. Cavers, An analysis of pilot symbol assisted modulation for
Rayleigh fading channels, IEEE Trans. Veh. Technol., vol. 40, pp. 686693, Nov. 1991.
[5] J. G. Proakis, Digital Communications. McGraw-Hill, third edition,1995.
[6] E. Zehavi, 8-PSK trellis codes for a Rayleigh channel, IEEE Trans.Commun., vol. 40, pp. 873887, May 1992.
[7] X. Li, A. Chindapol, and J. A. Ritcey, Bit-interleaved coded modulationwith iterative decoding and 8PSK modulation, IEEE Trans. Commun.,vol. 50, pp. 12501257, Aug. 2002.
[8] Z. Yang and X. Wang, A sequential Monte Carlo blind receiver for
OFDM systems in frequency-selective fading channels, IEEE Trans.Signal Process., vol. 50, no. 2, pp. 271280, Feb. 2002.[9] Y. Huang and J. Ritcey, 16-QAM BICM-ID in fading channels with
imperfect channel state information, IEEE Trans. Commun., vol. 2, pp.10001007, Sept. 2003.
[10] P. Garg, R. K. Mallik, and H. M. Gupta, Performance analysis of space-time coding with imperfect channel estimation, IEEE Trans. WirelessCommun., vol. 4, pp. 257265, Jan. 2005.
[11] M. Speth, S. A. Fechel, G. Fock, and H. Meyr, Optimum receiverdesign for wireless broadband systems using OFDM-part I, IEEE Trans.Commun., vol. 47, no. 11, pp. 16681677, Nov. 1999.
[12] A. Leke and J. M. Cioffi, Impact of imperfect channel knowledgeon the performance of multicarrier systems, in Proc. IEEE GlobalTelecommun. Conf. (Globecom), pp. 951955, Nov. 1998.
[13] K. Ahmed, C. Tepedelenhoglu, and A. Spanias, Effect of channelestimation on pair-wise error probability in OFDM, in Proc. IEEE Int.Conf. Acoustics, Speech and Signal Processing (ICASSP), pp. 745748,May 2004.
[14] A. Lapidoth and S. Shamai, Fading channels: how perfect need perfectside information be ? IEEE Trans. Inform. Theory, vol. 48, pp. 11181134, May 2002.
[15] V. Tarokh, A. Naguib, N. Seshadri, and A. R. Calderbank, Space-timecodes for high data rate wireless communication: Performance criteria inthe presence of channel estimation errors, mobility, and multiple paths,
IEEE Trans. Commun., vol. 47, pp. 199207, Feb. 1999.[16] G. Taricco and E. Biglieri, Space-time decoding with imperfect channel
estimation, IEEE Trans. Wireless Commun., vol. 4, no. 4, pp. 18741888, July 2005.
[17] J. K. Cavers and P. Ho, Analysis of the error performance of trellis-coded modulations in Rayleigh-fading channels, IEEE Trans. Commun.,vol. 40, no. 1, pp. 7483, Jan. 1992.
[18] P. Piantanida, G. Matz, and P. Duhamel, Estimation-induced outage ca-pacity of Ricean channels, in Proc. Signal Process. Advances Wireless
Commun. (SPAWC), July 2006.[19] S. M. S. Sadough, P. Piantanida, and P. Duhamel, MIMO-OFDM
optimal decoding and achievable information rates under imperfectchannel estimation, in Proc. Signal Process. Advances Wireless Com-mun. (SPAWC), June 2007.
[20] P. Piantanida, S. M. S. Sadough, and P. Duhamel, On the outagecapacity of a practical decoder using channel estimation accuracy, inProc. IEEE International Symp. Inform. Theory (ISIT), June 2007.
[21] E. Biglieri, J. Proakis, and S. Shamai, Fading channels: Information-theoretic and communications aspects, IEEE Trans. Inform. Theory,vol. 44, no. 6, pp. 26192692, Oct. 1998.
[22] S. M. S. Sadough and M. A. Khalighi, Optimal turbo-blast detectionof MIMO-OFDM systems with imperfect channel estimation, in Proc.
IEEE International Symposium on Personal, Indoor and Mobile Radio
Communications, pp. 16, Sept. 2007.[23] P. Magniez, B. Muquet, P. Duhamel, V. Buzenac, and M. de Courville,
Optimal decoding of bit-interleaved modulations: theoretical aspectsand practical algorithms, in Proc. Int. Symp. on Turbo Codes andrelated topics, pp. 169172, Sept. 2004.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:41 from IEEE Xplore. Restrictions apply.
8/14/2019 Improved Iterative Detection And
12/12
5050 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008
[24] L. Bahl, J. Cocke, F. Jelinek, and J. Raviv, Optimal decoding of linearcodes for minimizing symbol error rate, IEEE Trans. Inform. Theory,pp. 284287, Mar. 1974.
[25] N. Merhav, G. Kaplan, A. Lapidoth, and S. Shamai, On informationrates for mismatched decoders, IEEE Trans. Inform. Theory, vol. 40,pp. 19531967, Nov. 1994.
[26] T. Cover and J. Thomas, Elements of Information Theory. Wiley Seriesin Telecomunications. New York: Wiley, 1991.
[27] I. E. Telatar, Capacity of multi-antenna gaussian channels, AT&T BellLabs Tech. Memo., Tech. Rep., 1995.
[28] J. Hirriart-Urrty and C. Lemarchal, Convex Analysis and Minimization Algorithms I. Springer-Verlag, 1993.[29] L. Ozarow, S. Shamai, and A. Wyner, Information theoretic consider-
ations for cellular mobile radio, IEEE Trans. Inform. Theory, vol. 43,pp. 359378, May 1994.
[30] A. Batra, J. Balakrishnan, and A. Dabak, Multiband OFDM physicallayer proposal for IEEE 802.15 task group 3a, IEEE, Tech. Rep., July2003.
[31] B. Muquet, P. Magniez, P. Duhamel, M. de Courville, and G. Giannakis,Turbo demodulation of zero-padded OFDM transmissions, in Proc.
Asilomar Conf. on Signals, Systems and Computers, pp. 18151819,2000.
[32] Y. Huang and J. A. Ritcey, Optimal constellation labeling for iterativelydecoded bit-interleaved space-time coded modulation, IEEE Trans.
Inform. Theory, vol. 51, no. 5, pp. 18651871, May 2005.[33] J. Foerster, Channel modeling sub-committee report final,
IEEE802.15-02/490, Tech. Rep., 2003.[34] M. Bilodeau and D. Brenner, Theory of Multivariate Statistics. New
York: Springer, 1999.[35] M. Schwartz, W. R. Bennett, and S. Stein, Communications Systems and
Techniques. McGraw-Hill, 1966.
Seyed Mohammad Sajad Sadough (S04M08)was born in Paris in 1979. He received the B.Sc.degree in Electrical Engineering (electronic) fromShahid Beheshti University, Tehran, Iran, in 2002,the M.Sc. degree and the PhD degree in ElectricalEngineering (telecommunication), both from ParisSud 11 University, Orsay, France, in 2004 andJanuary 2008, respectively. During his PhD (200407), he has been with ENSTA (National Engineer-ing school in advanced techniques), Paris, Franceand Suplec-CNRS/LSS (Laboratory of Signals and
Systems), Gif-sur-Yvette, France. He has been a lecturer in the Electronics andComputer Engineering Department of ENSTA where his research activitieswere focused on improved reception schemes for Ultra Wideband commu-nication systems. From December 2007 to September 2008, he has been aPostdoctoral researcher at Suplec-CNRS/LSS, Gif-sur-Yvette, France, wherehe was involved in research projects with Alcatel-Lucent on satellite mobilecommunication systems.
Dr Sadough joined the Faculty of Electrical and Computer Engineering ofShahid Beheshti University, Tehran, Iran, in October 2008, as an AssistantProfessor in the Telecommunication Department. His current research interestsinclude Signal processing for wireless communications with particular empha-sis on multicarrier and MIMO systems, joint channel estimation and decoding,iterative reception schemes and interference cancellation under partial channelstate information.
Pierre Duhamel (F98) was born in France in1953. He received the Eng. Degree in ElectricalEngineering from the National Institute for AppliedSciences (INSA) Rennes, France in 1975, the Dr.Eng. Degree in 1978, and the Doctorat s sciencesdegree in 1986, both from Orsay University, Orsay,France.
From 1975 to 1980, he was with Thomson-CSF,Paris, France, where his research interests werein circuit theory and signal processing, including
digitalfi
ltering and analog fault diagnosis. In 1980,he joined the National Research Center in Telecommunications (CNET), Issyles Moulineaux, France, where his research activities were first concernedwith the design of recursive CCD filters. Later, he worked on fast algorithmsfor computing Fourier transforms and convolutions, and applied similartechniques to adaptive filtering, spectral analysis and wavelet transforms.From 1993 to Sept. 2000, he has been professor at ENST, Paris (NationalSchool of Engineering in Telecommunications) with research activities fo-cused on Signal processing for Communications. He was head of the Signaland Image processing Department from 1997 to 2000. He is now withCNRS/LSS (Laboratoire de Signaux et Systmes, Gif sur Yvette, France),where he is developing studies in Signal processing for communications(including equalization, iterative decoding, multicarrier systems, cooperation)and signal/image processing for multimedia applications, including sourcecoding, joint source/channel coding, watermarking, and audio processing. Heis currently investigating the application of recent information theory resultsto communication theory.
Dr. Duhamel was chairman of the DSP committee from 1996 to 1998, anda member of the SP for Com committee until 2001. He was an associateEditor of the IEEE T RANSACTIONS ON SIGNAL PROCESSING from 1989 to1991, an associate Editor for the IEEE S IGNAL PROCESSING LETTERS, anda guest editor for the special issue of the IEEE T RANSACTIONS ON SIGNALPROCESSING on wavelets.
He was Distiguished Lecturer, IEEE, for 1999, and was co-general chair ofthe 2001 International Workshop on Multimedia Signal Processing, Cannes,France. He was also co-technical chair of ICASSP 06, Toulouse, France.The paper on subspace-based methods for blind equalization, which he co-authored, received the Best paper award from the IEEE transactions on SPin 1998. He was awarded the Grand Prix France Telecom by the FrenchScience Academy in 2000.