2314 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO.

5, JUNE 2009

Approximate ML Decision-Feedback Block

Equalizer for Doubly Selective Fading Channels
Lingyang Song, Member, IEEE, Rodrigo Caiado de Lamare, Member, IEEE,
Are Hjørungnes, Senior Member, IEEE, and Alister G. Burr, Member, IEEE

Abstract—To effectively suppress intersymbol interference (ISI) many suboptimal low-complexity equalization techniques have
at low complexity, in this paper, we propose an approximate been proposed, such as the popular minimum-mean-square-
maximum-likelihood decision-feedback block equalizer (A-ML- error decision-feedback equalizer (MMSE-DFE), which is very
DFBE) for doubly selective (frequency- and time-selective) fading
channels. The proposed equalizer design makes efficient use of the effective in certain multipath environments and has a com-
special time-domain representation of multipath channels through plexity that is only dependent on forward and backward filter
a matched filter, a sliding window, a Gaussian approximation, and lengths [3]. However, there is a nonnegligible performance
a decision feedback. The A-ML-DFBE has the following features: loss of minimum mean square error (MMSE)-based equalizers,
1) It achieves a performance that is close that to that of maximum- compared with MLSE [4], [5].
likelihood sequence estimation (MLSE) and significantly outper-
forms minimum mean square error (MMSE)-based detectors. Further still, while much research have been conducted on
2) It has substantially lower complexity than conventional equal- time-domain equalization, few works take the special form of
izers. 3) It easily realizes complexity and performance tradeoff channel representation into good account. Two properties of
by adjusting the length of the sliding window. 4) It has a simple the channel matrix in time domain are effectively utilized in
and fixed-length feedback filter. The symbol error rate (SER) is this paper: 1) The Toeplitz-like channel matrix significantly
derived to characterize the behavior of the A-ML-DFBE and can
also be used to find the key parameters of the proposed equalizer. contributes to the equalizer design. 2) The large number of zero
In addition, we further prove that the A-ML-DFBE obtains full elements reduces the computational complexity. As a result, we
multipath diversity. propose a robust approximate maximum-likelihood decision-
Index Terms—Doubly selective fading channels, equalization, feedback block equalizer (A-ML-DFBE) to combat ISI over
linear minimum mean square error (MMSE), matched filter, doubly selective fading channels with low computational com-
maximum-likelihood sequence estimation (MLSE), MMSE plexity. The proposed equalizer exploits substantial benefit
decision-feedback equalizer (MMSE-DFE). from the special time-domain representation of the multipath
channels by using a matched filter, a sliding window, a Gaussian
I. I NTRODUCTION approximation, and a decision feedback. The main ideas are,
first, to subtract the effect of the already detected signals
W IRELESS communications often suffer from severe
intersymbol interference (ISI) due to doubly selective
fading. To suppress channel distortion, channel equalization
obtained from past decisions. This can be treated as a decision-
feedback process. Second, we apply Gaussian approximation
techniques are essential and, indeed, have received consider- [6]–[9] to realize near-maximum-likelihood (ML) detection.
able attention for many years. Maximum a priori probability The accuracy of this procedure can be improved by adjusting
(MAP) equalization is an optimum equalization procedure in the length of the sliding window using the central limit theorem.
terms of minimum symbol error rate (SER) [1] but requires Consequently, a complexity and performance tradeoff can be
a prohibitive computational complexity for many applications, realized, and a convergence in SER performance can also be
because it is exponential in channel length and constellation obtained by adjusting the length of the sliding window.
size. Maximum-likelihood sequence estimation (MLSE) can Note that [6] and [7] can only be used for frequency flat-
obtain a SER performance that is very close to that of MAP, fading channels and that [8] aims to recover signals for multi-
but its complexity is still extremely high [2]. As a result, user systems. Although a probabilistic-data-association-based
equalizer is reported in [9], there are several major differences
compared to the proposed approach: In [9], it is a requirement
Manuscript received January 31, 2008; revised June 22, 2008 and
September 3, 2008. First published November 21, 2008; current version
to update the mean and the variance for all detected symbols;
published May 11, 2009. This work was supported by the Research Council many iterations have to be used to make the performance
of Norway (NFR) under Project 176773/S10 called OptiMO. The review of converge; there is no feedback process; and no matched filter is
this paper was coordinated by Dr. X. Wang.
L. Song is with School of Electronics Engineering and Computer Science,
employed. In [10], a bidirectional arbitrated decision-feedback
Peking University, 100871, Beijing, China (e-mail: [email protected]). equalization (BAD) algorithm, which has a complexity of at
R. C. de Lamare and A. G. Burr are with the Department of Electronics, least twice that of the MMSE-DFE but can achieve better
University of York, YO 10 5DD York, U.K. (e-mail: [email protected];
[email protected]). performance, was presented. In [11], a class of block decision-
A. Hjørungnes is with the UniK-University Graduate Center, University of feedback equalizers (DFEs) is presented for frequency-domain
Oslo, 2007 Oslo, Norway (e-mail: [email protected]). equalization, but it assumes that the lengths of the channel,
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. forward filter, and backward filter are infinitely long, which is
Digital Object Identifier 10.1109/TVT.2008.2009758 not practical. In addition, it requires a large number of iterations

0018-9545/$25.00 © 2009 IEEE

Authorized licensed use limited to: DGIST. Downloaded on December 09,2023 at 13:30:20 UTC from IEEE Xplore. Restrictions apply.
SONG et al.: A-ML-DFBE FOR DOUBLY SELECTIVE FADING CHANNELS 2315

to make the performance converge, which increases the system they can also be decoded in complex form using standard zero-
delay and computational complexity. forcing or MMSE approaches, or linear or decision-feedback
The rest of this paper is organized as follows: In Section II, equalization [3].
we present the channel and signal models. The proposed
A-ML-DFBE scheme and complexity comparisons are dis-
III. D ESCRIPTION OF THE P ROPOSED M ETHOD
cussed in Section III. The performance is analyzed in
Section IV. Simulation results are presented in Section V. In A. A-ML-DFBE
Section VI, we draw the main conclusions. The proof is given
The proposed equalization algorithm can be summarized into
in the Appendix.
three steps: 1) forward process, which builds up the forward
Notation: Boldface upper-case letters denote matrices; bold-
filter using a temporal submatched filter; 2) decision-feedback
face lower-case letters denote vectors; Ci×j and Ri×j denote
process, which cancels the interference using a fixed-length
the set of i × j complex and real matrices, respectively; (·)T
backward filter, and 3) approximate ML process, which realizes
stands for transpose; (·)∗ denotes complex conjugate; (·)H
the final signal detection with the aid of Gaussian approxima-
represents conjugate transpose; Ii stands for an i × i identity
tion. The detailed description of each step is given here.
matrix; E denotes expectation; var represents variance; and
1) Forward Process: Supposing we start decoding sk , a
x2 = xH x.
temporal submatched filter (forward filter) is applied to (2), i.e.,

II. C HANNEL AND S IGNAL M ODELS HH H H

k r = Hk Hs + Hk n (3)
The doubly selective fading channel can be modeled using a where Hk is a matrix of size N × Lf , which comprises the
finite-impulse response (FIR) filter, i.e., entries in H, starting from the kth column to the (k + Lf −
1)th column and from the first row to the N th row. Lf (L ≤

L−1
H(z, t) = hk (t)z −k (1) Lf ≤ N ) is the length of the sliding window, which must be
k=0
equal to or larger than L for smaller ISI and larger diversity gain
and smaller than or equal to N . When Lf = N , the matched
where H(z, t) is the z transformation at time t, hi (t) is the filter becomes HH . For simplicity, we may rewrite (3) as
channel coefficient of the ith path, and L is the length of the FIR
filter. For simplicity, we only consider a single-input–single- yk = Js + nk (4)
output system. The received signals can be written in vector Lf ×1 Lf ×N
form as (for convenience, we drop the time index for each where yk = HH k r∈C , J = HH k H∈C , and nk =
Lf ×1
transmission frame) Hk n ∈ C
H
. We call this process horizontal slicing since it
takes Lf rows of H. J is given by
r = Hs + n (2) ⎡ ⎤
hHk h1 ··· hH
k hN
⎢ .. .. .. ⎥
where the received signals r = [r1 , . . . , rN +L−1 ]T , N is the J=⎣ . . . ⎦ (5)
length of s, transmitted signals s = [s1 , . . . , sN ]T , and n = hH
k+Lf −1 h1 ··· hH
k+Lf −1 hN
[n1 , . . . , nN +L−1 ]T , whose elements are independent samples
of a zero-mean complex Gaussian random variable with vari- where hi ∈ C(N +L−1)×1 denotes the ith column of matrix H.
ance σ 2 = E[|sk |2 ]/SNR, in which E[|sk |2 ] represents the aver- The length of the forward filter has been defined as Lf in (3).
age power of the transmitted symbols from constellation A. In 2) Decision-Feedback Process: The function of this step is
this paper, we set E[|sk |2 ] = 1. The time-domain representation to suppress the effects of the detected terms.
of the doubly selective fading channel H ∈ C(N +L−1)×N can To further decrease the complexity of (4), we can just con-
be written as sider a certain number of the transmitted symbols and have
⎡ ⎤
h1 (0) 0 0 ··· 0 yk ≈ Jk sk + nk (6)
⎢ .. ⎥
⎢ h2 (0) h1 (1) 0 ··· . ⎥
⎢ ⎥ where Jk ∈ CLf ×(k+Lf −1) can be constructed by taking the
⎢ .. ⎥
⎢ . h (1) h (2) · · · 0 ⎥ first column to the k + Lf − 1th column of J in (5), and sk =
⎢ 2 1 ⎥
⎢ . ⎥ [s1 , . . . , sk+Lf −1 ]T . We call this process as vertical slicing
H = ⎢ hL (0) .. h2 (2) · · · h1 (N − 1) ⎥ .
⎢ ⎥ since it takes k + Lf − 1 columns of J. Moreover, (6) can be
⎢ . ⎥
⎢ 0 hL (1) .. · · · h2 (N − 1) ⎥ decomposed with respect to each transmitted symbol
⎢ ⎥
⎢ . . . .. ⎥
⎣ .. .. .. ··· . ⎦ k+Lf −1

0 0 0 · · · hL (N − 1) yk ≈ ji si + nk (7)
i=1
Note that H has a structure that is similar to the Toeplitz
form, and some form of guard interval is necessary to avoid where ji ∈ CLf ×1 is the ith column of matrix Jk , and si
interblock interference between the received signals [5]. The is the ith transmitted symbol. Note that (7) is equivalent to
symbols in (2) can be recovered by MLSE [1]. Alternatively, (4) when the vertical slicing includes all the symbols in J;
Authorized licensed use limited to: DGIST. Downloaded on December 09,2023 at 13:30:20 UTC from IEEE Xplore. Restrictions apply.
2316 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 5, JUNE 2009

Lf = N + 1 − k, which implies that the length of Lf will have TABLE I

APPROXIMATE ML DECISION-FEEDBACK BLOCK
some effect on the system performance; and the effect of Lf EQUALIZATION ALGORITHM
will further be discussed in Sections IV and V. We can write
(7) as

k+Lf −1

k−1 
yk ≈ ji si + jk sk + ji si + nk (8)
i=1 i=k+1

where k−1 i=1 ji si denotes the detected terms that can be rebuilt
k+Lf −1
by past decisions, jk sk is the current target, and i=k+1 ji si
represents the undetected terms. The function of the feedback
process is to reconstruct k−1 i=1 ji si for later interference can-
cellation. Therefore, it is important to decide the length of the
backward filter Lb . Based on the expressions of H and (5),
we have j1 = j2 = · · · = jk−L−2 = 0; thus, the length of the
backward filter Lb can be fixed at L − 1, (L > 1) to reconstruct
the effects of past decisions. Equation (8) can be rewritten as
follows by simplifying the detected terms:

k+Lf −1

k−1 
yk ≈ ji si + jk sk + ji si + nk (9)
i=k−Lb i=k+1 where Jk+1 can be constructed using the (k + 1)th column as
k−1
the (k + Lf − 1)th column of Jk , and Jk can be obtained by
where Lb is equal to L − 1. As in (9), i=k−Lb ji si can taking the kth column to the (k + Lf − 1)th column of Jk .
be reconstructed from past decisions, and the following past According to the central limit theorem, the accuracy of the
decision cancellation process can be applied: Gaussian assumption can be improved by increasing the length
of the forward filter (sliding window) Lf .

k−1
As η k has an approximate Gaussian distribution, likelihood
yk = y k − ji si . (10)
function p(yk |sk ) is given by
i=k−Lb

The preceding process is very similar to the decision- p(yk |sk )∞ exp −(yk − jk sk )H Λ−1
k (yk − jk sk ) . (14)
feedback cancellation process, but, unlike MMSE-DFE, we do
not need to calculate the coefficients of the feedback filter. Finally, sk can be recovered by the following ML detector:
Moreover, Lb is fixed at L − 1, which means that only L − 1 
past decisions need to be fed back, which is much less than sk = arg min (yk − jk sk )H Λ−1
k (yk − jk sk ) . (15)
what is typically required by MMSE-DFE. sk ∈A

3) Approximate ML: This step aims to achieve near-optimal

detection by applying Gaussian approximation. We substitute At k = N − Lf + 1, there are no more new received signals
(9) into (10) and get outside the sliding window. Thus, we can then simply decode
each undetected symbol by treating the rest as Gaussian terms
k+Lf −1 and removing the effects of the detected symbols. This decod-

yk = jk sk + ji si + nk . (11) ing process is very similar to the case of k < N − Lf + 1 by
i=k+1 fixing the sliding window. The overall A-ML-DFBE algorithm
is summarized in Table I.
To decode sk with low computational complexity while
maintaining a performance comparable to that of the ML
k+Lf −1 B. Computational Complexity Analysis
decoder, we treat the undetected terms i=k+1 ji si and the
noise vector nk,k+Lf −1 together as a new complex-valued Before we show the complexity comparisons, we present
Gaussian vector with matching mean and covariance matrix, how to further reduce the proposed equalizer complexity.
k−1
such that (11) can be expressed as Note that, in (9), the detected terms i=k−Lb ji si can be
rewritten as Jk−1 sk−Lb with size Lf × 1, where sk−1 =
yk = jk sk + η k (12) [sk−Lb , . . . , sk−1 ]T is of size Lb × 1. With respect to the di-
agonal element hH g hg in J, when g > L, we can find that
where η k is a vector of zero-mean complex-valued Gaussian
random variables with size Lf × 1 and covariance hH i≥g+L
g hi = 0,
2 
Λk = Jk+1 JH
k+1 + σ Jk (13) hH
i hg = 0, i≤g−L (16)
Authorized licensed use limited to: DGIST. Downloaded on December 09,2023 at 13:30:20 UTC from IEEE Xplore. Restrictions apply.
SONG et al.: A-ML-DFBE FOR DOUBLY SELECTIVE FADING CHANNELS 2317

TABLE II that the matrix inversion lemma can be used to reduce the
COMPUTATIONAL COMPLEXITY OF VARIOUS SCHEMES FOR ONE SLIDING
WINDOW WITH LENGTH N , NUMBER OF PATHS L, BPSK complexity from cubic to quadratic order, but it does not affect
CONSTELLATIONS, AND LENGTHS OF THE FORWARD AND the aforementioned conclusions.
BACKWARD FILTERS Lf AND Lb , RESPECTIVELY

IV. P ERFORMANCE A NALYSIS

A. Analytical SER and BER Derivations
In this section, we analyze the SER and bit error rate (BER)
performances of the A-ML-DFBE. Note that tail detection only
comprises the operation of very few symbols; thus, the perfor-
mance is dominated by Step 2 of the A-ML-DFBE process in
Table I, which will now be analyzed. We assume that all the
decisions are accurate for analysis, which is a normal assump-
tion in decision-feedback theory [4]. In (12), which contains
−1/2
correlated noise η k , the prewhitening filter Ψk = Λk can
be applied to make the variance of the noise uncorrelated, i.e.,

and thus, Jk−1 has the following form: Ψk yk = Ψk jk sk + Ψk η k (17)

⎡ hH h ··· ··· hH ⎤
k k−Lb k hk−1 where Ψk η k with size Lf × 1 has a Gaussian distribution with
⎢ .. ⎥ zero mean, and all components have unit variance.
⎢ 0 hH
k+1 hk−Lb +1 ... . ⎥
⎢ ⎥ Since the noise now has become white Gaussian, the matched
⎢ .. .. .. .. ⎥
⎢ . . . . ⎥ filter (Ψk jk )H can be employed, and we have the following
Jk−1 = ⎢
⎢
⎥
⎢ 0 ··· 0 hk+Lb −1 hk−1 ⎥
H
⎥ received signal equation in scalar form:
⎢ 0 ··· ··· 0 ⎥
⎢ .. .. ⎥
⎣
.
..
.
..
. .
⎦ yk = ξk sk + υk (18)
0 ··· ··· 0
where yk = (Ψk jk )H Ψk yk , ξk = Ψk jk 2 , and υk =
which is of size Lf × Lb . We can observe that there are only (Ψk jk )H Ψk η k , which is a scalar with zero mean and variance
Lb Ψk jk 2 . The SER for M -ary phase-shift keying (PSK)
i=1 i = ((1 + Lb )/2)Lb nonzero elements in Jk−1 , so that
k−1 constellation is given by [13]
the reconstruction of the detected terms i=k−Lb ji si can
further be simplified. Similarly, in (15), the calculation of Λk (M −1)π
and jk can also be simplified. M
1 gpsk γk
Now, we discuss the complexity of the A-ML-DFBE, linear SERkM = exp − dθ (19)
π sin2 θ
MMSE [4], MMSE-DFE [4], and BAD [10] detectors in terms 0
of the number of additions and multiplications. The resulting
values are given in Table II, which were obtained by inspecting Δ Δ
where gpsk = sin2 (π/M ), γk = (|ξk sk |2 /var(υk )) = (ξk2 /
the relevant algorithms in Table I, [4], and [10]. Details of the ((Ψk jk )H (Ψk jk ))) = Ψk jk 2 , and M denotes the
computation of complexity, e.g., matrix inversion, can be found constellation size. The average BER for M − PSK can be
in [12]. The computational complexity of the A-ML-DFBE written as
algorithm is a function of frame length N , impulse response
N −Lf +2
length L, and the length of the forward filter Lf , which is 1 
obtained based on Table I. BERM = BERkM (20)
N − Lf + 2
From the table, we observe that the A-ML-DFBE has the k=1
same order of complexity as the linear MMSE and MMSE-
DFE. However, the A-ML-DFBE is less complex than the where BERkM ≈ (1/ log2 M )SERkM [1] for high signal-to-noise
MMSE-DFE since the A-ML-DFBE requires a smaller Lf ratio (SNR) and Gray mapping. Since the tail is normally short,
value and does not require building up the backward filter. which has a length of Lf − 2, compared with the whole frame
Compared with the linear MMSE, the A-ML-DFBE needs a length N , its effects can be neglected. Note that, in time-
N −L +2
relatively shorter forward filter and thus has lower complexity. invariant channels, SER1M = SER2M = · · · = SERM f due
The relation between the filter length and the performance can to the property of J (γ1 = γ2 = · · · = γN −Lf +2 ) by assuming
clearly be observed in Section V. BAD requires a complexity perfect decision feedback at high SNR.
of at least twice that of the MMSE-DFE. Note that, with regard
to computational complexity, we focus on time-domain imple-
B. Multipath Diversity Analysis
mentation, even though a low-complexity frequency-domain
implementation is also possible using the block-circulant struc- Next, we further analyze the behavior of the proposed A-ML-
ture that can be created by the guard interval. In addition, note DFBE at high SNR. Assuming perfect channel estimation at the
Authorized licensed use limited to: DGIST. Downloaded on December 09,2023 at 13:30:20 UTC from IEEE Xplore. Restrictions apply.
2318 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 5, JUNE 2009

receiver and taking (19) as an example, it can be upper bounded

by [1]
1 gpsk
SERkM ≤ exp − 2 γk
2 ⎛ sin θ ⎞
min(Lf −1,L−1)
1 gpsk 
≈ exp ⎝− 2 |hi (t)|2 ⎠
2 σ Lf sin2 θ i=0
(21)
min(Lf −1,L−1)
where γk ≈ (1/σ 2 Lf ) i=0 |hi (t)|2 at high SNR
(refer to Appendix for the derivation). To obtain good perfor-
mance in terms of multipath combining and ISI suppression,
we should choose Lf ≥ L. Then, by averaging (21) over the
Rayleigh probability density function [14], (21) becomes
 −L
1 gpsk SNR
SERM 
k (22) Fig. 1. Analytical BER performance of the A-ML-DFBE over a doubly selec-
2 L · Lf sin2 θ tive fading channel with perfect channel estimation (L = 5, fd Ts = 0.0001)
and simulated BER.
which indicates that the A-ML-DFBE achieves the maximum
multipath diversity order L. to make up for the performance loss caused by the slicing
processes to obtain the same performance. Note also that, if
C. Analysis of the Length of the Forward Filter the length of the forward filter is equal to N , the A-ML-DFBE
(Sliding Window) and Backward Filter directly enters the “Tail Detection” step (step 3) in Table I,
and since there is no slicing operations at all, whether or
It has been shown that forward-filter length Lf is a very not the matched filter is used will make no difference in the
important parameter in the proposed A-ML-DFBE. In this performance. However, the value of Lf is normally much less
section, we discuss the behaviors of Lf . 1) Increasing the than N . Theoretically, using the same methods, as shown in
value of Lf can improve the robustness of (15) due to the Appendix, it is easy to obtain the SNR for the A-ML-DFBE
following reasons: First, as shown in (5), a larger value of Lf when the matched filter is removed. Due to the space limitation,
can incorporate more received signals and channel information we drop the detailed derivation part. However, we can conclude
in the forward filter. Second, as shown in (13), increasing Lf that the performance of the A-ML-DFBE can be upper bounded
can make the Gaussian assumption more accurate. 2) While by the same equalize without using the matched filter.
the performance can be enhanced, as shown in Table II,
the complexity will correspondingly increase. Hence, for the
A-ML-DFBE, a complexity and performance tradeoff can be V. S IMULATION R ESULT
realized by adjusting Lf . 3) Performance gets converged by In all the simulations, binary PSK constellation is used to
increasing the value of Lf as the Gaussian assumption becomes generate a transmission with a rate of 1 b/s/Hz. We plot the BER
accurate enough. This implies that a moderate length of the versus the SNR. For the analytical results, we assume perfect
forward filter can deliver good performance. 4) As discussed decision feedback, but for the simulation results, we use the
in Section IV-B, Lf should be equal to or larger than L for feedback decisions. The performance is determined over dou-
maximum diversity order. 5) The length of the backward filter bly selective Rayleigh fading channels. The impulse response
Lb is always equal to L − 1 due to the special structure of H. length is L = 5; thus, the length of the backward filter of the
A-ML-DFBE can be fixed as Lb = L − 1 = 4. Jakes’ model is
D. Analysis of the Matched Filter in (3) applied to construct a time-selective Rayleigh fading channel
for each subpath. The carrier frequency fc = 2 GHz, and the
Note that the matched filter in (3) can obtain some addi- symbol period Ts = 128/c, where c is the speed of light. The
tional information from the received signals outside the slicing simulation results are plotted with two speeds: v = 5 km/h and
window. Recalling (4)–(9), yk can be written as 150 km/h (corresponding to fd Ts = 0.0001 and 0.0093, where
 T Doppler frequency fd = vfc /c). The frame length N is 128.
yk = hH H
k r, . . . , hk+Lf −1 r . (23) In Figs. 1 and 2, we examine the analytical BER perfor-
mance obtained in (20), assuming that the channel estimation
Although some information is lost after horizontal and verti- is perfect. The simulations are plotted with vehicle speed
cal slicing, some gains can still be realized by considering the v = 5 km/h. In Fig. 1, we compare the analytical BER with
whole received signal r. the simulated BER. It can be observed that the analytical
Supposing that the matched filter is removed, the detection BER is close and asymptotically converges to the simulated
procedures in Table I can be used but will lead to performance curves at high SNR. In Fig. 2, the analytical BER for the
degradation since only the received signals inside the sliding A-ML-DFBE is plotted, employing different forward-filter
window will be considered, where yk = [rk , . . . , rk+Lf −1 ]T . lengths. As previously discussed, the length of the forward filter
As a result, the length of the forward filter has to be increased Lf should at least be equal to L to realize a good performance.
Authorized licensed use limited to: DGIST. Downloaded on December 09,2023 at 13:30:20 UTC from IEEE Xplore. Restrictions apply.
SONG et al.: A-ML-DFBE FOR DOUBLY SELECTIVE FADING CHANNELS 2319

Fig. 2. Analytical BER performance of the A-ML-DFBE with various

forward-filter lengths over a doubly selective fading channel with perfect Fig. 4. Simulated BER performance of the A-ML-DFBE over a doubly selec-
channel estimation (L = 5, fd Ts = 0.0001). tive fading channel with perfect channel estimation (L = 5, fd Ts = 0.0001).
Comparisons of the A-ML-DFBEs with and without the matched filter (MF)
are shown.

see that the A-ML-DFBE with Lf = 5 can provide a much

better performance than BAD with Lf = 15. Note that our
A-ML-DFBE has lower complexity than MMSE-DFE and,
thus, BAD. Clearly, from Figs. 1–3, we can see that there exists
a complexity and performance tradeoff in terms of Lf . Perfor-
mance can be improved by increasing the length of the forward
filter (slicing window). In addition, performance convergence
can also be observed, which indicates that a limited value of Lf
is enough to deliver most of the performance gain.
In Fig. 4, simulation comparisons are made for the A-ML-
DFBE without using the matched filter. Perfect channel esti-
mation is assumed. Vehicle speed v = 5 km/h is adopted. We
choose different Lf values for the no-matched-filter case, i.e.,
Fig. 3. Simulated BER performance of the A-ML-DFBE over a doubly 5, 10, and 15, whereas Lb remains the same, i.e., 4. It is shown
selective fading channel with LS channel estimation (L = 5, fd Ts = 0.0001). that, at Lf = 5, the performance without the matched filter is
Comparisons of the linear MMSE, MMSE-DFE, BAD, and MLSE are shown.
worse than that with it. We can also observe significant perfor-
From Fig. 2, we can see that the proposed A-ML-DFBE with mance loss due to the small value of Lf . It is shown that Lf
Lf = 5 provides a much better performance than that with must be 15 for the system with no matched filter to provide the
Lf = 3, and as the value of Lf increases, the performance same performance as the matched filter system with Lf = 10.
begins to converge. It can also be seen that, for the A-ML- Hence, from the simulation results, we can see that the matched
DFBE, Lf = 10 (twice that of L) is enough to obtain a good filter is very important for system performance. Note that, as
BER performance. discussed in Section III-B, the forward and backward filter taps
In Fig. 3, the simulation results for the A-ML-DFBE detec- are actually fixed and can be obtained before the A-ML-DFBE
tor are illustrated in comparison with those for conventional detection. The complexity increase using the matched filter is
linear MMSE, MMSE-DFE, BAD, and MLSE decoders. The much more worthwhile than the increase in the length of the
simulations are plotted with vehicle speed v = 5 km/h. Least forward filter without using the matched filter.
square (LS) channel estimation [4] is used. From Fig. 3, it In Fig. 5, simulation results for the A-ML-DFE detector
can be observed that, at BER = 10−3 , the performance of the are illustrated in comparison with those of conventional linear
A-ML-DFBE with Lf = 5 is far better than that of the linear MMSE, MMSE-DFE, BAD, and MLSE decoders using LS
MMSE and MMSE-DFE equalizers. There is only 2-dB loss, channel estimation, and the vehicle speed is v = 150 km/h.
compared with that of the MLSE decoder at BER = 10−5 . At Here, we choose different values for Lf for the A-ML-DFBE.
Lf = 10, there is about 0.8-dB loss, compared with that of the From the simulation results, we can still observe that the
MLSE. Almost no difference can be observed for the A-ML- performance of the A-ML-DFBE converged at Lf = 10, and
DFBE when Lf is increased to 15 since Lf = 10 is sufficient no gain can be obtained at Lf = 15. Due to the time-variant
to make the performance converge. Note that, when Lf = 15, effects, the performance is degraded, compared with the results
the A-ML-DFBE gives almost the same performance as that in Fig. 3. We can see about 1-dB loss between the MLSE and
when Lf = 10, which demonstrates that only a small value of the A-ML-DFBE with Lf = 10 when BER = 10−5 . However,
Lf is required to achieve a good performance. We can also the proposed equalizer can still substantially outperform linear
Authorized licensed use limited to: DGIST. Downloaded on December 09,2023 at 13:30:20 UTC from IEEE Xplore. Restrictions apply.
2320 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 5, JUNE 2009

written as
−1
−1
γk = Ψk jk 2 = jH H 2
k Λk jk = jk σ X + YY
H
jk
(24)
Δ
where, for convenience, Y = Jk+1 is of size Lf × Lf − 1, and
Δ
X = Jk . By using the Kailath variant (A + BC)−1 = A−1 −
A−1 B(I + CA−1 B)−1 CA−1 [15], the inversion term on the
right side of (24) can further be written as
γk = σ −2 jH
k 
× X−1 − X−1 Y(σ 2 ILf −1 + YH X−1 Y)−1 YH X−1 jk .
(25)
At high SNR, as σ 2 → 0+ , the effect of σ 2 ILf −1 is compar-
atively small, which can be ignored from an asymptotic point
Fig. 5. Performance of the A-ML-DFBE over a doubly selective fading of view. Hence, we have the following approximation for the
channel with LS channel estimation (L = 5, fd Ts = 0.0093). Comparisons second term in (25):
of the linear MMSE, MMSE-DFE, BAD, and MLSE are shown.
σ −2 X−1 Y(YH X−1 Y)−1 YH X−1
MMSE and MMSE-DFE in all SNR regimes. About 8-dB
= σ −2 X− 2 Z(ZH Z)−1 ZH X− 2
1 1
(26)
performance gain can be obtained by the proposed scheme with
Lf = 5, compared with the BAD at BER = 10−3 . Δ
where Z = X−1/2 Y with size Lf × (Lf − 1), and (·)−1/2
represents the unique positive definite Hermitian root [15].
VI. C ONCLUSION Let Z+ be the Moore–Penrose inverse of matrix Z,
and Z+ = (ZH Z)−1 ZH with size (Lf − 1) × Lf . Note that
In this paper, we have proposed a simple approximate ML rank(ZZ+ ) = rank(Y) = Lf − 1 and that ZZ+ is of size
DFE for a doubly selective fading environment. From the Lf × Lf . By eigenvalue decomposition, we can get ZZ+ =
analytical and simulation results, we conclude that the A-ML- UΠUH , where U is the unitary eigenvector matrix and
DFBE significantly outperforms the linear MMSE, MMSE- Δ
DFE, and BAD detectors and provides a performance that is Π = diag{λ1 , . . . , λLf −1 , 0}. From the definition of ZZ+ ,
very close to that of the MLSE. We have shown that, when Lf is we have (ZZ+ )2 = ZZ+ . Therefore, ZZ+ is idempotent [15],
large enough, further increases in Lf do not improve the perfor- and any idempotent matrix has eigenvalue 1 or 0; thus, Π =
mance much. This implies that the proposed equalizer is quite diag{1, . . . , 1, 0}. We can then get
−2
robust against ISI. A tradeoff in terms of the complexity and Z(ZH Z)−1 ZH X− 2 jk = jH −2
UΠUH X− 2 jk
1 1 1 1
jH
k X k X
the performance can be achieved by adjusting the value of Lf . Lf − 1 H −1
Computational complexity comparison has demonstrated that ≈ j X jk . (27)
Lf k
the A-ML-DFBE requires fewer additions and multiplications
than MMSE-based schemes. In addition, the implementation From (24)–(27), at high SNR, we can obtain
of the matched filter is very important, and the A-ML-DFBE 1 H −1
obtains maximum diversity order when Lf ≥ L. γk ≈ j X jk . (28)
σ 2 Lf k
Due to the DFE processing, parallel computing is difficult to
achieve for the proposed equalizer. However, by adjusting the From (5), we can get
size of the data block or the filters (backward and forward), or −1
−1
both, the latency can be reduced. The proposed equalizer can jH H H
k X jk = hk Hk Hk Hk HH
k hk (29)
easily be used for radar communication systems as it is suitable
where jk = HH k hk , and X = Hk Hk is of size Lf × Lf and
H
for solving time-domain equalization problems. In current wire-
H −1 H
less systems such as Universal Mobile Telecommunications rank Lf . Since Hk (Hk Hk ) Hk has the same structure as
System, High-Speed Downlink Packet Access, or High-Speed Z(ZH Z)−1 ZH in (27), we can get the corresponding eigen-
Uplink Packet Access, the A-ML-DFBE can be used to recover values as
signals similar to MMSE or MMSE-DFE. For Long Term −1 H
Evolution (LTE) or LTE advance, the proposed algorithm can EIG Hk HH k Hk Hk
be extended to realize frequency-domain equalizations. = diag{0, . . . , 0, 1, . . . , 1, 0, . . . , 0 }. (30)
        
k−1 Lf N +L−Lf −k
A PPENDIX
D ERIVATION OF C LOSED -F ORM E XPRESSION OF Finally, combining (28) and (29), at high SNR, as σ 2 → 0+ ,
γk AT H IGH SNR we finally have
min(Lf −1,L−1)
1 
Now, the closed-form expression of γk at high SNR is γk ≈ |hi (t)|2 . (31)
2
σ Lf
derived in terms of Lf and L. From Section IV-A, γk can be i=0
Authorized licensed use limited to: DGIST. Downloaded on December 09,2023 at 13:30:20 UTC from IEEE Xplore. Restrictions apply.
SONG et al.: A-ML-DFBE FOR DOUBLY SELECTIVE FADING CHANNELS 2321

R EFERENCES Rodrigo Caiado de Lamare (M’05) received the

Diploma degree in electronic engineering from the
[1] J. G. Proakis, Digital Communications, 4th ed. New York: McGraw-
Federal University of Rio de Janeiro (UFRJ), Rio
Hill, 2001.
de Janeiro, Brazil, in 1998 and the M.Sc. and Ph.D.
[2] G. D. Forney, “Maximum-likelihood sequence estimation of digital se-
degrees in electrical engineering from the Pontifical
quences in the presence of intersymbol interference,” IEEE Trans. Inf.
Catholic University of Rio de Janeiro (PUC-Rio),
Theory, vol. IT-18, no. 3, pp. 363–378, May 1972.
Rio de Janeiro, in 2001 and 2004, respectively.
[3] G. K. Kaleh, “Channel equalization for block transmission systems,” From January 2004 to June 2005, he was a
IEEE J. Sel. Areas Commun., vol. 13, no. 1, pp. 110–121, Jan. 1995.
Postdoctoral Fellow with the Center for Studies
[4] S. Haykin, Adaptive Filter Theory, 3rd ed. Upper Saddle River, NJ:
in Telecommunications (CETUC), PUC-Rio. From
Prentice–Hall, 1996.
July 2005 to January 2006, he was a Postdoctoral
[5] G. Leus and M. Moonen, “Equalization techniques for fading channels,” Fellow with the Signal Processing Laboratory, UFRJ. Since January 2006,
in Handbook on Signal Processing for Communications, M. Ibnkahla, Ed.
he has been with the Communications Research Group, Department of Elec-
Boca Raton, FL: CRC, 2004.
tronics, University of York, York, U.K., where he is currently a Lecturer in
[6] L. Song and A.G. Burr, “Successive interference cancelation for space-
communications engineering. His research interests include communications
time block codes over time-selective fading channels,” IEEE Commun. and signal processing.
Lett., vol. 10, no. 12, pp. 837–839, Dec. 2006.
[7] Y. Jia, C. Andrieu, R. J. Piechocki, and M. Sandell, “Gaussian approx-
imation based mixture reduction for near optimum detection in MIMO
systems,” IEEE Commun. Lett., vol. 9, no. 11, pp. 997–999, Nov. 2005.
[8] J. Luo, K. R. Pattipati, P. K. Willett, and F. Hasegawa, “Near-optimal
multiuser detection in synchronous CDMA using probabilistic data as-
sociation,” IEEE Commun. Lett., vol. 5, no. 9, pp. 361–363, Sep. 2001.
[9] S. Liu and Z. Tian, “Near-optimum soft decision equalization for fre-
quency selective MIMO channels,” IEEE Trans. Signal Process., vol. 5, Are Hjørungnes (SM’06) received the
no. 3, pp. 721–733, Mar. 2004. Sivilingeniør (M.Sc.) degree (with honors) from
[10] J. K. Nelson, A. C. Singer, and U. Madhow, “BAD: Bidirectional arbi- the Norwegian Institute of Technology, Trondheim,
trated decision-feedback equalization,” IEEE Trans. Commun., vol. 53, Norway, in 1995 and the Doktor Ingeniør (Ph.D.)
no. 2, pp. 214–218, Feb. 2005. degree from the Norwegian University of Science
[11] A. M. Chan and G. W. Wornell, “A class of block-iterative equalizers for and Technology, Trondheim, in 2000.
intersymbol interference channels: Fixed channel results,” IEEE Trans. He is an Associate Professor with the UNIK-
Commun., vol. 49, no. 11, pp. 1966–1976, Nov. 2001. University Graduate Center, University of Oslo,
[12] G. H. Golub and C. D. Loan, Matrix Computations, 3rd ed. Baltimore, Oslo, Norway. From August to December 2000, he
MD: Johns Hopkins Univ. Press, 1996. was a Researcher with the Tampere International
[13] M. K. Simon and M.-S. Alouini, Digital Communications Over Fading Center for Signal Processing, Tampere University
Channels: A Unified Approach to Performance Analysis, ser. Wiley of Technology, Tampere, Finland. From March 2001 to July 2002, he was a
Series in Telecommunications and Signal Processing. Hoboken, NJ: Postdoctoral Fellow with the Signal Processing Laboratory, Federal University
Wiley, 2001. of Rio de Janeiro, Rio de Janeiro, Brazil. From September 2002 to August
[14] V. Tarokh, N. Seshadri, and A. R. Calderbank, “Space-time codes for 2003, he was a Postdoctoral Fellow with the Signal Processing Laboratory,
high data rate wireless communication: Performance criterion and code Helsinki University of Technology, Espoo, Finland. From September 2003 to
construction,” IEEE Trans. Inf. Theory, vol. 44, no. 2, pp. 744–765, August 2004, he was a Postdoctoral Fellow with the Digital Signal Processing
Mar. 1998. and Image Analysis Group, Department of Informatics, University of Oslo. He
[15] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. Cambridge, has held visiting appointments at the Image and Signal Processing Laboratory,
U.K.: Cambridge Univ. Press, 1991. Reprinted 1999. University of California, Santa Barbara; the Signal Processing Laboratory,
Federal University of Rio de Janeiro; the Mobile Communications Depart-
ment, Eurécom Institute, Paris, France; the University of Manitoba, Winnipeg,
MB, Canada; and the Department of Electrical and Computer Engineering,
Lingyang Song (M’07) received the B.S. degree University of Houston, Houston, TX. He has also been an Alcatel–Lucent Chair
in communication engineering from Jilin University, with SUPÉLEC, Gif Sur Yvette, France. His current research areas include
Changchun, China, in 2002 and the Ph.D. degree signal processing, communications, and wireless networks.
in differential space time codes and multiple-input– Prof. Hjørungnes has been an Editor for the IEEE TRANSACTIONS ON
multiple-output from the University of York, York, WIRELESS COMMUNICATIONS since March 2007. He was the recipient of
U.K., in 2007.
the Best Paper Award at the IEEE International Conference on Wireless
From January to September 2003, he was a Soft-
Communications, Networking, and Mobile Computing (WiCOM 2007).
ware Engineer with Hwasun Tomorrow Technology,
Beijing, China. From June to December 2005, he was
with Philips Research U.K., Cambridge, U.K., as a
Research Scientist. He was a Postdoctoral Research
Fellow with the University of Oslo, Oslo, Norway, until rejoining Philips
Research U.K. in 2008. He is now with School of Electronics Engineering
and Computer Science, Peking University, China. He was a Visiting Research
Fellow with Harvard University, Cambridge, MA, and the University of York.
He currently serves on the Editorial Board of the International Journal of Alister G. Burr (M’91) was born in London, U.K., in 1957. He received the
Communications, Network and System Sciences, and the International Journal B.Sc. degree in electronic engineering from the University of Southampton,
of Smart Home. He is also a Guest Editor for Elsevier Computer Communica- Southampton, U.K., in 1979 and the Ph.D. degree from the University of
tions and the EURASIP Journal on Wireless Communications and Networking. Bristol, Bristol, U.K., in 1984.
He is the author or coauthor of more than 50 journal and conference pro- From 1975 to 1985, he was with Thorn-EMI Central Research Laboratories,
ceeding papers. He is a coeditor of the book Orthogonal Frequency Division London. In 1985, he joined the Department of Electronics, University of York,
Multiple Access (OFDMA)-Fundamentals and Applications (to be published York, U.K., where he has been a Professor of communications since 2000.
by Auerbach, CRC, 2008). He is a coholder of a number of patents. His re- He has also been a Visiting Professor with Vienna University of Technology,
search interests include adaptive signal processing, estimation and optimization Vienna, Austria. His research interests include wireless communication sys-
theory, cognitive and collaborative communications, channel coding, wireless tems, particularly modulation and coding, turbo codes, and turbo processing
mesh/sensor networks, broadband wireless access, and future communication techniques, as well as multiple-input–multiple-output and cooperative systems.
systems. Dr. Burr is currently the Chair of Working Group 1 of the European COST
Dr. Song is a Member of the IEEE Communications Society. He serves 2100 Program “Pervasive Mobile and Ambient Wireless Communications.”
as a Member of the Technical Program Committee and a Co-Chair for sev- He was the recipient of the Senior Research Fellowship from the U.K. Royal
eral international conferences and workshops. He was the recipient of the Society in 1999 and the J. Langham Thompson Premium from the Institution
K. M. Stott Prize for excellent research from the University of York. of Electrical Engineers in 2002.
Authorized licensed use limited to: DGIST. Downloaded on December 09,2023 at 13:30:20 UTC from IEEE Xplore. Restrictions apply.