Iterative Channel Estimation with Robust Wiener

Filtering in LTE Downlink

Luis Ángel Maestro Ruiz de Temiño∗ , Carles Navarro i Manchón∗ , Christian Rom† ,
Troels B. Sørensen∗ and Preben Mogensen∗
∗ Department of Electronic Systems, Aalborg University
Niels Jernes Vej 12, 9220 Aalborg, Denmark
Email: {lam, cnm, tbs, pm}@es.aau.dk
† Infineon Technologies

Alfred Nobels Vej 25, 9220 Aalborg, Denmark

Email: christian.rom@infineon.com

Abstract—In this paper, an iterative enhancement of the robust
Wiener filter (RWF) estimator is presented for a turbo-coded or-
thogonal frequency division multiplexing (OFDM) system under
the umbrella of the 3GPP Long Term Evolution. The proposed
scheme can operate with uncoded or coded feedback, and
outperforms the conventional linear RWF in the whole signal-to-
noise ratio (SNR) range with both approaches. Results show that
most of the gain is obtained in the first iteration of the algorithm,
and better performance is achieved with the coded feedback
scheme. A good tradeoff between accuracy and complexity is
achieved by selecting a low number of turbo coding iterations
(TCI) in the iterative loop and concentrating most of them at the
final decoding stage. Following this design, cell spectral efficiency
gains of around 2.7 % and 6.5 % can be obtained with respect
to linear RWF for micro- and macro-cell scenarios respectively.
I. I NTRODUCTION
Over the recent years, an increasing demand for higher
data rates in wireless communications systems has arisen in
order to support broadband services. To achieve such high
data rates, wideband transmission over the dispersive mobile
channel is required. A highly efficient way to cope with
the frequency selectivity of wideband channels is orthogo-
nal frequency division multiplexing (OFDM). In OFDM, the
transmission bandwidth is divided into lower-rate narrowband
orthogonal subcarriers. This, together with the employment of
a proper cyclic prefix (CP), allows simple equalization of the
multipath channel [1]. The ability to easily cope with multipath
distortion and the high spectral efficiency has motivated the
election of OFDM in upcoming wireless standards, such
as the 3GPP Long Term Evolution (LTE) [2] or the IEEE
802.16e-2005 standard (WiMAX) [3]. In both standards, pilot-
assisted channel estimation (PACE) is defined in order to allow
coherent detection at the receiver.
When PACE is employed, pilot symbols, known by both the
transmitter and the receiver, are sent in pre-defined subcarrier
locations. By processing the received signal at these positions,
the receiver can estimate the whole channel response for
each OFDM symbol. Minimum mean-squared-error (MMSE)
interpolation of the channel response has been proposed as a
solution [4], known as Wiener filtering interpolation. Although
it exhibits the best performance among the existing linear
algorithms in literature, it requires accurate knowledge of
second order channel statistics, which is not always feasible
at a mobile receiver. To overcome this, a robust design of
the filter can be used [5], which eases the practical imple-
mentation while still keeping an acceptable performance. To
further improve the accuracy of the estimator, iterative channel
estimators can be employed. In these schemes, the estimates
are improved by feeding back data decisions to the channel
estimation block in an iterative fashion, as done for instance
in [6], [7]. However, these schemes usually suffer from a
prohibitive computational complexity.
In this paper, we propose an iterative enhancement of the
robust Wiener filter (RWF) in a turbo-coded OFDM system.
The complexity of the scheme can be tuned by varying
the number of channel estimation iterations as well as the
number of iterations inside the turbo decoder. The performance
of the estimator is evaluated for LTE parameters, showing
that improved accuracy can be achieved while keeping the
computational complexity at a reasonable level, making it
suitable for practical implementation in a mobile receiver.
The remainder of the paper is organized as follows: in
Section II, the considered OFDM system is described. The
proposed iterative scheme is presented in Section III, and its
performance is analyzed in Section IV. Finally, Section V
concludes the work.
II. S YSTEM D ESCRIPTION
A simplified block diagram of the considered OFDM sys-
tem, with classical linear reception, is depicted in Fig. 1. On
the transmitter side, the bit stream b is fed to the turbo encoder
block, which follows 3GPP Release 7 specifications [8]. This
block contains a rate 1/3 turbo encoder, a rate matching
module that performs repetition or puncturing of the coded bits
depending on the selected coding rate, and a bit interleaver.
The coded bits, denoted by the vector c, are then modulated
onto a QPSK, 16QAM or 64QAM constellation and pilot
symbols are inserted in the data stream. These modulated
symbols, both data and pilots, are mapped onto the Nu central
subcarriers of the system, and an inverse fast Fourier transform
(IFFT) of size Nf f t is performed to obtain the time-domain

978-1-4244-1722-3/08/$25.00 ©2008 IEEE. 1

Authorized licensed use limited to: Univ of Puerto Rico Mayaguez - Library. Downloaded on August 11,2010 at 13:10:01 UTC from IEEE Xplore. Restrictions apply.

Fig. 1. Block diagram of the considered system with classical linear reception
OFDM signal. After the addition of the CP, the signal is
transmitted over a wireless multipath channel characterized
by its channel impulse response:
N
t −1
g(τ, t) = αi (t)δ(τ − τi )
i=0
where αi (t) are wide sense stationary, uncorrelated complex
Gaussian random path gains at time instant t, with their
respective delays τi . Nt denotes the total number of paths.
At the receiver, the CP is first discarded, and a fast Fourier
transform (FFT) is performed to recover the frequency-domain
signal. Assuming that the channel is static over the duration
of one OFDM symbol and that the CP is longer than the
maximum excess delay of the channel τNt −1 , the signal
yu ∈ CNu can be described as:
yu = Xu hu + wu
Note that the subindex u indicates that only the Nu subcarriers
filled with symbols are considered. In (2), Xu ∈ CNu ×Nu
is a diagonal matrix containing the transmitted symbols and
hu ∈ CNu is a vector with the channel transfer function (CTF)
coefficients at each subcarrier:
Nt −1 −j2πkτi
hu [k] = αi e T0 Nf f t (3)
i=0
where k denotes the subcarrier index and T0 denotes the
sampling period. Finally, wu ∈ CN
u is an additive white
2 Nu ×Nu
Gaussian noise vector with variance σw . where X̂u ∈ C
In order to recover the transmitted bits, the channel estima-
tor block needs to obtain an estimate of hu , which is used
by the soft demodulator to derive the soft estimates c̃ of the
coded bits. To this end, Np pilot symbols are transmitted in
some predefined subcarrier positions. The received signal in
these locations can be written as:
yp = Xp hp + wp
with Xp ∈ CNp ×Np and (wp , hp ) ∈ CNp being subsets of
the corresponding matrices defined in (2). Finally, the soft
estimates of the coded bits pass through an iterative turbo
decoder, which is based on the max-log maximum a-posteriori
algorithm [9], [10], obtaining as output the hard decisions on
the information bits b̂.
Authorized licensed use limited to: Univ of Puerto Rico Mayaguez - Library. Downloaded on August 11,2010 at 13:10:01 UTC from IEEE Xplore. Restrictions apply.
III. I TERATIVE ROBUST W IENER F ILTER
In this section, we present an iterative enhancement of the
RWF in which the demodulated data symbols are re-used, as if
they were known, in order to improve the estimates’ accuracy.
The operation of the algorithm is depicted in Fig.2. In the first
stage, only the pilot symbols are used to obtain the estimate
of the CTF which can be expressed as:
(0) 2
ĥRW F = R̄hu hp (R̄hp hp + β · σw · Ip )−1 ĥLS,p (5)
where ĥLS,p = X−1 p yp ∈ C
Np
is the least squares channel
estimate in frequency-domain at pilot positions, β ∈ R is a
constant that depends on the modulation, Ip ∈ NNp ×Np is the
identity matrix and (R̄hu hp ∈ CNu ×Np , R̄hp hp ∈ CNp ×Np )
are subsets of the covariance matrix R̄hh ∈ CNf f t ×Nf f t . The
latter is defined assuming a uniform power delay profile (PDP)
with paths’ delays uniformly distributed between 0 and τNt −1
as [5]:
τN −1
(1) t
1 − e−2πj T0 (k−n)/Nf f t
R̄hh [k, n] = τ k−n
(6)
2πj NTt0−1 N
fft
(0)
The estimate of the CTF in the used subcarriers ĥRW F is
then employed, together with the received signal, in the soft
demodulator block to obtain an estimate of the transmit coded
bits c̃. Note that in this block, the pilot symbols have been
removed from the information stream. If the maximum number
of iterations has not been reached, the vector c̃ is decoded
using an iterative turbo decoder, with M turbo coding iterations
(TCI), thus obtaining a hard estimate of the uncoded transmit
(2) bits b̂. Finally, the vector b̂ is re-encoded and posteriorly re-
modulated to QPSK/QAM symbols in the mapper block. It is
important to highlight here that this block is also responsible
to re-multiplex the pilots with the data symbols in order to
keep the same structure as in the received signal. This point
is the end of the first stage, which is related to i = 0 with i
denoting the iteration index. From this point on (i > 0), all
demodulated transmit symbols, both data and pilots, are used
in the estimation process. Hence, it is appropriate to rewrite
the estimate of the CTF for i > 0 as:
(i) 2
ĥRW F = R̄hu hu (R̄hu hu + β · σw · Iu )−1 (X̂u )−1 yu (7)
is a diagonal matrix whose elements
are the estimated transmit symbols (both data and pilots).
This iterative process is repeated until the maximum number
of iterations Imax is achieved. Then, the vector c̃ is decoded
with N TCI, obtaining the output bit vector b̂. It is worth men-
tioning the employment of a different number of TCI inside
and outside the estimation loop since it will be demonstrated
(4) later that it plays an important role on the system performance.
IV. P ERFORMANCE E VALUATION
In the following, the performance of the iterative estimator
proposed in Section III will be evaluated by means of Monte
Carlo simulations. To this end, a single-input single-output
(SISO) downlink OFDM system based on the 10 MHz LTE
physical layer parameters [11] will be considered. These
2
 −10

−12

−14

Average MSE [dB]

−16

−18

Fig. 2. Iterative RWF Block Diagram i=0

−20
i = 1; M = 0
TABLE I i = 1; M = 2
S IMULATION PARAMETERS −22 i = 1; M = 4
i = 1; M = 6
Sampling frequency 15.36 MHz i = 1; M = 8
−24
Subcarrier spacing 15 KHz 5 6 7 8 9 10
SNR [dB]
Nf f t 1024
Nu 600
Fig. 3. Average MSE for RWF with QPSK 1/2
CP length 5.2a /4.68b µs
Pilot spacing 6 TABLE II
Pilot overhead 4.76% R ELATIVE SPECTRAL EFFICIENCY GAINS (%)
Slot duration 0.5 ms
Configuration M=2, N=6 M=1, N=7 M=3, N=5 M=4, N=4
OFDM symbols per slot 7
Relative gain 4.4 3.3 1.6 0.5
Antenna scheme SISO
User speed 3 kmph
MCS settings QPSK: 1/6, 1/3, 1/2, 2/3
16QAM: 1/2, 2/3, 3/4 (i = 0) has also been plotted. First of all, the employment
64QAM: 1/2, 2/3, 3/4, 4/5 of the iterative RWF allows for a reduced MSE compared
to the linear algorithm. Besides, it can be observed that the
a First OFDM symbol in a slot. estimates’ accuracy is highly improved with M = 2 TCI since
b 2th − 7th OFDM symbol in a slot. the MSE is significantly lowered. There is not appreciative
gain by increasing M further than 2. From the results, we can
conclude that it is worth employing a low number of TCI (M)
parameters are gathered in Table I as well as the modulation in the inner decoding while concentrating most of them (N)
and coding set (MCS) formats employed. QPSK modulation in the last one to correct as many errors as possible.
has been considered for pilot symbols, which are transmitted Table II gathers the relative gains in spectral efficiency of the
in the first and fifth OFDM symbol within a slot with an even iterative RWF at i = 1 with respect to the linear estimator with
frequency-domain spacing of 6. Furthermore, Typical Urban different combinations of M and N. In this case, 16QAM 2/3
20 taps channel model [12] is employed in the simulations and has been considered as well as an operating point at SN R =
low user speed is considered. Results are presented in terms 15 dB. The results confirm the statement above mentioned
of packet error rate (PER), spectral efficiency (bits/s/Hz) and since the configuration with M = 2 and N = 6 provides the
mean squared error (MSE) of the CTF as a function of the highest gain ∼ 4.4 %.
signal-to-noise ratio (SNR). Finally, Fig. 4 depicts PER vs SNR curves for the M =
2, N = 6 configuration for two selected MCS, namely QPSK
A. Turbo Coding Iterations Evaluation 1/2 and 16QAM 2/3. Results show that in both cases, this
The goal of this subsection is to provide the optimum configuration performs slightly better than uncoded feedback
distribution of TCI inside and outside the estimation loop. (i = 1, M = 0, N = 8), obtaining a gain around ∼ 0.3 dB at
Furthermore, in order not to increase the complexity at the P ER = 10−2 with respect to the linear estimator (i = 0).
receiver, while compared with the linear RWF, the maximum
number of TCI is fixed to be M + N = 8. In the rest of B. Spectral Efficiency Results
the document, only one iteration of the algorithm has been In this section, we extend the previous results and build
considered since it has been observed that most of the gain the link adaptation (LA) curves for the iterative RWF. Fig. 5
is obtained in this first stage. Fig.3 depicts the average MSE presents the spectral efficiency results for: uncoded feedback,
in dB for the proposed iterative estimator at the end of the coded feedback with M = 2, N = 6 and linear RWF. The
first iteration using QPSK 1/2. The number of TCI inside latter has been represented for comparison purposes. Results
the estimation loop (M) varies from 0 (uncoded feedback) to show that coded feedback with M = 2, N = 6 obtains
8. For the sake of comparison, the MSE of the linear RWF the highest gains with respect to the linear algorithm. At

3
Authorized licensed use limited to: Univ of Puerto Rico Mayaguez - Library. Downloaded on August 11,2010 at 13:10:01 UTC from IEEE Xplore. Restrictions apply.
 0
10 4
Known Channel
3.5
RWF i = 0; M = 0; N = 8

Spectral Efficiency (bits/s/Hz)

3 RWF i = 1; M = 0; N = 8
−1 RWF i = 1; M = 2; N = 6
Packet Error Rate

10
2.5

−2 Known Channel 1.5

10
RWF i = 0; M= 0; N= 8
RWF i = 1; M= 0; N= 8 1
RWF i = 1; M= 2; N= 6
0.5

−3
10 0
3 4 5 6 7 8 9 −5 0 5 10 15 20 25 30
SNR [dB] SNR [dB]
(a)
Fig. 5. Spectral efficiency curves for RWF

10
0 TABLE III
R ELATIVE CELL EFFICIENCY GAINS (%)

Cell environment M=0, N=8 M=2, N=6 M=2, N=8

Micro 0.8 2 2.7
−1
Packet Error Rate

10 Macro 2.2 5 6.5

presented in Table III, which also includes results for a new

10
−2
Known Channel
RWF i = 0; M= 0; N= 8
RWF i = 1; M= 0; N= 8
RWF i = 1; M= 2; N= 6
−3
10
11 12 13 14
SNR [dB]
(b)
Fig. 4. Packet error rate for: (a) QPSK 1/2; and (b) 16QAM 2/3
SN R = −2 dB, which corresponds to QPSK 1/6, a relative
gain of 36 % in spectral efficiency is achieved. At medium
SNR, e.g., SN R = 15 dB (16QAM 2/3), the gain is reduced
to 4.4 % whereas at SN R = 25 dB (64QAM 4/5) it is only
1.3 %. These gains are further reduced in the case of uncoded
feedback, being almost negligible at low SNR. This is due
to the fact that at this SNR range many errors committed in
the symbols’ detection are fed back to the channel estimator,
which will use this wrong information as correct symbol
decisions. However, we must highlight that no degradation in
performance is observed, i.e., the iterative algorithm is always
better than the linear one even at low SNR and considering
uncoded feedback.
Finally, the LA curves are mapped to cell spectral efficiency
by means of SNR distributions for macro- and micro-cell envi-
ronments [13]. The relative spectral efficiency gains obtained
with the iterative estimator compared to the linear one are
configuration with M = 2 and N = 8 TCI. As expected,
the gains are higher for coded feedback and for macro-cell
environment, since lower values of SNR are more likely to
happen in this scenario and, as shown before, the iterative
estimation is better exploited for medium-low SNR. Note that
15 16 17 18 the M = 2, N = 8 configuration obtains the highest gains
up to 6.5 % and 2.7 % in macro- and micro-cell scenarios,
respectively.
C. Complexity
Regarding the complexity of the iterative approach, it has
to be mentioned that each iteration of the algorithm has a
complexity of the same order as the one of linear RWF.
For this estimator, the main contribution to the complexity
comes from the term (R̄hu hu + β · σw 2
· Iu )−1 . This ma-
trix can be pre-calculated and stored for several scenarios,
thus only one run-time matrix multiplication is required (Nu2
complex multiplications). However, this complexity can be
significantly lowered by using a low-rank reduction based
on the singular value decomposition (SVD) of the channel
autocovariance matrix R̄hu hu [5]. This low-rank estimator
only requires 2pNu multiplications where p is the index of the
first singular value approximately equal to zero. Fig. 6 depicts
the amplitude of the singular values of R̄hu hu for different
channel profiles. Note that in scenarios with long maximum
path delay, such as the exponential decaying profile (EDP)
with τNt −1 = 17 μs, the number of singular values with
significant amplitude becomes larger; therefore, the number
of required multiplications (complexity) increases. Table IV

4
Authorized licensed use limited to: Univ of Puerto Rico Mayaguez - Library. Downloaded on August 11,2010 at 13:10:01 UTC from IEEE Xplore. Restrictions apply.
 180
linear RWF. Furthermore, by slightly increasing the number
of TCI in the last decoding to N = 8, cell spectral efficiency
160 Indoor A (τ
Nt−1
= 0.31 μs) gains of 2.7 % and 6.5 % are achievable over linear RWF
140 Typical Urban (τNt−1= 2.14 μs) for micro- and macro-cell environments, respectively. Finally,
Pedestrian B (τNt−1= 3.7 μs) using a low-rank estimator based on the SVD decomposition
120 of the channel autocovariance matrix, the complexity can
EDP (τNt−1= 17 μs)
Amplitude

100 be significantly reduced, especially for channels with short

maximum excess delay.
80
R EFERENCES
60
[1] L. Hanzo, M. Munster, B. Choi, and T. Keller, OFDM and MC-CDMA
40 for Broadband MultiUser Communications, WLANs and Broadcasting.
John Wiley - IEEE Press, 2003.
20 [2] “Evolved Universal Terrestrial Radio Access (E-UTRA); LTE Physical
Layer - General Description,” 3rd Generation Partnership Project, Tech.
0
20 40 60 80 100 120 140 160
Rep. TS 36.201, V8.1.0, Nov. 2007.
Singular value index [3] IEEE Standard for Local and metropolitan area networks Part 16: Air
Interface for Fixed and Mobile Broadband Wireless Access Systems
Amendment 2: Physical and Medium Access Control Layers for Com-
Fig. 6. SVD of R̄hu hu in 10 MHz bined Fixed and Mobile Operation in Licensed Bands and Corrigendum
1, IEEE Std. 802.16e-2005, 2006.
TABLE IV [4] P. Hoeher, S. Kaiser, and P. Robertson, “Two-dimensional pilot-symbol-
C OMPLEXITY R EDUCTION (%) aided channel estimation by wiener filtering,” Acoustics, Speech, and
Signal Processing, 1997. ICASSP-97., 1997 IEEE International Confer-
ence on, vol. 3, pp. 1845–1848 vol.3, 21-24 Apr 1997.
Channel profile Selected p Complexity reduction [5] O. Edfors, M. Sandell, J. van de Beek, and S. Wilson, “OFDM channel
Indoor A 10 97 estimation by singular value decomposition,” IEEE Transactions on
Typical Urban 28 91 Communications, vol. 46, no. 7, pp. 931–939, July 1998.
[6] S. ten Brink, F. Sanzi, and J. Speidel, “Two-dimensional iterative
Pedestrian B 40 87 APP channel estimation and decoding for OFDM systems,” Global
EDP 160 47 Telecommunications Conference, 2000. GLOBECOM ’00. IEEE, vol. 2,
pp. 741–745 vol.2, 2000.
[7] S. Y. Park, Y. G. Kim, and C. G. Kang, “Iterative receiver for joint
detection and channel estimation in OFDM systems under mobile radio
summarizes the complexity reduction achieved by means of channels,” Vehicular Technology, IEEE Transactions on, vol. 53, no. 2,
the low-rank estimator with respect to the full-rank RWF. This pp. 450–460, March 2004.
[8] “Evolved Universal Terrestrial Radio Access (E-UTRA); Multiplexing
reduction is strictly dependant on the length of the channel and channel coding (FDD),” 3rd Generation Partnership Project, Tech.
considered, being up to 97 % for Indoor A profile. Rep. TS 25.212, V7.6.0, Sep. 2007.
[9] W. Koch and A. Baier, “Optimum and sub-optimum detection of
V. C ONCLUSION coded data disturbed by time-varying inter-symbol interference,” IEEE
Globecom, pp. 1679–1684, Dec. 1990.
In this work, we have proposed an iterative enhancement of [10] A. Erfanian, S. Pasupathy, and G. Gulak, “Reduced complexity symbol
the RWF estimator with relatively low complexity. It has been detectors with parallel structures for ISI channels,” IEEE Transactions
seen that most of the gain is obtained in the first iteration for on Communications, vol. 42, no. 4, pp. 1661–1671, April 1994.
[11] “Evolved Universal Terrestrial Radio Access (E-UTRA); Physical Chan-
both uncoded and coded feedback. Even though the latter is nels and Modulation,” 3rd Generation Partnership Project, Tech. Rep. TS
shown to provide the best results, both approaches improve 36.211, V8.1.0, Nov. 2007.
the performance with respect to the linear estimator for all the [12] “Technical Specification Group Radio Access Networks; deployment
aspects,” 3rd Generation Partnership Project, Tech. Rep. TR 25.943,
SNR range. The optimal combination of TCI among the inner V7.0.0, Jun. 2007.
and outer turbo decoding has been investigated fixing a total [13] I. Z. Kovacs, K. I. Pedersen, J. Wigard, F. Frederiksen, and T. E. Kolding,
number of M + N = 8 for complexity constraints. Results “HSDPA performance in mixed outdoor-indoor micro cell scenarios,”
IEEE International Symposium on Personal, Indoor and Mobile Radio
show that selecting M = 2 and N = 6, relative spectral Communications, pp. 1–5, Sept. 2006.
efficiency gains up to 36 % can be obtained compared to the

5
Authorized licensed use limited to: Univ of Puerto Rico Mayaguez - Library. Downloaded on August 11,2010 at 13:10:01 UTC from IEEE Xplore. Restrictions apply.