Low-Complexity and Filter-Tap Memory Optimized Channel Estimation in Multi-Mode and Multi-Standard OFDM Systems

Low-complexity and filter-tap
memory optimized channel

estimation in multi-mode and
multi-standard OFDM systems
Farzad Foroughi Abari
Thesis for the degree of

Licentiate in Engineering
Lund University
Lund, Sweden
2011
Licentiate in Engineering (LicEng) is an intermediate academic
degree between MSc and PhD awarded by Swedish and Finnish
universities, likened to a Master of Philosophy (MPhil) degree
in the British system.
Department of Electrical and Information Technology

Radio Systems Group
Lund University
P.O. Box 118, S-221 00 Lund, Sweden
Series of licentiate and doctoral theses

No. 29
ISSN: 1654-790X
ISBN: 978-91-7473-119-4
c 2011 by Farzad Foroughi Abari.
Printed in Sweden by Tryckeriet i E-huset, Lund University, Lund.
November 2011
Abstract
Communication over orthogonal pulses in frequency (e.g., OFDM) has become a
dominant choice for the emerging high speed wireless standards. One of the main
processing blocks in such systems is dedicated to combating and compensating for
the fading frequency selective wireless channels. Thus, channel estimation is a cru-
cial signal processing operation in modern wireless transceivers. To assist channel
estimation, it is a common practice in wireless communications jargon to use pilots.
Scattered pilots have been used in a number of existing state-of-the-art commu-
nications systems. For instance, LTE and DVB-H benefit from different forms of
scattered pilots in their downlink transmission. Considering the very fact that it is
desired to use low-complexity algorithms to address the channel estimation needs
of the mobile terminals, a number of approaches to low-complexity estimators have
been probed and discussed in this thesis. The focus of the work has been on devel-
oping algorithms with low hardware requirements in terms of processing elements as
well as on-chip memory, when possible. In particular, it is shown how certain fami-
lies of linear estimators can be manipulated into low-complexity variants which can
be used as alternatives for channel estimation with a certain amount of performance
degradation. Besides, the estimators have been tailored to the needs of two existing
OFDM standards, i.e., LTE and DVB-H. In the event of co-integration of these stan-
dards, the innate analogy between them, in terms of OFDM system parameters and
their respective pilot patterns, is exploited to reduce the overall hardware require-
ments for channel estimation purposes. For instance, it is shown that by exploiting
the similarities in 2nd-order channel statistics, the employed channel estimators can
be shared between LTE and DVB-H across their modes of operation. Furthermore, a
certain strategy can be pursued for pilot pattern design in OFDM systems such that
it is possible to reuse the designed estimators in multi-mode or/and multi-standard
environments.
iii
iv
Contents
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
General Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Wireless OFDM systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Narrowband wireless channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Wideband wireless channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Second-order wireless channel statistics . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Multi-carrier communications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.5 OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.6 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3 Channel estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1 Pilot-aided channel estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Minimum mean square error (MMSE) channel estimation . . . . . . 16
3.3 Low-complexity MMSE channel estimation . . . . . . . . . . . . . . . . . . . . 18
3.4 Merged-pilot channel estimation strategy . . . . . . . . . . . . . . . . . . . . . . 19
3.5 Filter-tap memory optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4 Channel estimation in LTE and DVB-H . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1 LTE physical layer specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2 DVB-H physical layer specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.3 Multi-mode filter-tap memory optimization in LTE . . . . . . . . . . . . 26
4.4 Multi-mode filter-tap memory optimization in DVB-H. . . . . . . . . . 29
4.5 Multi-standard filter-tap memory optimization across LTE and
DVB-H . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5 Summary of included papers and technical reports . . . . . . . . . . . . . . . . . 30
5.1 Paper I - Channel estimation for a mobile terminal in a multi-
standard environment (LTE and DVB-H) . . . . . . . . . . . . . . . . . . . . . 31
5.2 Paper II - Low-complexity channel estimation for LTE in fast
fading environments for implementation on multi-standard plat-
forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.3 Paper III - On coefficient memory co-optimization for channel
estimation in a multi-standard environment (LTE and DVB-H) . 33
5.4 Technical Report - Channel estimation filter-tap memory-optimized
pilot pattern design for OFDM systems in multi-mode and multi-
standard environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
7 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
I Channel estimation for a mobile terminal in a multi-standard
environment (LTE and DVB-H) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2 Standards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.1 LTE (Long Term Evolution) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
v
2.2 DVB-H (Digital Video Broadcasting-Handheld) . . . . . . . . . . . . . . . . 44
3 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1 MMSE estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 Robust MMSE (RMMSE) estimator . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3 Modified Robust MMSE (MRMMSE) estimator . . . . . . . . . . . . . . . 51
4.4 DFT-based estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5 Hardware implementation issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.1 LTE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.2 DVB-H . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
II Low-complexity channel estimation for LTE in fast fading en-
vironments for implementation on multi-standard platforms. . . 63
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2 LTE (Long Term Evolution) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
III On coefficient memory co-optimization for channel estimation
in a multi-standard environment (LTE and DVB-H) . . . . . . . . . . . 77
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
2 LTE and DVB-H . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4 Coefficient memory co-optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
IV Channel estimation filter-tap memory optimized pilot pattern
design for OFDM systems in multi-mode and multi-standard
environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
2 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3 Pilot-pattern design in OFDM systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5 Filter-tap memory optimized pilot-pattern design in OFDM systems 99
5.1 Pilot-pattern design for an arbitrary OFDM system . . . . . . . . . . . . 100
5.2 Pilot pattern design for OFDM systems compatible with a stan-
dardized OFDM system, e.g., LTE . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
vi
Preface
This thesis consists of a General Introduction and the following scientific papers
which are referred to in the text by their roman numerals:1
I. Farzad Foroughi Abari, Johan Löfgren, and Ove Edfors, "Channel estimation
for a mobile terminal in a multi-standard environment (LTE and DVB-H)",
in Proc. of IEEE Int. Conf. on Signal Process. and Commun. Syst., Omaha,
NE, Sep. 2009.
II. Farzad Foroughi Abari, Farnaz Sharifabad, and Ove Edfors, "Low-complexity
channel estimation for LTE in fast fading environments for implementation
on multi-standard platforms", in Proc. of IEEE Veh. Tech. Conf., Ottawa,
Canada, Sep. 2010.
III. Farzad Foroughi Abari, Fredrik Rusek, and Ove Edfors, "On coefficient mem-
ory co-optimization for channel estimation in a multi-standard environment
(LTE and DVB-H)", in Proc. IEEE 8th Int. Workshop Multi-Carrier Syst.
and Solutions, Herrsching, Germany, May 2011.
IV. Farzad Foroughi Abari, Fredrik Rusek, and Ove Edfors, "Channel estimation
filter-tap memory optimized pilot pattern design for OFDM systems in multi-
mode and multi-standard environments", Tech. Rep., Dept. of Elec. and
Info. Technology (EIT), Lund, Sweden, June 2011.
Other publications by the author

I. Farnaz Sharifabad, Farzad Foroughi Abari, and Ove Edfors, "Low-complexity
channel estimation for low-mobility LTE using 4 × 4 MIMO", in Proc. the 6th
IEEE Conf. Ind. Electron. and Applicat., Beijing, China, June 2011.
II. Isael Diaz, Balaji Sathyanarayanan, Alirad Malek, Farzad Foroughi, and Joachim
Neves Rodrigues, "Highly Scalable Implementation of a Robust MMSE Channel
Estimator for OFDM multi-standard Environment", accepted to IEEE Workshop
Signal Process. Syst., Beirut, Lebanon, Oct. 2011.
1
The order of the authors names indicates their relative contributions to the publications.
vii
General Introduction
Farzad Foroughi Abari

1 Introduction 3
1 Introduction
Gradually replacing single-carrier communications systems, orthogonal frequency
division multiplexing (OFDM) has been a driving force behind the latest high speed
standards in wired as well as wireless communications. The ability to divide a wide
spectrum into narrowband subchannels, where data could possibly be multiplexed
for different users, has provided a number of different possibilities, including but not
limited to, reducing or eliminating the inter symbol interference (ISI), using simple
equalizers and employing flexible channel estimators.
Although the narrowband characteristics of tones in a given OFDM system sim-
plifies the equalization at the receiver side, the effects of the wireless channel ex-
perienced by the tones still needs to be compensated. As a result, channel state
information (CSI) needs to be supplied to or estimated at the receiver. A number
of approaches are suggested in literature to aid channel estimation at the receiver
in the presence of additive white Gaussian noise (AWGN). These approaches range
from simple linear to complex nonlinear alternatives, each having a certain com-
plexity/performance characteristic attached to them. Thus, the estimators can be
tailored to the characteristics of each standard. One of the dominating decision
criteria is the availability (and the pattern) of pilot data which can be used to aid
channel estimation in a given system.
The pilot patterns are usually categorized into continuous and scattered vari-
ants. For instance, continuous pilots in the form of preamble have been proposed
in IEEE 802.11n [1] for high speed data communications through Wireless Local
Area Network (WLAN). For systems experiencing higher amount of fading, such as
Long Term Evolution (LTE) [2], scattered pilots are usually employed. The presence
of scattered pilots not only aids instantaneous estimation of the channel but also
provides the possibility for tracking channel variations in time.
Scattered pilots can be thought of sampling a 2-D grid. For example, if a number
of consecutive OFDM symbols, extending over the time axis, are unfolded in the
frequency direction a 2-D time-frequency grid is constructed. The building blocks
in the constructed grid are tones where one tone is a subchannel in a given OFDM
symbol. In addition, if the grid is sampled according to some random or deterministic
pattern, a subset of points, called pilot tones, is collected which can be used to assist
a number of signal processing operations in a transceiver. For instance, one of the
main applications is to aid the estimation of the wireless channel, known as pilot-
aided channel estimation by convention.
Thus, pilot-aided channel estimation can be described as a mapping operation
between the sampled time-frequency space, i.e., the pilots space, to the full 2-D space
where all the required tones are estimated or interpolated. The mapping function
may denote a linear or nonlinear operation and depending on its characteristics, the
pilot pattern can be optimized for a certain system. For instance, the pilot sampling
frequency should comply with the Nyquist criteria if a linear mapping is desired.
On the other hand, Nyquist criteria may be violated for some nonlinear operations,
e.g., iterative joint channel estimation and data detection [3, 4].
Conventionally, the pilots are designed for a certain system constrained by a
4 General Introduction
number of requirements. The design parameters can cover a wide range of require-
ments such as high performance in fast-fading environments, low latency, and high
capacity, just to name a few. However, to the best of the author’s knowledge, little
work has been done with respect to designing pilot patterns which are optimized for
co-implementation and integration of several OFDM systems in one single platform.
The idea deserves a growing attention especially when the astonishing number of
different standards that future mobile terminals need to handle are taken into ac-
count. In fact, it is no longer far fetched to imagine a smart phone handling high
rate data communications under one standard while simultaneously streaming high
resolution multimedia content.
On the other hand, hardware technology does not usually keep pace with the
exponential growth in signal processing requirements in mobile terminals. Thus,
it is at times inevitable to relax the stringent requirements on the performance
of the optimal algorithms to achieve a certain performance/complexity trade-off.
For instance, optimal linear estimators designed for channel estimation purposes
may find little interest when their application in low-power terminals, with limited
hardware resources, is taken into consideration. In this thesis, an effort has been
made to address some of the performance/computational complexity trade-offs with
respect to low-complexity channel estimators.
The main contribution of this thesis covers the following.
• A number of low-complexity channel estimators for LTE and Digital Video
Broadcasting-Handheld (DVB-H) [5] are probed and their performance in
terms of mean square error and bit error rate is evaluated. The studied esti-
mators include the MMSE and DFT-based variants where MMSE estimators
have been more thoroughly analyzed.
• Depending on the fading characteristic of the wireless channel, a number of

strategies, e.g., pilot merging, for alleviating the time-domain filtering are
proposed. Furthermore, The effect of pilot merging on the performance of
the channel estimator is evaluated. Taking the performance degradation into
account, various strategies for pilot merging are proposed.
• It is shown how similarities between channel estimation filter taps can be used
to reduce the on-chip memory needed for the storage of the filter coefficients.
Moreover, it is shown that by seeking certain pilot merging schemes the on-chip
memory for the storage of filter taps can be optimized for LTE and DVB-H
across their various modes of operation.
• Last but not least, a new approach to pilot pattern design in OFDM systems
is proposed. The suggested strategy maximizes the number of OFDM systems
which can be co-integrated in the same platform by employing similar filters
for channel estimation purposes.
As a result, the following structure has been adopted in this thesis. In Sec. 2
a brief introduction to wireless OFDM systems is provided and their prominent
advantages and possible shortcomings are discussed. In Sec. 3, a mathematical
2 Wireless OFDM systems 5
description of minimum mean square error (MMSE) family of channel estimators is

given. Furthermore, it will be shown how a certain family of low-complexity estima-
tors can be derived through relaxing the stringent requirements on the estimators’
performance. Moreover, The performance of the estimators at various simplifying
stages is probed. Also, simple methods to avoid time-domain filtering for moder-
ately fast fading wireless channels are discussed. The nuts and bolts of filter memory
optimization are elaborated in Sec. 3.5.
After the general introduction three papers and a technical report, attached as
separate chapters, finalize the discussions in this thesis. Papers I-III elaborate on
the principles for channel estimation and filter memory optimization discussed in
the general introduction. The final chapter, i.e., the attached technical report, how-
ever, wraps up the thesis by providing a novel and methodological approach to pilot
pattern design for OFDM systems. The design criteria is to enable the proposed
estimators to be shared among a maximum number of modes in the same system.
Besides, it provides the feasibility to apply the filters to other compatible OFDM
systems co-implemented in multi-standard platforms. Furthermore, the pilot pat-
tern design for OFDM systems in a multi-standard platform, where one OFDM
system is already fixed (standardized), will be discussed in 5.2. Remaining a domi-
nant high-rate data communications standard for the foreseeable future, compatible
pilot-pattern design for systems co-implemented with LTE is discussed through an
example in Sec. 5.2.
2 Wireless OFDM systems

The communications channels in wireless environments are characterized by the
reflection of the transmitted radio signal from the objects in the environment. In
the event of a line-of-sight (LOS) communications, the secondary reflections from
the objects in the environment may have a negligible effect on the performance
of the radio communication. However in many scenarios, there is either no LOS
component or the reflections from the objects are too strong to be ignored.
There are different approaches to the characterization of the wireless channels.
In ray tracing [6], e.g., a deterministic channel model is usually sought. Although ray
tracing provides a good insight into certain properties of the wireless environment, it
cannot precisely describe and track the time-variations as well as object movements
in the environment. Specifically, it is quite complicated to characterize the reflections
when the number of existing objects increases. Thus, another approach, based on
probability theory, has been adopted to describe the wireless channels and their
properties. In the following, a concise and statistical description of narrowband and
wideband wireless channels is provided.
2.1 Narrowband wireless channels

In the event of no line of sight (NLOS), the simplest example of a narrowband chan-
nel is the existence of one reflecting object in the wireless environment. As a result,
the wireless channel response is composed of a single tap. However, in the majority
of practical cases, multiple reflecting objects exist. But, if the reflections from the
objects arrive at the receiver in a sufficiently narrow time interval such that the
receiver cannot resolve the individual reflections, the sum of the reflections consti-
tute a single tap which categorizes it as a narrowband channel. Thus, a simplified
baseband representation of a narrowband channel can be described as
hcont. (τ, t) = α0 (t)δ(τ − τ0 ), (2.1)
where α0 (t) = E |α0 (t)|2 exp (2πfD t − 2πfc τ0 ) and τ0 is the arrival time. Further-

more, fD and fc are the Doppler and carrier frequencies, respectively.
2.2 Wideband wireless channels

In the event of multiple resolvable reflections, the single-tap channel model does
not hold any more and a more general description needs to be sought. Thus, for
a wide sense stationary uncorrelated scattering (WSSUS) wireless channel with M
resolvable components, (2.1) becomes [7]
M
X −1
hcont. (τ, t) = αl (t)δ(τ − τl ), (2.2)
l=0
where the lth component in (2.2) is either a single reflection or the sum of a cluster
of non-resolvable reflections, as mentioned in Sec. 2.1. Besides, αl (t) are inde-
pendent zero-mean complex Gaussian processes with Rayleigh distributed ampli-
tudes. Furthermore,
PM −1 the2 power delay profile (PDP) associated with (2.2) is defined
2
as θ(τ ) = l=0 E |αl | δ(τ − τl ) where E |αl | is the variance of tap l.
Due to the fact that the majority of signal processing in today’s transceivers
is performed in discrete time, a discrete-time format of (2.2) needs to be sought.
Thus, if the wireless channel remains time-invariant during the transmission of a
single data symbol m, (2.2) can be written as
M
X −1
hcont. (τ, m) = αl (m)δ(τ − τln Ts ), (2.3)
l=0
where Ts is the sampling interval at the transceiver and τln is the normalized arrival
time, i.e., τln = τl /Ts . Moreover, if the available symbol bandwidth (BW) is divided
into N equal frequency intervals of ∆f , the discrete-time frequency response of (2.3)
at discrete time m becomes
M
X −1
h [k, m] = αl (m) exp(−2πj(k/N)τln ). (2.4)
l=0
where h [k, m] is the channel attenuation on discrete tone k at time m.

2.3 Second-order wireless channel statistics

For the presumed WSSUS channel in (2.4), the cross correlation between subcar-
riers k and k ′ corresponding to channels associated with data symbols m and m′
respectively is
rh [k − k ′ , m − m′ ] = E {h[k, m]h∗ [k ′ , m′ ]} , (2.5)
where (·)∗ denotes the complex conjugate operation.
After some mathematical manipulation it can be deduced that
rh [k − k ′ , m − m′ ] = rh,f [k − k ′ ]rh,t [m − m′ ] (2.6)
where rh,f is the correlation among subcarriers at a fixed time instance in frequency
direction and rh,t is the correlation between two instances of the same subcarrier at
times m and m′ . It can be observed from (2.6) that the channel correlation is the
product of correlation in time and frequency. This is known as separable correlation
in literature.
The channel correlation in frequency, rh,f [k − k ′ ], is a function of the PDP. For
example, it can be shown [8] that for a uniformly distributed PDP, θ(τ ), where all
taps constitute equal energy rh,f reduces to
1 − exp(−2jπL∆k/N)
rh,f [∆k] = , (2.7)
2jπL∆k/N
where ∆k = k−k ′ is the distance between two subcarriers, e.g., a pilot and estimated
tone, and L is the PDP spread in discrete time samples.
The channel correlation in time, rh,t [m − m′ ], may be characterized in various
forms depending on the properties of the fading environment. For instance, if the
zeroth-order Bessel function of the 1st kind is denoted as J0 (·), then
rh,t [m − m′ ] = J0 (2πfd (m − m′ )) (2.8)
for a 2-D scattering environment meeting certain conditions [9]. In addition, if

sinc(x) = sin(x)
x
it has been shown that
rh,t [m − m′ ] = sinc(2πfd (m − m′ )) (2.9)
for a 3-D scattering environment constraint to certain conditions [10]. Furthermore,

fd in (2.8) and (2.9) denotes the normalized Doppler spread, i.e., fd = Dp Tu /2,
where Tu is the data symbol duration and Dp is the Doppler spread. Besides, it is
worthwhile to remind the reader that the channel variations in time are presumed
to be negligible for Tu .
It can be observed from (2.9) that fd has a significant effect on the correlation in
time. In fact, the lower the fd , the higher the correlation in time. For instance, Fig.
1 illustrates the time correlation, for a 3-D scattering environment, versus time for
a number of different terminal velocities where the carrier frequency, fc = 2.6 GHz.
It can be seen that the wireless channel exhibits a quite strong time correlation
even for high terminal speeds. As will be elaborated in subsequent sections, LTE’s
1
v=30 km/h
v=60 km/h
0.8 v=120 km/h
0.6
0.4
rh,t
0.2
−0.2
−0.4
0 20 40 60 80 100
Tu
Figure 1: Time correlation versus time. The time axis is scaled by the LTE’s
symbol time duration, i.e., Tu = 66.7 µs, for illustration purposes. It can be observed
that for terminal velocities as high as 120 km/h, there is still a strong time correlation
among a few consecutive OFDM symbols.
symbol duration is Tu = 66.7 µs. Thus, the time correlation remains almost strong
during the transmission of a few OFDM symbols even when the terminal speed
approaches as high a velocity as 120 km/h. As will be discussed in the following
sections, this important property has been exploited to contemplate low-complexity
channel estimators where time-domain filtering may be avoided.
Meanwhile, Fig. 2 illustrates the absolute value of correlation in frequency asso-
ciated with (2.7) for a number of different NL , where N is the number of DFT points.
An interesting observation can be made from this figure. The subcarriers decorre-
late quite fast for the exemplary values, which are typical of the ones expected in
LTE and DVB-H, as elaborated in Secs. 4.1 and 4.2. The above mentioned fact
will be used in the subsequent sections to facilitate the derivation of low-complexity
channel estimators.
2.4 Multi-carrier communications

Single-carrier communications were once dominating the high-speed wireless commu-
nications arena. In a single-carrier system binary data, after going through various
source and channel coding stages, is modulated and and transmitted over the air.
To maximize the throughput the symbols are inserted sequentially in a compact
pulse train before transmission. In the meantime, the symbol duration in a con-
ventional single-carrier system is reduced to provide the opportunity for high rate
1
L/N=1/4
L/N=1/8
0.8 L/N=1/16
L/N=1/32
0.6
0.4
rh,f
0.2
−0.2
−0.4
0 10 20 30 40 50
∆k
Figure 2: Frequency correlation versus tone separation, ∆k. It can be observed

that there is a substantial correlation loss among subcarriers which are located far
apart, especially for larger values of NL .
signaling. Had the transmission been materialized in a flat-fading wireless channel,

this method would have provided an efficient alternative for fast data transmission.
However, the majority of wireless channels in practice are frequency selective and
exhibit a substantial amount of fading. As a result, the symbols experience severe
inter symbol interference (ISI) before arriving at the receiving end, requiring sophis-
ticated equalization techniques to retrieve the data. The effect of ISI would become
more detrimental in the event of a higher frequency selectivity and a shorter symbol
interval.
To combat the frequency selectivity in wide-band channels, multi-carrier mod-
ulation was proposed an an alternative [11]. The idea behind multi-carrier modu-
lation is to restrict the occupied BW of each symbol to a sufficiently narrowband,
called subchannel, subband, or tone so that each symbol experiences little or no
fading upon transmission through a frequency selective channel. In other words,
the symbol time is increased so much as the wireless channel’s delay spread be-
comes negligible in comparison. The original idea behind multi-carrier modulation,
as described above, suffered from the fact that one cannot arbitrarily narrow down
the symbol BW without increasing its length in time. Considering the fact that the
symbol duration in time has to be limited, a penalty factor in terms of increased BW
needs to be paid. Besides, such a scheme suffers from expensive narrowband filters
while requiring a large number of modulators/demodulators at the transceiver.
Figure 3: The concept of orthogonal signaling in frequency. The peculiar spacing

of signals insures no ICI.
2.5 OFDM
Following the multi-carrier concept, during the 60s [12, 13], it was realized that the
pulses (data symbols) may overlap each other in frequency without interference. The
overlap operation needs to be done in such a way that the pulses remain orthogo-
nal to one another. As a result, the penalty in BW for a sufficiently large number
of subchannels becomes insignificant [11]. Inserting a number of data symbols on
a number of overlapping orthogonal pulses is known as orthogonal frequency data
multiplexing, i.e., OFDM, and is the principle behind high rate multi-carrier com-
munications. Fig. 3 illustrates how a number of well-known sinc pulses can overlap
each other while remaining orthogonal. Moreover, N such symbols can be collected
into a single OFDM symbol and modulated to the desired carrier frequency for
transmission. In other words, OFDM converts a frequency selective channel into a
non-frequency selective one [14]. Considering the fact that the encapsulated OFDM
symbols can be much longer in time when compared to their single-carrier counter
parts, they experience less severe ISI when transmitted over a frequency selective
wireless channel. Moreover, to eliminate the effect of ISI, a cyclic prefix (CP), also
known as guard interval (GI), whose length is equal to the maximum expected de-
lay spread of the channel is appended to each symbol before carrier modulation and
final transmission [15].
The CP can appear in one of the following two forms. Either the last M samples,
where M equals the delay spread of the wireless channel, in each OFDM symbol
are copied and appended to the beginning of the symbol or M zeros are appended
instead. The former converts the linear convolution into a circular convolution
without additional operation at the receiver. In the latter, however, the receiver
needs to collect the last M samples of the received symbol and copy it into the
beginning of the corresponding symbol, converting a linear convolution to a circular
convolution [16]. Both approaches allow the partitioning of the frequency selective
channel into a number of parallel flat-fading subchannels without requiring the CSI
at the transmission side. In retrospect, there are other approaches such as vector
coding (VC), also known as precoding, where the CSI is needed at the transmitter
to allow a similar operation [16].
Although more robust against ISI, OFDM systems may suffer from a number
of drawbacks. For instance, the presence of inter carrier interference (ICI) results
in a loss of orthogonality among subcarriers in OFDM systems. In other words,
the orthogonality among the signals may be jeopardized due a multitude of reasons
including but not limited to fast fading environments, carrier frequency offset, and
timing jitter. If hardware imperfections at the transmitter or receiver, e.g., I/Q
imbalance [17], are neglected, fast fading is the main reason behind loss of orthogo-
nality resulting in performance degradation of the system. A simple way to combat
ICI is to adjust the OFDM symbol duration so that each symbol experiences little
or no fading during its transmission. Thus, fading places a limit on the duration of
OFDM symbols and its resilience to ISI. High peak-to-average power ratio (PAPR)
is another prohibiting factor in practical OFDM system implementations. High
PAPR can place stringent requirements on the linearity of power amplifiers both at
the receiving and transmitting sides requiring costly design and components.
Practical implementations of OFDM systems were not realized during the early
days of telecommunications due to hardware shortcomings. As digital hardware de-
sign became more mature and digital signal processors (DSP) became more powerful
and efficient, OFDM systems became more popular. Probably, the most critical bur-
den associated with the practical implementation of OFDM systems has been the
insertion of overlapping pulses in specific intervals such that the orthogonality of
the whole pulse train is secured. Implementing such system in analogue domain
is cumbersome due to deviations of analogue electronics from precise mathemati-
cal descriptions. However, if available, OFDM modulation/demodulation may be
carried out in digital domain through a conventional inverse fast Fourier transform
(IFFT)/fast Fourier transform (FFT) operations.
The baseband equivalent of OFDM modulation can be stated as
N
X −1
s(t) = x[k] exp(j2πk∆f t), (2.10)
k=0
where ∆f is the inter-carrier distance between the modulation frequencies, k refers

to the kth subchannel and x(k) corresponds to the data selected from a constellation
map, e.g., 16QAM. In the event of proper sampling, i.e., fs = N∆f , (2.10) becomes
an inverse discrete Fourier transform (IDFT) and may be written as

N
X −1
s[n] = x[k] exp(j2πkn/N), (2.11)
k=0
where N is the number of samples in an OFDM symbol. DFT/IDFT can be effi-

ciently carried out by FFT and IFFT algorithms. Furthermore, exponential growth
in integrating more area-efficient transistors on the same chip provided the oppor-
tunity to integrate dedicated fast and power-efficient FFT/IFFT accelerators on a
transceiver. Thus, commercial interest in OFDM was ignited as digital hardware
technology became more mature and made its practical implementation feasible.
As a result of orthogonal modulation, each transmitted data symbol experiences
flat fading due to its narrow-band characteristics. However, due to frequency-flat
fading, subcarriers experiencing low signal to noise ratios (SNR), result in consider-
able bit error rate (BER) degradation. This phenomenon is similar to single-carrier
systems in flat-fading environments. Thus, various coding and interleaving schemes,
over time and frequency, can be practiced to mitigate or reduce the detrimental
effects. Examples of such coding and interleaving schemes in practical applications
can be found in [2, 5]. A detailed description of such operations, however, is beyond
the scope of this thesis.
2.6 System model

The adopted system model in this thesis is a single-input single-output (SISO)
OFDM system operating in a fading and frequency-selective wireless environment.
The choice of SISO system is due to its simplicity for illustration of the introduced
concepts . However, the acquired results may be easily extended to multiple-input
multiple-output (MIMO) OFDM systems. The proposed OFDM system is charac-
terized by the following parameters,
• Tu , [µs] = useful OFDM symbol du- • Doppler spread (Dp ) and delay
ration, spread (Ds ),
• N = number of DFT points, where

• TCP , [µs] = CP duration, N = 2q and q is a positive integer.
Furthermore, other parameters of interest associated with the system can be

easily deduced from the above system specifications. For instance, one may figure
out that ∆f = 1/Tu , BW = fs = N/Tu , where ∆f is the subcarrier spacing, fs is
the transceivers sampling frequency, and BW is the occupied bandwidth.
Fig. 4 illustrates a possible configuration for the proposed system model. The
adopted wireless channel is a wide-band channel similar to (2.2). Following the signal
flow depicted in Fig. 4, data corresponding to one OFDM symbol are inserted on the
designated tones at the transmitter and fed into an N-point IFFT whose output is
given by (2.11). Then, the CP is inserted at the output of the IFFT block and after
going through digital-to-analogue (D/A) conversion and modulation to the proper
3 Channel estimation 13
Ant # T1 Ant # R1
Binary Data Wireless CSI

fc channel fc
DMUX
MUX
FFT
DMUX
IFFT
MUX
Interleaving
CP D/A ~ A/D CP
Deinterleaving
Decoding
+
Coding
+
Modulation Demodulation
Figure 4: The adopted SISO OFDM system model.
carrier frequency the outcome is transmitted over the wireless channel modeled by
(2.2).
Upon reception at the receiver, the signal is picked up by the receiving antenna
and after carrier demodulation, low-pass filtering, analogue-to-digital (A/D) conver-
sion, and CP removal, the result is fed into an N-point FFT. If the signal samples
before FFT block are collected into the column vector r of size N, then
r = Hs + η, (2.12)
where H is a N × N circulant matrix associated with (2.4) for the mth OFDM
symbol, s corresponds to the N × 1 vector of data in (2.11) and η is the N × 1
vector of AWGN. The circulant matrix H is the direct result of employing CP in
OFDM systems. Furthermore, if the unitary Fourier matrix of size N is denoted by
Tn , the output of FFT equals
y = Tn r = Tn Hs + Tn η. (2.13)
However, it can be shown that following (2.11), s = TH n x, where x is the column

vector of transmit data having size N. Considering the fact that the unitary Fourier
matrix Tn and its Hermitian, i.e., THn , diagonalize any circulant matrix [18], (2.13)
can be written as
y = Λx + η, (2.14)
where Λ = Tn HTH n is a diagonal matrix containing the channel attenuations corre-
sponding to (2.4) for a fixed OFDM symbol, i.e., discrete time instance m. Moreover,
(2.14) can be rewritten as
y = Xh + η, (2.15)
where X is a diagonal matrix of size N ×N encompassing the transmitted data x(k),
and h is a column vector of size N containing the channel attenuations in frequency.
3 Channel estimation
Channel estimation is an essential part of coherent data detection in OFDM sys-
tems [19]. A lot of work has been carried out for channel estimation in wireless
communications. With the advent of OFDM and its practical realization through
FFT/IFFT hardware accelerators, channel estimation research picked up for such
systems and ever since various methods have been suggested in literature [7, 8, 20–
22]. The proposed methods range from simple linear estimators to complicated
iterative estimation in the presence of channel coding [4, 23]. In general, one may
categorize the channel estimation algorithms as linear and nonlinear estimators.
Quite contrary to a linear approach where the CSI can be obtained through, e.g.,
a matrix multiplication, nonlinear methods are characterized by multi-stage process-
ing blocks where at least one of the stages encompasses certain nonlinear operations.
For instance, joint estimation and detection approaches comprise iterative informa-
tion passage between the decoder as well as the channel estimation blocks. These
approaches usually have a higher performance when compared to the conventional
linear methods at the expense of higher computational complexity. Meanwhile, in
the event of pilot-aided channel estimation, these approaches may need less pilot
overhead, even below the threshold set by the Nyquist sampling criteria [4]. An
elaborate analysis of nonlinear channel estimation approaches needs a book of its
own and is well beyond the scope of this thesis.
Due to their inherent simplicity and low complexity, linear estimators have been
of high interest in practical system implementations. Besides, their existence is
often independent of other transceiver subsystems, e.g., decoders when compared to
some nonlinear methods. A linear estimation operation can be described as a simple
matrix multiplication such as
ĥ = Wŷ, (3.1)
where ŷ is a column vector whose elements are the measured signal values and W
is the estimation filter matrix.
3.1 Pilot-aided channel estimation

To be able to capture the CSI in a communications system, one often needs to rely
on the (noisy) observations of the received signal values. Moreover, the estimator,
e.g., at the receiver, usually needs to know the transmitted signal values to estimate
the channel. Yet, there are classes of estimators, known as blind estimators, which
rely only on the received signal and its pseudo-stationary properties to do the es-
timation [24–27]. Their application, however, is prohibitive in low-power terminals
with limited processing resources due to their high computational complexity.
In the majority of practical cases the transmitter and receiver should have a priori
knowledge of the transmitted data for the purposes of channel estimation. Thus,
the transmitter and receiver agree on a subset of data which is called pilot-data by
convention and is extensively used in practical system implementations. The pilots
are inserted in specific locations in the transmitted signal. They appear either in
continuous or scattered patterns. The contiguous scheme is usually employed in
wireless systems where little fading is expected. For instance, IEEE 802.11n [1]
uses continuous transmission of pilots for the purposes of channel estimation in its
preamble. Contrary, if the wireless channel experiences a considerable amount of
fading, e.g., movements in the environment, scattered pilots distributed both in
time and frequency need to be used so that the channel variations over time and
frequency can be tracked.
Figure 5: An example of rectangular pilot pattern scattered in time and frequency.

α and β are pilot separations in frequency and time, respectively.
An example of scattered pilot pattern in OFDM systems is depicted in Fig. 5.

This is known as rectangular pilot pattern where pilots appear in specific intervals,
both in time and frequency. Furthermore, the pilots associated with different OFDM
symbols appear exactly on the same tones. On the other hand, the scattered pilots
do not necessarily need to appear on the same tones in each pilot carrying OFDM
symbol. In other words, the pilots of a specific pilot-carrying OFDM symbol may
experience a distinct frequency shift. This is known as skewed pilot pattern. If
the pilot carrying OFDM symbols are arranged such that the pilot frequency shift
increases monotonically with respect to the symbol time index, a distinct skewed
pilot pattern, demonstrated in Fig. 6, can be achieved. This is known as a primary
skewed pilot pattern in this thesis.
The scattered pilots can be viewed as a 2-D sampling of a time-frequency grid.
Thus, Nyquist’s sampling criteria, in time and frequency, needs to be met if a proper
reconstruction of the original channel is desired. Thus, the sampling can not fall
below a certain threshold set by the Nyquist’s sampling criteria as well as certain
system dynamics such as the presence of AWGN. On the other hand, the pilots
introduce spectral inefficiency in communications systems since a portion of the
BW needs to be allocated to pilots rather than data transmission. Besides, although
increasing the number of pilots results in better estimation performance, the overall
SNR scales down proportionally since more energy is allocated to the pilots rather
than useful data transmission.
Figure 6: A primary skewed pilot pattern. The pilot frequency offset increases
monotonically with respect to the symbol index time. α and β are the distances
between pilots in frequency and time respectively. γ is the pilot frequency offset
and Γ is the distance between two consecutive pilot-carrying OFDM symbols. The
above mentioned parameters are related through β = Γ( αγ ).
3.2 Minimum mean square error (MMSE) channel estima-

tion
Being a linear algorithm, a full linear MMSE (LMMSE) estimator provides the
ultimate performance in terms of MSE among its linear counterparts. The LMMSE
estimators make use of several parameters such as SNR as well as 2nd-order statistics
of the channel. In the following, the mathematical derivations for LMMSE channel
estimation in the event of rectangular pilot pattern is derived. The results can be
easily extended to skewed pilot scheme with minor mathematical manipulations.
Suppose in the system exemplified by Fig. 4, L consecutive OFDM symbols
are collected for channel estimation purposes. As a result, compared to (2.15), h
corresponds to a vector of size LN × 1 encompassing the channel tones associated
with L consecutive OFDM symbols stacked on top of each other. Besides, the pilot
spacing in terms of the number of subcarriers in time and frequency are α and β
respectively, see Fig. 5. If the variables associated with pilots are identified by (·)p ,
the linear channel estimation for the above mentioned grid can be written as
ĥ = Wŷp , (3.2)
where ŷp is a Np × 1 column vector containing the observations of transmitted

pilots at the receiver. In other words, for an observation interval of N × L tones,
yp is the column vector of collected pilots from the pilot carrying OFDM symbols
stacked on top of each other. In minimum mean square error (MMSE) estimators,
the LN × Np matrix W is designed to minimize the MSE between the actual and
estimated channel. In other words, the intention is to find
2
min E h − ĥ . (3.3)

It can be shown that the solution to the above optimization problem is
W = Rhyp R−1
yp , (3.4)

where Rhyp = E hypH is the LN × Np matrix of cross correlation values between

the channel and the received pilots, Ryp = E yp ypH is the Np × Np matrix of
autocorrelation associated with the received data on the pilot positions. Throughout
this thesis (·)H denotes Hermitian transpose operation.
After some mathematical manipulations and simplifications of (3.4), (3.2) be-
comes
2 H
−1 −1
ĥmmse = Rhhp Rhp + ση Xp Xp ĥp,ls , (3.5)
where ĥmmse is a column vector of estimated tones, ση 2 is the AWGN power and Xp
is a diagonal matrix containing the transmitted pilots stacked on its main diagonal.
Furthermore, ĥp,ls = X−1p yp is the least squares (LS) estimation of the pilots at
the receiver. Moreover, if the pilots are selected in a way that they all constitute
constant envelope, then XH p Xp may be replaced by an identity matrix scaled by the
energy of the pilots which can be further incorporated into the SNR. As a result,
(3.5) can be written as
−1
I
ĥmmse = Rhhp Rhp + ĥp,ls , (3.6)
SNR
where SNR is the signal-to-noise ratio, Rhhp is the cross-correlation matrix between
the pilot and data tones, and Rhp hp is the autocorrelation matrix of the pilot tones.
For the above mentioned 2-D filtering operation, the MSE performance associ-
ated with tone k in OFDM symbol l is
J(l, k) = σh2 − rhyp R−1 H

yp rhyp , (3.7)
where rhyp is the N(l − 1) + k row in Rhyp .

To tackle the computational complexity of the ideal 2-D filter, the filter can be
designed for a presumed 2nd-order channel statistics as well as SNR. For instance,
(3.4) can be designed for the worst-case PDP and Doppler spectrum. Thus, if filter
W is designed for a presumed 2nd-order statistics other than the actual values, the
MSE associated with the design mismatch equals
J(l, k) = σh2 − rhyp wH − wrH H

hyp + wRyp w , (3.8)
where w is the N(l − 1) + k row in W.

On the other hand, it has been shown that 2-D Wiener filters may be broken into
two 1-D Wiener filters [28, 29]. Although suboptimal, the performance of the filters
can be close to the performance of the 2-D filter while the computational complexity
is considerably reduced. For the purposes of channel estimation, the ideal 2-D filter
may be broken into Wt and Wf , i.e., filtering in time and frequency respectively.1
If filtering is first done in time domain, then
ĥ1 = Wt yt,p , (3.9)
where ĥ1 is the column vector of estimated tones on the pilot positions, yt,p is the
yt,p .
vector of observed pilot tones in time direction, and Wt = Rhyt,p R−1
Meanwhile, the estimated tone ĥ1 (l, k) can be written as
ĥ1 (l, k) = h1 (l, k) + η ′ (l, k), (3.10)
where η ′ (l, k) is a zero mean noise process with variance E |η ′ (l, k)|2 = J1 (l, k).

Moreover, J1 (l, k) is the MSE due to the first filtering operation, i.e., filtering in
time and can be expressed as
J1 (l, k) = σh2 − rht yt,p wtH − wt rH H

ht yt,p + wt Ryt,p wt , (3.11)
where the effects of model mismatch are included. Besides, wt is the kth row
in matrix Wt . Following the time-domain filter, the 2nd filtering operation, i.e.,
filtering in frequency is performed. As a result, the estimated tones in frequency for
one OFDM symbol may be written as
ĥ2 = Wf ĥ1 , (3.12)
where −1
Wf = Rhf hf,p Rhf,p + J1 I , (3.13)
and J1 is a diagonal matrix of MSE values from the time-domain filtering.
3.3 Low-complexity MMSE channel estimation

If the 1-D estimators in (3.9) and (3.12) are designed for actual channel statistics, a
formidable computational complexity still hits the system. On one hand, the chan-
nel statistics should be tracked and evaluated at each moment. On the other hand,
even if the statistics do not change fast and stay rather constant over a number
of OFDM symbols, the real-time matrix inversion consumes a lot of hardware re-
sources. Thus, it is but inevitable to design and pre-compute the 1-D filters for a
fixed set of variables in the wireless channel scenario to reduce the computational
complexity. For instance, the covariance matrices in Wf can be designed for a rect-
angular PDP whose length is equal to the maximum expected delay spread of the
1
Due to the linearity of the filters, the ordering can be done arbitrarily. This holds for a
rectangular pilot pattern and may not be true for a skewed pilot pattern. Besides, one should
be careful about changing the order of the filters when the calculations are carried out in block
matrices. Matrices do not generally commute.
wireless channel. Similarly, the covariance matrices in Wt may be designed for a

3-D scattering environment [10], i.e., a rectangular Doppler spectrum whose length
is equal to the maximum expected Doppler spread, Dp , in the channel. Throughout
this thesis, such estimators are called robust MMSE (RMMSE) estimators.
To further reduce the computational complexity as well as on-chip memory re-
quirements, one may alleviate the 1-D filtering in time domain. For instance, in
piece-wise constant approach, the channel estimation is performed for each pilot-
carrying OFDM symbol in frequency and the same estimation is used for equaliza-
tion purposes of the subsequent OFDM symbols until the next pilot carrying OFDM
symbol is received. In the event of skewed pilot pattern, another low-complexity esti-
mation strategy, called merged-pilot channel estimation, can be used. This strategy
usually outperforms the piece-wise alternative both in slow as well as moderately
fast fading environments and is elaborated in the following sections.
3.4 Merged-pilot channel estimation strategy

Imagine a primary skewed scattered pilot pattern illustrated in Fig. 6, where α is the
pilot tone separation in frequency, β is the pilot tone separation in time, γ is the pilot
frequency shift and Γ is the distance between two consecutive pilot-carrying OFDM
symbols. Besides, the pilots have been designed such that the Nyquist sampling
criteria, both in time and frequency, are met with an acceptable margin.2 For such
systems the pilots from a number of pilot carrying OFDM symbols can be collected
and merged into a pilot vector (PV) which is later fed into the frequency domain
channel estimator. The pilot spacing in PV depends on the merging pattern. The
estimated channel is further used for the equalization of the region encompassed by
the pilot contributing OFDM symbols. Fig. 7 illustrates an instance of a primary
skewed pilot pattern where pilot merging has been practiced.
The constructed PV in Fig. 7 (on the left) is one of the many ways pilot merging
can be practiced. For instance, if the pilots from alternate OFDM symbols are
collected into a PV (as seen on the right hand side in Fig. 7), the virtual pilot
spacing associated with the PV becomes 4 rather than 2. Furthermore, if the pilots
from OFDM symbols 1 and 4 are merged into a PV, an uneven virtual spacing is
attained. In this thesis, due to the discrimination of certain tones against others, the
merging schemes resulting in uneven spacing are not taken into consideration. Thus,
by convention, pilot merging is referred to the family of merging schemes where the
attained virtual pilot spacing remains constant throughout the constructed PV.
The pilot merging practice has a certain number of pros and cons attached to
it. However, depending on the merging pattern, it outperforms piece-wise constant
alternatives for the majority of merging scenarios. The merits are especially more
emphasized for slow as well as moderately fast fading environments. One could
attribute the performance gain to the higher number of available pilots for channel
estimation. In other words, the effective number of available pilots for channel
estimation in frequency direction is, at least, doubled in the event of merging. Thus,
2
Due to a limited signal observation interval, twice as fast as Nyquist criteria has been adopted
as a rule of thumb for proper sampling [29, 30].
Figure 7: An example of pilot-merging practiced for a primary skewed pilot

pattern.
a two-fold estimation performance increase in frequency direction is expected once

merging is done provided that the channel remains stationary. Meanwhile, the
estimation indirectly captures the time-variations of the wireless channel in fading
environments. In other words, merging can be imagined as a robust 2-D filtering
where the correlation coefficients in time domain associated with the pilots collected
from the pilot-carrying OFDM symbols all equal 1. Thus, it is expected that it
exhibits an acceptable performance up to a certain amount of fading, characterized
by a certain Doppler spectrum where the channel estimation error due to mismatch
surpasses a tolerable error threshold.
There are many ways to characterize the performance of the estimators. A simple
and convenient way is to quantify that through MSE. Thus, The performance of
pilot merging in various fading environments can be measured by MSE. Suppose
W denotes the robust 2-D filter designed for a certain fixed PDP and a stationary
channel in time, i.e., the channel is presumed to be time-invariant. In other words,
the effect of pilot merging can be modeled by setting rh,t [m − m′ ] in (2.6) equal to
1. As a result, (2.6) can be written as
M
X −1
rh [k − k ′ , m − m′ ] = θ(τln ) exp(−2πj(k − k ′ )τln /N). (3.14)
l=0
Furthermore, the MSE of the estimation operation can be measured by (3.11). Thus,
depending on the tolerable MSE, one may use the merging scheme in slowly as well
as moderately fast fading environments. In the event of bad MSE performance, an
additional Wiener filtering in time may kick in to reduce the error due to mismatch.
The merits of pilot merging are not limited to the above mentioned facts. In
skewed pilot patterns, one could attain several different virtual pilot separations
through merging or time-domain filtering. As will be discussed in the following
section, the possibility to vary the virtual pilot spacing, in the event of skewed pilot
pattern, allows filter tap memory optimization. Moreover, the same filter coefficients
may be re-used for channel estimation purposes in other OFDM systems, provided
that certain requirements are met. However, before moving on to the next sections
where the details are further elaborated, this section deserves one more discussion
before being wrapped up.
The number of available pilots in a pilot-carrying OFDM symbol equals N/α.
Thus, if the estimation is carried out in frequency domain only, while no merging
is practiced, N/α pilots can be used for channel estimation of each tone. However,
exploiting all pilots for estimation of one tone turns out to be unnecessary due to
the considerable correlation loss among the tones and pilots which are located far
apart, see Fig. 2. Thus, only a subset of pilots in the vicinity of each tone may
be used for channel estimation purposes. Furthermore, the pilots can be selected
symmetrically with respect to each tone. The combination of RMMSE estimators
as well as selecting a subset of pilots results in a new low-complexity channel esti-
mator called modified RMMSE (MRMMSE) estimator. The only problem with this
approach is the inability of symmetric pilot selection at the edge subcarriers where
a larger number of pilots may be used to compensate for the loss in estimation per-
formance. Meanwhile, the number of selected pilots can be optimized for a certain
metric, e.g., MSE. The proposed strategy in this thesis for channel estimation in
low-power terminals can be briefly described as,
• Choose an appropriate number of pilots in the vicinity of each tone for esti-
mation purposes. The selection criteria may be based on a certain tolerable
MSE, i.e., SNR degradation with respect to stationary environments,
• Pilot-merging should be practiced unless the performance degradation due to

mismatch surpasses a certain threshold. The threshold can be again based on,
e.g., a certain acceptable MSE. As soon as the performance criteria is not met,
additional 1-D time-domain filtering may be invoked to improve the mismatch
due to fading.
The above principle has been proposed for channel estimation in Multi-base
[31] where a number of high-speed standards need to be integrated into the same
platform. One may observe that due to selecting a subset of pilots, the amount
of on-chip memory for robust-filter tap storage decreases. In channels with high
frequency selectivity the size difference between the available pilots and the proposed
pilot subset may increase substantially. The combination of pilot-merging as well as
employing a subset of pilots for robust channel estimation is known as merged-pilot
modified robust MMSE (merged-pilot MRMMSE) channel estimation throughout
this thesis unless otherwise stated.
3.5 Filter-tap memory optimization

In section 3.4, it was discussed how robust channel estimators can be pre-computed
and stored in memory. The coefficients are later fetched during the estimation
process. Besides, it was shown how the computational complexity can be further
reduced if only a subset of pilots in the vicinity of each subcarrier is used for channel
estimation purposes while introducing a small loss in the estimation performance.
The loss can be quantified and optimized through calculations and Monte Carlo
simulations so that a desired number of pilots can be selected for a specific system
configuration.
For instance, imagine an OFDM system with N tones in one OFDM symbol
where zero-padding of the edge subcarriers has been ruled out for simplicity. Besides,
a pilot separation of α′ is achieved in the constructed PV where the number of
available pilots is Np,merg . Furthermore, it is decided that only a subset of the
available pilots, i.e., Np,merg
′
, where Np,merg
′
is a power of 2 and Np,merg
′
≪ Np,merg ,
may be used for estimation of each tone with an acceptable performance. As a result,
it is easy to conclude that the number of required filter taps for estimation of one
tone equals Np,merg
′
. A collection of such filter coefficients is called a filter set. Thus,
′
N ×α′
except 2 × p,merg. 2
− 1 edge subcarriers, a few filter sets can be shared among
all subcarriers for estimation purposes. More precisely, α′ filter sets are needed to
carry out the channel estimation for all tones other than the edge subcarriers.
Although the above discussed facts enable one to reduce the number of required
filter coefficients, the intension is to reuse a portion of these coefficients for channel
estimation purposes in other modes associated with the OFDM system of choice or
even in other compatible OFDM systems. OFDM system mode in here means the
one that is characterized by the employed CP.
Imagine the 1-D robust estimators designed for a rectangular PDP, i.e., the
worst-case frequency selectivity for a given delay spread. It can be observed from
(2.7) that the correlation coefficients are a function of N, L, and ∆k. Thus, as
long as the same ratio in L∆k/N is held, the same filter coefficient can be used for
different system configurations. In fact, it is desired to force
rh,f 1 [∆k] = rh,f 2 [∆k], ∀k, (3.15)
where index 1 refers to system configuration 1 and index 2 refers to system config-
uration 2. As a result, the estimators will be identical for similar values of SNR.
For example, imagine the OFDM system in Fig. 4. Suppose the system has two
different design CPs, L1 and L2, due to two different wireless channel environments.
Furthermore, suppose L2 = 2L1 . To address the channel estimation needs two differ-
ent RMMSE filters are designed, each corresponding to one mode of operation, i.e.,
one CP. If the robust design strategy as mentioned above is followed, the correlation
coefficients become,
rh,f 1 [∆k1 ] = f1 (L1 ∆k1 /N),
(3.16)
rh,f 2 [∆k2 ] = f1 (L2 ∆k2 /N),
4 Channel estimation in LTE and DVB-H 23
where f1 (L∆k/N) = . Thus, to force
2jπL∆k/N
rh,f 1 [∆k1 ] = rh,f 2 [∆k2 ] ⇒ f1 (L1 ∆k1 /N) = f1 (2L1 ∆k2 /N), (3.17)
the following relation between the pilot separations should hold,
∆k1 = 2∆k2 . (3.18)
As a result, the constructed PVs should have different virtual pilot spacing for
each case, i.e., the pilot spacing in the event of L2 should be twice the spacing in
the event of L1 . Thus, one may use the merging scheme proposed in this thesis or
use additional 1-D filtering, at the expense of more computational complexity. For
the above example, where the pilot pattern is similar to Fig. 7, the constructed PV
is fed into the designed MRMMSE filter when the system is configured to operate
in mode 1. However, to insure a pilot spacing twice as much, pilots from alternate
pilot-carrying OFDM symbols should be collected into the PV.
The application of the above strategy is not restricted to on-chip filter memory
optimization associated with one OFDM system, with different modes of operation,
only. As it will be discussed in the following sections, the same principle may be
extended such that a family of filters can be reused across different OFDM systems.
For instance, it will be shown how a novel selection of filter design parameters
provides a maximum memory optimization when LTE and DVB-H coexist on the
same platform. Finally, the examples will be used as the basis for a pilot design
strategy where the goal is to maximize the number of system configurations which
can employ the same filters for channel estimation purposes.
4 Channel estimation in LTE and DVB-H

4.1 LTE physical layer specifications
LTE has been proposed as the system of choice for high-speed wireless communica-
tions in the next generation of mobile terminals. The standardization was initiated
and pursued by the 3rd Generation Partnership Project (3GPP) [32] where groups of
telecommunications associations collaborate and contribute to the project develop-
ment. Being a complicated system with various building blocks, a detailed analysis
of LTE is well beyond the scope of this thesis. However, the following provides
a brief introduction to some of the physical layer characteristics and specifications
which are crucial from a channel estimation view point.
LTE is an OFDM system with a number of operation modes supporting different
BWs. The supported BW ranges from 2 MHz to 20 MHz and corresponds to DFT
sizes of 128 to 2048 points. One of the fundamental characteristics of LTE across its
different configurations is the tendency to keep a fixed subcarrier spacing of 15 kHz,3
3
The only exception is when LTE is used for multicast/broadcast over single frequency network
(MBSFN) where ∆f = 7.5 kHz. In this thesis, MBSFN associated with LTE has not been covered.
Figure 8: An example of time-frequency grid corresponding to one RBP for LTE.

The scattered pilots are identified by the black boxes.
i.e., ∆f = 15 kHz. As a result, the system’s sampling frequency scales according to

fs = N∆f samples/s. Besides, two different CPs are defined in LTE, called normal
and extended CPs. The normal CP, Lnormal , is 4.69 µs long while the extended CP,
Lextended , has a length of 16.67 µs. Depending on the operating environment, the
system is configured for one of the above GIs. For instance, in larger cells where
the delay spread of the channel is extensively longer, the extended CP can be used.
Please note that the actual OFDM symbol duration, without CP, is Tu = 66.7 µs.
LTE is an example of orthogonal frequency division multiple access where user
data in downlink is multiplexed for different users. As a result, data needs to be
encapsulated in specific units. There are various data encapsulations in LTE. For
example, an RB is the collection of 12 tones in 7 consecutive OFDM symbols if the
normal CP is employed. If the extended CP is used, the number of constituting
OFDM symbols reduces to 6 consecutive ones. Moreover, two consecutive RBs,
encompassing the same tones, constitute a RB pair (RBP). Fig. 8 illustrates an
example of RB and RBP. Although more coarse data encapsulations are also defined
in LTE, an RBP is the smallest block which may be allocated to a single user
during data transmission. Being properly encoded in the first few OFDM symbols
in each RBP, each terminal needs to actively receive and decode the control signaling
associated with each RBP. In the event of no data allocation, the terminal can enter
a stand-by state, to save energy, until a new RBP is scheduled for reception.
To assist channel estimation, a scattered pilot pattern has been designed and
proposed in LTE. Being a MIMO system, the pilots transmitted between a pair
of transmit/receive antennas have been designed to be orthogonal to the pilots
associated with other antenna pairs. As a result, MIMO channel estimation in LTE
reduces to several SISO estimations. For simplicity, the calculations are derived for
a single antenna port, i.e., antenna port 0. Fig. 8 shows the scattered pilot pattern
for antenna port 0 where pilot spacing in frequency is a constant value of 6 tones.
Moreover, the pilot frequency shift is 3 tones. In other words, compared to the
primary skewed pilot pattern in Fig. 6, α = 6, γ = 3 and β = 7. On the contrary,
the spacing between two consecutive pilot-carrying OFDM symbols associated with
a single RBP is not a constant value, i.e., Γ ∈ {3, 4}. The pilots in each RBP appear
in OFDM symbols 1, 5, 8 and 12 respectively.
Although the above described configuration is limited to a subset of antenna
ports, it provides an acceptable basis for the derivation of channel estimation algo-
rithms. Besides, the results can be easily extended to other antenna ports where
the pilot patterns slightly vary to accommodate pilot orthogonality across multiple
antennas. Meanwhile, in practical system implementations, certain OFDM symbols
are discriminated against others due to the sensitivity of the information they carry.
For instance, the first few OFDM symbols may undergo a different data modulation,
e.g., BPSK or QPSK, compared to the rest of symbols where up to 64 QAM may
be used for data modulation. However, for simplicity and without loss of general-
ity similar parameters across all OFDM symbols in an RBP have been used, when
applicable.
4.2 DVB-H physical layer specifications

Being a SISO OFDM system, DVB-H is an extension of digital video broadcast-
terrestrial (DVB-T) for transmission of multimedia content to wireless mobile ter-
minals. On top of various modifications to DVB-T, such as using different interleav-
ing schemes, the fundamental difference, at least from a physical layer perspective,
is the addition of 4K mode of operation in DVB-H. 4K mode provides a compro-
mise between data rate , hardware complexity, and overhead during implementation
when compared to the conventional 2K and 8K modes in DVB-T. Similar to LTE,
an elaborate description of DVB-H is not the intention of this thesis and the inter-
ested reader may refer to [5] for further information. However, a short overview of
DVB-H’s physical layer parameters fundamental to derivation of channel estimation
is provided as follows.
DVB-H supports three different BWs corresponding to each mode of operation,
i.e., 2k, 4k, and 8k. The size of each BW is simply controlled by scaling the sampling
frequency of the system. For a given BW, the sampling frequency remains constant
throughout various modes of operation. As a result, contrary to LTE, the subcarrier
spacing in DVB-H is not a constant value and varies inversely proportional to BW.
Moreover, four different GIs have been designed for DVB-H. The GIs are expressed
in terms of their length with respect to the useful OFDM symbol duration. As a
result, the GIs are known as 41 , 18 , 16
1 1
, and 32 .
Scattered pilots are also employed in DVB-H to address the channel estimation
Figure 9: An example of time-frequency grid corresponding to DVB-H. The

scattered pilots are identified by the black boxes.
needs. The pilots are evenly distributed in all OFDM symbols. Fig. 9 shows the
pilot pattern in an exemplary time-frequency grid for DVB-H. Moreover, mapping
the pilot scheme in DVB-H to the primary skewed pilot pattern in Fig. 6, α = 12,
β = 4, γ = 3 and Γ = 1 can be deduced. Meanwhile, to increase the channel
estimation performance at the expense of a reduced effective SNR, the pilots can be
transmitted with boosted power when compared to data symbols.4
4.3 Multi-mode filter-tap memory optimization in LTE

Following the discussion in Sec. 3 many different approaches can be pursued for
channel estimation in LTE. However, we will focus on the low-complexity robust
estimators, i.e., MRMMSE applied to antenna port 0. Doing so reduces MIMO
LTE channel estimation to a simpler SISO counterpart. The results can be easily
extended to MIMO LTE due to orthogonality of the pilots, though.
As already discussed, our general strategy is to alleviate the additional 1-D
filtering in time while achieving acceptable performance in moderately fast fading
environments. As a result, we have resorted to pilot-merging scheme to capture
the time variations of the wireless channel to a certain extent. Fig. 10 shows the
proposed methodology for pilot merging when an individual RBP has been selected
for channel estimation as well as the equalization. It can be seen from the figure that
we have relied not only on the pilot-carrying OFDM symbols within the specified
RBP but also on the 1st pilot-carrying OFDM symbol associated with the next
4
The boosted power option has not been employed for the deductions of the results in this
thesis.
Denotes one RB
Data Pilot symbol Collected Pilot
Time
Region 4
PV4
Region 3
PV3
Region 2
PV2
Region 1
PV1
k+12
k+11
k+10
k+9
k+8
k+7
k+6
k+5
k+4
k+3
k+2
k+1
Frequency
Figure 10: Pilot-merging strategy for channel estimation and equalization associ-
ated with LTE.
adjacent RBP. Although one may argue that there is no guarantee two consecutive
RBPs are allocated to the same user, the terminal needs to receive and decode the
first few symbols in each RBP to extract the scheduling information.
The constructed PV associated with channel estimation of each region is identi-
fied by brown tiles in Fig. 10. Considering PVs as such, the 1-D filter in frequency,
i.e., (3.12) needs to be designed. As discussed in Sec. 3.3, a robust MMSE filter de-
signed for the worst-case expected channel environment can be employed for channel
estimation purposes in LTE. Thus, the filters are designed for rectangular GIs with
delay spreads equal to the normal as well as the extended CPs. Besides, following
the discussion in Sec. 3.4, only a subset of pilots in the vicinity of each tone is
used for estimation purposes. Application of merged-pilot MRMMSE estimators
introduces a certain design mismatch compared to the full 2-D MMSE filter. The
amount of mismatch can be quantified and derived mathematically from (3.8) or
through Monte Carlo simulations.
Due to pilot merging, the mismatch in the event of high Doppler shifts becomes
more severe and dominates the PDP mismatch in fast-fading environments. As a
result, the merged-pilot MRMMSE channel estimation’s performance degrades for
high terminal speeds as a result of which higher SNR is needed to sustain the design
MSE. However, it can be shown in Sec. 5.2 that the BER performance degradation
due to Doppler is not as severe when various coding and interleaving are applied in
the practical system implementation. A thorough analysis of the performance for
various fading environments is discussed in Secs. 5.1, 5.2, and 5.3.
Designing the MRMMSE estimators for LTE system alone, one might need two
separate estimators, i.e., two estimators associated with the normal CP and extended
CP respectively. However, the memory optimization strategy discussed in Sec. 3.5
can be practiced to reduce the amount of on-chip memory. As a result, it is required
that
L1 ∆k1 /N1 = 2L1 ∆k2 /N2 (4.1)
for the filter coefficients associated with the different modes to be identical. Here,
index 1 and 2 refer to filter design in the event of normal and extended CPs, respec-
tively. However, it can be observed that ∆k 1
∆k2
= NL 2 N1
2 L1
≈ 3.5, when N1 = N2 , is not
an integer. A compromise in design CP has been proposed in Sec. 5.3, where the
design PDP, called L1,trunc. , is a truncation of normal GI such that
∆k1 L2 N1
= = 4, (4.2)
∆k2 N2 L1,trunc.
for N1 = N2 .
In other words, the robust filters in the event of normal CP may be designed
for a truncated GI such that (4.1) holds. Thus, to guarantee filter coefficient reuse,
the only possibility is to use pilot merging in the event of extended CP and ignore
alternate pilots in frequency for normal CP. By doing so the pilot spacing for normal
CP becomes 12 tones. However, the maximum distance for pilot spacing in frequency
is 14 tones if the Nyquist sampling criteria in frequency is to be met for LTE in
normal mode of operation. But Nyquist sampling criteria holds where theres is
an infinite (sufficiently large) signal observation interval in the absence of AWGN.
Thus, a pilot spacing of 12, just below the Nyquist threshold, results in degraded
channel estimation performance.
Had the pilot pattern for one RBP in LTE been designed according to Fig.
11, more flexibility in terms of filter memory optimization would be provided. In
fact, three different merging schemes could be achieved either through merging or
time-domain filtering while the pilot overhead would decrease by 25%. However,
due to the longer separation between pilots in time, almost twice as long compared
to the standardized pattern, the maximum supported mobility associated with the
system would scale by 1/2. Besides, the performance deterioration would be more
emphezised in the event of high mobility when pilot merging is practiced.
Alternatively, the robust estimator in the event of normal CP can be designed
for an extended normal GI.5 In other words, if the filter is designed for a rectangular
PDP with a delay spread of 8.3 µs rather than 4.7 µs,
∆k1 L2 N1
= = 2, (4.3)
∆k2 N2 L1,norm.,exten.
where L1,norm.,exten. is the design PDP for an extended normal CP.
Thus, the same filter coefficients may be shared between the two modes provided
that pilot merging in the event of extended CP is practiced while for operation in
5
Here the extended normal GI refers to the adopted design CP used in MRMMSE estimators
and should not be confused with the standardized extended CP in LTE, i.e., Lextended
Pilot tone Data tone
Time
Resource Block
Figure 11: An example of hypothetical pilot pattern for LTE to optimize the
filter-tap memory.
normal mode the pilot separation ought to remain 6, i.e., no pilot merging should
be sought.
Designing the filters for an extended normal GI is considered a safer approach
when compared to its truncated counterpart. Although, it introduces a considerable
amount of mismatch in connection with the majority of practical cases, it does not
discard the signal energy if the wireless channel constitutes high energy taps located
in the discarded region. Moreover, it allows reuse of the same filter coefficients
across the two modes of operation. On the contrary, if the estimator is viewed as
a filter, more noise passes through the filter designed for an extended normal CP,
when compared to the truncated normal CP, due to a wider filter BW.
4.4 Multi-mode filter-tap memory optimization in DVB-H

Due to a higher pilot density, merged-pilot MRMMSE channel estimation in DVB-H
can exhibit better performance in high terminal mobilities when compared to LTE.
Besides, more pilot merging patterns can be practiced. More precisely, 4 different
pilot merging schemes can be done with pilot separations of 3, 6, 9 and 12 tones.
Thus, for maximum filter memory optimization, one may use the same filter coeffi-
cients provided that pilot separations of 3, 6, 12, and 24 corresponding to 1/4, 1/8,
1/16, and 1/32 CPs, are forced in the constructed PVs. The drawback, however,
is a poor channel estimation performance in the event of 1/32 CP because a pilot
separation of 24 means ignoring alternate pilots in frequency.
4.5 Multi-standard filter-tap memory optimization across LTE

and DVB-H
Following the discussion in Secs. 4.3 and 4.4, it can be observed that one family of
filters may be used for channel estimation needs of both LTE and DVB-H provided
that LTE channel estimators are designed for an extended delay spread in the event
of normal CP. This is due to the fact that
Ldesign ∆k 3
= , (4.4)
N 4
for all possible configurations in LTE and DVB-H, provided that the parameters are
selected from Table 1. Besides, Ldesign corresponds to the presumed delay spread
of the wireless channel for which the estimators are designed. Table 1 shows the
different configurations for which channel estimation filter tap memory optimization
can be done for LTE and DVB-H.
Table 1: The configurations for which coefficient co-optimization is feasible according

to our proposed scheme.
LTE DVB-H
NDFT Operation Pilot BW Mode CP Pilot
mode Sep. (MHz) (k) len. Sep.
1
32
24
Normal CP 6 1
12
128-2048 (except 1536) 6-8 2-8 16
1
3 8
6
Extended CP 1
4
3
However, if the design criteria imposes the filters to be designed for a truncated
CP when LTE operates in the normal mode, then two separate filters need to be
designed and stored in memory to handle the channel estimation requirements when
the two standards coexist. Please refer to Sec. 5.3 for further discussion of the
technical details associated with such strategy.
5 Summary of included papers and technical re-

ports
5 Summary of included papers and technical reports 31
5.1 Paper I - Channel estimation for a mobile terminal in a

multi-standard environment (LTE and DVB-H)
Being two different OFDM systems, LTE and DVB-H have their own distinct speci-
fications and operation modes. For instance, DVB-H is a broadcast standard where
multimedia content is simultaneously broadcasted to a large number of users. As a
result, users have to receive and decode all the available data tones associated with
the transmitted OFDM symbols. However, in LTE, data can be encapsulated in
different data units bounded by a certain number of subcarriers in a given OFDM
symbol. Thus, the receiver does not necessarily need to estimate and decode all the
data associated with all the available data tones in a given OFDM symbol.
In this paper, two different approaches have been probed for channel estimation
when applied to LTE and DVB-H. One approach is the application of MRMMSE
algorithms as elaborated in this thesis. Contrary to the proposed MRMMSE algo-
rithm, time domain estimation through the application of DFT and IDFT has also
been probed. In other words, the performance of the above mentioned approaches
to LTE and DVB-H have been partly evaluated. Besides, the hardware complexity
in terms of the number of multiplications as well as memory requirements have been
evaluated.
Taking the fact that the terminal in LTE does not necessarily need to decoded all
the available data tones associated with an OFDM symbol into account, MRMMSE
best suits LTE in terms of hardware resources. On the other hand, DFT-based
algorithms exhibit much lower complexity when compared to the MRMMSE coun-
terpart. However, there is one big hurdle when using the DFT-based algorithms.
The majority of practical OFDM systems, including LTE and DVB-H, have spectral
masks through nulling the edge subcarriers. As a result, mixed-radix FFT and IFFT
algorithms need to be employed if DFT-based approach is to be applied.
Hardware implementation of mixed-radix FFT/IFFT algorithms seems to be far
more complicated when compared to radix 2 algorithms. On the other hand, ter-
minals usually have access to a radix-2 FFT for OFDM demodulation. As a result,
using mixed-radix FFT/IFFT blocks increases the power and hardware resource
requirements in a receiver. Thus, having separate MRMMSE and DFT-based esti-
mators may not be motivated in a multi-standard environment unless it is dictated
by the performance requirements. Meanwhile, MRMMSE algorithms provide a good
alternative with respect to DVB-H when both performance and hardware require-
ments are taken into account. As a result, the MRMMSE estimators can efficiently
address the channel estimation needs of the terminals both for LTE and DVB-H.
5.2 Paper II - Low-complexity channel estimation for LTE

in fast fading environments for implementation on multi-
standard platforms
The application of MRMMSE algorithms, when pilot merging is practiced, is an
efficient approach to low-complexity estimators in mobile terminals. The merits
of pilot merging in the event of fading environments are two-fold. First, nearly a
3-dB improvement in slowly-fading environments is achieved due to the fact that
the number of pilots for estimation purposes is doubled. Second, the pilot merging
partly captures the time-variations in fading environments. However, the introduced
errors due to mismatch in fast fading environments start to surpass the improvements
resulting from a larger number of available pilots for estimation. As a result, it is
necessary to characterize the fading environments where pilot merging would provide
an acceptable performance.
To reduce the complexity of the estimators, one needs to lower the number
of estimation operations. One way to achieve this is to widen the equalization
regions where the same estimated channel can be employed. However, in fast fading
environments the regions need to be divided into smaller subset for equalization
purposes. In this paper, it is shown that by using the available pilot pattern in
LTE, pilot merging may be practiced for different equalization regions. The merits
of such approach are observed in fast fading environments where a terminal speeds
up to 120 km/h are expected. Beyond that, one may need to employ additional
time-domain filtering to capture the time variations of the channel.
Meanwhile, it is shown that an additional simple filter in the direction of time
can be employed in the event of fast fading environments where pilot merging cannot
be practiced. The time-domain filter is designed for a worst-case Doppler spectrum,
i.e., a rectangular Doppler spectrum where the Doppler spread is at its maximum
value. The simulation results, show that the estimator exhibits a good performance
when the above mentioned additional filter kicks in. Although, the adopted standard
for investigation in this paper is LTE only, the proposed approaches can be easily
extended to DVB-H.
5 Summary of included papers and technical reports 33
5.3 Paper III - On coefficient memory co-optimization for

channel estimation in a multi-standard environment (LTE
and DVB-H)
In Sec. 5.1, 5.2, and 5.3 the merits of MRMMSE channel estimation algorithms
were discussed when applied to LTE and DVB-H. Meanwhile, the main emphasis
has been on reducing the computational complexity of such algorithms. Although,
computational complexity accounts for a large portion of the required hardware re-
sources, including power consumption, little attention was given to on-chip memory
for the storage of the channel estimation filter taps.
MRMMSE estimators alleviates the requirement for real-time calculation of filter
taps so they can be calculated off-line. However, the filter taps need to be stored
in memory such that the estimation block can efficiently fetch them during the
estimation operation. On the other hand, the number of filter taps increases as
more standards need to be integrated in the same platform. Even different modes
of operation associated with one standard may require additional filter memory.
In this paper, it is shown that by exploiting the similarities in pilot pattern in
LTE and DVB-H, one could design the MRMMSE estimators such that similar filter
taps could be employed for estimation purposes across LTE and DVB-H. Moreover,
the robust estimators may need to be designed for delay spreads other than what
is suggested by the standardized CPs so that similar filters can be shared among
the standards. As a result, the performance of the filters in such cases has been
evaluated in this paper. Specifically, the performance of the estimators in the event
of designing the estimators for a truncated normal CP in LTE has been discussed.
On the other hand, the pilot separation plays an important role if similar filter
taps need to be used. As a result, one may need to exercise pilot merging or addi-
tional time filtering to assure the compatibility among filter taps. Using different
pilot merging for different modes of operation in each standard might impose cer-
tain performance degradation in fading environments. This is more emphasized in
DVB-H where a number of different pilot merging schemes need to be practiced.
Thus, the channel estimator’s performance for various merging schemes has been
characterized with respect to different fading environments.
5.4 Technical Report - Channel estimation filter-tap memory-

optimized pilot pattern design for OFDM systems in multi-
mode and multi-standard environments
The possibility to optimize the on-chip memory may provide a 6-fold optimization
for the storage of channel estimation filter taps in a multi-standard environment
where LTE and DVB-H co-exist. The above mentioned optimization could not be
achieved if the divisor of the pilot separations in LTE and DVB-H was not the same
prime number, i.e., 3. Thus, one might ask if there is a possibility to integrate
more OFDM standards in the same platform while benefiting from similar channel
estimation filter taps. In a broader perspective, is it possible to maximize the number
of OFDM standards which can be integrated in the same platform, constrained to
the availability of a finite set of channel estimation filters, when designing the pilot
pattern.
In this paper, it is shown that by following a specific strategy in scattered pilot
pattern design for OFDM systems, the number of required filter taps upon integra-
tion of several systems in a multi-standard environment is minimized. The designed
pattern would still provide the possibility for channel estimation in the event of
applying linear estimators other than the proposed MRMMSE estimators. A pilot
pattern as such provides a maximum number of different frequency pilot spacings
when pilot merging or time filtering is practiced. As a result, the proposed approach
provides the flexibility for employing the MRMMSE channel estimators. Further-
more, it guarantees the proximity of the pilot carrying OFDM symbols such that
upon merging the shortest distance exists between the pilot contributing OFDM
symbols. This is especially important for equalization purposes in fast fading envi-
ronments.
On the other hand, there are situations where one OFDM system is already
standardized. So, how are the pilot patterns for new OFDM systems designed such
that they remain compatible with an existing system in a multi-standard environ-
ment? A strategy has also been suggested in this paper which addresses compatible
pilot pattern design for emerging OFDM systems with respect to an existing sys-
tem. Meanwhile, LTE is expected to remain a prominent OFDM standard for voice
and data communications in the foreseeable future. Thus, the report is wrapped
up through an example which elaborates the application of the above strategy to
pilot pattern design for the emerging OFDM systems which remain compatible with
LTE.
6 Concluding remarks 35
6 Concluding remarks
Channel estimation and pilot pattern design for OFDM systems is a well-established
area of research to which many researchers have contributed. However, as mobile
terminals tend to integrate more high speed OFDM standards there is still a lot
to do in terms of optimized signal processing algorithm design for such systems.
Although the optimization might be performed at the expense of some performance
degradation, a considerable reduction in the amount of hardware resources could
motivate their application in low-power terminals. In this thesis, it is shown how the
above principle can be applied to the channel estimation block in a multi-standard
receiver. By adopting some of the already existing standards such as LTE and
DVB-H, approaches to reduced complexity estimation algorithms are probed and
solutions are suggested.
Furthermore, the lessons learnt from such an approach provided the opportunity
to pursue a strategy for pilot pattern design for OFDM systems in multi-standard
environments. The above mentioned strategy not only maximizes the performance of
the receivers in the presence of merged-pilot MRMMSE channel estimators, but also
enables sharing the same filters among a maximum number of compatible OFDM
systems. Moreover, it is possible to design the pilot pattern in emerging OFDM
systems such that they are compatible, in terms of employing similar channel esti-
mation filters, with existing OFDM systems in a multi-standard platform.
To further explore the capabilities of such pilot pattern design, one could evaluate
the performance of the proposed estimators in the presence of correlation mismatch.
In other words, how much correlation deviation can be tolerated if the proposed
optimization schemes does not guarantee the similarity among channel estimation
filter taps. Last but not least, one may explore the potentials of the proposed
digital signal processing optimization techniques to other baseband digital signal
processing blocks. For instance, the possibilities for low-complexity digital signal
processing algorithms for estimation of I/Q imbalance, probably combined with
channel estimation, may be probed in multi-standard receivers.
7 References
References
[1] “IEEE Standard for Information technology–Telecommunications and informa-
tion exchange between systems–Local and metropolitan area networks–Specific
requirements Part 11: Wireless LAN Medium Access Control (MAC) and
Physical Layer (PHY) Specifications Amendment 5: Enhancements for Higher
Throughput,” IEEE Std 802.11n-2009, pp. c1–502, Sep. 2009.
[2] “Evolved Universal Terrestrial Radio Access (E-UTRA); LTE physical layer;
General description,” 3rd Generation Partnership Project, Tech. Specifications,
3GPP TS 36.201, Dec. 2009. Available: http://www.3gpp.org/ftp/Specs/html-
info/36201.htm.
[3] M. Gertsman and J. Lodge, “Symbol-by-symbol MAP demodulation of CPM

and PSK signals on Rayleigh flat-fading channels,” IEEE Trans. Commun.,
vol. 45, pp. 788–799, July 1997.
[4] C. Knievel et al., “2-D graph-based soft channel estimation for MIMO-OFDM,”
in Proc. IEEE Int. Conf. Commun., Cape Town, pp. 1–5, 2010.
[5] “Digital Video Broadcasting (DVB); Framing structure, channel coding and
modulation for digital terrestrial television,” European Telecommunications
Standards Institute (ETSI), Tech. Specifications Final Draft, 2004.
[6] J. McKown and R. Hamilton, “Ray tracing as a design tool for radio networks,”
IEEE Trans. Netw., vol. 5, pp. 27–30, Nov. 1991.
[7] M. Sandell and O. Edfors, “A comparative study of pilot-based channel estima-

tors for wireless OFDM,” Tech. Rep., Div. of Signal Process., Luleå Univ. of
Technology, Luleå, Sweden, 1996.
[8] O. Edfors et al., “OFDM channel estimation by singular value decomposition,”

IEEE Trans. Commun., vol. 46, pp. 931–939, July 1998.
[9] M. Gertsman and J. Lodge, “A statistical theory of mobile radio reception,”

Bell Syst. Tech. J., pp. 957–1000, July 1968.
[10] M. Gans, “A power-spectral theory of propagation in the mobile-radio environ-

ment,” IEEE Trans. Veh. Technol., vol. 21, pp. 27–38, Feb. 1972.
[11] A. Goldsmith, Wireless communications. New York, NY: Cambridge University

Press, 2005.
[12] R. W. Chang, “Synthesis of band-limited orthogonal signals for multichannel

data transmission,” Bell Syst. Tech. J., 45, pp. 1775–1796, 1996.
7 References 37
[13] R. Chang and R. Gibby, “A theoretical study of performance of an orthogo-

nal multiplexing data transmission scheme,” IEEE Trans. Commun., vol. 16,
pp. 529–540, Aug. 1968.
[14] Y. Li et al., “MIMO-OFDM for wireless communications: signal detection with

enhanced channel estimation,” IEEE Trans. Commun., vol. 50, pp. 1471–1477,
Sep. 2002.
[15] M. Engels, Wireless OFDM systems: How to make them work? Norwell, MA:
Kluwer Academic Publishers, 2002.
[16] G. Stuber et al., “Broadband MIMO-OFDM wireless communications,” Proc.

IEEE, vol. 92, pp. 271–294, Feb. 2004.
[17] M. Windisch, Estimation and Compensation of I/Q imbalance in Broadband

Communications Receivers. Dresden, Germany: Jorg Vogt Verlag, 2007.
[18] G. Strang, LINEAR ALGEBRA AND ITS APPLICATIONS. Belmont, CA:

Thomson Brooks/Coles, 2006.
[19] Y. Li et al., “Channel estimation for OFDM systems with transmitter diversity
in mobile wireless channels,” IEEE J. Sel. Areas Commun., vol. 17, pp. 461–471,
Mar. 1999.
[20] M. Ozdemir and H. Arslan, “Channel estimation for wireless OFDM systems,”
Commun. Surveys Tuts., IEEE, vol. 9, pp. 18–48, July 2007.
[21] M. Morelli and U. Mengali, “A comparison of pilot-aided channel estimation

methods for OFDM systems,” IEEE Trans. Signal Process., vol. 49, pp. 3065–
3073, Dec. 2001.
[22] V. Srivastava et al., “Robust MMSE channel estimation in OFDM systems with
practical timing synchronization,” in Proc. IEEE Wireless Commun. and Netw.
Conf., Atlanta, GA, pp. 711–716, 2004.
[23] W. Tianbin et al., “Graph-based soft channel and data estimation for MIMO
systems with asymmetric LDPC Codes,” in Proc. IEEE Int. Conf. Commun.,
Beijing, pp. 620–624, 2008.
[24] L. Tong et al., “Blind identification and equalization of multipath channels,” in

Proc. IEEE Int. Conf. Commun., Chicago, IL, pp. 1513–1517, 1992.
[25] L. Baccala and S. Roy, “A new blind time-domain channel identification method
based on cyclostationarity,” IEEE Signal Process. Lett., vol. 1, pp. 89–91, June
1994.
[26] H. Zeng and L. Tong, “Blind channel estimation using the second-order statis-
tics: Algorithms,” IEEE Trans. Signal Process., vol. 45, pp. 1919–1930, Aug.
1997.
[27] W. Songping and Y. Bar-Ness, “OFDM channel estimation in the presence of

frequency offset and phase noise,” in Proc. IEEE Int. Conf. Commun., Anchor-
age, AK, pp. 3366–3370, 2003.
[28] O. Edfors, Low-complexity algorithms in digital receivers. PhD thesis, Div. of

Signal Process., Luleå Univ. of Technology, Luleå, Sweden, 1996.
[29] P. Hoeher et al., “Two-dimensional pilot-symbol-aided channel estimation by

Wiener filtering,” in Proc. IEEE Int. Conf. Acous., Speech and Signal Process.,
Munich, pp. 1845–1848, 1997.
[30] R. Nilsson et al., “An analysis of two-dimensional pilot-symbol assisted mod-

ulation for OFDM,” in Proc. IEEE Int. Conf. Personal Wireless Commun.,
Mumbai, pp. 71–74, 1997.
[31] Ericsson AB et al., “System Requirements Specification (2nd version), incl.

scope and time plan,” Scalable Multi-tasking Baseband for Mobile Communica-
tions, Tech. Rep. D1.2, Jan. 2008. Available: http://www.multibase-project.eu.
[32] “Third generation partnership project (3GPP).”

Available: http://www.3gpp.org.
Channel estimation for a
mobile terminal in a
multi-standard environment
(LTE and DVB-H)
Farzad Foroughi Abari, Johan Löfgren, and Ove Edfors
Paper I
c 2009 IEEE. Reprinted, with permission, from Farzad Foroughi Abari, Johan Löf-
gren, and Ove Edfors, “Channel estimation for a mobile terminal in a multi-standard
environment (LTE and DVB-H),” in Proc. IEEE Int. Conf. Signal Process. and
Commun. Syst., Omaha, NE, pp. 1–9, Sep. 2009.
1 Introduction 41
Abstract
In this paper, we present an analysis of different channel estimator struc-
tures that can be used and efficiently implemented with minor configuration
changes of a common hardware for both LTE and DVB-H. Such common es-
timator structures for LTE and DVB-H allow algorithm and hardware reuse
when both standards are implemented on the same platform. In this approach,
a core estimator will be utilized to address parameter estimation for the two
standards. The estimators exploit similarities in pilot patterns between LTE
and DVB-H. The estimation techniques are discussed taking both algorithm
and complexity issues into account.
1 Introduction
Orthogonal Frequency Division Multiplexing (OFDM) is a special case of multi-
carrier transmission and it can accommodate high data rate requirement of mul-
timedia based wireless systems [1]. Due to its inherent structure, it is feasible to
achieve a multiple-access format while benefiting from various coding and frequency
hopping schemes. This will provide a high degree of diversity and inter-cellular inter-
ference suppression, etc. Another way of increasing the wireless system throughput
is the employment of multiple antennas both on transmitter and receiver sides. Since
parallel channels are established over the same time and frequency grid, high data
rates without the need of extra bandwidth are achieved [1].This is known as Multiple
Input Multiple Output (MIMO) in literature and has been the integral part of some
recent state of the art wireless systems such as IEEE 802.11n and LTE.
In OFDM, the bandwidth is divided into a number of orthogonal narrow band
subcarriers. The orthogonality of subcarriers is essential to OFDM which imposes
some constraints on their spacing. The spacing should be smaller than the coherence
bandwidth of the communications channel. An OFDM system is also prone to time
variations of the channel. The system, however, is simple to implement and enjoys
application of low complexity equalization techniques. Although equalization is
rather simple especially when compared to non-OFDM communications systems,
appropriate channel parameter estimation is still required to obtain the Channel
State Information (CSI). Accurate CSI is crucial for resource allocation, adaptive
modulation and coherent detection [2].
LTE is one of the evolved standards which promises a high consistent data rate
over the wireless channels. It has already been adapted as the next generation
standard for mobile communications. Similarly, DVB-H addresses the need for high
multimedia data rates over the wireless communications channels. Operating in a
multi-standard environment mobile terminals are soon expected to be able to receive
and decode data streams related to these two standards. Thus, it is desirable for
them to accommodate the essential hardware/software to handle the requirements
of the above mentioned standards. It is also envisaged that a typical terminal will
have to decode and process multiple parallel streams corresponding to concurrent
handling of LTE and DVB-H data at the same time. As a result, an increased
amount of processing power needs to be provided while keeping the power efficiency
Paper I: Channel estimation for a mobile terminal in a multi-standard environment
42 (LTE and DVB-H). . .
Figure 1: LTE, time-frequnecy grid. Each box corresponds to one Resource

Element(RE). A typical Resource Block is composed of 7 OFDM symbols in time
and 12 subcarriers.
as much as possible. Since the processing power cannot be surged indefinitely, novel
low complexity algorithms need to be investigated in connection with the hardware
design. We have probed the channel estimation issues in the scope of this paper.
More elaborately, new low complexity methods have been suggested which take both
LTE and DVB-H standard requirements into consideration. Although the topic of
channel estimation is both well-known and well-investigated to the communications
community, the authors have not found many works which address designing flexible
estimators that can be tailored to multi-standard processing specifically to support
coexisting of both LTE and DVB-H.
To assist channel estimation in DVB-H and LTE, scattered pilots have been
allocated throughout the time-frequency grid as defined in the following sections.
The pilot pattern is specific to each standard and largely depends on the typical
environment the terminal experiences in practice. Although there are differences in
the way pilots are scattered for LTE or DVB-H, there are many similarities among
them which can be exploited to develop a core estimator. Making good use of the
similarities which arise from the inherent nature of the standards, common channel
estimators can be developed. The proposed methods in this paper will rearrange
the pilots in LTE and DVB-H in such a way they will look similar in structure. As
a result, a number of flexible core estimators can be developed which can be equally
applied to both standards with minor modifications in some parameters such as FFT
2 Standards 43
Figure 2: DVB-H time-frequency structure.
size, number of pilots, etc.
2 Standards
The two different standards mentioned in the introduction have individually been
designed based on the specific requirements of the environment they are to oper-
ate. Even though the standards are both based on OFDM, they embody major
differences that need to be independently analyzed. In this section similarities and
differences among the standards are highlighted from a channel estimation perspec-
tive. Important parameters of interest in this paper are the pilot patterns as well as
resource allocations.
2.1 LTE (Long Term Evolution)

Mobile broadband encouraged the specification of LTE which became the basis for
the next generation of UMTS Mobile standards [3]. The standard is intended for
high data rates of at least 100Mbps in the downlink and 50Mbps in the uplink.
It also constitutes flexible carrier bandwidth allocations from below 5MHz up to
20MHz [4] .
The entire downlink chain has been designed to decrease the receiver complex-
ity. Several modes of operation have been designed to adapt to various channel
environments. The downlink chain in LTE has the following properties which are of
considerable importance when it comes to channel estimation.
2.1.1 Downlink Frame Structure

A time-frequency representation of the available spectrum is called a resource grid
[5], where the minimum unit is composed of one subcarrier in one symbol, named
Resource Element (RE). Another important definition is physical Resource Block

denoted as RB for simplicity. An RB is a measurement in time and frequency
corresponding to 6 or 7 OFDM symbols and 12 subcarriers, refer to Fig. 1.
2.1.2 Resource Allocations

LTE has a channel dependent scheduling [4]. This technique is specially tailored to
the needs of low speed terminals. The specific resources are allocated to a terminal
in relation to a channel condition measurement reported to the base station. There
are three different formats for downlink scheduling assignments but only one of them
supports frequency-contiguous allocation [4]. Generally, each terminal is guaranteed
at least two consecutive RBs once the access to downlink data is granted. Fig.
1 shows the structure of two RBs and their relative location in the defined time
frequency grid.
2.1.3 Pilot Pattern

LTE enjoys application of scattered pilots in the whole time-frequency grid. Being
a MIMO system, LTE standardization has specified independent pilot patterns for
each antenna port to facilitate parameter estimation. The LTE radio-access spec-
ifications refers to antenna ports rather than antennas to emphasize that what is
referred to does not necessarily correspond to a single physical antenna [4]. Fig. 1
shows the pilot pattern for antenna port0 [4]. Since scattered pilots do not cover the
whole time-frequency grid interpolation/extrapolation techniques should be prac-
ticed to derive a good channel estimate. The pilot patterns have been designed to
address the need for a maximum attainable spectrum efficiency while satisfying the
basic requirements for deriving a consistent channel estimate.
2.2 DVB-H (Digital Video Broadcasting-Handheld)

The Digital Video Broadcasting project (DVB) is an international industry-led con-
sortium committed to the development of technical standards related to Multimedia
broadcast. Being an evolution of DVB-T, DVB-H was specifically developed to ad-
dress the needs of a low power mobile terminal. Some of the characteristics of
DVB-H are as follows.
2.2.1 Frame Structure

DVB-H may operate in different modes, e.g., 2K, 4K, 8K. Irrespective of the mode
of operation, one frame in DVB-H is defined to constitute 68 OFDM symbols [6].
DVB-H is in fact a broadcast standard and in contrast to LTE, all subcarriers in a
symbol carry the information for all User Equipments (UEs). As stated earlier, in
LTE data is individually allocated to UEs. Besides, each UE is guaranteed as few
as 12 subcarriers in a given symbol.
Figure 3: The proposed system model. Zero padding is performed on both sides
of input data vector on the subcarriers corresponding to the guard bands.
2.2.2 Pilot Pattern

The pilots are intended to be used for time synchronization, frame synchronization,
channel estimation, transmission mode identification and phase noise tracking. Both
scattered as well as continuous pilots are transmitted in each symbol. While con-
tinuous pilots are primarily used for synchronization, scattered pilots are alloted for
channel estimation purposes. Scattered and continual pilot structure are shown in
Fig. 2
3 System model
Fig.3 demonstrates a simple baseband OFDM system model. This general descrip-
tion of system fits any of the two standards being investigated in this paper without
loss of generality. The uncoded data is modulated through one of the standardized
modulation schemes (e.g., 4QAM) and placed on Nc parallel streams where Nc de-
notes the number of used subcarriers (excluding guard band). The combination of
data and pilot subcarriers is fed into an IFFT block to finalize the OFDM modu-
lation. The modulated data is passed through a fading channel with AWGN and
OFDM demodulated through a FFT block at the other end. Demodulated data is
further processed to decode the transmitted bit stream. To preserve the orthog-
onality of the subcarriers a Cyclic Prefix (not shown in this system model) with
a length equal to the maximum expected delay spread is attached to the OFDM
modulated data on the transmitter and discarded before the demodulator on the
receiver. We assume perfect synchronization at the receiving side. The data is fed
to a channel estimator where data subcarriers indicated by white boxes in Fig.1 and
2 need to be estimated using the available scattered pilots.
Depending on the designated pilot pattern there are many approaches towards chan-
nel estimation for OFDM-based systems. Taking the fact that scattered pilots are
Figure 4: LTE, time-frequency grid. Mirrored pilots are collected from other
symbols to create the pilot vectors used in the estimators. Pilot vector 1 is used
for estimation over all subcarriers in RB1 while pilot vector 2 is used for estimation
over all subcarriers in RB2. The arrows indicate how the pilots from future symbols
are reflected into the pilot vectors.
standardized for the channel estimation in LTE as well as DVB-H into account, a
number of estimation/interpolation techniques can be applied for channel estima-
tion purposes in these systems. As already discussed in the previous section, Fig.1
and 2 illustrate how the pilots are placed in the time-frequency grid for these two
standards. Probably the best as well as the most complicated linear approach is
the application of a 2-D Wiener filtering [7]. Yet, the mentioned estimator can be
broken into two 1-D filters [7] to reduce the complexity. The 1-D filters correspond
to separate MMSE estimations one in time and the other in frequency directions.
The estimation in time is a function of channel variations in time. The variations
of channel taps over time can be described by the Doppler spread, Ds . Meanwhile
the Doppler spread shapes the spectrum [8] and affects the correlation properties of
the channel taps over time. Thus, the coherence time (Tc ) of a wireless channel can
be defined as the interval over which the channel taps don’t change significantly [9]
given as,
1
Tc = (4.1)
4Ds
where Ds is the Doppler spread in the channel. This definition is a somewhat
imprecise definition since the significant changes might belong to those taps which
have lower energy [9] or vice versa. It is expected that the correlation between two
realizations of a typical tap almost disappears beyond the coherence time.
Considering the restrictions wrapped around the coherence time, only the pilots
Figure 5: DVB-H time-frequency structure. Mirrored pilots are collected from the
past three symbols to create the pilot vectors used in the estimators. Contrary to
LTE, each pilot vector in DVB-H is used to estimate only one OFDM symbol (the
current symbol).
in the vicinity of the estimated channel attenuations should be used. The time
window in which the above mentioned method works shrinks as the Doppler spread
increases. The Doppler spread can be described by the equivalent terminal speed
relative to the transmitter (base-station). As the equivalent speed increases, the
coherence time decreases which results in lower correlation among subsequent chan-
nel realizations. There are a number of proposed approaches to exploit the time
correlation of various channel realizations. If the second-order channel statistics are
known, MMSE methods can be used to estimate/interpolate the pilots in time. For
instance, a robust MMSE estimator similar to the one applied for estimation in
frequency (as described in the following sections) can be used in time as well.
In our proposed approach, however, we have resorted to a simple method which
does not involve any estimations/interpolations in time. In fact, pilots from a num-
ber of neighboring symbols are collected and used to improve the channel estimation.
This has been done presuming stationary channel environment and as a result works
well in scenarios with low Doppler spreads. We have used the term mirrored pilots
referring to the ones collected from the neighboring symbols and inserted to the
pilot vectors used in the estimators. The mentioned pilot vectors as depicted in Fig.
4 and Fig. 5 also contain the existing pilots in each symbol.
To carry out the estimation in LTE the pilots from symbols 1 and 5 in each
RB are combined into a single pilot vector as depicted in Fig. 4. This pilot vector
is subsequently fed into the proposed channel estimators. The derived estimation
will be used for the whole OFDM symbols corresponding to that RB. Considering
Figure 6: Subcarrier estimation in LTE using only 8 pilots. Pilots from future
symbols are mirrored to the current symbol to assist the channel estimation. The
UE has access to all the pilots in each symbol irrespective of its allocated RBs.
MRMMSE with 8 pilots can be used to carry out the estimation.
the definition of the coherence time as explained previously, it is expected that this
method works for slow fading channels or in other words in channels characterized
with low Doppler. Furthermore, the farther the pilots are located from a target
symbol for estimation the lowest the cross-correlation they exhibit. As a result, the
largest Mean Squared Error (MSE) is expected for OFDM symbol 7 in each RB,
as it is furthest away from the pilots used in the estimation. Being the worst-case
scenario, the derived Mean Squared Error as well as uncoded BER have been derived
for symbol 7 as demonstrated in the simulations section.
The proposed approach is characterized with some drawbacks. For instance, data
decoding for each symbol cannot initiate until the 5th symbol in each RB has been
received and its pilots collected (mirrored). Thus, decoding has to be postponed
until the pilot-assisted channel estimation is carried out using the pilots collected
from symbols 1 and 5 in each RB. The associated memory overhead is expected to
be acceptable specially for the noncontiguous resource allocation in LTE.
Similarly, to estimate a symbol in DVB-H, the pilots from other symbols are
mirrored to facilitate the channel estimation. In other words, the pilots from the
past three OFDM symbols are saved in memory and consequently mirrored to the
current symbol for the estimation purposes. Thus, the largest distance between the
mirrored pilots and the current symbol is three OFDM symbols. Fig. 5 shows how
the symbol identified by the dashed line is estimated through collection of pilots from
the current as well as the past three symbols. It can be seen from the figure that the
Figure 7: The DFT-based estimator structure for the proposed estimators DFT-A,
DFT-B, DFT-C. The elements in a will be all ones in case of DFT-A.
resulting pilot spacing in frequency will be three subcarriers after the appropriate
pilots have been mirrored. This is quite similar to the witnessed pilot pattern in
the attained pilot vector in LTE. Refer to Fig. 4 and Fig. 5 to see how the derived
pilot vectors compare. In comparison to LTE, it is expected that application of
this method to DVB-H does not affect the quality of the estimators as adversely
in high Doppler scenarios. Besides, decoding of received symbols can immediately
start (provided that the UE has already collected the pilots from the past three
symbols) and the terminal does not need to wait for the additional pilots from the
future symbols.
Having collected the desired pilots in time, data subcarriers need to be estimated
in the frequency direction. Among the various investigated methods ML, MMSE
[10] and DFT-based [11] estimators are examples of estimation algorithms. Our
approach in this paper is the application of modified variants of MMSE and DFT-
based estimators. There are benefits to simplified estimators for a low power mobile
terminal due to the restricted amount of available processing power. If the estimators
can benefit from the commonly available hardware on the platform, like FFT blocks
or baseband processors, allocation of extra resources dedicated to channel estimation
becomes minimal.
4.1 MMSE estimator

The best linear estimator in terms of Mean Squared Error is the Linear Minimum
Mean Squared Error Estimator (LMMSE) [7]. Given the transmitted symbols X
and recieved symbols Y, under certain conditions the simplified MMSE estimation
of the channel h can be expressed as [12],
β
ĥmmse = Rhh (Rhh + I)−1 ĥls (4.2)
SNR
where
ĥls = X−1 Y (4.3)
is the Least Squares estimate of the channel attenuations h , Rhh is the autocorrela-
tion of the channel, SNR = E|xk |2 /σ 2 is the signal-to-noise ratio and β is a constant
0
10
DFT−A
DFT−B
DFT−C
MRMMSE,16 pilots
RMMSE
−1
10 Full MMSE
Mean Square Error

−2
10
−3
10
−4
10
−3 0 5 10 15 20 25
SNR, [dB]
Figure 8: LTE, MSE performance in a stationary environment.
which depends on the signal constellation of transmitted symbols, X. For a 4QAM

modulation scheme β equals 1.
In systems where scattered pilots are used to assist channel estimation the above
equation may be rewritten as,
β
ĥmmse = Rhm hp (Rhp hp + I)−1 ĥp,ls (4.4)
SNR
where Xp and Yp are the transmitted and received data sampled on the pilot posi-
tions and,
ĥp,ls = X−1
p Yp (4.5)
is the Least Squares estimate of the channel attenuations hp on the pilot tones,
Rhp hp is the autocorrelation of the sampled channel hp on pilot positions and Rhm hp
is the cross correlation matrix between the pilot tones and the data subcarriers.
The above equation system shows that for a channel containing Np scattered
pilots we need to have access to the second order channel statistics as well as the
noise variance at any given time. Moreover, a matrix inversion of Np × Np should be
carried out for any given SNR and a matrix calculation of Nm ×Np is required to the
interpolation over Nm data subcarriers. Thus, this estimator exhibits considerable
complexity [12].
4.2 Robust MMSE (RMMSE) estimator

As its name implies, RMMSE is designed to provide robustness of estimation for
various channel realizations. It should provide an acceptable performance irrespec-
tive of the environment in which the UE is about to operate. For this purpose, the
estimator could be designed for a worst-case scenario and its performance compared
to the full MMSE. It can be shown [13] that this estimator preforms well under
certain conditions. In this paper, we have chosen a uniform Power Delay Profile
0
10
Perfect Chan. Knowledge
RMMSE
MRMMSE,16 pilots
DFT−C
−1
BER 10
−2
10
−3
10
−4
10
−3 0 5 10 15 20 25
Eb/N0, [dB]
Figure 9: LTE, uncoded BER performance for 4QAM in a stationary environ-

ment.
(PDP) where all the channel taps are equally strong. The length of the channel has
been set to the length of the Cyclic Prefix. Consequently, the required correlation
matrices , Rhm hp and Rhp hp can be precalculated. Meanwhile, it has been shown [12]
that the performance loss is relatively acceptable for the lower SNRs if the estimator
is designed for the maximum SNR the system is expected to experience in practice.
Our simulation results also prove this point. Having fixed SNR in equation 4.4 the
required matrices can be precalculated and stored in memory. Since large matrix
inversions in 4.4 can be avoided the terminal can save a lot of processing as well as
battery power. In fact, the RMMSE estimator reduces the complexity to a mere
matrix multiplication of size Nm × Np .
4.3 Modified Robust MMSE (MRMMSE) estimator

Channel estimation using a RMMSE estimator makes use of all the pilot tones
allocated to an OFDM symbol and is still a rather costly approach when it comes
to hardware implementations. The implementation cost becomes even more crucial
when the number of pilots grows like in 8K DVB-H mode of operation where each
symbol constitutes at least 568 pilots. As a result, we have investigated alternative
approaches to the RMMSE estimator. To estimate each subcarrier only a few pilots
in the vicinity of that subcarrier might be used. Fig. 6 illustrates how, e.g., 8
pilots in the vicinity of two subcarriers k+10 and k+11, can be used and fed into
the MRMMSE to carry out the channel estimation. In other words, exploiting the
correlation for the full extent of the frequency axis might be of little use specially
when the estimator is designed for a worst-case PDP. The autocorrelation of a
system having a uniform PDP illustrates low correlation among subcarriers which
are located far apart. We have exploited this feature to develop an estimator which
uses only 16 neighboring pilots for estimation of each tone. As a result, in the
proposed MRMMSE, Rhp hp is a matrix of size 16 × 16. The pilots are preferably
0
10
MRMMSE,16 pilots (30 km/h)
DFT−C (30 km/h)
DFT−C (45 km/h)
−1
10
BER
−2
10
−3
10
−4
10
−3 0 5 10 15 20 25
Eb/N0, [dB]
Figure 10: LTE, uncoded BER performance for 4QAM corresponding to two
different terminal speeds. Simulations have been done for OFDM symbol 7 in a
typical RB which represents the worst-case scenario due to its low correlation with
the constructed pilot vector.
selected symmetrical with respect to the estimated tone. As stated earlier, not only
the pilots in the present symbol are used for estimation but also the pilots from other
symbols as depicted in Fig. 4 and Fig. 5 are mirrored to improve the performance.
This novel rearrangement of pilots proves to be performing well for low and medium
terminal speeds. This has been proved through simulations as illustrated in the
following sections. 1
4.4 DFT-based estimators

In OFDM systems where scattered pilots are used for channel estimation IDFT/DFT
interpolation is one of the alternative approaches towards channel estimation. In
these estimators the interpolation over all used subcarriers is obtained by [14],
1
ĥIDFT = FL FH
P ĥp,ls (4.6)
Np
where Np is the number of pilots FL is a N × L matrix constructed by taking the

first L columns of the full N × N Fourier transform matrix and FP is a subset of
the Fourier transform matrix on the pilot positions where N is the FFT size of the
system.
Since the Least Squares approach is applied to obtain noisy estimations of the
pilot subcarriers denoted as ĥp,ls , the interpolation over all subcarriers will produce
1
In this paper we have dropped the merged-pilot for MRMMSE and RMMSE estimators simply
because all the simulations have been carried out while merging the pilots. Thus, RMMSE and
MRMMSE estimators mentioned in this paper correspond to merged-pilot RMMSE and merged-
pilot MRMMSE estimators respectively.
noisy estimates of the channel subbands. This will result in a relatively high noise
floor which will end up in high Bit Error Rate after equalization. The associated
error covariance matrix can be derived as [14],
1
ĈHIDFT ≈ σH 2 FL FL H (4.7)
Np p
where σHp 2 is the variance of the AWGN on the pilot subcarriers. The above equation
indicates that the error covariance is a function of noise variance which will be
reflected into the estimation results if no measures are taken to reduce its effect.
Referred to as DFT-A estimator this method has been used in the context of channel
estimation in this paper.
Table 1: LTE,simulated system parameters

FFT size 512 δf, [kHz] 15
CP length,[n] 36 Used tones 300
Pilots/Symbol 50 PDP exp(−t/τrms )
Carrier Freq.,[GHz] 2.6 Doppler Spec. Two-dimensional, Jake’s
There are a handful of methods to remove the noise and decrease its variance
in the system. One simple low complex way (we call that DFT-B estimator) is
to capture the high energy taps of the channel after the IDFT and zero out the
rest [11]. This is equivalent to a simple vector multiplication. The time-domain
representation of noisy pilots fed into an IDFT of size Np is,
ĝls = IDFT(ĥp,ls ) (4.8)
Furthermore, the estimated channel can be described as,
ĥIDFT = DFT(ĝls × a) (4.9)
where vector a = [a1 a2 ...ap ] is a smoothing vector and the DFT size is equal to the
number of used subbands, Nc as depicted in Fig. 7. Due to the nature of wireless
channels the energy is usually concentrated in the first few taps. In case of sample-
spaced channels the energy is perfectly confined in a limited area in the beginning
of the CIR. In case of the nonsample-spaced channels, part of this energy is leaked
outside this area [11]. If the energy is confined to an area corresponding to the first
r taps and the last q taps(which is due to cyclic nature of the DFT), the elements
in vector a will be modified as follows,
a = [a1 a2 ...ar 0 0...0 ap−q ...ap ] (4.10)
where all nonzero elements in the above vector will be equal to 1. Fig. 7 shows the
general estimator structure we are using for our DFT-based estimators. The amount
of time-domain processing after IDFT in DFT-B method is minor since it is limited
to discarding the low energy channel taps and keeping the high energy ones [11].
Another low complexity DFT-based estimator (denoted as DFT-C estimator) is

to weight each channel tap by the following factor [11],
γi 2
(4.11)
γi 2 + σHp 2
where γi 2 is the tap energy and σHp 2 is the noise variance. Thus, vector a will
have Np nonzero elements each corresponding to a weighting factor. This factor
suppresses the noise energy on the channel taps and gives higher weight to useful
signal energy. This estimator has a higher complexity compared to DFT-B which
is proportional to the number of pilots. Weighting the taps in time-domain will
introduce 3 × NP additional real multiplications compared to DFT-A and DFT-
B. The performance/complexity trade-off, however, proves worthy specially when
operating in lower SNR range as illustrated in the simulations section.
5 Hardware implementation issues

Due to the availability of FFT/IFFT blocks on an OFDM-based UE, it is beneficial
to apply methods which might use the available hardware provided that it does
not jeopardize its accessibility to more critical system components (e.g., OFDM
demodulator, etc.). Since the focus of this paper is on algorithmic design, we will
leave the detailed analysis of hardware implementation issues to future work. The
nature of typical OFDM systems like LTE and DVB-H requires FFT transforms of
sizes other than radix 2. Thus, the mixed-radix FFT blocks have to be used instead.
A number of these FFT blocks have been proposed in literature [15]. In this paper we
have suggested application of a well-known mixed-radix Cooley-Tuckey algorithm
which has the radices 2,3 and 5. This suffices the requirements of our proposed
implementation for LTE and DVB-H. The block involves the following computations
for an FFT size of N [15],
N = 2p × 3q × 5r (5.1)
3 8
A(N) = 2N( p + q + 4r − 1) + 2 (5.2)
2 3
14
M(N) = 2N(p + 2q + r − 2) + 4 (5.3)
5
5 14 34
A(N) + M(N) = 2N( p + q + r − 1) + 2 (5.4)
2 3 5
where M(N) and A(N) denote the number of real multiplications and additions.
To measure the complexity of the proposed algorithms we have resorted to the
corresponding number of required multiplications. The above mentioned mixed-
radix FFT suits the LTE DFT-based estimators without any further modifications.
The number of pilots in different modes of LTE can be always broken into Np =
2p × 5r . Moreover, using the proposed pilot rearrangement in this paper the number
of subcarriers will always be 3 times as many as the number of pilots which brings
about the resulting number fed to the IFFT block being factorized into radices 2,3
5 Hardware implementation issues 55
Table 2: LTE, estimator complexity, number of real multiplications per OFDM

symbol. NRB refers to the number of allocated RBs
Estimator Total Complexity Complexity Per RB
RMMSE 75,000 3,000
MRMMSE 12,000 480
5,688
DFT-B 5,688 NRB
5,988
DFT-C 5,988 NRB
Table 3: LTE, required SNR to maintain a constant uncoded BER = 10−2 for a
number of terminal speeds. Simulations are associated with OFDM symbol 7 in a
typical RB. The figures have been derived from MRMMSE simulation results.
Speed, [km/h] Doppler spread, [Hz] SNR, [dB]
10 48 14.8
20 96 15
30 144 17
40 192 20
45 217 21
48 231 30
50 241 N.A.
and 5. Quite contrary to LTE, the standard number of pilots in DVB-H does not
allow a mixed-radix FFT as mentioned above. For instance in 2K mode, the number
of pilots in each symbol is 142 which after rearrangement (mirroring) the collected
number of pilots in the pilot vector will be 568 that cannot be factorized into the
desired radices (2,3 and 5). Thus, it is not possible to directly apply the desired
FFT block in this paper for DFT-based channel estimation in DVB-H. To overcome
this obstacle, we have extrapolated a minimum number of pilots outside the used
band such that the resulting number of total pilots can be factorized into the desired
radices. As a result, 8, 16 and 32 pilots (distributed symmetrically on each side of
the used band) need to be extrapolated corresponding to 2K, 4K and 8K modes of
operation in DVB-H. For the purposes of simulations in this paper we have resorted
to the proposed MMSE channel estimation methods (MRMMSE) to extrapolate 8
pilots in the unused subband of DVB-H which makes the total number of pilots in
the pilot vector be factorized as follows,
576 = 26 × 32 (5.5)
From a complexity overhead view point, the required additional processing for ex-
trapolation is negligible compared to the whole. A comparison among the complexity
requirements of the algorithms has been provided in the simulation section. We are,
however, aware that to extrapolate the desired pilots, estimators with a better MSE
performance exist compared to what we have applied here. These estimators can
be precalculated and fetched from memory when needed. Yet, we have resorted to
the estimators proposed in this paper since they provide acceptable performance as
evidenced in the simulations section and eliminate the need for added complexity.
To address the computational needs of the MMSE-based algorithms, a pool of
fast multipliers can be harnessed. In an efficient implementation each complex mul-
tiplication is carried out by 3 real multiplications. Similarly, a baseband processor
can provide the necessary processing power. The mentioned resources are usually
available as standard components on a typical OFDM-based UE and can be ex-
ploited for the purposes of channel estimation.
Table 4: DVB-H, simulated system parameters

FFT size 2048 δf, [kHz] 4.464
Pilots/Symbol 142 PDP exp(−t/τrms )
Carrier Freq.,[MHz] 500 Doppler Spec. Jake’s Spectrum
6 Simulation results
6.1 LTE
The LTE simulation system parameters were chosen according to Table 1. To illus-
trate the performance of the estimators, the widely used Mean Square Error as well
as the Bit Error Rate measures have been used. Furthermore, no coding or inter-
leaving schemes have been adopted. Thus, illustrated BERs correspond to uncoded
data. Besides, the BER simulations have been carried out for a SISO system so that
a fair comparison can be done among the performance of the estimators for both
LTE and DVB-H. Fig. 8 shows how our estimators perform in a Mean Square Error
sense for LTE.
As mentioned earlier, full MMSE estimator is the best existing linear estimator.
As a result, it has been selected as the reference against which other estimators are
compared. As expected, DFT-A estimator has the worst performance amongst our
proposed estimators due to its redundant noise variance both in low and high SNRs.
DFT-C performs best in low SNRs due to its capability to remove the additive noise
and capture the useful energy of the channel taps. The RMMSE as well as the
MRMMSE estimators display an acceptable performance in the MSE sense over the
whole SNR range.
Having discarded DFT-A and DFT-B, Fig. 9 depicts the performance of DFT-C,
RMMSE and MRMMSE estimators from a BER view point. It can be seen that
the vivid MSE differences occurring in the high SNR region shrink from a BER
point of view when the simulations are done for 4QAM data modulation. The es-
timators have comparable performances for a stationary environment (no Doppler).
Having a relatively equal performance, the RMMSE can be also eliminated due to
its complexity as elaborated in Table 2. Figure 10 displays the BER performance
of DFT-C and MRMMSE for two different environments. The environments have
6 Simulation results 57
Table 5: DVB-H, required SNR to maintain a constant uncoded BER = 10−2 for a
number of terminal speeds. The figures have been derived from MRMMSE simula-
tion results.
Speed, [km/h] Doppler spread, [Hz] SNR, [dB]
10 9 14
50 46 14.8
100 92 15.4
150 139 17
190 176 22
200 185 24
210 194 N.A.
0
10
DFT−A
DFT−B
DFT−C
MRMMSE,16 pilots
RMMSE
−1
10 Full MMSE
Mean Square Error
−2
10
−3
10
−4
10
−3 0 5 10 15 20 25
SNR, [dB]
Figure 11: DVB-H, MSE performance in a stationary environment.
been simulated using the classical Jake’s model [8] [16] as a reference. The variations
in this environment are simulated as an equivalent terminal speed. It can be seen
that the decreased coherence in time can be partially compensated by increasing
the Signal-to-Noise-Ratio. A quick look at Table 3 reveals the fact that it is not
possible to maintain a target uncoded BER = 10−2 when the terminal speed reaches
a breaking threshold value.
The result of our simulations suggest a terminal speed of around 50 km/h as
the breaking value. Any attempts to sustain the uncoded BER above this speed
fails even if the SNR is increased indefinitely. As discussed previously, the above
mentioned effect is mainly the result of the correlation loss among realizations of
channel taps due to the Doppler spread.
One should, however, consider the fact that the simulations have been done
for the worst-case scenario (OFDM symbol 7) and it is expected that on average
the breaking value for the terminal speed increases. For instance, our uncoded
BER simulations for OFDM symbol 1 indicates a speed of 90km/h as the breaking
threshold. We also know from experience that a mobile UE rarely operates in fast
fading channels characterized by high terminal speeds and as a result a top speed of
0
10
RMMSE
MRMMSE,16 pilots
DFT−C
−1
10
BER
−2
10
−3
10
−4
10
−3 0 5 10 15 20 25
Eb/N0, [dB]
Figure 12: DVB-H, uncoded BER performance for 4QAM in a stationary envi-
ronment.
around 50 km/h satisfies the user needs in a majority of situations. It should also be
pointed out that no coding/interleaving schemes have been used in the simulations
and the system might perform better if the complete chain as standardized in LTE
and DVB-H is implemented.
The complexity of DFT-based methods depend on the total number of available
pilots corresponding to the LTE mode of operation. For the system simulated in
this paper 100 scattered pilots have been used. Table 2 shows the required number
of multiplications to estimate the channel for the complete set of subcarriers corre-
sponding to each OFDM symbol. The figures suggest that these estimators have a
lower complexity compared to their MMSE counterparts. However, the overall cost
of these estimators increases in the mean sense for LTE. The underlying reason can
be described by the fact that the resource allocation in LTE is not always contigu-
ous. As stated earlier, as few as two consecutive Resource Blocks (RB) -one RB
in each symbol- can be allocated to each terminal. That alone implies the redun-
dancy of using the proposed DFT-based methods since they do the estimation over
all subccariers in each symbol. On the other hand, MRMMSE estimator proves to
provide a lower complexity in the mean sense since the complexity is a function of
the number of estimated subcarriers (allocated RBs). The estimation complexity
can be described and quantized on the basis of per Resource Block (RB). While the
complexity of DFT-based estimator remains the same irrespective of the allocated
number of subcarriers, the MRMMSE estimator’s complexity can range from 480 to
12,000 multiplications per RB.
6.2 DVB-H
The DVB-H simulation system parameters were according to Table 4. The sim-
ulation environment has been modeled similar to LTE except for the parameters
which are specific to DVB-H like PDP length, etc. Furthermore, MSE as well as
0
10
DFT−C (80 km/h)
DFT−C (200 km/h)
−1
BER 10
−2
10
−3
10
−4
10
−3 0 5 10 15 20 25
Eb/N0, [dB]
Figure 13: DVB-H, uncoded BER performance for 4QAM corresponding to two
different terminal speeds.
BER have been used as measures of estimator performance. Fig. 11 shows the MSE
performance while Fig. 12 displays the BER results for the proposed estimators in
a stationary environment. Similarly, we can witness the quality of these estimators
under varying channel conditions as depicted in Fig. 13. Analogous to LTE, we
have investigated the BER breaking terminal speed for the proposed algorithms.
Table 6: DVB-H, estimator complexity, number of real multiplications per OFDM

symbol
Estimator Complexity
RMMSE 2,661,648
MRMMSE 74,976
DFT-B 44,168
DFT-C 45,896
The results in Table 5 expose us to the fact that these algorithms perform well
for a larger range of terminal speeds when applied to DVB-H. This can be partly de-
scribed by the lower carrier frequency in DVB-H as well as the used pilot structure.
Contrary to LTE, DFT-based estimators provide lower complexity when compared
to MRMMSE for DVB-H systems. This is largely due to the fact that DVB-H is a
broadcast scheme in which each terminal needs to decode the entire subcarriers in
each OFDM at a time. Thus, the terminal needs to estimate all the data subccari-
ers to decode the received data. Table 6 reveals the required complexity for each
estimator. Hence, the DFT estimators provide a lower complexity in contrast with
MRMMSE.
7 Conclusion
In this paper, a number of novel methods have been proposed to address the channel
estimation needs for a mobile terminal operating in a multi-standard environment.
Exploiting pilots in past/future symbols makes a core estimator plausible. The re-
sultant pilot spacing for both LTE and DVB-H provides the opportunity to design a
single estimator. These estimators have a relatively low computational complexity
when compared to other investigated methods. The coexistence of these two low
complexity estimators can provide the reliability and flexibility needed for the LTE
as well as DVB-H systems. These estimators, however, lose their performance when
significant changes in the environment (as a function of Doppler spread) occur spe-
cially for LTE. Considering the fact that UEs on average operate in a low terminal
speed range as well as the reduced complexity which is gained by applying these
estimation methods, the performance loss due to Doppler spread becomes accept-
able in many situations. Meanwhile, these estimators perform well for relatively
high terminal speeds when implemented for DVB-H which is mainly the result of
the standardized pilot pattern as well as lower carrier frequency. Thus, the over-
all performance of these estimators when applied to fast fading channels becomes
acceptable when both LTE and DVB-H are taken into consideration.
To improve the performance of these estimators for higher Doppler spreads in
LTE, low complexity estimation techniques need to be applied to track the channel
variations in time. Designing new estimators as such does not void the application
of the proposed estimators in this paper. They will appear as added computational
overhead which is due to time-domain filtering for LTE.
References
[2] P. Fertl and G. Matz, “Multi-user channel estimation in OFDMA uplink systems
based on irregular sampling and reduced pilot overhead,” in Proc. IEEE Int.
Conf. Acous., Speech and Signal Process., Honolulu, HI, pp. III.297–III.300,
2007.
[3] “LTE: An introduction [Online],” Ericsson AB, 2011. Available:

http://www.ericsson.com/res/docs/2011/lte-an-introduction.pdf.
[4] E. Dahlman et al., 3G Evolution, Second Edition: HSPA and LTE for Mobile
Broadband. Burlington, MA: Academic Press, 2008.
[5] “Evolved Universal Terrestrial Radio Access (E-UTRA); Physical layer proce-
dures (Release 8),” 3rd Generation Partnership Project, Tech. Specifications
3GPP TS 36.213, May 2008. Available: http://www.3gpp.org/ftp/Specs/html-
info/36213.htm.
7 Conclusion 61
[7] O. Edfors, Low-complexity algorithms in digital receivers. PhD thesis, Div. of

Signal Process., Luleå Univ. of Technology, Luleå, Sweden, 1996.
[8] M. Gans, “A power-spectral theory of propagation in the mobile-radio environ-

ment,” IEEE Trans. Veh. Technol., vol. 21, pp. 27–38, Feb. 1972.
[9] A. Goldsmith, Wireless communications. New York, NY: Cambridge University

Press, 2005.
[10] M. Morelli and U. Mengali, “A comparison of pilot-aided channel estimation

methods for OFDM systems,” IEEE Trans. Signal Process., vol. 49, pp. 3065–
3073, Dec. 2001.
[11] O. Edfors, M. Sandell, J.-J. van de Beek, S. K. Wilson, and P. O. Börjes-

son, “Analysis of DFT-based channel estimators for OFDM,” Wireless Personal
Communications, vol. 12, no. 1, pp. 55–70, 2000.

[13] V. Srivastava et al., “Robust MMSE channel estimation in OFDM systems with
practical timing synchronization,” in Proc. IEEE Wireless Commun. and Netw.
Conf., Atlanta, GA, pp. 711–716, 2004.
[14] S. Omar, A. Ancora, and D. Slock, “Performance analysis of general pilot-

aided linear channel estimation in LTE OFDMA systems with application to
simplified MMSE schemes,” in Proc. IEEE 19th Int. Symp. Personal, Indoor
and Mobile Commun., Cannes, pp. 1–6, 2008.
[15] D. Takahashi and Y. Kanada, “High-performance radix-2, 3 and 5 parallel 1-D

complex FFT algorithms for distributed-memory parallel computers,” The J.
of Supercomputing, vol. 15, pp. 207–228, Feb. 2000.
[16] D. Young and N. Beaulieu, “The generation of correlated rayleigh random vari-
ates by inverse discrete Fourier transform,” IEEE Trans. Commun., vol. 48,
no. 7, pp. 1114–1127, 2000.
Low-complexity channel
estimation for LTE in fast
fading environments for
implementation on
multi-standard platforms
Farzad Foroughi Abari, Farnaz Sharifabad, and Ove Edfors
Paper II
c 2009 IEEE. Reprinted, with permission, from Farzad Foroughi Abari, Farnaz
Sharifabad, and Ove Edfors, “Low complexity channel estimation for LTE in fast fad-
ing environments for implementation on Multi-Standard platforms,” in Proc. IEEE
Veh. Tech. Conf., Ottawa, pp. 1–5, Sep. 2010.
1 Introduction 65
Abstract
In this paper a number of low-complexity channel estimator structures tai-
lored to the needs of LTE terminals have been investigated. The focus has been
on the algorithms exhibiting reduced complexity while providing an acceptable
level of performance. To design these estimators other factors such as inter-
operability with another OFDM-based standard, DVB-H, in a multi-standard
environment have been taken into account. The key underlying parameter has
been exploiting the pilot pattern in LTE providing the possibility of minimal
filtering in time domain while exhibiting an acceptable level of performance in
fast fading environments.
1 Introduction
Providing a multiple-access scheme, Orthogonal Frequency Division Multiplexing
(OFDM) has replaced many of the traditional single carrier alternatives where data
is encapsulated in symbols which are usually shorter than the time span of the
physical channel. As a result, the communication over single carrier systems is
usually prone to Inter Symbol Interference (ISI) and needs complicated equalization
structures to reliably retrieve and decode the data on the receiver side.
In an OFDM based systems, however, data is modulated on a number of subcar-
riers which occupy a predefined BandWidth (BW). The system is designed in such
a way that subcarrier spacing is chosen to be smaller than the coherence bandwidth
of the channel. As a result, simple equalization techniques can be carried out on the
receiver side. The system, however, can suffer from Inter Carrier Interference (ICI)
which results from the fast fading channel environments. To counter this problem,
the symbol time in OFDM systems is designed to be shorter than the coherence
time of the physical channel. Contrary to single carrier systems, parallel data mul-
tiplexing in the downlink for a number of User Equipments (UE) is rather simple
in OFDM. The original idea of OFDM dates back to 1966 [1] but implementing the
system was cumbersome due to technology shortcomings. The advances in digital
signal processing and circuit design, however, have enabled practical realization of
OFDM systems in recent years.
Initially, OFDM systems were used in combination with Single-Input-Single-
Output systems (SISO) such as Digital Audio Broadcast (DAB). With the intro-
duction of Multiple-Input-Multiple-Output (MIMO), opening a new horizon in the
area of digital communications, new standards benefiting from the combination of
both OFDM and MIMO emerged into the telecommunications world. The applica-
tion of multiple antennas on both transmitter and receiver enables the system to
enjoy various diversity, multiplexing and beam forming schemes.
Long Term Evolution (LTE) has been opted as the 4th generation standard
for mobile wireless communications. LTE combines both OFDM and MIMO in
downlink to materialize the desired data rates of up to 100 Mbits/Sec in fading
environments. Channel estimation in LTE is a complex but nonetheless important.
To enable proper channel estimation, scattered pilots have been employed in the
LTE downlink channel.
Paper II: Low-complexity channel estimation for LTE in fast fading environments
66 for implementation on multi-standard platforms. . .
Figure 1: LTE, scattered pilot structure. Each RBP consists of two consecutive
RBs and covers 12 subcarriers over 14 OFDM symbols.
Channel estimation based on scattered pilots is considered a challenging task

when limited processing power restricts the scope of the applied algorithms. The
estimator usually needs to interpolate the unknown subcarriers in two dimensions
(time and frequency) which adds to the inherent complexity. The computational
complexity becomes even more detrimental when the estimation needs to be carried
out for a multitude of channels when the system is configured with MIMO.
In this paper we, in addition to a comparative analysis between previously known
estimation techniques, also introduce a new channel estimation technique which ex-
hibits both very low complexity as well as satisfactory performance up to terminal
speeds of 120 km/h, as specified for LTE. The new estimation technique is de-
signed with multi-standard processing on a single terminal platform in mind, like
in [2] where common channel estimation structures for both LTE and Digital Video
Broadcasting-Handheld (DVB-H) are addressed. Comparisons are provided in terms
of the required number of multiplications as well as memory usage.
2 LTE (Long Term Evolution)

Mobile broadband encouraged the specification of LTE which became the basis for
the next generation of UMTS Mobile standards [3]. One of the motivations behind
LTE system design is to break the limitations of high data rates in wireless channels.
The standardization not only demands LTE to provide high data rates in stationary
channel environments but also demands that LTE based terminals provide a peak
data rate of 100 Mbits/Sec in fading environments characterized with an equivalent
terminal speed of up to 50 km/h while maintaining a moderate performance when
2 LTE (Long Term Evolution) 67
the terminal speed increases up to 120 km/h [4]. While many of the system compo-
nents in LTE have already been standardized, some of the intermediate components,
e.g., the channel estimator, have been left for the individual system designers and
manufacturers to contemplate. As a result, each vendor has the opportunity to
integrate its own designated algorithms into their products.
To assist channel estimation, pilot data known to both the transmitter and re-
ceiver are sent over the wireless channel. The pilots are either continual (e.g., in
IEEE 802.11n) or scattered (e.g., LTE). Depending on the expected behavior of
the wireless channel, one or a combination of both schemes will be employed for
channel estimation purposes. LTE standardization has utilized scattered pilots in
the downlink to assist the channel estimation. Fig. 1 illustrates how pilots have
been scattered in time and frequency grid for a single antenna, i.e., antenna port
0. In the LTE standard, the term antenna port is preferred over physical antenna
to emphasize the fact that multiple physical antennas can transmit the same data
and as a result the combination can be considered as one antenna port using the
same pilot pattern. Considering the fact that channel estimation for MIMO systems
can be more cumbersome, as compared to SISO systems, LTE has defined differ-
ent pilot patterns for the different antenna ports. To reduce the channel estimation
complexity at the expense of spectral efficiency, the subcarriers corresponding to the
pilot locations in one antenna port remain silent during data transmission on the
other antenna ports. As a result, the estimator blocks don’t need to take interfer-
ence among the antenna ports into consideration when using pilot assisted channel
estimation methods.
In contrast to some of the contemporary wireless communications systems such
as DVB-H, data can be multiplexed for different users in each OFDM symbol in
LTE. In other words, the Band Width (BW) of the system can be broken into a
number of units each carrying data for one single UE. In LTE a Resource Block (RB)
consists of twelve consecutive subcarriers in frequency and seven OFDM symbols in
time as depicted in Fig. 1. Meanwhile, resource allocation for each UE consists
of at least two consecutive RBs in time. Thus, the smallest allocable data unit
to a typical UE is 12 subcarriers in 14 consecutive OFDM symbols known as one
RB Pair (RBP). Fig. 1 shows how RB and RBP are defined over time-frequency
grid in LTE. Having mentioned that, the UE needs to receive the first three OFDM
symbols in each RBP to decode the control signaling. Among other things, control
signals carry information about data scheduling [5] for each UE. If no data has been
allocated for a certain UE, it can enter a standby mode for the next eleven OFDM
symbols to save power.
The pilots associated with one RBP in LTE are distributed in OFDM symbols 1,
5, 8 and 12 for a 2 × 2 MIMO scheme. The pilot spacing in frequency domain for the
pilot carrying OFDM symbols is 6 subcarriers. Fig. 1 displays an example of pilot
structure for antenna port 0. The same pattern is shifted in frequency for antenna
port 1. These pilots provide possibility of employing various channel estimation
techniques, depending on the accuracy/complexity requirements of the system, to
acquire the proper Channel State Information (CSI).
Figure 2: LTE, simplified MIMO system model.
3 System model
Fig. 2 shows a simple MIMO OFDM system model which can be used as a reference
for a simplified LTE system. For the purposes of simulation in this paper a 2 × 2
MIMO scheme with optional turbo coding, similar to the one standardized for LTE,
has been used. LTE benefits from two different Cyclic Prefixes (CPs) known as
normal and extended CPs [5] to prevent ISI. In this paper, we have resorted to the
application of the normal CP. The same analysis can, of course, be done with the
extended CP. This is, however, beyond the scope of this paper. Similar to a typical
OFDM-based system, information is modulated through one of the standardized
modulation schemes for LTE, e.g., 4QAM. The modulated data as well as the pilots
are placed on the corresponding subcarriers and after zero padding, as standardized
in LTE providing the necessary Guard Band (GB), the result is fed into an IFFT.
The IFFT output is then CP extended, A/D converted and transmitted over a time
dispersive wireless channel with AWGN. The wireless channel length does not exceed
the CP length. The received signal in the form of added multi-path components is
received on each receiving antenna and passed through a D/A. After CP removal,
the result is OFDM demodulated through an FFT block. It is assumed that the
channel does not vary during each OFDM symbol transmission and hence the there
is no ICI in the system. Besides, we have assumed perfect synchronization in the
system.
We assume that the wireless channels between each pair of transmit/receive
antennas have the same properties and each can be mathematically described as [6],
L−1
X
h(t, τ ) = αl (t)δ(τ − τl ) (3.1)
l=0
where L is the number of taps, αl is the complex path gain and τl is the path delay
corresponding to each tap. The above equation is the description of a time-varying
linear system in which the path gains, αl , are WSS complex Gaussian random pro-
3 System model 69
Figure 3: Merged-pilot estimation technique, merging (collecting) the pilots into

Pilot Vectors (PVs). This picture shows how pilots from the 1st and 5th OFDM
symbols are collected to create PV1. For estimation of subcarriers in each region, the
pilots from the two closest pilot carrying OFDM symbols are collected and fed into
the MRMMSE estimator. Hence, four PVs are constructed and channel estimation is
done four times per RBP each corresponding to one region for equalization purposes.
The depicted PV in this figure consists of eight pilots chosen symmetrically with
respect to subcarriers k + 7 and k + 8 and is used to estimate them in region 1. The
proposed estimator in this paper, however, uses sixteen pilots instead.
cesses. The Channel Frequency Response (CFR) of the wireless channel at time t
can be written as, Z +∞
H(t, f ) = h(t, τ ) exp(−2jπτ )dτ. (3.2)
−∞
Assuming WSS channel taps, the time-frequency correlation function of a channel
can be described as,
R(∆f, ∆t) = E[H(f1 , t1 )H(f2 , t2 )]. (3.3)
It can be shown that the frequency correlation function, which is obtained by setting
∆t = 0 in R(∆f, ∆t) is the Fourier transform of the Power Delay Profile (PDP) of the
channel, while the time correlation function, which is obtained by setting ∆f = 0
in R(∆f, ∆t) can be described by the inverse Fourier transform of the Doppler
spectrum of the channel. The inverse Fourier transform of the Doppler spectrum
for a two dimensional isotropic uncorrelated scattering environment is [7]
R(∆t) = J0 (2πφmax ∆t), (3.4)
where J0 (.) is the zeroth-order Bessel function of the first kind, φmax is the maximum
Doppler in Hertz and ∆t is the time difference between two observations of the
channel at two different time instances. Equivalently, the coherence time of the
channel, Tcoh , can be defined as the time interval over which the autocorrelation
falls 3 dB from its maximum value [8]. According to equation (3.4), Tcoh decreases
inversely with the maximum Doppler frequency.
The scattered pilot structure in LTE enables the application of various channel
estimation techniques. The proper choice of the estimators depends on many factors
among which the quality and the complexity of the estimators are the dominant
criteria. Meanwhile, a lot of research has been done in the area of channel estimation
for the wireless systems employing scattered pilots [6]. In one work [2], a number of
DFT based estimators have been analyzed when applied to both LTE and DVB-H
for a multi-standard terminal. The best linear estimator in terms of Mean Square
Error (MSE) is the well-known 2-D LMMSE which exploits the second order channel
statistics in time and frequency to estimate/interpolate the unknown subcarriers.
However, real time processing of such a 2-D filter is intensive and due to its inherent
complexity it is rarely used in practical systems. Meanwhile, it has been shown that
the 2-D filter, under certain conditions, can be broken into two 1-D filters one in
time and one in frequency with negligible performance loss [9]. The 1-D filter in
frequency can be mathematically described as [10],
β
ĥmmse = Rhm hp (Rhp hp + I)−1 ĥp,ls (4.1)
SNR
where Xp and Yp are the transmitted and received data sampled on the pilot posi-
tions and,
ĥp,ls = X−1
p Yp (4.2)
0
10
−1
10
BER
−2
10
perfect channel knowledge

MRMMSE, piec.−cons., V=50 km/h
−3 MRMMSE, merged−pilot, V=50 km/h
10
2−D MRMMSE, V=50 km/h
MRMMSE, piec.−cons., V=120 km/h
MRMMSE, merged−pilot, V=120 km/h
2−D MRMMSE, V=120 km/h
−4
10
−5 0 5 10 15 20 25 30
Eb/N0, [dB]
Figure 4: LTE, uncoded BER. The simulations have been done for a 2 ×2 MIMO
system configured for spatial multiplexing using a ZF receiver.
is the Least Squares estimate of the channel attenuations hp on the pilot tones,
Rhp hp is the autocorrelation of the sampled channel hp on pilot positions, Rhm hp
is the cross correlation matrix between the pilot tones and the data subcarriers,
SNR = E|xk |2 /σ 2 is the signal-to-noise ratio and β is a constant which depends on
the signal constellation of transmitted pilot symbols, X. For a 4QAM modulation
scheme β equals 1.
According to (4.1), a matrix inversion of Np × Np , where Np is the number of
available pilots, needs to be done each time the second order channel statistics or the
SNR value change. Thus, the complexity of the above estimator is still formidable.
Due to the fact that the processing power on the mobile terminal is usually limited,
a number of low complexity variants of the above estimator need to be derived.
Reducing the complexity of the above estimator has been investigated in a number
of previous works [6]. A simple method to overcome the matrix inversion problem,
is to use a robust variant of the above estimator, robust MMSE [6], designed for
a fixed SNR and a fixed PDP. Despite its performance loss, its reduced complex-
ity makes this estimator more attractive for practical applications. In this paper,
we have chosen a uniform design PDP where the worst case channel condition has
been presumed. Although the robust MMSE estimator is more computationally
efficient than the full MMSE, Np complex multiplications need to be done to esti-
mate each subcarrier in frequency. To counter the complexity problem even more,
it has been shown [2] [11] that not all the available pilots in one OFDM symbol
need to be used to estimate/interpolate each subcarrier. Acceptable performance
can be achieved by only using the pilots in the vicinity of each subcarrier due to
the correlation loss between subcarriers and pilots which are located far apart, es-
pecially in highly frequency selective channels. It has been shown [2] that for the
LTE system configuration simulated in this paper, as few as 16 pilots will provide
satisfactory performance in terms of BER. We call this estimator Modified Robust
MMSE (MRMMSE) as originally proposed in [2].
Table 1: LTE,simulated system parameters

FFT size 512 ∆f, [kHz] 15
Pilots/Symbol 50 PDP exp(−t/9)
Carrier Freq.,[GHz] 2.6 Doppler Spec. Jake’s
Similarly, a robust MMSE in time direction can be used to estimate/interpolate

the unknown subcarriers in time. The applied 1-D robust MMSE is designed based
on the second order channel statistics in (3.4). Thus, the simplified 2-D MMSE
(2-D MRMMSE) can be performed through two separate 1-D filtering first in time
or frequency, followed by the filtering in the other domain [6]. In order to lower the
complexity of the algorithm even further, we have adapted two more time domain
strategies which waive the 1-D filtering in time. One approach is based on the well-
known piecewise constant interpolation where the channel is estimated for each pilot
carrying OFDM symbol in frequency and used for all the other subsequent nonpilot
carrying symbols. As soon as the next pilot carrying symbol is received, a new
frequency domain estimation is done and used likewise. The third method, merged-
pilot 1 , uses the mutual correlation existing between two consecutive pilot carrying
OFDM symbols through merging (collecting) the pilots into a single Pilot Vector
(PV) as depicted in Fig. 3. The constructed PV is then fed into a channel estimator
and the corresponding CSI is estimated. The same CSI is used for equalization of
all symbols exhibiting the shortest distance from the pilot contributing symbols.
In other words, four CSIs corresponding to four PVs are obtained. CSI1 is used
for equalization of OFDM symbols 1-4 (region 1), CSI2 is used for OFDM symbols
5-8 (region 2), CSI3 is used for OFDM symbols 9-11 (region 3) and CSI4 is used
for OFDM symbols 11-14 (region 4). To achieve optimal results for region 4, the
pilots from the first OFDM symbol corresponding to the next consecutive RBP are
used to construct PV4 which is a valid assumption because the terminal always
needs to listen to the first three OFDM symbols of each RBP. The feasibility of the
above approach arises from a rather strong correlation between consecutive OFDM
symbols in LTE. For the presumed system parameters in this paper, the Tcoh derived
from equation (3.4) covers as many as 6 OFDM symbols for an equivalent terminal
speed of 120 km/h. However, the largest distance between any pilot carrying OFDM
symbols in this approach is 4 which is associated with the PV1.
The above three approaches are characterized by their own pros and cons. The
2-D MRMMSE provides the best performance, esp. in high Doppler scenarios. Yet,
it has a larger computational complexity and causes a delay in the equalization and
decoding since all the pilot symbols in the entire RBP are used for the estimation.
1
We have used the term mirrored-pilot in [2] which is literally the same as the merged-pilot.
The notation has been changed to eliminate the probable ambiguity/confusion associated with the
term, mirrored.
32
30
28
26
SNR, [dB]
24
22
20
18
16 Merged−pilot
piecewise−constant
14
20 40 60 80 100 120
V, [km/h]
Figure 5: LTE, SNR vs. velocity to sustain an uncoded BER = 10−2 for a 2 × 2
MIMO system configured for multiplexing with ZF receiver. The figure shows the
performance superiority of the merged-pilot over the piecewise constant for both
slowly and fast fading wireless channels.
The piecewise constant approach does not need much memory to store symbols be-
fore the estimation initiates. Hence, it does not cause any considerable delay in
the system. Besides, the complexity is only restricted to the MRMMSE applied in
the frequency domain. The performance of this approach, however, is quite limited
specially when used in fast fading environments. The last but not least approach,
merged-pilot, provides a good compromise between hardware resources and perfor-
mance as shown in the simulation section. This new approach has a considerably
higher performance compared to the piecewise constant in environments character-
ized with high Doppler spreads. The performance gain can be described by two
facts. First, the pilot spacing corresponding to the PV in the merged-pilot approach
is half the spacing in the piecewise constant. Consequently, the cross correlation
between the pilots and the estimated subcarriers is higher. Second, the mutual cor-
relation provided by the pilots merged from two different OFDM symbols increases
the reliability of the estimator in fading environments, esp. in areas located half way
between the two pilot contributing symbols. Meanwhile, computational complex-
ity of this estimator is identical to the piecewise constant’s because both methods
use the same MRMMSE (16 pilots) estimator in frequency. Yet, the merged-pilot
approach has higher memory requirements since all the intermediate OFDM sym-
bols need to be stored before the pilots from the target pilot carrying symbols are
merged. Thus, an equalization lag hits the system which is still considerably lower
than the simplified 2-D MRMMSE.
0
10
perfect chan. knowl., Div.
Merged−pilot, Div., V=50 km/h
Merged−pilot, Div., V=120 km/h
−1
10 perfect chan. knowl., Mult.
Merged−pilot, Mult., V=50 km/h
Merged−pilot, Mult., V=120 km/h
−2
10
BER
−3
10
−4
10
−5
10
0 5 10 15 20 25
Eb/N0, [dB]
Figure 6: LTE, coded BER performance corresponding to the merged-pilot es-

timator for two different terminal speeds. Simulations have been done for a 2 × 2
MIMO system configured for both multiplexing and diversity using a ZF receiver.
The diversity results have been simulated using Alamouti precoding scheme.
For the purposes of simulation in this paper, the LTE simulation system parameters
were chosen according to Table 1. To illustrate the performance of the estimators,
the widely used Bit Error Rate (BER) measure has been used. Although the choice
of estimators has been based on the conventional uncoded BER, coded BER using
the standardized turbo code in LTE [5] has also been simulated for the estimators of
choice proposed in this paper. Besides, the simulations have been mainly performed
for a 2 × 2 MIMO scheme configured for spatial multiplexing using a Zero Forcing
(ZF) receiver. Since the possibility of diversity has also been provided in the LTE
system design, some of the simulations have also been done for a 2 × 2 MIMO with
diversity. Fig. 4 shows the performance of the estimators for a fading environment
for three different terminal speeds from an uncoded BER view point. The BER
performance reveals the advantage of collecting the pilots in the merged-pilot over
the piecewise constant approach. The difference becomes even more significant
when the terminal speed increases. The simplest method, piecewise constant, does
not exhibit a good level of performance for the terminal speed of 120 km/h. The
merged-pilot, however, offers a better performance as expected.
Table 2 contains the complexity and memory requirements of each technique
per RBP. The complexity as well as memory requirements of the 2-D MRMMSE is
much higher than the other two alternatives. The excessive complexity becomes pro-
hibitive on mobile terminals especially in case of multi-standard platforms. On the
other hand, our alternative approach, the merged-pilot estimator, provides an ac-
ceptable level of performance while exhibiting a moderate complexity. The piecewise
constant estimator, however, has a significantly worse performance at high terminal
speeds. Fig. 5 further describes the merits of the merged-pilot approach over the
6 Conclusion 75
piecewise constant. It can be seen from the figure that the uncoded BER = 10−2
breaking terminal speed is approximately 130 km/h for the merged-pilot approach
Table 2: LTE, complexity versus memory requirements for the three estimation
techniques. The results are estimated per RBP.
Estimation technique Complexity, [real Mul.] Memory requirements, [byte]
piecewise-constant 2688 96
merged-pilot 2688 480
2-D MRMMSE 8424 1440
while that value drops to almost 80 km/h for the piecewise constant. Theoretically,
it is possible to sustain an acceptable uncoded BER = 10−2 by having SNR = 24 dB
for a terminal speed of 120 km/h in the merged-pilot case while the other estimator
cannot sustain the BER = 10−2 for the same speed even if the SNR is increased
indefinitely. Fig. 6 demonstrates the behavior of the merged-pilot estimator when
coding is used. It can be witnessed that a coded BER = 10−2 can now be reached
at SNR = 14 dB which is almost 10 dB lower compared to the uncoded BER when
the system is configured for multiplexing. If the system is configured for diversity,
on the other hand, the BER = 10−2 can be reached already at SNR = 5 dB which
is almost 20 dB lower than for the uncoded case. Fig. 6 also reveals that for most
SNRs the simulated LTE system using our proposed estimator is only a few dBs
away from one with perfect channel knowledge. We also know from experience that
it is possible to have as a high SNR as 25 dB in a mobile terminal. As a result, the
performance of the estimator is well within the operational SNR region of the UE
and can be reliably used in practical systems.
6 Conclusion
In this paper, the MRMMSE in frequency in conjunction with three time-domain
strategies has been analyzed. The performance of the estimators for fast fading
environments shows that the best trade-off between complexity and performance
corresponds to the merged-pilot method as proposed in this paper. It has been
shown that this low complexity estimator outperforms the conventional piecewise
constant approach in fast fading environments while exhibiting the same complexity
level and waives the need for computationally complex estimators which involve
additional time-domain filtering. More importantly, the structure of this estimator
enables its efficient application in a multi-standard environment where LTE and
DVB-H coexist [2]. Having a core estimator as such, shared between LTE and
DVB-H, enables efficient hardware design which results in reduced chip area, lower
power consumption, etc.
7 Acknowledgments
The authors would like to thank MULTI-BASE, a 7th Framework Program funded
by the European Commission, for supporting this work.
References
[2] F. Foroughi Abari et al., “Channel estimation for a mobile terminal in a multi-
standard environment (LTE and DVB-H),” in Proc. IEEE Int. Conf. Signal
Process. and Commun. Syst., Omaha, NE, pp. 1–9, Sep. 2009.
[3] “LTE: An introduction [Online],” Ericsson AB, 2011. Available:

http://www.ericsson.com/res/docs/2011/lte-an-introduction.pdf.
info/36201.htm.
[5] E. Dahlman et al., 3G Evolution, Second Edition: HSPA and LTE for Mobile
Broadband. Burlington, MA: Academic Press, 2008.
[7] D. Young and N. Beaulieu, “The generation of correlated rayleigh random vari-
ates by inverse discrete Fourier transform,” IEEE Trans. Commun., vol. 48,
no. 7, pp. 1114–1127, 2000.
[8] A. F. Molisch, Wireless communications. West Sussex, United Kingdom: John

Wiley & Sons Ltd., 2007.


[11] K. Manolakis et al., “Performance evaluation of a 3GPP LTE terminal receiver,”

in Proc. 14th European Wireless Conf., Prague, pp. 1–5, 2008.
On coefficient memory
co-optimization for channel
estimation in a
multi-standard environment
(LTE and DVB-H)
Farzad Foroughi Abari, Fredrik Rusek, and Ove Edfors
Paper III
c 2009 IEEE. Reprinted, with permission, from Farzad Foroughi Abari, Fredrik
Rusek, and Ove Edfors “On coefficient memory co-optimization for channel estima-
tion in a multi-standard environment (LTE and DVB-H),” in Proc. IEEE 8th Int.
Workshop on MultiCarrier Syst. and Solutions , Herssching, pp. 1–5, May 2011.
1 Introduction 79
Abstract
This paper presents a novel memory optimization technique to exploit the
analogy between correlation coefficients in MMSE-based channel estimators
across multiple standards. The need for coefficient storage for robust MMSE
channel estimators is expensive in terms of on-chip memory. Besides, the
memory requirements grow linearly by integrating more standards onto one
platform. Thus, it becomes inevitable to address the increased on-chip memory
problem for such estimators. In this paper, we have shown that by exploiting
the inherent similarities between LTE and DVB-H design parameters, a three-
fold memory optimization may be achieved with minimal performance loss.
1 Introduction
Orthogonal frequency division multiplexing (OFDM) is a driving force behind the
latest high speed standards in wired as well as wireless communications. The ability
to break a wide spectrum into narrowband subchannels where data could possibly
be multiplexed for different users has provided a number of different possibilities,
including but not limited to, reducing or eliminating the inter symbol interference
(ISI), using simple equalizers and employing flexible channel estimators.
Long Term Evolution (LTE) is the 4th generation standard for mobile wireless
communications. It provides high spectrum efficiency with data rates of up to 100
Mbits/s. To facilitate channel estimation, scattered pilots are used in the downlink
transmission. As a result, a variety of channel estimation algorithms may be used,
as suggested in literature [1].
Similar to LTE, Digital Video Broadcast-Handheld (DVB-H) employs OFDM for
data transmission. DVB-H is an evolution of Digital Video Broadcasting-Terrestrial
(DVB-T) to address the needs of mobile terminals in time-varying wireless channels.
To facilitate channel estimation, scattered pilots are also used in DVB-H. Although
the pilot patterns in LTE and DVB-H are different, they both exhibit similarities
which may be exploited for co-optimization.
This paper is the continuation of the work started in [2] and followed by [3].
Although the channel estimation techniques presented in this paper are similar to
[2, 3], the presented memory optimization principle across LTE and DVB-H in this
work binds the proposed estimators into a single channel estimation package. We
will show that by exploiting the similarity between the 2nd-order wireless channel
statistics across LTE and DVB-H a three-fold on-chip memory reduction associated
with the channel estimator filter coefficients can be achieved.
The structure of the paper is as follows. After giving a short introduction to
DVB-H and LTE standards in Sec. 2, the employed channel estimation approach is
briefly discussed in Sec. 3. In Sec. 4 we show how designing the proposed estimator
for specific system parameters allows us to reuse certain coefficients across LTE and
DVB-H. Furthermore, it is shown that by designing the estimators for a truncated
guard interval in LTE further memory optimization is achieved. Finally, in Sec. 5
simulation results for the proposed design methods in terms of channel estimation
performance are discussed and conclusive remarks are given in Sec. 6.
Paper III: On coefficient memory co-optimization for channel estimation in a
80 multi-standard environment (LTE and DVB-H). . .
Figure 1: An example of time-frequency grid corresponding to one RBP for LTE.

The scattered pilots are identified by the black boxes.
2 LTE and DVB-H

LTE exploits the spectrum efficiency provided by OFDM and promises high data
rates in wireless communications systems which can be harnessed for various pur-
poses, such as, wireless Internet, high-quality multimedia transmission, etc. It also
provides the flexibility for various system implementations through different modes
of operation, i.e., different bandwidths (BW). The supported BW range (2-20 MHz)
corresponds to DFT sizes of 128 − 2048 points.
To combat ISI two different cyclic prefixes (CP) are defined for LTE, i.e., normal
and extended CPs each having a length equal to that of the maximum expected
delay spread in the wireless channel. The normal CP has a fixed length of 4.69 µs
and the extended CP’s spread is equal to 16.67 µs.
To facilitate channel estimation for such a wide spectrum in a multi-path fading
environment, scattered pilots are introduced in the physical layer of LTE. An exam-
ple of scattered pilots for a resource block pair (RBP) [4] is depicted in Fig. 1. As
can be seen the pilots are scattered in both time and frequency, which provides the
opportunity to apply various channel estimation schemes.
DVB-H is also an OFDM based system. It is a refinement of DVB-T and is
adapted for the needs of mobile communications. Among many different modifica-
tions, one can mention the addition of 4k operation mode which offers an additional
trade-off between transmission cell size and mobile reception capabilities [5].
...
Time
...
Frequency
Data Scattered Pilot
Figure 2: An example of time-frequency grid for DVB-H. The scattered pilots are
identified by the black boxes.
Similar to many other OFDM-based systems different CPs are also standardized
for DVB-H. More precisely, there are four different CPs adopted in DVB-H which
are known as 1/4, 1/8, 1/16 and 1/32. The numbers correspond to the ratios of the
CP duration and the useful symbol duration.
Scattered pilots are also employed in DVB-H for channel estimation purposes.
Fig. 2 illustrates an example of scattered pilots distribution, in time and frequency,
for a typical DVB-H system. It can be observed that the pilot density is higher in
DVB-H when compared to LTE. This enables the possibility to adopt a wider set of
channel estimation techniques, such as various merging schemes elaborated in this
paper.
A number of techniques can be used for pilot-assisted channel estimation in LTE and
DVB-H. Depending on the estimators complexity-performance trade-off, several al-
gorithms may be used in practice [6]. In [2, 3] the concept of a modified robust MMSE
(MRMMSE) estimator is introduced. The MRMMSE enables low-complexity chan-
nel estimation with an acceptable performance in fading environments specified by
LTE standard [4].
Considering a wide sense stationary uncorrelated scattering (WSSUS) environ-
ment our wireless channel, composed of M impulses with a certain power delay
profile (PDP), Θ(τl ), is modeled as
M
X −1
hcont. (τ ) = αl δ(τ − τl Ts ), (3.1)
l=0
...
Time
Region 2
Region 1 ...
Frequency
Data Scattered Pilot
Figure 3: An example of time-frequency grid for DVB-H. The scattered pilots

are identified by the black boxes. The solid circles show how pilots from three
consecutive OFDM symbols are collected to construct a pilot vector with a pilot
separation of 6. Similarly, the dotted circles identify the pilots collected from four
consecutive OFDM symbols resulting in a pilot vector with a separation of 3.
where αl are zero-mean complex Gaussian random variables and Ts is the system’s
sampling interval. Furthermore, the channel attenuation on tone k for the above
channel model becomes,
M
X −1
h [k] = αl exp(−2πj(k/N)τl ). (3.2)
l=0
If the subcarriers, including both pilots and data carrying tones, corresponding
to one OFDM symbol are collected into a single column vector at the receiver, we
can model the received vector as,
y = Xh + w (3.3)
where y is the received data column vector of size N, X is a diagonal matrix of
size N × N encompassing the transmitted data, h and w are column vectors of
size N containing the channel attenuations and the independent and identically
distributed (IID) white Gaussian noise values respectively. Moreover, let (.)H denote
the Hermitian transpose, then the 1-D MMSE-based channel estimation in frequency
for the above model is given by the following equation,
−1
1
ĥmmse = Rhhp Rhp hp + I ĥp,ls (3.4)
SNR
where ĥmmse is a column vector of size N containing the estimated channel tones
and hp is a column vector of size Np encompassing the sampled channel at the
pilot positions. Moreover, Rhp hp = E[hp hH p ] is a matrix containing the autocor-

relation values of the sampled channel on the pilot positions, Rhhp = E hhH

p is
the cross-correlation matrix between the pilot tones and subcarriers and SNR is the
signal-to-noise ratio. Furthermore,
ĥp,ls = X−1
p yp (3.5)
is the least squares (LS) estimate of the channel attenuations hp on the pilot tones,
where Xp is a matrix containing the transmitted pilot data on its diagonal and yp
is the column vector of received pilot tones.
To avoid computationally expensive matrix inversions imposed by (3.4), a robust
estimator may be used which has been shown to exhibit acceptable performance
[2]. The robust estimators may be designed for a set of fixed channel estimation
parameters, e.g., SNR and 2nd order channel statistics. For instance, the frequency
correlation, corresponding to the channel model in (3.1), between two subcarriers,
with a spacing of ∆k, depends on the PDP, Θ(τl ), of the channel. Furthermore, for
a uniformly distributed PDP the frequency correlation amounts to [7]
rh [∆k] = , (3.6)
2jπL∆k/N
where L is the delay spread of the wireless channel in samples. It can be seen in
(3.6) that the correlation among subcarriers in a single OFDM symbol depends only
on the distance between the corresponding subcarriers (∆k), the PDP length (L),
and the number of DFT points (N). As we will see in the following sections, the
ratio between the above parameters facilitates co-optimization of filter coefficients
across LTE and DVB-H.
It can be shown [2, 3] that due to correlation loss among the subcarriers located
far apart, using only a few pilots in the vicinity of each subcarrier provides an
acceptable performance in terms of channel estimation. Doing the estimation as
such, not only lowers the computational complexity of the estimator but also reduces
the coefficient memory requirements [3].
In [3] it has been shown that by combining the above mentioned strategy, i.e.,
using a subset of pilots, and the pilots merging scheme, low-complexity estimators,
called merged-pilot modified robust MMSE (merged-pilot MRMMSE) can be devel-
oped. Moreover, these estimators exhibit an acceptable performance in moderately
fast fading environments for LTE [3]. Likewise, the same principle may be applied
to DVB-H [2] with different approaches to pilot merging. Fig. 3 shows how two
different pilot merging schemes can be performed for channel estimation using (3.4).
To construct a pilot vector similar in concept to the one in [3], where the pilot sep-
aration is 3, pilots from four consecutive OFDM symbols are collected and merged
into a single pilot vector. This is illustrated by dotted circles in Fig. 3. The con-
structed pilot vector is further used for channel estimation as well as equalization
of, say, region 2 in Fig. 3. Similarly, if a pilot separation of 6 is desired, pilots from
alternate OFDM symbols, as shown by solid-line circles in Fig. 3, are collected into
a pilot vector. The built pilot vector is further fed into the channel estimator whose
PDP
delay
Design CP CP=delay spread
Figure 4: An example of a wireless channel’s PDP. It is expected that the wireless

channel’s delay spread does not exceed the CP length. The dashed region shows
how part of channel energy, though minimal, is ignored during estimator design.
Table 1: The configurations for which coefficient co-optimization is feasible according

to our proposed scheme.
LTE DVB-H
NDFT Ncp,trunc. Ncp Pilot BW Mode CP Pilot
(normal) (Extended) Sep. (MHz) (k) len. Sep.
1
6-8 2-8 6
128-2048 8-128 N.A. 3 32
1
6-8 2-8 16
3
1
6-8 2-8 8
6
128-2048 N.A. 32-512 3 1
6-8 2-8 4
3
output is used for equalization of the region encompassed by the pilot contributing
OFDM symbols, e.g., region 1 in Fig. 3.
Each pilots merging scheme has its own pros and cons in terms of performance
and computational complexity. An elaborate analysis of the computational complex-
ity versus performance is beyond the scope of this paper. However, manipulating
the pilots spacing as such provides the opportunity for further coefficient memory
optimization discussed in the following sections.
4 Coefficient memory co-optimization

As discussed in previous section and elaborated in [2, 3], using a subset of pilots in
the vicinity of each subcarrier for estimation purposes already reduces the memory
storage requirements for the RMMSE filter coefficients. This may be practiced for
each standard, i.e., DVB-H and LTE in this paper, separately. In other words, using
the above mentioned approach, 6 different MRMMSE filter coefficient sets should
be generated each designed for one specific CP configuration in LTE and DVB-H,
i.e., normal and extended CP in LTE as well as 1/4, 1/8, 1/16 and 1/32 CPs in
4 Coefficient memory co-optimization 85
DVB-H.
To enable further memory optimization across the standards, however, one may
use the innate similarity between LTE and DVB-H. In Sec. 2, it was shown that
by fixing the variables in (3.4),(3.6) the robust estimator is a function of delay
spread of the channel (L), subcarrier spacing (∆k) and DFT length (N). Let the
robust estimators be designed for a uniform PDP with a length equal to that of the
maximum expected delay spread in the wireless channel, i.e., the CP length, then
(3.6) becomes
1 − exp(−2jπNCP ∆k/N)
rh [∆k] = , (4.1)
2jπNCP ∆k/N
where Ncp is the designated CP length.
Considering the fact that pilots from neighboring OFDM symbols, as discussed
in previous sections, may be merged to address the channel estimation needs of the
terminals, various merging patterns may be practiced in LTE and DVB-H to force
the corresponding correlation matrices to be identical. In fact, our intention is to
force
rh1 [∆k] = rh2 [∆k], ∀k, (4.2)
where index 1 refers to LTE and index 2 refers to DVB-H. As a result, the estimators
will be identical for similar values of SNR. By doing so, it is easy to show that it is
required that
∆k1 NCP1 ∆k2 NCP2
= (4.3)
N1 N2
for the estimators in LTE and DVB-H to become identical.
Table 2: System simulation parameters.

Standard N Fc , [GHz] V, [km/h] Modulation Channel model Fading model
LTE 512 2.2 50 4QAM 3GPPTUx [8] 2-D, Jake’s
DVB-H 2048 0.75 10-200 4QAM TU-6 [5] 2-D, Jake’s
One can see that by adjusting the three different variables in (4.3), i.e., N, Ncp
and ∆k, the same coefficients can be shared between the two standards. For in-
stance, by designing the estimator based on the extended CP in LTE, when the
pilot separation is 3, a certain ratio suggested by (4.3) is achieved which enables
reusing the same estimator in DVB-H, across all different BWs and modes of opera-
tion, provided that the system is configured for 1/4 and 1/8 CP lengths. Moreover,
constrained by (4.3), a pilot separation of 3 should be practiced for 1/16 CP while
the pilot separation in case of 1/8 CP should be kept at 6, see Table 1. By doing so
the required number of filter coefficient sets reduces to 4 rather than 6.
Nevertheless, to be able to maximize the coefficient co-optimization, even further,
the normal CP length in LTE needs to be truncated to Ncp,trunc. where Ncp,trunc. = 2i
is the largest value that is smaller or equal to Ncp and is further used as the basis
for the generation of the MRMMSE estimators. These estimators are identified
as truncated MRMMSE (TMRMMSE) throughout this paper. Table 1 shows the
different configurations for which the same coefficients across LTE and DVB-H may
be reused. It can be seen that by designing the estimators for the truncated normal
50 km/h
3gppTUx
0
10
MRMMSE
TMRMMSE
−1
10
MSE
−2
10
−3
10
−5 0 5 10 15 20 25
SNR in dB
Figure 5: LTE channel estimator performance in terms of MSE. The difference in

MSE between MRMMSE and TMRMMSE is minimal. The observed MSE floor in
the high SNR region is because of the channel estimators mismatch with respect to
both fading and PDP.
CP in LTE, the estimators can be shared with DVB-H provided that the system
is set-up for 1/16 and 1/32 CP lengths. Furthermore, a pilot separation of 6 and
3 needs to be practiced for 1/32 and 1/16 respectively. One may practice other
pilot merging schemes for LTE and DVB-H to force identical estimation coefficients.
However, the suggested configuration in Table 1 not only is optimal in terms of
maximum coefficient optimization but also is in-line with [3] in terms of the channel
estimator’s performance in fast fading environments.
Designing the robust estimator for a truncated CP provides a two-fold memory
co-optimization. In other words, it further reduces the number of filter coefficient
sets from 4 to 2. However, it may result in a loss in the channel estimator’s perfor-
mance if the delay spread of the wireless channel exceeds the length of the truncated
CP. Fig. 4 illustrates how part of channel energy is ignored during estimator design
for an exemplary continuous time wireless channel. However, we expect that the
discarded channel energy is minimal for the majority of the wireless channels in
practice, specially because the required truncation is minimal. On the contrary, the
performance of the estimators may even improve if the root mean squared (rms)
delay of the wireless channel is considerably smaller than the CP length. As a re-
sult, due to the mismatch reduction between the actual and design PDP the channel
estimator’s performance for such cases slightly improves.
Estimator performance results for LTE, when the pilot merging scheme with a sep-
aration of 3 is adapted, are discussed in [2, 3]. The provided simulations provide a
good insight into the performance of such estimators when coefficients are generated
50 km/h
3gppTUx
0
10
Perfect channel knowledge
MRMMSE
TMRMMSE
−1
10
BER
−2
10
−3
10
−4
10
−5 0 5 10 15 20 25
SNR in dB
Figure 6: LTE channel estimator performance in terms of uncoded BER.
for a normal CP without truncation. They also depict a detailed picture of the es-
timator’s performance, in different fading environments, when the proposed pilots
merging schemes are practiced.
Fig. 5 shows the MSE performance of the two estimators designed based on
the normal and truncated CP. It can be seen that the MSE difference between the
two is quite small. This is in agreement with the discussion about the minimal
mismatch introduced by truncation in section 4. In fact, TMRMMSE estimators
exhibits a better performance for this specific simulation environment where one
of the standard ITU-R 3G [8] channel models, i.e., 3GPPTUx has been adopted.
This is mainly due to the fact that not only the rms delay spread of the channel is
considerably smaller than the CP length but also there is no resolvable multi-path
component in the discarded area, as in Fig. 4. As a result, the mismatch between
the design and actual PDP is improved resulting in an slightly improved channel
estimation performance.
Fig. 6 shows how the performance of the estimator in terms of uncoded bit
error rate (BER) is affected. In fact, the TMRMMSE BER curve overlays the
MRMMSE which supports the minimal MSE difference observed in Fig. 5. Further-
more, although not included in this paper, the performance of TMRMMSE for all
the different standardized channel models, ITU-R 3G channel models [8], have been
performed and the comparisons show similar behavior. On the whole, unless the
wireless channel in practice has strong taps located in the to-be-discarded region,
as depicted in Fig. 4, the TMRMMSE performance takes after the MRMMSE, i.e.,
without any CP truncation.
Although, no CP truncation is required during DVB-H estimator design, the
estimator’s performance results need to be discussed due to the application of var-
ious pilot merging schemes. As elaborated in [3] for LTE, pilots merging results
in different performance and complexity trade-offs. The impact is specifically more
visible in fast fading environments. In Fig. 7 the performance of the MRMMSE
Mobile Channel, TU−6

25
20
SNR, [dB]
15
MRMMSE, Pilot Sep. 12

Perfect chan. knowl.
10
40 60 80 100 120 140 160 180 200
Velocity,[km/h]
Figure 7: The required SNR to sustain an uncoded BER of 10−2 for a range of
terminal speeds. The simulations have been done for a carrier frequency of 750 MHz
and 4QAM modulation scheme in DVB-H.
estimators for two different pilots merging schemes, i.e., pilot separations of 3 and
6 are illustrated. Although, coefficient optimization for pilot separation of 12, i.e.,
with no merging, is not discussed in this paper, its performance result provides a
valuable reference for comparison purposes.
It is evident from Fig. 7 that different merging schemes provide a certain
gain/compromise in terms of the required SNR for various fading environments.
Besides, there is a significant loss in the estimators performance when the terminal
speed increases beyond a certain limit. However, if the performance constraints for
the fast fading environments are arbitrarily based on the requirements set by LTE,
i.e., an acceptable performance up to terminal speeds of 120 km/h [4], all the merg-
ing schemes proposed in this paper satisfy the requirements of the multi-standard
platforms specified by [9]. On the other hand, the proposed practice in this paper,
enables a three-fold coefficient memory reduction, 2 rather than 6 filter coefficient
sets, when both LTE and DVB-H are co-implemented.
6 Conclusion
We have shown that by employing various pilots merging schemes in DVB-H while
designing the MRMMSE estimators for a truncated normal CP in LTE, it is possible
to optimally share the MRMMSE estimators between LTE and DVB-H. In other
words, having the estimators designed for LTE enables the integration of DVB-
H standard on the same platform without any additional filter coefficient memory
cost. In effect, the on-chip memory requirement for the storage of the robust es-
timation coefficients in a multi-standard platform considerably decreases. For co-
implementation of LTE and DVB-H the on-chip coefficient memory is reduced by
a factor of three. Although, the above mentioned procedure forces specific pilot
7 Acknowledgments 89
merging patterns for a number of configurations in a given standard, the simulation

results show that the performance loss is acceptable for slow as well as moderately
fast fading environments.
7 Acknowledgments
This project was funded by Multi-base, a 7th Framework Program by the European
Commission.
References
standard environment (LTE and DVB-H),” in Proc. IEEE Int. Conf. Signal Pro-
cess. and Commun. Syst., Omaha, NE, pp. 1–9, Sep. 2009.
[3] F. Foroughi Abari et al., “Low complexity channel estimation for LTE in fast
fading environments for implementation on Multi-Standard platforms,” in Proc.
IEEE Veh. Tech. Conf., Ottawa, pp. 1–5, Sep. 2010.
info/36201.htm.


[8] “Guidelines for Evaluation of Radio Transmission Technologies (RTTs) for IMT-
2000,” International Telecommunication Union, RECOMMENDATION ITU-R
M.1225, 1997.

Channel estimation filter-tap
memory optimized pilot
pattern design for OFDM
systems in multi-mode and
multi-standard environments
Farzad Foroughi Abari, Fredrik Rusek, and Ove Edfors
Technical Report
Reprinted, with permission, from Farzad Foroughi Abari, Fredrik Rusek, and Ove
Edfors, “Channel estimation filter-tap memory optimized pilot pattern design for
OFDM systems in multi-mode and multi-standard environments,” Tech. Rep., Dept.
of Elect. and Inform. Tech. (EIT), Lund Univ., Lund, Sweden, 2011.
1 Introduction 93
Abstract
As more and more OFDM standards are integrated in one single plat-
form, there is a growing interest to adopt low-complexity signal processing
algorithms to lower the hardware requirements. A challenging part of many
transceivers is the channel estimation block. As a result, it is desired that the
number of necessary channel estimators is minimized while addressing sev-
eral OFDM standards in one platform. In this paper, we have shown that
by following a specific strategy in pilot pattern design for OFDM systems in
multi-standard environments the number of OFDM systems sharing a lim-
ited number of channel estimation filters can be maximized. Besides, we have
shown that by pursuing a variant of the proposed strategy, pilot patterns com-
patible with an already existing standard of interest, can be designed. This is
specially of significant importance where a number of OFDM systems need to
be co-integrated with an existing/standardized one such as LTE.
1 Introduction
The wireless communications entered a new era when orthogonal frequency divi-
sion multiplexing (OFDM) became practically feasible. Although the main concept
was introduced in the 1960s [1, 2], its application was mostly limited to military
applications due to the required sophisticated radio frequency electronics. However,
as digital electronics matured, OFDM could be realized through fast and efficient
digital circuits. OFDM provides many advantages over its single-carrier commu-
nication counter part. One of its main advantages is dividing the available band
width (BW) into narrow band channels such that inter symbol interference (ISI)
can be mitigated. It also provides the possibility for implementation of simple fre-
quency equalization. Simple frequency equalization can also be realized through
single carrier with frequency domain equalization (SC-FDE) [3]. However, OFDM
and SC-FDE each have their own pros and cons making them suitable for specific ap-
plications. Besides, OFDM is mostly used in the downlink of a radio communication
while SC-FDE is used in the uplink.
The opportunity for simple frequency equalization, provided by the OFDM, can-
not be fulfilled unless channel state information (CSI) is available at the receiver.
Taking the fact that it is usually cumbersome or even impossible to feed the CSI
to the receiver, it needs to be estimated at the receiving side. As a result, channel
estimation becomes an inevitable part of any OFDM transceiver architecture. There
are many approaches to channel estimation. The methods cover a wide range from
blind estimation methods [4–7] to low-complexity approaches suitable for low-power
mobile terminals. Blind channel estimation, channel estimation in the presence of
channel coding, and minimum mean square error (MMSE) pilot-aided channel esti-
mation are examples of various existing approaches.
A common and widely acceptable way of performing channel estimation is to
use data known to both the transmitter and the receiver, i.e., pilot data. The
pilots are either contiguous, meaning they appear in the form of a preamble such
as in IEEE 802.11n [8] or are distributed over the time-frequency grid in OFDM
Technical Report: Channel estimation filter-tap memory-optimized pilot pattern
94 design for OFDM systems in multi-mode and multi-standard environments. . .
systems. The distribution of the pilots is based on many different factors such as
spectral efficiency and transmit power requirements. Moreover, the pilot pattern
may depend on the channel estimation algorithm. For instance, in joint channel
estimation and data detection the pilot density could be much lower than what is
required by the conventional linear estimators [9].
A lot of research has been carried out in the area of pilot pattern design for
OFDM systems [10–12]. One could refer to [13] where an intuitive description of
2-D pilot pattern design has been provided. The parameters of interest for pilot
pattern design are are associated with the most encountered issues in single standard
transceivers. Instances of such issues are the effect of pilots on spectral efficiency, the
peak to average power ration (PAPR), and signal to noise ratio (SNR). Meanwhile,
being a multi-variable optimization, the design should provide a proper balance
between estimation performance and pilot overhead. Moreover, when designed for a
practical implementation on systems with limited hardware resources, e.g., mobile
terminals, it should provide the possibility for proper application of low complexity
signal processing algorithms. Thus, depending on the target system requirements,
a number of different patterns may be adopted.
However, little research has been carried out to probe the pilot pattern design
for OFDM systems in multi-standard environments. As more OFDM systems are
gradually standardized for wireless communications, this topic deserves a more in-
depth investigation. Besides, the smart mobile terminals of today have to handle
several streams associated with different standards simultaneously. For instance, a
typical terminal could be hooked up to Long Term Evolution (LTE) network while
streaming multimedia through Digital Video Broadcasting-Handheld (DVB-H). It
is expected that more OFDM standards enter the wireless market in future. Thus,
mobile terminals have to support them to be able to stay on the competing edge.
On the other hand, mobile terminals have their own restrictions in terms of
hardware resources and power requirements. The hardware and battery technology
have not been able to keep pace with the intensive signal processing algorithmic
developments. As a result, it is but inevitable to dig into low-complexity alternatives
for the demanding and greedy algorithms. This has motivated, e.g., the research
results in [14–16] within Multi-base [17] project. The results indicate that there is
a good chance of exploiting the similarities in the existing OFDM standards such
as LTE and DVB-H so the available hardware resources can be shared among the
channel estimation blocks. For instance, it has been shown in [16] that by using the
similarity in the available pilot pattern in LTE and DVB-H a significant reduction,
in terms of the required on-chip memory for the storage of channel estimator filter
taps, can be done. In other words, a few filter sets can be reused in LTE and DVB-H
across their different modes of operation.
However, we have observed that the low-complexity channel estimators proposed
in [14, 15] could provide a higher performance in fading environments if the pilot pat-
tern was designed differently. For instance, with some modifications to the available
DVB-H pilot pattern as elaborated in Sec. 5.2, one could enhance the performance
of the proposed channel estimators in fast fading environments. Besides, one may
design the pilot pattern in future OFDM systems such that maximum compatibility
2 System model 95
Ant # T1 Ant # R1
Binary Data Wireless CSI

fc channel fc
DMUX
MUX
FFT
DMUX
IFFT
MUX
Interleaving
CP D/A ~ A/D CP
Deinterleaving
Decoding
+
Coding
+
Modulation Demodulation
Figure 1: The adopted SISO OFDM system model.
in terms of the proposed low-complexity channel estimators [14, 15] may be achieved
with respect to other OFDM standards. This is of significant importance in low-
power multi-standard environments where the hardware resources should be utilized
as efficiently as possible.
In this work, we will show a strategy for pilot pattern design in emerging OFDM
standards. The main criteria for the proposed design strategy focuses on pilot pat-
tern design for maximum compatibility with other OFDM standards, with respect
to channel estimation, as elaborated above. Meanwhile, the suggested strategy does
not restrict the estimation algorithms to the ones proposed in [14–16]. The proposed
pilot pattern strategy does not confine the possible estimators to MMSE estimators
only and other linear estimators can be invoked, if needed. In other words, it will
follow the Nyquist sampling criteria with a sufficient margin [13] to sample the avail-
able signal space to create a pilot subset. Moreover, it is expected that some of the
recently proposed OFDM standards such as LTE will prevail in the wireless mar-
ket for the foreseeable future and may remain as an integral part of multi-standard
platforms. Thus, we will also propose a pilot pattern design strategy for emerg-
ing OFDM systems provided that they remain compatible with an existing OFDM
standard, e.g., LTE.
We have adopted the following structure in this report. First, an OFDM sys-
tem model will be provided where our adopted transceiver structure, including the
wireless channel, is characterized. Then, a brief introduction to the low-complexity
MMSE channel estimation is given. This is the basis on which the pilot patterns
will be designed. In Sec. 5, we will elaborate our design strategy for pilot pattern
design in multi-standard OFDM environments. Finally, Sec. 6 will wrap up the
report through some concluding remarks.
2 System model
Our adopted system model is a single-input single-output (SISO) OFDM system
operating in a fading and frequency-selective wireless environment. We have chosen
a SISO OFDM model because it provides a simple approach for description of the
concepts discussed in this paper. However, the results may be easily extended
to multiple-input multiple-output (MIMO) OFDM systems. Fig. 1 illustrates a
simple transceiver structure for a typical OFDM system operating in a wireless
environment.
Presuming the wireless channel to be composed of M resolvable paths in a wide

sense stationary uncorrelated scattering (WSSUS) environment, it can be mathe-
matically described as
M
X −1
hcont. (τ, t) = αl (t)δ(τ − τl ), (2.1)
l=0
where αl s are independent zero-mean complex-valued Gaussian random variables,

with Rayleigh distributed amplitudes, and τl is the arrival time associated with the
lth component. Moreover, the 2nd-order channel statistics associated with (2.1)
in discrete-time can be described through the following separable autocorrelation
function
rh [k − k ′ , m − m′ ] = rh,f [k − k ′ ]rh,t [m − m′ ] (2.2)
where rh,f is the frequency correlation and rh,t is the time correlation. Moreover,
depending on the 2nd-order channel statistics in (2.2), the channel is characterized
by a maximum Doppler spread, Dp , and delay spread, Ds .
Following the signal flow depicted in Fig. 1, data corresponding to one OFDM
symbol are inserted on the designated tones at the transmitter and fed into an N-
point IFFT. Then, the GI (CP) is inserted at the output of the IFFT block and after
going through digital-to-analogue (D/A) conversion and modulation to the proper
carrier frequency the outcome is transmitted over the wireless channel modeled by
(2.1).
Upon reception at the receiver, the signal is picked up by the receiving antenna
and after carrier demodulation, low-pass filtering, analogue-to-digital (A/D) conver-
sion, and CP removal, the result is fed into an N-point FFT. If the signal samples
after the FFT block are collected into the column vector y of size N, then
y = Xh + η, (2.3)
where X is a N × N diagonal matrix containing the transmitted symbols, h corre-

sponds to the N × 1 vector of subchannel attenuations, and η is the N × 1 vector
of additive white Gaussian noise (AWGN).
3 Pilot-pattern design in OFDM systems

To be able to capture the CSI in a communications system, one often needs to rely
on the (noisy) observations of the received signal values. Moreover, the estimator,
e.g., at the receiver, usually needs to know the transmitted signal values to estimate
the channel. Yet, there are classes of estimators, known as blind estimators, which
rely only on the received signal and its pseudo-stationary properties to carry out the
estimation [4–7]. Their application, however, is prohibitive in low-power terminals
with limited processing resources due to their high computational complexity.
In the majority of practical cases the transmitter and receiver should have a pri-
ori knowledge of the transmitted data for the purposes of channel estimation. Thus,
the transmitter and receiver agree on a subset of data which is called pilot-data
3 Pilot-pattern design in OFDM systems 97
Figure 2: An example of rectangular pilot pattern scattered in time and fre-

quency.
by convention. This approach is extensively used in practical system implementa-

tions. The pilots are inserted in specific locations in the transmitted signal. They
appear either in continuous or scattered patterns. The contiguous scheme is usually
employed in wireless systems where little fading is expected. For instance, IEEE
802.11n [8] uses continuous transmission of pilots for the purposes of channel esti-
mation in its preamble. Contrary, if the wireless channel experiences a considerable
amount of fading, e.g., movements in the environment, scattered pilots distributed
both in time and frequency need to be used so that the channel variations over both
time and frequency can be tracked.
An example of scattered pilot pattern in OFDM systems is depicted in Fig. 2.
This is known as rectangular pilot pattern where pilots appear in specific inter-
vals, both in time and frequency. Furthermore, the pilots associated with different
OFDM symbols appear exactly on the same frequency tones. On the other hand,
the scattered pilots do not necessarily need to appear on the same tones in each
pilot carrying OFDM symbol. In other words, the pilots of a specific pilot-carrying
OFDM symbol may experience a distinct frequency shift. This is known as skewed
pilot pattern. If the pilot carrying OFDM symbols are arranged such that the pilot
frequency shift increases monotonically with respect to the symbol time index, a
distinct skewed pilot pattern, shown in Fig. 3, can be achieved. This is known as
primary skewed pilot pattern in this report.
The scattered pilots can be viewed as a 2-D sampling of a time-frequency grid.
Thus, Nyquist’s sampling criteria, in time and frequency, needs to be met if a proper
reconstruction of the original channel is desired. Thus, the sampling can not fall
below a certain threshold set by the Nyquist’s sampling criteria as well as certain
system dynamics such as the presence of AWGN. On the other hand, the pilots
Figure 3: A primary skewed pilot pattern. The pilot frequency offset increases
monotonically with respect to the symbol index time. α and β are the distances
between pilots in frequency and time respectively. γ is the pilot frequency offset and
Γ is the distance between two consecutive pilot-carrying OFDM symbols.
reduce the spectral efficiency in communication systems since a portion of the BW

needs to be allocated to pilot rather than data transmission. Besides, although
increasing the number of pilots results in better estimation performance, the overall
SNR scales down proportionally. Thus, a pilot pattern should provide an acceptable
balance among different optimization criteria.
Although there are various approaches to channel estimation for OFDM systems,
we have resorted to the application of low-complexity MMSE variants suggested
in [14, 15]. A full MMSE estimator can be mathematically described as,
−1
I
ĥmmse = Rhhp Rhp + ĥp,ls , (4.1)
SNR
where SNR is the signal-to-noise ratio, Rhhp is the cross-correlation matrix between
the pilot and data tones, Rhp hp is the autocorrelation matrix of the pilot tones, and
ĥp,ls is the least squares (LS) estimation of the channel on the pilot tones. It has been
shown in [14, 15] that the estimators can be designed for a worst-case channel envi-
ronment, i.e., rectangular Doppler spectrum and power delay profile (PDP) where
the Doppler and delay spreads are attributed to the maximum expected values.
Meanwhile, due to the correlation loss among the pilots and subcarriers which are
located far apart, only a subset of pilots in the vicinity of each tone can be used for
estimation purposes. This already results in a significant reduction in computational
complexity. To avoid additional time-domain filtering, the estimation is primarily
done in frequency direction. To capture the variations of the wireless channel in slow
and moderately fast fading environments, a pilot merging practice, only available for
skewed pilot patterns, has been suggested in [14–16]. In fast fading environments
an additional filtering may kick in to compensate for the time variations of the
channel. It can be observed that both alternatives provide a pilot vector which
may have different pilot spacing, depending on the merging or time-domain filtering
practice. For instance, Fig. 4 shows how a pilot separation of 3 can be attained if
all the pilots from all the pilot-carrying OFDM symbols, shown in the figure, are
merged or interpolated. A pilot spacing of 6 may be achieved if pilots from alternate
pilot-carrying OFDM symbols are merged or interpolated.
On the other hand, it has been shown in [18] and discussed in [16] that the
frequency correlation values when the estimators are designed for the worst-case
channel environment, i.e., a rectangular PDP, amount to
rh,f [∆k] = , (4.2)
2jπL∆k/N
where ∆k = k − k ′ is the distance between a pilot and estimated tone, N is the

DFT size, and L is the PDP spread in discrete time samples. Thus, the designed
filter taps will be a function of the pilot separation, as delay spread and DFT size,
i.e., δkL/N. Thus, by controlling the pilot separation, e.g., through merging, one
may be able to employ similar filter taps for estimation of the wireless channel. The
same principle has been suggested in [16] to use only two sets of filters to address
the channel estimation needs of LTE and DVB-H across their different modes of
operation. One may observe that if the estimators in [16] were designed based on
the extended CP in LTE, only one filter set could cover the channel estimation needs
of both standards. A detailed description of such approach is not the intension of
this report, though.
5 Filter-tap memory optimized pilot-pattern design

in OFDM systems
The presented approach for pilot pattern design is based on filter-coefficient memory
optimization in multi-standard environments. Following the memory optimization
strategy for the two exemplary standards, i.e., LTE and DVB-H, discussed in [16],
the intension is to propose a new approach towards skewed pilot pattern design. Our
goal in the suggested strategy is not to take all possible optimization criteria into
account. Rather, we would like to shed light on some parameters of interest which
can be taken into consideration during pilot pattern design for low-power terminals,
specially in multi-mode and multi-standard environments.
The presented ideas in this report for pilot pattern design, with respect to MR-
MMSE filter-coefficient optimization, cover two different scenarios. First, pilot pat-
tern design for an arbitrary OFDM systems is discussed in Sec. 5.1. The intension
is to provide maximum flexibility in terms of pilot merging schemes such that simi-
lar MRMMSE channel estimators can be used while the number of similar OFDM
systems selected for co-integration is maximized. Then, it will be shown in Sec.
5.2 that pilot pattern for an arbitrary OFDM system may be designed such that
its integration in a multi-standard platform, where one standard is already fixed,
makes it possible to employ similar MRMMSE channel estimators. Considering the
fact that LTE has been opted as the standard of choice for high speed communica-
tions, it is expected that it will be widely deployed and probably replace many of
existing commercial wireless communications systems. Thus, pilot pattern design
for an exemplary OFDM system, compatible with LTE, will finalize Sec. 5.2.
5.1 Pilot-pattern design for an arbitrary OFDM system

Consider a skewed pilot pattern specified by α, β, γ, and Γ. An example of such
pattern is the primary skewed-pilot pattern shown in Fig. 4. Following the merging
schemes practiced and elaborated in [14–16], we have restricted ourselves to the ones
where an even pilot spacing in the constructed PV can be achieved. As a result, the
only skewed patterns of interest are characterized by α/γ = k, where k is a positive
integer. In other words, γ has to be a divisor of α. On the other hand, it can be
observed that the number of possible merging schemes equals α/γ for such systems.
Thus, the smaller the value of γ, the higher the number of possible merging schemes.
On the other hand, due to efficient FFT/IFFT implementations, the majority

of practical OFDM systems are designed such that N = 2v , where v is a positive
integer. Also, it can be observed from a number of standardized OFDM systems
that the same principle holds with respect to the respective GIs. In other words,
the employed GIs in these systems satisfy
NCP = 2x , (5.1)
where x is a positive integer. For instance, all the GIs in DVB-H as well as the
extended one in LTE satisfy (5.1). In the event of normal GI in LTE, where (5.1)
is not held, the estimators need to be designed for a truncated or extended GI such
that Ldesign = 2x for maximum coefficient memory optimization [16]. Moreover, the
standardized GI in Worldwide Interoperability for Microwave Access (WiMAX) [19]
follows (5.1). Thus, it is a valid assumption to restrict the proposed pilot pattern
to OFDM systems where both N and GIs are powers of 2.
Taking the above mentioned facts into account, the optimal values for α and
γ are restricted to the powers of 2 only. Besides, unless prohibited due to fading
characteristics of the channel, γ = 2 is selected as a default design value. Selection
of γ = 2, as the smallest divisor of α, maximizes the number of possible merging
schemes which in turn increases the configurations for which similar MRMMSE
Figure 4: An example of pilot-merging practiced for a primary skewed pilot

pattern. The pilot vector to the left is constructed by merging all the available
pilots in the tile and exhibits a pilot separation of 2. The merged-pilot vector to the
right, however, is constructed by merging the pilots from alternate pilot carrying
OFDM symbols and has a pilot separation of 4.
filters can be used, this follows (4.2). Thus, the following procedure may be pursued
for pilot pattern design for an arbitrary OFDM system.
• Stage 1:
First, the parameters α, β, γ, and Γ need to be estimated with respect to the
design criteria, i.e., Ds , Dp , ∆f , and N. Taking the Nyquist sampling criteria
with a sufficient margin 1 into account and restricting α to powers of 2 only,
where
α = max 2q (5.2)
q
and
1
α6 , (5.3)
2Ds ∆f
it is trivial to conclude that
α = 2⌊log2 2Ds ∆f ⌋ .
1
(5.4)
Likewise, restricting β to the powers of 2 only, constraint to the adapted

sampling criteria, j k
∆f
log2
β=2 2Dp
. (5.5)
1
The sampling is set to twice as fast as the one required by the Nyquist sampling criteria. It has
been shown that [13, 20] it provides an acceptable balance in the channel estimation performance
versus pilot overhead for band-limited channels in the presence of AWGN.
Figure 5: An example of an elementary tile constructed from Fig. 4. Please note

that the pilot indexes in frequency monotonically increases as a function of symbol
time.
γβ
Besides, Γ should
be
l a positive
m integer such that Γ = α
> 1.Thus, γ = 2i ,
where i = max 1, log2 αβ .
• Stage 2:
Following the design parameters derived from Stage 1, suppose a primary
skewed pilot-pattern, similar to Fig. 3, is created. The pattern is arranged
such that the pilot frequency shift increases monotonically with respect to
the OFDM symbol time index. Although the constructed pattern provides an
appropriate alternative from a coefficient-memory optimization principle, one
may rearrange the indexes for the pilot-carrying OFDM symbols such that a
more optimized pattern with respect to channel estimation performance, in
the event of merging in fast fading environments, is provided.
For instance, if a pilot spacing of 6 in the constructed PV associated with Fig.
4 is desired, the pilots from OFDM symbols 1 and 7 should be merged. As
a result, upon the application of merged-pilot MRMMSE channel estimation
and respective equalization, a considerable performance degradation may hit
the system for fast-fading environments. If, however, the pilot-carrying OFDM
symbols selected for merging were placed next to each other, the performance
degradation would considerably decrease due to higher correlation among the
symbols in time. Meanwhile, the same principle applies in the event of merg-
ing more pilot-carrying OFDM symbols. Thus, a more optimized skewed-pilot
pattern may be sought such that upon pilot-merging the performance degra-
dation due to fast fading environments is minimized.
The 2nd stage of our proposed pilot pattern design procedure is illustrated
Figure 6: An example of a rearranged elementary tile constructed from Fig. 4.

Please note the difference in pilot pattern when compared to Fig. 5.
by the flow-chart shown in Fig. 8. However, it is worthwhile to provide an

insight into the operation of the proposed flow-chart by designing an alterna-
tive pilot pattern for the example shown in Fig. 4. Let’s start by collecting
the first four pilot-carrying OFDM symbols and remove the non-pilot carry-
ing tones. The result is a tile (called an elementary tile) depicted in Fig. 5.
The resultant elementary tile does not provide the closest distance between
pilot-contributing OFDM symbols when merging is practiced. If, instead, the
pilot carrying OFDM symbols were rearranged such that a modified elemen-
tary tile, shown in Fig. 6, was created, then channel estimation performance
in fading environments, in the event of merging, would improve. Finally, the
modified elementary tile is expanded by inserting a proper number of non-
pilot carrying subcarriers and OFDM symbols such that α = 8 and Γ = 3 are
satisfied. The result is a fundamental tile, seen in Fig. 7, which may be used
to span the whole time-frequency grid for a given OFDM system according to
its specifications.
The design procedure for an arbitrary OFDM system may be initiated by
creating an elementary tile by collecting the first α/γ pilot-carrying OFDM
symbols in time and removing the non-pilot carrying subcarriers associated
with the primary skewed-pilot pattern depicted in Fig. 3. Please observe that
the OFDM symbols are arranged such that the pilot frequency shift monoton-
ically increases with respect to the OFDM symbol time index. As mentioned
earlier, the outcome is known as an elementary tile and is used in the flow-chart
shown in Fig. 8 to finalize the pilot pattern design suggested by this litera-
ture. The flow-chart follows an iterative grid search where at each stage the
best candidates for pilot merging are grouped together. In fact, the algorithm
guarantees the proximity of pilot-contributing OFDM symbols in the event
Figure 7: An example of a fundamental tile constructed from Fig. 4. Please note

the difference in pilot pattern when compared to Fig. 4. To construct a merged-pilot
vector where a pilot separation of 4 is desired, the pilots from two consecutive pilot-
carrying OFDM symbols need to be merged. As a result, the performance of the
channel estimation, in the event of pilot merging, in fading environments increases
when compared to Fig. 4. If a pilot separation of 2 is desired, the channel estimation
performance for both patterns is the same.
of pilot merging. The novelty of such pilot arrangement not only guarantees
the proximity of two-consecutive pilot-carrying OFDM symbols but also pro-
vides the closest distance among pilot-carrying OFDM symbols in case a pilot
merging of higher order is sought. For instance, if four pilot-carrying OFDM
symbols should be merged, our algorithm makes sure the pilot-contributing
OFDM symbols exhibit the shortest distance from each other, when compared
to other possible pilot arrangements.
The discussion about the proposed pilot-pattern strategy described by the flow-
chart in Fig. 8 can be finalized through an example. Assume the proposed approach
in Stage 1 results in α = 16, γ = 2, and Γ = 3 for a given OFDM system. Fig. 9
illustrates the procedure mentioned in Stage 2 and elaborated by Fig. 8 to finalize
the proposed pilot pattern design for this system. It is worthwhile to point out that
the constructed tile can be used to fill in the time-frequency grid of arbitrary size
dictated by the OFDM system requirements and specifications.
Figure 8: Construction of an optimum tile in terms of filter-memory optimized

MRMMSE estimators as well as merging operation. The final pattern looks different
from the primary skewed pilot pattern exemplified by Fig. 3 and 4
Figure 9: An example of tile construction according to the procedure in Stage 2

as illustrated in Fig. 8.
An alternative approach to the application of the flow-chart shown in Fig. 8,

each time a new pilot pattern needs to be designed, is to create a library of mod-
ified elementary tiles. Please note that each modified elementary tile is distinctly
characterized by the number of pilots it accommodates. Besides, the number of
pilots in the fundamental tile for each design equals ( αγ ). Thus, after the proper
design values, i.e., α, γ, and Γ are estimated through Stage 1, the appropriate tile is
selected from the library and expanded, constraint to α and Γ, to create the target
fundamental tile.
5.2 Pilot pattern design for OFDM systems compatible with

a standardized OFDM system, e.g., LTE
With the advent of more standards as well as improvements in hardware technology,
there is a growing interest in integrating more wireless OFDM standards in one
platform. An example of such systems has been elaborated in [17]. Thus, it becomes
but inevitable to finalize the pilot pattern design for OFDM systems when the design
criteria is constraint by an existing OFDM system, e.g., LTE. However, as discussed
in Sec. 5.1 the OFDM systems of interest are restricted to the ones where the
number of DFT points as well as GI lengths are powers of 2. However, for systems
where some GI lengths happen to disagree with the above presumptions, e.g., the
normal CP in LTE, the MRMMSE estimators need to be designed for a truncated
or extended CP. The MRMMSE filter design procedure has been discussed in [16]
with respect to the normal CP in LTE. Besides, to practice the suggested pilot
merging as well as filter-tap memory optimization in the event of co-integration,
αstd. = k × γstd. should hold, where k is a positive integer and subscript (·)std.
denotes the standardized OFDM system parameters.
Imagine a standardized OFDM system with
v
γstd. = 2u r0v0 ...rl−1
l−1
, (5.6)
where {r0 , ..., rl−1 } belong to the set of prime numbers excluding 2 and {u, v0, ..., vl−1 } ∈
Z+ . Furthermore, Z+ is the set of nonnegative integers. As discussed above, the sys-
tem should satisfy αstd. = k×γstd. so that MRMMSE filter-coefficient co-optimization
can be performed. As a result, to design a flexible pilot pattern for the compatible
OFDM system, two design goals should be pursued in parallel. First, the proposed
pilot pattern should allow its co-integration with the standardized system while us-
ing similar MRMMSE filter coefficients. Second, it should maximize the number
of other OFDM systems which can be integrated in the same platform provided
that similar filter coefficients can be used. Moreover, a natural outcome of such
strategy is not only maximizing the number of compatible OFDM systems but also
maximizing the the number of operation modes, in terms of employing different GIs
associated with the same OFDM system.
Similar to Sec. 5.1, the pilot pattern design can be split into two different stages.
During the first stage, the parameters of interest, i.e., α, γ, and γ are estimated.
The second design stage, however, rearranges the locations of the pilot-carrying
OFDM symbols such that a higher performance in the event of merging in fast-
fading environments is achieved. The following procedure can be followed for the
first design stage. Suppose
α = max (γini 2q ) (5.7)
q
v
constraint to (5.3), where q is a positive integer and γini = r0v0 ...rl−1
l−1
. As a result,
j k
1
log2
α=2 2γini Ds ∆f
. (5.8)
Meanwhile, β should follow (5.5). Besides, Γ shouldl be a positive

m integer such that
γβ i α
Γ = α > 1.Thus, γ = γini 2 , where i = max(1, log2 γini β ). The second design
stage is consistent with the flow-chart in Sec. 5.1 as illustrated in Fig. 8.
An example of the above presented approach is pilot-pattern design for systems
compatible with LTE. Being the standard of choice for high-speed wireless communi-
cations, it is expected that LTE gradually replaces the existing standards in market
and remain a dominant one for the foreseeable future. As a result, it is not far
fetched to presume that many multi-standard wireless OFDM platforms integrate
LTE as one of their fundamental alternatives. Thus, the future OFDM systems can
be designed such that maximum flexibility in terms of MRMMSE channel estimators
when co-implemented with LTE is achieved.
For instance, suppose it is desired to design the pilot pattern for systems com-
patible with LTE. Furthermore, the given OFDM system specifications result in
α = 12, γ = 3, and Γ = 1. Thus, an elementary tile, similar to the one associated
with Step 1 in Fig. 10, may be constructed. Moreover, the elementary tile can be
further processed by the flow-chart depicted by Fig. 8. The result is a fundamental
tile which is more efficient in terms of employing MRMMSE estimators.
One might realize that the advantages are not limited to MRMMSE algorithms
only. In fact, in the event of slowly fading environments, the majority of pilot-
based channel estimation algorithms may benefit from merging the nearby pilots,
because not only the number of effective pilots increases but also the resultant pilot
spacing is reduced. Although, one might argue that the same principle holds for a
primary skewed pilot pattern, employing the proposed pilot re-indexing facilitates
the merging scheme for a higher amount of fading associated with a given wireless
environment.
Last but not least, had the pilot pattern for DVB-H been designed according to
the fundamental elementary tile shown in Fig. 10, the performance of the merged-
pilot channel estimation algorithms would have improved for fading environments.
Please observe that the design parameters, i.e., α, Γ, and γ are similar to the ones
standardized for DVB-H.
6 Conclusion
We have shown that it is possible to adopt a different approach for pilot pattern
design if the co-integration of several OFDM systems is sought. The proposed design
6 Conclusion 109
Figure 10: Construction of a fundamental tile for an OFDM system compatible

with LTE where α = 12, γ = 3, and Γ = 1. Compared to the standardized pilot
pattern for DVB-H, the result is more optimized for merging practices.
procedure maximizes the performance of the merged-pilot MRMMSE algorithms.

Meanwhile, it provides the possibility to employ other estimation approaches. The
main merits, however, result from the possibility to change the pilot spacing by either
merging or time-domain filtering. As a result, the MRMMSE filters can be shared
among a number of different standards. Our approach starts with the pilot pattern
design for an arbitrary OFDM system. However, it is shown that it can be modified
and adapted to design the pilots for systems which need to remain compatible with
a given OFDM standard, such as LTE. Moreover, we showed that by following the
proposed strategy, the pilot pattern for an exemplary DVB-H system would look
differently when co-implemented with LTE. In other words, the existing pilot pattern
may not be optimal if pilot merging upon channel estimation is sought. Although
the proposed approach is specifically developed for filter-tap memory optimization
as well as high performance in the event of merged-pilot MRMMSE estimation, it
provides an alternative insight into pilot pattern design for future OFDM systems.
References
[2] R. Chang and R. Gibby, “A theoretical study of performance of an orthogo-

nal multiplexing data transmission scheme,” IEEE Trans. Commun., vol. 16,
pp. 529–540, Aug. 1968.
[3] F. Horlin and A. Bourdoux, Digital Compensation for Analog Front-ends. West
Sussex, United Kingdom: John Wiley & Sons Ltd., 2008.
[4] L. Tong et al., “Blind identification and equalization of multipath channels,” in

Proc. IEEE Int. Conf. Commun., Chicago, IL, pp. 1513–1517, 1992.
[5] L. Baccala and S. Roy, “A new blind time-domain channel identification method
based on cyclostationarity,” IEEE Signal Process. Lett., vol. 1, pp. 89–91, June
1994.
[6] H. Zeng and L. Tong, “Blind channel estimation using the second-order statis-
tics: Algorithms,” IEEE Trans. Signal Process., vol. 45, pp. 1919–1930, Aug.
1997.
[7] W. Songping and Y. Bar-Ness, “OFDM channel estimation in the presence of

frequency offset and phase noise,” in Proc. IEEE Int. Conf. Commun., Anchor-
age, AK, pp. 3366–3370, 2003.
[8] “IEEE Standard for Information technology–Telecommunications and informa-

tion exchange between systems–Local and metropolitan area networks–Specific
requirements Part 11: Wireless LAN Medium Access Control (MAC) and
Physical Layer (PHY) Specifications Amendment 5: Enhancements for Higher
Throughput,” IEEE Std 802.11n-2009, pp. c1–502, Sep. 2009.
6 Conclusion 111
[9] C. Knievel et al., “2-D graph-based soft channel estimation for MIMO-OFDM,”
in Proc. IEEE Int. Conf. Commun., Cape Town, pp. 1–5, 2010.
[10] B. Hassibi and B. Hochwald, “How much training is needed in multiple-antenna

wireless links?,” IEEE Trans. Inf. Theory, vol. 49, pp. 951–963, Apr. 2003.

Mumbai, pp. 71–74, 1997.
[12] F. Tufvesson and T. Maseng, “Pilot assisted channel estimation for OFDM
in mobile cellular systems,” in Proc. IEEE Veh. Tech. Conf., Phoenix, AZ,
pp. 1639–1643, 1997.
[13] P. Hoeher et al., “Two-dimensional pilot-symbol-aided channel estimation by

Wiener filtering,” in Proc. IEEE Int. Conf. Acous., Speech and Signal Process.,
Munich, pp. 1845–1848, 1997.
standard environment (LTE and DVB-H),” in Proc. IEEE Int. Conf. Signal
Process. and Commun. Syst., Omaha, NE, pp. 1–9, Sep. 2009.
[15] F. Foroughi Abari et al., “Low complexity channel estimation for LTE in fast
fading environments for implementation on Multi-Standard platforms,” in Proc.
IEEE Veh. Tech. Conf., Ottawa, pp. 1–5, Sep. 2010.
[16] F. Foroughi Abari et al., “On coefficient memory co-optimization for channel
estimation in a multi-standard environment (LTE and DVB-H),” in Proc. IEEE
8th Int. Workshop on MultiCarrier Syst. and Solutions , Herssching, pp. 1–5,
May 2011.


[19] “IEEE802.16, in part 16: Local and metropolitian local area networks-air in-
terface for fixed and mobile broadband wireless access systems amendment 2:
Physical and medium access control layers for combined and fixed operation
in licenced bands and corrigendum 1,” Broadband Wireless Access Working
Group, Tech. Specifications Final Draft, 2006.

Mumbai, pp. 71–74, 1997.

Low-Complexity and Filter-Tap Memory Optimized Channel Estimation in Multi-Mode and Multi-Standard OFDM Systems

Uploaded by

Copyright:

Available Formats

Low-Complexity and Filter-Tap Memory Optimized Channel Estimation in Multi-Mode and Multi-Standard OFDM Systems

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Low-Complexity and Filter-Tap Memory Optimized Channel Estimation in Multi-Mode and Multi-Standard OFDM Systems

Uploaded by

Copyright:

Available Formats

Low-complexity and filter-tap

memory optimized channel

Farzad Foroughi Abari

Thesis for the degree of

Department of Electrical and Information Technology

Series of licentiate and doctoral theses

Other publications by the author

Farzad Foroughi Abari

• Depending on the fading characteristic of the wireless channel, a number of

description of minimum mean square error (MMSE) family of channel estimators is

2 Wireless OFDM systems

2.1 Narrowband wireless channels

hcont. (τ, t) = α0 (t)δ(τ − τ0 ), (2.1)

more, fD and fc are the Doppler and carrier frequencies, respectively.

2.2 Wideband wireless channels

where h [k, m] is the channel attenuation on discrete tone k at time m.

2.3 Second-order wireless channel statistics

rh [k − k ′ , m − m′ ] = rh,f [k − k ′ ]rh,t [m − m′ ] (2.6)

rh,t [m − m′ ] = J0 (2πfd (m − m′ )) (2.8)

for a 2-D scattering environment meeting certain conditions [9]. In addition, if

rh,t [m − m′ ] = sinc(2πfd (m − m′ )) (2.9)

for a 3-D scattering environment constraint to certain conditions [10]. Furthermore,

2.4 Multi-carrier communications

Figure 2: Frequency correlation versus tone separation, ∆k. It can be observed

signaling. Had the transmission been materialized in a flat-fading wireless channel,

Figure 3: The concept of orthogonal signaling in frequency. The peculiar spacing

where ∆f is the inter-carrier distance between the modulation frequencies, k refers

an inverse discrete Fourier transform (IDFT) and may be written as

where N is the number of samples in an OFDM symbol. DFT/IDFT can be effi-

2.6 System model

• N = number of DFT points, where

Furthermore, other parameters of interest associated with the system can be

Binary Data Wireless CSI

Figure 4: The adopted SISO OFDM system model.

However, it can be shown that following (2.11), s = TH n x, where x is the column

3.1 Pilot-aided channel estimation

Figure 5: An example of rectangular pilot pattern scattered in time and frequency.

An example of scattered pilot pattern in OFDM systems is depicted in Fig. 5.

3.2 Minimum mean square error (MMSE) channel estima-

where ŷp is a Np × 1 column vector containing the observations of transmitted

It can be shown that the solution to the above optimization problem is

J(l, k) = σh2 − rhyp R−1 H

where rhyp is the N(l − 1) + k row in Rhyp .

J(l, k) = σh2 − rhyp wH − wrH H

where w is the N(l − 1) + k row in W.

ĥ1 = Wt yt,p , (3.9)

ĥ1 (l, k) = h1 (l, k) + η ′ (l, k), (3.10)

J1 (l, k) = σh2 − rht yt,p wtH − wt rH H

ĥ2 = Wf ĥ1 , (3.12)

3.3 Low-complexity MMSE channel estimation

wireless channel. Similarly, the covariance matrices in Wt may be designed for a

3.4 Merged-pilot channel estimation strategy

Figure 7: An example of pilot-merging practiced for a primary skewed pilot

a two-fold estimation performance increase in frequency direction is expected once

• Pilot-merging should be practiced unless the performance degradation due to

3.5 Filter-tap memory optimization

rh,f 1 [∆k] = rh,f 2 [∆k], ∀k, (3.15)

the following relation between the pilot separations should hold,

∆k1 = 2∆k2 . (3.18)